Study On Prediction Method Of Industrial Time Series Data

The industrial field of our country tends to automation,large-scale and systematization gradually,and once special operating conditions appear in key production equipment,it will have an impact on the entire production system.Therefore,the safety of production equipment has gradually attracted attention.The popularity of sensor networks in many real-world systems(such as intelligent buildings,factories,power plants,and data centers)generates a large amount of multi-dimensional time series data for production equipment.Rich sensor data can be continuously monitored through anomaly detection.However,due to the dynamic complexity of these systems,traditional anomaly detection methods are inadequate.And the supervised machine learning methods cannot be used because of the lack of labeled data.Therefore,this dissertation studies the prediction of industrial time series data based on generative adversarial networks.This dissertation designs and implements a time series prediction algorithm for industrial time series data,and proposes to optimize the network training efficiency through transfer learning methods.The main work of this dissertation includes:First of all,a prediction method suitable for industrial time series is proposed.By improving the traditional Gated Recurrent Unit(GRU)algorithm,the prediction accuracy is improved.And compared with the classic time series prediction methods,the performance of the improved algorithm is evaluated.Secondly,an unsupervised multivariate time series prediction method is proposed,which is based on Generative Adversarial Networks(GAN).This method uses the gated recurrent neural network as the basic model(i.e.,generator and discriminator)to capture the time correlation of the time series distribution under the GAN framework.In addition,a method for detecting abnormal values of time series data by using a generator and a discriminator is proposed.Finally,a time series prediction transfer learning method is proposed,which is based on domain adaptation.In view of the fact that the new production line lacks sufficient training data,it is proposed to use other historical data of similar production lines for transfer learning,and to eliminate the distribution difference of the data set through the domain adaptive method,thereby improving the speed of model training and prediction performance.This subject is facing the increasingly strong working conditions of industrial manufacturing equipment and the needs of status data analysis.For the core problem of multi-parameter fusion status prediction of industrial manufacturing equipment,the following three methods are proposed,a time series prediction method based on GRU network and a time series anomaly detection algorithm based on GAN,and the time series prediction transfer learning method based on domain adaptation.Not only can adapt to the actual different types of time series data application scenarios,but also it effectively improves the performance of industrial time series prediction methods.