Early-warning industrial fault detection based on physics-guided residual learning and calibrated CRNNs

Industrial plants generate high volume multivariate sensor time series where early fault detection must be accurate, governable and explainable under regime changes, sensor correlations and drift. Many published Tennessee Eastman Process (TEP) studies report strong receiver operating characteristic (ROC) or accuracy but provide limited evidence for calibrated probabilities, uncertainty bounds and actionable sensor level diagnostics, so alarm thresholds remain ad hoc and difficult to defend in control rooms. To address this issue, we propose a physics guided pipeline that combines twin based residuals with compact rolling and spectral descriptors and an attention based convolutional recurrent neural network (CRNN). The proposed method is also compared against state-of-the-art (SOTA) baselines to establish its relative effectiveness under the same evaluation setting. Residuals highlight deviations from physics consistent behavior, while a one-dimensional convolutional neural network (Conv1D) and bidirectional gated recurrent unit (BiGRU) encoder with attention models short and long temporal structure over 120 step windows. SHapley Additive exPlanations (SHAP) guided feature selection reduces redundancy and improves interpretability, Optuna stabilizes training and Platt scaling calibrates anomaly probabilities for policy driven thresholding under a false alarm budget. Reliability is quantified with 1000 block bootstrap bias-corrected and accelerated (BCa) confidence intervals. On a run level 70/20/10 split of Tennessee Eastman Process runs, the method achieves about 99.0% accuracy with area under the receiver operating characteristic curve (AUC-ROC) and area under the precision-recall curve (AUC-PR) near 1.00, macro F1-score of 0.93 ± 0.02 and expected calibration error (ECE) reduced to ≈ 0.03 after calibration. Early warning governance improves the Numenta Anomaly Benchmark (NAB) score by 17% over an Isolation Forest baseline and supports low nuisance alarm operation. These results show that residual-guided learning plus calibrated decision governance can support deployment-oriented industrial monitoring by enabling auditable thresholds and operator-facing explanations within the Tennessee Eastman benchmark setting.