Soft sensor of multiple operating condition processes based on input–output correlated difference focusing networks

Abstract Aiming at the problems of information accumulation loss, quality correlation, and multiple operating condition feature extraction in soft sensor based on stacked networks, this paper proposes a method based on input–output correlation difference focusing networks (IO‐DFN). Constructing the input–output stacked isomorphic autoencoder (IOSIAE) network, it is proposed to reconstruct the original input variables and quality variables layer by layer in stacked autoencoder (SAE) to overcome the cumulative loss of the original input information and consider the quality correlation. A multi‐module architecture is constructed, where different modules all use input–output isomorphic autoencoders, and different loss functions are used during training to extract multiple operating condition features. It is proposed to use the inter‐module similarity self‐attention mechanism to highlight the differences of different module features, obtain the module focusing features, and establish multiple prediction models between them and the outputs. The correlation between each module focusing feature and the original input is calculated separately to further highlight the weights of the different module features, and the weights are used to fuse the different predicted values into the final predicted values. The results are validated by simulation of an industrial sulphur recovery process with a thermal power generation process, and the findings show the efficacy of the proposed approach.