Robust Broad Learning System And Its Network Structure Optimization Based On Unknown Noise Modeling

Deep neural network has developed rapidly in recent years.However,deep neural network has the defects of gradient disappearance and many structural parameters.Broad learning system expands in"wide direction",with fast convergence and high accuracy.Its unique feature layer structure ensures the ability of data mining,and the ridge regression algorithm is used to train the model quickly to ensure the efficiency of model learning.However,there are still some defects in the broad learning system:the width learning system can only have effects on pure Gaussian distribution data in the modeling process.When the data is damaged due to uncertainties or geometric errors,sampling errors,modeling errors,etc.occur in the modeling process,the broad learning system modeling is no longer stable,and its generalization performance will be reduced due to the influence of outliers or other unknown distribution noises.Secondly,due to the mechanism of randomly generating network parameters,the width learning system will have many problems,such as model over-fitting,structural redundancy,and generalization performance degradation.To solve the above problems,the main work of this thesis is as follows:（1）In view of the problem that the broad learning system is vulnerable to outliers or outliers that lead to the degradation of the model’s performance,based on the robust least squares,the aim is to improve the robustness of the model.L_2,P norm are introduced into the objective function of the model,and theL_2,P norm is used to punish the samples with large errors,and a lower weight is given to eliminate outliers through sparse structure.In addition,in view of the characteristics that individual outliers are difficult to eliminate and identify,the probability weight is introduced to remove outliers in a way similar to the attention mechanism,and the probability weight is added to each sample,and whether the sample is a normal sample is identified by assigning a value of 0-1 to further enhance the robustness of model modeling.（2）In the actual industrial process,the collected data will destroy the inherent distribution of the original data due to geometric errors,measurement errors,modeling errors and other reasons,and the actual collected data can no longer be described by a single distribution.In view of the problems of the standard broad learning system,such as poor modeling ability and reduced generalization performance in the face of unknown distributed noise data,Therefore,a robust broad learning system based on mixed Gaussian distribution modeling is proposed.Aiming at the problem that the improved objective function cannot be directly optimized,the model parameters are optimized using the expectation maximization algorithm,and the optimized parameters are solved using KKT theorem to improve the modeling ability of the model to unknown noise.In addition,an incremental algorithm version of the broad learning system is proposed under the framework of mixed Gaussian distribution,It can quickly add nodes or samples with noise without re-modeling,and improve the robustness of the model.The feasibility of the algorithm is discussed through open source data sets and actual data sets in the experiment.（3）Because of the mechanism that the broad learning system randomly generates network parameters due to its horizontal structure,The quality and quantity of network nodes depend on the distribution of random parameters,which will ultimately affect the prediction performance of the model.If there are too few nodes,the learning ability cannot be guaranteed.If there are too many nodes,the calculation amount of the model will be greatly increased,resulting in problems such as over-fitting of the model,precision reduction,etc.Aiming at the optimization of the broad learning system network structure,how to take into account the sparsity and generalization of the network structure,a broad learning system based on adaptive lasso is proposed.The adaptive lasso is used to allocate different adaptive coefficients for each output weight of the broad learning system,select important output weights in the network through adaptive parameters,avoiding over-fitting problems and improving the prediction performance of the model,Finally,the effectiveness of the algorithm is verified on several UCI data sets.