Classification-Driven Contribution Analysis with Median Calibration for fault class and variable co-identification

Simultaneous and accurate identification of fault classes and variables is essential for effective maintenance in industrial systems. However, traditional statistical methods for fault variable identification struggle with the complexities of deep learning-based classifiers. Existing deep learning approaches often approximate nonlinear operations with linear or local linear models, resulting in imprecise contribution evaluations. Moreover, they face challenges in identifying fault directions for multivariate faults due to the influence of noise and contributions from fault-free variables. To address these limitations, this paper proposes a novel framework, Classification-Driven Contribution Analysis with Median Calibration (CDCA-MC), for fault class and variable co-identification. The framework introduces a classification-driven contribution to quantify the influence of variable deviations on a classification-driven monitoring index. It further extends backpropagation to compute accurate derivatives within nonlinear feature extraction layers of deep learning-based classifiers. To enhance robustness, a median calibration strategy is employed to mitigate the impact of noise and disturbances on contribution analysis, ensuring reliable identification of fault variables. The proposed CDCA-MC framework is validated on two case studies: the Continuous Stirred Tank Reactor (CSTR) and the Fixed-Wing Unmanned Aerial Vehicle (FW-UAV), demonstrating its effectiveness in fault diagnosis.