Data-driven anomaly detection and diagnostics for changeover processes in biopharmaceutical drug product manufacturing

Abstract

A two-step approach is presented for anomaly detection and diagnosis in batch process systems. The approach was applied to a case study of a decontamination process in biopharmaceutical drug product manufacturing. Analysis of historical operation data from four consecutive years was combined with process knowledge to achieve a deeper understanding of process failures, their impacts, and precursors. For anomaly detection, failure boundaries were established within the data space for different types of errors and process steps. The failure boundaries helped in both the detection and diagnosis of faults as each fault occupied unique boundaries. Precursors were identified as data from “non-failed” runs start venturing into the failure boundaries. The use of a rolling window enabled the differentiation of shifts due to failure in operations from shifts due to regular maintenance and calibration works. For diagnosis, a root cause analysis was used to identify irregularities in “non-failed” runs and to detect underlying equipment problems requiring preemptive intervention. Overall, the approach successfully addressed challenges in anomaly detection for batch runs, including drifting, overlapping, and imbalanced data. The integration of process understanding into the data analysis played an important role in bridging the information gaps. The gained insights are valuable for operation support, leading to a reduction in downtime and higher production efficiency. The approach can be used to help identifying sources of failures and the required corrective actions as an intermediate-term goal between fault detection and predictive maintenance. The proposed approach is robust and could be extended to other changeover and production operations.