Quality prediction of industrial process based on Kolmogorov–Arnold graph convolution aggregation temporal convolution network

Abstract Industrial process data are essentially complex time series data with high nonlinearity and dynamic changes, which makes it difficult to extract features for quality prediction leading to poor prediction accuracy. To address the limitations of current data-driven methods for industrial process data in feature extraction and interpretability, an interpretable quality prediction model based on Kolmogorov–Arnold graph convolutional aggregation temporal convolutional network (KAGCN-KATCN) is proposed. Firstly, the KAGCN module is designed by improving the aggregation stage of graph convolutional network, which leverages the nonlinear decomposition of the Kolmogorov–Arnold network (KAN) to enhance the extraction of nonlinear features and adequately learn the key information. Secondly, the KATCN module is designed to enhance the extraction of dynamic dependencies, which improves the residual connectivity of TCN by filter response normalization to boost the training stability and generalization ability, and KAN is introduced to map the convolution results nonlinearly while enhancing the interpretability of the model. Then, the multi-head self-attention mechanism is integrated to aggregate the key features output from the KATCN module, which enhances the parallel computing capability and improves the accuracy of the final prediction results. Finally, the experimental validation on two industrial cases (a debutanizer column and a sulfur recovery unit) demonstrates that the proposed model outperforms several existing quality prediction models.