(672c) Data-Driven Predictive Monitoring and Operation Support for Change-over Processes in Biopharmaceutical Drug Product Manufacturing

In biopharmaceutical drug product manufacturing, product-change-over is a crucial but time-consuming process. This step is necessary to establish and maintain a sterile production environment, which is essential to guarantee the product quality. During the change-over process inside the isolator, both the exterior and interior surface of the production machinery require intensive cleaning/disinfection, achieved through hydrogen peroxide decontamination, and clean-in-place (CIP)/sterilization-in-place (SIP), respectively. Numerous sensors are available to constantly monitor the process and provide real-time measurements of process variables, whereby the stored sensor data is only used for backtracking actions. In our work, we exploit the available data to develop a system for predictive monitoring of the process and early failure detection.

The data is initially prepared for exploitation through several steps. First, a dimensionality reduction step through principal component analysis (PCA) is conducted followed by a data cleaning and classification step, where the runs are classified into failed and successful. The boundaries of successful runs are then defined and are further referred to as “Golden Zone.” Machine learning algorithms such as Random Forest and k Nearest Neighbor are used to predict the performance of a run following a few minutes of operation. The quality of the prediction is assessed, and the distance to the “Golden Zone” boundaries is measured to guide the decision-making process by the operators in case abnormalities are detected.

The application of this predictive monitoring approach for early failure detection to the historical data of a hydrogen peroxide decontamination process of an industrial filling line at Hoffmann La Roche in Switzerland yielded a 50% reduction in the number of repeated runs due to failures leading to significant time and financial savings.

A complementary root cause analysis is also conducted to deepen the understanding of the complex system where a limited understanding of the causes of failure is provided. Pattern recognition within the analyzed data is used to identify different sources of errors and to recognize the full impact of certain abnormalities or failures (such as leaks) on the system operation and performance.

We further explore opportunities to improve the machine learning algorithms to achieve the earliest possible failure detection points, i.e., to maximize the time distance from the predicted failure to the occurrence of a recorded alarm. This could be especially helpful in the case of hydrogen peroxide decontamination where failures can be detected before the introduction of the hydrogen peroxide to avoid the need to repeat the aeration process which is the most time-consuming step in the procedure.