Unsupervised Φ-OTDR anomaly detection based on feature-enhanced LSTM autoencoder

To address the challenge of labeling large datasets and scarcity of anomalous samples for phase-sensitive optical time-domain reflectometers (Φ-OTDR) in complex environmental monitoring, an unsupervised anomaly detection method based on dual sliding window preprocessing feature enhancement with long short-term memory (LSTM) auto-encoder (AE) was proposed. A preprocessing method containing dual sliding window operations was employed to effectively extract the feature-enhanced temporal sequences from phase signals, which were encoded and reconstructed by LSTM-AE, and the threshold was dynamically adjusted according to the reconstruction error (RE) of the original and output signals to achieve unsupervised anomaly detection. In the experiment, the vibration data of the suspended optical cable was collected for training the model to conduct the anomalous disturbance detection experiment. The results show that the proposed method without labeling the dataset achieves the detection accuracy of 94.85%, which is an improvement of 11.82% over the accuracy of the traditional AE. This research provides an effective unsupervised anomaly detection solution for the engineering applications of Φ-OTDR in the fields of power fiber optic cable monitoring and perimeter security.