Non-stationary Compensated Multi-current Pattern (NCMP): A novel image-based motor fault diagnosis framework

Image-based convolutional neural networks (CNNs) have been studied for motor fault diagnosis by learning discriminative patterns from two-dimensional signal representations. However, under non-stationary and data-limited conditions, speed and torque variations distort these representations, causing fault patterns to deform and CNNs to learn operating-condition-dependent characteristics rather than intrinsic fault information. Existing approaches remain limited in three respects. First, image representations often lack a stable structural reference, resulting in inconsistent fault patterns across operating conditions. Second, image augmentation strategies rely largely on data-driven expansion with limited physical consistency, making it difficult to preserve fault-related structures. Third, image learning schemes remain generic, using uniform learning or generic reconstruction objectives with limited guidance toward fault-related components. To overcome these limitations, this paper proposes a Non-stationary Compensated Multi-current Pattern (NCMP) framework with three modules. First, NCMP constructs a structurally consistent representation by compensating for operating-condition variations and aligning three-phase current patterns. Second, structure-preserving augmentation extends conventional pixel-level cut–paste to pattern-level components, increasing data diversity while preserving fault structures. Third, reconstruction-based pretraining with tri-factor masking and classification fine-tuning promotes fault-discriminative representation learning. By integrating these modules, NCMP reframes image-based diagnosis as a fault-oriented structural learning paradigm in which representation, augmentation, and pretraining are jointly designed to preserve and emphasize fault-related structures. Across three motor platforms, the proposed framework achieved diagnostic accuracies of 98.1%, 98.4%, and 98.6%, with reconstruction errors of 0.05, 0.06, and 0.08, respectively.