Resumen
In the actual operation of the machine, due to a large number of operating conditions and a wide range of operating conditions, the data under many operating conditions cannot be obtained. However, the different data distributions between different operating conditions will reduce the performance of fault diagnosis. Currently, most studies remain on the level of generalization caused by a change of working conditions under a single condition. In the scenario where various conditions such as speed, load and temperature lead to changes in working conditions, there are problems such as the explosion of working conditions and complex data distribution. Compared with previous research work, this is more difficult to generalize. To cope with this problem, this paper improves generalization method SAND-Mask (Smoothed-AND (SAND)-masking) by using the total gradient variance of samples in a batch instead of the gradient variance of each sample to calculate parameter s. The SAND-Mask method is extended to the fault diagnosis domain, and the DCNG model (Deep Convolutional Network Generalization) is proposed. Finally, multi-angle experiments were conducted on three publicly available bearing datasets, and diagnostic performances of more than 90%, 99%, and 70% were achieved on all transfer tasks. The results show that the DCNG model has better stability as well as diagnostic performance compared to other generalization methods.