A vibration-based fault diagnosis method for rolling bearings via optimized wavelet-SVM fusion

Rolling bearings are critical components of rotating machinery, and their health status directly affects the operational reliability of equipment. This paper proposes an optimized wavelet-SVM fault diagnosis method based on multi-source vibration signal fusion: Three-channel inputs are constructed by synchronously collecting vibration signals from the drive end and fan end, along with their differential signals; Wavelet packet decomposition is utilized to extract frequency-domain features such as unit node energy entropy and wavelet coefficient standard deviation, while dimensionless indicators independent of rotational speed (kurtosis factor/waveform factor/impulse factor) are introduced to enhance time-domain characterization; The fused features are input into an RBF-SVM classifier after dimensionality reduction via PCA (retaining ۹۹% variance, reducing dimensions from ۱۰۲ to ۴). Experiments indicate that on the CWRU dataset, this method achieves ۹۷.۰% precision, ۹۶.۹% recall, and an F۱-score of ۹۶.۹% (representing a ۲.۹% improvement over single-source input methods); Although there is a ۲.۴% absolute accuracy gap compared to deep learning solutions, it possesses significant edge advantages—memory usage is only ۱۲KB and inference latency is ۰.۶ms—providing a high-precision, low-cost embedded solution for rotating machinery fault diagnosis

کلیدواژه ها:

Statistical features ، wavelet packet decomposition ، Gaussian kernel function ، Principal component analysis ، Prediction accuracy