Clarke-Park Transformation-Based Input Configuration for CNN Diagnosis of Stator Faults in Three Phase Induction Motors Using Motor Current Signature Analysis

Motor current signature analysis (MCSA) with convolutional neural networks (CNNs) is widely applied to three-phase induction motor fault diagnosis. Conventional input configurations, such as direct three-phase RGB mapping or discrete wavelet transform (DWT), primarily depend on inter-phase amplitude and phase differences, limiting their ability to capture subtle physical responses, especially for inter-turn short circuits (ITSC). This study proposes a Clarke–Park transformation (CPT)-based input configuration that extracts $I_{\alpha }$ , $I_{\beta }$ and $I_{d}$ components and assigns them to the red, green, and blue channels, respectively. This design simulatneously represents spatial current distribution ( $I_{\alpha }$ , $I_{\beta }$ ) and dynamic response ( $I_{d}$ ) in the rotating reference frame, enhancing physical interpretability and discriminative capability. Comparative experiments using identical raw data and the ResNet50 model evaluated CPT against RGB mapping and DWT inputs. CPT achieved the highest test accuracy (99.51%), with notable consistency in the challenging ITSC class (F1-score: 0.993) and perfect classification of the 4Turn class. Grad-CAM visualizations confirmed that CPT produces more localized, fault-focused activations than RGB mapping, comparable to DWT, without requiring complex frequency transforms. These results demonstrate that CPT effectively organizes current signal features, enabling faster convergence, improved classification stability, and high diagnostic accuracy, offering a physically meaningful input design for reliable motor fault diagnosis.