Prediction Research Based On Security Monitoring Data Of Internet Of Things

With the continuous development of China's Internet of Things(Io T),5G and other technologies,Io T monitoring technology is also being applied in various industries,such as safety production,smart mines,and smart buildings.The application of the Io T monitoring technology is very extensive,and the monitoring data collected through the Io T devices has also shown a blowout growth.Massive amounts of data provide a data foundation for basic research such as big data and artificial intelligence.The node state in Recurrent Neural Network(RNN)is transmitted in its own network in a cyclic manner,so that the RNN has a better processing effect in the processing of time series data.Since the RNN uses a sharing mechanism to transfer its parameters during the training process,it is prone to gradient dispersion and gradient explosion problems during the training process.Therefore,it is difficult to capture longterm associations in the data.This paper conducts in-depth research on the RNN,analyzes the reasons for gradient dispersion and gradient explosion of the RNN,and proposes corresponding solutions.Long Short-Term Memory(LSTM)is also a recurrent neural network,which can be used to process individual data or the entire sequence of data.The LSTM solves the problem of gradient dispersion and gradient explosion in the training process of the RNN model,and also can capture long-term associations in data and has more advantages than the RNN.However,the LSTM still has certain shortcomings,such as low prediction accuracy,hyperparameters difficulty in selection and poor interpretability,etc.This paper takes the prediction of time series data collected by Io T monitoring equipment in the field of safety production as the research background,combined with the problems in actual safety production applications,and conducts theoretical research and modeling analysis based on recurrent neural networks and particle swarm optimization algorithms.The main content including the following three aspects:(1)Because the traditional single prediction model cannot effectively reduce the impact of random errors in the time series data on the prediction results,the prediction accuracy is low.Therefore,this paper proposes a combined prediction model of long short-term memory based on residual correction,which uses a neural network model to replace the traditional statistics-based prediction model,and improves its prediction accuracy by reducing random errors in time series data.First divide the time series data set into two parts: training set and test set,and then normalize them separately,then use the LSTM to train the training set data once and get the residual sequence on the training set,and finally the residual sequence is used to correct the secondary training of the LSTM.Experimental verification shows that the model has higher prediction accuracy,lower error and better prediction performance than single and combined prediction models.(2)The LSTM has a complex network topology and numerous hyperparameters,and there are problems such as difficulty of selecting hyperparameters,the hyperparameter adjustment of the neural network model is generally adjusted based on previous experience,resulting the interpretability of the final network model is poor.The hyperparameter optimization problem of the LSTM can be abstracted into a type of optimization problem.Therefore,this paper introduces particle swarm optimization(PSO)to solve the hyperparameter optimization problem of the LSTM model,but the Basic PSO(BPSO)also has certain defects and shortcomings.The BPSO is improved and optimized,and a dynamic adaptive particle swarm optimization algorithm based on chaos optimization(DAWCPSO)is proposed.First,a new dynamic adaptive inertial weight is proposed to balance the local and global search behavior of particles;when the particle is in the local optimal state,the chaos optimization strategy is used to expand the search range of the particle,so that the particle can continue to search for other solutions with higher accuracy in the feasible region.Theoretical analysis and experimental verification show that the DAWCPSO not only overcomes the premature convergence problem of the BPSO,but also improves the accuracy and convergence speed of the optimization result without affecting the time complexity of the BPSO.(3)In view of the difficulty of selecting hyperparameters in the LSTM,the PSO is introduced to optimize the seven hyperparameters of the LSTM.Firstly,it improves and optimizes the shortcomings of the BPSO,and uses the improved the APSO to optimize the hyperparameters of the LSTM.Theoretical analysis and experimental verification show that the APSO enables the LSTM to adaptively adjust its network topology and hyperparameters according to different time series data;compared with traditional parameter optimization algorithms,the APSO has the advantages of fast convergence,low computational overhead,and strong robustness;and compared with the prediction performance of single prediction model and combination prediction model,the APSOLSTM has better performance;in short-term prediction scenarios,the BPSO-LSTM has certain advantages,but the results are not very obvious,the prediction results of the LSTM are not much different,but the prediction performance of the APSO-LSTM is slightly better than the LSTM;compared with the other two models,the APSO-LSTM has lower error,higher accuracy and stronger robustness in medium and long-term prediction scenarios.Although the three prediction models have good prediction performance,the comprehensive prediction performance of the APSO-LSTM has stronger reliability and superiority,and also has higher reference value and application prospects.