Research On Structural Intelligent Control Algorithms Based On Long Short-term Memory Networks

The control algorithm directly determines the performance of the active control system and the vibration control effect.The traditional control algorithm needs to establish the accurate mathematical model of the controlled structure,which is complex and difficult to simplify.In order to have higher robustness,the structure vibration control algorithm tends to be intelligent gradually.The artificial neural network control algorithm has been widely used for its good adaptive and self-learning ability.However,traditional artificial neural network exposes many problems with the increasing complexity of the controlled structure and the increasing demand for control,such as no correlation between the same level data,insensitivity to time parameters,low performance of feature extraction,and so on.It is difficult to meet the needs of vibration control for the increase of building volume and the complexity of calculation data.To sum up,it is urgent to study a new neural network model which is more suitable for civil engineering vibration control.Deep learning,as a machine learning strategy for analyzing and interpreting data by imitating human brain mechanism,has many advantages over artificial neural network,such as the way to extract data features is closer to the process of brain learning and thinking;the extracted features have more essential expression to the original data;the prediction results are more practical,etc.In this paper,the long short-term memory(LSTM)network in deep learning is combined with the control theory,and a new intelligent control algorithm is proposed,which is introduced into the field of structural vibration control.Taking the high-rise building structure as the research object,the corresponding theoretical research,controller structure improvement and simulation analysis are carried out for the specific problems.The main research contents are as follows:1)The response prediction model of LSTM network is constructed,the model parameters are analyzed and optimized,the displacement values of the structural model under three different types of ground motions are extracted,and carry out training and prediction,which proves that LSTM can be used to predict the response of the building structure;2)The proposed response prediction model of LSTM network is applied to the design of intelligent control algorithm of structure.Taking the three-layer benchmark framework structure as the research object,a Simulink control system simulation platform is built to study its control effect.Compared with BP and RBF shallow learning neural network,it is found that the performance of LSTM intelligent control algorithm is better than shallow learning,which shows that the application of this algorithm to civil engineering domain is feasible;3)The convolution neural network(CNN)and the LSTM intelligent control algorithm are combined to visualize the feature data of each hidden layer,the logic and rationality of each layer are studied,the influence of CNN’s super parameters on the controller is analyzed,and the LSTM centralized controller which integrates CNN level feature learning is put forward,which solve the problems of high dimension of input data processing and large amount of calculation when the centralized control system is in the process of "distributed collection,centralized processing".so as to improve the operation accuracy and speed of high-rise building central controller;4)Based on the study of the above-mentioned centralized control algorithm of LSTM,four kinds of overlapping and independent decentralized LSTM controllers are designed.According to Lyapunov stability theory,the genetic algorithm(GA)is used to optimize the super parameters of LSTM.At the same time,aiming at the possible actuator failure and other problems in the control system,the system simulation of 20 layer benchmark frame structure is carried out.Compared with the results of centralized control,the problems of large information interference and poor reliability in the centralized control are solved,and the superiority of the intelligent distributed control of LSTM is verified.