Nonstationary Characteristics Analysis And Deep Learning Prediction Of Industrial Times Series Data

Modern industry is gradually moving towards automation,intelligence and datamation.With the extensive use of sensors,a large amount of industrial time series is generated every day in the production process,which contain the status information of the whole production process.Mining useful information from these massive data is a hot research issue at present.However,industrial time series are often characterized by high dimension,strong coupling and nonstationarity,which brings great challenges to the analysis and prediction of industrial time series data.In this paper,the correlation characteristics and nonstationarity of industrial time series are analyzed.And the research focuses on the deep learning prediction methods under nonstationary time series.The main work and contributions of this paper are as follows:(1)In order to solve the nonstationary industrial time series data analysis and forecasting problems with periodicity,a long short-term memory network(LSTM)prediction method based on correlation analysis is proposed.Firstly,hidden features are extracted by the autocorrelation graph among the real industrial time series.The correlation analysis and mechanism analysis contribute to finding the appropriate secondary variables as model input.In addition,the time variable is complemented to precisely capture the periodicity.Then a LSTM network is constructed to model and forecast sequential data.The experimental results on a certain cooling system demonstrate that the proposed method has higher accuracy and stronger long-term prediction ability compared with several traditional forecasting methods.(2)In order to solve the common nonstationary industrial time series data analysis and forecasting problems,a fractional stochastic configuration networks(FSCN)prediction method based on fractional order differential is proposed.In general,the integer order difference method is commonly used to eliminate the nonstationary characteristic of the sequence.However,this approach always leads to over-difference.Therefore,the fractional differential technique is applied to deal with the nonstationary time series analysis.First,the Hurst exponent of industrial time series data is calculated to determine the order of fractional difference.Then a FSCN network is constructed to model sequential data after difference.In addition,we also give an explicit mathematical analysis about the prediction uncertainty for the proposed FSCN model which provide an additional approach for calculating the confidence interval of prediction results.The experimental results on a certain cooling system data and the open nonstationary benchmark data set demonstrate that the proposed FSCN method has a high accuracy in both the training set and the testing set and effectively improves the prediction effect of the traditional SCN algorithm.