CN-115694934-B - Power data transmission abnormality detection method, system and medium based on federal learning

CN115694934BCN 115694934 BCN115694934 BCN 115694934BCN-115694934-B

Abstract

The invention discloses a federal learning-based power data transmission abnormality detection method, a federal learning-based power data transmission abnormality detection system and a federal learning-based power data transmission abnormality detection medium, wherein the federal learning framework with a parameter monitoring module is deployed at an edge gateway layer of a power system; when the abnormal information monitoring network of the parameter monitoring module detects new abnormal information, the new abnormal information is confirmed whether to be recorded in the buffer zone, if not, the new abnormal information is used for updating the buffer zone and filtering the abnormal information, and after the abnormal information is filtered, the abnormal information is updated by utilizing knowledge transfer to ensure accurate data transmission. The invention effectively detects and screens the abnormal information, accurately prevents the abnormal data transmission behavior in the network and ensures the data transmission safety.

Inventors

LU YOUFEI
CAI YANCHUN
LIU XUAN
LIU LUHAO
LIANG XUEQING
WU RENBO
ZHANG YANG
ZHAO HONGWEI
CHEN MINGHUI
ZHANG SHAOFAN
ZOU SHIRONG

Assignees

广东电网有限责任公司广州供电局

Dates

Publication Date: 20260508
Application Date: 20221019

Claims (8)

1. The method for detecting the abnormal transmission of the power data based on federal learning is characterized by comprising the following steps of: deploying a federal learning architecture with a parameter monitoring module at an edge gateway layer of the power system, wherein the parameter monitoring module comprises an abnormal information monitoring network and a buffer zone; When the edge gateway uploads the power data, the transmission data is used as a parameter set of an abnormal information monitoring network in the parameter monitoring module, abnormal information identification is carried out, abnormal information in the transmission process is obtained, and the abnormal information is transmitted into a buffer zone; When the abnormal information monitoring network of the parameter monitoring module detects new abnormal information, confirming whether the new abnormal information is recorded in the buffer area, if not, updating the buffer area by using the new abnormal information, and filtering the abnormal information; after filtering the abnormal information, updating transmission data by using knowledge transfer to ensure accurate data transmission; The training steps of the parameter monitoring module are as follows: Dividing a parameter set X of an abnormal information monitoring network into an unmarked parameter set U and a marked abnormal parameter set V, wherein the marked abnormal parameter set V comprises the characteristics and the labels of each marked parameter; Performing initial training on the abnormal information monitoring network by using a cross entropy loss function on the marked abnormal parameter set V, and training the classification capacity of the abnormal information monitoring network to obtain a class set of the marked abnormal parameter set; Clustering the class set of the marked abnormal parameter set by using a clustering algorithm, constructing an N multiplied by N prototype matrix by taking a clustering center as a prototype, and initializing the prototype matrix by using zero, wherein N is the number of classes in the class set of the marked abnormal parameter set; The method comprises the steps of carrying out joint training on a prototype matrix and an initial trained abnormal information monitoring network on a marking parameter set V to obtain a trained parameter monitoring module, wherein a loss function of the joint training is loss p = ω * loss 2 + loss d , wherein ω represents a weight parameter, loss 2 is used for describing characteristics of the prototype and the marked abnormal parameter to be learned, loss d is used for improving the classification capacity of the parameter monitoring module based on the classification loss of the distance; when the abnormal information monitoring network of the parameter monitoring module detects new abnormal information, initializing the weight parameters, specifically: And (3) initializing the weight parameters by setting A as a new abnormal information set detected by the parameter monitoring module, wherein the formula is as follows: , wherein ω n 、ω m is the weight columns of the nth class and the mth class, N is the number of classes stored in the buffer, and M is the number of abnormal classes.
2. The federal learning-based power data transmission anomaly detection method of claim 1, wherein the anomaly information monitoring network is constructed based on a neural network, wherein a parameter set of the anomaly information monitoring network is denoted as x= { X 1 , …, X i , …, X m+k }, an unlabeled parameter set is denoted as u= { X 1 , …, X m }, and a labeled anomaly parameter set is denoted as v= { X 1 , …, X k }, wherein X i is an ith parameter, m is a number of unlabeled parameters, and k is a number of labeled anomaly parameters; The cross entropy loss function is: , Wherein S is the batch size of the neural network, f i is the characteristic of the ith marked abnormal parameter extracted by the abnormal information monitoring network, and l i is the identification accuracy of the ith marked abnormal parameter.
3. The federal learning-based power data transmission anomaly detection method of claim 1, wherein the loss 2 is used for describing prototype and marked anomaly parameter features to be learned, expressed as: , Wherein f i is the characteristic of the ith marked abnormal parameter extracted by the abnormal information monitoring network, and p i is the prototype characteristic of the ith marked abnormal parameter; loss d represents a distance-based classification penalty to improve the classification capability of the parameter monitoring module, expressed as: , Wherein, l i is the identification accuracy of the ith marked abnormal parameter, D [ i ]: represents the ith row of the distance distribution matrix D; the distance distribution matrix D is used for calculating the Euclidean distance between the prototype and the marked abnormal parameter set, and the formula is as follows: , where i=1, 2,..s, j=1, 2,..n, epsilon is a non-zero constant.
4. The method for detecting abnormal power data transmission based on federal learning according to claim 1, wherein the step of identifying abnormal information to obtain the abnormal information in the transmission process and transmitting the abnormal information into the buffer zone is specifically as follows: Inputting the unlabeled parameter set into a trained parameter monitoring module, and obtaining a feature set, a category set, a label set and a matrix distribution matrix of a prototype of the unlabeled parameter set; Taking the class set of the marked abnormal parameter set as a sample set C of all correct classification; calculating average distance distribution according to the distance distribution matrix to obtain a threshold value, and detecting abnormal information in the unlabeled parameter set, wherein the formula is as follows: , μ i = ρ * η i , Wherein η i represents a distance threshold value of the ith class in the unlabeled parameter set classification set with respect to all correct classification sample sets C, μ i represents a confidence level of the ith class in the unlabeled parameter set classification set, M ij is a maximum confidence level value of the ith class in the jth correct classification sample set in the unlabeled parameter set classification set, C i is the number of the ith correct classification sample set C in the correct classification sample set C, ρ is an experience parameter, and when all confidence level values of one unlabeled parameter in the unlabeled parameter set are smaller than μ i , the unlabeled parameter is judged to be abnormal, and the class of the unlabeled parameter is regarded as an abnormal class; All abnormal classes are regarded as one class, and the abnormal unlabeled parameters are stored into a buffer area as a prototype through clustering, and are used for detecting abnormal information in the next incremental training, so that potential different classes are further searched, specifically: Assume the first The abnormal class set in the secondary training is U T , and K means are used for clustering U T into K prototypes { C 1 , C 2 , …, C k }; And taking each prototype as a category, automatically marking the corresponding unlabeled parameters, converting the unlabeled parameters into marked parameters, and adding the prototypes into a buffer area to help the detection of the next round of abnormal information.
5. The method for detecting abnormal power data transmission based on federal learning according to claim 4, wherein when the abnormal information monitoring network of the parameter monitoring module detects new abnormal information, determining whether the new abnormal information is recorded in the buffer, if not, updating the buffer with the new abnormal information, and filtering the abnormal information, specifically: If m new classes are found in the T-th detection and q new classes are found in the t+1th detection, the update mode of the buffer B is as follows: , , Where f u is a feature of the u-th class in the anomaly information, l u is a feature of the u-th class in the anomaly information, and W T and W T+1 represent weight vectors of the T-th and t+1th times, respectively.
6. The system is characterized by comprising a deployment module, an identification module, a filtering module and a transmission module; the deployment module is used for deploying a federal learning architecture with a parameter monitoring module at an edge gateway layer of the power system, wherein the parameter monitoring module comprises an abnormal information monitoring network and a buffer zone; The identification module is used for taking the transmission data as a parameter set of an abnormal information monitoring network in the parameter monitoring module when the edge gateway uploads the power data, identifying the abnormal information to obtain the abnormal information in the transmission process and transmitting the abnormal information into the buffer zone; The filtering module is used for confirming whether the new abnormal information is recorded in the buffer area or not when the abnormal information monitoring network of the parameter monitoring module detects the new abnormal information, and updating the buffer area by using the new abnormal information if the new abnormal information is not recorded, and filtering the abnormal information; The transmission module is used for updating transmission data by using knowledge transfer after filtering abnormal information, so as to ensure accurate data transmission; The training steps of the parameter monitoring module are as follows: Dividing a parameter set X of an abnormal information monitoring network into an unmarked parameter set U and a marked abnormal parameter set V, wherein the marked abnormal parameter set V comprises the characteristics and the labels of each marked parameter; Performing initial training on the abnormal information monitoring network by using a cross entropy loss function on the marked abnormal parameter set V, and training the classification capacity of the abnormal information monitoring network to obtain a class set of the marked abnormal parameter set; Clustering the class set of the marked abnormal parameter set by using a clustering algorithm, constructing an N multiplied by N prototype matrix by taking a clustering center as a prototype, and initializing the prototype matrix by using zero, wherein N is the number of classes in the class set of the marked abnormal parameter set; The method comprises the steps of carrying out joint training on a prototype matrix and an initial trained abnormal information monitoring network on a marking parameter set V to obtain a trained parameter monitoring module, wherein a loss function of the joint training is loss p = ω * loss 2 + loss d , wherein ω represents a weight parameter, loss 2 is used for describing characteristics of the prototype and the marked abnormal parameter to be learned, loss d is used for improving the classification capacity of the parameter monitoring module based on the classification loss of the distance; when the abnormal information monitoring network of the parameter monitoring module detects new abnormal information, initializing the weight parameters, specifically: And (3) initializing the weight parameters by setting A as a new abnormal information set detected by the parameter monitoring module, wherein the formula is as follows: , wherein ω n 、ω m is the weight columns of the nth class and the mth class, N is the number of classes stored in the buffer, and M is the number of abnormal classes.
7. An electronic device, the electronic device comprising: At least one processor, and A memory communicatively coupled to the at least one processor, wherein, The memory stores computer program instructions executable by the at least one processor to enable the at least one processor to perform the federal learning-based power data transmission anomaly detection method of any one of claims 1-5.
8. A computer-readable storage medium storing a program, wherein the program, when executed by a processor, implements the federal learning-based power data transmission abnormality detection method according to any one of claims 1 to 5.

Description

Power data transmission abnormality detection method, system and medium based on federal learning Technical Field The invention belongs to the technical field of data transmission abnormality detection, and particularly relates to a federal learning-based power data transmission abnormality detection method, a federal learning-based power data transmission abnormality detection system and a federal learning-based power data transmission abnormality detection medium. Background In order to comprehensively support the operation of a novel power system taking new energy as a main body, accelerate the digital construction of a power grid, adapt to the change of the operation characteristics of the power grid from the planned, centralized to open sharing and intelligent interaction directions, and the distributed architecture gradually takes the dominant role in the power grid. However, in the deep mining and analysis of the power data, there is a risk of revealing the privacy information of the machine learning participants, and therefore, how to ensure the privacy of the participants in the process of fully utilizing the power data for machine learning is a research subject of great attention at present. Federal learning architectures are widely used because of their superior privacy preserving capabilities, but in this architecture there remains a weak link that is vulnerable to attacks and intrusions. In the process of transmitting data at the edge, if the data is attacked abnormally, the convergence rate and performance of the federal model are poor. Disclosure of Invention The invention aims to overcome the defects and shortcomings of the prior art, and provides a federal learning-based power system data transmission abnormity method, a federal learning-based power system data transmission abnormity system and a federal learning-based power system medium. In order to achieve the above object, according to one aspect of the present invention, there is provided a federal learning-based power data transmission abnormality detection method, comprising the steps of: deploying a federal learning architecture with a parameter monitoring module at an edge gateway layer of the power system, wherein the parameter monitoring module comprises an abnormal information monitoring network and a buffer zone; When the edge gateway uploads the power data, the transmission data is used as a parameter set of an abnormal information monitoring network in the parameter monitoring module, abnormal information identification is carried out, abnormal information in the transmission process is obtained, and the abnormal information is transmitted into a buffer zone; When the abnormal information monitoring network of the parameter monitoring module detects new abnormal information, confirming whether the new abnormal information is recorded in the buffer area, if not, updating the buffer area by using the new abnormal information, and filtering the abnormal information; After filtering the abnormal information, updating the transmission data by using knowledge transfer to ensure the accuracy of data transmission. Preferably, the training steps of the parameter monitoring module are as follows: Dividing a parameter set X of an abnormal information monitoring network into an unmarked parameter set U and a marked abnormal parameter set V, wherein the marked abnormal parameter set V comprises the characteristics and the labels of each marked parameter; Performing initial training on the abnormal information monitoring network by using a cross entropy loss function on the marked abnormal parameter set V, and training the classification capacity of the abnormal information monitoring network to obtain a class set of the marked abnormal parameter set; Clustering the class set of the marked abnormal parameter set by using a clustering algorithm, constructing an N multiplied by N prototype matrix by taking a clustering center as a prototype, and initializing the prototype matrix by using zero, wherein N is the number of classes in the class set of the marked abnormal parameter set; And carrying out joint training on the prototype matrix and the initial trained abnormal information monitoring network on the marked parameter set V to obtain a trained parameter monitoring module. Preferably, the anomaly information monitoring network is constructed based on a neural network, wherein a parameter set of the anomaly information monitoring network is represented as X= { X 1,…,Xi,…,Xm+k }, an unlabeled parameter set is represented as U= { X 1,…,Xm }, a labeled anomaly parameter set is represented as V= { X 1,…,Xk }, wherein X i is an ith parameter, m is the number of unlabeled parameters, and k is the number of labeled anomaly parameters; The cross entropy loss function is: Wherein S is the batch size of the neural network, f i is the characteristic of the ith marked abnormal parameter extracted by the abnormal information monitoring network, and l