A binary neural network fault diagnosis method based on neural architecture search

Abstract The intelligent fault diagnosis method based on deep neural networks has too many model parameters, high computational complexity, and the parameters of the new architecture are manually designed by a series of experiments. Based on these deficiencies, a neural architecture search-based binary neural network (NAS-BNN) fault diagnosis method is proposed. In the proposed method, the NAS is used to enhance BNN’s automatic search capability. L1 and L2 path regularization are introduced to improve the stability of BNN search, and variance-based edge regularization is proposed to enhance the robustness of edge selection. Adaptive adjustment of the channel sampling ratio can quickly obtain a balance between different operations and search efficiency. The proposed NAS-BNN fault diagnosis method has been successfully applied to the experiment of aero-engine rolling bearings, and the effectiveness of the proposed method is verified from three aspects, i.e. complexity analysis, noise resistance performance analysis, and class imbalance analysis. The experimental results show that the proposed NAS-BNN method is superior to traditional BNN methods, the computational cost is greatly reduced, and the optimal models can be obtained with less time and low storage. The proposed method is superior to other NAS approaches and manually designed approaches in the accuracy and efficiency.