Semi-Supervised Incremental Soft Sensor Model With Spatiotemporal Graph Regularization for Process Industry

In some process industries, process variables are readily measurable, whereas real-time detection of industrial indices poses challenges. This discrepancy results in the lack of a one-to-one correspondence between process variables and indices, leading to a large number of incomplete samples in process industry datasets. This article introduces an incremental random recurrent neural network (IRRNN) that combines incremental architecture construction with random weights. Based on the IRRNN, a semi-supervised IRRNN (SS-IRRNN) approach for soft sensor modeling is proposed. SS-IRRNN leverages both labeled and unlabeled samples as inputs and utilizes resampling techniques to obtain labeled hidden states from the hidden layer of the IRRNN. These labeled states are then used to optimize the output weights of the IRRNN. Furthermore, SS-IRRNN with spatiotemporal graph (SGSS-IRRNN) regularization is designed to enhance the capture of spatiotemporal information in incomplete datasets. Numerical simulation experiments and soft sensing experiments on hematite grinding processes demonstrate that the proposed methods exhibit sound performance when dealing with incomplete datasets.