Acta Optica Sinica, Volume. 45, Issue 11, 1110002(2025)
Substation Intrusion Event Identification Method Based on DAS and GADF-CAFM-MSCNN
Figures & Tables(17)
Fig. 3. Structure of CAFM. (a) Channel attention module; (b) spatial attention module; (c) cross-attention module
Fig. 5. One-dimensional vibration curves for different intrusion events. (a) No invasion; (b) striking; (c) climbing; (d) trampling; (e) shoveling
Fig. 6. Corresponding GADF and GASF images of one-dimensional vibration curves. (a) No invasion; (b) striking; (c) climbing; (d) trampling; (e) shoveling
Fig. 9. Confusion matrices of CAFM-MSCNN for GADF and GASF testing sets. (a) GADF; (b) GASF
Fig. 10. Output characteristics of different parts of CAFM-MSCNN. (a) Original signal; (b) MSCNN; (c) CAFM; (d) output layer
Fig. 11. Accuracy curves and loss function curves of CAFM-MSCNN and MSCNN for training sets and validation sets. (a) CAFM-MSCNN, accuracy; (b) CAFM-MSCNN, loss function; (c) MSCNN, accuracy; (d) MSCNN, loss function
Fig. 12. Accuracy and loss functions for training sets and validation sets of four traditional models. (a) CNN, accuracy; (b) CNN, loss function; (c) TCN, accuracy; (d) TCN, loss function; (e) LSTM network, accuracy; (f) LSTM network, loss function; (g) CNN-LSTM, accuracy; (h) CNN-LSTM, loss function
Table 1. Compositions of training set, testing set and validation set
View tableTable 1. Compositions of training set, testing set and validation set
Category Total Training set Validation set Testing set No invasion 903 723 90 90 Striking 856 686 85 85 Climbing 1085 869 108 108 Trampling 1190 952 119 119 Shoveling 1344 1076 134 134
Table 2. Results of MSCNN and CAFM-MSCNN for training set and validation set
View tableTable 2. Results of MSCNN and CAFM-MSCNN for training set and validation set
Model Training loss Training accuracy Validation loss Validation accuracy MSCNN 0.014440 99.63% 0.07462 97.81% CAFM-MSCNN 0.005915 99.94% 0.04669 98.13%
Table 3. Results of training sets and verification sets for CNN, TCN, LSTM network and CNN-LSTM
View tableTable 3. Results of training sets and verification sets for CNN, TCN, LSTM network and CNN-LSTM
Model Training loss Training accuracy Validation loss Validation accuracy CNN 0.2148 93.13% 0.2141 92.84% TCN 0.3451 87.76% 0.3301 89.65% LSTM network 0.4462 83.20% 0.4499 84.02% CNN-LSTM 0.2831 91.26% 0.3319 89.31%
Table 4. Classification results of testing sets for different models
View tableTable 4. Classification results of testing sets for different models
Category Accuracy /% CNN LSTM network TCN CNN-LSTM MSCNN CAFM-MSCNN Average 95.22 83.33 90.43 91.37 97.66 98.49 No invasion 100.00 100.00 100.00 100.00 100.00 100.00 Striking 89.41 59.79 83.51 84.54 91.76 97.65 Climbing 95.37 88.70 89.57 87.83 98.15 97.22 Trampling 95.80 83.49 93.58 91.74 99.16 98.32 Shoveling 95.52 84.68 85.48 92.74 99.25 99.25
Table 5. Testing time of testing sets for different models
View tableTable 5. Testing time of testing sets for different models
Model CNN LSTM network TCN CNN-LSTM MSCNN CAFM-MSCNN Time /s 1.47 0.01 0.07 0.04 1.66 1.72
Tools
Get Citation
Copy Citation Text
Zhiniu Xu, Tianjie Ma, Yangyang Cai, Lijuan Zhao. Substation Intrusion Event Identification Method Based on DAS and GADF-CAFM-MSCNN[J]. Acta Optica Sinica, 2025, 45(11): 1110002
Paper Information
Category: Image Processing
Received: Nov. 26, 2024
Accepted: Apr. 27, 2025
Published Online: Jun. 24, 2025
The Author Email: Lijuan Zhao (hdzlj@126.com)
CSTR:32393.14.AOS241793
© Copyright 2018-2021 | Chinese Laser Press.
All Rights Reserved 沪ICP备15018463号-20