SmartFerment AI: Fault-Stratified Anomaly Detection with Conformalized Uncertainty Quantification for Fed-Batch Fermentation Monitoring: A Simulation Benchmark with Leakage-Free Evaluation

Abstract

SmartFerment AI is an anomaly classification, variational autoencoder (VAE) reconstruction scoring, fed-batch sequence forecasting, uncertainty quantification (with Isolation Forest and Conformalized Quantile Regression) and leakage-free monitoring framework. The framework is tested on 24 simulated E. coli fed-batch runs (Monod model, 2,328 time points) which are clearly divided in model fitting, ensemble weight selection, conformal calibration and sealed evaluation without temporal data leakage. The study focuses on four common gaps in bioprocess monitoring: (G1) temporal leakage, (G2) no evidence of ablation, (G3) no statistical verification and (G4) missing calibration uncertainty intervals. The performance of detection is measured using the ROC-AUC with AUC = 0.891 (BCa 95% CI: 0.743–0.989) for the sealed test set of 14 positive cases and 374 negative cases on the specific task of detecting dissolved-oxygen-crash. The aggregate AUC is higher (0.9438, BCa 95% CI: 0.896–0.981) but is significantly affected by a stress-test pH excursion fault and is considered secondary. One-step forecasting outperforms the other algorithms at shorter timescales (1-step) due to the high level of lag-1 autocorrelation, while XGBoost's performance is significantly better than Persistence, Linear Regression, Random Forest, and GRU at operationally relevant timescales of 5 and 10 steps (e.g., RMSE of 14.81 g/L for Persistence, 5.97 g/L for Linear Regression, 14.81 g/L for Random Forest, and 14.81 g/L for GRU, versus 1.01 g/L for XGBoost at 5 steps). Compared to Monte Carlo Dropout configurations with sub-nominal coverage (PICP = 92.1% and 87.4%) at a 95% target, CQR achieves a near-nominal coverage. Post-hoc isotonic calibration improves the Maximum Calibration Error by 40.2% and the Expected Calibration Error by 24.5% leading to better probability reliability of XGBoost anomaly scores. A six configuration ablation analysis and Two One-Sided Test (TOST) equivalence testing identify that XGBoost is the best model in the ensemble and is found to be statistically equivalent to the full ensemble model within ±0.05 AUC, thus it is recommended as the single-model detector. For the Monod-kinetics simulation, the most important features for making anomaly decisions are aeration efficiency, pH, and agitation rate, which yield process interpretable explanations and not causal claims for physical fermentation systems. All findings are restricted to simulated data, and external validation to PenSimPy and Tennessee Eastman Process benchmarks is identified as the major roadmap to industrial applicability, followed by physical bioreactor data.