Aiming at the low accuracy of traditional intelligent fault identification methods of rotating machinery with artificial features as input and the strong dependence of deep learning methods on data volume， Hu invariant moment has the characteristics of telescopic translation and rotation invariance and the advantages of unsupervised learning deep learning model in feature extraction of small sample data. A deep convolutional auto-encoder fault diagnosis model （DCAE-FDM） is proposed， which integrates Hu invariant moment and DCAE features. Firstly， the effective singular value method is used to purify the original vibration signals， and the purified axis orbits are obtained， which are divided into training set and test set in a certain proportion， and the Hu invariant moments of them are calculated respectively. Secondly， the constructed DCAE-FDM model is used to extract the deep auto encoding features. Thirdly， the two features are fused together， and the fusion features of training set are taken to train back propagation （BP） neural network. Finally， the fusion feature of test set is used to test the trained model. Results show that the recognition effect of the proposed method is significantly better than that of the deep learning method and the traditional recognition method. The average accuracy of the former is 98.5%， about 6 percentage points higher than that of the suboptimal model.