CN-122020261-A - Small sample cold source system anomaly detection method and device based on graph neural network

CN122020261ACN 122020261 ACN122020261 ACN 122020261ACN-122020261-A

Abstract

The invention provides a small sample cold source system anomaly detection method and device based on a graph neural network, and relates to the technical field of anomaly detection, comprising the steps of carrying out data processing on multi-source monitoring data to obtain time sequence data, constructing a graph structure according to physical coupling relations among various monitoring indexes in the time sequence data, and carrying out feature extraction processing on the graph structure through the graph neural network to obtain node embedded features; the method comprises the steps of carrying out time sequence modeling processing on node embedded features through a long-term and short-term memory network to obtain a prediction vector, carrying out self-supervision disturbance training on a graph structure to generate a simulated positive sample and a simulated negative sample so as to obtain a coupling sensitivity vector based on the prediction vector, the positive sample and the negative sample, calculating an abnormal sensitivity set based on the coupling sensitivity vector, and determining a multi-scale coupling abnormal score based on the abnormal sensitivity set so as to determine a target abnormal detection result according to the multi-scale coupling abnormal score. The invention can obviously improve the sensitivity of the anomaly detection of the cold source system.

Inventors

Lang Chaohao
HONG HAO
ZHANG SUJUN
WU JIANMING

Assignees

浙江远算科技有限公司

Dates

Publication Date: 20260512
Application Date: 20260413

Claims (10)

1. The small sample cold source system anomaly detection method based on the graph neural network is characterized by comprising the following steps of: carrying out data processing on multi-source monitoring data of a cold source system to obtain standardized time sequence data, and constructing a graph structure according to physical coupling relations among various monitoring indexes in the time sequence data so as to carry out feature extraction processing on the graph structure through a graph neural network to obtain node embedding features; performing time sequence modeling processing on the node embedded features through a long-short-term memory network to obtain a prediction vector, performing self-supervision disturbance training on the graph structure to generate simulated positive samples and negative samples, and performing optimization processing on potential feature space based on the prediction vector, the positive samples and the negative samples to obtain an enhanced coupling sensitivity vector; Calculating an abnormal sensitivity set based on the coupling sensitivity vector, and carrying out weighted fusion processing on each abnormal sensitivity in the abnormal sensitivity set to obtain a multi-scale coupling abnormal score so as to determine a target abnormal detection result according to the multi-scale coupling abnormal score, wherein the abnormal sensitivity set comprises local coupling sensitivity, global coupling sensitivity and univariate dynamic sensitivity.
2. The method for detecting the abnormality of a small sample cold source system based on a graph neural network according to claim 1, wherein the step of performing data processing on multi-source monitoring data of the cold source system to obtain standardized time series data comprises the steps of: performing linear interpolation restoration processing on short-time missing data in the multi-source monitoring data, and removing a time sequence window corresponding to the long-time missing data to obtain a complete continuous monitoring sequence; removing the acquisition error points in the monitoring sequence to obtain cleaned monitoring data; and carrying out normalization processing on the cleaned monitoring data according to the physical upper limit and the physical lower limit of each monitoring index in a preset standard operation interval, and mapping the monitoring indexes with different dimensions into a unified numerical value interval to obtain the standardized time sequence data.
3. The method for detecting the abnormality of the small sample cold source system based on the graph neural network according to claim 1, wherein the step of constructing the graph structure according to the physical coupling relation between each monitoring index in the time series data comprises the following steps: Acquiring preset initial side weights, wherein the initial side weights are used for reflecting initial physical coupling strengths among all monitoring indexes; In the running process of the system, carrying out dynamic calculation processing on the similarity among all monitoring indexes based on real-time data in a sliding time window, and obtaining a dynamic edge weight which is adaptively updated along with time according to a dynamic calculation result and the initial edge weight; And taking the monitoring indexes as graph nodes, taking the dynamic edge weights as connecting edges among the graph nodes, and constructing the graph structure reflecting the evolution of the multivariable coupling relation in real time.
4. The small sample cold source system anomaly detection method based on the graph neural network according to claim 1, wherein the step of generating simulated positive and negative samples by performing self-supervision disturbance training on the graph structure comprises the following steps: Applying a first disturbance to an original node or edge in the graph structure to simulate an abnormal state and various operation conditions of a data layer to obtain a first positive sample and a first negative sample; and applying a second disturbance to the node embedded feature in the graph structure to simulate the abnormal state and each operation condition of the coupling relation layer so as to obtain a second positive sample and a second negative sample.
5. The method for detecting the abnormality of the small sample cold source system based on the graph neural network according to claim 1, wherein the step of optimizing the potential feature space based on the prediction vector, the positive sample and the negative sample to obtain the enhanced coupling sensitivity vector comprises the following steps: And inputting the prediction vector, the features corresponding to the first positive sample and the first negative sample and the features corresponding to the second positive sample and the second negative sample into a contrast loss model for calculation processing, so that the features corresponding to the prediction vector are mutually gathered in a potential feature space, and the coupling sensitive vector after enhancement is obtained.
6. The small sample cold source system anomaly detection method based on the graph neural network according to claim 1, wherein the step of calculating an anomaly sensitivity set based on the coupling sensitivity vector comprises: Calculating characteristic differences between each node and neighbor nodes based on the coupling sensitivity vector and the dynamic coupling weight between the nodes to obtain the local coupling sensitivity; Calculating the deviation degree of each node relative to the overall operation state of the system based on the Euclidean distance between the coupling sensitivity vector and the global feature center, so as to obtain the global coupling sensitivity; And carrying out standardization processing by combining a preset standard deviation based on the deviation between the actual observed value of the time sequence data and the predicted value corresponding to the predicted vector to obtain the univariate dynamic sensitivity.
7. The small sample cold source system anomaly detection method based on the graph neural network according to claim 1, wherein the step of determining a target anomaly detection result according to the multi-scale coupling anomaly score comprises the following steps: acquiring a time sequence prediction error, and carrying out fusion processing on the multi-scale coupling anomaly score and the time sequence prediction error to obtain a comprehensive anomaly score; And carrying out self-adaptive threshold adjustment processing according to the comprehensive anomaly score in the historical sliding window, determining an anomaly judgment threshold value at the current moment, comparing the comprehensive anomaly score with the anomaly judgment threshold value, determining the anomaly grade at the current moment according to a comparison result, and outputting anomaly early warning information.
8. The utility model provides a little sample cold source system anomaly detection device based on picture neural network which characterized in that, the device includes: The system comprises a graph structure modeling module, a node embedding module and a node embedding module, wherein the graph structure modeling module is used for carrying out data processing on multi-source monitoring data of a cold source system to obtain standardized time sequence data, and constructing a graph structure according to physical coupling relations among various monitoring indexes in the time sequence data so as to carry out feature extraction processing on the graph structure through a graph neural network to obtain node embedding features; The self-supervision contrast learning module is used for carrying out time sequence modeling processing on the node embedded features through a long-term and short-term memory network to obtain a prediction vector, carrying out self-supervision disturbance training on the graph structure to generate simulated positive samples and simulated negative samples, and carrying out optimization processing on potential feature space based on the prediction vector, the positive samples and the negative samples to obtain an enhanced coupling sensitive vector; The anomaly judgment and early warning module calculates an anomaly sensitivity set based on the coupling sensitivity vector, and performs weighted fusion processing on each anomaly sensitivity in the anomaly sensitivity set to obtain a multi-scale coupling anomaly score so as to determine a target anomaly detection result according to the multi-scale coupling anomaly score, wherein the anomaly sensitivity set comprises local coupling sensitivity, global coupling sensitivity and univariate dynamic sensitivity.
9. A server comprising a processor and a memory, the memory storing computer executable instructions executable by the processor, the processor executing the computer executable instructions to implement the method of any one of claims 1 to 7.
10. A computer readable storage medium storing computer executable instructions which, when invoked and executed by a processor, cause the processor to implement the method of any one of claims 1 to 7.

Description

Small sample cold source system anomaly detection method and device based on graph neural network Technical Field The invention relates to the technical field of anomaly detection, in particular to a small sample cold source system anomaly detection method and device based on a graph neural network. Background The nuclear power cold source system is an important supporting link for safe operation of a nuclear power unit, the operation state monitoring and abnormality early warning mainly depend on parameter data collected by an operation monitoring system, and at present, related technologies propose that a common abnormality detection method mainly comprises manual rules and static threshold judgment, a supervised machine learning model and a traditional unsupervised detection algorithm, but the manual rules and the static threshold judgment are too dependent on experience, false alarm or missing alarm is easy to generate, a real fault sample required by the supervised machine learning model is more and the labeling cost is higher, the modeling capability of the traditional unsupervised detection algorithm on a multivariable dynamic coupling relation and a time sequence trend is insufficient, the generalization capability is limited, weak trend abnormality is difficult to capture, in addition, the cold source system always shows weak trend abnormality before abnormality occurs, the abnormality detection sensitivity of the prior art scheme is lower, and the recognition capability of fault precursors is generally lacking, so that the effective early warning is difficult to realize. Disclosure of Invention Therefore, the invention aims to provide a small sample cold source system abnormality detection method and device based on a graph neural network, which can remarkably improve the sensitivity of cold source system abnormality detection. According to the method, data processing is conducted on multi-source monitoring data of a cold source system to obtain standardized time sequence data, a graph structure is built according to physical coupling relations among monitoring indexes in the time sequence data, feature extraction processing is conducted on the graph structure through the graph neural network to obtain node embedded features, time sequence modeling processing is conducted on the node embedded features through a long-period memory network to obtain a prediction vector, self-supervision disturbance training is conducted on the graph structure to generate simulated positive samples and negative samples, the potential feature space is optimized based on the prediction vector, the positive samples and the negative samples to obtain enhanced coupling sensitivity vectors, an abnormal sensitivity set is calculated based on the coupling sensitivity vectors, weighting fusion processing is conducted on each abnormal sensitivity in the abnormal sensitivity set to obtain multi-scale coupling abnormal scores, and a target abnormal detection result is determined according to the multi-scale coupling abnormal scores, wherein the abnormal sensitivity set comprises local coupling sensitivity, global coupling sensitivity and single-variable dynamic sensitivity. In one embodiment, the method comprises the steps of carrying out data processing on multi-source monitoring data of a cold source system to obtain standardized time sequence data, carrying out linear interpolation restoration processing on short-time missing data in the multi-source monitoring data, carrying out rejection processing on time sequence windows corresponding to long-time missing data to obtain a complete continuous monitoring sequence, carrying out rejection processing on acquisition error points in the monitoring sequence to obtain cleaned monitoring data, carrying out normalization processing on the cleaned monitoring data according to physical upper limits and physical lower limits of various monitoring indexes in a preset standard operation interval, and mapping the monitoring indexes of different dimensions into a unified numerical value interval to obtain the standardized time sequence data. In one embodiment, the step of constructing the graph structure according to the physical coupling relation among the monitoring indexes in the time sequence data comprises the steps of obtaining preset initial edge weights, wherein the initial edge weights are used for reflecting initial physical coupling strength among the monitoring indexes, dynamically calculating the similarity among the monitoring indexes based on real-time data in a sliding time window in the running process of the system, obtaining dynamic edge weights which are adaptively updated along with time according to a dynamic calculation result and the initial edge weights, taking the monitoring indexes as graph nodes, taking the dynamic edge weights as connecting edges among the graph nodes, and constructing the graph structure reflecting the evolution of the multivariable coupling r