Shapelet Transformation of Multivariate Time Series for IoT Anomaly Detection

The proliferation of Internet of Things (IoT) devices has generated substantial volumes of multivariate time-series data in multiple domains. Such data are susceptible to anomalies that may indicate system malfunctions or security threats. This research introduces a novel shapelet transformation approach for classifying multivariate time-series data to improve anomaly detection in IoT systems. Our approach distinguishes itself from the classic Shapelet Transform by specifically optimizing the extraction process to handle high-dimensional IoT data more efficiently, contrasting with methods such as Fast Shapelets, which are primarily designed for speed without focusing on multivariate contexts. This methodology centers on extracting short, informative subsequences, known as shapelets, from time series to facilitate classification. This approach is validated using the industrial IoT fault detection dataset for predictive maintenance in automation, which contains 1,000 entries of sensor data collected from machines in an industrial automation environment. This dataset includes three main sensor measurements: vibration (mm/s), temperature (°C), and pressure (Bar). The evaluation process involves partitioning the data into training and test sets and employing cross-validation to ensure robustness. The performance of the proposed method is benchmarked against traditional algorithms. Results demonstrate notable improvements: the F1-score is 0.6451 for temperature, 0.5778 for vibration, and 0.6984 for pressure, with an overall accuracy of 94%. This study establishes a framework for enhancing IoT system reliability by advancing anomaly detection, data mining, and machine learning.