Substation equipment anomaly detection method based on confidence element learning framework

Abstract

The application discloses a substation equipment anomaly detection method based on a confidence meta-learning framework, relates to the technical field of substation equipment anomaly detection, and solves the problems that the existing method is susceptible to abnormal sample interference and poor in adaptability. The steps are as follows: a visible light image training set of normal substation equipment is constructed, an anomaly detection model is trained to evaluate the equipment state, meanwhile, a reconstruction, density and equipment key area abnormal loss function is constructed, the sample is converted into an abnormal score; the abnormal score distribution algorithm is used to identify a threshold and assign a sample weight, the training loss is fused, the covariance matrix is built by training and verifying the loss to evaluate the model uncertainty, the adaptive regularization is introduced to reduce the uncertainty, and then the model parameters are updated by using the meta-learning; the model with the updated parameters is used to process visible light images of equipment to be detected and output results. The application quantifies two types of uncertainty in a label-free unsupervised framework, reduces abnormal sample interference, improves the adaptability of the model by combining the meta-learning, and improves the detection accuracy.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of substation equipment anomaly detection, in particular to a substation equipment anomaly detection method based on a confidence meta-learning framework. BACKGROUND

[0002] A substation is a place where the voltage is changed, and it is a place where the voltage and current of electric energy are transformed, centralized and distributed in the power system. In order to ensure the quality of electric energy and the safety of equipment, voltage adjustment, power flow control and protection of transmission and distribution lines and main electrical equipment need to be carried out in the substation.

[0003] In the operation process of the substation, in order to ensure the normal operation of the substation, the state of the equipment in the substation needs to be detected, so as to facilitate the timely discovery of abnormal phenomena of the substation equipment. In the current anomaly detection method, due to the existence of abnormal data samples, the model will be seriously affected by abnormal samples or samples close to the decision boundary, causing inaccurate data in the detection process. In terms of adaptability, the adaptability of the current substation equipment anomaly detection method to the environment and equipment is relatively single, which limits its use in complex equipment and environment of the substation. SUMMARY

[0004] In view of the defects of the prior art, the present application provides a substation equipment anomaly detection method based on a confidence meta-learning framework, which aims to solve the problems in the background art.

[0005] To achieve the above purpose, the present application provides the following technical scheme: a substation equipment anomaly detection method based on a confidence meta-learning framework, comprising the following steps:

[0006] Step S1: constructing a training image set, the samples in the training image set being visible light images of normal equipment in the substation; training an anomaly detection model by using the training image set to evaluate the running state of the equipment, and constructing a loss function based on reconstruction, density and equipment key area anomaly as a model training target; evaluating the running state of the equipment is to convert each sample in the training image set into an anomaly score;

[0007] Step S2: calculating an abnormal sample recognition threshold according to the distribution characteristics of the anomaly score , distributing the sample weight based on the recognition threshold, and integrating the distributed sample weight into the training loss of the anomaly detection model; constructing a covariance matrix based on the training loss and the validation loss of the anomaly detection model, evaluating the uncertainty of the model through the covariance matrix, and introducing adaptive regularization to reduce the uncertainty of the model; updating the model parameters based on meta-learning;

[0008] Step S3: Acquire visible light images of the substation equipment to be inspected, and use the anomaly detection model updated in step S2 to detect the visible light images of the substation equipment to be inspected, and output the anomaly detection results.

[0009] Furthermore, by using the training image set The process of training an anomaly detection model to evaluate the device's operational status involves: defining the core mapping function of the anomaly detection model, using this function to convert each sample in the training image set into anomaly scores; validating the anomaly detection model after training; and defining the core mapping function of the anomaly detection model as follows: , The parameters of the anomaly detection model are represented as follows; the anomaly score is expressed as... , Represents the training image set The Middle One sample.

[0010] Furthermore, the specific process of converting each sample in the training image set into anomaly scores using the core mapping function is as follows: The core mapping function is... Calculate each sample abnormal scores , Indicates sample At the center point of the feature space, Indicates the first Weighting coefficients for key regions; This represents the set of critical areas of the equipment; the first anomaly detection threshold is set as follows: , This represents the mean of the sample anomaly scores. Indicates the standard deviation of the sample outlier scores. This represents the first sensitivity coefficient; if the sample abnormal scores Then the sample Mark as abnormal, otherwise mark as normal.

[0011] Furthermore, based on the distribution characteristics of the anomaly scores, the anomaly sample identification threshold is calculated. The specific process of assigning sample weights based on the identification threshold and incorporating the assigned sample weights into the training loss of the anomaly detection model is as follows:

[0012] Calculate the anomaly sample identification threshold based on the distribution characteristics of the anomaly scores. , is represented as:

[0013] ;

[0014] In the formula, , These represent the outlier scores. The first and third quartiles; Indicates an adjustable parameter;

[0015] Based on recognition threshold Calculate each sample Confidence weight ;

[0016] The confidence weights of the samples are incorporated into the training loss, defining a data-weighted loss function. As the training objective of the model, the data-weighted loss function is expressed as:

[0017] ;

[0018] In the formula, The basic loss represents the construction, including the reconstruction-based loss function. Density-based loss function and the abnormal loss function of critical equipment areas ; Represents the training image set The total number of samples in the middle.

[0019] Furthermore, a covariance matrix is ​​constructed based on the training loss and validation loss of the anomaly detection model. The uncertainty of the model is then evaluated using the covariance matrix, and adaptive regularization is introduced to reduce model uncertainty. The specific process is as follows: Based on the training loss of the anomaly detection model... With verification loss Constructing the covariance matrix Calculate the covariance matrix determinant Define adaptive regularization terms. , Represents the basic regularization coefficient; This represents the uncertainty sensitivity coefficient.

[0020] Furthermore, the specific process of updating model parameters based on meta-learning is as follows:

[0021] training image set The task is randomly divided into several non-overlapping subtasks, each containing an independent training subset and a validation subset, simulating detection scenarios under different devices or environments.

[0022] For each subtask, use a data-weighted loss function. Calculate the training loss gradient And update the model parameters: , This represents the learning rate of the inner loop. This represents the temporary parameters of the model after the subtask has adapted.

[0023] Combined with model temporary parameters Sample confidence weights With adaptive regularization term Define the total meta-learning loss:

[0024] ;

[0025] In the formula, Represents temporary parameters of the model The corresponding covariance matrix; Indicates L2 regularization terms;

[0026] Based on meta-learning total loss Calculate the gradient of parameters Update model parameters , This represents the learning rate of the outer loop. This process is iterated until all subtasks are trained, resulting in the optimized model parameters. .

[0027] Furthermore, the specific process of step S3 is as follows:

[0028] Step S3.1: Construct the test sample set. The samples in the image are visible light images of the substation equipment to be tested.

[0029] Step S3.2: Define the core mapping function of the optimized anomaly detection model as follows: Using the core mapping function Calculate the test sample set Anomaly score for each sample , Represents the test sample set The Middle One sample;

[0030] Step S3.3: Statistical Sample Mean of outlier scores with standard deviation Combined with the second sensitivity coefficient Calculate the second anomaly detection threshold. ;

[0031] Step S3.4: Compare outlier scores Compared with the second anomaly detection threshold ,like Greater than the second anomaly detection threshold Then determine the sample The corresponding device is in an abnormal state, and the output result is 1; if Less than or equal to the second abnormality determination threshold value If yes, the device is determined to be in a normal state, and the output result is 0.

[0032] Further, the specific process of constructing a loss function based on reconstruction, density and abnormality of a key area of the device as a model training target is as follows:

[0033] Construction of a loss function based on reconstruction , which is expressed as:

[0034] ;

[0035] In the formula, represents the output of the abnormality detection model after reconstruction of the sample ;

[0036] Construction of a loss function based on density , which is expressed as:

[0037] ;

[0038] In the formula, represents the distribution probability of the sample learned by the abnormality detection model; represents the natural logarithm;

[0039] Construction of an abnormality loss function of a key area of the device , which is expressed as:

[0040] ;

[0041] In the formula, represents the image part corresponding to the key area in the sample ; represents the center point of the key area in the sample in the feature space;

[0042] The parameters of the abnormality detection model , , are iteratively optimized by an end-to-end gradient descent algorithm to minimize until convergence.

[0043] An electronic device includes a processor, a memory and a bus, the processor and the memory are connected through the bus, wherein the memory is used to store a set of program codes, and the processor is used to call the program codes stored in the memory to execute a substation device abnormality detection method based on a confidence cell learning framework.

[0044] ​​The application discloses a nonvolatile computer storage medium, and the computer storage medium stores computer executable instructions.

[0045] Compared with the prior art, the application has the following beneficial effects: the running state of the equipment is evaluated by training the anomaly detection model, and in the anomaly detection process of the substation equipment, the model uncertainty and data uncertainty are quantified in the label-free unsupervised learning framework, and then the weight used in the training stage is estimated by using the uncertainty, so that the target of the confidence meta-learning is to make the model more dependent on the normal samples and less dependent on the abnormal samples or the samples close to the decision boundary, so as to improve the accuracy of the data in the detection; meanwhile, the adaptability of the anomaly detection model in the substation equipment anomaly detection is improved by combining the meta-learning method, so that the model can quickly adapt to different equipment and environmental conditions, and the adaptability of the substation equipment anomaly detection method is improved. BRIEF DESCRIPTION OF DRAWINGS

[0046] Figure 1 The method flowchart of the application. DETAILED DESCRIPTION

[0047] As shown in Figure 1 , the application provides a technical scheme: a substation equipment anomaly detection method based on a confidence meta-learning framework, comprising the following steps:

[0048] Step S1: constructing a training image set, the samples in the training image set are visible light images of normal equipment in a substation (the equipment in the substation includes transformers, switch cabinets, busbars and the like); the running state of the equipment is evaluated by training an anomaly detection model by using the training image set, and a loss function based on reconstruction, density and equipment key area anomaly is constructed as a model training target; the running state of the equipment is evaluated by converting each sample in the training image set into an anomaly score.

[0049] The training image set is used to train the anomaly detection model to evaluate the running state of the equipment. The process of training the anomaly detection model to evaluate the running state of the equipment is as follows: a core mapping function of the anomaly detection model is defined, each sample in the training image set is converted into an anomaly score by using the core mapping function; and the anomaly detection model is verified after being trained.

[0050] The core mapping function of the anomaly detection (AD) model is , , wherein f represents a parameter of the anomaly detection model; the anomaly score is represented as (the higher the anomaly score, the higher the possibility that the sample belongs to an abnormal state), , wherein the training image set The first sample.

[0051] Wherein, the specific process of converting each sample in the training image set into an anomaly score by the core mapping function is:

[0052] 1. Calculate the anomaly score of each sample through the core mapping function , , wherein, the center point of the sample in the feature space is represented by , and the weight coefficient of the th key region is represented by .

[0053] 2. Set the first anomaly determination threshold as , wherein, the mean of the sample anomaly score is represented by , the standard deviation of the sample anomaly score is represented by , and the first sensitivity coefficient is represented by

[0054] 3. If the anomaly score of the sample is , then the sample is marked as abnormal, otherwise it is marked as normal.

[0055] Wherein, the specific process of constructing a loss function based on reconstruction, density, and device key region anomaly as the model training target is:

[0056] 1. Construct a loss function based on reconstruction (suitable for reconstruction-based anomaly detection models), represented as:

[0057] .

[0058] In the formula, represents the output of the anomaly detection model after reconstructing the sample ; represents the total number of samples in the training image set . By minimizing this loss, the model learns the reconstruction rule of normal samples.

[0059] 2. Construct a loss function based on density (suitable for density estimation-based anomaly detection models), represented as:

[0060] .

[0061] In the formula, Anomaly score Distribution probability Natural logarithm; by minimizing the negative log-likelihood loss, the model fits the distribution characteristics of normal samples.

[0062] 3. Constructing the anomaly loss function of the key area of the device (Aiming at monitoring the core components of the device), it is represented as:

[0063] ;

[0064] In the formula, Anomaly score Corresponding key area Image part in the sample Anomaly score Key area Center point in the feature space; through this loss, the model focuses on the consistency of the features of each key area of the device.

[0065] 4. Minimize the above loss function as the goal, and iteratively optimize the parameters of the anomaly detection model through the end-to-end gradient descent algorithm Until the loss function converges, the trained anomaly detection model is obtained, which can initially realize the evaluation ability of the running state of the device.

[0066] Step S2: According to the distribution characteristics of the anomaly score, calculate the anomaly sample recognition threshold, assign the sample weight based on the recognition threshold, and integrate the assigned sample weight into the training loss of the anomaly detection model; Based on the training loss and the validation loss of the anomaly detection model, a covariance matrix is constructed, the uncertainty of the model is evaluated through the covariance matrix, and adaptive regularization is introduced to reduce the uncertainty of the model; Update the model parameters based on meta-learning.

[0067] Among them, according to the distribution characteristics of the anomaly score, the anomaly sample recognition threshold Based on the recognition threshold, the sample weight is assigned, and the specific process of integrating the assigned sample weight into the training loss of the anomaly detection model is as follows:

[0068] 1. According to the distribution characteristics of the anomaly score, calculate the anomaly sample recognition threshold , which is represented as:

[0069] ;

[0070] In the formula, , respectively represent the first quartile (25th percentile) and the third quartile (75th percentile) of the anomaly score ​denote the adjustable parameters, used to balance the influence of normal and abnormal samples on the model.

[0071] 2、based on the identification threshold calculate the confidence weight of each sample , when the abnormal score of the sample is less than or equal to the identification threshold , the weight tends to 1, and the model relies on the sample during training; when the abnormal score of the sample is greater than the identification threshold , the weight is less than 1, reducing its interference with model training.

[0072] 3、Integrate the confidence weight of the sample into the training loss to define the data weighted loss function as the model training target; the data weighted loss function is represented as:

[0073] ;

[0074] In the formula, denotes the basic loss constructed, including , and .

[0075] wherein, based on the training loss and validation loss of the anomaly detection model, a covariance matrix is constructed, the uncertainty of the model is evaluated through the covariance matrix, and the specific process of introducing adaptive regularization to reduce the uncertainty of the model is:

[0076] 1、based on the training loss and validation loss of the anomaly detection model, a covariance matrix is constructed:

[0077] ;

[0078] In the formula, denotes the covariance; the matrix reflects the correlation of loss fluctuations.

[0079] 2、calculate the determinant of the covariance matrix , the higher the determinant value, the stronger the variability of the training loss and the validation loss, the worse the fitting stability of the model to the sample (high uncertainty); the lower the determinant value, the stronger the robustness of the model.

[0080] 3、to reduce the uncertainty of the model, define an adaptive regularization term ,​​​ denotes the base regularization coefficient; denotes the uncertainty sensitivity coefficient; when the model uncertainty is high , the regularization term increases, constraining the model parameters from overfitting and improving the generalization ability.

[0081] wherein the specific process of updating the model parameters based on meta-learning is as follows:

[0082] 1. The training image set is randomly divided into several disjoint sub-tasks, each of which contains an independent training subset and a validation subset, simulating detection scenarios under different devices or environments.

[0083] 2. For each sub-task, the data weighted loss function is used to calculate the training loss gradient , and the model parameters are updated as follows: , denotes the inner loop meta-learning rate, denotes the temporary model parameters of the sub-task after adaptation, which are used to realize the rapid adaptation of the model to a single task scenario.

[0084] 3. Combined with the temporary model parameters , the sample confidence weight , and the adaptive regularization term , the meta-learning total loss is defined as:

[0085] ;

[0086] wherein denotes the covariance matrix corresponding to the temporary model parameters ; denotes the L2 regularization term, which is used to constrain the parameter size.

[0087] 4. Based on the meta-learning total loss , the parameter gradient is calculated, and the model parameters are updated as follows: , denotes the outer loop meta-learning rate, and the process is iterated until all sub-tasks are trained, obtaining the optimized model parameters with cross-device and cross-environment adaptation ability.

[0088] Step S3: Collecting the visible light image of the substation equipment to be detected, and detecting the visible light image of the substation equipment to be detected by the anomaly detection model whose model parameters are updated in step S2, and outputting the anomaly detection result.

[0089] Step S3.1: Constructing a test sample set, wherein the test sample set The sample in the step S3.1 is a visible light image of the substation equipment to be detected.

[0090] Step S3.2: defining the core mapping function of the optimized anomaly detection model as , calculating the anomaly score of each sample in the test sample set by using the core mapping function . , represents the i-th sample in the test sample set .

[0091] Step S3.3: calculating the second anomaly determination threshold by combining the mean value and the standard deviation of the anomaly scores of the samples with the second sensitivity coefficient (pre-set value, such as 1.5-2.5).

[0092] Step S3.4: comparing the anomaly score with the second anomaly determination threshold , if the anomaly score is greater than the second anomaly determination threshold , it is determined that the equipment corresponding to the sample is in an abnormal state, and the output result is 1; if the anomaly score is less than or equal to the second anomaly determination threshold , it is determined that the equipment is in a normal state, and the output result is 0.

[0093] An electronic device, comprising a processor, a memory and a bus, wherein the processor and the memory are connected through the bus, wherein the memory is used to store a set of program codes, and the processor is used to call the program codes stored in the memory to execute the substation equipment anomaly detection method based on the confidence cell learning framework.

[0094] A non-volatile computer storage medium, the computer storage medium stores computer executable instructions, characterized in that the computer executable instructions execute the substation equipment anomaly detection method based on the confidence cell learning framework.

[0095] Although the embodiments of the present application have been shown and described, it can be understood by those skilled in the art that various changes, modifications, replacements and variations can be made to the embodiments without departing from the principles and spirits of the present application, and the scope of the present application is defined by the appended claims and their equivalents.​

Claims

1. A method for detecting anomalies in substation equipment based on a confidence element learning framework, characterized in that, Includes the following steps: Step S1: Construct a training image set, where the samples in the training image set are visible light images of normal equipment in the substation; use the training image set to train an anomaly detection model to evaluate the operating status of the equipment, and at the same time construct a loss function based on reconstruction, density, and anomalies in key areas of the equipment as the training objective of the model; evaluate the operating status of the equipment by converting each sample in the training image set into an anomaly score; Step S2: Calculate the abnormal sample identification threshold based on the distribution characteristics of the abnormal scores, allocate sample weights based on the identification threshold, and incorporate the allocated sample weights into the training loss of the abnormal detection model. The covariance matrix is ​​constructed based on the training loss and validation loss of the anomaly detection model. The uncertainty of the model is evaluated through the covariance matrix, and adaptive regularization is introduced to reduce the uncertainty of the model. The model parameters are updated based on meta-learning. Step S3: Acquire visible light images of the substation equipment to be inspected, and use the anomaly detection model updated in step S2 to detect the visible light images of the substation equipment to be inspected, and output the anomaly detection results. By using the training image set The process of training an anomaly detection model to evaluate the device's operational status involves: defining the core mapping function of the anomaly detection model, using this function to convert each sample in the training image set into anomaly scores; validating the anomaly detection model after training; and defining the core mapping function of the anomaly detection model as follows: , The parameters of the anomaly detection model are represented as follows; the anomaly score is expressed as... , Represents the training image set The Middle One sample; The specific process of converting each sample in the training image set into anomaly scores using the core mapping function is as follows: The core mapping function is... Calculate each sample abnormal scores , Indicates sample At the center point of the feature space, Indicates the first Weighting coefficients for key regions; This represents the set of critical areas of the equipment; the first anomaly detection threshold is set as follows: , This represents the mean of the sample anomaly scores. Indicates the standard deviation of the sample outlier scores. This represents the first sensitivity coefficient; if the sample abnormal scores Then the sample Mark as abnormal, otherwise mark as normal; Calculate the anomaly sample identification threshold based on the distribution characteristics of the anomaly scores. The specific process of assigning sample weights based on the identification threshold and incorporating the assigned sample weights into the training loss of the anomaly detection model is as follows: Calculate the anomaly sample identification threshold based on the distribution characteristics of the anomaly scores. , represented as: ； In the formula, , These represent the outlier scores. The first and third quartiles; Indicates an adjustable parameter; Based on recognition threshold Calculate each sample Confidence weight ; The confidence weights of the samples are incorporated into the training loss, defining a data-weighted loss function. As the training objective of the model, the data-weighted loss function is expressed as: ； In the formula, The basic loss represents the construction, including the reconstruction-based loss function. Density-based loss function and the abnormal loss function of critical equipment areas ; Represents the training image set The total number of samples in the middle.

2. The substation equipment anomaly detection method based on the confidence element learning framework according to claim 1, characterized in that: The specific process of constructing a covariance matrix based on the training loss and validation loss of the anomaly detection model, evaluating the model's uncertainty using the covariance matrix, and introducing adaptive regularization to reduce model uncertainty is as follows: Based on the training loss of the anomaly detection model... With verification loss Constructing the covariance matrix Calculate the covariance matrix determinant Define adaptive regularization terms. , Represents the basic regularization coefficient; This represents the uncertainty sensitivity coefficient.

3. The substation equipment anomaly detection method based on the confidence element learning framework according to claim 2, characterized in that: The specific process of updating model parameters based on meta-learning is as follows: training image set The task is randomly divided into several non-overlapping subtasks, each containing an independent training subset and a validation subset to simulate detection scenarios under different devices or environments. For each subtask, use a data-weighted loss function. Calculate the training loss gradient And update the model parameters: , This represents the learning rate of the inner loop. This represents the temporary parameters of the model after the subtask has adapted. Combined with model temporary parameters Sample confidence weights With adaptive regularization term Define the total meta-learning loss: ； In the formula, Represents temporary parameters of the model The corresponding covariance matrix; Indicates L2 regularization terms; Based on meta-learning total loss Calculate the gradient of parameters Update model parameters , This represents the learning rate of the outer loop. This process is iterated until all subtasks are trained, resulting in the optimized model parameters. .

4. The substation equipment anomaly detection method based on the confidence element learning framework according to claim 3, characterized in that: The specific process of step S3 is as follows: Step S3.1: Construct the test sample set. The samples in the image are visible light images of the substation equipment to be tested. Step S3.2: Define the core mapping function of the optimized anomaly detection model as follows: Using the core mapping function Calculate the test sample set Anomaly score for each sample , Represents the test sample set The Middle One sample; Step S3.3: Statistical Sample Mean of outlier scores with standard deviation Combined with the second sensitivity coefficient Calculate the second anomaly detection threshold. ; Step S3.4: Compare outlier scores Compared with the second anomaly detection threshold ,like Greater than the second anomaly detection threshold Then determine the sample The corresponding device is in an abnormal state, and the output result is 1; if Less than or equal to the second anomaly detection threshold If the condition is met, the device is considered to be in normal condition, and the output result is 0.

5. The substation equipment anomaly detection method based on the confidence element learning framework according to claim 4, characterized in that: The specific process of constructing a loss function based on reconstruction, density, and anomalies in key areas of the device as the model training objective is as follows: Constructing a loss function based on reconstruction , represented as: ； In the formula, This indicates that the anomaly detection model is effective for samples. The reconstructed output; Constructing a density-based loss function , represented as: ； In the formula, This represents the samples learned by the anomaly detection model. Probability distribution; Represents the natural logarithm; Constructing anomaly loss functions for critical areas of the equipment , represented as: ； In the formula, Indicates sample Corresponding key areas The image portion; Indicates sample Central key areas The center point in the feature space; To minimize , , With the goal of optimizing the parameters of an anomaly detection model iteratively using an end-to-end gradient descent algorithm, we can achieve the desired results. Until it converges.

6. An electronic device, characterized in that, The device includes a processor, a memory, and a bus, wherein the processor and the memory are connected via the bus, the memory is used to store a set of program code, and the processor is used to call the program code stored in the memory to execute the substation equipment anomaly detection method based on the confidence element learning framework as described in any one of claims 1-5.

7. A non-volatile computer storage medium storing computer-executable instructions, characterized in that, The computer can execute instructions to perform the substation equipment anomaly detection method based on the confidence element learning framework as described in any one of claims 1-5.

Images