Bearing Fault Recognition Based on Signal Enhancement and an Improved Convolutional Neural Network

Abstract

To address the challenges of bearing fault recognition, a hybrid intelligent method based on signal processing and convolutional neural networks (CNN) is constructed. A test bench based on a worm gear reducer is built to collect vibration signals with four typical fault conditions: inner race crack, outer race crack, rolling element crack, and cage fracture. The raw signals are processed through slicing, IIR low-pass filtering, notch filtering, Fast Fourier Transform (FFT), and spectrum energy feature extraction to establish a fault feature database. A CNN-based fault recognition model is developed and trained using the AdamW optimization algorithm and Mean Squared Error (MSE) loss function. An improvement is proposed to reduce the overfitting problem of the neural network. The results show that after 100 iterations of training, the model achieves a recognition accuracy of nearly 100% for different bearing fault types, verifying the effectiveness of the proposed method for intelligent bearing fault diagnosis in complex noise environments.