Mc-rsgn: A Novel Soft Sensor Model Toward High-noise Data In Dynamic Industrial Polymerization Processes

Soft sensor technology is crucial for enhancing the real-time performance and operational efficiency of industrial systems, serving as a key driver for modern industrial automation and intelligence. However, due to the strong coupling and nonlinear characteristics of industrial systems, complex nonlinear dynamic responses and data instability arise under high-noise environments, posing severe challenges to the dynamic feature extraction of data. To address these issues, a novel soft sensor model based on the multichannel residual shrinkage gating network (MC-RSGN) is proposed in this study. First, a dynamic feature extraction module based on multichannel convolution is introduced. By reconstructing data to generate multichannel samples that encapsulate local correlations, this module overcomes the limitation of a single-channel convolution kernel that is unable to cover remote variables. Second, a deep noise reduction module is designed. The soft-threshold function is integrated into residual connections to efficiently filter out noise signals and reduce the network complexity. Furthermore, a novel soft sensor model combined with a gated recurrent unit (GRU) is developed, which not only effectively captures the dynamic characteristics of complex industrial data but also filters high-noise signals. Finally, the superiority of the proposed MC-RSGN model is verified through numerical simulations on a continuous stirred-tank reactor (CSTR) system and an industrial case study on a polymerization process.