In mechanical industry, about 30-40% of the failures of rotating machines are caused by bearing failures. Therefore, it is very important to monitor the condition of bearings in real time and detect faults in order to plausibly diagnose rotating machinery failures.
The ResNet-50 deep learning model has been widely considered for bearing defect diagnosis research. However, the ResNet-50 model requires a lot of computation time owing to its high computational complexity even though it generates good performance.
In this paper, we propose a lightweight CNN model that shows excellent bearing defect diagnosis performance in terms of both accuracy and computation time. In the proposed model, 1x1 convolution is applied to the step before entering the dense layer to reduce the complexity of the CNN model. As well, we consider various CNN models with different complexity by varying a number of the convolutional layers. In addition, input size reduction was applied to further reduce the complexity of the proposed CNN model. The experimental results using the CWRU Bearing Dataset show that the CNN lightweight model proposed in this paper has better classification performance than the defect diagnosis SOTA model implemented by the ResNet-50 model, and the computation time is significantly reduced compared to the SOTA model.
Using the model developed in this paper, it is expected that bearing defects are able to be accurately and quickly diagnosed without high specifications and high costs.