Abstract:
Due to the constraints of data format, using deep learning models directly after fast Fourier transform on fault data often fails to obtain sufficient features, making it difficult to further improve the accuracy and generalization of fault diagnosis under variable operating conditions. To this end, a lightweight fault diagnosis model is proposed that integrates two-dimensional frequency domain and improved residual networks. Before acquiring and learning more features, the first half of the symmetric spectrum converted by the FFT method is intercepted and unfolded in a row first manner on a two-dimensional surface to obtain frequency domain data feature information represented by a matrix; Then ReLU6 is used as the Activation function of the improved residual networks to learn the matrix features; Finally, the final state feature information is mapped into a one hot encoding vector. The test on the data set of Case Western Reserve University shows that the diagnostic accuracy of this method is improved by 7.07% compared with the one-dimensional diagnostic model; Compared with the standard residual network, the fluctuation in accuracy has been reduced by 0.14%, improving the accuracy and stability of fault prediction.