A semi-supervised underwater acoustic target recognition method based on domain adversarial neural networks

Existing underwater acoustic target passive recognition methods face the challenges of insufficient cross-domain generalization capabilities and the absence of adaptive mechanisms for unknown targets. To address these issues, this article proposes a joint-learning model for adversarial neural networks with maximum mean discrepancy (JT-MMD-AN). The method integrates the maximum mean discrepancy (MMD) metric into the adversarial domain adaptation framework, achieving cross-domain feature alignment by constraining the optimization process of domain classifiers, while the information contained in the one-dimensional spectrum and two-dimensional Mel-frequency cepstral coefficients (MFCC) features is learned simultaneously through the joint training mechanism. For open-set recognition, a three-stage identification mechanism is designed, comprising feature template construction, dynamic threshold determination, and similarity matching. In closed-set cross-domain tasks on the DeepShip dataset, the proposed algorithm achieves 73.83% average recognition accuracy using only limited labeled target data, representing a 11% improvement over classical domain adversarial neural network (DANN) methods. In open-set testing scenarios, it attains 63.44% accuracy for unknown category identification, outperforming traditional methods by approximately 35%. Experimental results demonstrate that the JT-MMD-AN method effectively mitigates domain shift issues while balancing the requirements of known-class discrimination and unknown-class detection under limited annotation conditions.