Cross-machine deep subdomain adaptation network for wind turbines fault diagnosis

Recently, subdomain adaptation has gained extensive interest in addressing the problem of wind turbine (WT) fault diagnosis. However, current methods mainly focus on the subdomain adaptation of statistical features and scenarios with constant rotation speed. To overcome these limitations, a new cross-machine deep subdomain adaptation network (CMDSAN) is proposed in this paper for fault diagnosis of WT under multiple operating conditions. CMDSAN contains an improved subdomain adaptive (ISA) mechanism. In ISA, a subdomain distribution shift measure of jointed statistical and geometric features is constructed to boost domain confusion. Meanwhile, to further capture fine-grained information and discriminative features, a local correlation alignment (LCA) strategy is proposed. Additionally, a two-stage training trade-off factor is designed for balancing classification and ISA loss during the training process to improve the transferability of features. Subsequently, test rigs are constructed, i.e., a planetary gearbox test rig and a scaled-down test rig for WT gearbox with a reduction ratio of 110.11, to validate the effectiveness and superiority of CMDSAN. The case studies conducted under constant rotation speed, acceleration, and deceleration demonstrate that the proposed CMDSAN exhibits better fault transfer diagnostic ability than other domain adaptation methods.