A cross domain processing deep transfer learning network for rotating machinery fault diagnosis

Abstract In the field of intelligent fault diagnosis of mechanical equipment, existing cross-domain diagnostic models based on transfer learning (TL) do not utilise the commonality information between the two domains in the data processing stage, which leads to the loss of transferable features that are essential for the cross-domain fault diagnostic task. To address this issue, this paper proposes a cross-domain processing deep TL network model (CDPDTLN), which consists of a cross-domain data processing (CDP) module, a feature extraction module and a domain-adaptive diagnostic module. In the CDP module, the adaptive multivariate variational modal decomposition algorithm is used to process the source and target domain data simultaneously, which preserving the common features between the two domains. In the feature extraction module, to realise the feature extraction work under various complex operating conditions, an improved multi-scale residual network is proposed to extract domain-invariant features. In the domain-adaptive diagnostic module, a combined domain distribution adaptation (CDDA) strategy is proposed to align the marginal and conditional distributions of the two domains. In the CDDA strategy, a weighted mean square discrepancy metric is defined by combining maximum mean discrepancy with maximum mean square discrepancy to enhance the distribution alignment and domain confusion capabilities. In multi-scenario cross-domain experiments, the diagnostic accuracy of the CDPDTLN model exceeds 95%. The results show that the proposed model can effectively retain and learn domain-invariant features, significantly improving the reliability and robustness of cross-domain diagnosis.