Time-series anomaly detection in machine monitoring data based on extended iForest

In the monitoring of industrial equipment operational conditions, sensor data frequently encounter challenges such as missing values, misalignment, and abnormal fluctuations, often resulting from harsh environmental factors or sensor malfunctions. These issues significantly impair data integrity and reduce the accuracy of equipment health assessments. Many existing anomaly detection approaches depend heavily on substantial volumes of labelled normal and faulty samples, and they exhibit limited capability in addressing the complex characteristics of high-dimensional signals encountered in practical industrial settings. To address these challenges, an anomaly detection method based on the extended iForest (EIF) is introduced, incorporating a sliding window mechanism to enhance adaptability to temporal patterns and improve detection stability. The proposed method facilitates dual-level identification by detecting both anomalous temporal segments and point-level anomalies within those segments. Experimental evaluation on both synthetic signals and real bearing vibration datasets demonstrates that the method attains high accuracy and robustness in identifying diverse types of anomalies, including missing data, signal shifts, and amplitude inflation.