Application Of Gated Recurrent Unit In Equipment Maintenance Under Industrial Big Data

In 2014,K.Cho proposed gated recurrent unit(Gru),as an important branch of recurrent neural network,which introduces long-term memory model and gate structure to alleviate the long-term dependence problem.Gru further optimizes the long short term memory model(LSTM)and only retains two gate operations: update gate and reset gate,which makes the structure more simple,so Gru is widely used in temporal data in recent years.At the same time,with the development of national new infrastructure,the implementation of a large number of digital transformation projects of manufacturing enterprises produces a large number of industrial data,which has great value in the whole industrial production process.How to apply appropriate models to mine the information in industrial big data is also a recent research hotspot.Based on the big data of the ice industry of the fan blade,the data preprocessing methods such as data missing value processing,abnormal value processing,balance processing and normalization processing are used to regulate the data,and the data imbalance is solved.The feature engineering methods such as feature selection and feature creation are used to select variables with strong correlation with modeling,which improves the upper limit of the model,and both of them jointly complete the preliminary task of data mining.Then,the gate recurrent unit is established under the tensorflow framework,and the highest prediction accuracy is achieved in the process of iteration.Finally,the comparison between the implementation process and the prediction effect is compared with the traditional machine learning model logic regression,random forest and support vector machine.The implementation process and prediction results show that GRU,with its simple structure and unique GPU operation mode,makes the operation speed exceed other machine learning models,and the total operation time of 20 iterations of the model is about 3min.The best logic regression in other machine learning models also needs to run for about 5min on the basis of characteristic engineering.The results show that the average accuracy of the prediction of GRU is 98%,the average recall rate is 96%,the F value is 0.97 and AUC value is up to 0.85,which is more than the machine learning model compared with it.The good performance of GRU model in the wind turbine blade icing data not only solves the problem of blade icing in wind power industry,but also shows the potential application value of the model in the field of industrial big data equipment maintenance,which lays a foundation for its realization in more application scenarios.