Industrial prediction method based on graph sampling and aggregation of temporal features

Abstract The soft sensor technique is extensively used to manage nonlinear and dynamic industrial process variables for predicting critical quality variables. Traditional soft sensor models face limitations in feature extraction and fail to capture the sequential dependencies among process variables. In order to resolve these difficulties, a graph sampling and aggregating temporal convolutional network (GTTL) is proposed that integrates a transformer encoder and a long short‐term memory (LSTM) network incorporating dual attention mechanisms. First, relevant feature variables with strong correlation are selected using the maximum information coefficient (MIC) and kernel density estimation (KDE) methods. Next, GraphSAGE is employed to sample and aggregate features from the local neighbourhood of nodes, serializing the extracted features. Subsequently, the filtered response normalization (FRN) technique is applied to optimize the residual structure in the temporal convolutional network (TCN), enhancing the model's generalization and robustness when processing heterogeneous data. The transformer encoder and the LSTM network with dual‐attention mechanisms effectively capture long‐range dependencies within the sequence, identifying process variables and time points associated with quality variable, thus enhancing the accuracy of quality variable predictions. Finally, the proposed soft sensor model's effectiveness and advantages are demonstrated using real industrial process datasets. The source code of this study is available at: https://github.com/Forgun/model.git .