Anomaly detection and critical attributes identification for products with multiple operating conditions based on isolation forest

Performance analysis of the existing mechanical products is critical to identifying design defects and improving product reliability. With the advances of information technologies, product operating data collected through continuous condition monitoring (CM) serve as main sources for analysis of performance and detection of anomaly. Most of the existing anomaly detection methods, however, are not effective when CM data are very high dimensional, leading to poor quality of assessment results. Besides, the effects of multiple operating conditions on anomaly detection are seldom considered in these existing methods. To solve these problems, an integrated approach for anomaly detection and critical behavioral attributes identification based on CM data is developed in this research. Gaussian mixed model GMM) is employed to develop a method for clustering of operating conditions. Isolation forest (iForest) method is used to detect anomaly instances, and further to identify the critical attributes related to product performance degradation. The effectiveness of the developed approach is demonstrated by an application with collected operating data of a wind turbine.