Rolling Bearing Fault Diagnosis based on ICEEMDAN-WTATD-DaSqueezeNet

Rolling bearing is a mechanical part of industrial scenarios which is extensively used. The working environment of the bearing is complex including high load, high pressure, and high temperature, so the failure probability is high. So the vibration signals of bearings are often accompanied by a large amount of noise, leading to unstable signal data characteristics. Aiming at the characteristics of data instability and a large number of complex noises, an adaptive threshold denoising (WTAT) method based on improved complete ensemble empirical mode decomposition with adaptive noise (ICEEMDAN) combined with wavelet transform is proposed. Then combined with an improved SqueezeNet model, an accurate diagnosis of rolling bearing fault was achieved. Firstly, the ICEEMDAN is used for mode decomposition to obtain a sequence of intrinsic mode functions (IMFs), which effectively solves the mode aliasing and pseudo-mode problems of traditional empirical mode decomposition. Moreover, the WTAT noise reduction is carried out for several IMFs with the least correlation, which effectively solves the problem of over-strangling and over-retention of traditional wavelet denoising. Then, in order to better integrate local features and global dependencies, the dual attention convolutional module is embedded in the fire module of SqueezeNet and a lightweight model named dasqueeezenet is built for bearing fault diagnosis. Finally, compared with common convolutional neural networks such as VGG16, ResNet and Xception, the presented method achieves remarkable results in terms of the accuracy with 96.17%, the precision with 96.3%, the recall rate with 96.42% and the F1-Score with 96.26%, and the presented model's effectiveness in diagnosing rolling bearing faults has been demonstrated.