Aiming at using autoencoder to extract process features for fault detection, the local structure information of data is not considered, a method of neighborhood denoising quadrature autoencoder is proposed. The neighborhood preservation embedding algorithm extracts the neighborhood information of the data as a weight to weight the process data and strengthen the local structure information of the data. The orthogonal autoencoder further extracts the nonlinear features of the process data with local information weighting. The robustness of the autoencoder is enhanced by adding noise, and the network parameters are trained by the back-propagation algorithm to obtain a robust autoencoder model that can capture the local and global characteristics of the data. T2 and SPE statistics are constructed in the latent feature and reconstructed residual space of the model, respectively, and the statistical control limits are calculated for fault detection. Simulation experiments are carried out on the Tennessee-Eastman process and the three-phase flow process, and the results show the effectiveness of the proposed algorithm.