An active signal-noise decoupling network with superior noise generalisation for fault diagnosis of rotating machinery

In fault diagnosis, models trained on source domain samples from idealised environments with low and stable noise struggle to adapt to target domain samples where the noise is higher and dynamically varying in intensity. Therefore, this paper proposes an active signal-noise decoupling network (ASNDN). The network leverages source domain samples (hereafter referred to as pseudo-clean samples) and samples generated by adding noise to the source domain samples (hereafter referred to as pseudo-noise samples) to decouple fault features and noise features, and captures domain-invariant fault features through saliency-guided feature fusion, thereby enabling knowledge transfer from the source domain to the target domain. Specifically, a multi-scale feature extraction module is proposed to extract the latent features within the samples. Subsequently, a dual-channel decoupling structure is proposed to decouple fault and noise features within pseudo-clean and pseudo-noise samples. Additionally, a dual-head feature integrator is proposed to further suppress the noise components within the fault features, thereby enhancing the extraction of domain-invariant fault features. Finally, a joint optimisation strategy is proposed to facilitate the learning of domain-invariant fault features. The experimental results indicate that the proposed method demonstrates superior noise generalisation across three datasets, providing an effective diagnostic tool for engineering applications.