Domain adaptation method based on pseudo-label dual-constraint targeted decoupling network for cross-machine fault diagnosis

Abstract

In recent years, domain adaptation methods have gained widespread traction for addressing domain-shift problems caused by distribution discrepancy across different domains in fault diagnosis. Nonetheless, the significant variations in data distribution among cross-machine scenarios present obstacles to extracting domain-invariant features, ultimately leading to suboptimal recognition performance. To tackle this issue, a novel domain adaptation method based on pseudo-label dual-constraint targeted decoupling network is proposed. Initially, the targeted decoupling network (TDNet) is presented, employing a decoupling strategy that integrates convolutional feature channel separation with feature-targeted constraints. This strategy aims to extract domain-invariant features while alleviating the directional biases introduced by excessive constraints. Subsequently, the pseudo-label dual-constraint feature alignment (PDFA) method is introduced to effectively utilizes pseudo-label information. The PDFA minimizes confusion in pseudo-labels within the target domain while enforcing alignment constraints on concatenated pseudo-label features, ensuring efficient and precise cross-domain alignment while preserving inter-class discriminability. Additionally, batch attention (BA) is introduced to learn the intricate interdependencies among samples from the same batch, enriching feature representations to facilitate effective knowledge transfer. Experimental results based on twelve cross-machine tasks demonstrate the superiority of the proposed method in cross-machine fault diagnosis in comparison to existing domain adaptation techniques.