Distance-Based Explainable Anomaly Detection for Electrical Machines Through Vibration Signals

Electrical machines are an important part of industrial applications like automated production lines such as conveyor belt drives. Mechanical failures are a major cause of shutdowns. To reduce downtime or worsening of the damage, a condition monitoring system is required. Therefore, analysing vibration data for condition monitoring in rotating machines at an early stage using data-driven methods has become an important topic, with many possible solutions available. This paper focuses on anomaly detection using distance-based classifiers and statistical features from the time domain to detect anomalous behaviour. The presented method can be applied to various motor types and can be trained directly on the process after designing a new drive. The example of bearing anomaly detection is provided solely for illustrative purposes. It should be noted that the process can be applied in a multitude of ways. A method for selecting the appropriate distance, based on the hypersphere volume was presented, and its performance was compared with typical one-class classifiers such as one-class support vector machine (OC-SVM) and isolation forest (IF). In addition, this paper presents two methods for determining the features that have the greatest impact on classifier prediction in the case of anomalous behaviour. The methods were evaluated for detecting bearing faults using the Case Western Reserve University (CWRU) bearing data set and an own data set, achieving nearly error-free classifier performance.