Abnormal Event Identification via Deep Spatio-Temporal Feature Extracting of Imbalanced Distributed Optical Fiber Vibration Signals

Distributed Acoustic Sensor (DAS) technology possesses considerable potential for perimeter security and diverse surveillance monitoring applications. By capitalizing on fiber vibration signal data examination and artificial intelligence methodologies, the detection and classification of abnormal vibrations becomes feasible. This paper introduces an abnormal event identification technique for distributed optical fiber sensing signals, which display interconnected temporal series and spatial structure attributes. The suggested method surpasses conventional statistical machine learning algorithms in recognition and classification precision. In this strategy, a Convolutional Long Short-Term Memory (ConvLSTM) unit is utilized to devise an encoder-decoder network for the extraction of spatiotemporal features. Subsequently, these extracted elements are input into an Extreme Gradient Boosting Tree (XGBoost) classifier for abnormal event classification. Moreover, a Time Variation Autoencoder (Time-VAE) based temporal series data generation model is proposed to tackle the challenge of imbalanced abnormal samples in optical fiber vibration signal data. A pragmatic case study corroborates the efficacy of the proposed technique in augmenting the classification accuracy of irregular optical fiber vibration signals. It provides invaluable theoretical backing for the wider adoption of DAS technology in security monitoring applications.