A Deep Learning Method for Bearing Fault Diagnosis Based on Time-Frequency Image

Rolling element bearing is a critical component in rotating machinery that reduces the friction between moving pairs. Bearing fault diagnosis is always considered as a research hotspot in the field of prognostics and health management, especially with the application of deep learning. Deep learning, such as a convolutional neural network (CNN), can extract features automatically compared with traditional methods. However, the construction of the CNN model and the training process still need a lot of prior knowledge, and it takes a lot of time to build an optimal model to achieve a high classification accuracy. In addition, great challenges of universal applicability exist when different input forms (e.g., different sampling lengths or signal forms) are considered. This paper presents a universal bearing fault diagnosis model transferred from a well-known Alexnet model, and only the last fully connected layer needs to be replaced, which could reduce prior knowledge and extra time in establishing a new model. Accordingly, it is necessary to convert a raw acceleration signal to a uniform-sized time-frequency image, even when these data have different sizes. Furthermore, standardized images created by eight time-frequency analysis methods are applied to validate the effectiveness of the proposed method in two case studies. The results indicate that this method can be applied in bearing fault diagnosis, and t-SNE helps to understand the process of feature extraction and condition classification.