Bearing Diagnosis Accuracy Comparison Using Convolutional Neural Network with Time/Frequency Domain Signals

Deep learning is the most attractive topic in the field of machine learning and relevant applications. Owing to the strong learning ability of the convolutional neural network (CNN), it integrates the feature extraction from raw data and classification as a complete learning process and makes the bearing fault diagnosis intelligent. In the published results, the inputs of the CNN may be the raw temporal waveform of vibration, its processed waveform or converted 2D images. In this paper, focusing on the diagnosis accuracy of rolling bearings, a comparative study is conducted among the inputs using the raw temporal waveform, the frequency spectrum, and the envelope spectrum of a vibration signal. First, an appropriate classification model based on the CNN is constructed. Then, experimental data from bearing with real damages are collected and then transformed and converted into some small gray pixel images for training and testing the CNN model. Finally, the classification accuracies using three signals are compared. The results indicate that the diagnosis performances using the above three signals are close when the trained CNN models are stable; among them the model using the frequency spectrum of the vibration signal is a little better than the models using the other two signals, which may be a reference for further investigating the deep learning used in the field of bearing diagnosis.