Semantic-Based Anomaly Detection Approach for Large-Scale Time Series Data in Acceleration Events

Automobile manufacturers face challenges in detecting errors due to the complexity of interconnected vehicle components and rely on rapid analysis of large datasets. During endurance testing, vehicle fleets are driven under various conditions while measurement data is collected. These measurements help experts develop engines and components with precision. Data mining methods compress large datasets, extract key information for anomaly detection, and save analysis time. This article introduces the use of the Dynamic Time Warping (DTW) approach for analyzing measurement signals, enabling insights into event curve shapes. A novel artifact was developed to integrate this approach into the engine development process, streamlining anomaly detection for practical application. Individual events are stored in subsets, allowing them to be used for further training in supervised learning approaches. The study focuses on unlabeled time series data from Mercedes-Benz endurance testing, particularly on anomalous decelerations (hesitations), using clustering-based anomaly detection.