Research On Disturbance Signal Classification Of φ-OTDR Optical Fiber Sensing System

Distributed optical fiber sensing system based on phase-sensitive optical time-domain reflectometry(φ-OTDR)has been widely used in real-time dynamic monitoring fields such as perimeter security due to its long monitoring range and high sensitivity.With the improvement of optical systems,how to quickly and accurately classify disturbance events and reduce system false alarms in φ-OTDR has become the focus and difficulty in practical applications and theoretical research.After thoroughly investigating signal processing literature of φ-OTDR,for solving existing problem of current classification algorithm,an end-to-end hybrid disturbance event classification model is designed based on temporal convolutional network(TCN),bidirectional long short-term memory(Bi LSTM)and dual attention mechanism to improve the classification performance from the perspective of signal processing.Semi-supervised learning is also combined to further improve the classification performance on small number of labeled data,and solve the problem of difficult and time-consuming data labeling in the system.(1)Aiming at extracting spatio-temporal features of φ-OTDR signal,a disturbance classification model based on TCN-Bi LSTM is designed.It use dilated causal convolution of TCN to extract temporal causal features and Bi LSTM to extract bidirectional correlations in space domain.The final spatio-temporal features is used as the basis for classification.Experimental results show that TCN effectively improves the ability to distinguish between disturbance and non-disturbance events with 0% nusiance alarm rate(NAR)and 90.1% classification accuracy,better than other three models which are representative and perform well on classification.(2)After introducing dual attention mechanism,a disturbance classification network combining channel attention based TCN with spatial attention and Bi LSTM is designed to improve the overall classification performance,at the expense of increasing small number of network parameters.The extraction of time domain features is further optimized by introducing a channel attention mechanism.The model can process long spatial sequence which contains redundant information,and focus on real disturbance information when extracting spatial features due to spatial attention mechanism.The classification accuracy is 93.4% increased by 3.3% compared with TCN-Bi LSTM,and NAR is 0% with barely increased total training time.(3)Since deep learning relies on large amount of labeled data,while in φ-OTDR,the unlabeled data is easy to collect and label data is time-consuming and difficult,pseudosupervised initialization based on label propagation algorithm(LPA)and semi-supervised self-training methods are designed,combining semi-supervised learning with the existing disturbance event classification model.Experimental results show that when reducing the proportion of labeled data,both methods can improve classification performance.To be more specific,when the proportion of labeled data is less than 50%,semi-supervised self-training model performs better,accuracy is increased by up to 11.4%,which can still maintain a low NAR with 0%;when the proportion of labeled data is higher than50%,pseudo-supervised initialization method performs better on classification accuracy,with the maximum increase of 1.8%.However,semi-supervised self-training model still performs best on NAR,as low as 0%.