Simulation-trained AI-System for Two-Stage Fault Detection and Diagnosis of Rolling Bearings in Industrial Applications

Abstract

Rolling bearings are an essential component in manufacturing equipment and can lead to production losses if a failure is not detected in time. Furthermore, the data obtained in many industrial applications is not labelled, so that a classification of intact and defective bearings is often not possible. This paper presents an approach for an AI-system that is trained with a bearing fault simulation and can then be deployed in industrial applications. A method for bearing fault simulation is presented which is based on the superposition of characteristic fault frequencies of a bearing. With the data from the simulation, an AI-system is implemented with a two-stage approach for fault detection and diagnosis. First, a semi-supervised k-nearest neighbor distance measure is calculated using synthetic data from the simulation. From this, a suitable value for the number of neighbors k is defined and the threshold at which a sample is considered faulty is determined. Then the AI-system is deployed in an industrial application and real world data is analysed with the given value for k and the threshold. In the second stage, the detected faulty samples are classified with regard to the type of fault using a decision tree, which is also trained with the data from the simulation. This approach is validated with two different real world bearing datasets and the results show the effectiveness of the presented approach.