Anomaly Detection in Time Series Data Using Unsupervised Machine Learning and Statistical Methods

Anomaly detection in signal data is crucial for identifying unexpected patterns and deviations without labeled data, making it suitable for real-world scenarios with infrequent and diverse anomalies. The implementation of machine learning-based detection methods is becoming popular due to their efficiency and time-saving nature. Effective anomaly detection is in constant need in the automotive industry, where it can be applied in functional integrity checks, calibration optimization, failure detection, regulatory compliance standards, and more. Due to NOx emissions from combustion engines having detrimental effects on the environment, this paper evaluates the effectiveness of unsupervised learning methods, including one-class SVM and isolation forest, alongside statistical techniques such as modified z-score and empirical analysis, in detecting anomalies within the engine out NOx mass flow signal. By providing a detailed comparison and analysis, this paper offers insights into their practical applications in time series data.