Research On Anomaly Detection Method Of Main Pump State Data In Nuclear Power Station

In the process of power generation,the use of traditional fossil fuels has brought serious environmental pollution and global warming.As a kind of clean energy,nuclear power is quite mature in technology and has strong supply capacity.But when nuclear power fails,it is an enormous accident.Throughout the history of several major nuclear accidents,each time both caused immeasurable damage.Therefore,we study the technology of anomaly detection,fault diagnosis and other technologies related to the safe operation of nuclear power plant.These technologies have great practical significance and application value for promoting the further development of nuclear power industry in the direction of safety and efficiency.In the actual operation and maintenance of the main pump,most of them use a simple threshold determination method to detect anomalies,which is based on the information of some measuring points.This traditional threshold judgment has limitations.On the one hand,the monitoring information utilization rate and state evaluation accuracy of the main pump are low.On the other hand,the threshold method is difficult to detect the latent faults and types of fault during the operation of the main pump.Due to the difference of equipment,the threshold is often set according to experience.Based on the above problems,this paper proposes an anomaly detection model which based on the combination analysis of single-dimensional state data feature extraction and multi-dimensional state data feature extraction.For the one-dimensional state parameters,the auto regressive model AR(1)is used to obtain the model parameters,then the time-dependent transition probability sequence of the one-dimensional state parameters is obtained by combining the quantitative results of SOM neural network.For multi-dimensional state parameters,unsupervised clustering algorithm OPTICS is used to cluster to generate different pattern groups.Then,according to the comprehensive analysis of the two kinds of feature extraction results,the anomaly detection model of this paper is obtained.Finally,the anomaly detection model proposed in this paper is applied to the anomaly detection of the state data of the main pump in a nuclear power plant.Experimental results show that the proposed model based on feature extraction of single-dimensional state data and multi-dimensional state data has good performance in real-time and accuracy.In order to further verify the performance of the method,we compare this method with the traditional method.These traditional method include the multi-dimensional state data combination analysis method based on k-means(AR(1)+SOM+ k-means)and the multi-dimensional state data combination analysis method based on DBSCAN(AR(1)+SOM+DBSCAN).The experimental results once again show that the method in this paper does have good real-time performance and accuracy.