Self-aligning ball bearing fault classification using selective kernel modules in resnext architecture

Self-aligning ball bearings are widely used in many applications including industrial manufacturing systems. Therefore, it is important to assess the faults that may occur in a self- aligning bearing. Deep learning techniques can be applied to various types of machinery for predictive maintenance. We can train various deep learning models to analyze the patterns of vibrations obtained from the machinery during their operations and use the trained model to predict the fault when any kind of fault occurs in the machinery. This study aims at finding a novel method for bearing fault analysis and classification using several convolutional neural network algorithms to find out the algorithm that gives the highest classification accuracy. The vibration data has been denoised using Daubechies wavelet denoising and then converted to vibration plot in time domain for all the faults. The denoised vibration has been rescaled to a common range using Min-Max normalization. The normalized vibration plots have been segmented in time intervals and split into training and testing image datasets. These datasets have been put in different advanced CNN models to analyze the performance of the model in the classification of the bearing faults. A comparative study has been done amongst five different convolutional neural networks. These networks are ResNet50, ResNeSt50, ResNext50_32x4d, DenseNet121, and SKResNext50_32x4d, the last one being the proposed model. The accuracy of the bearing fault classification using ResNext is 97.51% which has been improved further when selective kernel modules have been integrated into it. The highest accuracy attained by the ResNext architecture integrated with Selective Kernel attention modules, called SkResNext, has come out to be 98.51%.