A hybrid deep learning network for diagnosing multipoint faults in rolling bearings under variable operating conditions

Rolling bearings are essential in rotating machinery, but their operation under diverse conditions like load, speed, and temperature can lead to complex faults. While efforts have addressed multiple component faults, identifying and classifying multipoint faults in rolling bearings remains challenging. In this paper, a hybrid deep learning network (HDLNet) for multipoint fault diagnosis in rolling bearings is proposed under variable operating conditions. Firstly, in the time domain, a two-layer bidirectional long short-term memory (Bi-LSTM) network is utilized to extract the temporal features of vibration signals. Secondly, in the spatial domain, a multi-scale dynamic snake convolution with fast spatial pyramid pooling attention (MDSC-FSPPA) network is proposed to capture the spatial characteristics of vibration signals. Following this, fault features extracted from both time and spatial dimensions are combined, and improved residual network (ResNet) is leveraged to further enhance the feature representation and suppress noise. Furthermore, visualization techniques are employed to intuitively show the relationships and clustering structures among them. Experimental dataset of rolling bearing with multipoint faults are used to evaluate the proposed HDLNet. The findings indicate that the proposed HDLNet has good noise resistance generalization performance and superior diagnostic performance under different rotational speeds compared to other established intelligent fault diagnosis approaches.