Anomaly Detection of Three-Phase Induction Motor Based on Unsupervised Learning Using Frequency Distortion Characteristics

Anomaly detection and process monitoring are crucial technologies for controlling industrial site processes. The application of predictive maintenance technology, which has the potential to revolutionize the operations of factories and industrial organizations, is a prime example of the Industrial Internet of Things. This technology is currently attracting significant attention and acting as a catalyst for the growth and development of smart factories. The detection of defects in induction motors, which are widely used as essential components in various industrial fields, is crucial. Through the implementation of an anomaly detection model, maintenance costs can be reduced. Additionally, the lifespan of the motor can be extended through timely replacement and supplementation of parts. In this paper, current data samples from a three-phase induction motor are used to detect motor anomalies by analyzing their spectral characteristics. The distortion of the spectrum that occurs when an induction motor malfunctions is formulated and utilized as a feature for this purpose. An algorithm for anomaly detection, based on unsupervised learning and multivariate kernel density estimation (MKDE), is adopted as a model. The performance of the algorithm is evaluated based on accuracy, the area under the ROC curve (AUC), and the F1 score. In the evaluation process, the Youden index is used to determine the optimal threshold value. In order to determine the bandwidth of the Gaussian kernel used in MKDE, various existing approaches are applied to evaluate its performance. The MKDE model developed through the experiment shows an AUC of 0.9663, an accuracy of 0.8971, and an F1 score of 0.9041.