Research On Multivariate Time Series Anomaly Detection Method Based On Federated Learning

Anomaly detection of time series data has been widely studied and applied in network security detection,autonomous vehicles,daily maintenance of large-scale industrial equipment and other fields.Anomaly detection of multivariate time series data is faced with several challenges.First,the existing anomaly detection methods have a single model,and most of them do not consider the characteristics of time and variables at the same time,resulting in poor robustness of the model.Second,for the sake of data security and privacy,data agencies are reluctant to share data with each other,and the phenomenon of "data islands" is serious,which significantly reduces the performance of anomaly detection.This paper focuses on anomaly detection of multiple time series data and federal learning,and the main work is summarized as follows:Firstly,aiming at the current challenge of multivariate timing data anomaly detection,a multivariate timing data anomaly detection model based on dual-channel Transformer encoder is designed.Aiming at the problem of multi-temporal data anomaly detection,a multi-temporal data anomaly detection method TcTAD based on dual-channel Transformer encoder is proposed.In this method,dual-channel Transformer encoder is used to learn time dimension features and spatial dimension features of time series at the same time,and the features of the two dimensions are connected through gating network.After feature fusion,the features are input to the linear layer and the prediction of timing data is output.Meanwhile,aiming at the problem that timing data is sometimes difficult to predict,Transformer decoder is used to reconstruct time series data,and the prediction error and reconstruction error sum are taken as the criterion of anomaly judgment.Finally,the two loss functions are jointly optimized to improve the performance of anomaly detection.Experimental results show that the model has superior performance in anomaly detection.Secondly,aiming at the problem of high communication cost in federated learning,an anomaly detection method FedADD based on federated self-attention distillation is designed.A privacy-oriented multivariate time-series data anomaly detection scenario,integrating TcTAD into the federal learning framework for data privacy issues.At the same time,aiming at the problem of high communication cost in federated learning and the problem of Transformer model with many parameters,the deep self-attention distillation is introduced to compress the model and the adaptive distillation method is proposed.The distillation intensity of teacher model to student model is positively correlated with the performance of teacher model,and the teacher model is updated locally.Student models are trained cooperatively among different users through federated learning.Experimental results show that model compression can effectively reduce communication overhead and allow heterogeneity of models.Finally,two anomaly detection methods FedTAD and FedMAD based on personalized federated learning are designed to solve the problem of data heterogeneity in federated learning.Firstly,the multiple time-series data anomaly detection scenario is privacyoriented.Aiming at the performance loss caused by data heterogeneity in federated anomaly detection,FedTAD introduces freeze fine tuning technology in migration learning based on federated learning framework,which can further improve the performance of federated anomaly detection.Second,privacy protection oriented multivariate time-series data anomaly detection scene,as the data in the work of the abnormal heterogeneous model of the problem of low convergence FedMAD combined federal and meta learning model in the learning process of training a applies to all user model initialization parameter,improve the model convergence.