A Collaborative Modeling Method Based On Two-Dimensional Broad Learning System

With the development of artificial intelligence and the appearance of various machine learning algorithms,the artificial neural network has emerged as one of the most popular research fields nowadays.It simulates the operation mechanism and organization structure of the nervous system and utilizes its own excellent information processing ability to achieve learning in complex environments.Broad Learning System(BLS)is an emerging flat structured artificial neural network proposed in recent years,which has an excellent performance in the fields of regression and classification with its flexible structure and fast dynamic update,and it is thus widely implemented in image recognition and industrial inspection.However,when it comes to the large-scale tasks,the current BLS algorithms require conventional vectorization operations,which cannot directly handle 2D input data,thereby decreasing the efficiency of model construction;besides,BLS often faces ‘islanded’ data scenarios in practical applications,and it is difficult to effectively integrate data for modeling due to privacy and other limitations.Therefore,on the basis of existing outcomes of BLS,this thesis focuses on the collaborative modeling method based on the two-dimensional broad learning system.This method not only achieves faster and more efficient model construction in image modeling tasks,but also realizes multi-party collaborative modeling without sharing the local data.The main work contents and innovations are as follows:(1)Two-dimensional broad learning system is first proposed.The broad learning system requires vectorization operations in large-scale image data modeling tasks,which leads to complex and time-consuming modeling and even ‘dimensional explosion’.To address this problem,two-dimensional broad learning system(2D-BLS)is proposed by introducing the left and right projection vectors,and the related incremental learning algorithms are designed.2D-BLS can directly process 2D matrix image data,reduce computational complexity and significantly improve the modeling efficiency.The theoretical foundation is laid for the research in the subsequent chapters.(2)The collaborative modeling method based on two-dimensional broad learning system(BC-2D-BLS)is proposed.In practical ‘islanded’ data scenarios,it is hard to improve the performance of local 2D-BLS models,which cannot share data with other clients for joint modeling due to privacy and other limitations.In this thesis,the overall network weights can be updated iteratively by introducing consistency strategy to establish a centrality-free collaborative modeling network architecture,constructing incremental Lagrangian function,and using the alternating direction multiplier method.Furthermore,different incremental strategies are introduced in the collaborative modeling method to achieve dynamic updates while improving the overall performance of the BC-2D-BLS network.(3)The collaborative modeling method based on two-dimensional broad learning system proposed in this thesis is applied to an industrial example of circuit board defect detection,and a model for identification and classification of circuit board defects is established.Experiments are carried out in terms of modeling efficiency and collaborative modeling effectiveness.And the results show that the collaborative modeling method based on two-dimensional broad learning system can effectively meet the practical industrial demands.In summary,this thesis focuses on a collaborative modeling method based on two-dimensional broad learning system,and the developed two-dimensional broad learning algorithm significantly improves the learning efficiency in image data model construction.On this basis,the proposed collaborative modeling method not only protects data privacy,achieves multi-client collaborative modeling,and improves the performance of the overall model,but also provides a novel and applicable learning method in the field of machine learning.