Feature fusion-based fiber-optic distributed acoustic sensing signal identification method

Abstract

Fiber-optic distributed acoustic sensing (DAS) systems based on phase-sensitive optical time-domain reflection (Φ-OTDR) technology have been widely used for perimeter security and oil and gas pipeline safety monitoring. To address the problem of low recognition accuracy of high-sampling-rate long-sequence signal data (length greater than or equal to 1000 points) collected by the DAS system, we propose a CDIL-CBAM-BiLSTM network model based on feature fusion. The model uses a modified circular dilated Convolutional Neural Network (CDIL-CNN) to extract detailed temporal structure information from each signal node, and combines it with bidirectional long short-term memory (Bi-LSTM) network using feature fusion to dig deeper into the data. Meanwhile, a convolutional block attention module (CBAM) was introduced to improve the model performance. The experimental results based on 5040 training samples and 2160 test samples show that the proposed model can achieve an average recognition accuracy of more than 99% for six real disturbance events under perimeter security scenarios, and the recognition time was less than 2 ms. In addition, our method achieved the highest recognition accuracy compared with other methods used in the experiments and can be extended to other areas, such as pipeline safety monitoring and industrial inspection measurements.

Improved Deep Learning Model for Workpieces of Rectangular Pipeline Surface Defect Detection

“…This is very useful in detection and segmentation tasks. On the one hand, larger receptive fields can detect and segment large targets, and on the other hand, higher resolutions can accurately locate targets [21]. Hollow convolution can arbitrarily increase the receptive field without introducing additional parameters.…”