A Semi-Supervised Stochastic Configuration Network with Incremental Self-Training

Modern industrial processes are characterized by nonlinearity, complex multivariate relationships, and a scarcity of labeled data, all of which present significant challenges to the data-driven modeling. To address the issue of limited labeled data, this paper introduces a semi-supervised stochastic configuration network model with incremental self-training. The proposed model incrementally selects unlabeled samples and generates pseudo-labels, thereby enhancing its adaptability in data-scarce scenarios. To begin with, weighted clustering is applied to the labeled samples to improve the differentiation of sample importance. Then, a confidence-based decision mechanism is employed to identify high-confidence pseudo-labeled samples, with dynamic adjustments the model based on the quality of these pseudo-labels. Finally, the proposed algorithm is validated using a publicly available industrial dataset and applied to a semi-autogenous ball mill crusher system. Experimental results demonstrate that, under conditions of limited labeled data, the proposed algorithm outperforms existing methods in terms of robustness and prediction accuracy, showcasing its promising potential for applications, particularly in the domain of soft sensor.