Optimized Multiclass Support Vector Machines Using Empirical Features for Compound Fault Diagnosis in Roller Element Bearings

Abstract Taper roller bearings are crucial components in rotating machinery. Undetected faults like pitting, unbalance, and misalignment can degrade performance and cause unexpected failures. Conventional diagnostic approaches often treat these faults in isolation or rely heavily on data-driven models without integrating physical insight. To overcome these limitations, this study proposes a novel hybrid method. It combines Dimension Theory Modeling (DTM) with Support Vector Machines (SVM) for robust fault diagnosis and classification. Vibration signals are collected from an instrumented bearing test rig under controlled fault scenarios, including single and compound faults. Fast Fourier transform (FFT) extracts frequency-domain features linked to bearing defect frequencies and modulation effects. A DTM-based dynamic model is formulated to identify faults that capture interactions among fault types and reflect severity. Experimental results demonstrate that the DTM-SVM framework achieves precise fault quantification with an average estimation error of 3.42%. The extracted features are classified using a multiclass SVM with an radial basis function (RBF) kernel. It achieves 99.16% accuracy across healthy, misaligned, unbalanced, inner race, outer race, clearance, and compound fault conditions. The proposed approach offers a reliable and physically interpretable solution for real-time condition monitoring of taper roller bearings in industrial settings.