Benchmarking Control Charts and Machine Learning Methods for Fault Prediction in Manufacturing

This paper examines and benchmarks different approaches in their ability to detect and predict faults in manufacturing processes, based on real-world use cases and with respect to their differing dataset properties. Knowing about the occurrence of faults becomes more and more important in manufacturing due to increasing quality demands and legal guidelines. In addition, the complexity of manufacturing processes is constantly increasing. This stems from a higher product variance resulting from individual and customized products as well as additional external influences such as human errors, environmental factors and tool wear. As a result, today’s process data is often no longer normal distributed. Furthermore, data volume steadily increases, thereby opening new opportunities for data-driven analytics approaches. Frequently applied control charts for statistical process control (SPC) often lack the ability to deal with multiple variables and non-normal distributed data at the same time, since multivariate and nonparametric control charts are underrepresented in past research. Consequently, there is a need for new process control methods in manufacturing that are suitable for large amounts of data and cover diverse and dynamic distribution models. Therefore, machine learning models have been recognized as feasible approaches to meet these requirements. For comparison a Hotelling’s T² control chart, a K-Chart, an Isolation Forest, an ARIMAX model and a Neural Network have been implemented. We evaluate each method by missed detection rate (MDR), false alarm rate (FAR) and whether signals occurred before or after the faults. Real-world data sets of a commercial vehicle manufacturer serve as benchmarking basis.