Rolling Bearing Fault Diagnosis Based on 2D CNN and Hybrid Kernel Fuzzy SVM

Abstract

Addressing the issues of poor anti-noise performance and insufficient mining of fault information in vibration signals in traditional fault diagnosis methods, a network structure algorithm (2DCNN-HKFSVM) based on the 2D convolutional neural network (CNN) and the hybrid kernel fuzzy support vector machine (HKFSVM) is proposed. First, the original bearing vibration signals are converted into 2D grayscale images; then, these grayscale images are used as inputs to the 2D convolutional neural network for feature extraction and dimensionality reduction; finally, the obtained feature vectors are passed to the hybrid kernel fuzzy support vector machine for fault detection. Compared with the support vector machine (SVM), the fuzzy support vector machine (FSVM) assigns different weights to different bearing fault samples through the fuzzy membership function, thereby reducing the impact of noise on the classification results. Furthermore, the hybrid kernel function combined according to Mercer's theorem enables the FSVM to take both global and local fitting into account, further improving the classification performance of the FSVM. Compared with some existing fault diagnosis models that combine CNN with machine learning algorithms such as SVM and random forests (RF), 2DCNN-HKFSVM exhibits better generalization ability and anti-noise performance.