Author Affiliations
1State Key Laboratory of Precision Measuring Technology and Instruments, School of Precision Instrument and Opto-Electronics Engineering, Tianjin University, Tianjin 300072, China2Ningxia Hui Autonomous Region Water Conservancy Engineering Construction Center, Yinchuan 750004, Ningxia , Chinashow less
Fig. 1. Structure diagram of φ-OTDR
Fig. 2. Comparison of the data presentation form between the time-space map and the time-domain curve
Fig. 3. Operation process of CNN and LSTM. (a) 1D convolution; (b) 2D convolution; (c) LSTM
Fig. 4. Schematic of the internal structure of φ-OTDR integrated system
Fig. 5. Structure diagram of BRNN (LSTM)
Fig. 6. Training situation and test results of LSTM. (a) Curve of loss function with dynamic learning rate; (b) curve of loss function with fixed learning rate; (c) background noise; (d) excavation signal; (e) motor vibration signal; (f) walking signal
Fig. 7. STFT results of target signals. (a) Background noise; (b) excavation signal; (c) motor vibration signal; (d) walking signal
Fig. 8. Network structure of LSTM-CNN
Fig. 9. Comparison of overall evaluation indicators. (a) Train accuracy; (b) test accuracy; (c) train loss; (d) test loss
Fig. 10. Recall comparison of test set. (a) Background noise; (b) excavation signal; (c) motor vibration signal; (d) walking signal
Fig. 11. Precision comparison of test set. (a) Background noise; (b) excavation signal; (c) motor vibration signal; (d) walking signal
Fig. 12. Precision of validation sample. (a) Background noise; (b) excavation signal; (c) motor vibration signal; (d) walking signal
Fig. 13. Recall of validation sample. (a) Background noise; (b) excavation signal; (c) motor vibration signal; (d) walking signal
Fig. 14. Precision of validation sample of models in Table 3. (a) Background noise; (b) excavation signal; (c) motor vibration signal;(d) walking signal
Fig. 15. Recall of validation sample of models in Table 3. (a) Background noise; (b) excavation signal; (c) motor vibration signal; (d) walking signal
Net number | Net structure | Input | Number of CNN layers | Layer node | Learning rate |
---|
1 | ANN | Time-domain sequence | 0 | 36 | 0.01 | 2 | ANN | STFT | 2 | 80 | 0.05 | 3 | ANN | Time-domain sequence + STFT | 2 | 88 | 0.03 | 4 | CNN | Time-domain sequence + STFT | 4 | 90 | 0.001 | 5 | LSTM-CNN | Time-domain sequence + STFT | 4 | 90 | 0.001 |
|
Table 1. Important parameters of different neural network structures
Sample type | Number of sample points | Accuracy / % |
---|
Noise | Excavation | Motor vibration | Walking |
---|
Original | 327 | 218 | 478 | 710 | | Net 1 | 252 | 167 | 720 | 468 | 35.77 | Net 2 | 340 | 180 | 550 | 663 | 82.28 | Net 3 | 136 | 140 | 815 | 373 | 47.39 | Net 4 | 328 | 193 | 493 | 695 | 89.47 | Net 5 | 315 | 224 | 493 | 695 | 94.43 |
|
Table 2. Distribution of identification results and accuracy of the verification samples
Model type | Model 1 | Model 2 | Model 3 | Model 4 | Model 5 |
---|
Structure | LSTM-CNN | SVM | KNN | Decision-tree | Random-forest | Test accuracy /% | 94.60 | 83.65 | 83.71 | 78.17 | 87.38 | Validation accuracy /% | 94.43 | 81.98 | 65.66 | 75.24 | 83.26 |
|
Table 3. Comparison of LSTM-CNN and common machine-learning models