Fault diagnosis of gearbox based on attention residual network and multi-sensor data fusion

Aiming at the problems that gearbox fault information is susceptible to strong noise interference, single sensor cannot meet the requirements of reliability and accuracy of gear fault diagnosis (GFD) in complex scenarios, and fault feature extraction and fusion of gearbox are difficult in GFD, a novel and intelligent improved attention feature fusion (IAFF) residual network (IAFFRNet) is proposed to mine global and local gearbox fault feature and obtain superior FD results. First, the acoustic and vibration signals extracted by various types of sensors are converted into two-dimensional images by Gramian angular difference fields encoding, which are further feature-spliced and fused. Then, a multiscale large kernel convolution module is constructed to capture and fuse different-scale image features, and the multiscale features are extracted, and the location information is aggregated by the designed attention residual module. Furthermore, the IAFF module assigns different weights to the fused features, thus fusing the multi-sensor features extracted at different stages. Finally, the better effectiveness and generalization ability of the presented IAFFRNet are comprehensively verified by the created NEU dataset and the publicly available SEU dataset. The experimental results indicate that the IAFFRNet method can accurately classify the gear fault and possesses excellent FD ability compared with other methods.