A New Deep Learning Model for Semi-supervised Soft-sensing of an Industrial Production Process

Soft sensing offers a promising solution for predicting key quality variables in various production industries. One of the major challenges in developing an effective data-driven soft sensor is the scarcity of labeled data and the obstacle of extracting useful information from unlabeled samples. To address this issue, this work proposes a new deep learning-based soft sensor model called a spatiotemporal deep learning network. It leverages an encoder-decoder structure to explicitly exploit the spatial and temporal information in both labeled and unlabeled data, enabling efficient utilization of the latter to facilitate prediction performance. The encoder realizes more detailed spatiotemporal dependencies extraction by the proposed gated recurrent unit-based attention mechanism and the channel-calibration attention-based one. Finally, the extracted spatiotemporal features are fed to a multi-layer perceptron-based prediction head for soft sensor modeling. A mixed form loss is employed in the decoder to train our proposed model which incorporates both labeled and unlabeled data. Experiments are conducted on a real-life industrial process, demonstrating the feasibility and effectiveness of the proposed method.