Study on Interpretability of Convolutional Neural Network for Electrical Characteristics Prediction of Automotive Wire Harness

The evaluation of automotive wire harnesses requires both interpretability and diversity in judgment criteria to enhance the practicality and reliability of diagnostic models. This study primarily focuses on optimizing heatmap generation for model interpretability by introducing score class activation mapping (Score-CAM) as a replacement for the previously used gradient-weighted CAM (Grad-CAM) in wire harness diagnosis. Score-CAM provides a more robust and accurate representation of feature importance as interpreted by a convolutional neural network (CNN), addressing limitations in stability and reliability inherent to Grad-CAM. Additionally, crosstalk voltage (XTV) is proposed as a supplementary diagnostic criterion alongside common mode current (CMC), enriching the evaluation framework by providing an additional perspective for comprehensive diagnosis. Experimental results demonstrate that the integration of Score-CAM significantly enhances heatmap quality and model interpretability, while the inclusion of XTV offers valuable complementary insights. These advancements collectively contribute to the development of more effective and interpretable diagnostic systems for automotive wire harnesses.