A CLSTM-CNN-Attention hybrid model and its application in fault diagnosis

Autonomous underwater vehicles (AUVs) are indispensable equipments in underwater detection, surveying, and investigation. As the main power source of AUV, the accurate and timely fault diagnosis of thruster plays key role in ensuring its safe navigation. However, this task is full of challenges due to the complication, vagueness, and randomness of underwater settings. To address the issues, a hybrid diagnosis model named CLSTM-CNN-Attention is proposed, which combines convolutional long short term memory (CLSTM), convolutional neural network (CNN) and attention mechanism. Specifically, it combines an improved hybrid network to realize the connection of long-term and short-term memory (LSTM) with CNN in parallel, which can capture the time-related and space-related fault information of input signal simultaneously. At the same time, a new linear rectification function is also introduced into the hybrid model to enhance its anti-interference capability. Finally, the diagnostic performance of the hybrid model is further improved by adding attention mechanisms, which could better focus on the fused information. Experimental and comparison results indicate that the suggested approach has remarkable interference suppression capacity and surpasses other relevant methods, demonstrating good performance in fault diagnosis of AUV thruster.