Trend Following Of Time Series Based On The Improved Federated Learning And Its Application

Trend following of time series has attracted lots of attention in the field of industrial security monitor,for instance,the trend following of Electromagnetic Radiation Intensity Time Series Data(ERI-TSD)collected from coal mine can provide an important basis for the safety forecast of rockburst.At present,the definations of industrial time series are relative simple and the accuracy of trend following needs to be improved.In addition,there are still little research concentrate on the trend following considering the needs of multi-sensor time series complementation and privacy protection.Federated Learning(FL)proposed in recent years,can effectively protect data security while improving the complementarity of data feature in the aggregation processing,however,existing related studies seldom focus on the trend following of industrial time series.To this end,this paper studies the federated learning based industrial time series trend following algorithm.The main contents are as follows:(1)An Enhanced LSTM for Trend Following of Time Series: The definitions of the time series trends are first given by referring to the volatility representations in stock option,and convert the trend following into an essential new form of time series prediction.For time series data from single sensor,an enhanced Long Short-Term Memory(LSTM)by effectively combining the particle swarm optimization algorithm and gradient descent method is proposed,and a trend following framework is subsequently built based on the enhanced LSTM.The proposed algorithm is applied to the trend following of ERI-TSD from coal mine and PM2.5-TSD from UCI dataset,respectively,and the experimental results prove that the proposed algorithm can effectively promote the accuracy of trend following.(2)Trend Following of Multi-sensor Time Series based on the Aggregated Feature Obtained Through Federated Learning: Based on research(1),this section further considers the trend following of multi-sensor industrial time series.First,each sensor is regarded as an independent client in FL,and the enhanced LSTM proposed in content(1)will be built on each client to extract the trend features of local time series.Then,the data features and model parameters are both uploaded to the central server at the same time,in addition a feature aggregation strategy for each client is given.Subsequently,based on the aggregated features and the real values of time series stored in clients,the Echo State Network(ESN)is utilized to realize the accurate trend following of each local dataset.Finally,the proposed algorithm is applied to the trend following of multi-sensor collected ERI-TSD,the experimental results illustrate that our algorithm can promote the accuracy of trend following under the premise of protecting data privacy.(3)Trend Following of Multi-sensor Time Series based on the Enhanced Federated Learning with Client-side Heterogeneous Models and Key Parameter Aggregation: Considering the scale difference of local dataset on each client and the high communication cost of the FL algorithm,we further propose an enhanced FL with heterogeneous client models and low communication cost to realize the trend following.First,according to the scale difference,the criteria for constructing client models with shallow or deep neural networks are determined.The singular value decomposition is then used to extract key features of the fully connected layer parameters of client models to reduce the communication cost and to enhance the learning capability of client.The aggregation mechanisim with multiple models on the central server is presented based on the heterogeneous characteristics of the uploaded parameters and models.The proposed algorithm is applied to three benchmark classification datasets and the trend following of multi-sensor ERI-TSD.The results demonstrate that the proposed method can effectively improve the accuracy of local learning models and greatly reduce the communication cost.The research objects in the above research range from single-sensor time series to multi-sensor collected ones,and different trend following strategies are designed for different research objects.The experimental results demonstrate that the study here can greatly promote the accuracy of trend following of time series.This thesis contains 27 figures,21 tables and 101 references.