Rolling Bearing Fault Detection Based on Evolving Approaches

The detection of faults in rolling bearings is essential for ensuring the reliability of industrial systems, given the critical role these components play in rotary equipment. Rolling bearings are particularly susceptible to degradation under harsh operating conditions, which can lead to severe mechanical failures. This paper presents two complementary, evolving-based approaches for detecting bearing faults using the Typicality and Eccentricity Data Analytic (TEDA) algorithm. TEDA is a dynamic, unsupervised method tailored for low-latency outlier detection, leveraging two core measures: typicality, which assesses similarity to historical data, and eccentricity, which quantifies deviation. Approach 1 follows a three-stage pipeline: (1) segmenting vibration signals into time windows, (2) extracting statistical features from each window, and (3) applying TEDA for fault detection. Approach 2 adopts a more direct strategy: (1) organizing raw vibration signal segments into time windows and (2) applying TEDA without explicit feature extraction. Both methods were evaluated using the widely recognized Case Western Reserve University (CWRU) bearing dataset. Experimental results show that both approaches deliver competitive and, in some cases, superior performance compared to traditional methods, even without a post-processing stage. Notably, Approach 1 achieved an average precision of 1.0000 and an F1-score of 0.9750 for class 1, outperforming a referenced state-of-the-art method that reached an F1-score of 0.9638. In addition to proposing two complementary detection pipelines, this work contributes by modifying the TEDA algorithm to enhance detection efficiency. Specifically, updates to TEDA’s internal variables are halted for detected outliers, streamlining processing and eliminating the need for additional post-processing stages.