Abstract: Rolling bearing is a critical component of wind turbine. Because of its complex operating conditions, it is difficult to accurately identify the type of fault. In order to solve the problem of insufficient feature learning ability of traditional deep neural network in strong noise environment, an improved deep residual shrinkage network based on dense block (DB-DRSN) is proposed to realize efficient fault diagnosis of rolling bearing under strong noise and different load conditions. First of all, the vibration signals with different levels of noise are sampled at intervals and matrixed, and a two-dimensional grayscale image is constructed as the input sample. Then, the dense connection residual shrinkage block unit based on Dense block (DB-RSBU) is constructed. The Bottleneck layer is used to replace the convolution hidden layer in the residual shrinkage block unit, and the concat connection is added to make full use of the shallow and deep features. After each dense connection, the dimension of feature maps is reduced by 1×1 convolution, and the attention module and soft threshold function are used to assign different thresholds to the channel-by-channel features and reduce noise. Finally, the input samples go through the network stacked by convolution layer, pooling layer and DB-RSBU layer to get the classification results. The experimental results show that the average diagnostic accuracy of DB-DRSN model under different noise levels on CWRU and PU rolling bearing data sets are 99.80% and 96.44% respectively, which has higher accuracy, faster convergence speed and stronger anti-interference ability than other models. The improvement of network structure by introducing the core idea of dense connection can provide a new method for data-driven fault diagnosis of wind turbine rolling bearings.

Key words: rolling bearing, fault diagnosis, improved deep residual shrinkage network, dense block, attention mechanism