A multi‐feature space constrained stacked autoencoder and its application for uncertain process monitoring

Abstract Industrial equipment measurement data are often subject to errors and uncertainties due to factors such as environmental conditions and equipment aging, posing significant risks to operational safety. To mitigate these issues, we propose a novel process monitoring method based on a multi‐feature space constrained stacked autoencoder (MFSCSAE), designed to reduce the impact of uncertainties. In real‐world industrial processes, uncertain data typically fluctuate within an interval centred around the true value. The MFSCSAE model incorporates multiple feature space constraints, using the upper and lower bounds of this interval as inputs, with the true measurement data serving as the reconstruction target. A new loss function is derived by combining the deviation between the model's output and the true target with the deviation between the features of the hidden layers. The model is trained on normal operational data, and control limits are determined using support vector data description (SVDD). These control limits are then used to assess whether the industrial process is functioning within acceptable bounds. The proposed method is applied to both the Tennessee‐Eastman (TE) process and a real industrial fluid catalytic cracking (FCC) process, demonstrating the effectiveness of the MFSCSAE model in monitoring uncertain processes.