Content uploaded by Jingyue Li
Author content
All content in this area was uploaded by Jingyue Li on Mar 04, 2020
Content may be subject to copyright.
Smart Maintenance in Asset Management – application with Deep
Learning
Harald Rødseth1, Ragnhild J. Eleftheriadis2, Zhe Li3 and Jingyue Li3
1 Department of Mechanical and Industrial Engineering, Norwegian University of Science and
Technology (NTNU), 7491 Trondheim, Norway
harald.rodseth@ntnu.no
2 Sintef Manufacturing AS, Product and Production Development
Ragnhild.Eleftheriadis@ntnu.no
3 Department of Computer Science, Norwegian University of Science and Technology
(NTNU), 7491 Trondheim, Norway
zhel, jingyue.li}@ntnu.no
Abstract. With the onset the digitalization and Industry 4.0, the maintenance
function and asset management in a company is forming towards Smart Mainte-
nance. An essential application in smart maintenance is to improve the mainte-
nance planning function with better criticality assessment. With the aid from ar-
tificial intelligence it is considered that maintenance planning will provide better
and faster decision making in maintenance management. The aim of this article
is to develop smart maintenance planning based on principles both from asset
management and machine learning. The result demonstrates a use case of criti-
cality assessment for maintenance planning and comprise computation of anom-
aly degree (AD) as well as calculation of profit loss indicator (PLI). The risk
matrix in the criticality assessment is then constructed by both AD and PLI and
will then aid the maintenance planner in better and faster decision making. It is
concluded that more industrial use cases should be conducted representing dif-
ferent industry branches.
Keywords: Smart Maintenance, Anomaly detection, Asset Management.
1 Introduction
The mission of asset management can be comprehended as the ability to operate the
physical asset in the company through the whole life cycle ensuring suitable return of
investment and meeting the defined service and security standards [1]. Further, it is also
stated in ISO 55000 that asset management will realize value from the asset in the or-
ganization where asset is a thing or item that has potential or actual value for the
2
company [2]. The role of the maintenance function in asset management has been fur-
ther detailed in EN 16646 standard for physical asset management considers the rela-
tionship between operating and maintaining the asset is documented [3]. In particular
it is recommended in this standard that dedicated key performance indicator (KPI) can
be applied in physical asset management. A proposed KPI that improves the integrated
planning process between the maintenance and the production function in asset man-
agement is denoted as profit loss indicator (PLI). This indicator evaluates the different
types of losses in production from an economic point of view. Also, PLI has been tested
in different industry branches such as the petroleum industry [4] and manufacturing
industry [5].
With the onset of digitalization in industry enabled by breakthrough innovations
from Industry 4.0 changes the maintenance capability in the company. The shift is from
a “off-line” maintenance function where data is collected and analyzed manually, to-
wards a digital maintenance [6] and is often denoted as smart maintenance [7, 8].
Artificial intelligent (AI) and machine learning which is a central part of smart
maintenance is considered as a fundamental way to process intelligent data. Yet, there
is a difference between traditional machine learning and data driven artificial intelli-
gence [9]. The difference lies in the performance of feature extractions, in manufactur-
ing often mentioned as machine learning or Advanced Manufacturing. In this article
anomaly detection for smart maintenance will be investigated more in details.
Application of AI is also relevant in order to improve the plant uptime. Anomaly in
mechanical systems usually cause equipment to breakdown with serious safety and eco-
nomic impact. For this reason, computer-based anomaly detection systems with high
efficiency are imperative to improve the accuracy and reliability of anomaly detection,
and prevent unanticipated accidents [10].
From a smart maintenance perspective, the result of a more digitalized asset man-
agement should also include that maintenance is planned with insight from the individ-
ual equipment in combination of the system perspective of the asset [6]. This need is
further supported with empirical studies that points out the necessity for criticality as-
sessment when increasing the productivity through smart maintenance planning [8]. In
fact, maintenance planning is regarded to be unlikely to achieve optimum maintenance
planning without a sound criticality assessment of the physical asset such as the ma-
chines [8].
Smart maintenance has also been denoted with other terms such as deep digital
maintenance (DDM) [11] where application of PLI is of relevance. In DDM it still re-
mains to investigate in appropriate scenarios for the planning capabilities in smart
maintenance that includes anomaly detection and criticality assessment.
The aim of this article is to develop an approach for decision support in smart mainte-
nance planning based on principles both from asset management and machine learning-
based anomaly detection and criticality assessment.
The future structure of this article is as follows: Section 2 presents relevant literature
in smart maintenance whereas Section 3 demonstrates an essential application in smart
maintenance planning where criticality assessment is conducted based on anomaly de-
tection and PLI calculation. Finally, Section 4 discuss the results with concluding re-
marks.
3
2 Smart Maintenance in Asset Management
2.1 The trend towards Smart Maintenance
To succeed with a successful asset management strategy it has been considered that
it is vital to include PLI as an output for the strategy [12]. This has also been included
in maintenance planning in the concept deep digital maintenance (DDM) [11]. In DDM
it has also been demonstrated maintenance planning for one component. It remains to
evaluate several work orders in maintenance planning in DDM. In asset management
the machine learning method such as deep learning has gained popularity [13] where
e.g. diagnostics of health states of power transformers has been applied [14]. In smart
asset management a three-steps approach is proposed [13]:
1. Data gathering from observational data to evaluate the component condition
and defining threshold rules.
2. Analysis of historical data to identify patterns that support in predictions of
future failures.
3. Leverage the component condition with the defect of the failure. This step will
also evaluate the economic perspective in the analysis.
Also smart maintenance is outlined as a key element in the Industry 4.0 roadmap for
Germany [15]. In this strategic roadmap, smart maintenance is considered to improve
the competitive advantage for the maintenance function in the company and is an “en-
abler” itself for successful Industry 4.0 implementation where maintenance data is
shared between manufacturer, operator, and industry service. Furthermore, smart
maintenance has also other important characteristics:
• A common “language” of maintenance processes defined in EN 17007 [16].
• Maintenance technology support with e.g., artificial intelligence (AI) [7].
In smart maintenance this has been addressed with the need for artificial intelligence
(AI) [7]. Despite that maintenance work supported by AI still has barriers to overcome,
it is considered to be an effort worth taking. With support from deep learning, we can
create knowledge of extracted features in an end-to-end process [9]. For instance, neu-
ral networks make the smart data to predict what will happen and take proactive actions
based on improved pattern.
To succeed with smart maintenance in asset management, the emphasizes of specific
plans for maintenance over a long span of time is expected to ensure the greatest value
of equipment over its life cycle [7].
In smart maintenance it is also stressed that it is vital to have established criticality
assessment in maintenance planning [6, 8]. In particular it is concluded that data-driven
machine criticality assessment is essential for achieving smart maintenance planning.
2.2 Smart Maintenance Framework
To ensure value creation in smart maintenance it is necessary to devise a sound frame-
work in smart maintenance. Figure 1 illustrates our proposed framework and is inspired
from [17] and [18]. The starting point in this framework is the data source and includes
external data such as inventory data of spare parts from suppliers. In addition, product
4
data is from the equipment such as condition monitoring data, as well as enterprise data
from computerized maintenance management system (CMMS). All the raw data
sources are then aggregated in multiple formats in a data cloud.
Fig. 1. Value creation with data in smart maintenance framework inspired from [17] and [18].
The raw data is further applied as for smart data analytics including both predictive
and prescriptive analytics. In the maintenance field, the predictive analytics will com-
prise e.g. forecasts of the technical condition of the machine. To ensure value creation
of the physical asset it is also important to include prescriptive analytics that supports
in recommended actions in maintenance planning. This will include anomaly detection
to evaluate the probability of future machine breakdowns. In addition, it is also neces-
sary to evaluate the consequences of the machine failure. In DDM, the PLI seems prom-
ising for this purpose [11]. To assess the criticality of the machine in maintenance plan-
ning, both the probability and the consequence can be combined in a risk matrix.
The result of smart data is deeper insight in the business, where e.g., the plant capac-
ity has increased as well as deeper insights of the partners where e.g., spare part supply
is improved.
3 Smart Maintenance Planning with Criticality Assessment
As shown in Figure 1 and explained above, criticality assessment [8] could be an es-
sential element of prescriptive analysis in our smart maintenance framework. In overall
a criticality assessment should evaluate both the probability and consequences for each
failure of the equipment. We hereby use a demonstration use case to explain our pro-
posal of criticality assessment with three steps: structured approach (1) PLI estimation;
(2) the anomaly degree calculation; (3) the criticality assessment. So far, few companies
have collected data that can be used for PLI estimation and anomaly detection, we could
not get both data from the same machine. The data we present in the case study come
from two different machines. However, for the demonstration purpose and for explain-
ing our ideas, we believe it is applicable to merge the data to explain our idea by as-
suming that the data come from the same machine.
5
Step 1: PLI estimation
The calculation of profit loss indicator is applied based on earlier case study from both
[11] and [5]. The case study considers the malfunction of an oil cooler in a machine
center. The malfunction was first observed when the machine cantered produced scrap-
page. A quality audit meeting evaluated economic loss of this scrappage. In addition,
maintenance personnel conducted inspection on the machine center and found that the
cause of this situation was due to malfunction of the oil cooler. This oil cooler was
replaced, and the machine center had in total 6 days with downtime. In addition to
scrappage it was also evaluated that the machine had lost revenue due to the downtime.
Table 1 summarizes the different type of losses that occurred due to this situation of the
malfunction of machine center.
Table 1. Expected PLI of malfunction of a machine center based on both [11] and [5].
Situation
Type of loss
PLI value/ NOK
Damaged part (Scrappage)
Quality loss
120 000
Quality audit meeting
Quality loss
3 500
Maintenance labor costs
Availability loss
21 570
New oil cooler
Availability loss
47 480
Loss of internal machine revenue
Availability loss
129 600
Sum
322 150
When the consequences for the failure has been estimated, the next step is to calcu-
late the anomaly degree (AI) for the physical asset and the industrial equipment’s.
Step 2: Anomaly degree (AD) Calculation
Figure 2 shows the obtained anomaly degree (AD) of one machine. An increasing AD
will indicate an increasing probability of equipment failure. When maintenance plan-
ning is conducted, updated information about the anomaly degree for each equipment
should be collected and analyzed.
Likewise the calculation of PLI, the data used for calculating AD is also from an
actual industry equipment. However, the data is not from a machine center and repre-
sents another industry branch. The primary datasets include failure records and meas-
urement data from the monitoring system. The target is to obtain the anomaly degree
of the equipment by using machine learning based analysis approaches. In the experi-
ment, we labelled both failure and normal records. Thus, the obtained anomaly degree
can describe the difference between the target observation and normal samples.
6
Fig. 2. Change of anomaly degree (AD)
During the experiment to calculate AD, we adjusted the measurement data collected
in different scales to a common scale. Then, we applied standard normalization to pre-
process the raw data. The applied machine learning model is constructed through a fully
connected deep neural network with seven hidden layers. SoftMax is used as the acti-
vation function of the final output layer. Leaky Relu is applied as the activation func-
tions of the hidden layers. The number of nodes in hidden layers of the constructed deep
neural network is 64, 32, 32, 16, 16, 8, 2, respectively, to train the neural network
smoothly. We selected Adam and categorical cross-entropy as the optimizer and loss
function during the training process. Results in Figure 2 demonstrates the obtained
anomaly degree of the machine from the analysis using deep neural network, which
represents the degradation of the machine’s health state along the time.
Step 3: Criticality assessment
When both the probability and the consequences have been evaluated, the criticality
assessment can be performed in a risk matrix evaluated both the probability and the
consequence for a future malfunction in the machine center. This will support in priority
of maintenance work orders. For example, a risk matrix with specified categories can
support priority of corrective work orders [19].
Figure 3 illustrates a proposed risk matrix in smart maintenance that supports plan-
ning of preventive work orders. In the consequence category, the PLI is established for
the physical asset and classified as “medium, high” in. The probability category is eval-
uated with AD. By trending AD in the risk matrix it is possible to evaluate when a
preventive maintenance work order should be executed and the possible costs and con-
sequences
The color code is following a traffic-light logic; if the equipment is located in green
zone, no further actions are necessary. If the equipment is in a yellow zone, it is an early
warning where maintenance actions should be executed. If the equipment is in the red
zone, it is an alarm where immediate maintenance actions should be executed.
In addition to the color-code system each field in the matrix is marked with a number
indicating a priority number. The criticality is of the machine has a yellow code in the
start but will have a red color code if no maintenance actions are performed. When the
maintenance planner has several machines that are being criticality assessed, it will be
possible to prioritize which machine that should be maintained first.
0
0.1
0.2
0.3
020 40 60 80 100
Anomaly Degree (AD)
Time (measuring poitns)
7
Fig. 3. Risk matrix as criticality assessment for equipment
4 Discussion and concluding remarks
This article has demonstrated application of criticality assessment, which can be an
essential element of our proposed smart maintenance framework, with application of
both PLI as well as anomaly degree calculation. The benefit of this system is that the
maintenance planner will have a “digital advisor” for evaluating the anomaly that can
enable a faster and better decision-making process in maintenance planning. It is ex-
pected that the deep learning method with deep neural network will be further investi-
gated and developed due to it’s promising results in AD.
Also, with the aid from PLI calculations, it is possible to improve the evaluation of
the consequences of e.g., machine breakdown. In a risk matrix it is then possible to
establish a work priority system where some equipment with anomaly should be prior-
itized before others. For example, if there are future work orders both categorized in
yellow sector andred sector, it would recommend to prioritize the work in the red sector.
There are also some challenges with the criticality assessment that should be ad-
dressed in future research in contribution to theory of criticality assessment. First, it is
of importance to improve the accuracy of both anomaly degree calculation as well as
calculation of PLI. Second, it will be of importance to evaluate sound criteria for each
category in the risk matrix. Yet, this seems to also be a challenge in existing risk matri-
ces. A more practical aspect that needs to be investigated is to evaluate how the digital
approach of the risk matrix will interfere with existing criticality assessment and still
not reduce the performance of the physical asset.
Although a use cases have been applied with data from different industry branches
to demonstrate the criticality assessment, further research will also require a coherent
demonstration in several industry sectors, including both manufacturing industry as
well as the process industry.
Acknowledgements The authors wish to thank for valuable input from both the research project CPS-plant
(grant number: 267750), as well as the research project CIRCit – Circular Economy Integration in the Nordic
Industry for Enhanced Sustainability and Competitiveness (grant number: 83144).
8
Reference
1. Schneider, J., et al., Asset management techniques. International Journal of Electrical
Power & Energy Systems, 2006. 28(9): p. 643-654.
2. ISO, ISO 55000 Asset management - Overview principles and terminology. 2014:
Switzerland.
3. CEN, EN 16646: Maintenance - Maintenance within physical asset management.
2014.
4. Rødseth, H., et al., Increased Profit and Technical Condition Through New KPIs in
Maintenance Management, in Proceedings of the 10th World Congress on Engineering
Asset Management (WCEAM 2015), K.T. Koskinen, et al., Editors. 2016, Springer
International Publishing: Cham. p. 505-511.
5. Rødseth, H. and P. Schjølberg, Data-driven Predictive Maintenance for Green
Manufacturing, in Advanced Manufacturing and Automation VI. 2016, Atlantis Press.
p. 36-41.
6. Bokrantz, J., et al., Maintenance in digitalised manufacturing: Delphi-based scenarios
for 2030. International Journal of Production Economics, 2017. 191: p. 154-169.
7. Yokoyama, A. Innovative Changes for Maintenance of Railway by Using ICT-To
Achieve "smart Maintenance". in Procedia CIRP. 2015.
8. Gopalakrishnan, M., et al., Machine criticality assessment for productivity
improvement: Smart maintenance decision support. International Journal of
Productivity and Performance Management, 2019.
9. Wang, J., et al., Deep learning for smart manufacturing: Methods and applications.
Journal of Manufacturing Systems, 2018. 48: p. 144-156.
10. Li, Z., et al., A deep learning approach for anomaly detection based on SAE and LSTM
in mechanical equipment. The International Journal of Advanced Manufacturing
Technology, 2019. 103(1): p. 499-510.
11. Rødseth, H., P. Schjølberg, and A. Marhaug, Deep digital maintenance. Advances in
Manufacturing, 2017. 5(4): p. 299-310.
12. Rødseth, H. and R. Eleftheradis, Successful Asset Management Strategy
Implementation of Cyber-Physical Systems, in WCEAM 2108. 2019.
13. Khuntia, S.R., J. Rueda, and M. Meijden, Smart Asset Management for Electric
Utilities: Big Data and Future. 2017.
14. Tamilselvan, P. and P. Wang, Failure diagnosis using deep belief learning based
health state classification. Reliability Engineering & System Safety, 2013. 115: p. 124-
135.
15. DIN, German Standardization Roadmap - Industry 4.0, in Version 3. 2018: Berlin.
16. CEN, EN 17007: Maintenance process and associated indicators. 2017.
17. Porter, M.E. and J.E. Heppelmann, How smart, connected products are transforming
companies. Harvard Business Review, 2015. 2015(October).
18. Schlegel, P., K. Briele, and R. H. Schmitt, Autonomous Data-Driven Quality Control
in Self-learning Production Systems: Proceedings of the 8th Congress of the German
Academic Association for Production Technology (WGP), Aachen, November 19-20,
2018. 2019. p. 679-689.
9
19. NORSOK, Z-008- Risk based maintenance and consequence classification 2017,
Oslo: Standard Norge. 56 s.
