Comparative Diagnosis of SVM, Random Forest and kNN for Bevel Gear Fault Detection Based on MODWT Denoising and Multi-Domain Features

This study compares three classical classifiers—support vector machine (SVM), random forest (RF) and k-nearest neighbors (kNN)—for bevel gear fault diagnosis based on maximal overlap discrete wavelet transform (MODWT) features. Tri-axial vibration signals were acquired for four pinion conditions (healthy, chipped tooth, broken tooth and worn gear) at four speeds and four loads, giving 64 recordings. The sym8 wavelet was selected objectively by combining the maximum energy to Shannon entropy ratio with a correlation criterion. After wavelet-domain thresholding, 75 multi-domain features were reduced to 12 by Fisher score ranking and correlation pruning, repeated inside every validation fold. Under leave-one-out cross-validation, the radial basis function SVM reached 98.4% accuracy (95% confidence interval, 91.7–99.7%), against 92.2% for kNN and 90.6% for RF, but these differences are not statistically significant (exact McNemar test, p = 0.125). When an entire speed was withheld from training, the ranking reversed: RF reached 93.8%, against 81.3% for kNN and 76.6% for SVM, and the RF–SVM difference was significant (Holm-adjusted p = 0.003). Withholding a load level affected the three classifiers similarly (85.9–89.1%). The most suitable classifier therefore depends on whether the training data cover the operating envelope of the monitored machine.