Power data anomaly detection system
By adopting a distributed anomaly detection architecture that collaborates between edge and cloud in the power system, using local and global models for abnormal detection of power data, the problems of low detection efficiency and high computing power requirements are solved, and efficient and accurate power is achieved Data abnormality detection.
Patent Information
- Application Number
- CN202410875858.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-07-02
- Publication Date
- 2024-11-12
AI Technical Summary
The existing power data abnormality detection scheme has low detection efficiency and high requirements for cloud computing power, which affects the accuracy and reliability of the power system management and decision-making.
Adopt a distributed exception detection system, deploy local exception detection models on the edge, and deploy global exception detection models on the cloud. Upload data to the edge through the device side for division and local detection. Then upload the results to the cloud for global detection, and use the global model The final abnormality detection result is generated.
It improves detection efficiency, reduces the computing power requirements for the cloud, enhances the accuracy and reliability of detection, and avoids the risk of data privacy leakage.
Smart Images

Figure CN118940083A_ABST
Abstract
Description
Technical Field
[0001] The embodiments of the present invention relate to the field of electric power technology, and in particular to a power data anomaly detection system. Background Art
[0002] The power system is an electric energy production and consumption system consisting of power plants, transmission and transformation lines, power distribution stations, and power consumption. In other words, the power system is a collection formed by integrating a large number of distributed power equipment. This collection can operate and control these distributed power equipment as a whole by integrating and coordinating them.
[0003] On this basis, the power system cloud-edge collaborative architecture refers to the collaborative cooperation between the cloud and the edge in the power system. The cloud is responsible for centralized power management and decision-making, while the edge is responsible for actual power production, transmission and consumption, and realizes remote monitoring and control through communication with the cloud.
[0004] However, in the cloud-edge collaboration of the power system, due to the diversity and dispersion of distributed power equipment, as well as the uncertainty of communication networks and sensor equipment, this will cause anomalies in the collection and transmission of power data, which will in turn affect the accuracy and reliability of power equipment management and decision-making. Therefore, the importance of anomaly detection of power data for the cloud-edge collaborative business processing of the power system cannot be ignored.
[0005] However, the current power data anomaly detection solution has problems such as low detection efficiency and high requirements for cloud computing power, which need to be solved urgently. Summary of the invention
[0006] The embodiment of the present invention provides a power data anomaly detection system, which solves the problems of low detection efficiency and high requirements on cloud computing power.
[0007] According to one aspect of the present invention, a power data anomaly detection system is provided, comprising: a device end, an edge end and a cloud end; wherein at least two edge nodes are deployed on the edge end, local anomaly detection models are respectively deployed on the at least two edge nodes, and a global anomaly detection model is deployed on the cloud end; wherein,
[0008] The device side is used to upload the collected target power data to the edge side;
[0009] The edge end is used to divide the received target power data and distribute the obtained divided data to at least two edge nodes;
[0010] Each of the at least two edge nodes is configured to perform anomaly detection on the received partitioned data using the deployed local anomaly detection model, and upload the obtained local anomaly detection result to the cloud;
[0011] The cloud end is used to use a global anomaly detection model to perform anomaly detection based on local anomaly detection results received and uploaded by at least two edge nodes, and obtain a target anomaly detection result of the target power data according to the obtained global anomaly detection result.
[0012] Optionally, the global anomaly detection model is obtained by nonlinearly combining the local anomaly detection models based on the target coefficients corresponding to the local anomaly detection models; in the cloud, it is specifically used for:
[0013] According to the target coefficients corresponding to the local anomaly detection models, anomaly detection is performed on the local anomaly detection results uploaded by at least two edge nodes, and a target anomaly detection result of the target power data is obtained according to the obtained global anomaly detection result.
[0014] On this basis, the cloud is also used for:
[0015] Obtain local anomaly detection models uploaded by at least two edge nodes;
[0016] Based on each local anomaly detection model and the original coefficients corresponding to each local anomaly detection model, a target formula is constructed;
[0017] With the goal of minimizing the value of the target formula, each original coefficient is solved to obtain the target coefficient corresponding to each local anomaly detection model.
[0018] Optionally, edge nodes are also used to:
[0019] Constructing a training data set, wherein the training data set includes multiple groups of training data, each group of training data includes data features of training power data and training anomaly labels of the training power data, wherein the training anomaly labels represent whether the training power data is abnormal;
[0020] The training data set is analyzed to obtain the functional relationship between the training anomaly labels and the data features, and based on the functional relationship, a local anomaly detection model deployed on the edge node is obtained.
[0021] On this basis, the edge node is optionally used to:
[0022] Using function mining based on gene expression programming, the training data set is analyzed to obtain the functional relationship between abnormal labels and data features;
[0023] Based on the functional relationship, a local anomaly detection model deployed on the edge node is obtained.
[0024] On this basis, the edge node is optionally used to:
[0025] Obtaining function mining factors, wherein the function mining factors may include at least one of a population with a preset size, a fitness function, a selection operation with a selection probability, a mutation operation with a mutation probability, a recombination operation with a recombination probability, and an insertion operation with a string insertion probability;
[0026] According to the function mining factors, the training data set is analyzed using function mining based on gene expression programming to obtain the functional relationship between abnormal labels and data features.
[0027] On this basis, the edge node is optionally used to:
[0028] The functional relationship f between the abnormal label and the data feature is obtained through the following formula:
[0029] f←G(.P op (N),F,S(P s ),M(P mu ),C(P c ),T(P t ) / |X);
[0030] Among them, G is the function mining based on gene expression programming, N is the preset size, and P is op (N) is the population, F is the fitness function, P s is the selection probability, S(P s ) is the selection operation, P mu is the mutation probability, M(P mu ) is a mutation operation, P c is the recombination probability, C(P c ) is the reorganization operation, P t is the insertion probability, T(P t ) is a string insertion operation, and X is a training data set.
[0031] Optionally, edge nodes are also used to:
[0032] Constructing a test data set, wherein the test data set includes multiple groups of test data, each group of test data includes data features of test power data and test anomaly labels of the test power data, wherein the test anomaly labels represent whether the test power data is abnormal;
[0033] The test data set is used to test each local anomaly detection model and the global anomaly detection model to obtain the anomaly detection effects of each local anomaly detection model and the global anomaly detection model.
[0034] On this basis, optionally, the anomaly detection effect is represented by anomaly detection accuracy and / or anomaly detection time consumption, and the anomaly detection accuracy is represented by at least one of accuracy rate, false detection rate and missed detection rate.
[0035] Optionally, the device side includes at least two data acquisition devices; wherein,
[0036] Each of the at least two data collection devices is used to upload the collected target power data to the edge end;
[0037] The edge node can be used to aggregate the received target power data, divide the aggregated data, and distribute the divided data to at least two edge nodes;
[0038] The cloud end is specifically used to use a global anomaly detection model to perform anomaly detection based on local anomaly detection results uploaded by at least two edge nodes, and obtain a target anomaly detection result of the aggregated result based on the obtained global anomaly detection result.
[0039] The power data anomaly detection system described in the embodiment of the present invention includes a device end, an edge end and a cloud end; wherein at least two edge nodes are deployed on the edge end, each edge node is respectively deployed with a local anomaly detection model, and a global anomaly detection model is deployed on the cloud end; through the device end, the collected target power data is uploaded to the edge end; through the edge end, the received target power data can be divided, and the obtained divided data can be distributed to each edge node, so that each edge node, using the local anomaly detection model deployed on itself, performs anomaly detection on the received divided data, and uploads the obtained local anomaly detection result to the cloud end, thereby realizing distributed anomaly detection of the target power data; then, through the cloud end, using the deployed global anomaly detection model, anomaly detection is performed based on the local anomaly detection results uploaded by each edge node, and then the target anomaly detection result of the target power data is obtained according to the obtained global anomaly detection result. The above technical scheme, by performing local anomaly detection on the edge end first, and then uploading each local anomaly detection result to the cloud end for global anomaly detection, this cloud-edge collaborative anomaly detection scheme improves the detection efficiency and reduces the computing power requirements on the cloud end.
[0040] It should be understood that the contents described in this section are not intended to identify the key or important features of the embodiments of the present invention, nor are they intended to limit the scope of the present invention. Other features of the present invention will become easily understood through the following description. BRIEF DESCRIPTION OF THE DRAWINGS
[0041] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.
[0042] Figure 1a is a schematic diagram of a framework of optional examples of various power data anomaly detection systems provided in embodiments of the present invention;
[0043] Figure 1b is a schematic diagram of an optional example architecture of various power data anomaly detection systems provided in an embodiment of the present invention;
[0044] Figure 2 is a structural block diagram of a power data anomaly detection system provided according to an embodiment of the present invention;
[0045] Figure 3 is a structural block diagram of another power data anomaly detection system provided according to an embodiment of the present invention;
[0046] Figure 4 is a structural block diagram of another power data anomaly detection system provided according to an embodiment of the present invention;
[0047] Figure 5a is a schematic diagram of another accuracy comparison example in a power data anomaly detection system provided according to an embodiment of the present invention;
[0048] Figure 5b is a schematic diagram of another example of false detection rate comparison in a power data anomaly detection system provided according to an embodiment of the present invention;
[0049] Figure 5c is a schematic diagram of another example of missed detection rate comparison in a power data anomaly detection system provided according to an embodiment of the present invention;
[0050] Figure 5d 4 is a schematic diagram of another example of average time consumption comparison in a power data anomaly detection system provided according to an embodiment of the present invention. DETAILED DESCRIPTION
[0051] In order to enable those skilled in the art to better understand the scheme of the present invention, the technical scheme in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work should fall within the scope of protection of the present invention.
[0052] It should be noted that the terms "first", "second", etc. in the specification and claims of the present invention and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchangeable where appropriate, so that the embodiments of the present invention described herein can be implemented in an order other than those illustrated or described herein. The situations of "target", "original", etc. are similar and will not be repeated here. In addition, the terms "including" and "having" and any variations thereof are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device comprising a series of steps or units is not necessarily limited to those steps or units clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.
[0053] Before introducing the embodiments of the present invention, an exemplary description of its application scenarios is given. Figure 1a The distributed power data anomaly detection system under the cloud-edge collaboration of the power system described in this embodiment may include a data source, a device end (i.e., a data acquisition layer), an edge end (i.e., an edge computing layer), and a cloud end (i.e., a cloud aggregation layer), wherein the edge end and the cloud end are the core contents. The edge end mainly performs distributed anomaly data detection (i.e., local data detection) on multiple edge nodes, and transmits the obtained multiple local anomaly detection models to the cloud end for fusion, thereby obtaining a global anomaly detection model.
[0054] For details, see Figure 1b , the device side mainly completes the collection of power data by various sensors under the power equipment. The edge side mainly receives the power data collected by each sensor, realizes the aggregation or preprocessing of the power data, and constructs the training data set and the test data set, especially the training data set and the test data set of abnormal power data. Finally, the generation training of the local anomaly detection model is performed, and the trained local anomaly detection model is uploaded to the cloud. The cloud obtains multiple local anomaly detection models, and generates a global anomaly detection model according to the global model fusion algorithm. At the same time, the model is tested based on the test data set, and the model output results are finally fed back to the corresponding power equipment for further processing, decision-making and adjustment.
[0055] Figure 2 This is a structural block diagram of a power data anomaly detection system provided by an embodiment of the present invention. This embodiment is applicable to the case where power data anomaly detection is performed under the cloud-edge collaboration of the power system, and is particularly applicable to the case where power data anomaly detection is performed first on the edge and then on the cloud.
[0056] See also Figure 2The power data anomaly detection system according to the embodiment of the present invention includes: a device end 10, an edge end 11 and a cloud end 12; wherein at least two edge nodes 111 are deployed on the edge end 11 (two are shown as an example in the figure), local anomaly detection models 1111 are respectively deployed on the at least two edge nodes 111, and a global anomaly detection model 121 is deployed on the cloud end 12; wherein,
[0057] The device end 10 is used to upload the collected target power data to the edge end 11;
[0058] The edge node 11 is used to divide the received target power data and distribute the obtained divided data to at least two edge nodes 111;
[0059] Each edge node 111 of the at least two edge nodes 111 is used to perform anomaly detection on the received partitioned data using the deployed local anomaly detection model 1111, and upload the obtained local anomaly detection result to the cloud 12;
[0060] The cloud 12 can be used to use the global anomaly detection model 121 to perform anomaly detection based on the local anomaly detection results uploaded by at least two edge nodes 111, and obtain the target anomaly detection result of the target power data according to the obtained global anomaly detection result.
[0061] Among them, at least two edge nodes 111 are deployed on the edge 11. In this way, the target power data uploaded by the device end 10 can be divided through the edge 11. For example, the target power data can be divided according to the number of edge nodes 111 to obtain at least two divided data. The number of divided data and the number of edge nodes 111 can be the same or different, especially can be the same, which is not specifically limited here. Then, through the edge 11, the at least two divided data obtained by the division are distributed to each edge node 111. For example, one or more divided data are allocated to each edge node 111. The number of divided data allocated to each edge node 111 can be the same or different, especially can be the same, which is not specifically limited here. Through the edge 11, distributed anomaly detection of target power data is achieved.
[0062] Furthermore, since each edge node 111 is respectively deployed with a local anomaly detection model 1111, each edge node 111 of at least two edge nodes 111 can use the local anomaly detection model 1111 deployed on itself to perform anomaly detection on the partitioned data assigned to itself, obtain the anomaly detection result of the partitioned data, and upload it to the cloud 12 for further processing. It can be understood that since the partitioned data is part of the target power data, the anomaly detection result can also be called the local anomaly detection result of the target power data.
[0063] Furthermore, through the cloud 12, after receiving the local anomaly detection results uploaded by each edge node 111, the global anomaly detection model 121 deployed on itself can be used to perform anomaly detection on these local anomaly detection results (that is, perform anomaly detection through comprehensive analysis) to obtain global anomaly detection results, and then the target anomaly detection results of the target power data can be obtained based on the global anomaly detection results to complete the anomaly detection of the target power data.
[0064] It should be noted that, considering the wide distribution of data acquisition devices on the device side, if the target power data collected by them are transmitted to the cloud through the network for anomaly detection, this will cause a certain network delay, which will lead to low detection efficiency, and the cloud will need to perform anomaly detection on massive target power data, which will lead to higher computing power requirements on the cloud.
[0065] Correspondingly, the above technical solution first performs distributed anomaly detection at the edge, and then only uploads the local anomaly detection results with limited data volume to the cloud through the network for anomaly detection, thereby reducing the probability of network congestion, improving detection efficiency, and reducing the high computing power requirements of the cloud. In addition, it can also avoid the occurrence of data privacy leakage caused by direct transmission of original data.
[0066] The power data anomaly detection system described in the embodiment of the present invention includes a device end, an edge end and a cloud end; wherein at least two edge nodes are deployed on the edge end, each edge node is respectively deployed with a local anomaly detection model, and a global anomaly detection model is deployed on the cloud end; through the device end, the collected target power data is uploaded to the edge end; through the edge end, the received target power data can be divided, and the obtained divided data can be distributed to each edge node, so that each edge node, using the local anomaly detection model deployed on itself, performs anomaly detection on the received divided data, and uploads the obtained local anomaly detection result to the cloud end, thereby realizing distributed anomaly detection of the target power data; then, through the cloud end, using the deployed global anomaly detection model, anomaly detection is performed based on the local anomaly detection results uploaded by each edge node, and then the target anomaly detection result of the target power data is obtained according to the obtained global anomaly detection result. The above technical scheme, by performing local anomaly detection on the edge end first, and then uploading each local anomaly detection result to the cloud end for global anomaly detection, this cloud-edge collaborative anomaly detection scheme improves the detection efficiency and reduces the computing power requirements on the cloud end.
[0067] An optional technical solution, the device side includes at least two data acquisition devices; wherein,
[0068] Each of the at least two data collection devices is used to upload the collected target power data to the edge end;
[0069] The edge node can be used to aggregate the received target power data, divide the aggregated data, and distribute the divided data to at least two edge nodes;
[0070] The cloud end is specifically used to use a global anomaly detection model to perform anomaly detection based on local anomaly detection results uploaded by at least two edge nodes, and obtain a target anomaly detection result of the aggregated result based on the obtained global anomaly detection result.
[0071] The above technical solution realizes the distributed collection of target power data by using at least two data collection devices, thereby ensuring the comprehensiveness and rapidity of the target power data collection. On this basis, the target power data collected by each data collection device is aggregated through the edge to obtain a whole, so that the data of the whole can be divided, thereby ensuring the uniformity of the data division.
[0072] Figure 3 It is a structural block diagram of another power data anomaly detection system provided by an embodiment of the present invention. This embodiment is optimized based on the above-mentioned technical solutions. In this embodiment, optionally, the global anomaly detection model is obtained by nonlinearly combining the local anomaly detection models based on the target coefficients corresponding to the local anomaly detection models; in the cloud, it is specifically used to: perform anomaly detection on the local anomaly detection results uploaded by at least two edge nodes according to the target coefficients corresponding to the local anomaly detection models, and obtain the target anomaly detection result of the target power data according to the obtained global anomaly detection result. Among them, the explanations of the terms that are the same or corresponding to the above-mentioned embodiments are not repeated here.
[0073] For details, see Figure 3 The power data anomaly detection system described in this embodiment includes: a device end 20, an edge end 21 and a cloud end 22; wherein at least two edge nodes 211 are deployed on the edge end 21, and local anomaly detection models 2111 are respectively deployed on the at least two edge nodes 211, and a global anomaly detection model 221 is deployed on the cloud end 22. The global anomaly detection model 221 can be obtained by nonlinearly combining the local anomaly detection models 2111 based on the target coefficients corresponding to the local anomaly detection models 2111; wherein,
[0074] The device end 20 is used to upload the collected target power data to the edge end 21;
[0075] The edge 21 is used to divide the received target power data and distribute the obtained divided data to at least two edge nodes 211;
[0076] Each edge node 211 of the at least two edge nodes 211 is used to perform anomaly detection on the received partitioned data using the deployed local anomaly detection model 2111, and upload the obtained local anomaly detection result to the cloud 22;
[0077] The cloud 22 can be used to perform anomaly detection on the local anomaly detection results uploaded by at least two edge nodes 211 according to the target coefficients corresponding to each local anomaly detection model 2111, and obtain the target anomaly detection result of the target power data according to the obtained global anomaly detection result.
[0078] Among them, each local anomaly detection model 2111 has a corresponding target coefficient, and the global anomaly detection model 221 can be obtained by nonlinearly combining each local anomaly detection model 2111 based on these target coefficients. For example, assuming that there are m edge nodes 211, the local anomaly detection model 2111 on the i-th edge node 211 among the m edge nodes 211 can be obtained by f i (X i ), i∈[1,m], the target coefficient corresponding to the local anomaly detection model 2111 on the i-th edge node 211 can be expressed by α i , i∈[1,m], then the global anomaly detection model 221 can be To express.
[0079] Further, as described above, since the global anomaly detection model 221 is obtained by nonlinearly combining the local anomaly detection models 2111 based on the target coefficients corresponding to the local anomaly detection models 2111, the cloud 22 performs anomaly detection on the received local anomaly detection results according to the target coefficients corresponding to the local anomaly detection models 2111 to obtain a global anomaly detection result.
[0080] The technical solution of the embodiment of the present invention constructs a global anomaly detection model through a nonlinear combination of local anomaly detection models. The global anomaly detection model can comprehensively consider the local anomaly detection results output by each local anomaly detection model, thereby ensuring the accuracy of the subsequent global anomaly detection results.
[0081] In other words, the above technical solution realizes the rapid training of local anomaly detection models on the edge and the collaborative fusion of global anomaly detection models on the cloud.
[0082] An optional technical solution, the cloud, is also used for:
[0083] Obtain local anomaly detection models uploaded by at least two edge nodes;
[0084] Based on each local anomaly detection model and the original coefficients corresponding to each local anomaly detection model, a target formula is constructed;
[0085] With the goal of minimizing the value of the target formula, each original coefficient is solved to obtain the target coefficient corresponding to each local anomaly detection model.
[0086] Among them, after generating the local anomaly detection model, each edge node on the edge can upload the local anomaly detection model to the cloud, so that the cloud can obtain the local anomaly detection model.
[0087] On this basis, in order to solve the target coefficients corresponding to each local anomaly detection model, a target formula can be constructed based on each local anomaly detection model and their corresponding original coefficients. The original coefficient is the target coefficient to be solved. Then, with the goal of minimizing the value of the target formula, each original coefficient is solved to obtain the corresponding target coefficient.
[0088] Exemplarily, based on the above examples and genetic algorithm, it is assumed that the target formula is expressed by the following formula: Among them, j is the number of the population, which needs to be inherited from the first population. On this basis, mS ize is the number of iterations required, i.e. the maximum value of j; y j is the fitness of the jth population. In practical applications, it can also be represented by f.
[0089] On this basis, assuming In order to solve the corresponding α i , can be made That is, for each α in the above formula i Find the partial derivatives to get m about α i , and set it equal to 0, construct m equations about α i By solving the m equations about α i The system of equations can be obtained as m α i , which is the target coefficient after solution.
[0090] The above technical solution, combined with the abnormal algorithm, realizes the effective solution of the target coefficient.
[0091] Before introducing the following embodiments, its application scenarios are first exemplified. For example, in practical applications, anomaly detection can be implemented based on human experience and clustering algorithms. However, human experience is based on prior knowledge to subjectively judge whether the target power data is abnormal, which has the problems of low efficiency and strong subjectivity; while the clustering algorithm is prone to fall into local optimality, and the random generation of cluster centers makes the anomaly detection accuracy low, the false detection rate and missed detection rate high, and the algorithm complexity is large.
[0092] To this end, the following embodiment combines genetic algorithms and genetic coding to propose an anomaly detection scheme, which can solve complex problems based on simple coding. The core of the scheme is to mine the relationship between the independent variable X and the dependent variable Y from a set of training data sets D, especially to mine the functional relationship between the abnormal label Y and the abnormal data feature X from a set of abnormal data sets, and finally judge whether there is an anomaly based on the function.
[0093] Figure 4 It is a structural block diagram of another power data anomaly detection system provided by an embodiment of the present invention. This embodiment is optimized based on the above-mentioned technical solutions. In this embodiment, optionally, the edge node can also be used to: construct a training data set, wherein the training data set includes multiple groups of training data, each group of training data includes data features of training power data and training anomaly labels of training power data, wherein the training anomaly labels represent whether the training power data is abnormal; analyze the training data set to obtain the functional relationship between the training anomaly labels and the data features, and based on the functional relationship, obtain a local anomaly detection model deployed on the edge node. Among them, the explanations of the terms that are the same as or corresponding to the above-mentioned embodiments are not repeated here.
[0094] For details, see Figure 4 The power data anomaly detection system described in this embodiment may include: a device end 30, an edge end 31 and a cloud end 32; wherein at least two edge nodes 311 are deployed on the edge end 31, and a global anomaly detection model 321 is deployed on the cloud end 32; wherein,
[0095] Each edge node 311 of the at least two edge nodes 311 is used to construct a training data set, wherein the training data set includes multiple groups of training data, each group of training data includes data features of training power data and training anomaly labels of training power data, wherein the training anomaly labels represent whether the training power data is abnormal;
[0096] Each edge node 311 of the at least two edge nodes 311 is further used to analyze the constructed training data set to obtain a functional relationship between the training anomaly label and the data feature, and based on the functional relationship, obtain a local anomaly detection model 3111 deployed on the edge node 311;
[0097] The device end 30 is used to upload the collected target power data to the edge end 31;
[0098] The edge end 31 is used to divide the received target power data and distribute the obtained divided data to at least two edge nodes 311;
[0099] Each edge node 311 of the at least two edge nodes 311 is further used to perform anomaly detection on the received partitioned data using the deployed local anomaly detection model 3111, and upload the obtained local anomaly detection result to the cloud 32;
[0100] The cloud 32 can be used to use the global anomaly detection model 321 to perform anomaly detection based on the local anomaly detection results uploaded by at least two edge nodes 311, and obtain the target anomaly detection result of the target power data according to the obtained global anomaly detection result.
[0101] Among them, a training data set is constructed through the edge node 311, and each group of training data in the training data set may include data features of training power data and training anomaly labels that characterize whether the training power data is abnormal. Further, the training data set can be analyzed to obtain a functional relationship between the training anomaly labels and the data features, and based on the functional relationship, a local anomaly detection model 3111 deployed on the edge node 311 can be obtained to complete the training of the local anomaly detection model 3111.
[0102] The technical solution of the embodiment of the present invention constructs a training data set and analyzes the training data set to mine the functional relationship between data features and abnormal labels, thereby achieving effective generation of a local anomaly detection model.
[0103] An optional technical solution, edge node, is also specifically used for:
[0104] Using function mining based on gene expression programming, the training data set is analyzed to obtain the functional relationship between abnormal labels and data features;
[0105] Based on the functional relationship, a local anomaly detection model deployed on the edge node is obtained.
[0106] The above technical solution utilizes function mining based on Gene Expression Programming (GEP) to achieve accurate mining of functional relationships.
[0107] On this basis, the edge node is optionally used to:
[0108] Obtaining function mining factors, wherein the function mining factors may include at least one of a population with a preset size, a fitness function, a selection operation with a selection probability, a mutation operation with a mutation probability, a recombination operation with a recombination probability, and an insertion operation with a string insertion probability;
[0109] According to the function mining factors, the training data set is analyzed using function mining based on gene expression programming to obtain the functional relationship between abnormal labels and data features.
[0110] In practical applications, optionally, the functional relationship f between the abnormal label and the data feature can be obtained by the following formula:
[0111] f←G(.P op (N),F,S(P s ),M(P mu ),C(P c ),T(P t ) / |X);
[0112] Where G is the function mining based on gene expression programming; N is the preset size, P is op (N) is a population of size N; F is the fitness function; P s is the selection probability, S(P s ) is the probability P s The selection operation of P mu is the mutation probability, M(P mu ) is the probability P mu The mutation operation of P c is the recombination probability, C(P c ) is the probability P c Recombination operation; P t is the insertion probability, T(P t ) is the probability P t The insertion operation of ; X is the training data set.
[0113] On this basis, in order to better understand the above technical solution, it is further necessary to explain that in practical applications, the training data set can also be called a sample data set. Assume that the sample data set X = {X1, X2, ..., X n |A}, where {X1,X2,…,X n |A} is the data feature of n sample data, and A represents the abnormal label of the sample data set. Then the sample data set X is called an abnormal data set, denoted by X A .
[0114] Let E=(E1,E2,…,E m ) represents a set of m edge nodes, represents the abnormal data set on the i,i∈[1,m]th edge node, then f i (X i ) is the local anomaly detection model of the i-th edge node. On this basis, let is the abnormal data set on the i-th edge node, A={1,0}, 1 is the abnormal label, and 0 is the normal label. If f i (X i ) is closer to 1, indicating that the power data on the i-th edge node is abnormal power data, otherwise it is regarded as normal power data.
[0115] On this basis, in order to better mine the local anomaly detection model on the edge node, optionally, F can be expressed by the following formula, which can also be called f itteess :
[0116]
[0117] Among them, S ize is the size of the training data set on each edge node; D is a constant; C (i,j) It is expressed as the value of the i-th individual for the j-th sample data; Indicates the true value of the jth sample. Specifically, in the genetic algorithm, the individuals of the next population need to be selected and inherited from the previous population. Then the jth sample expresses the number of samples in the previous population, and i expresses the individuals selected and inherited based on the sample j of the previous population. Therefore, C (i,j) is a relative value, and It is a definite value.
[0118] Based on any of the above technical solutions, optionally, the edge node is further used to:
[0119] Constructing a test data set, wherein the test data set includes multiple groups of test data, each group of test data includes data features of test power data and test anomaly labels of the test power data, wherein the test anomaly labels represent whether the test power data is abnormal;
[0120] The test data set is used to test each local anomaly detection model and the global anomaly detection model to obtain the anomaly detection effects of each local anomaly detection model and the global anomaly detection model.
[0121] The above technical solution can effectively verify the anomaly detection effect of the model using the test data set. In practical applications, optionally, the anomaly detection effect can be represented by anomaly detection accuracy and / or anomaly detection time consumption, and the anomaly detection accuracy is represented by at least one of accuracy rate, false detection rate and missed detection rate.
[0122] On this basis, in order to better understand the above technical solution and prove the effectiveness of the anomaly detection solution described above, the following is an exemplary description of it in combination with specific experiments.
[0123] (1) Simulation environment:
[0124] The entire experimental platform is equipped with 6 edge nodes and 1 cloud server deployed in the cloud, as well as 6 test data sets. The scales of the three test data sets A, B, and C in the six test data sets are 160, 410, and 166 rows respectively, the feature size is 12, and 1 point is sampled in 1 hour. 80% of the data is used as training data, and the remaining 20% is used as test data. Each load data set in the three test data sets D, E, and F in the six test data sets contains 365 data, and the feature size is 24.
[0125] It should be noted that Table 1 shows the original test data set. On this basis, in order to verify the anomaly detection effect of the model, abnormal power data is randomly injected into the test data set shown in Table 1, but the existing scale and feature size are not changed. At the same time, according to the injected abnormal power data, anomaly annotation is performed manually to form the processed test data set shown in Table 2.
[0126] Table 1 Original test dataset
[0127] name Scale / Line Feature size / piece D1 160 12 D2 410 12 D3 166 12 D4 365 24 D5 365 24 D6 365 24
[0128] Table 2 Processed test data set
[0129] name Scale / Line Feature size / piece Abnormal label D1 160 12 2 D2 410 12 2 D3 166 12 2 D4 365 24 2 D5 365 24 2 D6 365 24 2
[0130] (2) Performance indicators
[0131] In order to evaluate the anomaly detection effect of the model, it is assumed that Tn represents the number of normal power data identified as normal power data, Fn represents the number of abnormal power data identified as normal power data, Ta represents the number of abnormal power data identified as abnormal power data, and Fa represents the number of normal power data identified as abnormal power data.
[0132] On this basis, Represents the accuracy rate. The higher the Pr, the better the anomaly detection effect. It represents the false positive rate. The smaller Fr is, the better the anomaly detection effect is. represents the missed detection rate, M r The smaller it is, the better the anomaly detection effect is. is the average time-cost acceleration ratio, S pThe larger it is, the smaller the overall average computing overhead is, where S1 is the average computing time of a single edge node, and Mt is the average computing time of multiple edge nodes.
[0133] In order to verify the anomaly detection effect of the model and highlight its advantages, this experiment compares it (CSDAD-GEP) with the genetic programming (GP) algorithm and the decision tree (DT) algorithm. The comparison results are shown in the figure. Figure 5a , Figure 5b , Figure 5c and Figure 5d It should be noted that Figure 5a-5d The left side is the comparison result of the training data set and the right side is the comparison result of the test data set.
[0134] According to the above Figure 5a-5d It can be seen that whether it is a training data set or a test data set, the CSDAD-GEP used in this experiment is superior to the GP algorithm and the DT algorithm in terms of accuracy, false positive rate, missed positive rate and average time consumption, and has good application prospects.
[0135] The above specific implementations do not constitute a limitation on the protection scope of the present invention. It should be understood by those skilled in the art that various modifications, combinations, sub-combinations and substitutions can be made according to design requirements and other factors. Any modification, equivalent substitution and improvement made within the spirit and principle of the present invention should be included in the protection scope of the present invention.
Claims
1. A power data anomaly detection system, characterized in that: include: Device side, edge side and cloud side; wherein at least two edge nodes are deployed on the edge side, local anomaly detection models are deployed on the at least two edge nodes respectively, and a global anomaly detection model is deployed on the cloud side; wherein, The device end is used to upload the collected target power data to the edge end; The edge is used to divide the received target power data and distribute the obtained divided data to the at least two edge nodes; Each of the at least two edge nodes is configured to perform anomaly detection on the received partitioned data using the deployed local anomaly detection model, and upload the obtained local anomaly detection result to the cloud; The cloud end is used to utilize the global anomaly detection model to perform anomaly detection based on the local anomaly detection results received and uploaded by the at least two edge nodes, and obtain a target anomaly detection result of the target power data according to the obtained global anomaly detection result.
2. The system according to claim 1, characterized in that The global anomaly detection model is obtained by nonlinearly combining the local anomaly detection models based on the target coefficients corresponding to the local anomaly detection models respectively; the cloud end is specifically used for: According to the target coefficients corresponding to the local anomaly detection models, anomaly detection is performed on the local anomaly detection results respectively uploaded by the at least two edge nodes, and according to the obtained global anomaly detection result, a target anomaly detection result of the target power data is obtained.
3. The system according to claim 2, characterized in that The cloud is also used for: Obtaining local anomaly detection models uploaded by the at least two edge nodes respectively; Constructing a target formula based on each of the local anomaly detection models and the original coefficients corresponding to each of the local anomaly detection models; With the goal of minimizing the value of the target formula, each original coefficient is solved to obtain the target coefficient corresponding to each local anomaly detection model.
4. The system according to claim 1, characterized in that The edge node is further used for: Constructing a training data set, wherein the training data set includes multiple groups of training data, each group of the training data includes data features of training power data and training anomaly labels of the training power data, wherein the training anomaly labels indicate whether the training power data is abnormal; The training data set is analyzed to obtain a functional relationship between the training anomaly label and the data feature, and based on the functional relationship, a local anomaly detection model deployed on the edge node is obtained.
5. The system according to claim 4, characterized in that The edge node is further specifically used for: Analyzing the training data set by using function mining based on gene expression programming to obtain a functional relationship between the abnormal label and the data feature; Based on the functional relationship, a local anomaly detection model deployed on the edge node is obtained.
6. The system according to claim 5, characterized in that The edge node is further specifically used for: Obtaining function mining factors, wherein the function mining factors include at least one of a population with a preset size, a fitness function, a selection operation with a selection probability, a mutation operation with a mutation probability, a recombination operation with a recombination probability, and an insertion operation with a string insertion probability; According to the function mining factors, the training data set is analyzed using function mining based on gene expression programming to obtain a functional relationship between the abnormal label and the data feature.
7. The system according to claim 6, characterized in that The edge node is further specifically used for: The functional relationship f between the abnormal label and the data feature is obtained by the following formula: f←G(.P op (N),F,S(P s ),M(P mu ),C(P c ),T(P t ) / |X); where G is the function mining based on gene expression programming, N is the preset size, P op (N) is the population, F is the fitness function, P s is the selection probability, S(P s ) is the selection operation, P mu is the mutation probability, M(P mu ) is the mutation operation, P c is the recombination probability, C(P c ) is the recombination operation, P t is the insertion probability, T(P t ) is the string insertion operation, and X is the training data set.
8. The system according to claim 1, characterized in that The edge node is further used for: Constructing a test data set, wherein the test data set includes multiple groups of test data, each group of the test data includes data features of test power data and a test abnormality label of the test power data, wherein the test abnormality label indicates whether the test power data is abnormal; The test data set is used to test each of the local anomaly detection models and the global anomaly detection model to obtain anomaly detection effects of each of the local anomaly detection models and the global anomaly detection model.
9. The system according to claim 8, characterized in that The anomaly detection effect is represented by anomaly detection accuracy and / or anomaly detection time consumption, and the anomaly detection accuracy is represented by at least one of accuracy rate, false detection rate and missed detection rate.
10. The system according to claim 1, characterized in that The device end includes at least two data acquisition devices; wherein, Each of the at least two data acquisition devices is used to upload the collected target power data to the edge end; The edge is specifically used to aggregate the received target power data, divide the obtained aggregated result, and distribute the obtained divided data to the at least two edge nodes; The cloud is specifically used to use the global anomaly detection model to perform anomaly detection based on the local anomaly detection results received and uploaded by the at least two edge nodes, and obtain the target anomaly detection result of the converged result according to the obtained global anomaly detection result.
