Research on Dual Extreme Learning Machine Based on Residual Structure and its application to process modeling

In order to deal with the high dimensional and high complex process data, a residual structure based dual extreme learning machine model (R-DELM) is proposed. This method proposed in this paper uses dual ELM model: one ELM model is used to establish the process model, which can be viewed as a primary regression model; another one is used to compensate the first ELM model by establish residual-ELM model, which can be viewed as a secondary regression model. In the proposed method, the Kendall correlation coefficient (KCC) is firstly utilized to select the relevant process variables for soft sensor modeling, then the proposed R-DELM model is utilized to compute the key quality output. A real-world production process called Purified Terephthalic Acid (PTA) is applied to investigate the model performance of the R-DELM based soft sensor model. Compared with some other machine learning models, simulation results indicate that R-DELM has the characteristic of high modeling accuracy.