ArticlePDF Available

Abstract and Figures

In this design study, we present IRVINE, a Visual Analytics (VA) system, which facilitates the analysis of acoustic data to detect and understand previously unknown errors in the manufacturing of electrical engines. In serial manufacturing processes, signatures from acoustic data provide valuable information on how the relationship between multiple produced engines serves to detect and understand previously unknown errors. To analyze such signatures, IRVINE leverages interactive clustering and data labeling techniques, allowing users to analyze clusters of engines with similar signatures, drill down to groups of engines, and select an engine of interest. Furthermore, IRVINE allows to assign labels to engines and clusters and annotate the cause of an error in the acoustic raw measurement of an engine. Since labels and annotations represent valuable knowledge, they are conserved in a knowledge database to be available for other stakeholders. We contribute a design study, where we developed IRVINE in four main iterations with engineers from a company in the automotive sector. To validate IRVINE, we conducted a field study with six domain experts. Our results suggest a high usability and usefulness of IRVINE as part of the improvement of a real-world manufacturing process. Specifically, with IRVINE domain experts were able to label and annotate produced electrical engines more than 30% faster.
Content may be subject to copyright.
IRVINE: A Design Study on Analyzing Correlation Patterns of
Electrical Engines
Joscha Eirich, Jakob Bonart, Dominik J ¨
ackle, Michael Sedlmair,
Ute Schmid, Kai Fischbach, Tobias Schreck, and J¨
urgen Bernard
Fig. 1. The IRVINE system. Users have an overview over clusters in (A). They can select clusters in (A) and engines in (B). After
selecting an engine in (B), the acoustic signature of the engine is displayed in (C) and respective raw acoustic measurements in (D).
Detailed information about selections from (C) is shown as line chart and scatter-plot and bar chart in (F). After the analysis of an
engine, the user can assign a label in (E) and provide an annotation for the label in (D).
—In this design study, we present IRVINE, a Visual Analytics (VA) system, which facilitates the analysis of acoustic data
to detect and understand previously unknown errors in the manufacturing of electrical engines. In serial manufacturing processes,
signatures from acoustic data provide valuable information on how the relationship between multiple produced engines serves to
detect and understand previously unknown errors. To analyze such signatures, IRVINE leverages interactive clustering and data
labeling techniques, allowing users to analyze clusters of engines with similar signatures, drill down to groups of engines, and select
an engine of interest. Furthermore, IRVINE allows to assign labels to engines and clusters and annotate the cause of an error in
the acoustic raw measurement of an engine. Since labels and annotations represent valuable knowledge, they are conserved in
a knowledge database to be available for other stakeholders. We contribute a design study, where we developed IRVINE in four
main iterations with engineers from a company in the automotive sector. To validate IRVINE, we conducted a field study with six
domain experts. Our results suggest a high usability and usefulness of IRVINE as part of the improvement of a real-world manufactur-
ing process. Specifically, with IRVINE domain experts were able to label and annotate produced electrical engines more than 30% faster.
Keywords: Design study, interactive labeling, interactive clustering.
Index Terms
: H.5.2 [Information Interfaces and Presentation]: User Interfaces—Graphical user interfaces (GUI); User-centered design
J. Eirich, U. Schmid, and K. Fischbach are with University of Bamberg
J. Bonart is with Fraunhofer IWU
M. Sedlmair is with University of Stuttgart
T. Schreck is with Graz University of Technology
J. Bernard is with University of Zurich
J.Eirich, J.Bonart and D.J¨
ackle are with BMW Group
Manuscript received xx xxx. 202x; accepted xx xxx. 202x. Date of Publication
xx xxx. 202x; date of current version xx xxx. 202x. For information on
obtaining reprints of this article, please send e-mail to:
Digital Object Identifier: xx.xxxx/TVCG.202x.xxxxxxx
The automotive industry is currently in the midst of its greatest change
in the last 100 years, namely in the direction of electromobility [28].
With the setup of electrical engines (from here on: engines), of course,
new, previously unknown errors are introduced. For car manufacturers,
it is extremely important to identify and understand these errors to
meet high-quality standards. However, detecting and understanding
such errors is currently hampered by two major challenges:
manufacturing processes produce a large quantity of parts, hence, re-
sulting in a vast amount of recorded sensor data during testing. On-site
interviews with test engineers showed that they are required to analyze
more than forty interdependent signals from hundreds of produced
engines manually and per hour. Furthermore, the measured signals
can be related to multiple sub-components of an engine. This situation
poses a major challenge and as a result test engineers are capable to
only analyze few parts in detail in the given time.
Because only
a few engines can be analyzed and their errors classified in detail, the
gained knowledge is particularly precious for the overall improvement
of the testing procedures. Yet, it is currently unknown how the
gained knowledge can be stored or transferred. To address the named
challenges, we present IRVINE (InteRactiVe cluseterINg labEling),
a Visual Analytics (VA) system that results from a design study [49]
project carried out together with automotive engineers at BMW.
We base our study on acoustic data that is collected through the
propagation of sound inside the engines. The analysis of acoustic data
allows specific errors to be recognized. For example, a dirty bearing
sounds differently from a clean bearing. The interaction between
different acoustic frequencies forms so-called signatures. Signatures
typically comprise one primary and multiple secondary error symptoms.
On top of raw acoustic data, we also include the derived correlations in
acoustic signatures, which enable interdependent error analyses. Given
the collected acoustic data, IRVINE provides capabilities for clustering
similar signature correlations. To gain a fine-grained overview of sim-
ilar signatures, clustering can be performed interactively by engineers
on a subset of clusters or engines with previously provided labels.
Engineers can then select specific engines of a cluster for an in-depth
analysis. To record their findings we introduce knowledge capturing
capabilities through labeling and annotation. In this regard, we refer
to labeling as the error class (e.g. B-Bearing error), which serves as
a basis for the classification of further errors. In contrast, we refer to
annotations as a means to record the reason for an error within the raw
sensor data (e.g. Threshold violation in B-Bearing measurement).
To better support the goals and tasks of engineers, we first familiar-
ized with their domain and gained an understanding of their problems.
The IRVINE system was iteratively designed, developed, and evaluated
in close collaboration with engineers at BMW. The overall goal of the
system is to support engineers in the detection and analysis of error-
prone produced engines. IRVINE was hereby tailored to fit their domain-
specific requirements and designed to be seamlessly integrated into their
daily work. As a result, the systems can handle vast amounts of data
enabling engineers to easily analyze multiple engines, their signatures,
and acoustic raw data. Labels and annotations are stored in a knowledge
base, which is used to both enhance IRVINE and to be available for
different stakeholders, who can benefit from labels and annotations.
In summary, our contributions are: (1) The problem characterization
and abstraction of the studied use case; (2) the reporting of the inter-
active design of IRVINE; and (3) the evaluation of IRVINE together
with six automotive engineers and reflections of our design process.
We start by giving a brief summary on related work about interactive
clustering (Section 2.1), interactive labeling (Section 2.2), and
previous design studies in the automotive sector (Section 2.3).
2.1 Visual Interactive Clustering
Interactive clustering approaches have been proposed for various data
types, including trajectory data [3], text documents [40], (social) net-
works [37], and time series data [8]. Furthermore, other visualization
approaches focus specifically on the visual analysis of sensor data [1]
including a variety of analysis goals, such as segmentation, clustering,
or classification [38]. In scenarios where very large data sets are
analyzed, some approaches make use of matrix-based visualizations,
for instance to group similar matrices that show changes in brain
connectivity networks [4]. With respect to our clustering scenario, we
focus on the grouping of large amounts of acoustic signatures of en-
gines on their similarities, with the help of matrix visualizations. Some
approaches explicitly leverage domain knowledge, e.g. Yang et al. [59],
focusing on interactive steering methods to visually constrained clus-
tering using both user and publicly available knowledge. The selection
of specific subsets of interest where another subsequent computation
of the clustering is applied plays also an important role [42].
Interactive clustering is often combined with dimensionality reduc-
tion techniques [43] to benefit from the complementary strengths of a)
reducing the number of instances from many to a few (via clustering)
and b) reducing multiple data dimensions to a low-dimensional visual
representation (dimensionality reduction). We use the Self-Organizing
Maps (SOM) [32] clustering algorithm, which naturally combines
both steps. Vesanto was one of the pioneers who showed the benefits
of SOMs for visual cluster analysis [55], and many methodological
contributions to interactive clustering followed from VA research.
Schreck et al. [46] presented a VA system that allowed the interactive
visual initialization, quality assessment, and refinement of SOMs.
IRVINE also supports the interactive refinement of SOMs [41], with
both with different parameters and data subsets of interest. The SOM
has also been used for the visual analysis of sensor data, in MotionEx-
plorer [12] to facilitate exploratory search and in FuryExplorer [57] for
the comparison of cluster patterns with attached metadata. Finally, the
SOMFlow system serves as a platform for the interactive exploration
of multiple SOMs, the analysis and refinement of intermediate results,
and the back-and-forth navigation along the analytical workflow [42].
In our design study, we use interactive clustering to support the
engineers in the analysis of acoustic data. Particularly inspiring for the
specific design of IRVINE are the iterative training and refinement, the
training of subsets of interest, and the analysis of individual clusters
and cluster elements in detail for downstream labeling tasks.
2.2 Visual Interactive Labeling
As well as interactive clustering, labeling is also a frequently supported
task in VA [20]. In this regard, Bernard et al. [13] propose the concept
of Visual-Interactive-Labeling (VIAL), to label yet unknown data in
an interactive exploratory setup. VIAL hereby bridges the gap between
active learning [50] and advanced visualization concepts, where the
type of labels depends on the given task and approach. Categorical
labels count to the more common types, which can either be of binary
or multi-valued nature [9]. While binary labels would allow simple
user feedback, such as “ok vs. not ok“, IRVINE supports multi-valued
labels, enabling a more specific tagging of different classes for a data
instance. Other examples for systems that enable categorical labeling
are provided for textual documents [25], bio images [7], handwritten
digits [9], or video streams [27]. Considering our use case, we focus
on the assignment of categorical multi-valued labels.
A second important type of labels is continuous labels, which are
often applied when a more fine-grained degree of interestingness is
required. Here, systems exist for candidate rating and evaluation [56]
or to choose between irrelevant and relevant views [6]. Yet other
labeling approaches allow the comparison of pairs of objects [10] or
groups of objects [16] in combination with algorithmic models using
this implicit feedback to adjust the attribute weightings or the feature
space. Finally, another type of user feedback is to annotate features or
data attributes directly. For instance, approaches exist that support the
dynamic evaluation of feature subsets and resulting models [60] or to
annotate images while relating model results to their input features [34].
With respect to our use case, engineers will be able to annotate local
regions of interest within the analyzed feature space.
Many of the labeling and annotation approaches were built to train
some kind of more or less explainable machine learning model, such
as decision trees [44], or support vector machines [25] to name a few.
However, especially when users are not familiar with machine learning,
previously labeled and annotated data instances, may already provide
valuable information to guide analyses. Thus instead of creating a
“black-box” classifier [39], we will focus on the storage and availability
of labels and annotations to guide user analyses.
2.3 Design Studies in the Manufacturing Sector
Design studies in the automotive sector are mostly carried out for
engineering design and anomaly detection. In this context, efforts were
carried out to visualize the exploration of multi-criteria alternatives
for rotor designs [19], in-car communication networks [47, 48], or
anomaly detection with test stations from large scale manufacturing
processes [21, 53]. While some of the mentioned studies acknowledge
the need for storing expert knowledge [21], the only systems we
found which addresses the problem of specifically leveraging that
kind of knowledge is Cardiogram [48]. Cardiogram addresses the
problem of debugging masses of traces from in-car communications
networks to become error-free. In turn, the problem at hand which
IRVINE addresses is the analysis of acoustic data, which necessitates
different kinds of visualization approaches. Thus, grounded on
previous findings, we did build a system that allows domain experts
to externalize their knowledge from the analysis of acoustic data,
supporting them in their high cognition task of analyzing engines.
During this study, we primarily followed Sedlmair et al.’s nine-stage
framework for design studies [49]. In addition, we used the Nested
Model for visualization design and validation by Munzner [35]. The
nested model guides a more detailed problem characterization, the data
operation and abstraction, the visual encoding and interaction design,
and the algorithm design.
Our system development went through four main iterations, during
which we interviewed engineers, tested design alternatives, and held
critical discussions with visualization experts. Each iteration was
carried out in close collaboration with one engineer at BMW with
extensive experience in the optimization and design of test procedures
for engines. He accompanied the system development with knowledge
about the problem domain (Section 4.2), the data abstraction (Section
4.3), and resulting tasks (Section 4.4). He also gave a constant stream
of feedback on the visual design of our system. The exchange with
the engineer took place on up to four meetings per week ranging from
30-60 minutes. During these meetings, fundamental characterization
and design aspects were discussed and open issues clarified, while he
provided us with feedback about relevant tasks and the system design.
An evaluation of the system was carried out with four engineers at
BMW. Based on the feedback from the evaluation, we refined IRVINE,
which resulted in increased labeling and annotation speed with two
more engineers. Methodological details for this downstream evaluation
are provided in Section 8.
We report the characterization of the problem domain in the field of
quality control in the manufacturing sector. First, we provide an intro-
duction to automated part testing in automotive engineering (Section
4.1), followed by a description of the domain of our collaborators
(Section 4.2). Finally, we report on the task abstraction, supporting
engineers with the analysis of error-prone engines (Section 4.4).
4.1 The Automotive Engineering Domain
The manufacturing of engines requires testing each engine in various
stations along the assembly line. To that end, engineers analyze acoustic
data by recording the noises of engines when being simulated with real-
world conditions on a test bench. With IRVINE, we build upon the com-
mon case of acceleration tests, where a speed ramp is running from 2500
to 14000 revolutions per minute (rpm) to capture a sonic image of the
engine at all relevant rotations. The measurement of these acoustic mea-
surements enables the detailed analysis of engine sub-components (e.g.
the gear or the bearing) and allows to allocate the primary symptom
of an error (e.g. a gear with a scratched surface sounds different from
an undamaged gear). According to the measurements of acoustic data,
engineers can exactly identify the sub-components, which are known
a priori from lab experiments carried out at earlier development stages.
What is challenging for engineers though, is the existence of sec-
ondary error mechanisms accompanying each specific primary error. In
fact, fault mechanisms are visible in the primary error source, but also
partially in their secondary error sources. The engineers reported that
a good understanding of the interplay between primary and secondary
errors sources is extremely valuable to further improve automated part
testing with acoustic measurements. At BMW, engineers have devel-
oped a procedure for the analysis of relations between primary and
secondary error symptoms, based on the systematic correlations of vast
amounts of pairs of measurement values. Aggregated forms of these
correlation results are called the
of an engine. A detailed de-
scription of the signature computation process is outlined in Section 5.1.
Fig. 2. Data analysis workflow of engineers. Engineers are either con-
firmers or explorers with the shared goal to select, analyze, label, and
annotate a single engine.
Based on the observation of engineers and interviews conducted in
the early stages of the project, we abstracted the principal workflow of
engineers for the analysis of engines through acoustic data. In general,
engineers switch between two main goals referring to (1) identifying
engines of interest as well as (2) analyzing, labeling, and annotating
engines. The workflow is shown in Figure 2, offering interesting
nuances with respect to engine identification. Currently, the dominating
information-seeking behavior [29,51] of engineers is of confirmatory
nature. Confirmers already know engines of interest, aiming for the
validation of hypotheses about errors in an engine. In addition, the
characterization of the domain problem in Section 4.2 shows the need
for exploratory analysis support. Explorers are open to large varieties
of engines and respective signatures, aiming at formulating new
hypotheses about unknown engines and signatures. The workflow for
unknown engines roughly ranges from multiple engines, over a small
selection of engines, down to single engines of interest. Both explorers
and confirmers share the same goal downstream: engineers review the
signature of an engine in detail to allocate symptoms and compare it to
the raw acoustic measurements to validate these symptoms. As a result,
a label can be assigned to the engine. Finally, engineers also make
annotations directly in the acoustic raw data to externalize knowledge
about the source of an error.
4.2 Domain Problem
Both the confirmation and the exploration of engines as outlined in
Figure 2 are time-consuming and tedious processes due to the following
two main problems: First, large amounts of available signatures and
recorded acoustic raw data measurements make it almost impossible
for engineers to keep track of all engines at a desired level of detail.
As a result, the detailed analysis, labeling, and annotation of engines
can only be performed in an anecdotal manner. Only the confirmation
of specific engines is supported, while an exploration of unknown sig-
natures from unknown engines remains an open issue. Here, engineers
reported that above all the efficient grouping of similar signatures from
engines is not possible at the moment. The systematic analysis of all
engines and error types falls short. Second, the knowledge, which is
created during the detailed analysis of engines has a high impact on
the improvement of automated part testing procedures. For engineers,
it is especially useful to know, which error was observed in engines
and where the source of the error can be seen in the data. However,
occurred errors in engines are labeled too seldom by engineers, while
the annotation of the cause of an error is not supported at all.
4.3 Data Abstraction
Engineers record their data inside a test bench with a sensor for mea-
suring the noise of the engine. For that purpose, the engine’s rpm is
steered under controlled conditions to analyze the behavior of engines.
Measuring the noise allows the evaluation of the acoustic properties
of components and their technical condition. The result of such acous-
tic measurements is a three-dimensional data structure consisting of
loudness measured across all possible combinations of rpm and orders.
To shift focus on anomalies, the loudness values are often replaced by
residuals, e.g., by the deviation from measured values to mean values
of an ensemble of engines (expected value), divided by the standard
(see also eq.
) A 2D pixel-based visualization of these
measurements often used by engineers is the
, as shown
in Figure 3.
Fig. 3. Spectrogram of acoustic measurements for an engine. Rpm is
shown on the y- and orders on the x-axis. The color indicates the volume
of the measurement compared to the mean distribution of engines.
value is displayed on the y-axis, describing the acceleration
of the engine up to a maximum speed. The
is displayed on the
x-axis, which is the relation between a measured frequency and the
speed of the engine during measurement, consequently describing how
often an excitation occurs per revolution. Orders are beneficial, as they
allow the fine-grained analysis of sub-components through an expert’s
eye. By analyzing the geometry of the rotating engine, engineers are
able to derive the measured sub-component of an engine up to a very
detailed level (e.g. 24th teeth of a gear).
are shown on a
color scale, where brighter colors represent particularly loud orders.
According to the engineers, in serial manufacturing regions of loud
frequencies are a good indicator for different types of errors. The
three-dimensional measurement data forms the basic representation of
an engine through abstract data, providing the raw data for any analysis
scenario on engine errors in IRVINE.
Figure 3 shows a residual order-spectrogram. Specifically, 512
order lines (each column in Figure 3) ranging from 2500 - 14000
rpm are recorded for each engine. To identify signatures in order
combinations that might result in a faulty engine, theoretically every
possible order combination has to be analyzed manually for each engine.
For engineers, this would result in 262,144 possible combinations,
which is not feasible for a manual analysis. Hence, the engineers
narrowed down the orders to the 41 most relevant ones. Based on the
informed selection of orders, they can be connected to the seven main
sub-components of an engine (rotor-shaft-, electromagnetic-, first gear-,
second gear-, A-bearing-, B-bearing-, and C-bearing orders). Knowing
the connected sub-components for individual orders allows the fine-
grained analysis and labeling of engine errors at a sub-component level.
4.4 Task Abstraction
The intensive collaboration with engineers at BMW helped us to under-
stand the domain, the domain problem, and the desired workflow for
the analysis.In the following, we present the task abstraction that will
support engineers in reaching their goals. Likewise, these tasks will
serve as primary design targets for IRVINE.
Gain overview of engines: By taking the role of explorers, en-
gineers need a structured overview of the data, in our case of engines
represented by their acoustic signals. Grouping engines by the similar-
ity of acoustic signals using clustering algorithms would be desirable.
In addition, interactive grouping to adapt to the information need of
individual engineers would be useful, e.g., based on different steering
parameters (such as the number of groups) or filtered subsets of engines.
Drill-down to engines: Engineers exploring large numbers of
engines need support for the drill-down to engines of interest. The
information need may differ between clusters of engines, the labeling
status, or the degree of the anomaly of engines.
Identify engine of interest: The workflow of both explorers and
confirmers includes the identification of single engines as an entry point
into an in-depth analysis. Identification may be through knowledge
about a particular engine, by special or even unique acoustic signatures,
or by traversing several (similar) engines in the exploration process.
Analyze single engine: Engineers need to analyze individual
engines in detail for being able to assign labels on a profound basis.
Single engines are both represented by their acoustic signature and the
raw acoustic measurements. The analysis is also supported by stored
domain knowledge from the systems knowledge base.
Assign label to engine: Engineers need to assign labels of error
categories. In close collaboration, we formed a default alphabet of
label categories as follows: Error/Electromagnetic-Field,Error/First
Gear,Error/Second Gear,Error/A-Bearing,Error/B-Bearing,
Error/C-Bearing,No error. In addition, engineers also expressed the
need to leave the label alphabet open for modifications, as knowledge
about error variations will constantly grow as IRVINE is used. Finally,
to enhance efficiency in combination with
, it would be desirable
to label single engines but also multiple similar engines at a glance.
Annotate acoustic measurements: The cause of a labeled engine
can be annotated by marking the respective region inside of the acoustic
raw data of an engine (the spectrogram) serving two purposes. First,
annotations can be used to review how similar labels were annotated by
different users. Second, a sufficient amount of annotations for a given
label will allow building thresholds for semi-automatic error detection.
In this section, we present the measurement of acoustic data and
pre-processing steps, carried out to prepare the data. Based on the data
abstraction (Section 4.3), we describe the computation of an engines
signature (Section 5.1). Next, we describe details on the feature
extraction and model training for our clustering approach (Section 5.2).
5.1 The Computation of Signatures
To analyze relations between primary and secondary symptoms of
errors, engineers calculate correlations in-between spectrograms. The
motivation is given in Section 4.3 and the schematic process for each
engine is outlined in Figure 4. An exemplary process is described for
a single signal pair (signal A and B) and one engine to increase clarity.
This process is consequently applied to all available signals and engines.
Fig. 4. Computation of our Hypermatrix. From a given spectrogram in (A),
we extract two columns that are correlated with each other in (B). This is
done for each signal pair over all engines in (C). Next, we subtract the
resulting mean signal combination in (D) from each signal combination
of a single engine which results in the deviation of a signal pair from one
engine to all other engines in (E). Each signal pair is then aggregated
and stored in a new matrix in (F).
As a first step, we calculate the mean and standard deviation for
the ensemble of engines at hand. Consequently, the relative deviation
of the
-th engine to the mean
for each (rpm, order)-tuple is then
expressed in units of the standard deviation σ:
Residual(ord,r pm)i=value(ord,rpm)iµ(ord,rpm)
σ(ord,r pm).(1)
An example of the resulting residual spectrogram described in
Section 4.3 is given in (A). Next, we extract two measurements A
and B (being 1D- curves each), and calculate the outer product for
the pair, resulting in a 2D-matrix (see (B)). The choice of possible
pairs is restricted to 41 relevant orders from the data abstraction,
which are used to derive relevant order combinations. Calculating
the mean value of the 2D-matrices over all engines for each entry
(C) results in the correlation matrix (D). This correlation matrix
effectively consists of Pearson correlations for pairs of rpm-values
of two extracted orders. Therefore, the resolution regarding different
engine speeds and the corresponding orders is still retained. To extract
the difference in the correlation, we subtract the correlation matrix
from each outer product resulting in a matrix describing correlations
inside each measurement (E). This matrix is then reduced onto its
cumulated deviation using the Frobenius-norm. Consequently, each
cumulated deviation for each pair of orders is then ordered into the
reduced-Difference-Correlation-Matrix, which we call Hypermatrix.
Figure 5 shows three exemplary Hypermatrices of different engines
and their corresponding signature patterns. The first belongs to an
OK engine with no errors. The second represents an engine with
an anomaly in the B-Bearing but no error and the third an error in
the B-Bearing. Each matrix was build with 41 signals, provided
by our domain expert, which can be related to one of an engine’s
sub-components as described in the data abstraction. The region for
each sub-component was marked by our domain expert with black
lines. Hence, secondary error sources in sub-components can be
identified by reviewing correlations in each region. Especially, in the
second and third Hypermatrix one can clearly see which order is the
primary error source in the signature and which orders resonate with
the excitation thus representing secondary error sources.
Fig. 5. Hypermatrices showing acoustic signatures. (1) with no error, (2)
an anomaly in the B-Bearing, and (3) an error in the B-Bering.
5.2 Feature Engineering and Model Training
To group similar acoustic signatures, we extract features from each en-
gine’s Hypermatrix. The feature extraction process is shown in Figure
6. The combination of all seven sub-components results in 28 possible
combinations and is depicted as Region (R) in Figure 6. From each R,
the sub-matrix is extracted. After experimenting with different feature
sets, we received the best clustering results extracting the maximum of
each of the 28 resulting matrices. This results in a 28x1 feature vector.
The engineers agreed on the usefulness of this approach, as in their
application domain, louder noises tend to be the cause for an error.
However, different features can also be used as input by applying
minor changes to the feature extraction process.
To cluster similar Hypermatrices, the 28x1 feature vector is
used as input for a Self-Organizing Map (SOM) [32], as it nicely
combines clustering with dimensionality reduction functionality. For
the computation of the SOM, we follow a simplified version of the
standard training process described by Kohonen [31]. We set the
SOM grid size such as to expect at least one data vector per node,
which accounts for very specific error types. We initialize the SOM
prototype vector dimensions with random numbers between 0 and 1
and train the SOM by iterating over the input data vectors and adjusting
the SOM nodes. Specifically, we find for each input data vector the
best matching SOM prototype unit (BMU) according to Euclidean
distance. We then adjust the BMU and its neighborhood according
to a linearly decreasing learning rate and circular neighborhood kernel.
This configuration comprises the initial implementation of our SOM
and can be changed by the user as described in detail in Section 6.2.
We note that our heuristic setting of parameters already gave us robust
results for our application, hence, we did not see the need for parameter
optimizations. In principle, also other visual clustering techniques may
be applied besides SOM. We particularly chose SOM because of its
robustness in our application domain, and as it gives an overlap-free
rectangular layout that well supports visual comparison tasks.
Fig. 6. Feature engineering process. All sub-component combinations
in the in the Hypermatrix form 28 Regions (R) of interest. From each R,
the sub-matrix is extracted and from each sub-matrix, the maximum is
computed. All maximas of regions form the input vector for the SOM.
Figure 1 shows an overview of the main views of IRVINE, as also out-
lined in Section 6.1. Each view is marked from A-E, where we will use
this notation hereafter to refer to IRVINE’s individual views. Due to non-
disclosure agreements with BMW, we are not able to show engine ids.
6.1 Overview of the System
In (A), all clusters of similar Hypermatrices as outlined in the Sec-
tions 5.1 and 5.2 are displayed. Each Hypermtarix in the grid is rep-
resented by the mean aggregation of all Hypermatrices in a cluster
of engines. The grid view serves to get an overview over all clus-
ters and their individual properties (
) and to support the decision
of which cluster to select (
). To visually encode a Hypermatrix,
we use a sequential color scale. This is appropriate since all Hy-
permatrix values range from 0-1, where 1 represents a high corre-
lation between a pair of sub-components and 0 a low correlation.
Each grid in the cluster view also contains three
additional rectangles. The first shows the number
of engines in a cluster. The second represents the
number of already labeled engines as a pie chart,
where green are labeled engines and grey not labeled
ones. The aggregated deviation of all engines in a
cluster to the serial distribution of signatures from all available engines
is displayed as two colored glyphs. Here, red arcs represent engines,
which contain a deviation greater than zero and blue lower zero. This
separation is important because engines that are deviating upwards
are louder and downwards are quieter. Hence, users get an immediate
overview of how many engines are in each cluster, how many engines
are already labeled, and which cluster contains the most anomalous
engines. In Figure 1 the selected cluster is shown by its grey stroke.
(B) shows engines in a cluster, the Hypermatrix of each engine as
small multiple, and the aggregated deviation of the signature of a single
engine to all other engines. This view is designed to get an overview
of the engine’s properties and compare them to other engines in the
cluster. Thus, it supports the user in selecting an engine (
). If the user
has a specific engine id of interest the engine list can also be filtered
accordingly and thus allows for confirmation as shown in Figure 2. The
Hypermatrix of each engine is represented in the engine list view as a
small multiple, while the same glyph representation as in the cluster
view is used. Initially, the list is sorted according to the euclidean
distance of engines to their centroids in each cluster. However, the user
can sort the list according to the deviation of engines greater than zero.
In Figure 1 the selected cluster contains 9 engines, sorted according to
their anomaly score and contains similar Hypermatrices. By clicking
on “Zoom in Cluster” a detailed overview of all Hypermatrices in the
cluster replaces the view in (A), which is demonstrated in Figure 8-1.
Of course, users have the option to return to the initial view. The user
can also retrain clusters by clicking on “Train new SOM”. The process
to train clusters is outlined in detail in Section 6.2.
In (C), the detailed Hypermatrix of a selected engine is shown.
Additional information about single selected cells in the matrix is
displayed in the upper left triangle. This view is designed to support the
user in the analysis of a selected engine (
). A matrix representation
is adequate to represent the relation between pairs of sub-components
(e.g. Gear and rotor shaft). As pointed out in Section 5.1, the regions of
sub-components in the Hypermatrix are marked with additional black
lines. The same color scheme as in the cluster view for Hypermatrices
is applied. The selection of a cell in the Hypermatrix is supported by
additional lines and triangles (dark and light grey) for each axis.
(D) visualizes a spectrogram of the selected engine according to
Section 4.3. This type of visualization is the same one, engineers ana-
lyze during their daily routines and is used to support the analysis of a
single engine (
). Here, a diverging color scale is used. Blueish colors
represent acoustic measurements, which are more quiet compared to all
other engines, and reddish colors louder ones. This kind of color scale
is appropriate since all values spread around 0 and deviate in different
directions. By hovering over a cell in (C) the according orders of a pair
of sub-components are displayed in the spectrogram. In the example
in Figure 1, the sub-component pair (B-Bearing and A-Bearing) is
selected. The former is shown with a triangle in dark grey and the
latter with light grey. An engine can be annotated by clicking on “Add
Annotation” (T6), which is outlined in detail in Section 6.4
(E) shows the distribution of already assigned labels as a bar chart.
The main purpose of the view is to support the labeling of (selections
of) engines (
), as outlined in detail in Section 6.3. To show label
distributions, a bar chart is an obvious choice. It would have been
possible to use a pie chart, but for the labels that would have broken
the guideline that there should be no more than six segments [24]. In
the example in Figure 1, all but one engine in the cluster are labeled as
B-Bearing error.
(F) shows three different views. First, the line chart displays two
orders from the spectrogram across their rpm values. Second, the scat-
terplot shows the correlation of the selected order pair. Third, the bar
chart shows the aggregated deviations for a region in the Hypermatrix
to all other regions as indicated in Figure 6. These views are designed
to facilitate the analysis of an engine (
), where their input data are
displayed when hovering over a cell in (C). In our application domain,
deviations above and below three tend to be reasons for errors in the
selected part. Thus, lines above and below this limit are marked as red
for (+3) and blue for (-3). The purpose of the scatterplot is to provide
additional information on how the two selected order lines correlate
with each other. An additional overview about the five most deviating
sub-component pairs in (B) is provided as a bar chart view in (E). To
be consistent with our use of colors, red bars represent aggregated
deviations greater than zero and blue ones lower zero.
6.2 Interactive Clustering
Fig. 7. Dialog to train SOM. Users can select SOM parameters in (1),
choose input clusters in (2), and filter for relevant labels in (3).
To facilitate and speed up labeling, users can perform clustering based
on a subset of engines in (A) - “Train new SOM”. Figure 7 shows
the dialog window with the three possible options for user interaction
marked from 1-3. In (1), the user can define the input parameters to
train a new SOM, such as batch size or learning rate.
As input data, the user can filter all engines based on their cluster (2)
or label (3). In (2), the user selects engines by clicking on a cluster in the
grid to form a subset of new engines. In (3), the user selects relevant la-
bels of interest by clicking on a bar in the bar chart. After the parameter-
ization, the SOM is trained as described in Section 5.2 and replaces (A)
in Figure 1. However, users always can return the initial visualization.
6.3 Interactive Labeling
After an analysis is complete, the user can assign a label to the com-
ponent in (E) with the two list views (
). Labels can also have
subcategories (e.g. “B-Bearing/Inner ring”). The user is able to create
new categories and subcategories or update and delete them in (E).
After a label is selected, users can either label a single engine or the
entire cluster. However, in our design study, we experienced cluster
labeling only for very small not-OK or large OK clusters. Provided
labels serve for the retraining of the clustering as shown in Section
6.2 and are stored in the system’s database to be available to other
engineers. Entered labels further support three tasks. First, they give
an additional overview over groups of engines (
), because clusters
which are already completely labeled are less interesting for an analysis.
Second, they support the selection of an engine (
), since engines that
contain a label are also less probable to be selected. Third, they help in
the analysis of a selected engine (
). This is because they are immedi-
ately displayed in the engine list view in (B) and thus give hints on the
probability of a label for the selected engine. If for example, 4 out of
5 engines in a cluster contain the same label, it is probable that the last
engine also contains the same error and thus should be labeled equally.
6.4 Annotation of Sensor Data
When an engine is labeled, the cause of an error can further be anno-
tated by the user in the spectrogram in (D). Here, the user can select the
specific region in the spectrogram by a rectangle selection as shown in
Figure 8-2 and -3. By dragging the edges of the rectangle, users are pro-
vided with feedback in a tooltip to which order the left and right side and
to which rpm value the upper and lower side of the rectangle belongs.
As well as labels, annotations can support analyses in multiple ways.
Fig. 8. (1) Drill-down of Hypermatrices from the selected cluster in
Figure 1(A) and errors in the A-Bearing in (2) and a B-Bearing in (3) with
according annotations.
First, users are able to review annotations from previous analyses
of other engineers. By selecting a label from the bar chart view in
Figure 1-E all other annotations for the same label are displayed in the
spectrogram. This supports the user in the analysis of an engine (T4).
Second, when an engine was annotated, the user can request a
label and annotation suggestion from the system. Here, all labels
and respective annotations are queried from the database. Since each
annotation exactly specifies the range of columns (Left Order/ right
Order) and rows (Upper rpm/ Lower rpm), it allows to narrow down
the search space for an anomaly in the spectrogram. Next, the three
annotations and according labels with the highest threshold violations
are displayed to the user. If there are no violations, the system suggests
that the selected engine seems to be OK. We choose to give the user
only the opportunity to request an annotation from the system after
an annotation input was manually made by the user. With this design
choice we intent to avoid blind trust in a system recommendation and
thus a decision bias. The more annotations are made, the better the sug-
gestion of further annotations becomes. All labels and annotations are
stored in the system’s database and can be used for different purposes,
such as model training. Thus, a continuous stream of knowledge from
domain experts is stored [36] with IRVINE via labeling and annotating.
6.5 Implementation
The system is a single-page web application written in Typescript,
HTML, and CSS using the framework Angular Js. All views are based
on D3.js [14]. To improve rendering speed, all SVGs are rendered
as canvas elements. IRVINE runs on a Docker Container on a virtual
machine, so that each employee inside the BMW network can access it
via a public URL. The input data is processed in a separate Python ap-
plication and stored in a SQL Database. All API calls run on a separate
Python Flask application hosted on a virtual machine from BMW.
This section outlines two usage scenarios for the interactive clustering
and labeling with IRVINE, in line with the groups of required
functionalities of Section 4.4. The first scenario (Section 7.1) shows
how a new clustering can be performed to retrieve a group of similar
signatures. The goal hereby lies in the fast detection of interesting
signatures for labeling. The second scenario (Section 7.2) shows how
users perform a detailed analysis of an engine, provide a label for a
B-Bearing error, and annotate the respective acoustic measurement.
7.1 Scenario 1: Interactive Clustering
The engineer Alexandra is an explorer. For her analysis, she has 434
engines available, all of which are unknown to her at the start. Her
goal is to assign labels to unlabeled engines (
). She starts IRVINE,
leading to the analysis state as shown in Figure 1 with a SOM cluster-
ing with 5x5 grid cells. Alexandra is interested in engines with a clear
signature. She selects four grid cells with an overall of 28 engines and
four different labels, namely the second gear (GBX2), the rotor shaft,
the B-Bearing, and the magnetic field, as shown in Figure 7-3. To gain
an overview of the characteristics of only these 28 engines (
), she de-
cides to re-train clustering with 4x4 cells only using these engines. Due
to the small number of selected engines, she chooses a relatively small
batch size, while for the other parameters she keeps the default settings.
Fig. 9. Clusters resulting from a SOM training with 28 engines. The
selected cluster contains three engines, where two are already labeled
with an error in the second gear.
Alexandra analyzes the clustering result as shown in Figure 9 and
realizes that, as she expected, all clusters contain a small number of
engines between zero and three engines. In this way, she can focus
on very small subsets of engines, when browsing through individual
grid cells. At a glance, she identifies many different signatures, while
stronger errors (with a high degree of red colors) align on the left of the
SOM. Interestingly, the two clusters with the most apparent signatures
were already labeled by other engineers before.
She decides to browse further and selects the top left cluster (
containing engines with well-distinguishable signatures. Two of the
three engines already contain labels, hinting at errors in the second gear.
The Hypermatrix of the unlabeled engine in the engine list view in
Figure 9 is very similar to the labeled engines. Alexandra thus comes to
the conclusion to label this engine also with an error in the second gear
). The proper validation of an assigned label requires a detailed
analysis of the selected engine. This scenario is outlined in Section 7.2.
7.2 Scenario 2: Interactive Labeling and Annotating
The engineer Thomas is a confirmer. He uses the same data set as
Alexandra but wants to analyze engines with a specific signature he
discovered in previous analyses (
). The cluster with the interesting
signature is outlined in Figure 1 with a grey stroke in (A). For this
cluster, only one engine has not been labeled yet, which is why Thomas
selects this engine for a detailed analysis (
). Being a confirmer,
Thomas has a hypothesis that an error in the B-Bearing can also be
seen in the acoustic measurement of an A-Bearing. By reviewing the
Hypermatrix of the selected engine in Figure 1 (C), Thomas notices
that in fact there seems to be an anomaly originating in the B-Bearing,
indicated by the dark red colors. This assumption is supported by
reviewing the bar chart in (F), where also the B-Bearing is marked as
the biggest anomaly in the engine. By hovering over the Hypermatrix,
he selects orders that can be related to the inner ring of the B-Bearing
and the outer ring of the A-Bering as shown in the upper white triangle
in (C). He notices that in the line chart and scatterplot the selected
order line for the inner ring of the B-Bearing is above a threshold of
three, which is marked as red in (F). He can also see that there seems
to be a week negative linear relationship between the inner ring of the
B-Bearing and the outer ring of the A-Bearing. Thus, he decides to
label the engine as B-Bearing error (
), which is supported by the fact
that all other engines in the cluster are also labeled as B-Bearing error
as shown in the engine list view in (B). In the spectrogram in (D) the
region where the order of the inner ring of the B-Bearing is shown with
a black line and dark grey triangle, he identifies the highest residual
values. Therefore, he clicks on “Add Annotation” in (D) and annotates
the respective region of the error (
). He now confirms his hypothesis
that in fact for the sample of 434 engines, B-Bearing errors can also be
seen in acoustic measurements of A-Bearings.
In this section, we introduce our evaluation methodology (Section 8.1)
and present our study findings (Section 8.2).
8.1 Methodology
Our evaluation has two goals. First, to validate the usefulness and
usability of the proposed technical considerations and the resulting
visualization in terms of effectiveness for automotive engineers. Sec-
ond, to evaluate if labeling speed increased after implementing the user
feedback from the first evaluation round as part of our iterative design
process. The analysis of complex data, where domain knowledge is es-
sential is a high-level cognitive task. Such tasks are however difficult to
measure quantitatively and objectively [54]. As for real-world scenarios,
data, users, and tasks are important, we perform a qualitative field study
to evaluate the usefulness and usability of IRVINE. In this type of study,
qualitative coding of user feedback is combined with a quantitative
usability scale [19]. To evaluate, whether labeling speed increased after
including user feedback, we measure how many labels and annotations
users made by using the system for twenty minutes. We then compare
the results to self-reported labeling speed before the system introduc-
tion, before we implemented user feedback, and after the user feedback.
: The study was carried out with six engineers (others
than the lead user) from BMW, responsible for the development of
testing procedures for the manufacturing of electrical vehicles. They
were all male between 23 and 33 years old and had a mean working
experience of 4 years in the problem domain, all with a background
in mechanical engineering. Inside BMW only very few engineers with
a sufficient level of knowledge about the analysis of acoustic measure-
ments of engines exist. Thus, only a low number of potential candidates
are able to properly evaluate IRVINE. However, this is rather common in
design studies, where presented visualizations often tackle very specific
problems, which can be addressed by only a few users [11,19,47,48].
: For the first round of interviews, we used acoustic data from
434 randomly selected engines over a period of six months. The data
did not contain any labels nor annotations. For the second round of
interviews, acoustic data from 308 completely new engines from a
period of four months were uploaded to the system.
: The following task was given to each user: “Please find
error-prone engines and provide labels and annotations for each
engine.Engineers had 20 minutes to find as many error-prone engines
as possible. An exemplary execution that we observed during the
development with our lead engineer can be the following: First, select
a cluster and then an engine from the cluster. Next, analyze the engines’
Hypermatrix, by hovering over its cells and find a sub-component pair
of interest. Then, make a hypothesis, for example, the resonance of
the first gear can also be observed in the rotor-shaft. Next, inspect
the spectrogram, the line charts, the scatterplot, and the bar chart, to
accept or reject this hypothesis. If an engine contains an error, select
a label from existing labels in the engine list view or create and save
a new label as free text and annotate its cause in the spectrogram.
: As the concept of our Hypermatrix and resulting
clustering is rather complex, a system introduction was carried out
in a kick-off group session. Here, two researchers were present, one
tacking notes and the other explaining the system components to the
engineers. The session took place online and lasted for 80 minutes.
Next, interviews were scheduled with each participant. To evaluate the
usability and usefulness of IRVINE, four interviews were conducted in
the form of a think-aloud session [52]. Each session took on average 60
minutes and involved a short walk-through of the system, open-ended
questions [45] about the usage, and a usability questionnaire. All
interviews were held online, where each engineer executed the same
predefined task. The notes from the kick-off and think-aloud study were
analyzed using a qualitative coding methodology [18]. Repeated ideas
or statements in the feedback were assigned with codes extracted from
the data. Afterward, the codes were grouped into abstract categories to
summarize the study results that are also aligned to a set of questions
proposed by Lam et al. [33], in the context of user experience. To
quantitatively assess the usability of our system, we applied the System
Usability Scale (SUS) [45]. Due to our small sample size, the SUS scale
does not provide empirical evidence of the usability of our visualization
but rather a rough direction to support our assumptions of our design
choices. In addition, we measured how many labels and annotations
were made by the users during the task execution. To evaluate if label-
ing speed did improve after the implementation of the user feedback,
additional two interviews were conducted in the form of a think-aloud
session [52]. Here, engineers executed the previously defined task,
where we again measured how many labels and annotations were made.
8.2 Findings
We observed that all engineers used the general system workflow as
outlined in Section 8.1. First, all engineers selected a cluster, where
one engineer also retrained a new SOM and noted, “I want to have
a more detailed view from these similar clusters to better see emerging
signatures”. Next, engineers selected an engine from the engine list.
Here, one engineer also did zoom in the cluster, noting that “for me
it is easier, to see all Hypermatrices in one view, to choose a relevant
engine”. All engineers noted that the engine list is very helpful to
identify relevant engines. One expert also noted that “It is good to
know that I can filter this list, because I often have specific engine IDs
beforehand, which I need to evaluate in detail”.
Next, engineers turned to the Hypermatrix and specifically looked
for cross structures in the matrix. This was reported as relevant,
where one engineer explained “This view helps to evaluate how this
resonance affects other parts of an engine”. Next, engineers made
hypotheses regarding the signatures of engines and verified them using
the spectrogram, the line charts, or the scatterplots. Here, the use of
residual values instead of absolute values in decibel was noted to be
especially helpful (“Normally I have to compare two engines, where
it is often tedious to find a perfect engine as ground truth.)
Engineers then assigned labels to the selected engines, where two of
them created new label categories in the list view. Three engineers also
requested a label from the system after their label input and in all three
cases agreed with the systems’ suggestion. Next, engineers annotated
the cause for the detected error in the spectrogram and reported that
it is very helpful that their previous inputs are stored in the system
to guide their analyses. One engineer also mentioned that this kind
of knowledge helps for discussions with other stakeholders, such as
analysts (“It is good to either print it out or use the tool itself to show
the analysis to other non-domain related stakeholders”).
Engineers also recommended some system improvements, which
we implemented in a fourth development cycle in our system. Apart
from minor issues, for example, small font sizes, engineers requested
to include subcategories for each label. Furthermore, they requested
features to better detect relevant engines of interest. This resulted for
example in the creation of pie charts that show how many labels exist
in each cluster or glyphs, which show the anomaly score of clusters
and engines as one can see in Figure 1.
Q1 Q2 Q3 Q4 Q5 Q6 Q7 Q8 Q9 Q10 Total
Expert 1 7.5 10 7.5 5 7.5 7.5 10 10 5 7.5 77.5
Expert 2 7.5 10 7.5 10 7.5 10 5 10 7.5 10 85
Expert 3 10 10 7.5 10 7.5 10 10 7.5 10 10 92.5
Expert 4 7.5 7.5 7.5 7.5 7.5 10 10 7.5 10 10 85
Avg. 8.1 9.4 7.5 8.1 7.5 9.4 8.8 8.8 8.1 9.4 85
Table 1. Results of the System Usability Scale [45] with four engineers.
Considering the quantitative results of the usability survey, our sys-
tem provides excellent usability according to the adjective equivalent
of the achieved SUS score [5]. We found that our systems’ usability is
with 85 highly above the average score of 68 [45]. The individual scores
are outlined in Table 1. Even though we evaluated the system with only
four engineers, we were confident that our system did reach a sufficient
level of usability and thus focused in later interviews only on evaluating
labeling speed. This assumption was supported since the remaining two
engineers did not report any recommendations to improve the system.
Fig. 10. Improvement of labeling speed compared to the status quo
before the introduction of IRVINE, before Iteration 4 and after Iteration 4.
Figure 10 outlines the improvement of labeling speed, which was
reached by using IRVINE. Before the systems’ introduction engineers
reported to be able to provide one or two labels in a manual analysis in
twenty minutes. However, they noted that it is rather difficult to provide
an exact number of labeling speeds since errors and resulting analy-
ses differ significantly. Here, one engineer noted “There are errors,
which can be detected in minutes and others that take up to two hours.
However, generally speaking, I normally can provide up to three labels
within an hour”. Even though this number of labels is rather vague, we
can use it as a rough direction on how IRVINE improved labeling speed.
During our study, we were able to measure the exact number of pro-
vided labels during the task execution. As Figure 10 shows, engineers
were able to label engines 33.3% faster than before the system intro-
duction and 75.0% after we included the user feedback. If we compare
labeling speed to the final system, we can see an improvement of 133%.
Even though engineers noted that each analysis heavily depends on the
engine and thus labeling speed is generally speaking hard to quantify,
we state that with IRVINE labeling speed, however, improved.
In this design study, we contribute the problem characterization and
abstraction of the analysis of acoustic signatures in the manufacturing
of electrical engines. We further report the interactive design of the
presented VA system IRVINE, which we evaluated together with six
automotive engineers. Our design study represents a very detailed
view of the problem domain. From a more abstract point of view,
we integrated a VA system into the workflows of domain experts to
support them with the interactive clustering and labeling of sensor data.
To the best of our knowledge, the problem we described in Section 4.2
is unique inside BMW and has not been addressed by other researchers
so far. We thus argue that a design study with a detailed analysis
of the problem domain and resulting abstractions was necessary to
successfully support domain experts with their very specific tasks.
The overall success of our approach is demonstrated by the fact that
study participants rated IRVINE with a high usability and usefulness
in combination with an increased labeling speed. Our evaluation,
however, had a number of limitations. Probably the most important one
is the fairly small sample size of participants. However, as mentioned
in Section 8.1 it is rather unusual to build solutions for specific domain
problems, which address a community bigger than a couple of dozen
users. A second limitation is the quantification of labeling speed. As
mentioned in our Evaluation, almost every engine has to be treated
more or less differently. Especially regarding electrical engines, where
product behavior is not that well understood compared to more mature
technologies, such as petroleum engines, it is thus hard to determine
beforehand how long it will take to produce a label or to carry out a
detailed analysis. Nevertheless, we believe that our system increases
labeling and analysis speed compared to a manual analysis of the data
and serves well to better detect and understand errors in engines.
Although the design was specific to a particular domain, there are
some aspects that provide guidance to the design to other domains.
Some of that guidance results in the following suggestions to transform
high-dimensional data and design similar systems:
1) Cluster Signatures
: The SOM visualization helps engineers to
immediately detect clusters of similar signatures. The provided small
multiples also support the decision, which clusters to analyze first. The
detection of a group of potential anomalous engines was known to be
an important problem, but before the introduction of IRVINE, engineers
had to manually compare single engines in a slow error-prone process.
In contrast, our SOM visualization allows for a robust and fast identifica-
tion of relevant engines of interest. The benefits of SOM grid visualiza-
tions were previously noted in a use case to analyze speech signals [42].
2) Use Hypermatrices
: The Hypermatrix view helps experts in
the fast allocation of local anomalies in engines. Our approach to
compute and visualize Hypermatrices can further be used to design
systems, where similar data to spectrograms are used as input data.
For example, apart from our application domain, spectrograms are
applied in the analysis and synthesis of speech signals [23], seismic
activities [17], or the medical sector [26], where often sonograms - a
similar representation - are applied.
3) Combine Labeling and Annotations
: Apart from providing
labels for single data instances, as shown in other approaches [44,58],
annotating specific regions directly in matrix views of high-dimensional
data can provide more detailed information about the cause of an error.
One might argue that suggested labels and annotations from the system
are preferable to omit human biases. In our case, however, the primary
goal is not only to detect errors but also to understand the reason for
their occurrence. This kind of gained knowledge [9] is more relevant
than algorithms, which are maybe capable to detect errors fast, but
often remain a black box [39]. Thus, we believe that is important to
keep experts in an active role in high stakes decision making instead
of degrading them to validate or reject model suggestions [2].
4) Make Externalized Knowledge Available
: We present guided
workflows with our visualization on how to externalize and store
knowledge. Here, matrix views, line charts, bar charts, and scatterplots,
helped in creating and validating hypotheses. IRVINE aggregates
knowledge in the form of a label and annotations. This knowledge
does not only help domain-related experts to guide their analyses but
also other stakeholders outside the application domain [22]. Labels,
for example, can be queried by analysts to train machine learning
classifiers and annotations can provide valuable insights on the relevant
feature space of each label.
We believe that these four recommendations can help other re-
searchers when investigating complex data or domain problems. For
instance, Kim et al. [30] and Brattain et al [15] both performed studies
on the analysis of sonogram data. To identify secondary symptoms of
illnesses they can also compute Hypermatrices as demonstrated in our
study. Theses Hypermatrices could then be clustered by using a similar
SOM implementation as demonstrated in Section 5.2. Resulting similar
clusters can then be labeled by medical experts and regions of interest
directly annotated in the sonograms. This externalized knowledge aids
to train other medical experts in the analysis of sonogram data.
The design also had some limitations, which we briefly summarize
here. In our visualization, we use data of only one sensor in one test
bench. However, using data from different test stations or sensors
provides new challenges for visualization designs. One challenge
for example is the right mapping of produced components to their
recorded data across multiple stations and the visualization of multiple
test stations. A solution for this can be the detection of anomalies
for multiple stations with network views, and the visualization of
according Hypermatrices and spectrograms in separate views.
Further, annotations are limited to a rectangular selection. In our
case, this is appropriate, since data in the spectrogram has either a
vertical or horizontal relation. In use cases where patterns are vertically
and horizontally dependent at the same time, annotating data via lasso
functionalities might be more adequate.
Finally, to get a comprehensive overview for the Hypermatrix, our
lead engineer narrowed down 41 relevant orders from the 512 available
ones in the spectrograms. This effort was carried out manually and
based on years of experience of the engineer in the problem domain.
However, this effort can be supported automatically, for example, by
computing most deviating orders of a sample of spectrograms to support
the selection of important orders before computing the Hypermatrix.
This paper presents a design study on the development of a visual-
ization approach to analyze signatures and acoustic measurements of
engines. The resulting VA system IRVINE leverages interactive data
labeling and clustering approaches to facilitate the analysis of high
amounts of acoustic data to detect and understand previously unknown
errors. IRVINE comprises five different core visualizations. (1) The
cluster view and an engine list give an overview over similar signatures
of engines. Furthermore, they allow for drilling down to a group of
engines and the selection of an engine of interest. IRVINE also allows
a more fine-grained clustering by retraining the initial SOM based on
a selection of sub-clusters and labels. (2) The Hypermatrix view allows
the analysis of signatures and the creation of hypotheses regarding
anomalies in the data. (3) The spectrogram view represents the ground
truth of the data to validate previously made hypotheses. (4) Line
charts, scatterplots, and bar charts, map Hypermatrix selections to
spectrograms. (5) Labeling and annotation features allow users to tag
specific engines and store knowledge in the system. IRVINE evolved
iteratively, where we closely worked with engineers from BMW, tested
design alternatives, and held critical discussions. The success of our
design is shown by the high usability scores, high reported usefulness,
and an increased labeling and annotation speed.
There are several avenues for our research. One is to investigate
the connection of sensor data across multiple testing stations and
the resulting challenges for future visualizations. A second one is to
investigate how externalized knowledge enables other stakeholders
outside the application domain, for example, data analysts, to improve
their work. Finally, the effectiveness of our visualization concept
should be investigated in other application domains.
W. Aigner, S. Miksch, H. Schumann, and C. Tominski. Visualization of
Time-Oriented Data. Springer, 1st ed., 2011.
S. Amershi, M. Cakmak, W. B. Knox, and T. Kulesza. Power to the
people: The role of humans in interactive machine learning. AI Magazine,
35(4):105–120, 2014.
G. Andrienko, N. Andrienko, S. Rinzivillo, M. Nanni, D. Pedreschi, and
F. Giannotti. Interactive visual clustering of large collections of trajectories.
In Visual Analytics Science and Technology, pp. 3–10, 2009.
B. Bach, N. Henry-Riche, T. Dwyer, T. Madhyastha, J.-D. Fekete, and
T. Grabowski. Small multipiles: Piling time to explore temporal patterns
in dynamic networks. Computer Graphics Forum, 34, 2015.
A. Bangor, P. Kortum, and J. Miller. Determining what individual sus
scores mean: Adding an adjective rating scale. Journal of Usability
Studies, 4:114–123, 2009.
M. Behrisch, F. Korkmaz, L. Shao, and T. Schreck. Feedback-driven
interactive exploration of large multidimensional data supported by visual
classifier. In Visual Analytics Science and Technology, pp. 43–52, 2014.
S. Berg, D. Kutra, T. Kroeger, C. Straehle, B. Kausler, C. Haubold,
M. Schiegg, J. Ales, T. Beier, M. Rudy, K. Eren, J. Cervantes, B. Xu,
F. Beuttenmueller, A. Wolny, C. Zhang, U. K
othe, F. Hamprecht, and
A. Kreshuk. ilastik: interactive machine learning for (bio)image analysis.
Nature Methods, 16:1–7, 2019.
J. Bernard, D. Daberkow, D. Fellner, K. Fischer, O. Koepler, J. Kohlham-
mer, M. Runnwerth, T. Ruppert, T. Schreck, and I. Sens. Visinfo: a digital
library system for time series research data based on exploratory search—a
user-centered design approach. International Journal on Digital Libraries,
16(1), 2015.
J. Bernard, M. Hutter, M. Zeppelzauer, D. Fellner, and M. Sedlmair. Com-
paring visual-interactive labeling with active learning: An experimental
study. Transactions on Visualization and Computer Graphics, PP:1–1,
J. Bernard, C. Ritter, D. Sessler, M. Zeppelzauer, J. Kohlhammer, and
D. Fellner. Visual-interactive similarity search for complex objects by
example of soccer player analysis. In Computer Vision, Imaging and
Computer Graphics Theory and Applications, vol. 3, pp. 75–87, 2017.
J. Bernard, D. Sessler, J. Kohlhammer, and R. A. Ruddle. Using dashboard
networks to visualize multiple patient histories: A design study on post-
operative prostate cancer. Transactions on Visualization and Computer
Graphics, 25(3):1615–1628, 2019.
J. Bernard, N. Wilhelm, B. Kr
uger, T. May, T. Schreck, and J. Kohlhammer.
Motionexplorer: Exploratory search in human motion capture data based
on hierarchical aggregation. Transactions on Visualization and Computer
Graphics, 19(12):2257–2266, 2013.
J. Bernard, M. Zeppelzauer, M. Sedlmair, and W. Aigner. Vial: a unified
process for visual interactive labeling. The Visual Computer, 34, 2018.
M. Bostock, V. Ogievetsky, and J. Heer. D-3: Data-driven documents.
transactions on visualization and computer graphics, 17:2301–9, 2011.
L. J. Brattain, B. A. Telfer, M. Dhyani, J. R. Grajo, and A. E. Samir.
Machine learning for medical ultrasound: status, methods, and future
opportunities. Abdominal Radiology, 43(4):786–799, 2018.
E. T. Brown, J. Liu, C. E. Brodley, and R. Chang. Dis-function: Learning
distance functions interactively. In Visual Analytics Science and Technol-
ogy, pp. 83–92, 2012.
V. W. Chao, Y. Wu, L. Zhao, V. Tsai, and C.-H. Chen. Seismologically
determined bedload flux during the typhoon season. Scientific Reports, 5,
K. Charmaz. Constructing grounded theory : a practical guide through
qualitative analysis. Sage Publications, London; Thousand Oaks, Calif.,
L. Cibulski, H. Mitterhofer, T. May, and J. Kohlhammer. Paved: Pareto
front visualization for engineering design. Computer Graphics Forum,
39:405–416, 2020.
J. J. Dudley and P. O. Kristensson. A review of user interface design
for interactive machine learning. Transactions on Interactive Intelligent
Systems, 8(2), 2018.
J. Eirich, D. J
ackle, T. Schreck, J. Bonart, O. Posegga, and K. Fischbach.
Vima: Modeling and visualization of high dimensional machine sensor
data leveraging multiple sources of domain knowledge. In Visualization
in Data Science at IEEE VIS, 2020.
J. Eirich, D. J
ackle, S. Werrlich, and T. Schreck. Visual analytics in
organizational knowledge creation: A case study. In European Conference
on Information Systems, 04 2021.
J. L. Flanagan. Speech analysis; synthesis and perception. Springer, 2nd
ed., 1972.
M. Hardin, D. Hom, and L. Williams. Which chart or graph is right for
you?, 2016.
F. Heimerl, S. Koch, H. Bosch, and T. Ertl. Visual classifier training for
text document retrieval. Transactions on Visualization and Computer
Graphics, 18(12):2839–2848, 2012.
S. Hughes. Medical ultrasound imaging. Physics Education, 36:468, 2001.
B. H
oferlin, R. Netzel, M. H
oferlin, D. Weiskopf, and G. Heidemann.
Inter-active learning of ad-hoc classifiers for video visual analytics. In
Visual Analytics Science and Technology, pp. 23–32, 2012.
A. Kampker, K. Kreiskother, N. Lutz, V. Gauckler, and M. Hehl. Re-ramp-
up management of scalable production systems in the automotive industry.
In Industrial Technology and Management, pp. 137–141, 2019.
D. A. Keim. Information visualization and visual data mining. Transac-
tions on Visualization and Computer Graphics, 8(1):1–8, 2002.
S. Kim, B. K. Seo, J. Lee, K. Cho, K. Lee, B.-K. Je, H. Y. Kim, Y.-S.
Kim, and J.-H. Lee. Correlation of ultrasound findings with histology,
tumor grade, and biological markers in breast cancer. Acta oncologica,
47:1531–8, 2008.
T. Kohonen. Essentials of the self-organizing map. Neural Networks,
37:52–65, 2013.
T. Kohonen, M. R. Schroeder, and T. S. Huang. Self-Organizing Maps.
Springer, Berlin, Heidelberg, 3rd ed., 2001.
H. Lam, E. Bertini, P. Isenberg, C. Plaisant, and S. Carpendale. Empirical
studies in information visualization: Seven scenarios. Transactions on
visualization and computer graphics, 18, 2011.
J. Moehrmann, S. Bernstein, T. Schlegel, G. Werner, and G. Heidemann.
Improving the usability of hierarchical representations for interactively la-
beling large image data sets. In Human computer interactions, p. 618–627.
Springer, Berlin, Heidelberg, 2011.
T. Munzner. A nested model for visualization design and validation.
Visualization and Computer Graphics, 15:921 – 928, 2010.
I. Nonaka and H. Takeuchi. The knowledge-creating company: How
japanese companies create the dynamics of innovation. Oxford University
Press, New York, 1995.
A. Pister, P. Buono, J.-D. Fekete, C. Plaisant, and P. Valdivia. Integrating
prior knowledge in mixed initiative social network clustering, 2020.
C. Ratanamahatana, J. Lin, D. Gunopulos, E. Keogh, M. Vlachos, and
G. Das. Mining Time Series Data, pp. 1049–1077. Springer, 2010.
C. Rudin. Stop explaining black box machine learning models for high
stakes decisions and use interpretable models instead. Nature Machine
Intelligence, 1:206–215, 2019.
T. Ruppert, M. Staab, A. Bannach, H. L
ucke-Tieke, J. Bernard, A. Kuijper,
and J. Kohlhammer. Visual interactive creation and validation of text
clustering workflows to explore document collections. Electronic Imaging,
2017(1):46–57, 2017.
D. Sacha, Y. Asano, C. Rohrdantz, F. Hamborg, D. Keim, B. Braun, and
M. Butt. Self organizing maps for the visual analysis of pitch contours.
Computational Linguistics, 2015.
D. Sacha, M. Kraus, J. Bernard, M. Behrisch, T. Schreck, Y. Asano, and
D. A. Keim. Somflow: Guided exploratory cluster analysis with self-
organizing maps and analytic provenance. Transactions on Visualization
and Computer Graphics, 24(1):120–130, 2018.
D. Sacha, L. Zhang, M. Sedlmair, J. A. Lee, J. Peltonen, D. Weiskopf,
S. C. North, and D. A. Keim. Visual interaction with dimensionality
reduction: A structured literature analysis. Transactions on Visualization
and Computer Graphics, 23(1):241–250, 2017.
A. Sarkar, M. Spott, A. Blackwell, and M. Jamnik. Visual discovery and
model-driven explanation of time series patterns. In Visual Languages and
Human-Centric Computing, pp. 78–86, 2016.
J. Sauro. Measuring usability with the system usability scale (sus), Ac-
cessed 05.11.2020.
T. Schreck, J. Bernard, T. von Landesberger, and J. Kohlhammer. Vi-
sual cluster analysis of trajectory data with interactive kohonen maps.
Information Visualization, Palgrave Macmillan, 8(1):14–29, 2009.
M. Sedlmair, A. Frank, T. Munzner, and A. Butz. Relex: Visualization
for actively changing overlay network specifications. Visualization and
Computer Graphics, 18:2729–2738, 2012.
M. Sedlmair, P. Isenberg, D. Baur, M. Mauerer, C. Pigorsch, and A. Butz.
Cardiogram: Visual analytics for automotive engineers. In Human Factors
in Computing Systems, pp. 1727–1736, 2011.
M. Sedlmair, M. Meyer, and T. Munzner. Design study methodology:
Reflections from the trenches and the stacks. Visualization and Computer
Graphics, 18:2431–2440, 2012.
B. Settles. Active learning literature survey. Computer Sciences Technical
Report 1648, University of Wisconsin–Madison, 2009.
B. Shneiderman. Inventing discovery tools: Combining information visu-
alization with data mining. Information Visualization, 1(1):5–12, 2002.
M. Someren, Y. Barnard, and J. Sandberg. The Think Aloud Method - A
Practical Guide to Modelling CognitiveProcesses. Academic Press, 1994.
J. Suschnigg, B. Mutlu, A. Fuchs, V. Sabol, S. Thalmann, and T. Schreck.
Exploration of anomalies in cyclic multivariate industrial time series data
for condition monitoring. In Big Data Visual Exploration and Analytics,
M. Tory and T. M
oller. Evaluating visualizations: Do expert reviews
work? Computer graphics and applications, 25:8–11, 2005.
J. Vesanto. Som-based data visualization methods. Intelligent Data
Analysis, 3(2):111–126, 1999.
N. Weber, M. Waechter, S. C. Amend, S. Guthe, and M. Goesele. Rapid,
detail-preserving image downscaling. Transactions on Graphics, 6, 2016.
N. Wilhelm, A. V
ogele, R. Zsoldos, T. Licka, B. Kr
uger, and J. Bernard.
Furyexplorer: Visual-interactive exploration of horse motion capture data.
Visualization and Data Analysis, 2015.
D. Wu, X. Wang, J. Su, B. Tang, and S. Wu. A labeling method for
financial time series prediction based on trends. Entropy, 22(10), 2020.
W. Yang, X. Wang, J. Lu, W. Dou, and S. Liu. Interactive steering
of hierarchical clustering. Transactions on Visualization and Computer
Graphics, 2020.
J. Zhao, M. Karimzadeh, A. Masjedi, T. Wang, X. Zhang, M. M. Crawford,
and D. S. Ebert. Featureexplorer: Interactive feature selection and explo-
ration of regression models for hyperspectral images. In Visualization
Conference, pp. 161–165, 2019.
... Recent success stories show how such analysis efforts result in substantial manufacturing improvements. Examples are the visual exploration of assembling line performance to detect inefficiencies [55], the analysis of acoustic signatures of electrical engines to improve manufacturing quality [12], or the identification of patterns in machine repair logs to decrease maintenance costs [20]. However, data-driven analyses exist that cannot be executed by only one stakeholder group alone. ...
... There, design studies and resulting VA applications mainly support product design, condition monitoring of stations, the optimization of testing procedures, or the visual support of high cognition tasks. Efforts were carried out to visualize in-car communication networks [42][43][44], to facilitate the exploration of multi-criteria alternatives for rotor designs [7], to detect and analyze anomalies in test stations [14,51], the visual exploration of assembling data to detect inefficiencies [55], and to support mechanical engineers in the analysis of acoustic signatures of electrical engines [12]. ...
... Some of the mentioned studies explicitly acknowledge the need for externalizing tacit expert knowledge [14]. For example, the Cardiogram system [44] stores externalized expert knowledge in the form of state machine diagrams, while IRVINE [12] stores expert knowledge in the form of labels for electric engines and annotations in the raw sensor data. All mentioned studies succeeded in creating insights for engineering experts based on machine sensor data. ...
We present ManuKnowVis, the result of a design study, in which we contextualize data from multiple knowledge repositories of a manufacturing process for battery modules used in electric vehicles. In data-driven analyses of manufacturing data, we observed a discrepancy between two stakeholder groups involved in serial manufacturing processes: Knowledge providers (e.g., engineers) have domain knowledge about the manufacturing process but have difficulties in implementing data-driven analyses. Knowledge consumers (e.g., data scientists) have no first-hand domain knowledge but are highly skilled in performing data-driven analyses. ManuKnowVis bridges the gap between providers and consumers and enables the creation and completion of manufacturing knowledge. We contribute a multi-stakeholder design study, where we developed ManuKnowVis in three main iterations with consumers and providers from an automotive company. The iterative development led us to a multiple linked view tool, in which, on the one hand, providers can describe and connect individual entities (e.g., stations or produced parts) of the manufacturing process based on their domain knowledge. On the other hand, consumers can leverage this enhanced data to better understand complex domain problems, thus, performing data analyses more efficiently. As such, our approach directly impacts the success of data-driven analyses from manufacturing data. To demonstrate the usefulness of our approach, we carried out a case study with seven domain experts, which demonstrates how providers can externalize their knowledge and consumers can implement data-driven analyses more efficiently.
... A common technique in condition testing are acoustic testing procedures. The IRVINE system [18] is based on acoustic analysis of test objects, e.g., engines during analysis in a testbed. The system visualizes obtained spectrograms, and allows to compare these across different tests to find deviations which may explain product errors. ...
... (a) The IRVINE system[18] supports inspecting large amounts of acoustic test data, over-viewing clusters of acoustic profiles of test objects. Test engineers can compare acoustic profiles, annotate observations for hints of errors, and compare measurement details. Figure courtesy of Joscha Eirich. ...
Advances in sensor and data acquisition technology and in methods of data analysis pose many research challenges but also promising application opportunities in many domains. The need to cope with and leverage large sensor data streams is particularly urgent for industrial applications due to strong business competition and innovation pressure. In maintenance, for example, sensor readings of machinery or products may allow to predict at which point in time maintenance will be required and allow to schedule service operations respectively. Another application is the discovery of the relationships between production input parameters on the quality of the output products. Analysis of respective industrial data typically cannot be done in an out-of-the-box manner but requires to incorporate background knowledge from fields such as engineering, operation research, and business to be effective. Hence, approaches for interactive and visual data analysis can be particularly useful for analyzing complex industrial data, combining the advantages of modern automatic data analysis with domain knowledge and hypothesis generation capabilities of domain experts. In this chapter, we introduce some of the main principles of visual data analysis. We discuss how techniques for data visualization, data analysis, and user interaction can be combined to analyze data, generate and verify hypotheses about patterns in data, and present the findings. We discuss this in the light of important requirements and applications in the analysis of industrial data and based on current research in the area. We provide examples for visual data analysis approaches, including condition monitoring, quality control, and production planning.
... Observational studies [37,40,82], think-aloud methods [8], and contextual inquiry [74] are useful methods for identifying tasks while using visualization systems, presuming that a solution for a problem already exists. In particular, domain experts can be a valuable source for task abstractions, through interviews [21,22,39,73], surveys [3,7] or other inquiries. Shortcomings include the limited availability of experts [70] and the risk of skewing the task set because of a low number of experts [42]. ...
... Characterizing the problem well requires close and long-term collaboration with experts [177]. Such a challenge generally exists in the development of other visual analytics methods [178,179]. ...
Full-text available
Developing effective visual analytics systems demands care in characterization of domain problems and integration of visualization techniques and computational models. Urban visual analytics has already achieved remarkable success in tackling urban problems and providing fundamental services for smart cities. To promote further academic research and assist the development of industrial urban analytics systems, we comprehensively review urban visual analytics studies from four perspectives. In particular, we identify 8 urban domains and 22 types of popular visualization, analyze 7 types of computational method, and categorize existing systems into 4 types based on their integration of visualization techniques and computational models. We conclude with potential research directions and opportunities.
Visualization techniques are useful in the analysis and insight generation for applications in computing in science and engineering. In this article, we describe the importance of visualization to a digital twin (DT), a virtual representation of a physical object, process or system that can be applied for different tasks, such as data-driven simulation, analysis or monitoring. We illustrate tasks in DTs and give examples of how visualization techniques can be applied for DTs in different application areas.
In recent years, supervised machine learning models have become increasingly important for the advancing digitalization of the manufacturing industry. Reports from research and application show potentials in the use for application scenarios, such as predictive quality or predictive maintenance, that promise flexibility and savings. However, such data-based learning methods require a large training sets of accurately labeled sensor data that represents the manufacturing process in the digital world and allow model to learn corresponding behavioral patterns. Nevertheless, the creation of these data sets cannot be fully automated and requires the knowledge of process experts to interpret the sensor curves. Consequently, the creation of such a data set is time-consuming and expensive for the companies. Existing solutions do not meet the needs of the manufacturing industry as they cannot visualize large data sets, do not support all common sensor data forms and offer little support for efficient labeling of large data volumes. In this paper, we build on our previously presented visual interactive labeling tool Gideon-TS that is designed for handling large data sets of industrial sensor data in multiple modalities (univariate, multivariate, segments or whole time series, with and without timestamps). Gideon-TS also features an approach for semi-automatic labeling that reduces the time needed to label large volumes of data. Based on the requirements of a new use case, we extend the capabilities of our tool by improving the aggregation functionality for visualizing large data queries and by adding support for small time units. We also improve our labeling support system with an active learning component to further accelerate the labeling process. We evaluate the extended version of Gideon-TS on two industrial exemplary use cases by conducting performance tests and by performing a user study to show that our tool is suitable for labeling large volumes of industrial sensor data and significantly reduces labeling time compared to traditional labeling methods.KeywordsLabelingTime seriesUnsupervised learningActive learningSensor dataVisual analytics
Conference Paper
Full-text available
The conversion between tacit and explicit knowledge remains an often-discussed and highly relevant topic in organizational knowledge creation. Although prior research addresses this process, it primarily focuses on the conversion between tacit and explicit knowledge through social processes. This work discusses theories of organ-izational knowledge creation in the light of sociotechnical systems, and specifically extends them to the interaction between individuals and visual analytics systems that afford analytical decision making based on interactive visu-alization and knowledge discovery mechanisms. Based on related work, we develop a theoretical framework to explain novel mechanism for knowledge creation afforded by visual analytics systems. We evaluate our framework with a case study with one of the leading organizations in the automotive industry. Over the course of the case study, we observe and analyze interactions between domain experts and a newly introduced visual analytics sys-tem. Through our case study findings, we reveal novel mechanisms of organizational knowledge creation and discuss their implications.
Full-text available
Time series prediction has been widely applied to the finance industry in applications such as stock market price and commodity price forecasting. Machine learning methods have been widely used in financial time series prediction in recent years. How to label financial time series data to determine the prediction accuracy of machine learning models and subsequently determine final investment returns is a hot topic. Existing labeling methods of financial time series mainly label data by comparing the current data with those of a short time period in the future. However, financial time series data are typically non-linear with obvious short-term randomness. Therefore, these labeling methods have not captured the continuous trend features of financial time series data, leading to a difference between their labeling results and real market trends. In this paper, a new labeling method called “continuous trend labeling” is proposed to address the above problem. In the feature preprocessing stage, this paper proposed a new method that can avoid the problem of look-ahead bias in traditional data standardization or normalization processes. Then, a detailed logical explanation was given, the definition of continuous trend labeling was proposed and also an automatic labeling algorithm was given to extract the continuous trend features of financial time series data. Experiments on the Shanghai Composite Index and Shenzhen Component Index and some stocks of China showed that our labeling method is a much better state-of-the-art labeling method in terms of classification accuracy and some other classification evaluation metrics. The results of the paper also proved that deep learning models such as LSTM and GRU are more suitable for dealing with the prediction of financial time series data.
Conference Paper
Full-text available
The highly integrated design of the electrified powertrain creates new challenges in the holistic testing of high-quality standards. Particularly test technicians face the challenge, that lots of machine-sensor data is recorded during these tests that needs to be analyzed. We present VIMA, a VA system that processes high dimensional machine-sensor data to support test technicians with these analyses. VIMA makes use of the concept of interactive labeling to train machine learning models and the process model of knowledge creation in visual analytics to create new knowledge through the interaction with the system. Its usefulness is demonstrated in a qualitative user study with four test technicians. Results indicate that through VIMA, previously undetected abnormal parts, could be identified. Additionally, a model trained with labels generated through VIMA, was deployed on a test station, that outperforms the current testing procedure, in detecting increased backlashes and improved the test benches output by 15%.
Full-text available
Design problems in engineering typically involve a large solution space and several potentially conflicting criteria. Selecting a compromise solution is often supported by optimization algorithms that compute hundreds of Pareto-optimal solutions, thus informing a decision by the engineer. However, the complexity of evaluating and comparing alternatives increases with the number of criteria that need to be considered at the same time. We present a design study on Pareto front visualization to support engineers in applying their expertise and subjective preferences for selection of the most-preferred solution. We provide a characterization of data and tasks from the parametric design of electric motors. The requirements identified were the basis for our development of PAVED, an interactive parallel coordinates visualization for exploration of multi-criteria alternatives. We reflect on our user-centered design process that included iterative refinement with real data in close collaboration with a domain expert as well as a summative evaluation in the field. The results suggest a high usability of our visualization as part of a real-world engineering design workflow. Our lessons learned can serve as guidance to future visualization developers targeting multi-criteria optimization problems in engineering design or alternative domains.
How has Japan become a major economic power, a world leader in the automotive and electronics industries? What is the secret of their success? The consensus has been that, though the Japanese are not particularly innovative, they are exceptionally skilful at imitation, at improving products that already exist. But now two leading Japanese business experts, Ikujiro Nonaka and Hiro Takeuchi, turn this conventional wisdom on its head: Japanese firms are successful, they contend, precisely because they are innovative, because they create new knowledge and use it to produce successful products and technologies. Examining case studies drawn from such firms as Honda, Canon, Matsushita, NEC, 3M, GE, and the U.S. Marines, this book reveals how Japanese companies translate tacit to explicit knowledge and use it to produce new processes, products, and services.
We propose a new approach—called PK-clustering—to help social scientists create meaningful clusters in social networks. Many clustering algorithms exist but most social scientists find them difficult to understand, and tools do not provide any guidance to choose algorithms, or to evaluate results taking into account the prior knowledge of the scientists. Our work introduces a new clustering approach and a visual analytics user interface that address this issue. It is based on a process that 1) captures the prior knowledge of the scientists as a set of incomplete clusters, 2) runs multiple clustering algorithms (similarly to clustering ensemble methods), 3) visualizes the results of all the algorithms ranked and summarized by how well each algorithm matches the prior knowledge, 4) evaluates the consensus between user-selected algorithms and 5) allows users to review details and iteratively update the acquired knowledge. We describe our approach using an initial functional prototype, then provide two examples of use and early feedback from social scientists. We believe our clustering approach offers a novel constructive method to iteratively build knowledge while avoiding being overly influenced by the results of often randomly selected black-box clustering algorithms.
Hierarchical clustering is an important technique to organize big data for exploratory data analysis. However, existing one-size-fits-all hierarchical clustering methods often fail to meet the diverse needs of different users. To address this challenge, we present an interactive steering method to visually supervise constrained hierarchical clustering by utilizing both public knowledge (e.g., Wikipedia) and private knowledge from users. The novelty of our approach includes 1) automatically constructing constraints for hierarchical clustering using knowledge (knowledge-driven) and intrinsic data distribution (data-driven), and 2) enabling the interactive steering of clustering through a visual interface (user-driven). Our method first maps each data item to the most relevant items in a knowledge base. An initial constraint tree is then extracted using the ant colony optimization algorithm. The algorithm balances the tree width and depth and covers the data items with high confidence. Given the constraint tree, the data items are hierarchically clustered using evolutionary Bayesian rose tree. To clearly convey the hierarchical clustering results, an uncertainty-aware tree visualization has been developed to enable users to quickly locate the most uncertain sub-hierarchies and interactively improve them. The quantitative evaluation and case study demonstrate that the proposed approach facilitates the building of customized clustering trees in an efficient and effective manner.
Conference Paper
Industrial product testing is frequently performed in cycles, resulting in cycle-dependent test data. Monitoring the condition of products under test involves analysis of large and complex test data sets. Main tasks are to detect anomalies and dependencies between observation variables, which appears to be challenging to engineers. In this paper, we present a flexible and extend-able visual analytics approach for anomaly detection focusing on cycle-depended data. It is based on a glyph representation to visualize anomaly scores of cycles with respect to interactively selected reference data. Our approach is built on a design study in collaboration with an industrial engineering corporation, and is demonstrated on real data from engines tested on automotive testbeds. Based on findings from evaluation results, we provide a discussion and an outlook for future work.
We present ilastik, an easy-to-use interactive tool that brings machine-learning-based (bio)image analysis to end users without substantial computational expertise. It contains pre-defined workflows for image segmentation, object classification, counting and tracking. Users adapt the workflows to the problem at hand by interactively providing sparse training annotations for a nonlinear classifier. ilastik can process data in up to five dimensions (3D, time and number of channels). Its computational back end runs operations on-demand wherever possible, allowing for interactive prediction on data larger than RAM. Once the classifiers are trained, ilastik workflows can be applied to new data from the command line without further user interaction. We describe all ilastik workflows in detail, including three case studies and a discussion on the expected performance. ilastik is an user-friendly interactive tool for machine-learning-based image segmentation, object classification, counting and tracking.