Toward Digital-Twin-Enabled Bioprocess Monitoring: Fault-Inclusive Soft Sensing of Penicillin Concentration Under Process Deviations

Data-driven soft sensors can estimate fermentation product concentrations from routinely recorded process variables, but strong performance during normal operation does not establish reliability during process deviations. This study evaluated current-time penicillin-concentration estimation using 100 simulated IndPenSim batches comprising 90 normal-operation batches and ten documented deviation batches, with 113,935 observations in total. Initial normal-trained models and five follow-up experiments examined complete-batch validation, dependence on batch-progress features, phase-specific error, model-family comparisons, early out-of-distribution warnings and empirical prediction ranges. These analyses motivated a matched comparison between normal-only and fault-inclusive HistGradientBoosting regressors using 36 current and causal history-based inputs. Normal performance was evaluated in five regime-balanced complete-batch folds, and deviation performance by leave-one-fault-batch-out evaluation; every tested deviation batch was excluded from its own model fit. Batch-balanced sample weights were used, with a factor of three assigned to permitted deviation batches in fault-inclusive fitting. On held-out deviation batches, pooled RMSE decreased from 3.195 to 2.564 g/L, a reduction of 19.73%, while MAE decreased from 2.076 to 1.441 g/L and R2 increased from 0.8580 to 0.9085. Normal-operation RMSE was nearly unchanged at 1.981 and 1.983 g/L. Eight of ten deviation batches improved. The mean paired fault-batch RMSE difference was -0.7063 g/L, with a descriptive 95% batch-bootstrap interval of -1.2844 to -0.2194 g/L. Improvement was largest within the first-to-last recorded fault-reference window, but late-stage errors persisted. Batch 100 remained poorly predicted, with fault-inclusive RMSE of 6.631 g/L and R2 of -2.4837. A fault-risk classifier, Isolation Forest OOD detector and empirical error ranges offered incomplete reliability information. Fault-inclusive training therefore improved this benchmark on average, but did not establish generalization to unseen fault mechanisms, calibrated safety warnings or deployment in physical fermentation.