Abstract: In industrial processes, machine learning methods based on neural networks are commonly used for fault diagnosis. However, such methods suffer from poor transferability and usually require the design of distinct neural networks for different tasks. The optimization of network model structures and parameters mainly relies on repeated trial and error. This process not only depends on expert knowledge and experience but also consumes substantial time. As model depth increases, the difficulty of model design and the workload of manual parameter tuning multiply dramatically. Unlike manually designed neural networks, Neural Architecture Search (NAS) can automatically design and generate network structures, which reduces the reliance on expert experience and knowledge to a certain extent, simplifies manual optimization procedures, and cuts down the human and time costs incurred by structural design and parameter adjustment. Nevertheless, networks generated by conventional NAS methods usually have a single branch with a fixed structure. To address this issue, this paper proposes a fault diagnosis method based on multi-path neural architecture search. Firstly, a controller is designed to automatically generate neural network structures with graph convolutional networks as the basic unit. Secondly, a weight-sharing mechanism is adopted to substantially reduce network evaluation time. Furthermore, the ε-greedy strategy is introduced to prevent the controller from falling into local optima. Finally, the effectiveness of the proposed method is verified on the Tennessee Eastman (TE) process simulation platform. Compared with other classical methods, the proposed method achieves an average accuracy improvement of over 3.8% and a search speed improvement of more than 71.2%.