A Regularized Semi-supervised Broad Learning System-based Industrial Soft Sensor With Limited Labeled Data

As an efficient machine learning approach, broad learning system (BLS) is suitable for soft sensing of nonlinear industrial processes. However, for practical soft sensing tasks, process data incorporates a large number of unlabeled data that lead to an over-fitting problem. To handle this challenging problem, this paper proposes a regularized semi-supervised broad learning system-based industrial soft sensor with limited labeled data. To fully leverage the unlabeled data, a manifold regularization term is embedded in the objective of BLS. In this way, the information contained in the unlabeled data can be extracted and used for accurate parameter learning. In addition, an L1 and an L2 regularization terms are also utilized to mitigate the over-fitting issue, thus enhancing the generalization ability of BLS. Accounting for the non-smoothness of the designed objective, this paper derives a new augmented Lagrange method-based model optimization algorithm, based on which the proposed model can be well trained. An industrial case is conducted, and the experimental results demonstrate the feasibility and efficacy of the proposed soft sensor.