Deep Rational Attention Network With Threshold Strategy Embedded for Mechanical Fault Diagnosis

Abstract

Benefiting from the progress of artificial intelligence, the deep learning-based data-driven fault diagnosis procedure has gradually become a preferred and reliable option. Although this kind of end-to-end diagnosis mode can achieve satisfactory results in most conventional conditions, the existing methods may not be able to guarantee the diagnostic accuracy when noise is a prominent issue. To address the above problem, the deep rational attention network (DRANet) is developed in this article. In the DRANet, a pseudo soft threshold function is developed to further modulate the attention weights in a convolutional neural network. Besides, the threshold is automatically determined during network optimization. The attention weight modulate strategy driven by the pseudo soft threshold function accords with the idea of threshold denoising in signal processing and thus endows the deep architecture with the ability to further neglect the features with little contribution. Experimental vibration signals coupled with different types of noise are employed to evaluate the efficacy of the proposed DRANet. Results show that in comparison with other methods, the method of this work holds a more outstanding diagnostic performance, especially in the case of low SNR.