A Multi-Scale Deep Learning Attention-based Feature Method for Rolling Elements Bearing Fault Detection in Industrial Motor Drives

In the last decade, convolutional neural networks have achieved great success in the automated fault diagnosis of rotating equipment in electrical machines. However, the application of convolutional models encounters some challenges to deal with such as (i) the requirement of a vast amount of training data and (ii) the selection of the neural architecture, and particularly the sizes of the convolutional kernels that effectively extract features from the raw input signal. To alleviate the above challenges, we propose a deep learning network consisting of multiple independent densely connected convolutional streams with different sizes of kernels and of a simple attention mechanism that fuses the extracted features, producing a feature mapping with generalization and discrimination power. Simulation cases with a widely used bearing fault detection benchmark show the effectiveness of the proposed approach, especially in cases of a restricted amount of training samples.