Semi-supervised Extreme Learning Machine For Soft Sensor Modeling Of Chemical Processes

Soft sensor modeling of chemical processes has played an important role in ensuring the reliable and continuous operations of process equipments and further enhancing the quality of products. However, semi-supervised learning has rarely been applied to soft sensor modeling of chemical processes. As a result, a large amount of information hidden in unlabeled samples cannot be utilized in soft sensor models. In this thesis, two soft sensor modeling methods using the semi-supervised learning are proposed. The main contributions are as follows:(1) Traditional soft sensor models are not able to make good use of unlabeled samples. The semi-supervised learning machine(SELM) is introduced into the soft sensor modeling of chemical processes. SELM can effectively enhance the prediction accuracy and reliability by using unlabeled samples. The SELM soft sensor modeling method is applied to online prediction of the gasoline yield in a fluid catalytic cracking unit process. Compared with the backing propagation neural network and extreme learning machine, better prediction results have been obtained.(2) The traditional leave-one-out-cross-validation for selection of the penalty parameter l in SELM is computationally expensive. To overcome this shortcoming, a fast-leave-one-out-cross-validation SELM(FSELM) is proposed. The FSELM-based soft sensor modeling method is applied to online prediction of the 19 composition variables in the Tennessee Eastman process. Compared with the principal component regression and extreme learning machine, more accurate and reliable prediction results have been obtained.(3) The number of hidden layer nodes is selected by experience in FSELM, which would be over-fitting if the number of nodes is not suitable. Furthermore, the FSELM model has to be retrained in an inefficient manner if the number of nodes is changed. To select a suitable number of nodes automatically and enhance the efficiency of modeling, a recursive FSELM(RFSELM) is proposed. The RFSELM soft sensor modeling method is applied to online prediction of the flooding velocity in packed columns and the gasoline yield in a fluid catalytic cracking unit process. The obtained results indicate that the RFSELM soft sensor is reliable and accurate.