Research On The Condition Evaluation Technology Of The Complex System Based On Recurrent Neural Network

Due to the rapid development of modern industrial technology,the structure of industrial systems is gradually moving from simple to complex,and the operating rules are becoming more indescribable.The operating conditions of complex systems are directly related to productivity,and unexpected failures can cause not only unplanned system downtime but also significant economic losses.Therefore,it is important to make an accurate evaluation of the operation status in complex systems and achieve warning of system fault anomalies in time,which ensures the system is in a normal operating state.As a typical complex system,wind turbines are used as the object of this paper to investigate the condition evaluation using recurrent neural network(RNN)and its variant networks.Considering the high dimensionality and a large number of samples of SCADA data from wind turbines,an improved random forest algorithm is proposed to identify the features that are selected to be most closely correlated with the operating status of wind turbines.The algorithm mainly uses the Euclidean distance for distinguishing the weights of the same feature among different samples first,after that,the random forest algorithm is used to calculate the importance of the features,and finally feeds the selected features into a two-layer gated recurrent unit(GRU)neural network for training.Furthermore,to more accurately evaluate the condition of wind turbines,an end-to-end training AE-GRU neural network model is proposed that integrates Auto-Encode(AE)and GRU neural networks.The model can not only realize the mining of SCADA data time series features,but also perform deep representation learning on data features.The experimental results show that the wind turbine condition assessment model based on the improved random forest algorithm surpasses the models using the variance threshold method,the L1-SVM algorithm,and the random forest algorithm in terms of accuracy and F1Score performance metrics on both SCADA datasets.Whereas compared with RNN,LSTM,and GRU,the proposed AE-GRU neural network model,on SCADA dataset 1,its accuracy and F1-Score reach 0.9764 and 0.9790,which exceed the second place(GRU neural network model)by about 3.32%and 2.83%respectively;On SCADA dataset 2,the accuracy and F1-Score were 0.9752 and 0.9763,which outperform the second place(GRU neural network model)by about 2.86%and 2.69%,respectively.In addition,the parameters of the models are compared,and the results prove that the AEGRU neural network model also has higher computational efficiency.