Research On Multiple Models Adaptive Control Based On Neural Networks

With the rapid development of the economy as well as the improvement in technology,the manufacturing processes have gradually become complicated and changeful,and relevant work conditions have been increasingly diversified,which has placed new,strict requirements on the quality control.During certain practical control processes,some occasional situations(such as a part abraded or coming off)will instantaneously and dramatically change the model of a controlled object.With the traditional adaptive control method,a controlled plant is generally studied based on a model whose parameters remain unchanged or are slowly changing,and the operating conditions is time-invariant or slow time-variant.However,in the case of component failure or other unexpected faults,the dynamic model of the system will be abruptly changed,which normally leads to significant transient errors.The traditional adaptive control algorithm has a low convergence rate,resulting in poor control effect.The multiple models adaptive control(MMAC)is considered the most effective method to solve the above problem.Key points of this method include:establish a model set with multiple models according to different working points that the controlled plant possibly has to cover all the uncertainties of this object;design a corresponding controller set based on each model in the model set;define the switching rules according to the identification errors between each model and the controlled object.Once the parameters of the controlled system are changed,based on the switching rules,the system will select,in the model set,the model most applicable to the current controlled system,and then switch to the controller of the selected model.With that in mind,this paper focuses on the numerous nonlinear systems during practical production,establishes a neural network-based multiple models adaptive controller,as well as multiple models at different working points of the controlled object,to convert the uncertainties in the parameters of the controlled object to different weight coefficients of the neural network model.Major achievements are:1.Based on the dynamic neural networks(DNN),this paper sets up multiple DNN identification models(adaptive models,fixed models and re-initialized adaptive models)taking into consideration the unmodeled dynamics in the system respectively from aspects of two typical DNN models:the parallel model and series-parallel model.It also combines such models to establish the model sets,the controller sets and relevant switching rules under different DNN combinations,and then compare the control effects under various combinations.Meanwhile,it proves the system stability and the switching stability.2.Based on the static neural network,this paper proposes a multiple models adaptive control that's based on the OS-ELM neural network,establishes corresponding model set,controller set and switching rules,and proves the stability of the system and the switching.3.Based on the OS-ELM and EM-ELM neural networks,this paper proposes a self-organizing neural network,i.e.,the OEM-ELM neural network.The fundamental elements of the OEM-ELM algorithm include:online learning,network performance evaluation and dynamic increase of hidden layer nodes.The combined advantages of both OS-ELM and EM-ELM not only improves the OEM-ELM's capacity of network identification but also avoids the redundancy of network nodes.This paper also shows application of the OEM-ELM neural network-based adaptive control,and analyzes the node changes' influences on the system.4.Applying the achievements in the production system where one recycles steel slag to produce slag powder,this paper analyzes the production processes of slag powder and sorts out key factors influencing the specific surface area of slag power as well as the pressure difference of the mill.Due to limitations of the actual verification,this paper establishes a series of DNN model cover containing A variety of working conditions based on a large amount of on-site data of practical production to simulate the production conditions,and based on which,the paper develops an OEM-ELM neural network-based adaptive controller to further test the proposed algorithm.