Research on Underwater Gliders Progressive Anomaly Detection Method Based on Multi-head Attention Mechanism

Progressive anomalies have an incalculable impact on long-endurance platforms such as Autonomous Underwater Gliders (AUGs), making their detection essential. This paper constructs an unsupervised learning framework based on a multi-head attention mechanism to analyze multi-source sequential data and develop an algorithm for detecting progressive anomalies in AUGs. Firstly, an orthogonalized bidirectional Long Short-Term Memory (LSTM) algorithm is introduced to remove the strong correlations in feature data and achieve source decoupling. Secondly, a random mask strategy is incorporated into the encoder to enrich the data features. Then, prior and sequential associations are added to the encoder-decoder, and a minimax strategy is employed to amplify the association discrepancy, making the anomaly features more obvious. Finally, the model is trained and tested using actual experimental data to validate the detection effectiveness of the proposed algorithm. Compared with the existing unsupervised anomaly detection algorithms, the algorithm in this paper has a high detection rate, with an average value of more than 98%, which can provide a strong basis for the further processing of the subsequent anomalies.