An improved model combining sparse filtering and deep dilated gate convolutional network was proposed in order to solve the problem that feature extraction relies on manual experience when rolling bearing faults occur, and that the fault category was difficult to automatically and accurately identify. Sliding window was used to sample bearing vibration signals with time series characteristics in order to avoid over fitting. Heteroscedasticity was eliminated and data features were extracted by improving the sparse filtering of the objective function in order to shorten the calculation time and improve the accuracy of classification. The fault classification model was established by combining the dilated gate convolution and the bidirectional LSTM network, and the data noise can be filtered out. Data experiments from Case Western Reserve University and laboratory power equipment were compared. Results show that the fault diagnosis accuracy rate of this model can reach 98%. Different load and different SNR experiments with bearing vibration signals show that the model has generalization and anti-noise performance.