Deep Multiscale Subdomain Adaptive Network for Cross-Condition Interturn Short Circuit Fault Diagnosis

With its powerful feature extraction and early fault identification capabilities, deep learning-based diagnosis of interturn short circuit (ITSC) faults has gained increasing attention to ensure the reliable and safe operation of induction motors. However, its practical deployment in industrial settings faces a significant challenge: performance degradation due to distribution shifts in current signals under different working conditions. To overcome this issue, we propose a deep multiscale subdomain adaptive network (DMSAN) for cross-condition ITSC fault diagnosis. The framework first employs parallel multiscale convolutional kernels with residual connections to capture multiscale fine-grained information from raw current signals. Subsequently, a subdomain adaptation layer integrated with multikernel local maximum mean discrepancy (MK-LMMD) is applied to align feature distributions across domains. Furthermore, a dynamic weight strategy is designed to optimize the classification loss and subdomain loss, thereby ensuring model convergence stability. DMSAN demonstrates superior performance and stability across fifteen cross-condition fault diagnosis transfer tasks, achieving an accuracy of 99.59%. This represents a substantial improvement compared to existing methods, including conditional domain adversarial network (CDAN) (91.63%), generative adversarial networks (GANs) (87.76%), domain adversarial neural network (DANN) (88.05%), and deep adaptive network (DAN) (82.83%).