In the actual multimode chemical processes, new modes are usually generated due to the adjustments of product requirements. The existing fault diagnosis methods based on deep learning have limitations, such as the insufficient use of the design experience from the existing modes and the difficulty of training models with small samples. Aiming at the above problems, a fault diagnosis method named MetaNAS based on meta learning (ML) and neural architecture search (NAS) is proposed for new modes. The proposed method first uses NAS to automatically obtain the network models with the best performances for the existing modes. Then, ML is used to learn the design experience of the fault diagnosis models from the NAS processes of the existing modes. Finally, when new modes are generated, gradient update is carried out on the basis of the learned design experience, that is, the fault diagnosis models of new modes can be quickly obtained under the condition of small samples. The effectiveness and feasibility of the proposed method are fully verified by the simulation experiments of a numerical system and the Tennessee Eastman (TE) chemical process.