Process mechanisms fusion enhanced spatially scalable convolution network for multi-indicator prediction in process industries

Abstract

In process industries, the detection of interrelated production indicators, including throughput, quality, etc., is often delay due to the inherent continuity, causing production disruptions and quality issues. Accurate prediction of multi-indicator is crucial, but intricate nonlinear correlations among parameters and indicators pose significant challenges. Data-driven prediction can overcome the challenge of accurately constructing process mechanism models covering the entire production process and all elements, but struggle to infer cross-space migration patterns of parameters’ impacts on indicators. To address this issue, this study proposes a process mechanism fusion enhanced intelligent multi-indicator prediction method for process industries, using the polyester fiber polymerization process as an illustrative case. Firstly, a process mechanism model is established to generate mechanism data encapsulating process mechanisms like coupled relationships and spatial correlations, and these mechanisms are extracted as mechanism features, which are fused with data features to enhance the model’s performance. Secondly, a spatially scalable convolutional neural network is raised, which extracts the implicit deep data features and mechanism features between parameters and indicators from both within-process and cross-process dimensions, utilizing both real and mechanism data. Furthermore, a multi-head self-attention mechanism is employed to adaptively adjust the self-attention weights of the fused features, guiding the model to learn the complex relationships between fused features and enhancing the ability to learn complex coupled relationships and spatial correlations. Finally, the proposed prediction method is validated using polymerization process data and demonstrated superior performance in achieving accurate multi-indicator prediction compared to both efficient machine learning and advanced deep learning methods.