Content uploaded by Fanghao Tian
Author content
All content in this area was uploaded by Fanghao Tian on Jan 26, 2024
Content may be subject to copyright.
Automatic Data Extraction based on
Semiconductor Datasheet for Design Automation
of Power Converters
Fanghao Tian1∗, Diego Bernal Cobaleda1, Wilmar Martinez1
1 Dept of Electrical Engineering (ESAT), KU Leuven-Energyville, Diepenbeek-Genk, Belgium
∗E-mail: fanghao.tian@kuleuven.be
Abstract—The application of artificial intelligence (AI) to
design automation (DA) is a novel research area in power
electronics, owing to the complexity of power converter de-
sign, power loss modeling on magnetic and semiconductors,
and the large number of components to choose from. This
paper presents a tool for extracting dynamic and complex
nonlinear characteristics from semiconductor datasheets in
order to improve the power loss estimation model, resulting
in an optimal power converter design. An object recognition
neural network, CenterNet, is trained to extract figures
from datasheets. The dynamic data is then extracted from
figures utilizing Optical Character Recognition (OCR) and
morphological image processing techniques. Finally, the
acquired data is used to enhance the dynamic properties
of power switches and to develop a more accurate power
loss model for use as input to design automation tools.
Keywords—Artificial Intelligence, design automation, im-
age processing, power converters, optimization
I. INTRODUCTION
With the development of decarbonized systems, such
as electric cars, electric aviation, and photo-voltaic sys-
tem, the demand for power electronic converters has in-
creased significantly due to their widespread applications
[1]. Increasingly strict requirements for the design of
power converters in terms of efficiency, power density,
volume, and mass are proposed [2]. However, designing
optimal power converters is difficult due to the abundance
of available power electronics components [3]. Further-
more, as it is impossible to test each design experi-
mentally and choose the one with optimal performance,
preselection of devices and simulations of the possible
designs become necessary. Semiconductors simulations
are frequently based on a simplified model which does
not take into account all of the semiconductors’ dynamic
characteristics. Thus, it is critical to enhance the accuracy
of the semiconductor loss calculation model in order to
perform reliable power converter design optimization.
The study on information extraction has made some
progress as the deep learning technique develops. Clark
and Divvala proposed a method for detecting chunks of
body text and locating the figures or tables by reason-
ing about empty regions [4] [5]. Siegel et al. proposed
FigureSeer, an end-to-end framework for automatically
locating figures in academic papers, particularly those
with subfigures, and classifying them into figures in the
result section and other sections using Convolutional
Neural Networks (CNNs). Then, using Optical Charac-
ter Recognition (OCR) algorithms that recognize text
contained within figures such as titles and legends, all
information contained within the figures is extracted [6].
Another powerful tool named Diag2graph was proposed
in [7]. The same algorithm for figure localization as [6]
was used. Various components such as node, text, and
arrow were detected during the figure content analysis.
Additionally, their relations are analyzed. ChartOCR in
[8] is a universal model that extracts key point infor-
mation and determines the chart type at the same time,
then predicts the data range and extracts the text using
Microsoft OCR. The extraction of data from line charts
was studied in [9], while the segmentation of multiple
curves was explored in color plots.
There are few tools especially for data extraction from
power electronics datasheets. In comparison to standard
academic papers, datasheets have several distinguishing
characteristics. (1) There are intensive figures on some
pages. (2) Some figures are in grayscale while others
are in color. (3) The figures are line charts and usually
have grids. On this basis, a special object recognition
algorithm focusing on complex line charts in datasheets
is developed to extract data with high accuracy. This
paper focuses on creating an AI-based tool to extract
information from figures of semiconductors datasheets to
develop a method to improve semiconductors power loss
calculation modeling for simulation in power converter
design automation.
The paper is organized as follows. In section II, A
CenterNet-based object detection algorithm is utilized to
detect and separate the figures from datasheets. In section
III, the same neural network with differently trained
weight parameters is implemented for key elements de-
tection of the figures. Combined with the results of text
recognition by OCR, data from line charts are extracted.
Then, the power loss model of power switches is enriched
by the obtained dynamic data in section IV. The results
and conclusion are in section V and section VI separately.
II. FIGURE SE PARATION
A. Task Overview
To automate the data extraction process from
datasheets, it is necessary to separate the figures in the
datasheets prior to conducting figure analysis. Image
processing algorithms are used to extract data from each
separate figure. Object detection, text recognition, and
line data extraction are the primary tasks. The flowchart
is as shown in Fig.1.
Fig. 1: Overview of data extraction task.
Detailed procedures for figures separation will be de-
scribed as follows.
B. Dataset Construction
The positions of figures are recognized using a CNN-
based object detection algorithm. Therefore, pages of
transistor datasheets are collected to construct the dataset
for training by manually annotating original datasheet
pages. An example of the annotated page is as shown in
Fig.2, where figures on the datasheet page are annotated
with a figure label.
Fig. 2: Examples of annotated page [10].
Different manufacturers’ datasheets have a variety of
different styles. Notably, some pages contain six figures,
while others may only contain four. Additionally, figures
sometimes occupy only a portion of the page. To increase
the object detection algorithm’s universality, datasheets
from several different manufacturers are gathered. Differ-
ent datasheets from the same manufacturer, on the other
hand, have a similar layout. As a result, CNN training
requires only a small amount of training data.
C. CenterNet
Object detection is a challenging task in computer vi-
sion. Numerous algorithms have been developed, such as
YOLO, RCNN, Faster RCNN and etc [12]. CenterNet is
a novel anchor-free object detection algorithm with good
performance for object detection on multiple scales [11].
To begin, a multiple layer deep CNN network serves as
the backbone for feature extraction from original pictures.
Corner pooling and center pooling are implemented on
top of the backbone to obtain two corner heatmaps and a
center key point heatmap, which detect a pair of corners
and a center, respectively, thus determine the bounding
box for the object. The architecture of CenterNet is shown
in Fig.3.
Fig. 3: Architecture of CenterNet.
The choice of backbone CNN is determined by the
complexity of dataset. Due to the similarity of datasheets
pages and figures, ResNet50 is chosen to be the backbone
CNN [13].
D. Training
Following the preparation of the labeled dataset and
the CenterNet algorithm, the training for 100 epochs
begins. A frozen scheme is used to speed up the training
process. Firstly, a well-trained ResNet50 model by a large
classic dataset is loaded, the parameters of which stay
unchanged for the first 50 training epochs. The training
process is then continued with the ResNet50 parameters
unfrozen. The labeled pages are randomly divided into
training and validation dataset with percentages of 90%
and 10%, respectively.
III. DATA EXTRACTION
Individual figures are separated from datasheets and
analyzed to extract detailed information. Similar to the
figure separation task, object detection is the initial step
in locating and recognizing the critical elements of a
figure, such as the title, legend, coordinate origin corner,
(a) (b) (c)
Fig. 4: Line extraction process (a) original image. (b) binarized image. (c) line skeleton.
label, and values of x-axis and y-axis marks. Then text
recognition is required to identify text and numbers,
which are required to comprehend figure information.
Finally, the line data with correct coordinate values are
extracted.
A. Dataset Construction
To train the CenterNet for line chart recognition,
a dataset must be constructed by manually annotating
figures cut from transistor datasheets. Six key elements
are indicated, which are title, legend, label, value, corner,
and other (if exists). Similarly, in order to increase the
universality of the algorithm, figures in a variety of
styles from a variety of manufacturers are gathered. And
they are randomly divided into training and validation
datasets with percentages of 90% and 10%, respectively.
An example of the annotated figure is as shown in Fig.
5.
Fig. 5: Example of annotated figure [10].
The same CenterNet is trained by the line chart dataset
for 100 epochs.
B. Data Extraction
After recognizing the key elements of figures, data can
be extracted by text recognition and line data processing.
1) Text recognition: Optical character recognition
(OCR) has been developing dramatically recently. Tesser-
act, an open-source OCR engine, is utilized in this paper.
Combined with label information from object detection,
recognized texts are classified as title, label and etc.
2) Line data extraction: Prior to digitizing data in a
line chart, it is necessary to specify the coordinate infor-
mation along with the mark values. To begin, recognized
values are classified into x-axis and y-axis values. Then,
because the majority of datasheet figures contain grids,
the figure is binarized to detect vertical and horizontal
lines that correspond to mark values. In this way, an
accurate coordinate is established.
The following step is to remove the background grid
and extract data from individual lines. Since the lines
are generally thicker than the grid, morphological image
processing techniques are applied to eliminate the grid.
More specially, morphological closing, which is a dilation
operation followed by an erosion, is implemented to
remove the thin black grid on the background. Then the
skeleton of lines is extracted to reduce line width into a
single pixel, making data extraction easier.
Finally, for line charts with multiple lines, an addi-
tional step of data correlation is required. After collecting
all pixel coordinates, they are grouped according to their
x axis values, like [(xi, yi1),(xi, yi2),· · · ,(xi, yin)]. The
next dot can be estimated using the coordinates of the
previous two dots on a line as 1:
ˆyi+1 =yi−yi−1
xi−xi−1
(xi+1 −xi) + yi(1)
Then the dot closest to ˆyi+1 is considered to be the
correlated dot for this line. And to prevent incorrect
correlation or noise interference, the distance between
the estimated position and the closest dot is limited to
a specific range.
The process of converting the original figure to a
binarized figure for the purpose of matching mark values,
then to a line skeleton for the purpose of extracting
line data is illustrated in Fig. 4. The title and labels
are removed during the binarization process to eliminate
background noise.
IV. ANALYTICAL SEMICONDUCTOR POW ER L OS SE S
MOD EL
In the power electronics design process, the selection
of semiconductors components significantly impacts the
performance of power converters. Traditionally, engineers
tend to choose the components based on experiences since
It is impractical to manually inspect thousands of semi-
conductor components, and the market is filled up with
semiconductor switching devices with similar voltage and
current characteristics. As a result, additional time is
spent selecting a device. Additionally, the parameters and
variables of semiconductors are highly relevant. To be
able to determine which of them, or which combination of
them will improve the overall efficiency of the converter,
an accurate and fast model is required. However, dynamic
data from datasheets can improve the analytical power
loss model on field effect transistor (FET) to get a more
precise estimation model. Thus, power converter designs
can be thoroughly evaluated and optimized.
Fig. 6: FET capacitances.
Fig.6 Shows the representation of the intrinsic capac-
itance of a conventional FET. Where; input capacitance
Ciss=Cgd +Cg s, output capacitance Coss =Cds +Cgd, and
reverese transfer capacitance Crss=Cgd .
A. Power Loss model
Power losses on semiconductor switches, generally
consist of conduction losses, gate losses, switching losses
and reverse recovery losses. To estimate them, common
used calculation models are shown in (2)-(5) [15]. Addi-
tionally, when calculating power loss, the parameters in
these formulations are typically fixed to typical values
rather than dynamic data from datasheet plots, which
implies a more accurate behavior of the device. In Fig. 7
two screenshots from a power FET are shown, Fig. 7a
depicts the typical parameters usually taken for the design
while Fig. 7b depicts the behaviour of the dynamic
parameters depending on the operation point of the FET.
Pc=RDS(on)) IDRM S
2(2)
Pg=CissV2
GS fs(3)
Psw =VDS ID
2(tr+tf)fs(4)
Prr =CossV2
DS fs(5)
where trand tfare turning on and turning off time, which
can be calculated by formulations in [16].
(a)
(b)
Fig. 7: Parameters for power loss calculation from
datasheets (a) Typical parameter values (b) Dynamic
values [10].
B. Improved power loss model
As previously stated, dynamic data from datasheets
are extracted and stored, which can be used to augment
the model for calculating power loss. For conduction loss
formulation (2), conventional method is to choose a fixed
value for RDS(on). However, according to datasheets,
conduction resistance is a variable that is dependent
on the junction temperature. As a result, the specific
value of turn on resistance should be determined by the
temperature. A dynamic conduction loss map is shown
as Fig. 8 with junction temperature and drain current
changing between [0 −150] and [0 −50] separately.
For reverse recovery loss (5), the output capacitance
Coss is a variable depending on drain to source voltage
VDS as is shown in 7b. Instead of regarding Coss as a
constant, an improved calculation method is shown in (6).
The integration can be estimated by data from the line
chart.
P
′
rr =VDS fsZVDS
0
CossV dV (6)
As for the switching loss (4), similar dynamic data can
be implemented in the process of calculating trand tf.
Fig. 8: Conduction loss vs. RMS current and junction
temperature.
V. RE SU LTS
A total of 200 pictures for the figure separation task
and 300 pictures for the line data extraction task have
been trained and tested. As is shown in Fig.9, the training
loss reduces dramatically and stays at a small value. The
final training losses are 2.59 and 0.87 for both tasks, while
the validation losses are 2.15 and 0.57 separately.
(a)
(b)
Fig. 9: Training and validation losses for (a) figure
separation (b) line data extraction.
With the small training and validation losses, the
trained CenterNet can detect the key elements from a
datasheet page or line chart. Two examples of the results
are illustrated in Fig. 10a. Every figure on this page is
detected and selected by the red bounding boxes. With
the bounding box information, the figures can be cut
into individual figures. Additionally, for the line chart
data extraction task, 10b displays the critical elements
highlighted by bounding boxes and classified according to
the pre-defined classes. After applying the OCR for text
recognition and traditional image processing algorithms
for curve extraction, the dynamic data are extracted from
datasheets successfully.
(a)
(b)
Fig. 10: Object detection results for (a) figure separation
(b) line data extraction [10].
VI. CONCLUSIONS
Power loss simulation is critical for design automation
of power electronics. Semiconductor components like
power FETs have unavoidable power loss, especially at
high frequencies. Thus, an accurate and precise mathe-
matical model of semiconductor power losses can result
in the optimal selection of a semiconductor component
for a particular design requirement. By acquiring dynamic
data from datasheets, it is possible to accurately estimate
the power losses of a specific component under specified
operating conditions.
In this paper, a CenterNet based tool for data extrac-
tion of semiconductors datasheets is proposed. Firstly,
CenterNet was trained for two tasks, which are figure
separation from original datasheets and key elements
recognition for line charts. Secondly, data extraction
from line charts is accomplished using image processing
techniques. Finally, an improved dynamic power loss
model is explained, considering junction temperature for
conduction losses and dynamic intrinsic capacitance for
reverse recovery losses. The same method can be utilized
to enhance the switching loss model.
REFERENCES
[1] W. Martinez, C. Cortes, A. Bilal and J. Kyyra, ”Finite Element
Methods for Multi-objective optimization of a High Step-up Inter-
leaved Boost Converter,” in 2018 International Power Electronics
Conference (IPEC-Niigata 2018 ECCE Asia), pp. 2193-2198,
2018.
[2] W. Martinez, M. Yamamoto, J. Imaoka, F. Velandia and C. A.
Cortes, ”Efficiency optimization of a two-phase interleaved boost
DC-DC converter for Electric Vehicle applications,” in 2016
IEEE 8th International Power Electronics and Motion Control
Conference (IPEMC-ECCE Asia), pp. 2474-2480, 2016.
[3] Mouser Electronics. Accessed on: October 10, 2021. [Online].
Available: https://eu.mouser.com/c/semiconductors/discrete-
semiconductors.
[4] C. Clark and S. Divvala. ”Looking Beyond Text: Extracting
Figures, Tables, and Captions from Computer Science Papers,”
in AAAI Workshops at the Twenty-Ninth AAAI Conference on
Artificial Intelligence, pp. 2-8, 2015.
[5] C. Clark and S. Divvala. ”PDFFigures 2.0: Mining figures from
research papers,” in 2016 IEEE/ACM Joint Conference on Digital
Libraries, pp. 143-152, 2016.
[6] N. Siegel, Z. Horvitz, R. Levin, S. Divvala, and A. Farhadi, ”Fig-
ureseer: Parsing result-figures in research papers,” in European
Conference on Computer Vision, pp. 664–680, 2016.
[7] A. Roy, I. Akrotirianakis, A. Kannan, D. Fradkin, A. Canedo, K.
Koneripalli, and T. Kulahcioglu. ”Diag2graph: Representing Deep
Learning Diagrams in Research Papers as Knowledge Graphs,” in
2020 IEEE International Conference on Image Processing (ICIP),
pp. 2581-2585, 2020.
[8] J. Luo, Z. Li, J. Wang and C. -Y. Lin, ”ChartOCR: Data
Extraction from Charts Images via a Deep Hybrid Framework,”
in 2021 IEEE Winter Conference on Applications of Computer
Vision (WACV), pp. 1916-1924, 2021.
[9] S. Choudhury, S. Wang, P. Mitra, and C. Giles, ”Automated
data extraction from scholarly line graphs,” in Eleventh IAPR
International Workshop on Graphics Recognition (GREC), 2015.
[10] EPC Products datasheets. Accessed on: Febru-
ary 28, 2022. [Online]. Available: https://epc-
co.com/epc/Products/eGaNFETsandICs.aspx.
[11] K. Duan, S. Bai, L. Xie, H. Qi, Q. Huang and Q. Tian, ”Center-
Net: Keypoint Triplets for Object Detection,” in 2019 IEEE/CVF
International Conference on Computer Vision (ICCV), pp. 6568-
6577, 2019.
[12] L. Jiao, F. Zhang, F. Liu, S. Yang, L. Li, Z. Feng, and R. Qu, “A
Survey of Deep Learning-Based Object Detection”, IEEE Access,
vol. 7, pp. 128837–128868, 2019.
[13] K. He, X. Zhang, S. Ren and J. Sun, ”Deep Residual Learning for
Image Recognition,” 2016 IEEE Conference on Computer Vision
and Pattern Recognition (CVPR),pp. 770-778, 2016.
[14] R. Smith, ”An Overview of the Tesseract OCR Engine,” Ninth
International Conference on Document Analysis and Recognition
(ICDAR 2007), pp. 629-633, 2007.
[15] Z. J. Shen, Y. Xiong, X. Cheng, Y. Fu and P. Kumar, ”Power
MOSFET Switching Loss Analysis: A New Insight,” Conference
Record of the 2006 IEEE Industry Applications Conference
Forty-First IAS Annual Meeting, pp. 1438-1442, 2006.
[16] Vishay Siliconix, Appl. Note AN608A, pp.1-6. Available:
https://www.vishay.com/docs/73217/an608a.pdf.
