A Data Scientific Approach Towards Predictive Maintenance Application in Manufacturing Industry

Paper i proceeding, 2022

Most industries have recently started to harness the power of data to assess their performance and improve their production systems for future competitiveness and sustainability. Therefore, utilization of data for obtaining insights through data-driven approaches is invading every domain of industrial applications. Predictive maintenance (PdM) is one of the highest impacted industrial use cases in data-driven applications due to its ability to predict machine failures by implementing machine learning algorithms. This study aims to propose a systematic data scientific approach to provide valuable insights by analysing industrial alarm and event log data, which might further be used for investigation in root cause understanding and planning of necessary maintenance activities. To do that, a Cross-Industry Standard Process for Data Mining (CRISP-DM) is followed as a reference model in this study. The results are presented by first understanding the relationship between alarms and product types being processed in the selected machines by using exploratory data analysis (EDA). Along with this, the behavior of problematic alarms is identified. Afterward, a predictive analysis formulated as a multi-class classification problem is performed using various Machine Learning (ML) models to predict the category of alarm and generate rules to be used for further investigation in maintenance planning. The performance of the developed models is evaluated based on the different metrics and the decision tree model is selected with the higher accuracy score among them. As a theoretical contribution, this study presents an implementation of predictive modeling in a structured way, which uses a systematic data scientific approach based on industrial alarm and event log data. On the other hand, as a practical contribution, this study provides a set of decision rules that can act as decision support for further exploration of possible in-depth root causes through the other contextual data, and hence it gives an initial foundation towards PdM application in the case company.