Multiscale pyramidal convolutional attention network for intelligent fault diagnosis of gearboxes

In practical scenarios, the advanced intelligent fault diagnosis (IFD) technology of gearboxes is urgently needed. Recently, convolution neural networks (CNNs) have shown good application effects in fault diagnosis. However, they still face two major technical challenges: (1) Noise-polluted vibration signals increase the difficulty of precise IFD; (2) CNNs can only capture local data features and cannot extract global features. To deal with these challenges, this study proposes a multiscale pyramidal convolutional attention network (MPCANet) for the IFD of gearboxes. MPCANet can extract pyramid-type features, endowing it with strong noise robustness. In addition, multiple attention mechanism methods are employed in MPCANet. These methods enable the network to capture local and globle features of input data simultaneously, further improving feature richness. Two fault diagnosis experimental cases on wind turbine gearboxes are conducted, and the indicator values of the average diagnostic accuracy, recall, and F1 score for MPCANet are 99.71 %, 99.70 %, and 99.70 % in case I and 99.19 %, 99.20 %, and 99.19 % in case II; these values are better than those of other multiscale models. Furthermore, a noise robustness experiment shows that the proposed approach can work effectively in various noisy environments.