Abstract: To overcome the defects of autoencoder-based process monitoring in fault isolation and identification, we propose a denoising autoencoder (DAE)-based fault isolation and identification method by introducing unknown fault subspace into the optimized objective function of DAE. Considering the characteristics of the faults, we present an l₁ regularization term to achieve sparse isolation and design an optimization problem-solving method based on adaptive moment estimation. Compared with traditional methods, the DAE-based fault isolation and identification method are more effective in fault diagnosis of nonlinear processes. The application in the Tennessee Eastman (TE) process and blast furnace ironmaking process validate the feasibility of the proposed method.