(109h) Data-Driven Approaches Towards Real-Time Equipment Health Monitoring and Predictive Maintenance in Drug Product Manufacturing

Digitalization is the central pillar towards the introduction of Industry 4.0 in manufacturing. However, the pharmaceutical industry lags behind in this regard and sophisticated approaches in terms of data analysis are still seldom used despite the availability of an abundance of recorded process data. Highly regulated fields such as drug manufacturing require manufacturers to store process data over many years for backtracking purposes. The analysis of such stored data highlights the potential to use it for predictive maintenance purposes.

This work presents a data-driven approach for equipment monitoring and the implementation of predictive maintenance (PdM) in an aseptic filling line in drug product manufacturing. Multiple years’ worth of industrial data are used for this purpose. The data mainly consists of Supervisory Control and Data Acquisition (SCADA) variables which incorporate parameters such as pressure, temperature or vibration data. The goal of this work is to replace currently implemented time-based maintenance strategies with more efficient predictive and condition-based ones. The data obtained from the production facility varies in their source and structure based on their main intended purposes. The data is therefore first preprocessed and prepared. In this work, different machine learning algorithms are compared to identify the most suitable given the structure of the available data (i.e. the installed sensors at different locations in the line), the information content as for certain areas specific measurements are recorded (e.g. changeover processes), and the resulting prediction accuracy. Random forest algorithms are the most commonly applied in this regard due to their high performance potential with limited information incorporated in the data and an avoidance for over-fitting. Artificial neural networks are also popularly applied for a diverse set of PdM problems followed by support vector machines and k-means models. Ensemble learning or combinations of the different algorithms are also tested. A detailed analysis of the results obtained by the different algorithms is provided. The best strategies for each equipment are determined in terms of maintenance frequency.

A decision support tool is proposed to determine if an equipment is expected to fail in the foreseeable future. Expert knowledge is also then incorporated in the tool to determine the appropriate course of action identified with each type of failure or countermeasures needed to prevent failure. Different maintenance strategies are compared based on the expected downtime, uncertainty of the decision and the associated costs. Such a tool can be easily integrated with other facility scheduling decision tools.