Intelligent Diagnosis of Dual-Channel Parallel Rolling Bearings Based on Feature Fusion

In practical engineering, due to the complex and variable working conditions of rolling bearings and the highly nonlinear characteristics of fault signals, especially in the cases of limited fault samples, it is very difficult to achieve satisfactory diagnostic results with the traditional rolling bearing fault diagnosis method. Therefore, in this article, a two-way parallel rolling bearing intelligent diagnosis method based on multiscale center-cascaded adaptive dynamic convolutional residual network (MCADCRN) and Swin transformer (SwinT) is proposed. First, the original signals are transformed into the 2-D time–frequency map by using continuous wavelet transform to preserve the time–frequency characteristics of the original signals. Second, a multiscale center-cascaded dynamic convolutional residual block (MCDCRB) and a multidimensional coordinate attention mechanism (MDCAM) are designed to extract the fault features. Through multiscale convolutional operations, MCDCRB can capture the feature information in different frequency ranges and use a cascade structure to progressively extract higher level features. At the same time, the MDCAM dynamically selects and fuses the features of different scales to reduce the information redundancy and capture the key features; next, the MCADCRN network is constructed by multiple MCDCRBs and an MDCAM to capture the local features; then, the global features of the fault information are captured by using the mechanism of the moving window self-attention in the swin transformer network. Finally, the local features are fused with the global features and the recognition results are the output. The experimental validation is carried out with two different bearing datasets, and the average diagnostic accuracy of the proposed method under variable operating conditions is 99.64%, which is 1.97%, 1.53%, 1.71%, 1.16%, and 2.84% points higher than that of the five advanced methods, respectively. Under limited sample conditions, especially when there are only 50 samples, the diagnostic accuracies of the proposed method are 96.42% and 90.89%, respectively. The results verify the effectiveness of the proposed method.