Gradient-Weighted Feature Attribution Drift in Transformer-Based Fault Diagnosis Systems: An Empirical Analysis of Industrial Conveyor Belt Anomaly Detection

Abstract

Transformer-based deep learning architectures have demonstrated significant potential in industrial fault diagnosis; however, the stability of gradient-weighted feature attribution maps under distributional shift remains insufficiently characterized. This study presents an empirical analysis of attribution drift phenomena in a Vision Transformer (ViT) model deployed for conveyor belt anomaly detection across three manufacturing facilities. Utilizing Grad-CAM++ and Integrated Gradients frameworks applied to vibration spectrogram inputs, we quantify attribution instability across 14,200 labeled fault samples spanning five fault classes. Results demonstrate that attribution drift correlates strongly (r = 0.81) with inter-facility sensor calibration variance, and that domain-adaptive fine-tuning reduces mean attribution entropy by 34.7%. The findings provide actionable calibration protocols for industrial AI deployment and establish a reproducible benchmark for explainability-aware fault diagnosis systems.