Pattern Recognition in Distributed Optical Fiber Acoustic Sensing Based on Stacked Long Short-Term Neural Networks

A stacked long short-term memory neural network (Dual-LSTM) is proposed to address the challenges of strong noise and recognition difficulties in Distributed Optical Fiber Acoustic Sensing (DAS) signals. Based on the Long Short-Term Memory (LSTM) architecture, the Dual-LSTM utilizes both amplitude and phase information of signals, enhancing its capability in noise processing and classification. Experimental datasets were constructed using optical fiber cable intrusion event data collected by a DAS system from FreeSound audio resources. Results demonstrate that the Dual-LSTM achieves an average test accuracy of 99%, outperforming the I-LSTM (using only intensity information) and P-LSTM (using only phase information) by more than 50%. This indicates that the Dual-LSTM exhibits superior performance in signal denoising and pattern recognition, enabling significantly higher identification accuracy.