Multi-modal Time Series Data Error Discovery Algorithm Based On Hybrid Attention Mechanism

In recent years,the development of the Internet of Things has accelerated the increase in the number of Internet access devices,which generate a large amount of multi-modal time series data every day.Compared with single-modal data,multi-modal data can describe more abundant scenarios,so there are a lot of hidden relationships between different modal data to be tapped.Error discovery is an important part of the data quality control process.In the process of data collection,transmission,and processing,data errors may o ccur due to equipment reasons,system reasons,external environment reasons,and human factors.For multi-modal time series data,when the data of a certain modal is wrong,we can judge not only by the data around the time of the modal error data,but also by the data of other modals at the same time.Therefore,studying the multi-modal time series data error discovery algorithm can detect the quality of multi-modal time series data and evaluate whether there are errors in the data.The research content of this paper includes three parts: multi-modal time series data set construction,preprocessing and preliminary feature extraction research,multi-mode time series data fusion algorithm based on deep learning,and multi-mode time series data error discovery model.First,this paper proposes a multi-modal time series data error addition algorithm,which can add specified types of errors to a specified proportion of data in a specified mode.In this paper,data alignment between modalities is achieved,and pre liminary features of different modal data are extracted.Secondly,this paper proposes a feature extraction method based on Bi LSTM,and then proposes a multi-modal feature fusion method based on hybrid attention.First,the multi-output of the Bi LSTM network of each modal data is fused,and then the deep and shallow Layer features are fused.And discusses the advantages and disadvantages of the multi-modal feature fusion method based on hybrid attention and the existing Tensor-Fusion and Cascade-Fusion methods.Finally,this paper proposes a multi-modal time series data error discovery model based on hybrid attention,and analyzes the performance of each module of the model in detail through comparison with existing methods.Four different data sets were used in the experiment.The experiment shows that the multi-modal time series data error detection model proposed in this paper is higher than other comparison models in all aspects.