A power grid fault data processing method, device and medium

Abstract

The application discloses a kind of power grid fault data processing method, equipment and medium, the method includes the detection data acquisition of multi-dimensional to target power grid obtains power grid image and sensing parameter, and the environmental parameter in the collection target power grid environment;The historical detection data of target power grid is obtained and the fault sample proportion coefficient is extracted, the historical detection data is combined and divided and is integrated image fault identification branch and integrated sensing fault identification branch training is carried out;According to environmental parameter, obtain the influence coefficient of failure rate, image influence coefficient and sensing influence coefficient are carried out failure rate influence analysis and detection data influence analysis;According to failure rate, image and sensing influence coefficient, obtain the number of image identification branch and sensing identification branch number are calculated, and the number of image identification branch and sensing identification branch are obtained, and the technical problem that machine learning identification failure perception rate and accuracy rate are low is solved, which is caused by the few fault data samples of power grid line inspection.

Description

Technical Field

[0001] This invention belongs to the field of data processing technology, and in particular relates to a method, device and medium for processing power grid fault data. Background Technology

[0002] As modern society becomes increasingly reliant on electricity, the safe and stable operation of power systems is of paramount importance. As the core infrastructure for carrying, transmitting, and distributing electrical energy, the power grid comprises a vast number of widely distributed devices that operate under complex and ever-changing natural and operational conditions. These factors inevitably expose power grid equipment to various potential fault risks, such as insulation aging, loose connections, external damage, or the impact of extreme weather. Therefore, real-time and accurate monitoring and fault diagnosis of the power grid, and timely detection and handling of potential hazards, are of core value in ensuring power supply reliability, reducing economic losses, and improving operation and maintenance efficiency.

[0003] Existing technologies still have some problems in practical applications: First, power grid fault data has typical small sample characteristics, that is, the number of fault samples in historical data is far less than the number of samples in normal operation. This data imbalance problem seriously affects the learning effect and generalization ability of many intelligent algorithms, and is prone to missing the detection of rare faults.

[0004] Therefore, how to improve the model's ability to perceive abnormal features in the context of sparse fault samples, thereby improving the accuracy of power grid fault identification, has become an urgent problem to be solved. Summary of the Invention

[0005] To address the problems existing in the prior art, this invention proposes a power grid fault data processing method, equipment, and medium.

[0006] The technical solution of the present invention is as follows:

[0007] A method for processing power grid fault data, the method comprising:

[0008] Multidimensional detection data is collected from the target power grid to obtain power grid images and sensing parameters, and environmental parameters within the target power grid environment are also collected.

[0009] Historical detection data of the target power grid is obtained, and the fault sample proportion coefficient is extracted. The historical detection data is combined and divided, and the integrated image fault recognition branch and the integrated sensor fault recognition branch are trained. The branch fault sample proportion coefficient in the branch training data of each fault recognition branch is a preset fault sample proportion coefficient.

[0010] Based on the environmental parameters, failure rate impact analysis and detection data impact analysis are performed to obtain failure rate impact coefficient, image impact coefficient and sensing impact coefficient;

[0011] Based on the failure rate influence coefficient, image influence coefficient, and sensor influence coefficient, the number of image recognition branches and sensor recognition branches are calculated, and the fault recognition branches are invoked and faults are identified to obtain the power grid fault identification results.

[0012] Furthermore, the specific method for acquiring multi-dimensional detection data of the target power grid to obtain power grid images and sensing parameters, and for acquiring environmental parameters within the target power grid environment, includes:

[0013] Multidimensional detection data is collected from the target power grid to obtain power grid images and sensing parameters;

[0014] Environmental parameters within the target power grid environment are collected, including meteorological parameters, temperature parameters, and humidity parameters.

[0015] Furthermore, the specific method for obtaining historical detection data of the target power grid and extracting the fault sample proportion coefficient includes:

[0016] Obtain historical detection data of the target power grid and similar power grids within a historical time period;

[0017] Within the historical detection data, a set of sample power grid images and a set of sample sensing parameters are collected, and a corresponding set of sample fault identification results is collected, wherein the sample fault identification results include normal or abnormal.

[0018] The fault identification results of the samples that are selected as abnormal are filtered and the proportion is calculated to obtain the fault sample proportion coefficient.

[0019] Furthermore, the specific method for combining and dividing historical detection data, and training the integrated image fault recognition branch and the integrated sensor fault recognition branch includes:

[0020] The set of sample fault identification results is divided to obtain a set of normal sample fault identification results and a set of abnormal sample fault identification results.

[0021] Calculate the ratio of the preset fault sample proportion coefficient to the fault sample proportion coefficient to obtain the total number of branches K, where K is a positive integer;

[0022] Random selection with replacement is performed within the set of normal fault identification results and the set of abnormal fault identification results to obtain the sample fault identification results. The corresponding sample power grid image and sample sensing parameters are combined to obtain the first image fault identification training data and the first sensing fault identification training data. The proportion of abnormal fault identification results is a preset fault sample proportion coefficient.

[0023] The first image fault recognition training data and the first sensor fault recognition training data are used to train the first image fault recognition branch and the first sensor fault recognition branch.

[0024] Continue to select and combine K image fault recognition training data and K sensor fault recognition training data according to the preset fault sample proportion coefficient, and train to obtain integrated image fault recognition branch and integrated sensor fault recognition branch.

[0025] Furthermore, the specific method for training the first image fault recognition branch and the first sensor fault recognition branch using the first image fault recognition training data and the first sensor fault recognition training data includes:

[0026] Machine learning is used to construct the first image fault recognition branch and the first sensor fault recognition branch, respectively.

[0027] The first image fault recognition training data and the first sensor fault recognition training data are used respectively to conduct supervised training and testing on the first image fault recognition branch and the first sensor fault recognition branch. Training is completed when the error is less than the error threshold.

[0028] Furthermore, the specific methods for performing failure rate impact analysis and detection data impact analysis based on the environmental parameters to obtain the failure rate impact coefficient, image impact coefficient, and sensor impact coefficient include:

[0029] The failure rate impact analyzer is invoked, wherein the failure rate impact analyzer is trained using a sample environmental parameter set and a sample failure rate impact coefficient set;

[0030] The environmental parameters are input into the failure rate impact analyzer, which outputs the failure rate impact coefficient.

[0031] The image influence analysis branch and the sensor influence analysis branch are invoked. The image influence analysis branch is trained using a sample environmental parameter set and a sample image influence coefficient set, and the sensor influence analysis branch is trained using a sample environmental parameter set and a sample sensor influence coefficient set.

[0032] The environmental parameters are input into the image influence analysis branch and the sensor influence analysis branch respectively, and the image influence coefficient and sensor influence coefficient are obtained by identification output. The image influence coefficient includes the influence magnitude of image quality, and the sensor influence coefficient includes the influence magnitude of sensor parameter changes.

[0033] Furthermore, the specific method for calculating the number of image recognition branches and the number of sensor recognition branches based on the fault rate influence coefficient, image influence coefficient, and sensor influence coefficient, and then calling the fault recognition branches and performing fault recognition to obtain the power grid fault recognition result includes:

[0034] Based on the failure rate influence coefficient and the image influence coefficient, the image fault recognition coefficient is calculated.

[0035] The number of image recognition branches is obtained by multiplying the image fault recognition coefficient by the preset number of branches and rounding down, wherein the number of image recognition branches is greater than or equal to 1 and less than or equal to K, and K is the total number of branches;

[0036] Based on the failure rate influence coefficient and the sensing influence coefficient, the sensing fault identification coefficient is calculated.

[0037] The number of sensor identification branches is obtained by multiplying the sensor fault identification coefficient by the preset number of branches and rounding down.

[0038] Call the image fault identification branches of the specified number of image recognition branches, input the power grid images respectively, and obtain the image power grid fault identification results of the specified number of image recognition branches;

[0039] Call the number of sensor fault identification branches of the specified number of sensor identification branches, input the sensor parameters respectively, and obtain the sensor power grid fault identification results of the specified number of sensor identification branches;

[0040] Filter the power grid fault identification results that appear most frequently.

[0041] Furthermore, the specific method for calculating the image fault recognition coefficient based on the fault rate influence coefficient and the image influence coefficient includes:

[0042] The sum of 1 and the failure rate influence coefficient and the image influence coefficient is used as the image fault identification coefficient.

[0043] An electronic device, comprising:

[0044] Memory, used to store computer software programs;

[0045] A processor is used to read and execute the computer software program, thereby implementing the power grid fault data processing method as described in any of the preceding claims.

[0046] A non-transitory computer-readable storage medium storing a computer software program that, when executed by a processor, implements the power grid fault data processing method as described in any of the preceding claims.

[0047] Compared with the prior art, the present invention has the following beneficial effects:

[0048] This invention provides a method for processing power grid fault data. This method involves multi-dimensional data acquisition from the target power grid, including grid images, sensor parameters, and environmental parameters. It then extracts the proportion coefficient of fault samples by combining historical detection data, trains and integrates image and sensor fault recognition branches, and dynamically adjusts the number of fault recognition branches based on the impact of environmental parameters on fault rate and detection data. This improves the accuracy and efficiency of power grid fault identification. Attached Figure Description

[0049] Figure 1 This is a flowchart illustrating the power grid fault data processing method provided by the present invention.

[0050] Figure 2 This is a schematic diagram of the structure of the electronic device provided by the present invention.

[0051] Figure 3 This is a schematic diagram of the structure of a computer-readable storage medium provided by the present invention.

[0052] Explanation of reference numerals in the attached drawings: electronic device 500, memory 510, processor 520, first computer program 511, computer-readable storage medium 600, second computer program 611. Detailed Implementation

[0053] The present invention will be further illustrated below with reference to the accompanying drawings and specific embodiments. It should be understood that these embodiments are for illustrative purposes only and are not intended to limit the scope of the invention. After reading this invention, any modifications of the invention in various equivalent forms by those skilled in the art will fall within the scope defined by the appended claims.

[0054] Example 1:

[0055] This invention provides a method for processing power grid fault data, such as... Figure 1 As shown, the method of the present invention includes:

[0056] S10. Perform multi-dimensional detection data acquisition on the target power grid to obtain power grid images and sensing parameters, and collect environmental parameters within the target power grid environment;

[0057] For example, in the process of power grid fault data processing, the first step is to collect multi-dimensional detection data from the target power grid. Specifically, the collection work includes acquiring power grid images, which can intuitively display the current state of the power grid, such as the connection status of lines and the operating status of equipment. Simultaneously, sensor parameters are collected. These parameters are obtained through various sensors installed in the power grid, enabling real-time monitoring of key indicators such as current, voltage, and temperature. In addition, to gain a more comprehensive understanding of the power grid's operating environment, environmental parameters within the target power grid environment are also collected, such as meteorological parameters (e.g., wind speed, wind direction, rainfall), temperature parameters, and humidity parameters. These environmental parameters are crucial for analyzing the causes and impacts of power grid faults. For instance, under extreme weather conditions, such as high temperatures, strong winds, or heavy rain, the operating status of the power grid may be affected, leading to an increased fault rate. By collecting these environmental parameters, the correlation between power grid faults and environmental factors can be more accurately determined, providing strong support for subsequent analysis and processing.

[0058] S20. Obtain historical detection data of the target power grid, extract the fault sample proportion coefficient, combine and divide the historical detection data, and train the integrated image fault recognition branch and the integrated sensor fault recognition branch. The branch fault sample proportion coefficient in the branch training data of each fault recognition branch is the preset fault sample proportion coefficient.

[0059] Preferably, historical monitoring data of the target power grid is further obtained, and a fault sample proportion coefficient is extracted from it. Assuming a fault sample proportion coefficient of 1%, this means that the number of fault samples in the historical monitoring data is relatively small, which may pose a challenge to subsequent fault identification. To overcome this problem, the historical monitoring data is combined and divided to better utilize these limited data.

[0060] In the specific combination and partitioning, the sample fault identification results in the historical detection data are divided into a set of normal fault identification results and a set of abnormal fault identification results. This is done to allow for targeted selection of fault samples for subsequent training, improving the model's fault identification capability. Then, based on the ratio of the preset fault sample proportion coefficient to the actual fault sample proportion coefficient, the number of branches K to be trained is calculated. Here, a branch represents a model, and each branch corresponds to an independent model. This ratio reflects how many times the original data needs to be amplified or partitioned to achieve the preset fault sample proportion level. That is, if the preset fault sample proportion coefficient is 10% and the actual fault sample proportion coefficient is 1%, then 10 fault identification branches will be trained. The preset fault sample proportion coefficient can be set based on actual needs.

[0061] Next, samples are selected from both the normal fault identification result set and the abnormal fault identification result set in a random selection process with replacement. These samples are then combined with the corresponding power grid images and sensor parameters to form the first set of image fault identification training data and the first set of sensor fault identification training data. During this process, special attention is paid to controlling the proportion of abnormal fault identification results in the training data, ensuring it equals a preset fault sample proportion coefficient (e.g., 10%). This increases the visibility of fault samples during training. In other words, setting the proportion of fault samples in each training data set to the preset fault sample proportion coefficient (e.g., 10%) significantly increases the proportion of fault sample data in the training data, allowing the model to learn more fault features during training, thereby improving the perception rate and accuracy of fault identification.

[0062] After combining the first set of training data, the same method and preset fault sample ratios are used to randomly select and combine the remaining K-1 sets of image fault recognition training data and sensor fault recognition training data. Each set of training data contains normal and abnormal fault recognition results divided according to a preset ratio, along with their corresponding power grid images and sensor parameters. Finally, these combined and divided training data are used to train the integrated image fault recognition branch and the integrated sensor fault recognition branch, respectively. Because each set of training data has undergone combination and division and adjustment of fault sample ratios, the trained model can more accurately identify fault features in power grid images and sensor parameters, improving the perception rate and accuracy of fault recognition.

[0063] S30. Based on environmental parameters, conduct failure rate impact analysis and detection data impact analysis to obtain failure rate impact coefficient, image impact coefficient and sensor impact coefficient;

[0064] Furthermore, to more accurately assess fault risks and optimize fault identification, it is necessary to conduct fault rate impact analysis and detection data impact analysis based on current environmental parameters. This process first considers environmental parameters such as meteorological conditions (wind speed, temperature, humidity, etc.), geographical location, and surrounding terrain. These parameters may directly affect the operating status of the power grid and the fault incidence rate. Fault rate impact analysis can quantify the magnitude of the impact of these environmental factors on the fault rate, i.e., the fault rate impact coefficient. For example, in areas prone to storms, increased humidity and wind speed may lead to a decrease in the insulation performance of power grid equipment, thereby increasing the fault rate.

[0065] Data impact analysis primarily focuses on the influence of environmental parameters on image quality and sensor parameters. Image impact analysis assesses the extent to which current environmental conditions affect the quality of power grid image capture, outputting an image impact coefficient. For example, in heavily foggy weather, images may become blurry, affecting the accuracy of fault identification. Sensor impact analysis, on the other hand, aims to quantify the measurement errors of sensor parameters caused by environmental factors, i.e., the sensor impact coefficient. For example, high temperatures may cause thermal noise in sensors, leading to increased fluctuations in sensor data and thus affecting the reliability of fault identification.

[0066] By comprehensively considering the failure rate influence coefficient, image influence coefficient, and sensor influence coefficient, we can gain a more comprehensive understanding of the role of environmental factors in power grid fault identification and detection, providing a scientific basis for subsequent fault identification branch calls and fault identification result optimization.

[0067] S40. Based on the failure rate influence coefficient, image influence coefficient, and sensor influence coefficient, calculate the number of image recognition branches and sensor recognition branches, call the fault recognition branches and perform fault recognition to obtain the power grid fault recognition result.

[0068] Specifically, the number of image recognition branches and sensor recognition branches to be called for fault identification is determined based on the previously calculated fault rate impact coefficient, image impact coefficient, and sensor impact coefficient. The core principle is that when the fault rate impact coefficient, image impact coefficient, or sensor impact coefficient is large, it means that the current power grid operating environment is relatively harsh, and the difficulty and fault rate of fault identification may increase accordingly. In order to identify faults more accurately, more identification branches need to be called for parallel processing.

[0069] Specifically, the influence coefficients are used to calculate the image fault identification coefficient and the sensor fault identification coefficient. These two coefficients directly determine the number of image recognition branches and sensor recognition branches to be invoked. For example, if at a certain point in time, due to the influence of severe weather, both the fault rate influence coefficient and the image influence coefficient increase significantly, then the image fault identification coefficient calculated based on these two coefficients will also increase accordingly, leading to the need to invoke more image recognition branches. Similarly, if the sensor influence coefficient is large, the number of sensor recognition branches will also increase.

[0070] After invoking the appropriate number of identification branches, the power grid image and sensor parameters are input into these branches for fault identification. Since each branch is based on different training data and algorithms, they may produce different fault identification results. To obtain the final power grid fault identification result, a majority decision principle can be adopted, selecting the most frequently occurring identification result as the final result (normal or abnormal). This method fully utilizes the parallel processing capabilities of multiple identification branches, improving the accuracy and reliability of fault identification.

[0071] In a preferred embodiment, multi-dimensional detection data acquisition is performed on the target power grid to obtain power grid images and sensing parameters, and environmental parameters within the target power grid environment are collected, including: multi-dimensional detection data acquisition of the target power grid to obtain power grid images and sensing parameters; and collection of environmental parameters within the target power grid environment, wherein the environmental parameters include meteorological parameters, temperature parameters, and humidity parameters.

[0072] Specifically, image acquisition devices (such as high-definition cameras and infrared thermal imagers) deployed at key nodes of the power grid work in concert with sensing devices (such as current / voltage transformers, vibration sensors, and partial discharge monitors) to collect real-time image data and electrical / mechanical parameters of power grid equipment. For example, infrared thermal imaging technology can quickly locate local overheating areas of transformers, while vibration sensors can capture weak mechanical fluctuations caused by conductor galloping or insulator cracks. The combination of the two can intuitively reflect the surface condition and potential defects of the equipment. At the same time, environmental parameters of the target power grid area are collected simultaneously. Meteorological parameters (such as strong wind data monitored by anemometers and lightning strike frequency recorded by lightning location systems) can reveal the threat of extreme weather such as typhoons and thunderstorms to the structural safety of the power grid; temperature parameters (such as ambient temperature sensors and temperature measuring patches on equipment surfaces) can quantify the impact of high-temperature environments on the current-carrying capacity of conductors and the attenuation of equipment heat dissipation efficiency; humidity parameters (such as air hygrometers and insulator salt / ash density monitoring devices) are used to assess the risk of flashover and corrosion rate of metal components caused by humid environments. For example, in coastal power grids, coupled analysis of humidity parameters and salt spray concentration can accurately predict the probability of insulator flashover due to pollution accumulation, while the spatiotemporal correlation between meteorological parameters during typhoons and conductor galloping images can dynamically assess the risk of tower overturning. By integrating multi-dimensional detection data from the equipment side with multi-parameter monitoring results from the environmental side, a three-dimensional dataset covering "equipment-environment-operating conditions" can be constructed, providing high-resolution feature inputs for subsequent fault identification and significantly improving the comprehensiveness of power grid condition perception and the accuracy of anomaly early warning.

[0073] In a preferred embodiment, acquiring historical detection data of the target power grid and extracting the fault sample proportion coefficient includes: acquiring historical detection data of the target power grid and similar power grids within a historical time period; collecting a set of sample power grid images and a set of sample sensing parameters within the historical detection data, and collecting a corresponding set of sample fault identification results, wherein the sample fault identification results include normal or abnormal; filtering out abnormal sample fault identification results and calculating the proportion to obtain the fault sample proportion coefficient.

[0074] Preferably, complete monitoring data of the target power grid and similar power grids over a historical time span is obtained through a power grid data management platform. This data includes equipment operation records, inspection logs, and historical fault files, forming a composite dataset spanning multiple regions and operating conditions. Taking a regional power grid as an example, its data may include transmission line monitoring records under different climatic conditions (such as the plum rain season in the south and the snowstorm season in the north), as well as historical fault cases of similar power grids under similar topologies, providing rich scenario samples for subsequent analysis.

[0075] Next, the composite dataset is subjected to structured analysis to extract a set of sample power grid images (such as infrared thermal images of insulators and scanned images of tower structures), a set of sample sensor parameters (such as conductor sag monitoring values ​​and surge arrester leakage current curves), and a corresponding set of sample fault identification results (labeled as "normal" or "abnormal" by manual inspection or early diagnosis systems). For example, the sample data for a 220kV line may contain 1000 sets of infrared images and temperature rise curves, of which 80 sets are labeled as "abnormal," corresponding to fault types such as joint overheating and insulator deterioration.

[0076] After data extraction, a labeling and filtering mechanism is used to accurately locate fault samples. Specifically, SQL queries or machine learning classifiers are used to quickly extract all records labeled "abnormal" from the fault identification result set, establishing a fault sample subset. Then, the proportion of this subset in the overall data is calculated based on the proportion coefficient formula (i.e., number of fault samples / total number of samples × 100%). For example, if there are 80 fault samples in the above 220kV line sample and the total number of samples is 1000, the fault sample proportion coefficient is 8%. This coefficient directly reflects the sparsity of fault occurrence in historical data, providing a quantitative basis for subsequent training strategies. When the proportion coefficient is below a threshold (e.g., 5%), a data augmentation mode is activated, using a generative adversarial network (GAN) to synthesize fault samples or employing oversampling techniques to replicate rare fault cases to balance the ratio of positive and negative samples, avoiding overfitting to normal samples due to data bias. This process ensures that historical data retains the true fault distribution characteristics during training while also having sufficient density to support feature learning.

[0077] In a preferred embodiment, historical detection data is combined and divided, and integrated image fault recognition branches and integrated sensor fault recognition branches are trained, including: dividing the sample fault recognition result set to obtain a sample normal fault recognition result set and a sample abnormal fault recognition result set; calculating the ratio of a preset fault sample proportion coefficient to the fault sample proportion coefficient to obtain the total number of branches K, where K is a positive integer; performing random selection with replacement within the sample normal fault recognition result set and the sample abnormal fault recognition result set to obtain sample fault recognition results, combining the corresponding sample power grid image and sample sensor parameters to obtain first image fault recognition training data and first sensor fault recognition training data, wherein the proportion of abnormal fault recognition results is a preset fault sample proportion coefficient; using the first image fault recognition training data and the first sensor fault recognition training data, training the first image fault recognition branch and the first sensor fault recognition branch; continuing to randomly select and combine K image fault recognition training data and K sensor fault recognition training data according to the preset fault sample proportion coefficient to train and obtain the integrated image fault recognition branch and the integrated sensor fault recognition branch.

[0078] For example, combining and integrating historical detection data for training is a core step in improving model robustness. First, based on the labels of sample fault identification results, historical data is decomposed into a set of normal fault identification results (such as records of normal conductor temperature fluctuations and equipment vibration spectra without abnormal peaks) and a set of abnormal fault identification results (such as partial discharge exceeding limits and cases of sudden conductor sag), achieving initial decoupling of data in the fault state dimension. For instance, the historical data of a 500kV substation may contain 3000 sets of normal samples (such as daily inspection records and steady-state operation data) and 200 sets of abnormal samples (such as surge arrester operation logs and insulator flashover accident data).

[0079] Subsequently, to balance the sparsity of fault samples in the dataset, a ratio of a preset fault sample proportion coefficient (e.g., 15%) to the previously calculated fault sample proportion coefficient (e.g., 6.67%) is introduced. This determines the total number of branches, K = 15% / 6.67% ≈ 2.25, which, rounded up, yields K = 3, meaning three ensemble learning branches need to be constructed. This coefficient design is based on engineering experience and model convergence testing. If the preset proportion coefficient is too high (e.g., 30%), it may dilute the features of normal samples; if it is too low (e.g., 5%), it will fail to fully capture the diversity of fault modes.

[0080] During the data combination phase, a bootstrap sampling strategy with replacement is employed to extract data from both the normal and abnormal sample sets according to preset proportions. For example, when generating the first set of training data, 255 sets (85%) are randomly selected from 3000 normal samples, and 45 sets (15%) are randomly selected from 200 abnormal samples. These are then correlated with corresponding power grid images (such as abnormal hotspot areas in infrared thermal imaging) and sensing parameters (such as feature clusters in partial discharge PRPD spectra) to form the first image fault identification training data and the first sensor fault identification training data. This process ensures that each branch training set is stably exposed to the fault feature space by controlling the proportion of abnormal samples.

[0081] Furthermore, based on the initial training data, a hybrid architecture combining Convolutional Neural Network (CNN) and Long Short-Term Memory (LSTM) networks was used to train the first image fault recognition branch (focusing on spatial feature extraction) and the first sensor fault recognition branch (adept at temporal pattern recognition), respectively. Subsequently, the above sampling-combination-training process was repeated twice, generating a total of three independent training sets and corresponding branches. This branch-level data augmentation strategy enables each branch to learn differentiated representations of fault features: for example, branch 1 might focus on capturing the ultraviolet image features of conductor corona discharge, branch 2 focuses on the visible light image texture of insulator contamination accumulation, and branch 3 focuses on the statistical characteristics of partial discharge phase distribution.

[0082] Finally, through an ensemble learning voting mechanism, the identification results of the three branches will be determined by majority vote. For example, if the infrared image branch identifies a set of test data as "abnormal," the vibration sensing branch as "normal," and the ultraviolet image branch as "abnormal," the overall result will be classified as "abnormal." This strategy effectively mitigates misjudgments caused by data noise or insufficient feature coverage in a single branch, significantly enhancing the model's adaptability under complex working conditions.

[0083] In a preferred embodiment, training the first image fault recognition branch and the first sensor fault recognition branch using the first image fault recognition training data and the first sensor fault recognition training data includes: constructing the first image fault recognition branch and the first sensor fault recognition branch respectively using machine learning; and performing supervised training and testing on the first image fault recognition branch and the first sensor fault recognition branch respectively using the first image fault recognition training data and the first sensor fault recognition training data, and completing the training when the error is less than the error threshold.

[0084] Furthermore, based on machine learning theory, the model architectures for the first image fault recognition branch and the first sensor fault recognition branch are initialized respectively. For the image branch, a convolutional neural network (CNN) is used as the backbone network. It forms a deep learning model with spatial feature extraction capabilities through the stacking of convolutional layers, pooling layers, and fully connected layers. For example, lightweight networks such as ResNet-18 or EfficientNet are selected to balance computational efficiency and feature representation capabilities. For the sensor branch, a long short-term memory network (LSTM) or a one-dimensional convolutional neural network (1D-CNN) is used to capture the temporal dependence and local patterns in sensing parameters (such as partial discharge amplitude sequences and wire temperature time series curves). For example, a bidirectional LSTM structure is used to enhance the understanding of historical data context.

[0085] After model initialization, the supervised training and testing phase begins. For the first image-based fault recognition branch, the input is a preprocessed (e.g., size normalization, histogram equalization) infrared thermal image or visible light image of the power grid equipment, labeled with the corresponding fault type (e.g., insulator flashover, conductor strand breakage). For the first sensor-based fault recognition branch, the input is a standardized time series of sensor parameters (e.g., leakage current pulse waveform, vibration acceleration spectrum), also labeled with the fault category. During training, a dynamic adjustment strategy is employed: for example, the image branch uses the cross-entropy loss function to quantify the classification error, supplemented by Focal Loss to address class imbalance; for the sensor branch, mean squared error (MSE) and contrastive loss may be combined to enhance sensitivity to weak fault signals.

[0086] During training, error metrics (such as classification accuracy, F1 score, and root mean square error RMSE) are monitored in real time and compared with preset error thresholds (e.g., image branch classification accuracy must be ≥95%, and sensor branch RMSE must be ≤0.05). If the error of the current branch on the validation set does not meet the threshold, an adaptive optimization mechanism will be automatically triggered. For example, the learning rate may be adjusted for the image branch (e.g., using cosine annealing) or the data augmentation intensity may be increased (e.g., random rotation, elastic deformation); for the sensor branch, the network depth may be optimized (e.g., increasing or decreasing the number of LSTM layers) or an attention mechanism may be introduced (e.g., a Transformer encoder module). When the error is below the threshold for N consecutive training epochs (e.g., N=5), the training of that branch is considered to have converged, and the optimal model weights are saved.

[0087] In a preferred embodiment, based on environmental parameters, failure rate impact analysis and detection data impact analysis are performed to obtain failure rate impact coefficients, image impact coefficients, and sensing impact coefficients. This includes: calling a failure rate impact analyzer, which is trained using a sample environmental parameter set and a sample failure rate impact coefficient set; inputting environmental parameters into the failure rate impact analyzer and outputting failure rate impact coefficients; calling an image impact analysis branch and a sensing impact analysis branch, where the image impact analysis branch and the sensing impact analysis branch are trained using a sample environmental parameter set and a sample sensing impact coefficient set; inputting environmental parameters into the image impact analysis branch and the sensing impact analysis branch respectively, and identifying and outputting the image impact coefficient and the sensing impact coefficient, where the image impact coefficient includes the magnitude of the influence of image quality, and the sensing impact coefficient includes the magnitude of the influence of changes in sensing parameters.

[0088] In detail, a pre-trained failure rate impact analyzer is invoked, which is built based on historical environmental data and failure rate statistics. For example, by collecting meteorological parameters (such as typhoon paths and lightning density), temperature parameters (such as the duration of extreme high / low temperatures), and humidity parameters (such as the number of consecutive rainy days) of a power grid in a certain area over the past five years, and associating them with equipment failure rate data for the corresponding time periods, a non-linear mapping relationship from environmental features to failure rates is formed using Gradient Boosting Decision Tree (GBDT) or Deep Neural Network (DNN). When real-time environmental parameters (such as the current typhoon wind speed reaching level 15 and humidity consistently above 90%) are input into the analyzer, it outputs a failure rate impact coefficient through feature cross-referencing and attention mechanisms. For example, it quantifies that the current environment increases the probability of equipment failure by 3.2 times, providing a basis for prioritizing maintenance.

[0089] Simultaneously, the image impact analysis branch and the sensor impact analysis branch are invoked in parallel. The image branch is constructed based on the perturbation law of environmental parameters on image quality. For example, by simulating the contrast attenuation curve of insulator infrared images under different haze concentrations (PM2.5 values ​​0-500 μg / m³) and the changes in conductor galloping blur caused by strong winds (wind speed 10-30 m / s), a convolutional neural network (CNN) is trained to achieve regression prediction of environmental parameters to image quality indicators (such as structural similarity SSIM values). When real-time environmental parameters (such as the current PM2.5 concentration of 280 μg / m³ and wind speed of 18 m / s) are input into this branch, its output image impact coefficient is 0.65 (the smaller the value, the more severe the image degradation), indicating that image denoising needs to be strengthened or multispectral fusion technology should be adopted.

[0090] The sensing branch focuses on the dynamic impact of environmental parameters on sensor measurement accuracy. For example, it experimentally calibrates the ratio and angle drift of current transformers at different temperatures (-40℃ to +85℃) and the influence curves of humidity (10%-98%RH) on the sensitivity of vibration sensors, constructing prediction models based on support vector regression (SVR) or random forest (RF). When real-time environmental parameters (such as the current equipment surface temperature of 65℃ and ambient humidity of 82%RH) are input into this branch, its output sensing influence coefficient is 1.28 (indicating that the fluctuation range of sensing parameters is 28% larger than that of standard operating conditions), driving the system to activate the temperature and humidity compensation algorithm or switch to a backup sensor channel.

[0091] For example, in a typhoon defense scenario for a power grid in a coastal area, when the failure rate impact analyzer output coefficient is 3.2, the image branch output coefficient is 0.65, and the sensor branch output coefficient is 1.28, a comprehensive judgment is made to activate a Level III emergency response: on the one hand, the frequency of inspections in key areas is increased based on the failure rate coefficients; on the other hand, ultraviolet imaging is used to assist in detection due to image quality degradation, and Kalman filtering is applied to smooth the leakage current sensor data, which is greatly affected by humidity. This process achieves dynamic decoupling between environmental disturbances and equipment status, significantly enhancing the power grid's adaptability in complex environments.

[0092] In a preferred embodiment, the number of image recognition branches and the number of sensor recognition branches are calculated based on the fault rate influence coefficient, the fault sample proportion coefficient, the image influence coefficient, and the sensor influence coefficient. Fault recognition branches are then invoked and fault recognition is performed to obtain power grid fault recognition results. This includes: calculating the image fault recognition coefficient based on the fault rate influence coefficient and the image influence coefficient; multiplying the image fault recognition coefficient by a preset number of branches and rounding to obtain the number of image recognition branches, where the number of image recognition branches is greater than or equal to 1 and less than or equal to K, where K is the total number of branches; calculating the sensor fault recognition coefficient based on the fault rate influence coefficient and the sensor influence coefficient; multiplying the sensor fault recognition coefficient by a preset number of branches and rounding to obtain the number of sensor recognition branches; invoking the image fault recognition branches of the specified number of image recognition branches, inputting the power grid image respectively, and obtaining the image power grid fault recognition results of the specified number of image recognition branches; invoking the sensor fault recognition branches of the specified number of sensor recognition branches, inputting the sensor parameters respectively, and obtaining the sensor power grid fault recognition results of the specified number of sensor recognition branches; and filtering the power grid fault recognition results with the highest frequency.

[0093] Optionally, an image fault identification coefficient can be calculated through a collaborative analysis of environment, fault rate, and image quality. For example, when the fault rate impact coefficient is 3.2 (indicating that the current environment causes a 3.2-fold increase in the probability of faults) and the image impact coefficient is 0.65 (reflecting a 35% decrease in image clarity), a coefficient of 2.57 can be obtained by using a weighted formula (e.g., image fault identification coefficient = α × fault rate coefficient + β × (1 - image coefficient), where α and β are weighting factors, set here as α = 0.7 and β = 0.3). This coefficient includes both the direct impact of the environment on the fault occurrence rate and the indirect constraints of image degradation on feature extraction, providing a quantitative basis for the dynamic adjustment of the number of branches.

[0094] Subsequently, using a preset baseline number of branches (e.g., an initial configuration of 5 branches) as a base, the number of branches is flexibly expanded based on image fault recognition coefficients. For example, after rounding down the number of image branches by 2.57 × 5, 13 branches are obtained. However, due to the upper limit of the total number of branches K (e.g., K = 20), it is ultimately determined that 13 image fault recognition branches will be used. This process uses a coefficient-branch mapping mechanism to match the branch size with the task complexity. In extreme weather conditions such as typhoons, when image blurring makes single-branch feature extraction unreliable, multiple branches can improve fault tolerance through complementary perspectives (e.g., infrared cameras with different focal lengths) and complementary algorithms (e.g., a CNN + Transformer hybrid architecture).

[0095] Similarly, for scenarios where sensing parameters are affected by environmental interference, the system calculates the sensing fault identification coefficient based on joint modeling of failure rate and sensing stability. For example, when the failure rate coefficient is 3.2 and the sensing influence coefficient is 1.28 (reflecting a 28% increase in sensing fluctuation), the sensing fault identification coefficient = γ × failure rate coefficient + δ × sensing coefficient (γ = 0.6, δ = 0.4) is used to synthesize a coefficient of 2.91, which is then used to calculate the number of sensing branches = round(2.91 × 5) = 15, also subject to the constraint K = 20. This branch number design ensures that in high temperature and high humidity environments, problems such as current transformer ratio drift and decreased vibration sensor sensitivity are addressed by suppressing noise interference through multi-branch redundant sampling (such as multi-channel synchronous acquisition) and heterogeneous algorithm fusion (such as LSTM + physical model constraints).

[0096] During the branch invocation phase, the multimodal recognition engine is triggered in parallel. For the image branch, 13 branches receive images of power grid equipment (such as infrared thermal images of insulators) and output 13 sets of fault probability vectors through differentiated feature extraction strategies (such as some branches focusing on edge sharpening and some branches enhancing texture analysis). For the sensing branch, 15 branches simultaneously process multi-source sensor data (such as leakage current PRPD spectrum and conductor tension time series curves) and generate 15 sets of fault discrimination results by combining branch-specific threshold settings (such as using conservative thresholds for branches with strong interference).

[0097] Finally, the multi-branch outputs are integrated through a dynamic voting decision-making mechanism. For example, the frequency of fault labels is counted for 13 sets of image recognition results and 15 sets of sensor recognition results. If "insulator flashover" appears 9 times in the image branch and "partial discharge exceeding the standard" appears 11 times in the sensor branch, both significantly higher than other labels, then it is determined to be the final fault type.

[0098] In a preferred embodiment, the image fault identification coefficient is calculated based on the fault rate influence coefficient and the image influence coefficient, including: using 1 plus the sum of the fault rate influence coefficient and the image influence coefficient as the image fault identification coefficient, that is, using 1 plus the fault rate influence coefficient and the image influence coefficient.

[0099] Specifically, the image fault recognition coefficient is constructed using a normalized compensation strategy, namely: Image Fault Recognition Coefficient = 1 + Fault Rate Influence Coefficient + Image Influence Coefficient, achieving nonlinear fusion of multi-dimensional influences. Here, 1 serves as the basic weight, representing the baseline recognition capability under standard operating conditions; the fault rate influence coefficient (e.g., 3.2 under typhoon conditions) reflects the direct driving effect of the environment on the probability of equipment failure; and the image influence coefficient (e.g., 0.45 due to dense fog, representing a 55% decrease in image contrast) reflects the indirect constraint of the environment on imaging quality. For example, under typhoon and dense fog conditions, the calculated coefficient is: 1 + 3.2 + 0.45 = 4.65. This value includes both the risk weight of frequent faults and the compensation requirement for image degradation.

[0100] Understandably, this coefficient design carries dual engineering significance. Firstly, it enhances fault sensitivity. The coefficient value increases exponentially with the severity of the environment (e.g., the coefficient may reach 6.8 under heavy rain conditions), directly driving the dynamic expansion of the number of image branches. For example, when the coefficient exceeds the threshold of 4.0, the number of image branches is automatically expanded from the basic 5 to 22, enhancing the fault feature capture capability through multi-view imaging (e.g., visible light + infrared + ultraviolet multimodal fusion) and heterogeneous algorithms (e.g., CNN feature extraction + Transformer temporal modeling). Secondly, it compensates for quality degradation. The image impact coefficient is calculated in the form of "1 - quality degradation rate." When the image clarity drops below the threshold (e.g., SSIM < 0.3), the quality compensation term in the coefficient dominates the branch strategy adjustment, such as forcibly enabling super-resolution reconstruction preprocessing or switching the anti-blur feature extraction module. This design ensures the coefficient's dynamic adaptability in complex environments.

[0101] The power grid fault data processing method provided in this embodiment of the invention has at least the following technical effects:

[0102] 1. A dynamic branch adaptive adjustment mechanism driven by multi-dimensional environmental parameters achieves a dual improvement in fault identification accuracy and robustness under complex operating conditions. Specifically, the number of image and sensor branches called is dynamically calculated based on real-time meteorological parameters, temperature parameters, and image / sensor quality coefficients. By integrating multi-modal recognition results, noise interference is effectively suppressed, significantly enhancing the power grid's adaptability to extreme environments.

[0103] 2. By dynamically allocating training data based on failure rate-quality coupling, efficient matching of training sample distribution with actual working conditions is achieved. First, the failure sample proportion coefficient is extracted from historical data, and then the weights of abnormal samples in the training data are dynamically adjusted according to the failure rate influence coefficient and image / sensor influence coefficient of the current environmental parameters. Based on this, K training sets required for ensemble learning are generated through random sampling with replacement, ensuring the sensitivity of each branch model to high-risk failure features and avoiding overfitting due to sample imbalance.

[0104] 3. By using a quantitative fusion method of the triple impact coefficients of environment, fault, and data quality, a refined and dynamic optimization of fault diagnosis strategies is achieved. The fault rate impact coefficient, image impact coefficient, and sensor impact coefficient are normalized and integrated to generate image / sensor fault recognition coefficients, which serve as the core basis for adjusting the number of branches. This achieves real-time decoupling of environmental disturbances and diagnostic strategies, significantly reducing the risk of misjudgment in complex power grid scenarios.

[0105] Example 2:

[0106] Figure 2 This is a schematic diagram illustrating an embodiment of the electronic device provided in this invention. For example... Figure 2 As shown, an embodiment of the present invention provides an electronic device 500, including a memory 510, a processor 520, and a first computer program 511 stored in the memory 510 and executable on the processor 520. When the processor 520 executes the first computer program 511, it implements the power grid fault data processing method as described in Embodiment 1.

[0107] Example 3:

[0108] Figure 3 This is a schematic diagram illustrating an embodiment of a computer-readable storage medium provided by an embodiment of the present invention. For example... Figure 3 As shown, this embodiment provides a computer-readable storage medium 600, on which a second computer program 611 is stored. When the second computer program 611 is executed by a processor, it implements the power grid fault data processing method as described in Embodiment 1.

[0109] It should be noted that the descriptions of each embodiment in the above embodiments have different focuses. For parts that are not described in detail in a certain embodiment, please refer to the relevant descriptions in other embodiments.

[0110] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0111] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded computer, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0112] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0113] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0114] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of this invention and its equivalents, this invention also intends to include these modifications and variations.

Claims

1. A power grid fault data processing method, characterized by, The method comprises: Multi-dimensional detection data acquisition is performed on the target power grid to obtain power grid images and sensing parameters, and environmental parameters within the target power grid environment are collected; Historical detection data of the target power grid are acquired, and a fault sample proportion coefficient is extracted, the historical detection data are combined and divided, and training of an integrated image fault identification branch and an integrated sensing fault identification branch is performed, wherein a branch fault sample proportion coefficient in branch training data of each fault identification branch is a preset fault sample proportion coefficient; According to the environmental parameters, fault rate influence analysis and detection data influence analysis are performed to obtain a fault rate influence coefficient, an image influence coefficient, and a sensing influence coefficient; According to the fault rate influence coefficient, the image influence coefficient, and the sensing influence coefficient, the number of image identification branches and the number of sensing identification branches are calculated to obtain a power grid fault identification result through calling and fault identification of the fault identification branches; The specific method of acquiring the historical detection data of the target power grid and extracting the fault sample proportion coefficient comprises: The historical detection data of the target power grid and the same type of power grid within a historical time are acquired; Sample power grid image sets and sample sensing parameter sets are collected in the historical detection data, and corresponding sample fault identification result sets are collected, wherein the sample fault identification result includes normal or abnormal; Sample fault identification results that are abnormal are screened and a proportion is calculated to obtain a fault sample proportion coefficient; The specific method of combining and dividing the historical detection data and training the integrated image fault identification branch and the integrated sensing fault identification branch comprises: The sample fault identification result sets are divided to obtain sample normal fault identification result sets and sample abnormal fault identification result sets; The ratio of the preset fault sample proportion coefficient to the fault sample proportion coefficient is calculated to obtain a total branch number K, K being a positive integer; Sample fault identification results are randomly selected with replacement from the sample normal fault identification result sets and the sample abnormal fault identification result sets, sample power grid images and sample sensing parameters corresponding to the sample fault identification results are combined, and first image fault identification training data and first sensing fault identification training data are obtained, wherein the proportion of abnormal fault identification results is the preset fault sample proportion coefficient; The first image fault identification branch and the first sensing fault identification branch are trained by using the first image fault identification training data and the first sensing fault identification training data; K image fault identification training data and K sensing fault identification training data are randomly selected and combined according to the preset fault sample proportion coefficient, and the integrated image fault identification branch and the integrated sensing fault identification branch are trained.

2. The power grid fault data processing method of claim 1, wherein, The specific method of performing multi-dimensional detection data acquisition on the target power grid to obtain power grid images and sensing parameters, and collecting environmental parameters within the target power grid environment comprises: Multi-dimensional detection data acquisition is performed on the target power grid to obtain power grid images and sensing parameters; Environmental parameters within the target power grid environment are collected, wherein the environmental parameters include meteorological parameters, temperature parameters, and humidity parameters.

3. The power grid fault data processing method of claim 2, wherein, The specific method of training the first image fault identification branch and the first sensor fault identification branch by using the first image fault identification training data and the first sensor fault identification training data comprises: Respectively constructing the first image fault identification branch and the first sensor fault identification branch by using machine learning; Respectively performing supervised training and testing on the first image fault identification branch and the first sensor fault identification branch by using the first image fault identification training data and the first sensor fault identification training data, and completing the training when the error is less than the error threshold.

4. The power grid failure data processing method of claim 1, wherein, The specific method of performing fault rate influence analysis and detection data influence analysis according to the environmental parameter, and obtaining a fault rate influence coefficient, an image influence coefficient, and a sensor influence coefficient comprises: Calling a fault rate influence analyzer, wherein the fault rate influence analyzer is trained by using a sample environmental parameter set and a sample fault rate influence coefficient set; Inputting the environmental parameter into the fault rate influence analyzer to output a fault rate influence coefficient; Calling an image influence analysis branch and a sensor influence analysis branch, wherein the image influence analysis branch is trained by using a sample environmental parameter set and a sample image influence coefficient set, and the sensor influence analysis branch is trained by using a sample environmental parameter set and a sample sensor influence coefficient set; Respectively inputting the environmental parameter into the image influence analysis branch and the sensor influence analysis branch to identify and output an image influence coefficient and a sensor influence coefficient, wherein the image influence coefficient comprises an influence amplitude of image quality, and the sensor influence coefficient comprises an amplitude of influence change of a sensor parameter.

5. The power grid failure data processing method of claim 1, wherein, The specific method of calculating an image identification branch number and a sensor identification branch number according to the fault rate influence coefficient, the image influence coefficient, and the sensor influence coefficient, calling a fault identification branch, and obtaining a power grid fault identification result comprises: Calculating an image fault identification coefficient according to the fault rate influence coefficient and the image influence coefficient; Multiplying the image fault identification coefficient by a preset branch number and taking an integer to obtain an image identification branch number, wherein the image identification branch number is greater than or equal to 1 and less than or equal to K, and K is a total branch number; Calculating a sensor fault identification coefficient according to the fault rate influence coefficient and the sensor influence coefficient; Multiplying the sensor fault identification coefficient by a preset branch number and taking an integer to obtain a sensor identification branch number; Calling image fault identification branches of the image identification branch number, respectively inputting the power grid image to obtain image grid fault identification results of the image identification branch number; Calling sensor fault identification branches of the sensor identification branch number, respectively inputting the sensor parameter to obtain sensor grid fault identification results of the sensor identification branch number; Screening the power grid fault identification result with the largest number.

6. The power grid failure data processing method of claim 5, wherein, The specific method of calculating an image fault identification coefficient according to the fault rate influence coefficient and the image influence coefficient comprises: Taking 1 and the sum of the fault rate influence coefficient and the image influence coefficient as the image fault identification coefficient.

7. An electronic device, comprising: Comprise: A memory for storing a computer software program; A processor for reading and executing the computer software program to implement the power grid fault data processing method of any one of claims 1-6.

8. A non-transitory computer-readable storage medium, comprising: A storage medium having stored therein a computer software program which, when executed by a processor, implements the power grid fault data processing method of any one of claims 1-6.

Images