A Comparative Study on Multi-Condition Bearing Fault Diagnosis Methods Based on Machine Learning

Abstract

In industrial fields, bearing fault diagnosis often faces problems such as noise interference, variable operating conditions, and scarce samples. Based on the multi-load bearing dataset from Case Western Reserve University, this paper extracts 15 time-domain and frequency-domain features, and compares the diagnostic performance of random forest, support vector machine, prototypical network and a 1D-CNN deep learning baseline. To avoid overestimating generalization ability caused by random data splitting, a leave-one-load-out cross-validation experiment is further designed for cross-working-condition evaluation. The experimental results show that: (1) Under noise-free conditions, both traditional machine learning models achieve 100% accuracy, while the 1D-CNN baseline reaches 96.25%. (2) The anti-noise ability and load adaptability of random forest are significantly better than SVM and 1D-CNN; it maintains 90.42% accuracy at 0 dB strong noise and remains 100% stable across all 0-3 HP load conditions even in cross-load generalization tasks. (3) The prototypical network performs best in extreme small-sample scenarios, with 99.71% 1-shot accuracy under clean conditions. (4) Feature importance analysis confirms that frequency-domain features have stronger robustness in noisy environments, while some impact-sensitive time-domain features show negative contribution under strong noise.