Intelligent fault diagnosis of rotating machinery using a new ensemble deep auto-encoder method

Abstract

In traditional intelligent fault diagnosis methods of rotating machinery, features are designed manually by experts, which makes these methods less automatic. Deep auto-encoder (DA) provides an effective way to learn discriminative features. However, individual DA is of low generalization and not robust. This work develops an ensemble DA(EDA) method for intelligent fault diagnosis. EDA is constructed by combining sparse DA, denoising DA and contractive DA. Thus, EDA can effectively handle redundant information, noisy corruption and signal perturbation. To ensure the feature learning performance of each DA model, a self-adaptive fine-tuning is designed to obtain stable convergence. To enhance the discriminative features, a dynamic weighted average method is designed to aggregate these learned features. EDA is verified on three public datasets, and achieves the testing accuracies of 100%, 99.69% and 99.92%. Comparisons with other methods, including both traditional methods and individual DA methods, demonstrate that EDA obtains higher diagnosis accuracy.