Skip to main navigation Skip to main content
  • E-Submission

JKSPE : Journal of the Korean Society for Precision Engineering

OPEN ACCESS
ABOUT
BROWSE ARTICLES
EDITORIAL POLICIES
FOR CONTRIBUTORS
SPECIAL

고장 데이터 부재 및 부족 상황에서의 딥러닝 기반 기계시스템의 고장진단 방법론

Methods for Fault Diagnosis in Mechanical Systems based on Deep Learning in the Absence or Class Imbalance of Fault Data

Journal of the Korean Society for Precision Engineering 2023;40(5):345-351.
Published online: May 1, 2023

1 성균관대학교 대학원 기계공학과

2 성균관대학교 기계공학부

1 Department of Mechanical Engineering, Graduate School, Sungkyunkwan University

2 School of Mechanical Engineering, Sungkyunkwan University

#E-mail: sangwonl@skku.edu, TEL: +82-31-290-7467
• Received: March 13, 2023   • Revised: April 4, 2023   • Accepted: April 4, 2023

Copyright © The Korean Society for Precision Engineering

This is an Open-Access article distributed under the terms of the Creative Commons Attribution Non-Commercial License (http://creativecommons.org/licenses/by-nc/3.0) which permits unrestricted non-commercial use, distribution, and reproduction in any medium, provided the original work is properly cited.

  • 7 Views
  • 0 Download
  • 1 Crossref
  • 1 Scopus
prev next

Citations

Citations to this article as recorded by  Crossref logo
  • Distribution of Force Applied to a Lateral Damper during EMU Operation
    Hyun Moo Hur, Kyung Ho Moon, Seong Kwang Hong
    Journal of the Korean Society for Precision Engineering.2024; 41(9): 673.     CrossRef

Download Citation

Download a citation file in RIS format that can be imported by all major citation management software, including EndNote, ProCite, RefWorks, and Reference Manager.

Format:

Include:

Methods for Fault Diagnosis in Mechanical Systems based on Deep Learning in the Absence or Class Imbalance of Fault Data
J. Korean Soc. Precis. Eng.. 2023;40(5):345-351.   Published online May 1, 2023
Download Citation

Download a citation file in RIS format that can be imported by all major citation management software, including EndNote, ProCite, RefWorks, and Reference Manager.

Format:
Include:
Methods for Fault Diagnosis in Mechanical Systems based on Deep Learning in the Absence or Class Imbalance of Fault Data
J. Korean Soc. Precis. Eng.. 2023;40(5):345-351.   Published online May 1, 2023
Close

Figure

  • 0
  • 1
  • 2
  • 3
  • 4
  • 5
  • 6
  • 7
Methods for Fault Diagnosis in Mechanical Systems based on Deep Learning in the Absence or Class Imbalance of Fault Data
Image Image Image Image Image Image Image Image
Fig. 1 Robotic spot welding monitoring system
Fig. 2 Spot welding sensor data
Fig. 3 Histogram of reconstruction and thresholds
Fig. 4 Bearing simulator and sensor attachment [17] (Adapted from Ref. 17 on the basis of OA)
Fig. 5 STFT-based preprocessing (Applied to normal data)
Fig. 6 Architecture of the cGAN model
Fig. 7 Architecture of 2DCNN and cGAN-2DCNN diagnostic model
Fig. 8 Confusion matrix of 2DCNN and cGAN-2DCNN diagnostic model
Methods for Fault Diagnosis in Mechanical Systems based on Deep Learning in the Absence or Class Imbalance of Fault Data

Information on failure mode

Failure mode Description
Angular misalignment 5o between specimen and electrode tip
Electrode tip wear Tip area ≥19 mm2
Shunting effect 109 mm welding space

Spot welding parameter settings

Welding current [kA] 5.5
Electrode force [kgf] 150
Welding time [sec] 0.2167 (13 cycles)
Holding time [sec] 0.25 (15 cycles)

Architecture and hyperparameters of the Auto-encoder

  Encoder Decoder
First layer 1D conv. 1D conv. transposed
for all layers
Second layer Max pooling
Third layer 1D conv.
Fourth layer Max pooling
Fifth layer 1D conv.
Num. of filters 32, 128, 512 512, 128, 64, 32, 3
Num. of kernel size 3 for all layers 4, 3, 5, 3, 5
Num. of strides 3 for all layers
Activation function ReLU for all layers,
Sigmoid for Decoder’s last layer
Epochs 1,000
Batch size 32
Optimizer Adam (Learning rate = 0.001)
Loss function Mean Absolute Error (MAE)

Z-score and threshold for each confidence level

Confidence level Z-score Threshold
0.9 1.65 μ + σ × 1.65
0.95 1.96 μ + σ × 1.96
0.99 2.56 μ + σ × 2.56

Accuracy of the anomaly detection model

Failure mode Threshold (Confidence level)
0.9 0.95 0.99
Normal [%] 92 96 98
Misalignment [%] 96 94 34
Electrode tip wear [%] 96 94 48
Shunt effect [%] 100 98 86
Average [%] 96 95.5 66.5

Configuration of class imbalance dataset

Fault mode Number of data Damage type
Normal 500 None
Ball 25 1-line scratch
Cage 25 Crack
Outer 25 2-line scratch
Inner 25 2-line scratch

Parameter settings for STFT

Window size 400
Overlap size 80
Frequency range [Hz] 1-400

Architecture and hyperparameters of the cGAN model

  Generator Discriminator
Conv. layer Strided transposed conv. Strided conv.
Pooling Not used
Activation function Leaky ReLU (α = 0.2) for all conv. layers,
Sigmoid for discriminator’s last layer
Stride 2, 2
Kernel 3, 3
Epochs 1,000
Batch size 100
Optimizer Adam (Learning rate = 0.0002)

Configuration of augmented class imbalance dataset

Fault mode Number of real data Number of fake data
Normal 500 None
Ball 25 75
Cage 25 75
Outer 25 75
Inner 25 75
Table 1 Information on failure mode
Table 2 Spot welding parameter settings
Table 3 Architecture and hyperparameters of the Auto-encoder
Table 4 Z-score and threshold for each confidence level
Table 5 Accuracy of the anomaly detection model
Table 6 Configuration of class imbalance dataset
Table 7 Parameter settings for STFT
Table 8 Architecture and hyperparameters of the cGAN model
Table 9 Configuration of augmented class imbalance dataset