Multi-size wide kernel convolutional neural network for bearing fault diagnosis

The bearing is an indispensable part of mechanical systems. Fault diagnosis of bearing faults is vital for uninterrupted operations of the system, and to prevent catastrophic failure. Artificial intelligence implementation has revolutionized the bearing fault diagnosis method. Application of deep learning has eliminated manual feature extraction and selection requirements. While conventional convolutional neural networks have demonstrated potential in diagnosing faults, considering a more extensive variety of spatial variables can further optimize their performance. This paper proposes a multi-wide-kernel convolutional neural network-based model for bearing fault diagnosis. We propose wide kernels in the neural network's convolutional layers, which enable the model to learn broader patterns from the input for bearing fault diagnosis. The wide-kernel design enables the network to obtain local and global features more effectively, improving the network's capacity to distinguish between healthy and faulty bearings. We train and validate the proposed multi-wide-kernel convolutional neural networks using an extensive dataset of vibration signals collected from bearings under diverse scenarios. Because of its increased sensitivity to subtle fault patterns, the proposed model offers better accuracy. The model's efficacy is further confirmed by comparing it with existing cutting-edge techniques for diagnosing bearing faults.