Few-shot substation equipment state adaptive inspection system based on meta-learning

Abstract

The application relates to the technical field of intelligent inspection of substations, in particular to a few-sample substation equipment state adaptive inspection system based on meta-learning, which comprises the following modules: a state acquisition module that acquires the current feature vector and environmental parameter data of a target node; a drift detection module that compares environmental parameters to judge data drift and dynamically adjusts a confidence threshold; a risk assessment module that inputs feature data into a meta-learning model to output an initial risk probability and generates an effective risk probability based on threshold filtering; a blind area measurement module that acquires unobserved nodes and calculates system state blind area entropy; a scheduling decision module that compares blind area entropy with a threshold to generate a blind area entropy reduction bottom instruction or an adaptive inspection distribution instruction; and a strategy updating module that extracts actual inspection results and feeds them back to the model for parameter updating; the application solves the scheduling dilemma when resources are limited and improves the adaptive ability of the system under different environmental disturbances.

Description

Technical Field

[0001] This invention relates to the field of intelligent substation inspection technology, specifically to a few-sample substation equipment status adaptive inspection system based on meta-learning. Background Technology

[0002] With the continuous expansion of substation scale and the continuous improvement of intelligence level, equipment condition inspection has become a key link in ensuring the safe operation of the power grid. Under the constraints of limited inspection resources and complex and ever-changing environment, how to achieve high efficiency and accuracy of equipment monitoring has become a core problem that the current power grid operation and maintenance urgently needs to solve.

[0003] Therefore, enhancing the adaptability and anti-interference capabilities of inspection systems while ensuring high power grid reliability has become a crucial research direction in smart grid construction. Traditional substation equipment inspection and condition assessment mainly rely on the following methods: comprehensive inspection based on fixed cycles, standard deep learning fault diagnosis relying on a large number of samples for training, and risk-priority scheduling strategies based on static thresholds. These methods all have certain drawbacks. For example, fixed-cycle inspections cannot dynamically allocate inspection resources based on the real-time status of equipment, which not only easily leads to resource waste but also makes it difficult to prevent sudden faults during the inspection interval. Standard deep learning diagnostic methods lack the ability to generalize with few samples when facing rare new faults in substations, and when processing multimodal features, the average weight allocation results in insufficient extraction of feature dimensions sensitive to fault identification. Static threshold scheduling has insufficient adaptability to data drift caused by sudden environmental changes, and is prone to generating a large number of false alarms due to environmental noise, thus exhausting inspection resources. Risk-priority scheduling methods, due to the lack of measurement of potential risk accumulation in unobserved areas, are prone to falling into the dilemma of local observation when inspection resources are limited, making it difficult to prevent sudden faults caused by long-term blind spots. Summary of the Invention

[0004] The purpose of this invention is to provide a few-sample substation equipment status adaptive inspection system based on meta-learning, which solves the following technical problems: it can accurately identify new faults and suppress environmental noise interference under few-sample conditions, and can dynamically balance risk investigation and blind zone coverage through state blind zone entropy when resources are limited.

[0005] The objective of this invention can be achieved through the following technical solutions:

[0006] A meta-learning-based adaptive inspection system for substation equipment status using few samples includes:

[0007] The status acquisition module is used to acquire a preset substation equipment network topology, take the substation equipment in the preset substation equipment network topology as target nodes in the target node set, and acquire the current feature vector and current environmental parameter data of each target node in the target node set.

[0008] The drift detection module is used to perform baseline comparison on the current environmental parameter data, determine whether data drift has occurred, and dynamically adjust the confidence threshold corresponding to the preset meta-learning risk assessment model based on the results in the target node set.

[0009] The risk assessment module is used to input the current feature vector into the preset meta-learning risk assessment model, output the initial risk probability corresponding to each target node, and filter the initial risk probability based on the confidence threshold to generate an effective risk probability.

[0010] The blind zone measurement module is used to obtain the set of unobserved nodes and calculate the current state blind zone entropy of the system based on the historical unobserved duration and basic risk weight corresponding to the set of unobserved nodes.

[0011] The scheduling decision module is used to compare the state blind zone entropy with a preset entropy threshold. If the state blind zone entropy is greater than or equal to the preset entropy threshold, the effective risk probability is ignored and a fallback inspection instruction aimed at reducing the state blind zone entropy is generated. If the state blind zone entropy is less than the preset entropy threshold, an adaptive inspection allocation instruction is generated based on the effective risk probability and a preset resource coverage limit.

[0012] The strategy update module is used to extract the actual inspection result data returned after executing the fallback inspection instruction or the adaptive inspection allocation instruction, and feed the actual inspection result data back to the preset meta-learning risk assessment model for parameter update.

[0013] Optionally, methods for obtaining the current feature vector and current environment parameter data of each target node in the target node set include:

[0014] Infrared feature data and partial discharge feature data corresponding to each target node are obtained through preset detection resources and combined into the current feature vector;

[0015] Micro-meteorological data, including temperature, humidity, wind speed, and light intensity, are acquired through preset meteorological sensors and used as the current environmental parameter data.

[0016] Optionally, methods for determining whether data drift has occurred and dynamically adjusting the confidence threshold corresponding to the preset meta-learning risk assessment model based on the determination result include:

[0017] Obtain preset environmental baseline data;

[0018] Calculate the absolute value of the difference between the current environmental parameter data and the preset environmental baseline data, and use it as the environmental deviation value;

[0019] The environmental deviation value is compared with a preset drift determination threshold;

[0020] If the environmental deviation value is greater than or equal to the preset drift determination threshold, then data drift is determined to have occurred, and the preset initial confidence threshold is added to the preset penalty step size to obtain the confidence threshold.

[0021] If the environmental deviation value is less than the preset drift determination threshold, it is determined that no data drift has occurred, and the preset initial confidence threshold is used as the confidence threshold.

[0022] Optionally, methods for generating effective risk probabilities include:

[0023] Obtain a preset few-sample abnormal pattern library, wherein the preset few-sample abnormal pattern library includes one or more abnormal pattern features;

[0024] The feature distance between the current feature vector and the features of each abnormal pattern in the preset few-sample abnormal pattern library is calculated using the preset meta-learning risk assessment model.

[0025] The feature distance is mapped to a value that is negatively correlated with the feature distance through a preset normalization function, and the value is used as the initial risk probability corresponding to each target node;

[0026] Compare the initial risk probability with the confidence threshold;

[0027] If the initial risk probability is greater than or equal to the confidence threshold, then the initial risk probability is taken as the effective risk probability.

[0028] If the initial risk probability is less than the confidence threshold, then the effective risk probability is set to zero.

[0029] Optionally, methods for calculating the current state blind zone entropy of the system include:

[0030] The nodes in the target node set that have not been pre-allocated detection resources are extracted as the unobserved node set;

[0031] Obtain the historical unobserved duration and the basic risk weight for each unobserved node in the set of unobserved nodes;

[0032] The product of the historical unobserved duration and the basic risk weight is used to obtain the single-point blind zone risk value corresponding to each unobserved node.

[0033] The single-point blind zone risk values ​​corresponding to all unobserved nodes in the set of unobserved nodes are summed to obtain the state blind zone entropy.

[0034] Optionally, a method for generating fallback inspection instructions aimed at reducing the state blind zone entropy includes:

[0035] Based on the magnitude of the single-point blind zone risk value, the unobserved nodes in the unobserved node set are sorted in descending order to generate a high blind zone node sequence;

[0036] Obtain the preset resource coverage limit;

[0037] Extract unobserved nodes from the beginning of the high blind zone node sequence in sequence, with a number equal to the preset resource coverage limit. If the total number of unobserved nodes in the high blind zone node sequence is less than the preset resource coverage limit, then extract all the unobserved nodes to construct a fallback target set.

[0038] The bottom-line inspection instruction is generated based on the bottom-line target set.

[0039] Optionally, methods for generating adaptive inspection assignment instructions include:

[0040] Extract all the effective risk probabilities from the target node set;

[0041] The target nodes with an effective risk probability greater than zero are extracted and sorted in descending order according to the magnitude of the effective risk probability to generate a high-risk node sequence.

[0042] Extract nodes sequentially from the beginning of the high-risk node sequence until the number of extracted nodes equals the preset resource coverage limit or all nodes in the high-risk node sequence are extracted, and construct the primary target set.

[0043] Calculate the difference between the preset resource coverage limit and the number of nodes in the primary target set, and use it as the resource difference;

[0044] Compare the resource difference with zero;

[0045] If the resource difference is greater than zero, the nodes that already exist in the primary target set in the high blind zone node sequence are removed and a new candidate blind zone node sequence is generated. The number of nodes equal to the resource difference is extracted from the candidate blind zone node sequence and added to the primary target set. The adaptive inspection allocation instruction is generated based on the supplemented primary target set.

[0046] If the resource difference is equal to zero, the adaptive inspection allocation instruction is generated directly based on the primary target set.

[0047] Optionally, the method for feeding back the actual inspection results data to the preset meta-learning risk assessment model for parameter updates includes:

[0048] Extract the actual fault labels from the actual inspection result data;

[0049] Calculate the difference between the actual fault label and the corresponding effective risk probability as the prediction error value;

[0050] Based on the prediction error value, the gradient descent algorithm is used to calculate the gradient of the network parameters of the preset meta-learning risk assessment model.

[0051] Based on the network parameter gradient, the network parameters of the preset meta-learning risk assessment model are iteratively updated to complete the parameter update.

[0052] The beneficial effects of this invention are:

[0053] 1. This invention constructs a pre-defined meta-learning risk assessment model and adopts a weighted Euclidean distance-based metric, which can adaptively amplify the feature dimensions sensitive to fault identification, effectively overcoming the problem of weight averaging in standard prototype networks when processing heterogeneous multimodal features. This mechanism enables the system to accurately identify new fault modes in substations even with only a small number of samples, significantly improving the generalization ability and accuracy of equipment condition diagnosis.

[0054] 2. This invention introduces a drift detection module to perform baseline comparison of environmental parameters. When data drift is detected, the confidence threshold is dynamically adjusted to actively raise the alarm threshold and suppress false alarms caused by environmental noise. This dynamic gating mechanism solves the problem that static threshold scheduling cannot adapt to sudden environmental changes, avoids the exhaustion of inspection resources by false alarms due to meteorological interference, and effectively ensures the system robustness and resource utilization in harsh environments.

[0055] 3. This invention innovatively introduces state blind zone entropy as the core constraint of scheduling, quantifying the total amount of potential risks accumulated in unobserved areas due to resource constraints; by comparing state blind zone entropy with the entropy critical threshold, the system can flexibly switch between precise investigation mode and basic support mode. When the blind zone risk is too high, it automatically degrades to a fallback inspection strategy, breaking the dilemma of traditional risk-based scheduling being prone to local observation and preventing sudden failures caused by long-term blind zones.

[0056] 4. The present invention sets up a strategy update module to construct a closed-loop feedback mechanism for prediction and execution, which can extract real fault labels from actual inspection results data, calculate prediction errors, and use gradient descent algorithm to update the parameters of the meta-learning risk assessment model; this mechanism enables the system to have online evolution capabilities, and can use the actual inspection results to correct the model parameters in situ, continuously improving the model's adaptability to the aging evolution of substation equipment and new environmental characteristics.

[0057] 5. This invention integrates infrared feature data and partial discharge feature data through a status acquisition module, and combines micro-meteorological data as environmental parameters. This multi-dimensional sensing method not only realizes a comprehensive feature mapping of the equipment status from thermal and electromagnetic dimensions, but also provides a physical benchmark for drift detection, effectively distinguishes between real faults and environmental thermal effects, and verifies the high-fidelity characteristics and anti-interference ability of the system under multi-source heterogeneous data acquisition. Attached Figure Description

[0058] Figure 1 This is a schematic diagram of the structure of the few-sample substation equipment status adaptive inspection system based on meta-learning in this application embodiment. Detailed Implementation

[0059] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0060] Please see Figure 1 As shown, the meta-learning-based few-sample substation equipment status adaptive inspection system includes: a status acquisition module, used to acquire a preset substation equipment network topology, taking the substation equipment in the preset substation equipment network topology as target nodes in the target node set, and acquiring the current feature vector and current environmental parameter data of each target node in the target node set; and a drift detection module, used to perform baseline comparison on the current environmental parameter data, determine whether data drift has occurred, and dynamically adjust the confidence threshold corresponding to the preset meta-learning risk assessment model based on the judgment result.

[0061] The risk assessment module is used to input the current feature vector into the preset meta-learning risk assessment model, output the initial risk probability corresponding to each target node, and filter the initial risk probability based on the confidence threshold to generate an effective risk probability.

[0062] The blind zone measurement module is used to obtain the set of unobserved nodes and calculate the current state blind zone entropy of the system based on the historical unobserved duration and basic risk weights corresponding to the set of unobserved nodes.

[0063] The scheduling decision module is used to compare the state blind zone entropy with the preset entropy threshold. If the state blind zone entropy is greater than or equal to the preset entropy threshold, the effective risk probability is ignored and a fallback inspection instruction to reduce the state blind zone entropy is generated. If the state blind zone entropy is less than the preset entropy threshold, an adaptive inspection allocation instruction is generated based on the effective risk probability and the preset resource coverage limit.

[0064] The strategy update module is used to extract the actual inspection result data returned after executing the fallback inspection command or the adaptive inspection assignment command, and feed the actual inspection result data back to the preset meta-learning risk assessment model for parameter update.

[0065] This embodiment provides a few-sample substation equipment status adaptive inspection system based on meta-learning; the status acquisition module is connected to the substation's sensor network in real time to acquire the target node set. Each target node The current feature vector Compared with current environmental parameter data ;

[0066] in, It covers key equipment in the substation, such as main transformers, circuit breakers, and disconnect switches. Indicates that the target node is in the set The index in the range of values ​​is: , For the target node set The total number of nodes in the module is used to construct a digital mapping of the physical world; regarding the integrity logic of data acquisition, if the status acquisition module fails to acquire the target node due to reasons such as device offline or not being covered, etc. The current feature vector The system will generate a specific null value identifier;

[0067] This identifier will be recognized in the subsequent risk assessment process and trigger the system to adjust the corresponding initial risk probability. By forcibly assigning a value of zero, the node is automatically classified as an unobserved node and included in the blind zone measurement range in subsequent steps, thus forming a logical closed loop for data missing handling.

[0068] The drift detection module analyzes the current environmental parameter data. Baseline comparison is performed to determine if data drift has occurred, i.e., a phenomenon where drastic changes in environmental parameters cause the statistical distribution of device feature data to deviate from the preset training distribution. If data drift occurs, this module dynamically adjusts the confidence threshold corresponding to the preset meta-learning risk assessment model. The system automatically increases It becomes more conservative to suppress false alarms caused by environmental noise; the risk assessment module will use the current feature vector Input a pre-defined meta-learning risk assessment model, which is trained through meta-learning and has the ability to identify novel failure modes with only a small number of samples.

[0069] The specific meta-learning training process employs a plot-based meta-training strategy, in which a random sample is drawn from the historical fault dataset during each meta-training iteration. There are several fault categories, and each category is extracted. A support set is constructed from 1 sample, and then extracted The query set is constructed from the samples, forming a training plot of N classes and K samples; the objective function of the meta-training adopts the cross-entropy loss function, which aims to minimize the classification error of the query set samples on the prototype vectors generated by the support set;

[0070] The optimization process adopts a two-layer optimization architecture with inner and outer loops: the inner loop does not update parameters, but only calculates the class prototype vector of the current plot through the support set; the outer loop adopts an adaptive moment estimation optimizer with an initial learning rate of 0.001 and a weight decay coefficient of 1e-4, where the weight decay coefficient is the penalty of the L2 regularization term. Its function is to limit the excessive growth of network weights by introducing the sum of squares of the parameters into the loss function, thereby effectively preventing the model from overfitting during training with few samples. The gradient is calculated based on the classification loss of the query set and the global network parameters of the model are updated.

[0071] In this process, to overcome the problem of weight allocation averaging when standard prototype networks handle heterogeneous multimodal infrared and partial discharge characteristics of substations, this embodiment introduces a learnable feature dimension weight vector when calculating feature distance. Among them, the feature dimension weight vector The initial value is set to The total feature dimension is assumed to contribute equally to fault detection in the initial state, and its calculation formula is as follows:

[0072]

[0073] in, For weighted Euclidean distance; and They are the feature vectors and In the The numerical value of the dimension; The feature dimension weight vector In the The numerical value of the dimension allows the model to adaptively amplify the feature dimension that is sensitive to fault detection;

[0074] Meanwhile, to avoid the training process from getting stuck in an infinite loop, this embodiment clearly sets the training stopping condition: when the number of training iterations reaches the preset maximum value, such as 20,000 training rounds, or when the loss value on the validation set does not decrease within 50 consecutive checkpoints, that is, when the early stopping criterion is met, the system automatically terminates the training loop and saves the optimal model parameters.

[0075] Meanwhile, in order to achieve adaptive adjustment of feature dimension weights, the system calculates the weight vector. The gradient of the weights, according to the chain rule, is calculated using the following formula:

[0076]

[0077] in, For the first The initial risk probability of each inspected node; distance Regarding the first Dimensional weights The partial derivatives are:

[0078]

[0079] in, For comprehensive feature distance; For the current feature vector After mapping, in the first The values ​​of each dimension; For the first Prototype feature vectors of various abnormal patterns In the The numerical values ​​in each dimension are specifically generated by extracting feature vectors from a small number of support set samples belonging to this type of abnormal pattern and calculating their arithmetic mean; this prototype feature vector The initial value source is defined as follows: During the initialization phase before system startup, the system traverses the preset few-sample abnormal pattern library. All support set samples of the anomaly pattern are used with the initial feature mapping function obtained after the offline meta-learning training has converged. Perform forward inference and calculate the feature mean to complete all The first assignment;

[0080] Based on the above gradient, the system adjusts the weight vector. The update is performed using the following formula:

[0081]

[0082] in, For weighted learning rate, and These are the weights before and after the update, respectively; This represents the total number of inspection nodes. Labels for actual faults; For the first The effective risk probability of each inspected node;

[0083] To simplify the calculation, the Jacobian matrix effect introduced by the subsequent normalization function is ignored when calculating and updating the weights using gradient descent. An approximate gradient obtained by directly calculating partial derivatives is used for the initial iteration. After the update, a normalized exponential function is applied to... Normalization is performed to satisfy the non-negativity constraint and the condition that the sum of the weights is 1. The formula is as follows:

[0084]

[0085] in, For the summation index of the feature dimensions, The base of the natural logarithm is used, thus enabling the model to learn a feature metric space with good generalization ability.

[0086] This module outputs the initial risk probability for each target node. And based on confidence threshold right Filtering is performed to generate effective risk probabilities. Based on this, the blind spot measurement module obtains the set of unobserved nodes. and based on Historical unobserved duration corresponding to the middle node With basic risk weights Calculate the current state blind zone entropy of the system. This indicator aims to measure the total potential risk accumulated in unobserved areas of the entire substation network due to resource constraints; the dispatch decision module will consider the state blind zone entropy. Compared with the preset entropy threshold Comparison:

[0087] when When this occurs, it indicates that the system's control over the power grid state has dropped to a dangerous level. At this point, the system enters a priority mode that ensures basic system security, ignoring specific details. Generate an output designed to reduce The bottom-line inspection instruction;

[0088] when When this time is reached, it indicates that the overall system status is under control, and the system enters the precise troubleshooting mode, based on... With the preset resource coverage limit Generate adaptive inspection allocation instructions;

[0089] Among them, the preset entropy critical threshold It is an empirical value set based on historical operation and maintenance safety baseline assessment data of substations, and a preset upper limit for resource coverage. The maximum number of nodes that can be inspected at one time is determined by reading the current available inspection robot's power status and work schedule in real time; the strategy update module extracts the actual inspection result data returned after the execution command and feeds it back to the meta-learning risk assessment model for parameter update, so as to improve the model's adaptability to new environments and new faults.

[0090] This embodiment demonstrates the scheduling robustness of a few-sample substation equipment status adaptive inspection system based on meta-learning in a substation inspection scenario. By introducing state blind zone entropy as the core constraint of scheduling, the state acquisition module and the scheduling decision module work together to solve the problem that traditional risk-based scheduling is prone to getting stuck in local observation when resources are limited.

[0091] If environmental interference leads to unreliable data or excessively high blind zone risk, the system automatically downgrades to a fallback strategy, ensuring the robustness of power grid monitoring. Under normal conditions, the system utilizes the few-sample reasoning capability of meta-learning to achieve agile capture of new types of faults, verifying the adaptive capability of this technical solution under different environmental interferences.

[0092] In a preferred embodiment of the present invention, the method for obtaining the current feature vector and current environmental parameter data of each target node in the target node set includes: obtaining infrared feature data and partial discharge feature data corresponding to each target node through preset detection resources, and combining them into a current feature vector; obtaining micro-meteorological data including temperature, humidity, wind speed and light intensity through preset meteorological sensors, as current environmental parameter data.

[0093] This embodiment details the execution logic of the status acquisition module in obtaining the current feature vector and current environmental parameter data of each target node in the target node set; the status acquisition module obtains the preset substation equipment network topology. ,Will The substation equipment in the middle is used as the target node set The target node in the process is selected, and this operation aims to ensure the spatial correlation of the data; the status acquisition module acquires the infrared feature data corresponding to each target node through preset detection resources, namely the dual-light pod carried by the inspection robot. With partial discharge characteristic data ;

[0094] Specifically, the original infrared sensor image and the original partial discharge time-series signal are first acquired, and then mapped into one-dimensional infrared feature vectors through a pre-defined feature extraction network. With partial discharge eigenvector Specifically, for the original infrared sensing image, the preset feature extraction network uses a 50-layer residual network with fully connected layers removed as the backbone structure, and outputs a fixed-dimensional infrared feature vector through global average pooling. To reflect the thermal defect state of the equipment; for the original partial discharge time-series signal, a one-dimensional convolutional neural network containing multiple one-dimensional convolutional and max-pooling layers is used to extract time-series features and output partial discharge feature vectors. To reflect the insulation defect status of the equipment;

[0095] The status acquisition module combines these two by concatenating the feature vectors bit by bit to form the current feature vector. A unified high-dimensional multimodal vector is constructed, and the combination of this multimodal data provides richer fault discrimination basis for the meta-learning model; the status acquisition module obtains micro-meteorological data, including temperature, humidity, wind speed and light intensity, through micro-meteorological stations deployed in the substation, as current environmental parameter data;

[0096] This embodiment demonstrates the multidimensional sensing robustness of the status acquisition module in the complex physical environment of a substation. By fusing infrared feature data and partial discharge feature data, the status acquisition module can comprehensively sense the equipment status from both thermal and electromagnetic dimensions. At the same time, micro-meteorological data is introduced as current environmental parameter data, providing a physical benchmark for subsequent drift detection. This effectively distinguishes between real faults and thermal hysteresis effects caused by environmental thermal effects, verifying the high-fidelity characteristics of this technical solution under multi-source heterogeneous data acquisition.

[0097] In a preferred embodiment of the present invention, the method for determining whether data drift has occurred and dynamically adjusting the confidence threshold corresponding to the preset meta-learning risk assessment model based on the determination result includes:

[0098] Obtain preset environmental baseline data; calculate the absolute value of the difference between the current environmental parameter data and the preset environmental baseline data as the environmental deviation value; compare the environmental deviation value with the preset drift judgment threshold.

[0099] If the environmental deviation value is greater than or equal to the preset drift judgment threshold, data drift is determined to have occurred, and the preset initial confidence threshold is added to the preset penalty step size to obtain the confidence threshold.

[0100] If the environmental deviation value is less than the preset drift judgment threshold, it is determined that no data drift has occurred, and the preset initial confidence threshold is used as the confidence threshold.

[0101] This embodiment details the internal logic of the drift detection module in determining whether data drift has occurred and dynamically adjusting the confidence threshold; the drift detection module acquires preset environmental baseline data. The baseline data is a set of multi-dimensional vectors covering temperature, humidity, wind speed and light intensity, usually derived from historical average environmental parameters for that season or time period.

[0102] Given the multidimensional vector nature of environmental parameters, the drift detection module calculates the current environmental parameter data. Compared with the preset environmental baseline data The absolute value of the generalized difference between them, i.e., the vector norm, is used as the environmental deviation value. To eliminate dimensional differences across different environmental dimensions, such as temperature being measured in °C and wind speed in m / s, and to prevent the denominator from being zero, this embodiment specifically employs a weighted normalized Euclidean distance formula with a smoothing term for calculation.

[0103]

[0104] in, This represents the total number of dimensions for environmental parameters, sourced from system presets, such as... ;

[0105] For the current environment vector at the th The dimension values ​​are derived from data collected by weather sensors.

[0106] For the environmental baseline vector at the th The values ​​for the dimensions are specifically selected from historical statistical averages that match the season and hourly time period of the current detection time, in order to eliminate normal time-periodic environmental fluctuations.

[0107] For the first The standard deviation of the dimension parameter, specifically the standard deviation of historical data for the same time period mentioned above, is used for normalization.

[0108] For very small positive numbers, such as This is used to prevent calculation overflow anomalies where the denominator is zero when the historical fluctuation of a certain environmental parameter approaches zero.

[0109] For the first The weighting coefficients of the dimensional parameters on the impact of the device's operating status satisfy... ;

[0110] The drift detection module will Compared with the preset drift detection threshold A comparison is made; among which, the preset drift determination threshold is used. It is a statistical benchmark value determined by collecting a set of environmental deviation values ​​under fault-free conditions during the normal operation cycle of the substation in history and extracting the 95th percentile of the set;

[0111] when When the drift detection module determines that data drift has occurred, environmental noise may significantly interfere with the meta-learning model's judgment. The drift detection module then executes an adjustment formula:

[0112]

[0113] in, The adjusted confidence threshold is dimensionless. The initial confidence threshold is dimensionless. The penalty step size is dimensionless and its specific value is a fixed constant determined by simulating environmental noise of different intensities on the validation set, with the objective function being to maximize the model's F1 score in noisy environments. This operation raises the threshold for model alarms, ensuring that alarms are only issued when the model is extremely confident.

[0114] when At this point, the drift detection module determines that no data drift has occurred. The environment is stable, and the drift detection module maintains high agility, setting the preset initial confidence threshold. As a confidence threshold ;

[0115] This embodiment demonstrates the dynamic gating characteristics of the drift detection module under severe weather interference scenarios. By calculating the weighted normalized environmental deviation value and dynamically adjusting the confidence threshold, the drift detection module actively sacrifices some sensitivity in exchange for extremely high specificity. This overcomes the data synchronization deviation problem in the multi-dimensional sensor acquisition process, avoids the exhaustion of inspection resources due to false alarms caused by sudden environmental changes, and thus protects the availability of the system. It also verifies the adaptive noise resistance robustness of this technical solution in non-stationary environments.

[0116] In a preferred embodiment of the present invention, the method for generating an effective risk probability includes: acquiring a preset few-sample abnormal pattern library, wherein the preset few-sample abnormal pattern library includes one or more abnormal pattern features; calculating the feature distance between the current feature vector and each abnormal pattern feature in the preset few-sample abnormal pattern library using a preset meta-learning risk assessment model; mapping the feature distance to a value negatively correlated with the feature distance using a preset normalization function, and using the value as the initial risk probability corresponding to each target node; comparing the initial risk probability with a confidence threshold; if the initial risk probability is greater than or equal to the confidence threshold, then the initial risk probability is used as the effective risk probability; if the initial risk probability is less than the confidence threshold, then the effective risk probability is assigned a value of zero.

[0117] This embodiment details the method by which the risk assessment module generates effective risk probabilities; the risk assessment module obtains a preset library of few-sample abnormal patterns. The pre-defined few-sample anomaly pattern library includes more than one anomaly pattern feature. This library stores several typical faults, namely, meta-feature vectors of insulator damage or clamp overheating. The risk assessment module calculates the current feature vector through a pre-defined meta-learning risk assessment model, which is a model based on the existing prototype network architecture. and The feature distances for each abnormal pattern feature are calculated using a weighted Euclidean distance that includes the weights of the learnable feature dimensions, and the minimum feature distance is extracted as the comprehensive feature distance. The formula is as follows:

[0118]

[0119] in, To retrieve all exception pattern categories The minimum distance in the matrix is ​​used to find the known anomaly pattern that is most similar to the features of the current node;

[0120] This is a feature mapping function used to map input data to a high-dimensional feature space; specifically, this feature mapping function... A multilayer perceptron network structure is adopted, consisting of an input layer, two hidden layers, and one output layer; the number of neurons in the input layer is proportional to the current feature vector. The dimensions are consistent;

[0121] The first hidden layer contains 256 neurons, the second hidden layer contains 128 neurons, and the output layer contains 64 neurons, thus mapping the input data to a 64-dimensional feature space. Each hidden layer is followed by a modified linear unit activation function to introduce non-linear expressive power, and a random deactivation layer with a deactivation rate of 0.3 is added to prevent overfitting. The output layer uses linear output without using an activation function.

[0122] The network parameters were initialized using the Ho's normal distribution initialization method. Before the system went online, the network structure was trained offline using a large-scale historical public substation normal and basic fault dataset to obtain initial network parameters with few-sample generalization ability.

[0123] It should be noted that the offline meta-learning training process adopts the same plot-based meta-training strategy as described above, that is, by constructing training plots of N classes and K samples to perform inner and outer loop double-layer optimization, thereby ensuring the consistency of feature measurement mechanism between the pre-training stage and the online application stage.

[0124] The risk assessment module will use feature distance The value is mapped to a value negatively correlated with the feature distance using a preset normalization function, and this value is used as the initial risk probability. The formula is as follows:

[0125]

[0126] in, The scaling parameter is dimensionless, and its specific value is an empirical constant determined by performing a hyperparameter grid search on a historical small sample validation set. To ensure the negative correlation mapping logic that the smaller the feature distance, the closer the risk probability is to 1, mathematical constraints must be satisfied. ;

[0127] The offset parameter is dimensionless, and its specific value is an empirical constant determined by performing a hyperparameter grid search on a historical small sample validation set. To ensure the upper limit of the basic probability at zero distance in accordance with the aforementioned mapping logic, it must satisfy mathematical constraints. ;

[0128] Specifically, to handle data loss, if the status acquisition module fails to obtain a valid current feature vector for a target node, the risk assessment module will directly remove the node's current feature vector. Set to zero to ensure that the blind zone measurement module takes over subsequently; these two parameters together aim to adjust the slope and offset of the activation function, ensuring that the smaller the distance, the closer the probability is to 1; the risk assessment module will set the initial risk probability. Compared with the current dynamically adjusted confidence threshold Compare; when At that time, the risk assessment module will set the initial risk probability. As effective risk probability ;when At that time, the risk assessment module will effectively determine the risk probability. The value is assigned to zero;

[0129] This embodiment calculates the feature distance and combines it with a dynamically adjusted confidence threshold for threshold filtering. The risk assessment module can quickly identify similar fault features based on very few historical samples, effectively eliminating low-confidence predictions in the ambiguity zone, and constructing a cascaded defense mechanism for risk identification. This ensures that subsequent scheduling decisions are based only on high-confidence risk signals, verifying the accurate screening capability of this technical solution under conditions of few samples.

[0130] In a preferred embodiment of the present invention, the method for calculating the current state blind zone entropy of the system includes: extracting nodes in the target node set that have not been pre-allocated detection resources into an unobserved node set; obtaining the historical unobserved duration and basic risk weight corresponding to each unobserved node in the unobserved node set; multiplying the historical unobserved duration and basic risk weight to obtain the single-point blind zone risk value corresponding to each unobserved node; and summing the single-point blind zone risk values ​​corresponding to all unobserved nodes in the unobserved node set to obtain the state blind zone entropy.

[0131] This embodiment details the method by which the blind zone measurement module calculates the current state blind zone entropy of the system; the blind zone measurement module sets the target node set. Nodes that have not been pre-allocated detection resources, i.e., those not yet planned for coverage within the current scheduling cycle, are extracted as the initial set of unobserved nodes; simultaneously, the target node set is detected. Although resources have been allocated, the probability of effective risk remains. Nodes that were forcibly assigned a value of zero due to missing data are added to the initial set of unobserved nodes, thus constructing a complete set of unobserved nodes. ;

[0132] The blind zone metric module obtains the set of unobserved nodes. Each unobserved node Corresponding historical unobserved duration With basic risk weights ;in This represents the time interval since the node was last inspected. If the node is newly connected to the system and has never been inspected before, then its historical unobserved duration is used. Initialize to the preset maximum tolerance time, such as the total time since the equipment was put into operation;

[0133] The constants are preset based on the equipment type and importance of the node. Specifically, they are obtained by: pre-establishing an attribute lookup table containing the mapping relationship between equipment type and basic risk weights; assigning higher risk weight values ​​to core-level hub equipment such as main transformers and lower risk weight values ​​to ordinary peripheral equipment; and automatically assigning initial values ​​to each node by searching this attribute lookup table. The blind spot measurement module will record the historical unobserved duration. With basic risk weights Perform a product operation to obtain the single-point blind zone risk value for each unobserved node. The formula is as follows:

[0134]

[0135] For more critical equipment, the potential risk escalates faster when it is unmonitored, and therefore a higher risk growth rate is assigned, with its dimension naturally set as the reciprocal of time. This ensures the dimensionless nature of the state blind zone entropy as an information measure or a measure of system uncertainty.

[0136] The calculation logic shows that the more important the equipment and the longer the disconnection time, the greater its potential blind zone risk. Furthermore, through dimensional cancellation, the single-point blind zone risk value becomes a dimensionless quantity. The blind zone measurement module measures the set of unobserved nodes. All single-point blind zone risk values Perform summation to obtain the state blind zone entropy. The formula is as follows:

[0137]

[0138] in, For the first The risk value of a single blind zone for an unobserved node, dimensionless; This represents the total number of unobserved nodes, expressed in units of nodes.

[0139] This embodiment demonstrates the robustness of the blind zone measurement module in quantifying unknown risks under extremely limited inspection resources. By fusing historical unobserved duration with basic risk weights having the inverse dimension of time to calculate state blind zone entropy, the blind zone measurement module not only eliminates the influence of the time dimension, ensuring matching with subsequent dimensionless thresholds, but also no longer focuses solely on known faults, but mathematically quantifies the accumulation of unknown risks. This provides the system with a global safety baseline indicator, preventing sudden and serious faults caused by long-term neglect of certain areas, and verifying the foresight of this technical solution in macro-risk control.

[0140] In a preferred embodiment of the present invention, the method for generating a fallback inspection instruction to reduce state blind zone entropy includes: sorting unobserved nodes in the unobserved node set in descending order according to the magnitude of the single-point blind zone risk value to generate a high blind zone node sequence; obtaining a preset resource coverage upper limit; sequentially extracting unobserved nodes from the beginning of the high blind zone node sequence with a number equal to the preset resource coverage upper limit; if the total number of unobserved nodes in the high blind zone node sequence is less than the preset resource coverage upper limit, then extracting all unobserved nodes to construct a fallback target set; and generating a fallback inspection instruction based on the fallback target set.

[0141] This embodiment details the method by which the scheduling decision module generates fallback inspection instructions to reduce state blind zone entropy; if the state blind zone entropy Greater than or equal to the entropy critical threshold If the system enters basic support mode, the scheduling decision module will then make a decision based on the single-point blind zone risk value. The size of the set of unobserved nodes The unobserved nodes in the array are sorted in descending order;

[0142] To avoid situations where the complementary relationship between historical unobserved duration and basic risk weights leads to equal single-point blind zone risk values ​​for different nodes, making it impossible to distinguish priorities, when multiple nodes exist... At the same time, the scheduling decision module further determines the basic risk weight corresponding to each node. Sort the values ​​in descending order of size. If they are still the same, they are arranged according to the preset numbering order of the nodes in the target node set, thereby generating a unique high-blind-zone node sequence. ;

[0143] The scheduling decision module obtains the preset resource coverage limit. This parameter typically represents the total number of nodes that all currently available robots can cover; the scheduling decision module starts from the sequence of high blind zone nodes. The number of resources extracted sequentially from the starting end is equal to the preset resource coverage limit. Unobserved nodes; if the high blind zone node sequence The total number of unobserved nodes is less than the preset resource coverage limit. The scheduling decision module then extracts all unobserved nodes to construct a fallback target set. The scheduling decision module is based on a fallback target set. Generate a fallback inspection instruction;

[0144] This embodiment demonstrates the uncertainty control capability of the scheduling decision module in critical scenarios with high blind spots in the system. By constructing a sequence of nodes with high blind spots and extracting a set of fallback targets, the fallback inspection command forces the system to prioritize the observation of the most untested and most important devices. This strategy reduces the overall uncertainty of the system at the fastest speed, prevents cascading failures, and verifies the bottom-line defense capability of this technical solution under extreme resource constraints.

[0145] In a preferred embodiment of the present invention, the method for generating adaptive inspection allocation instructions includes: extracting all effective risk probabilities from the target node set; extracting target nodes with effective risk probabilities greater than zero, and arranging them in descending order according to the magnitude of the effective risk probabilities to generate a high-risk node sequence;

[0146] Extract nodes sequentially from the beginning of the high-risk node sequence until the number of extracted nodes equals the preset resource coverage limit or all nodes in the high-risk node sequence are extracted, and construct the primary target set; calculate the difference between the preset resource coverage limit and the number of nodes in the primary target set as the resource difference;

[0147] The resource difference is compared with zero; if the resource difference is greater than zero, the nodes that already exist in the primary target set in the high blind zone node sequence are removed and a new candidate blind zone node sequence is generated. The number of nodes equal to the resource difference is extracted from the candidate blind zone node sequence and added to the primary target set. An adaptive inspection allocation instruction is generated based on the supplemented primary target set.

[0148] If the resource difference is zero, then an adaptive inspection allocation instruction is generated directly based on the primary target set.

[0149] This embodiment details the method for the scheduling decision module to generate adaptive inspection and allocation instructions; if the state blind zone entropy Less than the entropy critical threshold The system then enters a precise screening mode; the scheduling decision module extracts all valid risk probabilities from the target node set. ;

[0150] The scheduling decision module will effectively determine the risk probability. Target nodes with values ​​greater than zero are extracted and, based on... The values ​​are sorted in descending order; to ensure the uniqueness of the algorithm's execution result and avoid random allocation of resources under equal risk probabilities, when there are multiple target nodes with effective risk probabilities... At the same time, the scheduling decision module further extracts the single-point blind zone risk values ​​corresponding to these nodes. Perform secondary descending sorting to prioritize nodes with both high risk and high blind spot. If the risk value of a single blind spot is low... If they are still the same, then the target nodes are arranged in the final order according to the preset numbering order in the preset substation equipment network topology, thereby generating a uniquely determined sequence of high-risk nodes. ;

[0151] The scheduling decision module selects high-risk node sequences. Nodes are extracted sequentially from the starting point until the number of extracted nodes equals the preset resource coverage limit. or high-risk node sequence All nodes in the dataset are extracted to construct the primary target set. Based on this, the scheduling decision module calculates the preset resource coverage limit. With the primary target set The difference in the number of nodes is used as the resource difference. The formula is as follows:

[0152]

[0153] in, This represents the upper limit of resource coverage, expressed in units of [number]. The number of nodes in the primary target set, in units of;

[0154] The scheduling decision module will use the resource difference Compare with zero; when If resources are still available after processing all high-risk nodes, the scheduling decision module will prioritize high-blind-zone node sequences. It already exists in the primary target set. The nodes in the list are deduplicated to generate a candidate blind zone node sequence. Starting from the beginning of this candidate blind zone node sequence, nodes are extracted in sequence with a number equal to the resource difference. If the total number of nodes in the candidate blind zone node sequence is less than Then extract all nodes and add them to the primary target set. And based on the supplemented primary target set Generate adaptive inspection assignment instructions; when When this occurs, it indicates that high-risk nodes have occupied all resources, and the scheduling decision module directly bases its decisions on the primary target set. Generate adaptive inspection assignment instructions;

[0155] This embodiment demonstrates the refined scheduling robustness of the scheduling decision module in a conventional substation inspection scenario. By calculating resource differences and executing a dynamic replenishment mechanism, the scheduling decision module achieves a perfect combination of risk priority and reserve blind spot filling. The adaptive inspection allocation instruction prioritizes concentrating resources to solve explicit faults discovered by the meta-learning model, and automatically covers the nodes with the highest risk in the blind spot when resources permit, thus delaying the growth of blind spot entropy to the maximum extent and verifying the multi-objective optimization capability of this technical solution under complex constraints.

[0156] In a preferred embodiment of the present invention, the method for feeding back actual inspection result data to a preset meta-learning risk assessment model for parameter updating includes: extracting real fault labels from the actual inspection result data; calculating the difference between the real fault label and the corresponding effective risk probability as a prediction error value; calculating the gradient of the network parameters of the preset meta-learning risk assessment model using a gradient descent algorithm based on the prediction error value; and iteratively updating the network parameters of the preset meta-learning risk assessment model based on the network parameter gradient to complete the parameter update.

[0157] This embodiment details the closed-loop logic of the strategy update module feeding back the actual inspection result data to the preset meta-learning risk assessment model for parameter updates; the strategy update module extracts the real fault labels corresponding to each inspected node from the actual inspection result data. The label usually comes from the judgment result after the photos returned by the inspection robot are confirmed by humans or high-precision offline models. A confirmed fault is marked as 1, and a false alarm is marked as 0.

[0158] The strategy update module calculates the actual fault label corresponding to each inspected node. and the corresponding effective risk probability The difference between them is used as the single-point prediction error value. And based on this error value, a single-point loss function is constructed. Further inspection of all items in this patrol Construct a global mean squared error loss function for each node. The formula is as follows:

[0159]

[0160]

[0161]

[0162] in, For the first The single-point prediction error value of each inspected node; The global mean squared error loss function; For the first The single-point mean square error loss of each inspected node, dimensionless.

[0163] The policy update module calculates the gradient of the network parameters of the pre-defined meta-learning risk assessment model based on the prediction error value using the gradient descent algorithm. It should be noted that, due to It is after the confidence threshold The filtered result, i.e. when hour Direct differentiation will cause the gradient to vanish;

[0164] Therefore, this embodiment employs a pass-through estimator technique, treating the threshold function as an identity mapping during backpropagation gradient calculation, thereby... The error gradient is backpropagated to the initial risk probability. The corresponding network parameters, and by analyzing all The global gradient is calculated by averaging the gradients of each node, as shown in the following formula:

[0165]

[0166] in, For network parameter gradients; This indicates that the pass-through estimator approximation was applied, i.e., partial derivatives were set during backpropagation. It ignores the truncation property of the threshold function, ensuring that false negatives are detected. but Even under these conditions, the model can still obtain update motivation; this embodiment adopts a gradient-sensitive window strategy, and the specific gradient calculation formula is as follows:

[0167]

[0168] in, This is a preset window radius, for example, a value of 0.2; The preset linear decay coefficient is used, for example, a value of 0.1; this strategy allows the model to focus on optimizing ambiguous samples near the decision boundary, improving the parameter convergence stability during the threshold adaptive adjustment process; that is, only when the initial risk probability... At the threshold Within the preset neighborhood, for example Only then can it be fully applied. Approximation; if If the distance is far from this window, a linear decay coefficient, such as 0.1, is introduced, so that the model focuses on optimizing ambiguous samples near the decision boundary, thereby improving the parameter convergence stability during the threshold adaptive adjustment process.

[0169] The policy update module is based on network parameter gradients. The network parameters of the preset meta-learning risk assessment model are iteratively updated using the following formula:

[0170]

[0171] in, The updated network parameters are dimensionless. These are the network parameters before the update, dimensionless; The learning rate is dimensionless.

[0172] Based on the normalization function defined in the embodiment, the chain rule is applied as follows:

[0173]

[0174] The partial derivative of the probability with respect to the feature distance is:

[0175]

[0176] Through the above explanation, the system can accurately transmit the prediction error signal to the feature extraction network. Deep parameters This ensures the feature space optimization capability of the meta-learning model under conditions of few samples;

[0177] Furthermore, given the network parameters The update changes the mapping rules of the feature space. To ensure the accuracy of the distance metric, after the parameter update is completed, the system uses the updated network parameters. Feature extraction is performed again on the support set samples stored in the pre-defined few-shot anomaly pattern library, and the prototype feature vector of each anomaly pattern is dynamically updated by calculating the mean of the within-class features. The specific mathematical update formula is as follows:

[0178]

[0179] in, For belonging to the first The set of all support sets of original feature samples for each anomaly pattern is derived from a pre-defined few-sample anomaly pattern library.

[0180] For set The total number of samples included, in units of individuals;

[0181] To use the updated network parameters For the sample The high-dimensional feature vector after feature mapping; through this determined update step, the system eliminates the benchmark bias caused by parameter drift and realizes the synchronous evolution of the feature extractor and the prototype library;

[0182] This embodiment demonstrates the online learning robustness of the strategy update module in the scenario of substation equipment aging and evolution. By constructing a closed loop of prediction execution feedback and combining iterative updates with the calculation of network parameter gradients using a pass-through estimator, the strategy update module can use the measured results of each inspection to construct and correct the meta-learning model in situ. Especially for scenarios with few samples, this continuous parameter update mechanism allows the preset meta-learning risk assessment model to quickly adapt to new environments and new fault characteristics, verifying the performance self-evolution capability of this technical solution under long-term operation.

[0183] The foregoing has provided a detailed description of one embodiment of the present invention, but this description is merely a preferred embodiment and should not be construed as limiting the scope of the invention. All equivalent variations and modifications made within the scope of the present invention should still fall within the scope of the present invention.

Claims

1. A few-sample substation equipment status adaptive inspection system based on meta-learning, characterized in that, include: The status acquisition module is used to acquire a preset substation equipment network topology, take the substation equipment in the preset substation equipment network topology as target nodes in the target node set, and acquire the current feature vector and current environmental parameter data of each target node in the target node set. The drift detection module is used to perform baseline comparison on the current environmental parameter data, determine whether data drift has occurred, and dynamically adjust the confidence threshold corresponding to the preset meta-learning risk assessment model based on the determination result. The risk assessment module is used to input the current feature vector into the preset meta-learning risk assessment model, output the initial risk probability corresponding to each target node, and filter the initial risk probability based on the confidence threshold to generate an effective risk probability. The blind zone measurement module is used to obtain the set of unobserved nodes and calculate the current state blind zone entropy of the system based on the historical unobserved duration and basic risk weight corresponding to the set of unobserved nodes. The scheduling decision module is used to compare the state blind zone entropy with a preset entropy threshold. If the state blind zone entropy is greater than or equal to the preset entropy threshold, the effective risk probability is ignored and a fallback inspection instruction aimed at reducing the state blind zone entropy is generated. If the state blind zone entropy is less than the preset entropy threshold, an adaptive inspection allocation instruction is generated based on the effective risk probability and a preset resource coverage limit. The strategy update module is used to extract the actual inspection result data returned after executing the bottom-line inspection instruction or the adaptive inspection allocation instruction, and feed the actual inspection result data back to the preset meta-learning risk assessment model for parameter update; Methods for calculating the current state blind zone entropy of a system include: The nodes in the target node set that have not been pre-allocated detection resources are extracted as the unobserved node set; Obtain the historical unobserved duration and the basic risk weight for each unobserved node in the set of unobserved nodes; The product of the historical unobserved duration and the basic risk weight is used to obtain the single-point blind zone risk value corresponding to each unobserved node. The single-point blind zone risk values ​​corresponding to all unobserved nodes in the set of unobserved nodes are summed to obtain the state blind zone entropy. Methods for generating fallback inspection instructions designed to reduce the entropy of the state blind zone include: Based on the magnitude of the single-point blind zone risk value, the unobserved nodes in the unobserved node set are sorted in descending order to generate a high blind zone node sequence; Obtain the preset resource coverage limit; Extract unobserved nodes from the beginning of the high blind zone node sequence in sequence, with a number equal to the preset resource coverage limit. If the total number of unobserved nodes in the high blind zone node sequence is less than the preset resource coverage limit, then extract all the unobserved nodes to construct a fallback target set. The bottom-line inspection instruction is generated based on the bottom-line target set.

2. The few-sample substation equipment status adaptive inspection system based on meta-learning according to claim 1, characterized in that, Methods for obtaining the current feature vector and current environment parameter data of each target node in the target node set include: Infrared feature data and partial discharge feature data corresponding to each target node are obtained through preset detection resources and combined into the current feature vector; Micro-meteorological data, including temperature, humidity, wind speed, and light intensity, are acquired through preset meteorological sensors and used as the current environmental parameter data.

3. The few-sample substation equipment status adaptive inspection system based on meta-learning according to claim 2, characterized in that, Methods for determining whether data drift has occurred and dynamically adjusting the confidence threshold of the preset meta-learning risk assessment model based on the determination result include: Obtain preset environmental baseline data; Calculate the absolute value of the difference between the current environmental parameter data and the preset environmental baseline data, and use it as the environmental deviation value; The environmental deviation value is compared with a preset drift determination threshold; If the environmental deviation value is greater than or equal to the preset drift determination threshold, then data drift is determined to have occurred, and the preset initial confidence threshold is added to the preset penalty step size to obtain the confidence threshold. If the environmental deviation value is less than the preset drift determination threshold, it is determined that no data drift has occurred, and the preset initial confidence threshold is used as the confidence threshold.

4. The few-sample substation equipment status adaptive inspection system based on meta-learning according to claim 3, characterized in that, Methods for generating effective risk probabilities include: Obtain a preset few-sample abnormal pattern library, wherein the preset few-sample abnormal pattern library includes one or more abnormal pattern features; The feature distance between the current feature vector and the features of each abnormal pattern in the preset few-sample abnormal pattern library is calculated using the preset meta-learning risk assessment model. The feature distance is mapped to a value that is negatively correlated with the feature distance through a preset normalization function, and the value is used as the initial risk probability corresponding to each target node; Compare the initial risk probability with the confidence threshold; If the initial risk probability is greater than or equal to the confidence threshold, then the initial risk probability is taken as the effective risk probability. If the initial risk probability is less than the confidence threshold, then the effective risk probability is set to zero.

5. The few-sample substation equipment status adaptive inspection system based on meta-learning according to claim 1, characterized in that, Methods for generating adaptive inspection assignment instructions include: Extract all the effective risk probabilities from the target node set; The target nodes with an effective risk probability greater than zero are extracted and sorted in descending order according to the magnitude of the effective risk probability to generate a high-risk node sequence. Extract nodes sequentially from the beginning of the high-risk node sequence until the number of extracted nodes equals the preset resource coverage limit or all nodes in the high-risk node sequence are extracted, and construct the primary target set. Calculate the difference between the preset resource coverage limit and the number of nodes in the primary target set, and use it as the resource difference; Compare the resource difference with zero; If the resource difference is greater than zero, the nodes that already exist in the primary target set in the high blind zone node sequence are removed and a new candidate blind zone node sequence is generated. The number of nodes equal to the resource difference is extracted from the candidate blind zone node sequence and added to the primary target set. The adaptive inspection allocation instruction is generated based on the supplemented primary target set. If the resource difference is equal to zero, the adaptive inspection allocation instruction is generated directly based on the primary target set.

6. The few-sample substation equipment status adaptive inspection system based on meta-learning according to claim 5, characterized in that, The method for feeding back the actual inspection results data to the preset meta-learning risk assessment model for parameter updates includes: Extract the actual fault labels from the actual inspection result data; Calculate the difference between the actual fault label and the corresponding effective risk probability as the prediction error value; Based on the prediction error value, the gradient descent algorithm is used to calculate the gradient of the network parameters of the preset meta-learning risk assessment model. Based on the network parameter gradient, the network parameters of the preset meta-learning risk assessment model are iteratively updated to complete the parameter update.

Images