Research On Incremental Bayesian Broad Learning System And Its Industrial Application

The prediction of key variables in the industrial production process is helpful to optimize energy allocation and improve energy use efficiency,which is an important way to save energy and reduce carbon emissions.A large amount of data acquired through industrial processes makes it feasible to build a data-based model.However,on the one hand,due to the complex situation in the industrial process,the prediction model needs to be constantly updated quickly according to the ever-changing demand;on the other hand,the collected data often contains random missing values,which brings great challenges to the prediction work.As a neural network with a broad structure,a broad learning system(BLS)can update the model quickly and effectively,which is suitable for the prediction of industrial process variables.But it is difficult to determine the ridge parameters for calculating the output weight in a BLS.Therefore,an incremental Bayesian broad learning system(IBBLS)is proposed to solve the above problems.For the problems of industrial process variables prediction and model updating,IBBLS is proposed,where the posterior mean and covariance over the output weights are both derived and updated in an incremental manner,and the hyper-parameters are simultaneously updated by maximizing the evidence function.In such a way,the scale of matrix operations is capable of being effectively reduced.To verify the performance of the proposed method,a number of experiments by using 6 benchmark datasets and an industrial case are carried out.The experimental results demonstrate that the proposed method can not only achieve a better outcome compared to the classical BLS and other comparative algorithms but also incrementally remodel the system.For the problem of incomplete time series prediction,a Bayesian broad learning system for the dataset with missing data is proposed based on the Expectation-Maximization(EM)algorithm,where the output weight and the missing points are taken as unknown variables and the training process consists of two steps.In the first step,the expectation of the logarithmic likelihood function of the complete data is calculated with the posterior probability distribution of unknown variables.In the second step,the logarithmic likelihood function is maximized to update the value of the unknown variable.The above two steps are repeated continuously until the model converges,and the distribution of output weights is iteratively estimated while the missing points are filled.To verify the performance of the proposed method,a number of experiments by using 5 comparative method on synthetic datasets,benchmark datasets,and industrial experiment datasets.The results show that the proposed method can have better performance whether on dataset with low missing percent or with high missing percent,and has more advantages in training speed compared with the other EM-based algorithm.