Squeeze-and-excitation attention residual learning of propulsion fault features for diagnosing autonomous underwater vehicles

Abstract

Given the demanding and unpredictable operational conditions, autonomous underwater vehicles (AUVs) often encounter different propulsion faults, leading to significant economic losses and mission impairments. To address this challenge, vibratory time-series features can be extracted for the precise propulsion fault diagnosis of AUVs. A squeeze-and-excitation (SE) attention residual network (SEResNet) is therefore put forward to enhance the feature extraction for AUV propulsion fault diagnosis. By leveraging the vibratory time-series data obtained from the AUV, an SE attention mechanism is embedded into a residual network. This integration facilitates the extraction of pertinent vibratory fault features, subsequently utilized for accurate diagnosis of any propulsion faults. The effectiveness of the proposed SEResNet was validated through its application to an actual experimental AUV, with comparison against the state-of-the-arts. The results reveal that the present SEResNet outperforms all other comparison methods in terms of diagnosis performance for AUV propulsion faults.