Soft sensor model for nonlinear dynamic industrial process based on GraphSAGE-IMATCN

Industrial process data are closely related to production conditions and are essentially complex time series with high nonlinearity and dynamics. To solve the challenge of insufficient feature extraction of industrial process data, resulting in poor real-time monitoring of key quality variables, we propose an interpretable industrial soft sensor based on Graph Sampling and Aggregation Temporal Convolutional Network Improved by Multi-head Self-Attention (GraphSAGE-IMATCN) for predicting the trend of key quality variables in real time. Firstly, a three-dimensional data development strategy for batch processing is designed, and the maximum information coefficient (MIC) is introduced, and the threshold function is established by combining kernel density estimation to extract the characteristic variables with high quality correlation, and the explanatory and reliability of the model are enhanced by statistical methods. Secondly, a deep graph sampling aggregation (GraphSAGE) structure is designed for industrial big data, which aggregated features based on adjacent nodes and captured the context information of key nodes and serialized the extracted features to improve the computing speed of the model by combining the parallel computing advantages of the time convolutional network. Then, to overcome the data of different batch sizes and production scales, the residual structure of the Temporal Convolutional Network (TCN) is optimized by using Filter Response Normalization (FRN) to enhance the generalization and robustness of the model. Then, the multi-head self-attention mechanism (MHSA) is introduced to enhance the parallelism of the model, and the inference speed of the model is optimized to meet the key requirements of real-time performance for industrial process monitoring. Finally, the effectiveness of the proposed model is verified through experiments on the penicillin fermentation process and the debutanizer column.