Key Techniques Of Time Series Prediction For Industrial Big Data

In recent years,with the advent of the Industry 4.0,industrial big data has become an important research area.Due to the complex production process,a large number of sensors,rapid sampling frequency,and industrial equipment can easily accumulate large amounts of data in a short period,industrial big data has features like complicated mechanism models,time-series data,strong data dependence,high data dimensions,and massive data without labels.When there are special operating conditions,it often has large economic losses,so if an accurate prediction of the abnormalities in the production process is made,it will increase efficiency of the entire production process,contribute to large practical value.This topic mainly target at modeling algorithms for time series data of industrial big data.The traditional analysis methods of industrial big data often emphasize the statistical model and ignores the time correlation of industrial data.Therefore,from the perspective of the time sequence of data,this paper proposes time series data prediction algorithms from three aspects.The main work of this paper includes:First,a yield prediction algorithm based on time series data-a LSTM algorithm based on multi-variable tuning is proposed.The algorithm improves the traditional LSTM algorithm and converts the time series data into supervised learning sequences using periodicity,improving the prediction accuracy.Secondly,a fault prediction algorithm based on transfer learning-a transfer learning algorithm based on time window is brought up.This algorithm puts forward the concept of time window,using transfer learning between different machines with different sampling frequency,solving the learning problem of data without labels in industrial data.Thirdly,a prediction algorithm based on lifelong learning –a data update model is proposed for the lifelong learning prediction algorithm.It updates the existing model combining the data update model to automatically update model parameters over time.It can effectively identify the changes of data and iteratively update the model.To sum up,we have established a time series forecasting model system,which solves the problems of continuous variable prediction,discrete variable prediction and model self-learning updating in time series data.It mainly includes the following innovations: introducing the concept of periodic measurement and time window in the industrial prediction problem,especially for ndustrial data with time series characteristics;introducing transfer learning into time series data of different devices in the same production process;establishing data update model replacing mechanism model to simplify model update.Experimental results show that the time-series yield prediction algorithm improves the prediction accuracy by 54.05%,the transfer learning based fault prediction algorithm has a transfer accuracy about 97%,and lifelong learning prediction algorithm can update the data model effectively with at least 33% accuracy.