Causal Localization of Process Anomalies in Biopharmaceutical Fermentation

Abstract

Biopharmaceutical fermentation processes require continuous monitoring of temperature, pH, dissolved oxygen, agitation speed, feed rate, biomass concentration, pressure, conductivity, and metabolite levels. These variables interact through complex biochemical and mechanical mechanisms, making it difficult to identify whether an abnormal signal is a true process fault or a downstream response to another variable. This study proposes a fine-grained causal anomaly localization method for biopharmaceutical fermentation time series. The method combines domain-constrained causal discovery, temporal attention, and intervention-based residual scoring. First, a causal structure is learned under process constraints such as feed-to-growth dependency, oxygen-transfer relationships, and temperature-control feedback. Second, intervention-based residuals are computed by replacing suspected abnormal variables with causally expected values. Finally, abnormal variables are ranked according to their causal contribution to batch-level process deviation. Experiments are conducted on a fermentation monitoring dataset containing 1,120 production batches, 54 process variables, and one-minute observations collected from pilot-scale and production-scale bioreactors over 19 months. The dataset includes 72 million process records and 1,380 expert-labeled abnormal segments, including dissolved-oxygen control failure, feed-pump instability, pH drift, excessive foam formation, abnormal biomass growth, and temperature-control delay. The proposed method reduces median abnormal variable localization time from 16.9 minutes to 5.4 minutes compared with a reconstruction-only Transformer baseline. The false alert rate decreases to 1.8 cases per batch. The top-ranked causal variable matches expert investigation notes in 1,026 abnormal segments, and average event-boundary deviation is reduced by 12.7 minutes. Full batch assessment requires 4.6 minutes on a single GPU. The results demonstrate that causal inference can strengthen fine-grained anomaly localization in non-linear and highly coupled biopharmaceutical time series.