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

Area of Science:

• Artificial Intelligence
• Optical Sensing Technologies
• Signal Processing

Background:

• Distributed acoustic sensing (DAS) using phase-sensitive optical time-domain reflectometry (Φ-OTDR) leverages AI for vibration event recognition.
• Existing methods face challenges in further enhancing recognition accuracy for complex vibration signals.

Purpose of the Study:

• To propose a novel hierarchical fusion vibration event recognition method for Φ-OTDR systems.
• To improve feature extraction and temporal modeling for enhanced recognition performance.

Main Methods:

• Developed the MSCSE-BiLSTM architecture, integrating a multi-scale convolutional neural network (CNN), squeeze-and-excitation (SE) attention, and a bidirectional long short-term memory (BiLSTM) network.
• Utilized a dataset of 14,603 Φ-OTDR vibration event samples (car, manual tapping, road breaker, excavation).

Main Results:

• The MSCSE-BiLSTM method achieved an average validation accuracy of 99.02% and an average F1-score of 99.28%.
• Demonstrated superior performance compared to traditional models like Support Vector Machine (SVM), CNN, and CNN-LSTM.
• t-distributed stochastic neighbor embedding (t-SNE) visualization confirmed improved class separability in the feature space.

Conclusions:

• The proposed MSCSE-BiLSTM method offers a significant advancement in Φ-OTDR vibration event recognition.
• The hierarchical fusion approach effectively enhances feature representation and sequence modeling capabilities.
• The method provides a robust and accurate solution for classifying diverse field engineering vibration events.