Event- vs. time-driven: comparing applicability of sequential data types for machine learning-based reactive fault detection in manufacturing

The implementation of reliable machine learning for reactive fault detection in manufacturing requires large volumes of high-quality data. Time-series, one of the main data types in modern manufacturing, is available in high quantities, but also leads to high computational effort in model training and large number of duplicates in the data. Event-driven sequential data is another data type, used in manufacturing, reducing duplicates in data. A comparison of both sequential data types in this context is not fully explored in the literature. This paper explores a set of experiments to evaluate both sequential data types using various ML methods and windowing mechanisms on a real-world manufacturing system. Based on the results, recommendations for practitioners are derived and an outlook to possible future investigation is given.