Abnormal Analysis Of γ Dose Rate Based On Dynamic Threshold

The safety management of nuclear power is the focus of the development of nuclear power in the world.Through the establishment of radiation environment monitoring network,the monitoring system has accumulated a large amount of data for several years,and the data is still growing.It is a crucial method to improve the safety and reduce the loss by effectively detecting the abnormal data of γdose rate and identifying the potential risks of the abnormal performance of monitoring equipment and the irregular operation status of nuclear power plant.The abnormal pattern of the fluctuation of γdose rate data presents the characteristics of time correlation and aggregation,and there is no abnormal label data for reference.It is necessary to use unsupervised method to identify the abnormal from the perspective of data driven.In order to solve the problem of large amount of single variable time series data of γ dose rate and unstable fluctuation of data,this paper uses the Long-Short Term Memory network in the recurrent neural network solve the problem.This model performs a strong fitting and can train any length of single variable time series.By introducing the gate control mechanism to accumulate the original information,the problem of gradient disappearance or gradient explosion can be avoided,improving the accuracy of prediction.The prediction value reflects the fluctuation level of the normal value.We also use the dynamic threshold method to detect local anomaly as the reference baseline.The univariate prediction model uses its own change to predict,without considering the influence of external factors on numerical fluctuation,but the abnormal fluctuation of γ dose rate is closely related to environmental variables.In this paper,we respectively discuss the relationship between temperature,humidity,air pressure,rainfall and γdose rate,and introduce influence factors into the prediction model to further explain the numerical fluctuation.Influence factors also eliminate the influence of environmental variables on the numerical fluctuation as far as possible.By comparing the evaluation indexes of multiple multivariable prediction models,the Gradient Boosting Dicision Tree(GBDT)is used to predict γ dose rate,and then the dynamic threshold method is used to identify local anomalies.