For batch process data with multivariate, nonlinear, and high dimensional characteristics, the early fault signals are easy to be disturbed by noise and the fault amplitude is low, which lead to poor fault monitoring performance, we propose an early fault detection method for batch processes based on stack-robust neighbourhood preserving autoencoder(S-RNPAE) in this paper. Firstly, the objective function of the autoencoder is redesigned by the $L_{2,1