Adaptive Wavelet Transform Front End Integrated With Multitask Learning for Underwater Acoustic Target Recognition

Underwater acoustic target recognition (UATR) faces challenges from complex, non-stationary ship-radiated noise, which comprises a mixture of broadband, narrowband, and transient components requiring multi-scale analysis. To address this, we propose an end-to-end framework named adaptive wavelet transform with multi-task learning for underwater acoustic target recognition (AWTM-UATR). This end-to-end framework integrates an adaptive wavelet feature extraction (AWFE) front-end with learnable parameters and a joint recognition-reconstruction network (JRRN) employing multi-task learning (MTL). Additionally, it incorporates an attention-gated interaction module (AGIM) for synergistic feature fusion. The AWFE provides adaptive time-frequency analysis tailored to signal characteristics within this framework. The JRRN then learns robust features, leveraging the reconstruction task as a regularizer for the primary classification task. The AGIM further enhances classification by enabling attention-gated information flow from the reconstruction pathway. Meanwhile, the channel attention module and sparse regularization constraint (CAM & SRC) refine the features entering the JRRN. Experiments on the ShipsEar and DeepShip datasets show that AWTM-UATR achieves highly competitive performance. Furthermore, ablation studies and visualization analyses confirm the effectiveness of the key components and demonstrate the framework’s capability for robust UATR.