Co-training Based Semi-supervised Soft Sensor Modeling

Process industries seeks to keep the production process stable and trace the condition of primary product components. As a result, real-time monitoring is very essential in the production process. Great progress has been made in both science and technology, which makes it possible to detect many undetectable process variables. However, due to the existence of many physical reactions, biochemical reactions and energy exchange, the difficulties for variable detection are still enormous. Soft sensor is proposed to deal with this problem. By selecting some closely linked variables of the dominant variable, soft sensor builds a mathematical model to predict the dominant variable by input these auxiliary variables. Online monitoring of primary variables can be fulfilled in this way.However, the integrity of input and output data is essential in soft sensor modeling. Generally, primary variables are acquired through off-line techniques. Although the laboratory analytical instrument may be placed for the measurement of those impossible-to-be-detected variables, data is quite limited in consideration of the exorbitant expenses and long analytic time. Conventional soft sensor models are built on the basis of data with primary variables information. The short of training data leads to the inaccuracy of soft sensor models.In contrast, there exists large amount of easy-to-measured data which can’t be used in building the models in the industrial process. This abomination causes great waste. There’s no doubt that soft sensor models can be definitely improved assuming that valuable information can be extracted from those wasted easy-to-measured data.Co-training style soft sensors are built to solve the data-imbalance and non-linear problems in process industry set forth hereinabove. The main contributions of this paper are listed below:(1) In this paper, a new Co-training PLS based soft sensor is proposed for small samples size problems with many input variables. By splitting the input variables into two parts, two diverse models are built. By helping each other in the iteration rounds, the model update the labeled datasets with unlabeled data. At the end of the algorithm, two models are built based on the full-dimension data. Prediction result is accumulated by averaging the results of these two regressors. The superiority of Co-training PLS over Co-training kNN and conventional PLS is validated in the simulation.(2) Another Co-training style soft sensor called Co-training LWPLS is also proposed in this paper. In order to resolve the non-linear data-imbalance problems, Co-training LWPLS builds two models on the same training data by using different similarity calculation. The training procedure is similar to Co-training PLS, and the performance is validated on simulations, too.At the end of this paper, a summary is presented about the research results nowadays. Some personal ideas and expectations of future works are listed for the readers.