Research On Data Fusion Method Based On Deep Learning

As the technology of obtaining information developing,the data that people need to process is also increasing,which comes up with the problems of heterogeneous data type,poor data quality,fast data generation and so on.These problems have already become one of the main obstacles in restricting the data processing techniques.Currently,as one of the important method of data fusion techniques,the data fusion method based on deep learning has already become an effective way to solve the multisource heterogeneous data processing in the context of big data,having the significance of study.Based on the summary of previous research,this thesis conducts research on data fusion methods based on deep learning.This thesis first proposes a data fusion model based on LSTM-Attention to achieve multi-source time series data fusion;then in order to achieve multi-source heterogeneous data fusion,propose Multi-source heterogeneous data fusion model based on FC-SAE.The main research contents of this thesis are as follows:1.In the method of data fusion which based on the deep learning,the quality of generating the data features plays an essential role on the effect of fusion subsequently.From the perspective of data feature extraction,this thesis focuses on two common data types,time series data and text data,to study its deep learning feature extraction methods.For time series data,first analyze the characteristics of time series data,and study the feature extraction principles of common time series feature extraction models:ARIMA model and RNN model,and analyze the applicable scenarios and model deficiencies of each method;for text data,This thesis mainly studies the feature extraction principle of TF-IDF algorithm and Word2 vec algorithm,and provides theoretical support for subsequent research work.2.Multi-source data contains a more comprehensive descriptions on the object.Compared with the traditional data,decisions using the multi-source data tend to achieve a better result.In this thesis,in order to to increase the prediction accuracy of time series and make full use of the multi-source time series data,a multi-source data fusion model based on the LSTM-Attention was proposed.The model has used the LSTM algorithm to model the time series,extract the potential features of time series data,and filter the data of significance to the time series prediction in multi-source data through the Attention mechanism,so as to realize the data fusion of multi-source time series,establishing the LSTM-Attention multi-source data fusion model.3.There exists the heterogeneous data type problem between the time series data and textual data,it is difficult to efficiently fuse them using the traditional data fusion methods.In order to take full advantage of the information contained in the textual data and improve the accuracy of time series prediction,the thesis has put forward a multisource heterogeneous data fusion model which was based on the FC-SAE.The model has adopted the Glo Ve word embedding model and CNN to extract the features of textual data,using the FC neural network to extract underlying features of time series data.It has used the SAE model to fuse the multi-source heterogeneous data features,fully mining the data association relationship through,and further enhancing the accuracy of time series prediction.