Dual‐Constraint Network‐Based Soft Sensor for Industrial Processes

Abstract

ABSTRACTTo address information loss during network stacking and capture the temporal characteristics of industrial process data, this paper proposes a soft sensor method named Self‐Supervised Stacked Isomorphic Hierarchical Mapping Autoencoder Based on Dual‐Constraint Mechanism (SSIHMAE). The model's decoder directly reconstructs the original input. Additionally, it projects each encoder layer's output back to the corresponding input dimension through a mapping layer. By incorporating the cosine similarity between the mapped values and the input into the loss function, the model effectively mitigates information loss in deep networks. At the top layer of the network, self‐supervised temporal contrast learning is introduced. Data pairs adjacent in time serve as positive samples, while interval‐separated data pairs act as negative samples. The model minimizes the distance between the features of anchor samples and temporally adjacent samples by optimizing the InfoNCE objective function, while rejecting samples separated by intervals. In this way, multi‐scale temporal features are captured. Finally, a fully connected layer constructs the prediction model. Simulation results on sulfur recovery and coal‐fired power generation datasets demonstrate that the proposed method achieves 12.4% and 5.1% higher prediction accuracy than the traditional SAE model. These results validate the method's efficacy.