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

Anomaly detection for time series data of industrial equipment can provide an important basis for fault early warning.Due to the serious imbalance of industrial equipment time series data,the traditional supervised anomaly detection methods will be invalid,and the accuracy of anomaly detection models need to be improved.Using data from multiple equipment can improve the accuracy,but data privacy leakage may occur in the process of data aggregation.Federated Learning can train a model using data from multi nodes while protecting data privacy.Therefore,a Federated Learning-based time series anomaly detection method（FL-TSAD）is proposed,which contains:（1）The time series anomaly detection model of FL-TSAD.Aiming at the problems of serious imbalance of data labels,unknown prior of abnormal behavior and long-time dependence of time series data,an unsupervised detection model is proposed,where LSTM-DVAE is used to reconstruct data and the reconstruction error is used to detect anomaly.The experiments are carried out on four datasets.Compared with the other three models,the average effect on F₁-Score of the model was increased by 43.95%,36.84%and 21.54%respectively.The results prove the superiority of the proposed method.（2）The distributed training framework of FL-TSAD.Aiming at the problem of data privacy leakage that may occur when data from multiple equipment are aggregated for anomaly detection,Fed Avg algorithm is combined with the model proposed on research（1）and a distributed training framework is proposed.During the training process,only the model is uploaded,and the original data is kept locally,so the data privacy is guaranteed.The experiments are carried out on the above four datasets.Compared with the other three models,the average effect on F₁-Score of the framework was increased by 40.37%,35.51%and 7.91%respectively.The results illustrate that the proposed framework can improve accuracy of the model and guarantee data privacy.（3）The communication-efficient algorithm of FL-TSAD.Aiming at the problem of high communication cost of the traditional Fed Avg algorithm,a gradient compression based Fed Avg algorithm（GC-Fed Avg）is proposed.During the training process,only the gradient exceeding the threshold is uploaded,therefore,the communication cost is reduced.The experiments are carried out on the above four datasets.Compared with Fed Avg algorithm,F₁-Score is reduced by 2.71%and running time is reduced by 49.80%on average.The results demonstrate that the proposed algorithm can reduce the communication cost and shorten running time on the basis of small loss of accuracy.Experiments on four datasets show that the proposed method has good generality not only in the field of industrial equipment anomaly detection,but also in other fields.