Abstract: In the environment of low speed, heavy load and strong noise of heavy equipment, it is difficult for a single sensor to fully obtain bearing fault diagnosis information, resulting in low fault identification rate and unstable identification, resulting in bearing fault migration diagnosis failure under changing working conditions. Aiming at the above problems, a bearing fault migration diagnosis method based on multi-sensor information fusion was proposed. Firstly, multiple convolution neutral networks (MCNNs) was constructed combined with the number of sensor channels to extract the fault features of each channel.Then, the least absolute shrinkage and selection operator (Lasso) was introduced into MCNNs, and the feature weight was updated through network back propagation. Thus multi-channel feature fusion was realized. Finally, the source domain data was used to train the model and extract fusion features, and the model parameters were optimized through loss function. The model results obtained from the source domain training were taken as the initial model of the target domain, and the parameters of the initial model were fine-tuned through the target domain samples, so as to realize model migration. Finally, the performance experiments of information fusion effect, method comparison and sensor information acquisition attribute were done. The experimental results show that the sensor installation position has a great influence on information fusion, multiple convolution neutral networks combined with least absolute shrinkage and selection operator (MCNNs+Lasso) has a good feature fusion effect. The average diagnostic accuracy of migration is 99.03%, and the partial accuracy can reach 99.97%. The proposed method shows high migration accuracy and good generalization performance in multiple migration tasks with varying working conditions.
Key words:  rolling bearing; fault diagnosis; multisensor information fusion; multiple convolution neutral networks(MCNNs); least absolute shrinkage and selection operator (Lasso); transfer learning