Abstract
In modern process industries, it is of significance to build data-driven soft sensors to predict key performance indicators that are difficult to measure directly. However, the industrial data obtained are usually characterized by uncertain time series, different degrees of outliers, multiple redundant variables, and abundant unlabeled data, presenting difficulties in data-driven modeling. To address these difficulties, a semi-supervised and robust data-driven modeling algorithm is proposed. First, the t-distributed stochastic neighbor embedding (t-SNE) is applied to reduce the dimensionality of unlabeled samples. Second, a bidirectional long short-term memory (BiLSTM) network based a capped Huber loss is developed to deal with outliers, and the least absolute shrinkage and selection operator (LASSO) is introduced to remove redundant variables. Third, experiments on an artificial dataset and an industrial dataset demonstrated that the developed algorithm had a higher prediction accuracy than did other state-of-the-art methods. Furthermore, ablation studies were conducted to evaluate the contributions of different techniques on the model performance.
Funding Information
  • Shandong Provincial Natural Science Foundation of China (ZR2021MF022)

This publication has 31 references indexed in Scilit: