An unknown-condition-aware soft sensor for multi-condition industrial processes

Soft sensors have been widely used to estimate hard-to-measure quality variables in complex industrial processes. However, conventional soft sensors rely on a single global model, whose performance degrades markedly when the process switches among multiple operating conditions or enters operating regimes unseen during training. To address this issue, this paper proposes an unknown-condition-aware soft sensor framework for multi-condition industrial processes. Nonlinear multi-view clustering (NMVC) is combined with an evidential K-nearest neighbor (EKNN) classifier: NMVC partitions the operating conditions by fusing heterogeneous similarity views, while EKNN quantifies classification uncertainty through evidential reasoning so that unknown operating conditions are effectively recognized. Based on the operating condition recognition results, condition-specific local regression models are constructed by combining a convolutional neural network (CNN) for extracting coupling relationships among variables with Informer for capturing long-term temporal dependencies. The proposed method is validated on two industrial datasets: the Tennessee Eastman (TE) benchmark dataset and a real-world power-plant dataset. Comprehensive comparisons with classical baseline models and recent soft sensing methods demonstrate that the proposed method achieves superior performance in unknown operating condition recognition and multi-condition prediction. In particular, it improves unknown-condition recognition accuracy by 27.0% relative to K-means-KNN and increases the R 2 for prediction by 7.9% relative to Informer.