To address the issue that traditional Park's vector method is susceptible to interference from non-fault factors such as voltage unbalance when identifying interturn short-circuit faults in stator windings

a fault diagnosis method integrating data-driven preprocessing and deep learning was proposed. Firstly

a negative-sequence current compensator based on fully connected neural network was constructed to filter out the negative-sequence current components caused by non-fault factors through a data-driven approach. A Park's transformation was then applied to the processed current signals to extract fault features. Subsequently

a convolutional neural network(CNN)model integrated with convolutional block attention module(CNN-CBAM)was established for fault classification. The attention mechanism was utilized to enable model to adaptively focus on key features. A 5.5 kW prototype motor capable of simulating inter-turn short circuits of varying severity was constructed and employed to conduct experimental validation. Results demonstrate that the proposed method achieves a diagnostic accuracy of 97.31% for inter-turn short-circuit faults of varying severity under voltage imbalance scenarios. This research provides an effective method for the identification of inter-turn short circuit faults in induction motors during the latent period.