Multi-model Approach Based Soft Sensor Development Using Gaussian Process Regression

Many key variables in industrial processes are difficult to measure online because of technical or economic limitations, which brings difficulty in control and optimization of industrial production. To solve this problem, the soft sensor technique was proposed. Currently, the most common method of modeling a complex nonlinear system is constructing a single soft sensor model using partial least square, neural networks, or support vector machine. However, it can’t give the prediction accuracy with predictions. Besides, industrial processes always have features of strong nonlinearity, multiple modes and multiple phases, which results in complexity and low prediction accuracy of models.To this end, Gaussian process regression(GPR) and the multiple model approach were used to construct soft sensor models in this thesis. To verify the effectiveness of these proposed methods, process data from a real Erythromycin fermentation process were adopted. The main work in this thesis is as follows(1) Many traditional soft sensors cannot provide prediction accuracy, which results in difficulty in real applications. Therefore, a soft sensor based on the GPR model was proposed. In the method, a variable selection method based the principal component analysis was used to reduce the dimension of inputs of soft sensors. Results on data set from a real Erythromycin fermentation process show the effectiveness of the proposed method. Moreover, it was observed that the prediction accuracy of the GPR soft sensor is sensitive to the distributions of inputs.(2) It was pointed out that a single soft sensor model cannot perform well in multimode and multiphase processes. To this end, a multi-model GPR soft sensor based on a novel fusion mechanism on predicted variances(PV-MGPR) is proposed. This method not only has advantage of traditional multi-model methods, but also owns distinguished merits like convenience and having obvious physical interpretation when designing weights of sub-models. Predicted variances in this method can be understood as uncertainty of estimates. The predicted variance of the multi-model soft sensor can be calculated by predicted variances of local models using uncertainty combination methods. This method has been applied in a real Erythromycin fermentation process and achieves better performance than single soft sensors and several traditional multi-model soft sensors.(3) As we know, common multi-model soft sensors just consider clustering feature of data set, but ignore statistical characteristics of sub-models. To make the best of the statistical characteristics to improve prediction accuracy of soft sensors, a multi-model soft sensor method based on Dempster-Shafer theory and GPR(DS-PV-MF-MGPR) was proposed. When designing the weights of sub-models, the proposed method not only considers the membership function of clustering algorithm, but also takes statistical characteristics of sub-models into account. The proposed method is validated on the industrial data from a real Erythromycin fermentation process. For comparisons, the single GPR model based soft sensor and traditional multi-model soft sensors were also studied. Simulations show that the proposed method has better prediction accuracy and lower predicted uncertainty.