High-performance event recognition method with MSCSE-BiLSTM for the Φ-OTDR sensing system

Benefiting from advances in artificial intelligence algorithms, distributed acoustic sensing based on phase-sensitive optical time-domain reflectometry (<i>Φ</i>-OTDR) has attained high event recognition accuracy through the application of diverse learning models. Nevertheless, further improving recognition accuracy remains a persistent challenge. In this paper, we propose a hierarchical fusion vibration event recognition method for a <i>Φ</i>-OTDR sensing system, which integrates a multi-scale convolutional neural network, a squeeze-and-excitation attention mechanism, and a bidirectional long short-term memory (BiLSTM) network. The resulting architecture, referred to as MSCSE-BiLSTM, is designed to enhance feature extraction and temporal sequence modeling for improved recognition performance. Experimental results based on 14,603 samples of a four-class field engineering vibration event dataset collected by <i>Φ</i>-OTDR, containing car, manual tapping, road breaker, and excavation, demonstrate that the proposed method achieves an average validation accuracy of 99.02% and an average F1-score of 99.28%, respectively, which exhibits significant advantages compared to typical models, including support vector machine, convolutional neural network (CNN), and CNN long short-term memory. Additionally, through comparative <i>t</i>-distributed stochastic neighbor embedding visualization reveals that the proposed method can transform originally overlapping and entangled categories in <i>Φ</i>-OTDR signals into well-separated and compact clusters in the low-dimensional feature space, which further offers an intuitive validation of its recognition capability.