Learnable wavelet packet kernel guided deep discriminative dynamic joint domain adaptation network for cross-machine...

Recently, intelligent diagnosis methods based on unsupervised domain adaptation have played an important role in ensuring the long-term safe operation of mechanical systems. However, these methods still face the challenges of poor interpretability of network models, low quality of predicted pseudo-labels, and insufficient learning of discriminative features, which restrict the improvement of fault diagnosis accuracy and reliability. To tackle these issues, this study proposes a learnable wavelet packet kernel-guided deep discriminative transfer network (WPKDDTN) for cross-machine fault diagnosis under strong noise conditions. Firstly, to better capture domain-invariant and discriminative time-frequency fault features, a learnable wavelet packet kernel transform (LWPKT) layer is proposed, which integrates trainable wavelet filters and wavelet threshold functions into the wavelet packet transform. Secondly, an entropy-enhanced dynamic joint distribution alignment (EDJDA) criterion is proposed to optimize the joint distribution discrepancies between the source and target domains, where the improved entropy-enhanced local maximum mean discrepancy (ELMMD) incorporates information entropy to enhance the confidence of pseudo-labels. Then, considering the intra-class compactness and inter-class separability of the classification features learned by the network, an additional large margin (ALM) Softmax loss is designed to enhance the discriminative feature learning ability of the classifier. Finally, experimental results from cross-machine transfer tasks under varying noise levels demonstrate that the proposed WPKDDTN method outperforms other leading fault diagnosis methods, exhibiting strong noise robustness and generalization capability.