Mechanical Abnormal Sound Detection Based on Self-Supervised Feature Extraction

Yingjie Xue; Qi Chen; Songbin Zhou; Yisen Liu; Wei Han

doi:10.3788/LOP202259.1215013

Laser & Optoelectronics Progress, Volume. 59, Issue 12, 1215013(2022)

Mechanical Abnormal Sound Detection Based on Self-Supervised Feature Extraction

Yingjie Xue¹, Qi Chen¹, Songbin Zhou^2、*, Yisen Liu², and Wei Han²

¹Faculty of Information Engineering and Automation, Kunming University of Science and Technology, Kunming 650500, Yunnan , China

²Institute of Intelligent Manufacturing, Guangdong Academy of Sciences, Guangdong Key Laboratory of Modern Control Technology, Guangzhou 510070, Guangdong , China

show less

Abstract Get PDF(in Chinese)

Figures & Tables(15)

Fig. 1. Process of self-supervised feature extraction

Download full size

Fig. 2. Power spectrum of pink noise

Download full size

Fig. 3. Process of pink noise generation by Matlab

Download full size

Fig. 4. Flowchart of deeply separable convolution

Download full size

Fig. 5. Structure of Bottleneck module

Download full size

Fig. 6. Structure diagram of AE

Download full size

Fig. 7. Flowchart of abnormal sound detection for mechanical equipment

Download full size

Fig. 8. Time-frequency spectra of test sets of normal sound samples and abnormal sound samples

Download full size

Fig. 9. Time-frequency spectra of generating abnormal samples

Download full size

Fig. 10. Visualization of original time-frequency characteristics of test samples

Download full size

Fig. 11. Visualization of test sample self-supervised feature extraction

Download full size

Fig. 12. AUC results of six feature extraction methods for anomaly detection

Download full size

Table 1. Network structure of MoblienetV2
View table
Table 1. Network structure of MoblienetV2
Operator Expansion factor Channel Repeated times Stride
Conv2D - 16 1 2
Bottleneck 1 8 1 1
Bottleneck 6 16 2 2
Bottleneck 6 16 3 2
Bottleneck 6 32 4 2
Bottleneck 6 48 3 1
Bottleneck 6 80 3 2
Bottleneck 6 160 1 1
Conv2D - 1280 1 1
Avg pool - 1280 1 -

Table 2. Data partition of four kinds of machine sound
View table
Table 2. Data partition of four kinds of machine sound
Type Number of samples in training sets Number of normal samples in testing sets Number of abnormal samples in testing sets
Slider 2804 400 890
Valve 3291 400 479
Pump 3349 400 456
Fan 3675 400 1475

Table 3. Performance comparison of different anomaly detection models

View table

Table 3. Performance comparison of different anomaly detection models

Model	Slider D_AUC /%	Valve D_AUC /%	Pump D_AUC /%	Fan D_AUC /%	Average D_AUC /%
SSFE-IF	94.6	85.2	82.9	74.5	84.3
SSFE-Kmeans	93.8	82.7	80.2	73.7	82.6
SSFE-OCSVM	85.6	78.8	74.5	65.3	76.1
SSFE-GMM	94.5	89.9	87.6	76.8	87.2
SSFE-DAGMM	94.5	91.3	88.5	78.1	88.1
SSFE-CAE	94.8	91.2	89.0	78.2	88.3
SSFE-AE	95.0	92.7	88.2	78.0	88.5

Tools

Get Citation

Copy Citation Text

Yingjie Xue, Qi Chen, Songbin Zhou, Yisen Liu, Wei Han. Mechanical Abnormal Sound Detection Based on Self-Supervised Feature Extraction[J]. Laser & Optoelectronics Progress, 2022, 59(12): 1215013

Download Citation

EndNote(RIS)BibTex Plain Text

Set citation alerts for article

Save article for my favorites