RBC-AD: conformal anomaly detection with explicit false-alarm control for the Tennessee Eastman Process

Fault detection in the Tennessee Eastman Process (TEP) is challenged by transient dynamics, post-injection regime transitions, and score drift that destabilize fixed-threshold detectors and increase false alarms. In the proposed Risk-Budgeted Conformal Anomaly Detection (RBC-AD) framework, a calibration-based anomaly detector integrates a causal forecasting backbone, frequency-aware scoring, and conformal decisioning. The forecaster is a dilated temporal convolutional network with causal self-attention, trained on fault-free runs to predict one-step-ahead Gaussian parameters by minimizing negative log-likelihood. Detection then scores each window using a fused nonconformity measure that combines the window-averaged predictive log-likelihood with an FFT-based frequency deviation term, using robust normalization fitted on a held-out calibration split and tuning the fusion weight under leakage-controlled splits. Conformal quantiles at miscoverage level α=0.05 define normal, uncertain, and anomaly states with finite-sample correction. Run-level partitioning over 500 runs per split and post-injection sampling (3,000 normal windows, 8,000 fault windows) achieves ROC-AUC = 1.000 and PR-AUC = 1.000 under the post-injection-only protocol; full-stream results provide the deployment-relevant assessment. Static conformal thresholds (θu=0.1977, θa=0.3033) yield F1 = 0.9989 and F1 = 0.9999. Deployment selects an uncertainty threshold to satisfy an empirical normal alarm-rate target of 0.05. Full-stream evaluation reports uncertainty rate and detection delay under k-consecutive alarm logic, with optional drift-triggered recalibration using normal-stream scores.