Research On Distributed Fiber Optic Sensing Event Recognition Method Based On Deep Learning

Distributed fiber optic sensing systems have been widely used in perimeter security for their ability to achieve remote location monitoring with high sensitivity and spatial resolution.However,in practical applications,the complex application environment can lead to problems such as low accuracy of event recognition and poor real-time performance.Currently,researchers mainly classify distributed fiber optic sensing event recognition into two types of data features: digital signals and images.Based on this research background,this article builds a Distributed Acoustic Sensing(DAS)system and collects six common events.By using digital signal and image features,the article conducts research on distributed fiber optic sensing event recognition based on deep learning and achieves some important progress.The main research content is as follows:(1)The research and development status of distributed fiber optic sensing event recognition was investigated and summarized,with a focus on the limitations of machine learning in fiber optic sensing event recognition,including low detection efficiency and the need for manual feature extraction.Additionally,the use of deep learning methods in fiber optic sensing event image recognition was reviewed,revealing issues such as highly complex models,significant computational and hardware resource requirements for training and inference,and high time costs.(2)A DAS system was constructed,and the collected event data was preprocessed.Based on the DAS technology principle,the DAS system was set up,and six events were collected,including car passing,walking,foot stomping,knocking,digging,and rainfall.The collected event data was preprocessed using frame segmentation and normalization methods to obtain the event digital signal data set.Additionally,the Markov Transition Field(MTF)method was used to convert the digital signals into images,resulting in an event image data set.(3)A 1DCNN-BiLSTM method for distributed fiber optic sensing event digital signal recognition was proposed.Traditional machine learning methods for event digital signal recognition have limitations such as low detection efficiency and the need for manual feature extraction.In this study,a one-dimensional convolutional neural network(1DCNN)and a bidirectional long short-term memory recurrent neural network(BiLSTM)were combined for event digital signal recognition.The 1DCNN automatically extracted event signal features,and the BiLSTM was used to explore the time correlation of event digital signals.To improve the event classification performance of the model,appropriate Re LU activation functions and Adam optimization algorithms were selected.The experimental results showed that the1DCNN-BiLSTM method was significantly superior to 1DCNN,BiLSTM,1DCNN-LSTM,LSTM deep learning methods,as well as traditional machine learning methods such as SVM,RVM,and BP neural network.The recognition accuracies for the six events,including car passing,walking,foot stomping,knocking,digging,and rainfall,were 95.8%,95.1%,94.8%,96.3%,98.3%,and 100%,respectively,and the average recognition accuracy reached 96.7%.(4)We propose a fiber optic sensing event image recognition method based on Markov Transition Field(MTF)and knowledge distillation.In the event image recognition method,one of the main bottlenecks of fiber optic sensing event image recognition is how to achieve both the accuracy of large-scale models and the efficiency of small models.In this study,we used MTF-processed event images as the dataset,Mobile Net V3-Large as the teacher model,and two knowledge distillation methods,namely,intermediate feature layer and output feature layer,to compress our customized lightweight student model(KDS).Experimental results show that the KDS model has significantly higher recognition accuracy and efficiency compared to the other five baseline models.The recognition accuracy for the six event images of the car,walking,stomping,knocking,digging,and rainfall is 95.5%,95.8%,94.0%,100%,94.3%,and 100%,respectively,with an average accuracy of 96.6%,and the model has only 0.35 M parameters,demonstrating good event recognition accuracy while maintaining lightweight performance.