Anomaly Detection in Predictive Maintenance using Dynamic Time Warping

Abstract

Manufacturing systems face the fundamental challenge of efficient operation by leveraging vast amounts of real-time data collected through technological advancements such as artificial intelligence and machine learning. Maintenance systems have evolved to predict and manage equipment failures in advance, with data-driven fault detection being a crucial technology. However, most related research has been limited to single equipment for specific processes, making the direct application in actual manufacturing settings that use various equipment models or types challenging. When using multifacility models, the most crucial aspect is the analysis of variations and errors in the data collected from each facility. To mitigate the risk associated with a sole vendor, different models of equipment is used strategically, even for the same functionality. Consequently, collecting temporally mismatched data is prevalent. The current methodology, which has been predominantly focused on a single-facility approach, faces limitations in its application when dealing with unstructured, unlabeled data, or temporally mismatched data obtained across multiple facilities. This study employed the dynamic time warping (DTW) method to analyze discrepancies in time-series data obtained from multiple equipment groups by leveraging similarity analysis of data peak matching for anomaly detection. Specifically, an approach called auto time windowing is adopted to extract signal periods based on the detailed signal analysis results of the process, enabling the application of DTW. The auto time windowing allows for the accurate automated analysis of signal period by overcoming the limitations of analysis errors caused by noise in the existing data using the threshold of the actual signal. This methodology is validated for two different equipment groups involved in a real-world production process, where parts are attached to products. The results of this study demonstrated an improvement over conventional time-series analysis methods such as the Euclidean method, addressing errors that may occur. This research enhances the analysis theory using DTW for the actual problem of data discrepancies among multiple equipment groups in the manufacturing field, which is not previously considered in existing predictive maintenance (PdM) theories. This validation through case studies effectively contributes to expanding the utilization of PdM.