Soft-sensor Modeling Based On Machine Learning And Its Application

Data-driven modeling is an important method of soft-sensor modeling technology. Because of the development of industrial technology, the nonlinearity, multiple working conditions and other complex features of modeling objects request data-driven soft-sensor models to have better generalization ability. Building soft-sensor models with good generalization ability and estimation accuracy for complex industrial process is a difficult problem to solve, and the application of machine learning theory in data-driven soft-sensor modeling provides an effective way to solve that problem. In order to improve the generalization ability of soft-sensor models for complex industrial process, this paper researches on soft-sensor modeling based on machine learning, and mainly studies on utilizing multi-modeling based on clustering analysis or ensemble learning and local modeling to improve the generalization ability of soft-sensor models. Main results are as follows:1. It’s usually difficult to describe the complex features of working conditions for single model. Multi-modeling based on clustering is a common way to solve that problem, and is able to improve the generalization ability of soft-sensor models. The result of clustering has an important influence on the estimation ability of multi-model which is based on clustering, so a multi-model soft-sensor modeling method based on improved affinity propagation clustering algorithm is proposed. In order to improve the effect of clustering, the artificial fish-swarm algorithm is applied to optimize the preference parameter and damping parameter in affinity propagation clustering algorithm. And Gaussian process regression algorithm is utilized to build regression submodels for every cluster. The simulation results for a standard data set and data of an industrial production plant show the effectiveness of the method.2. For multi-model soft-sensor modeling based on clustering, the boundaries of clusters may be not obvious, which will often result in classification error and inaccurate estimation. A multi-model soft-sensor modeling method based on clustering and support vector data description algorithm is proposed. Firstly the data are divided into several clusters using affinity propagation clustering, and then support vector data description algorithm is utilized to define the boundaries of clusters, thus the position relations between test samples and the clusters can be clearly acquired. Based on the information of the position relations, appropriate models are chosen from Gaussian process regression submodels, global model and local model to estimate the output of test samples. The multi-model estimation precision is improved shown by the result of simulation using data from an industrial production plant.3. In order to improve the generalization ability of soft-sensor models for complex industrial process, a Gaussian process ensemble soft-sensor modeling method based on improved bagging algorithm is proposed. This method uses Gaussian process regression algorithm to build base learners, and uses the resample method of bagging algorithm to form training subsets of base learners, and a criteria for feature ordering base on normalized mutual information is proposed to select input features of base learners, which can implement supervised feature perturbance in the ensemble modeling for the sake of improving the diversity between base learners. When estimating the output of a test sample, according to the output variances given by Gaussian process base learners, several base learners are selected adaptively to calculate the output of the ensemble model. A soft-sensor modeling simulation using data from the reactors of an industrial production process shows the effectiveness of the method.4. For an industrial process object with strong nonlinearity and changeful working conditions, building online local model can improve the generalization ability of soft-sensor models. A local weighted mixed kernel partial least squares algorithm is proposed for online soft-sensing modeling and the effectiveness of the algorithm is shown by simulation. The original inputs are mapped into a high dimensional feature space by a kind of mixed kernel function which is made up of several kernel functions with different properties. In the high dimensional feature space, the mapped data are weighted by the weights of each sample which are calculated using the locally weighted learning algorithm. Then, the kernel partial least squares algorithm is used to build an online local soft-sensor model.