Run-Level Fault Detection and SHAP-Based Diagnosis of Persistent Classification Difficulty in the Tennessee Eastman Process

Reliable fault detection in nonlinear process systems requires both accurate classification and interpretable analysis of faults with weak or near-normal signatures. This study develops an explainable, data-driven, and temporally informed multiclass framework for the Tennessee Eastman Process (TEP), retaining all 21 operating conditions (20 fault types and the normal operating condition). Six sliding-window statistics were extracted from 52 process variables and classified using Extreme Gradient Boosting (XGBoost). Performance was evaluated at both sample and run levels through fault-specific window analysis, an a priori validation-driven hierarchical decomposition, SHapley Additive exPlanations (SHAP), a Relative Sensitivity Index (RSI) based on detection-delay sensitivity, and computational benchmarking. Aggregating sample-level predictions (macro F1 = 0.8823) into run-level decisions via majority voting improved performance substantially (macro F1 = 0.9515). Across five seeded repetitions, the mean run-level macro F1-score was 0.9522±0.0035 (95% CI: ±0.0044). Under the primary evaluation, 18 of 21 classes achieved F1 ≥0.97. Fault 3 benefited strongly from extended temporal context, whereas Normal operation, Fault 9, and Fault 15 retained a structured but asymmetric confusion pattern dominated by Normal–Fault 9 errors. SHAP identified model-specific attribution patterns associated mainly with cooling-water-related variability features, while RSI indicated greater prediction-stream sensitivity for the historically difficult faults. Feature extraction and inference required approximately 6.2 ms on CPU and 35.3 ms on GPU, negligible relative to the 180 s sampling interval. These findings indicate that temporal context, run-level aggregation, and explainability can jointly support accurate data-driven fault diagnosis while revealing persistent fault-specific ambiguity.