Optimization automating monitoring based on classification for rolling bearing

This work presents an automated and optimized methodology for detection and tracking real-time rolling bearings defects. It contains three sequential loops. The first loop is the initialization, as well as the constitution of a class called healthy. It starts by extraction of features from the acquisition signals, then select the most correlated features to reduce the unimportant features by the Relief method, normalization data by z-score, and reduce the dimension by t-SNE method, ended with a calculation of OPTICS parameter’s Epsilon and Minpts. The second loop is the detection, which has the same initialization steps, besides, OPTICS for data classification. The third loop is the follow-up: this phase consists of projecting the matrix of defect features into the final base that has made it possible to detect the defect class and to track the evolution of this class over time. The proposed methodology has been validated numerically and experimentally for rolling bearing in the outer race.