φ-OTDR Pattern Recognition Based on LSTM-CNN

Ming Wang; Zhou Sha; Hao Feng; Lipu Du; Dunzhe Qi

doi:10.3788/AOS221468

Abstract

Results and Discussions By comparing accuracy, precision, recall, and other parameters of test set (Figs. 9-11), it is found that the convolution operator and LSTM have significantly improved the performance of ANN. In specific event recognition, LSTM-CNN is always in the best state. According to the comparison of the validation set (Table 2, Fig. 12, and Fig. 13), the generalization and classification accuracy of LSTM-CNN are proved. For instance, in the recognition of excavation and walking signals, LSTM-CNN shows significant advantages, and the precision and recall rates exceed CNN by 10% to 15%. By comparing LSTM-CNN with four machine learning algorithms including SVM, KNN, decision tree, and random forest (Table 3, Fig. 14, and Fig. 15), it is proved that deep learning is effective in multi-classification problems with large batches of samples and with great superiority in generalization and classification accuracy. However, in the recognition of noise, SVM, decision tree, and random forest all achieve precision and recall of 100%. Therefore, the machine learning algorithms can be designed as a front-end procession mechanism of the neural network, and the non-threatening signals can be eliminated by taking advantage of its short response time and small data processing volume, so as to improve the efficiency of pattern recognition.Objective

With the advantages of long monitoring range, excellent anti-interference ability, accurate event location, and mature measurement principle, phase sensitive optical time domain reflectometer (φ-OTDR) has become a non-destructive detection method that has emerged in recent years and is widely used in pipeline safety maintenance, pig positioning and tracking, intrusion warning, and other fields. Threat warning and pattern recognition are the two main tasks of optical fiber sensing. On the one hand, it needs to quickly respond to possible threat events. On the other hand, countermeasures for different types of threats should be different. Therefore, it is necessary to identify event types. To simulate common threat events in engineering, we apply a self-developed φ-OTDR integrated chassis to collect four types of excitation signals including background noise, excavation, motor vibration, and walking. The established goal of pattern recognition is accomplished by using neural networks and machine learning classifiers, respectively. In addition, deep learning and machine learning have different characteristics in pattern recognition. Deep learning requires multiple rounds of iterative learning on a large batch of samples and converges to an optimal state through error back-propagation and parameter update. In contrast, machine learning has the advantages of fewer training samples, fast response time, and short training period, but it is highly dependent on the effectiveness of feature extraction. Therefore, the performance difference between the two kinds of classifiers is also one of the focuses for comparison in this paper.

Methods

We employed a short-time Fourier transform (STFT) to extract the time-frequency domain features of the signal while utilizing the long short-term memory network (LSTM) to extract the time-series features. On this basis, a one-dimensional convolution operator was employed to extract contour features, and a two-dimensional convolution operator was used to extract spatial features. Usually, the training period of a recurrent neural network (RNN) is long, so the LSTM needed to be trained separately to determine the optimal model parameters, and thus the LSTM parameters were directly called in the subsequent classifier training. With the comparison of the accuracy, precision, recall, and other evaluation indicators among an artificial neural network (ANN), traditional convolutional neural network (CNN), and LSTM-CNN (Table 1), the positive improvement of LSTM and CNN for classification performance was illustrated. The same training set and validation set were used to compare LSTM-CNN with four machine learning algorithms (Table 3), including support vector machine (SVM), K-nearest neighbor (KNN), decision tree, and random forest, and the superiority of deep learning compared with machine learning was analyzed in terms of generalization and classification accuracy. The time domain features were extracted by empirical mode decomposition (EMD), and the frequency domain energy features of the signal were extracted by discrete wavelet transform (DWT). The energy of decomposition curves of each layer of EMD and DWT was summed and normalized to construct eigenvectors, in which EMD and DWT were decomposed into six layers, respectively, and thus the eigenvectors contained 12 dimensions in total.

Conclusions

Aiming at the problem of φ-OTDR pattern recognition, we propose a neural network with LSTM and CNN as the main framework in this paper. With the time domain curve as the network input, we extract its time-series features through LSTM and then its contour and energy features via CNN and STFT, respectively. LSTM-CNN shows different degrees of superiority compared with ANN and traditional CNN. The accuracy of LSTM-CNN on the training set and validation set is 97.8% and 94.6%, respectively, and the accuracy of the validation sample set is 94.43%, which accomplishes the established goal of φ-OTDR pattern recognition. To make up for the drawback of the RNN in data processing speed, we utilize TensorFlow2.4GPU and CUDA11.1 to improve the data processing capability of LSTM-CNN, which provides a proof of concept and demonstration examples for actual engineering product development.