Defect diagnosis method, system and equipment for electrical equipment of distribution network and medium

Abstract

The invention discloses a defect diagnosis method, system and device for distribution network electrical equipment and a medium, and relates to the technical field of distribution network electrical equipment, and the method comprises the steps: an edge node obtains a local distribution network electrical equipment diagnosis sample; the cloud performs simulation gradient updating on parameters of the defect diagnosis model of the edge node according to the real label and the diagnosis sample; dividing the diagnosis sample based on the high and low confidence thresholds to generate a multilateral experience conduction set, and performing federal aggregation optimization to generate a global optimization model; the edge node obtains difference measurement of the defect diagnosis model and the global optimization model to construct a comparison sample; and the cloud end updates a quality evaluation function according to the comparison sample construction model to generate an evaluation result, and finally analyzes whether the high and low confidence thresholds are updated or not according to the evaluation result so as to obtain a final defect diagnosis model. The method solves the problems that an existing model is difficult to give accurate judgment with high confidence, and is low in accuracy and robustness and poor in adaptive capacity.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of distribution network electrical equipment, and particularly relates to a defect diagnosis method, system, device and medium for distribution network electrical equipment. BACKGROUND

[0002] With the deepening of the construction of smart grid, the safe and reliable operation of distribution network electrical equipment has become a key link to ensure power supply. Traditional equipment inspection mainly relies on periodic manual inspection, which is inefficient and difficult to find early hidden dangers. In recent years, the equipment defect diagnosis technology based on artificial intelligence has developed rapidly. By fusing multi-source sensing data such as infrared thermal imaging, partial discharge voiceprint and electrical parameters, the intelligent identification of equipment operating state can be realized. The introduction of cloud-edge collaborative architecture enables the edge side to deploy lightweight diagnosis models to realize real-time response, while the cloud end is responsible for model training and optimization, forming a distributed intelligent diagnosis system. However, the edge device is limited by the local data size and computing resources, and its diagnosis model often faces the problem of insufficient generalization ability, especially when encountering rare defect types or complex operating conditions, the confidence of the model output is low, and the diagnosis accuracy decreases significantly.

[0003] At present, the existing defect diagnosis method of distribution network electrical equipment has many limitations in actual application. First, the diagnosis model deployed on the edge is usually trained based on a limited and homogeneous data set. When it faces heterogeneous data of different regions and different types of equipment or low-confidence samples lacking clear features, the generalization ability of the model will be significantly insufficient, resulting in a decrease in diagnosis accuracy. Especially for some early and highly concealed defects, the existing model often has difficulty in giving accurate judgments with high confidence. Second, due to the differences in geographical environment, equipment aging degree and fault mode of each edge node, simply mixing and training all edge experiences or updating them separately may lead to poor learning effect of the model, and even introduce irrelevant experience. This inefficient collaborative use of multi-edge experience seriously restricts the accuracy and robustness of the diagnosis system in complex and variable distribution network environments. In addition, the current model updating and parameter configuration process mostly relies on manual experience optimization, which is difficult to quickly adapt to the data distribution drift and new defect modes brought by the continuous evolution of the distribution network system, greatly limiting the adaptive ability and long-term operation efficiency of the diagnosis system. SUMMARY

[0004] The present application provides a defect diagnosis method, system, device and medium for distribution network electrical equipment, which solves the problems of existing models that are difficult to give accurate judgments with high confidence, low accuracy and robustness in complex and variable distribution network environments, and poor adaptive ability.

[0005] Therefore, the first aspect of the present application provides a defect diagnosis method for distribution network electrical equipment, which comprises:

[0006] S1, the edge node diagnoses the operation state data of the local distribution network electrical equipment within a preset small time scale by a defect diagnosis model issued by the cloud, obtains a first defect category probability distribution, generates a diagnosis sample, and uploads the diagnosis sample to the cloud;

[0007] S2, the cloud updates the parameters of the defect diagnosis model of the edge node by simulated gradient based on the real labels fed back by the operation inspection and the first sub-diagnosis sample received within a preset medium time scale;

[0008] S3, the cloud divides the second sub-diagnosis sample received within a preset large time scale based on a low confidence threshold and a high confidence threshold, and generates a multi-edge experience conduction set according to the division result;

[0009] S4, the cloud performs federated aggregation optimization on the parameters of the defect diagnosis model updated by simulated gradient, generates a global optimization model by combining the knowledge distillation in the multi-edge experience conduction set, and issues the global optimization model to the edge node;

[0010] S5, the edge node obtains the difference measure of the defect diagnosis model and the global optimization model, and when the difference measure is greater than a preset threshold, constructs a comparison sample and uploads the comparison sample to the cloud;

[0011] S6, the cloud constructs a model update quality evaluation function according to the comparison sample to generate an evaluation result, and when the evaluation result is not greater than 0, updates the low confidence threshold and the high confidence threshold and returns to step S3, and when the evaluation result is greater than 0, takes the global optimization model as the final defect diagnosis model and issues it to the edge node for diagnosing the operation state data.

[0012] Optionally, in step S1, the edge node comprises:

[0013] S11, feature extraction is performed on the operation state data within a small time scale to obtain operation state features;

[0014] S12, inputting the operation state features into the defect diagnosis model issued by the cloud, outputting an identification result including a first defect category probability distribution and a diagnosis confidence, and generating a predicted defect category according to the defect category probability distribution;

[0015] S13, constructing a diagnosis sample by combining the operation state features, the identification result and the predicted defect category, and uploading the diagnosis sample to the cloud.

[0016] Optionally, the process of updating the parameters of the defect diagnosis model of the edge node by simulated gradient is represented as:

[0017] ;

[0018] wherein, ;

[0019]

[0020] wherein, a defect diagnosis model of an edge node in a medium time scale loss function; a defect diagnosis model of an edge node in a medium time scale uploaded sample subset; cross-entropy loss; regularization coefficient; a defect diagnosis model of an edge node in a medium time scale under the parameter; initial parameters of the defect diagnosis model of an edge node; learning rate; parameters of the defect diagnosis model updated by the simulated gradient; true label obtained by the operation and inspection feedback; the first defect category probability distribution; diagnostic confidence; operation state data; parameters of the judgment model gradient operator of the parameters

[0021] Optionally, the cloud in step S3 comprises:

[0022] S31, based on a low confidence threshold and a high confidence threshold, the second sub-diagnosis sample received within a preset large time scale is divided into a low confidence sample set and a high confidence correct sample set;

[0023] S32, according to the similarity of the operation state features in the low confidence sample set, a high confidence correct sample is matched for each low confidence sample to obtain a matched sample pair;

[0024] S33, the edge nodes corresponding to the matched sample pair are obtained and combined with the matched sample pair to determine the conduction relationship between the edge nodes, and a multi-edge empirical conduction set is constructed according to the conduction relationship of each edge node.

[0025] Optionally, the cloud in step S4 comprises:

[0026] S41, based on the multi-edge empirical conduction set, an edge model subset participating in knowledge conduction is extracted, and the aggregation weight of each edge model in the multi-edge empirical conduction in the edge model subset is calculated;

[0027] ​S42. For each edge model, knowledge distillation optimization is performed based on the multilateral experience transmission set to obtain the optimized edge model parameters.

[0028] S43. Combining the aggregated weights, the edge model parameters are aggregated into a global optimization model and distributed to the edge nodes.

[0029] Optionally, the edge nodes in step S5 include:

[0030] S51. The real-time collected operating status data are diagnosed using the defect diagnosis model and the global optimization model respectively, to obtain the second defect category probability distribution and the third defect category probability distribution.

[0031] S52. Calculate the difference measure of the model diagnostic distribution based on the probability distribution of the second defect category and the probability distribution of the third defect category;

[0032] S53. When the difference measure is greater than the preset threshold, generate comparison samples based on the probability distribution of the second defect category and the probability distribution of the third defect category, and upload them to the cloud.

[0033] Optionally, in step S6, the step of updating the quality assessment function based on the comparison samples to generate the assessment result includes:

[0034] A comparison sample set is generated based on the comparison samples of each edge node, and a model update quality assessment function is constructed by combining the real labels of operation and maintenance feedback over a large time scale. The assessment result is generated through the model update quality assessment function.

[0035] The model update quality evaluation function is expressed as follows:

[0036] ;

[0037] In the formula, For indicator functions, The number of edge nodes. For a large time scale For edge nodes The updated defect diagnosis model compared to the sample The results of the diagnosis For edge nodes The defect diagnosis model before the update compared to the sample Results of the diagnosis This provides a true label for operational and maintenance feedback over a large timescale.

[0038] A second aspect of the present invention provides a defect diagnosis system for power distribution network electrical equipment, the system comprising:

[0039] Unit 1: Edge nodes use the defect diagnosis model issued by the cloud to diagnose the operating status data of local power distribution network electrical equipment within a preset small time scale, obtain the probability distribution of the first defect category, generate diagnostic samples, and then upload them to the cloud.

[0040] The second unit: The parameters of the defect diagnosis model of the edge node are simulated and gradient updated by combining the real labels of operation and maintenance feedback and the first sub-diagnostic samples received within the preset medium time scale in the cloud.

[0041] Unit 3: The cloud-based system divides the second sub-diagnostic samples received within a preset large time scale based on low-confidence and high-confidence thresholds, and generates a multilateral experience transmission set based on the division results.

[0042] Unit 4: The cloud performs federated aggregation optimization on the parameters of the defect diagnosis model after the simulated gradient update, and generates a global optimization model by combining the knowledge distillation in the multilateral experience transmission set, and distributes it to the edge nodes;

[0043] Unit 5: Edge nodes acquire the difference measure between the defect diagnosis model and the global optimization model. When the difference measure is greater than a preset threshold, a comparison sample is constructed and uploaded to the cloud.

[0044] Unit 6: The cloud constructs a model based on the comparison samples to update the quality assessment function and generate assessment results. When the assessment result is not greater than 0, the low confidence threshold and the high confidence threshold are updated and the third unit is triggered. When the assessment result is greater than 0, the global optimization model is used as the final defect diagnosis model and sent to the edge nodes for diagnosis of the running status data.

[0045] A third aspect of the present invention provides a defect diagnosis device for power distribution network electrical equipment, the device comprising a processor and a memory:

[0046] The memory is used to store program code and transmit the program code to the processor;

[0047] The processor is used to execute the steps of the defect diagnosis method for distribution network electrical equipment as described in the first aspect above, according to the instructions in the program code.

[0048] A fourth aspect of the present invention provides a computer-readable storage medium for storing program code for executing the defect diagnosis method for distribution network electrical equipment described in the first aspect above.

[0049] As can be seen from the above technical solutions, the present invention has the following advantages:

[0050] This invention provides a defect diagnosis method for distribution network electrical equipment. Firstly, it designs a defect diagnosis method for distribution network electrical equipment based on multilateral experience transmission (including steps S1 to S4). Firstly, edge nodes utilize a defect diagnosis model distributed from the cloud to identify and diagnose defects in multi-source operating status data of local distribution network electrical equipment, extracting features and forming diagnostic samples which are then uploaded to the cloud. Secondly, the cloud collects these diagnostic samples from edge nodes over a medium timescale and, combined with real labels obtained from actual operation and maintenance, performs independent simulated gradient updates on the defect diagnosis model of each edge node, allowing it to initially adapt to its local scenario. This simulated update avoids the negative impact of heterogeneous data aliasing. Next, on a larger timescale, the cloud intelligently filters low-confidence samples and high-confidence correct samples from all uploaded diagnostic samples. By calculating the similarity of operating status features, it matches each low-confidence sample with an optimal high-confidence sample. Based on these matching relationships, the cloud performs knowledge distillation optimization on the edge models involved in the transmission, and then integrates these optimized model parameters into a global optimized model through federated aggregation, which is then distributed to each edge node to update its local diagnostic capabilities. Through the transfer of high-confidence knowledge across the edge, low-confidence samples gain learning opportunities, thereby significantly improving the accuracy, robustness, and collaborative efficiency of the overall diagnosis, making the diagnosis of defects in power distribution network electrical equipment more intelligent and reliable.

[0051] On the other hand, an optimization mechanism for updating the defect diagnosis model of distribution network electrical equipment based on multilateral feedback was designed (including steps S5 to S6). First, after a new globally optimized model is deployed to the cloud, edge nodes do not immediately replace the original model. Instead, they simultaneously run both the pre- and post-update diagnostic models to diagnose the newly collected data. When there is a significant difference in the diagnostic results of the two models, the edge nodes package the operating status characteristics and the diagnostic outputs of the two models into a comparison sample and upload it to the cloud. Second, the cloud collects comparison samples from multiple edge nodes and, combined with real labels obtained from actual operation and maintenance, constructs an evaluation function that quantifies the quality of the model update. This evaluation function reflects the change in diagnostic effect brought about by the model update. Finally, the cloud uses multilateral feedback information to adaptively adjust the high and low confidence thresholds in the model update method. By comparing the diagnostic performance before and after the model update in real time and integrating multilateral feedback, the system can self-evaluate the update effect and dynamically adjust the optimization strategy, achieving intelligent and automated model update. This not only improves the continuous optimization capability of the diagnostic method but also ensures that the model maintains optimal diagnostic performance and adaptability in complex and ever-changing real-world application scenarios. Attached Figure Description

[0052] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0053] Figure 1 A flowchart illustrating a defect diagnosis method for power distribution network electrical equipment provided in an embodiment of the present invention;

[0054] Figure 2 This is a flowchart illustrating a defect diagnosis system for power distribution network electrical equipment, provided as an embodiment of the present invention. Detailed Implementation

[0055] To make the objectives, features, and advantages of this invention more apparent and understandable, the technical solutions of the embodiments of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the embodiments described below are only some embodiments of this invention, and not all embodiments. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention.

[0056] Please see Figure 1 The present invention provides a method for diagnosing defects in power distribution network electrical equipment, comprising:

[0057] It should be noted that, addressing the issues of insufficient generalization ability of edge device models and low diagnostic accuracy for low-confidence samples in the defect diagnosis of distribution network electrical equipment, this invention proposes a diagnostic method based on multilateral experience transfer, as detailed in steps 101 to 104. This method leverages the high-confidence diagnostic experience of multiple edge nodes in the cloud to provide cross-edge knowledge transfer for low-confidence samples, achieving collaborative optimization of edge models and thereby improving the accuracy and generalization ability of distribution network electrical equipment defect diagnosis.

[0058] Step 101: The edge node diagnoses the operating status data of the local power distribution network electrical equipment within a preset small time scale using the defect diagnosis model issued by the cloud, obtains the probability distribution of the first defect category, generates diagnostic samples, and uploads them to the cloud.

[0059] In one embodiment, the edge node in step 101 performs the following steps, including:

[0060] Step 1011: Extract features from the operational status data within a small time scale to obtain operational status features;

[0061] It should be noted that this invention defines small, medium, and large time scales. Small time scale refers to the time granularity of real-time data acquisition and inference at edge nodes; this invention uses the second level for rapid response and generation of preliminary diagnostic samples. Medium time scale refers to the time granularity of simulated gradient updates and preliminary aggregation in the cloud, encompassing multiple small time scales; this invention uses the minute level to adapt to changes in local data distribution at the edge. Large time scale refers to the time granularity of multilateral experience transfer, global federated aggregation, and parameter tuning in the cloud, encompassing multiple medium time scales; this invention uses the hour level to complete cross-node knowledge transfer and model version iteration.

[0062] Step 1012: Input the running status features into the defect diagnosis model issued by the cloud, and output the identification results including: the probability distribution of the first defect category and the diagnostic confidence, and generate the predicted defect category based on the probability distribution of the defect category;

[0063] It should be noted that the first step is to implement localized execution of the pre-trained model and data upload at the edge nodes. This is done at each small time scale (time step). ), edge nodes Receive and process newly acquired operating status data of power distribution equipment These data are from multiple sources, including sensor data such as images, partial discharge acoustic signatures, and electrical parameters. Edge nodes. Run the defect diagnosis model distributed from the cloud. ,right Defect identification and diagnosis are performed. The specific process is as follows: First, through a feature extraction function... Extract from raw data Dimensional operating state characteristics Subsequently, the defect diagnosis model outputs the probability distribution of the first defect category based on these features. and diagnostic confidence , represented as

[0064] (1)

[0065] In the formula, For small timescale indexing; For medium-time scale indexing; For feature extraction functions; for 3D operating state feature vector; Defect diagnosis model for edge nodes In the medium timescale The following parameters; The probability distribution of the first defect category ( (Probability vector of class defects); For diagnostic confidence level.

[0066] Step 1013: Combine the operating status characteristics, identification results and predicted defect categories to construct diagnostic samples and upload them to the cloud.

[0067] It should be noted that edge nodes The operational status characteristics at the current small time scale and recognition results (including) , and predicting defect categories Organize into a diagnostic sample And upload it to the cloud.

[0068] Step 102: The cloud combines the real labels from operation and maintenance feedback with the first sub-diagnostic samples received within a preset medium time scale to simulate gradient updates of the parameters of the defect diagnosis model for edge nodes.

[0069] It should be noted that mixing the experience of all edge nodes into a single global model for learning can lead to inefficiency and decreased accuracy, as different edge nodes may face significant differences in the operating status of distribution network equipment, environmental noise, and defect patterns. The cloud-based defect diagnosis model, by simulating and updating the defect diagnosis model of each edge node at a mid-timescale, allows the edge node's defect diagnosis model to learn and adapt more focusedly to defect states similar to its own scenario, effectively improving diagnostic accuracy and robustness.

[0070] Cloud in the medium time scale Internally, based on from A sample set of edge nodes (i.e., the first subsample set); where medium time scale Includes small time steps, combined with real tags obtained from operation and maintenance feedback. (The method for obtaining real tags could be: after handling a defect, on-site maintenance personnel enter the diagnosed defect type and handling result into the work order management system via a handheld terminal. The cloud periodically retrieves archived defect elimination work orders from the work order management system and extracts the fault cause field as the real tag.) For edge nodes... Defect diagnosis model In the medium timescale Parameters on Simulated gradient updates are performed to allow the model to initially adapt to the local feature distribution. This process is represented as:

[0071] (2)

[0072] (3)

[0073] In the formula, Defect diagnosis model for edge nodes In the medium timescale The loss function; Cross-entropy loss; For edge nodes In the medium timescale Uploaded sample subset; The regularization coefficient is used. These are the initial parameters for the defect diagnosis model of the edge nodes; The learning rate; The parameters are used to simulate the defect diagnosis model after gradient update.

[0074] Step 103: The cloud platform divides the second sub-diagnostic samples received within a preset large time scale based on low confidence threshold and high confidence threshold, and generates a multilateral experience transmission set based on the division results.

[0075] In one embodiment, step 103 in the cloud performs the following steps, including:

[0076] Step 1031: Based on the low confidence threshold and the high confidence threshold, the second sub-diagnostic samples received within a preset large time scale are divided into a low confidence sample set and a high confidence correct sample set.

[0077] It should be noted that on a large time scale The cloud is based on all edge nodes in the current large time scale. All uploaded sample sets (i.e., the second sample set), based on diagnostic confidence. Based on prediction accuracy, all sample sets are divided into low-confidence samples and high-confidence correct samples. Specifically, when the sample confidence is below a threshold... When the confidence level of a sample is higher than the threshold, it is classified as a low-confidence sample; when the confidence level of a sample is higher than the threshold, it is classified as a low-confidence sample. When the predicted result matches the true label, it is classified as a high-confidence correct sample. These samples are organized into the following two sets, denoted as follows:

[0078] (4)

[0079] In the formula, For large time-scale indexing; For the first A low-confidence sample, satisfying ; For the first Diagnostic confidence of a low-confidence sample; For the first A number of high-confidence correct samples, satisfying ; For the first Diagnostic confidence of a high-confidence correct sample; The logical symbol "AND"; For the first Predicted defect category labels for a high-confidence, correctly identified sample; For the first The true defect category label of a high-confidence correct sample; and These represent the number of samples in the corresponding set; The low confidence threshold; This is the high confidence threshold.

[0080] Step 1032: Match a high-confidence correct sample to each low-confidence sample based on the similarity of the running status features in the low-confidence sample set, and obtain a matching sample pair;

[0081] It should be noted that, next, the cloud will calculate the similarity of the running status features for each low-confidence sample. Match an optimal high-confidence correct sample The matching process is represented as follows:

[0082] (5)

[0083] In the formula, Low confidence samples The corresponding running state feature vector; High-confidence correct samples The corresponding running state feature vector; The similarity metric function combines cosine similarity and Gaussian kernel. For kernel width parameters; for Norm.

[0084] To form a multilateral experience transfer set, each matching sample pair is recorded in the cloud. And its corresponding edge nodes. This invention constructs a sample model localization function. ,in yes The source edge node, yes The source edge node. Based on this localization function, a multilateral empirical transmission localization function is obtained. (When a match occurs).

[0085] Step 1033: Obtain the edge nodes corresponding to the matching sample pairs and combine the matching sample pairs to determine the transmission relationship between the edge nodes. Construct a multilateral empirical transmission set based on the transmission relationship of each edge node.

[0086] It should be noted that the multilateral experience transmission set Edge nodes are derived from all low-confidence samples that meet the matching criteria. and high-confidence correct sample source edge nodes The system is composed of components to capture the transmission relationships between edge nodes, thereby obtaining a multilateral experience transmission set. Represented as:

[0087] (6)

[0088] Step 104: The cloud performs federated aggregation optimization on the parameters of the defect diagnosis model after the simulated gradient update, and generates a global optimization model by combining the knowledge distillation in the multilateral experience transmission set, and distributes it to the edge nodes.

[0089] It should be noted that the time scale is in the cloud. Parameters of the updated defect diagnosis model after simulation Perform federated aggregation optimization and combine it with a multilateral experience transfer set. Knowledge distillation is performed to ultimately form a global optimization model.

[0090] In one embodiment, step 104 in the cloud performs the following steps, including:

[0091] Step 1041: Based on the multilateral experience transmission set, extract all marginal model subsets participating in knowledge transmission, and calculate the aggregation weight of each marginal model in the multilateral experience transmission.

[0092] It should be noted that, firstly, based on the multilateral experience transmission set The cloud first extracts a subset of all edge models involved in knowledge transfer. For each edge model in the subset Calculate its aggregate weight in multilateral empirical transmission. This weight takes into account both the importance of the sample as a source of knowledge (high confidence) and a recipient of knowledge (low confidence). The weight calculation is based on all relevant sample pairs, not just the matching relationships between nodes; therefore, the original sample index is used, represented as...

[0093] (7)

[0094] In the formula, This is a normalization constant; For indicator functions; Confidence of matching samples as a source of knowledge; These are the weighting coefficients for low-confidence samples.

[0095] Step 1042: For each edge model, perform knowledge distillation optimization based on the multilateral experience transmission set to obtain the optimized edge model parameters;

[0096] It should be noted that, next, for each edge model Based on experience transmission set Knowledge distillation (KD) optimization is performed, and the optimized edge model parameters are... Represented as

[0097] (8)

[0098] In the formula, medium time scale Simulate the updated parameters; To optimize the learning rate; For edge model In the medium timescale Simulate the updated parameters; In order to be with the first The first low-confidence sample matched The predicted probability distribution of a high-confidence correct sample; For knowledge distillation loss, such as measuring student models based on KL divergence. Teacher Model The degree of fit of the prediction.

[0099] Step 1043: Combine the aggregated weights to aggregate the edge model parameters into a global optimization model and distribute it to the edge nodes.

[0100] It should be noted that, finally, the cloud performs federated aggregation, combining these optimized edge model parameters into a global optimized model. , represented as

[0101] (9)

[0102] In the formula, These are the aggregated global model parameters. These are the optimized model parameters. The cloud will then globally optimize this model. The data is distributed to each edge node to update the local diagnostic model for application in the next medium timescale.

[0103] Step 105: Obtain the difference measure between the defect diagnosis model and the global optimization model at the edge node. When the difference measure is greater than the preset threshold, construct a comparison sample and upload it to the cloud.

[0104] It should be noted that, to address the problem that parameter configuration during model updates relies on manual tuning and is difficult to adapt to dynamic scenarios, this invention proposes an adaptive parameter optimization mechanism based on multilateral feedback. This mechanism compares the diagnostic differences between the model before and after optimization, and constructs parameter optimization objectives by combining operational feedback, thereby achieving adaptive adjustment of the update method and ensuring that the model maintains optimal performance in complex and ever-changing real-world applications. Please refer to steps 105 to 106.

[0105] In one embodiment, the edge node in step 104 performs the following steps, including:

[0106] Step 1051: Diagnose the real-time collected operating status data using the defect diagnosis model and the global optimization model respectively, and obtain the probability distribution of the second defect category and the probability distribution of the third defect category;

[0107] It should be noted that when the cloud performs a global optimization model... Distribute to edge nodes Afterwards, the edge node will not immediately replace the original model; it will simultaneously run the model before the update. (i.e., defect diagnosis model) and updated model (i.e., the global optimization model). For newly collected operational status data... (in (For the new time step index after model update), both models perform diagnosis simultaneously, outputting their respective defect category probability distributions. and , respectively, are represented as (the probability distribution of the second defect category).

[0108] (10)

[0109] In the formula, The model parameters before local updates at the edge nodes (corresponding to the previous large time scale) ); The updated parameters issued this time ( ).

[0110] Step 1052: Calculate the difference measure of the model diagnostic distribution based on the probability distribution of the second defect category and the probability distribution of the third defect category;

[0111] It should be noted that edge nodes are used to calculate the difference in diagnostic distributions between the two models. Using KL divergence, it is expressed as:

[0112] (11)

[0113] In the formula, For the first The probability of a class of defects; Total number of defect categories; and They are edge nodes No. The first and second recognitions are updated continuously. Time of the first The probability of a class of defects.

[0114] Step 1053: When the difference measure is greater than the preset threshold, generate comparison samples based on the probability distribution of the second defect category and the probability distribution of the third defect category, and upload them to the cloud.

[0115] It should be noted that when measuring differences... Exceeding the preset threshold When the update of the defect diagnosis model has significantly impacted the diagnostic results of the sample, cloud monitoring is required. The edge node packages the runtime features and the diagnostic outputs of the two models into a comparison sample. Specifically, the edge node constructs a comparison sample. The generated comparison sample set Uploaded to the cloud.

[0116] Step 106: The cloud builds a model based on the comparison samples and updates the quality assessment function to generate assessment results. When the assessment result is not greater than 0, the low confidence threshold and high confidence threshold are updated and the process returns to step 103. When the assessment result is greater than 0, the global optimization model is used as the final defect diagnosis model and distributed to the edge nodes for diagnosis of the running status data.

[0117] In one embodiment, step S6, constructing a model update quality assessment function based on the comparison samples to generate assessment results, includes: generating a comparison sample set based on the comparison samples of each edge node, constructing a model update quality assessment function by combining the real labels of operation and maintenance feedback over a large time scale, and generating assessment results through the model update quality assessment function.

[0118] It should be noted that the data collected in the cloud comes from... Comparison sample set of edge nodes Combined with real labels from operation and maintenance feedback We construct a parameter optimization objective. To evaluate the quality of model updates, we define a model update quality evaluation function. ,in The low-confidence threshold and high-confidence threshold involved in step 103 are represented as follows:

[0119] (12)

[0120] In the formula, This is an indicator function used to determine the accuracy of predictions; it provides quantitative metrics when the model's performance improves or deteriorates after an update. The cloud platform relies on this. The comprehensive evaluation results, the evaluation process can be understood as: calculating based on the comparison sample set This value represents the difference between the number of correctly predicted predictions by the new model and the number of correctly predicted predictions by the old model, reflecting the effectiveness of the model update. Then, the two thresholds are updated based on the numerical results. Specifically, if... This indicates that the model update has a good overall effect across multiple edge nodes, and no adjustment of the current parameters is required. ,if This indicates that the effect of this model update is poor or stagnant, and the parameters need to be adjusted. Adjustments will be made in hopes of achieving better results in the next round.

[0121] The specific adjustment amount can be set to a fixed step size. , represented as

[0122] (13)

[0123] In the formula, , , and These are the high and low confidence thresholds before and after the update, respectively.

[0124] The cloud will optimize the parameters. Applied to the next large timescale The 103 process enables adaptive improvement and optimization of diagnostic methods.

[0125] The above is a defect diagnosis method for distribution network electrical equipment provided in the embodiments of the present invention. The following is a defect diagnosis system for distribution network electrical equipment provided in the embodiments of the present invention.

[0126] Please see Figure 2 The present invention provides a defect diagnosis system for distribution network electrical equipment, comprising:

[0127] Unit 1 201: Edge nodes use the defect diagnosis model issued by the cloud to diagnose the operating status data of local power distribution network electrical equipment within a preset small time scale, obtain the probability distribution of the first defect category, generate diagnostic samples, and then upload them to the cloud.

[0128] Unit 202: The cloud combines the real labels from operation and maintenance feedback with the first sub-diagnostic samples received within a preset medium time scale to simulate gradient updates of the parameters of the defect diagnosis model for edge nodes.

[0129] Unit 3 203: The cloud-based system divides the second sub-diagnostic samples received within a preset large time scale based on low confidence thresholds and high confidence thresholds, and generates a multilateral experience transmission set based on the division results.

[0130] Unit 4, Question 204: The cloud performs federated aggregation optimization on the parameters of the defect diagnosis model after the simulated gradient update, and generates a global optimization model by combining the knowledge distillation in the multilateral experience transmission set, and distributes it to the edge nodes;

[0131] Unit 5, Question 205: Obtain the difference measurement between the defect diagnosis model and the global optimization model at the edge node. When the difference measurement is greater than the preset threshold, construct a comparison sample and upload it to the cloud.

[0132] Unit 6, 206: The cloud builds a model based on the comparison samples and updates the quality assessment function to generate assessment results. When the assessment result is not greater than 0, the low confidence threshold and high confidence threshold are updated and Unit 3, 203, is triggered. When the assessment result is greater than 0, the global optimization model is used as the final defect diagnosis model and distributed to the edge nodes for diagnosis of the running status data.

[0133] Furthermore, this embodiment of the invention also provides a defect diagnosis device for power distribution network electrical equipment, the device including a processor and a memory:

[0134] The memory is used to store program code and transmit the program code to the processor;

[0135] The processor is used to execute the steps of the defect diagnosis method for distribution network electrical equipment as described in the above method embodiments, according to the instructions in the program code.

[0136] Furthermore, this embodiment of the invention also provides a computer-readable storage medium for storing program code, which is used to execute the defect diagnosis method for distribution network electrical equipment described in the above method embodiments.

[0137] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0138] In the embodiments provided by this invention, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces, indirect coupling or communication connection between apparatuses or units, and may be electrical, mechanical, or other forms.

[0139] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0140] Furthermore, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0141] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0142] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A defect diagnosis method of a distribution network electrical device, characterized by, Comprise: S1, the edge node diagnoses the operation state data of the local distribution network electrical equipment in a preset small time scale by the defect diagnosis model issued by the cloud, obtains the first defect category probability distribution and generates a diagnosis sample, and then uploads to the cloud; S2, the cloud updates the parameters of the defect diagnosis model of the edge node by simulated gradient combining the real label fed back by the operation and inspection and the first sub-diagnosis sample received in a preset medium time scale; S3, the cloud divides the second sub-diagnosis sample received in a preset large time scale based on a low confidence threshold and a high confidence threshold, and generates a multi-edge experience conduction set according to the division result; S4, the cloud performs federal aggregation optimization on the parameters of the defect diagnosis model updated by simulated gradient, and generates a global optimization model combining the knowledge distillation in the multi-edge experience conduction set, and issues to the edge node; S5, the edge node obtains the difference measure of the defect diagnosis model and the global optimization model, and when the difference measure is greater than a preset threshold, constructs a comparison sample and uploads to the cloud; S6, the cloud constructs a model update quality evaluation function according to the comparison sample to generate an evaluation result, and when the evaluation result is not greater than 0, updates the low confidence threshold and the high confidence threshold and returns to step S3, and when the evaluation result is greater than 0, takes the global optimization model as the final defect diagnosis model and issues to the edge node for diagnosing the operation state data.

2. The defect diagnosis method of the power distribution electrical equipment according to claim 1, characterized by, The edge node of step S1 comprises: S11, feature extraction is performed on the operation state data in a small time scale to obtain operation state features; S12, input the operation state features into the defect diagnosis model issued by the cloud, output the recognition result including: first defect category probability distribution and diagnosis confidence, and generate predicted defect category according to defect category probability distribution; S13, combine the operation state features, the recognition result and the predicted defect category to construct a diagnosis sample and upload to the cloud.

3. The defect diagnosis method of the power distribution electrical equipment according to claim 2, characterized by, The process of simulated gradient updating the parameters of the defect diagnosis model of the edge node is represented as: ； wherein ; In the formula, Defect diagnosis model for edge nodes In the medium timescale The loss function; For edge nodes In the medium timescale Uploaded sample subset; Cross-entropy loss; The regularization coefficient is used. For defect diagnosis model In the medium timescale The following parameters; These are the initial parameters for the defect diagnosis model of the edge nodes; The learning rate; The parameters of the defect diagnosis model after the gradient update are used to simulate the process. The actual label obtained from the operation and maintenance feedback; This represents the probability distribution of the first defect category; The confidence level of the diagnosis; This is runtime status data; To judge the model parameters The gradient operator.

4. The network electrical equipment defect diagnosis method according to claim 3, wherein The cloud of step S3 comprises: S31, based on the low confidence threshold and the high confidence threshold, the second sub-diagnosis sample received in a preset large time scale is divided into a low confidence sample set and a high confidence correct sample set; S32, according to the similarity of the operation state features in the low confidence sample set, each low confidence sample is matched with a high confidence correct sample to obtain a matched sample pair; S33, obtain the corresponding edge node of the matched sample pair and combine the matched sample pair to determine the conduction relationship between the edge nodes, and construct a multi-edge experience conduction set according to the conduction relationship of each edge node.

5. The network electrical equipment defect diagnosis method according to claim 4, wherein The cloud of step S4 comprises: S41, based on the multi-edge experience conduction set, extract all edge model subsets participating in knowledge conduction, and calculate the aggregation weight of each edge model in the multi-edge experience conduction in the edge model subset; S42, for each edge model, perform knowledge distillation optimization based on the multi-edge experience conduction set to obtain the optimized edge model parameters; S43, combine the aggregated weights to aggregate the edge model parameters into a global optimization model, and distribute to the edge nodes.

6. The network electrical equipment defect diagnosis method according to claim 5, wherein The edge node of step S5 includes: S51, respectively diagnose the real-time collected operation state data by the defect diagnosis model and the global optimization model to obtain a second defect category probability distribution and a third defect category probability distribution; S52, calculate a difference measure of the model diagnosis distribution according to the second defect category probability distribution and the third defect category probability distribution; S53, when the difference measure is greater than a preset threshold, generate a comparison sample according to the second defect category probability distribution and the third defect category probability distribution and upload to the cloud.

7. The network electrical equipment defect diagnosis method according to claim 6, wherein In step S6, the model update quality evaluation function is constructed according to the comparison sample to generate an evaluation result, which includes: Generate a comparison sample set according to the comparison samples of each edge node, and construct a model update quality evaluation function combined with the real labels of the operation inspection feedback in a large time scale, and generate an evaluation result through the model update quality evaluation function; The model update quality evaluation function is expressed as: ； In the formula, is an indicator function, is the number of edge nodes, is a large time scale, is an edge node The updated defect diagnosis model diagnoses the contrast sample The result of diagnosis, is an edge node The updated defect diagnosis model diagnoses the contrast sample The result of diagnosis, is a true label of the operation and inspection feedback in a large time scale.

8. A defect diagnosis system of a distribution network electrical equipment, characterized by, It includes: The first unit: the edge node diagnoses the operation state data of the local distribution network electrical equipment in a preset small time scale by the defect diagnosis model distributed by the cloud, obtains a first defect category probability distribution, generates a diagnosis sample, and uploads to the cloud; The second unit: the cloud updates the parameters of the defect diagnosis model of the edge node by simulated gradient combined with the real labels of the operation inspection feedback and the first sub-diagnosis sample received in a preset medium time scale; The third unit: the cloud divides the second sub-diagnosis sample received in a preset large time scale based on a low confidence threshold and a high confidence threshold, and generates a multi-edge experience conduction set according to the division result; The fourth unit: the cloud performs federal aggregation optimization on the parameters of the defect diagnosis model updated by simulated gradient, and generates a global optimization model combined with the knowledge distillation in the multi-edge experience conduction set, and distributes to the edge node; The fifth unit: the edge node obtains the difference measure of the defect diagnosis model and the global optimization model, and when the difference measure is greater than a preset threshold, constructs a comparison sample and uploads to the cloud; The sixth unit: the cloud constructs a model update quality evaluation function according to the comparison sample to generate an evaluation result, and when the evaluation result is not greater than 0, updates the low confidence threshold and the high confidence threshold to trigger the third unit, and when the evaluation result is greater than 0, takes the global optimization model as the final defect diagnosis model and distributes to the edge node for diagnosing the operation state data.

9. A defect diagnosis device of a power distribution electrical equipment, characterized by, The device includes a processor and a memory: The memory is used to store program code and transmit the program code to the processor; The processor is used to execute the defect diagnosis method of the distribution network electrical equipment according to the instructions in the program code.

10. A computer-readable storage medium, characterized in that, The computer readable storage medium is used to store program code, and the program code is used to execute the defect diagnosis method of the distribution network electrical equipment. The computer readable storage medium is used to store program code, and the program code is used to execute the defect diagnosis method of the distribution network electrical equipment.

Images