Method and system for detecting burning loss fault of meter box in low-voltage transformer area

Abstract

The invention relates to a low-voltage transformer area meter box burning loss fault detection method and system, and the method comprises the following steps: obtaining the time sequence data of a low-voltage meter box, detecting an abnormal behavior through a trained PidanGAE-U model, recognizing the state of the low-voltage transformer area meter box, and generating a preliminary risk assessment result; obtaining multi-dimensional historical and real-time data, carrying out classification division and preliminary evaluation on the identified faults, and constructing a risk factor evaluation system; and based on a risk factor evaluation system, carrying out risk factor correction on a preliminary risk evaluation result, introducing a comprehensive weight evaluation algorithm to carry out quantitative calculation on the influence degree of each corrected factor, and finally generating a risk level of the burning loss fault. Potential faults can be identified and positioned in advance, and the stability of a power system and the power utilization safety of a user are ensured.

Description

Technical Field

[0001] This invention relates to the field of meter box fault detection, and in particular to a method and system for detecting burn-out faults in meter boxes in low-voltage distribution areas. Background Technology

[0002] A low-voltage distribution area is the smallest management unit centered on the distribution transformer. It is responsible for distributing low-voltage electricity to the user side and managing load, controlling line losses, and monitoring energy consumption within the area. It is a crucial link in ensuring the stable and efficient operation of the local power grid. Within the distribution area, the meter box is an important terminal device for users connecting to the power grid, typically integrating an energy meter, protective switches, and communication modules. It is mainly used for user electricity metering and data collection, while also providing line protection functions such as short circuits and overloads. The operating status of the meter box directly affects the accuracy of metering data and the safety of power supply, making it a key device in smart grid construction to protect user rights and ensure the stable operation of the distribution system. Therefore, when a meter box burns out, it can easily lead to accidents such as short circuits and power outages, and even fires, affecting the operation of the distribution area and the safety of user electricity consumption, causing a decline in power grid stability and economic losses. Therefore, an intelligent and efficient low-voltage meter box monitoring and early warning system is of great significance for achieving multi-dimensional status perception, rapid anomaly identification and early warning, improving fault location and handling efficiency, reducing the scope of power outages, shortening recovery time, and ensuring the safe and reliable operation of the power system.

[0003] Existing fault detection technologies for low-voltage meter boxes mainly rely on manual inspections and traditional metering data monitoring. Manual inspections are not only time-consuming and labor-intensive, but also prone to missed detections. Traditional fault detection often depends on staff conducting regular inspections, using visual inspection and experience to observe and judge the operating status of the meter box and metering device. Once a fault occurs, only post-fault damage detection can be performed. This means higher grid operating costs and potentially greater losses during the fault period. Summary of the Invention

[0004] To address the aforementioned problems, the present invention aims to provide a method and system for detecting burn-out faults in low-voltage distribution area meter boxes, which can identify and locate potential faults in advance, ensuring the stability of the power system and the electricity safety of users.

[0005] To achieve the above objectives, the present invention adopts the following technical solution:

[0006] A method for detecting burn-out faults in meter boxes in low-voltage distribution areas includes the following steps:

[0007] Acquire time-series data of low-voltage meter boxes, and use the trained PidanGAE-U model to detect abnormal behavior, identify the status of low-voltage meter boxes in the distribution area, and generate preliminary risk assessment results.

[0008] Acquire multi-dimensional historical and real-time data, classify and preliminarily assess identified faults, and construct a risk factor assessment system;

[0009] Based on the risk factor assessment system, the preliminary risk assessment results are corrected by risk factors. A comprehensive weight evaluation algorithm is introduced to quantify the impact of each corrected factor and finally generate the risk level of burn-out failure.

[0010] Furthermore, the PidanGAE-U model integrates U-Net and the U-GAEchasm architecture of GAE as the backbone network, and combines the preserved egg texture focusing module PidanGAF.

[0011] Furthermore, the PidanGAE-U model maps the time-series data X of the low-pressure meter box to a GAF image using the Gramian angular field algorithm. The time-series data X is first normalized numerically and mapped to obtain the angle sequence θ; then the original time-series data is mapped to... Interval:

[0012] ;

[0013] in, Let X be the i-th element of the time series data. The normalized value. , These represent the maximum and minimum values ​​of the time series data, respectively.

[0014] Mapping normalized values ​​to angles using the inverse cosine function:

[0015] ;

[0016] in, This represents the data value after normalization. For normalized data values The angle value obtained by performing an inverse cosine operation;

[0017] The angular feature matrix of a GAF image is generated by correlation operations on the angle sequence:

[0018] ;

[0019] ;

[0020] in, ), These represent the elements in the angular feature matrices of the Gramian Angular Summation Field and the Gramian Angular Difference Field images, respectively. This represents the i-th angle. This represents the j-th angle.

[0021] Texture angle mapping:

[0022] ;

[0023] in, It is the angle value after texture angle mapping. The angle value of the Gram angle field. , It is a coefficient parameter, L is related to the texture structure. To take the maximum value of i and j.

[0024] Anomalies can cause a sharp increase in the rate of change of texture angles. By calculating and processing the ratio of this rate to the global average rate of change, the weight of the anomaly region can be amplified.

[0025] ;

[0026] in, It amplifies the weight of abnormal regions. It is an activation function. It is a coefficient. The absolute value of the gradient, It is the mean of the gradient.

[0027] In abnormal situations, the periodic changes in angle are more pronounced. By combining the original image features with weighted processing, the unique alternating light and dark texture of preserved egg yolk is simulated, making the yolk features in abnormal areas clearer and more prominent in the final encoded features.

[0028] .

[0029] in, It is the final result obtained after relevant calculations. This is the basic input data. It is used to amplify the weights of abnormal regions. These are coefficient parameters used to adjust the contribution of the cosine term. It is the angle value obtained after texture angle mapping.

[0030] Furthermore, the GAF image feature G is first obtained through eggshell texture encoding and convolution transformation to obtain the latent mean matrix. Latent variance matrix Together with the skip connection feature matrix F, these three matrices have dimensions [L, C, S], where L is the image side length, C is the number of channels, and S is the scale level;

[0031] These features were fed into the variational fault:

[0032] ;

[0033] Where z is the final value of the variable; It is the mean of a normal distribution; It is the standard deviation of the normal distribution; These are standard normal distribution sample values;

[0034] After cross-scale fracture fusion:

[0035] ;

[0036] in, This represents the feature matrix at the s-th scale. This represents attention gating at scale S. This represents the element-wise product; thus, a potential vector that fuses multi-scale features is obtained.

[0037] The output of the variational fault is added to the original skip features via a residual variational connection, and then the GLU is focused through the eggshell texture:

[0038] ;

[0039] Through convolution (Conv) and activation functions In addition, it incorporates angle-related cosine operations to process feature F. This enhances the discriminative power of texture patterns and the stability of the model, outputting an enhanced feature matrix.

[0040] Layer normalization is performed, and the normalized features are then fed into the progressive decoder.

[0041] ;

[0042] in, , These are learnable parameters. Through deconvolution... Linear transformation Activation function and fusion features

[0043] Furthermore, by acquiring multi-dimensional historical and real-time data, the identified faults are classified and preliminarily assessed, and a risk factor assessment system is constructed, as follows:

[0044] An experimental platform simulating an actual distribution transformer area was built to reproduce fault scenarios. Multi-dimensional parameters such as voltage, current, and contact resistance were collected simultaneously to form an experimental database of fault scenarios. Field data was acquired through intelligent monitoring terminals. After outlier removal and missing value completion by edge computing units, the data was spatiotemporally aligned with the experimental data to construct a dual-source dataset of experimental and field data.

[0045] Using the occurrence of centralized burnout faults in dual-source datasets as the objective variable, and the degree of poor contact, duration of current overload, equipment operating years, ambient temperature, and deviation of the live-neutral current ratio as input features (i.e., initial risk factors), an XGBoost model is constructed:

[0046] The XGBoost model employs 5-fold cross-validation, using AUC (Area Under ROC Curve) as the evaluation metric to optimize hyperparameters. After training, the model outputs the importance scores of each feature, which are then normalized to obtain initial weights.

[0047] ;

[0048] in, Let be the initial weight of the i-th initial risk factor. denoted as the overall importance score of the i-th risk factor.

[0049] Furthermore, based on the risk factor assessment system, the preliminary risk assessment results are revised according to risk factors, as follows:

[0050] Based on the preliminary risk assessment results output by the PidanGAE-U model, the abnormal texture types include radial pine flower textures, dense alternating light and dark textures, and blurred and diffuse textures, which correspond one-to-one with burn-off fault causes such as poor contact, current overload, and equipment aging. Abnormal time periods are also identified by timestamps. Mark the start time of the anomaly. Mark the end time and severity of anomalies: quantified by combining model reconstruction error and anomaly texture weights;

[0051] The preliminary risk assessment results are spatiotemporally aligned with the real-time data of each initial risk factor collected by the on-site monitoring terminal (matching with the station area number and collection timestamp as key fields) to form an abnormal event-risk factor data pair. This data pair is then used as the analysis unit to construct a correlation calculation model.

[0052] Using the monitoring data sequence of a certain initial risk factor within an abnormal period as the independent variable:

[0053] ;

[0054] in, This is the start timestamp of the anomaly. This is the timestamp of the abnormal end. This is the real-time monitoring value of the risk factor at time t.

[0055] Based on the sequence of abnormal intensity during this abnormal period:

[0056] ;

[0057] in, The anomaly intensity value at time t is obtained by coupling calculation using anomaly degree quantification parameters:

[0058] ;

[0059] in, The PidanGAE-U model reconstruction error at time t reflects the model's deviation from the reconstruction of the GAF image (converted from current and voltage time series data). The larger the deviation, the higher the degree of anomaly. The abnormal texture weight at time t reflects the saliency of the texture in abnormal regions of the GAF image; a larger weight indicates a more prominent abnormal region. The calculation formulas for the above two are as follows:

[0060] Model reconstruction error calculation:

[0061] ;

[0062] in The original GAF image at time t is the first pixel value, The GAF image reconstructed for the model Pixel values, L is the side length of the GAF image, and C is the number of image channels;

[0063] Abnormal texture weight calculation:

[0064]

[0065] in, The weight of the preserved egg texture at time t is used to amplify the weight ratio of abnormal regions in the GAF image.

[0066] Furthermore, the correlation calculation model uses outlier-factor data pairs as the analysis unit, and the specific calculation formula is as follows:

[0067] ;

[0068] Based on correlation coefficient Based on the absolute value of the correlation between risk factors and abnormal events, the relationship can be divided into three categories: no correlation, weak correlation, and strong correlation.

[0069] A low-voltage distribution area meter box burnout fault detection system includes a processor, a memory, and a computer program stored in the memory. When the processor executes the computer program, it specifically performs the steps in the low-voltage distribution area meter box burnout fault detection method described above.

[0070] The present invention has the following beneficial effects:

[0071] 1. This invention addresses the complex texture features and diverse anomaly patterns of the GAF (Glass Airway Image) data in meter box operation. It designs the PidanGAE-U model based on GAE (Glass Airway Image). By introducing a preserved egg texture encoding module and a U-shaped cross-scale fusion structure, it leverages the targeted extraction of pine flower radial features and multi-scale texture association modeling advantages of both to perform in-depth analysis of abnormal texture information. Through iterative optimization of the encoding and decoding networks, combined with texture angle mapping and dynamic calibration of risk factors, it fully utilizes the preserved egg-like texture characteristics during anomalies, integrating global distribution and local detail features. It demonstrates good adaptability in identifying different anomaly types, more accurately capturing texture features and distribution information of anomaly areas, and improving the robustness and reliability of anomaly detection.

[0072] 2. This invention combines the target data detected and identified by the PidanGAE-U model. For low-voltage distribution substations in target scenarios, it specifically analyzes factors that may cause burn-out faults in the electricity meters within the target low-voltage metering boxes, such as abnormal equipment heating, abnormal voltage drops, the integrity of the equipment casing, and the number of years of operation. Machine learning and other technologies are used to study the correlation between relevant target data and the target. An experimental platform is built to verify the weight of different risk factors on burn-out under multiple dimensions. Multiple possible risk factors are combined into a new matrix, which serves as the input to a fuzzy comprehensive evaluation model. The difference between each factor and the extreme value in its respective set is used as the main basis and indicator for risk rating. The output risk level is fed back into the system, providing support for subsequent electricity meter burn-out protection and early warning work, and providing a reference for modifying the system's early warning scheme. Attached Figure Description

[0073] Figure 1 This is a schematic diagram of the method flow of the present invention;

[0074] Figure 2 This is a structural diagram of the PidanGAE-U model in one embodiment of the present invention;

[0075] Figure 3 This is a multi-dimensional risk factor analysis combined with an experimental platform in one embodiment of the present invention. Detailed Implementation

[0076] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments:

[0077] In this embodiment, firstly, a physical experimental platform is built to realistically reproduce the electrical connections, installation layout, and data communication involved in the actual operation of the low-voltage meter box. Based on this, various meter fault scenarios such as overload, short circuit, and poor contact are simulated. Simultaneously, different environmental conditions such as high temperature, high humidity, and low temperature are simulated. During the simulation, electrical parameters (current, voltage, power, etc.) and multi-dimensional environmental data such as temperature, humidity, and noise are collected synchronously to ensure comprehensive simulation of various abnormal situations. Furthermore, the data collected in the simulation experiment is analyzed in depth to identify the key risk factors affecting meter box burnout and their interactions.

[0078] Based on data obtained from the experimental platform, a deep learning model was used to identify the burn-out condition of low-pressure meter boxes. Focusing on the abnormal performance of various dimensions of data during low-pressure meter box burn-out faults, an innovative PidanGAE-U (Egg-shaped StructureGraph Autoencoder Chasm) model was designed. This model uses a U-GAEchasm network as its backbone architecture and integrates an egg-shaped texture focusing module (PidanGAF) for feature extraction, thereby achieving accurate identification of the normal and abnormal states of low-pressure meter boxes. Multi-source data collected from various sensors, such as electrical parameters and environmental parameters, were input into the model. The model performed comprehensive deep analysis and computation on these input data. With its powerful feature extraction and pattern recognition capabilities, it can accurately detect whether there is a burn-out fault in the low-pressure meter box, providing accurate identification results for subsequent fault handling and other processes.

[0079] After detecting burn-out faults in low-voltage meter boxes, the system first utilizes multi-dimensional historical and real-time data collected by the experimental platform to classify and preliminarily assess the identified faults. Subsequently, combining the status identification results output by the PidanGAE-U model, it dynamically corrects key risk factors such as abnormal temperature changes, physical damage to the box, and poor electrical contact. A comprehensive weighted evaluation algorithm is introduced to quantify the impact of each corrected factor, ultimately generating detailed analysis conclusions such as the risk level and core causes of the burn-out fault. Based on this, the system further integrates the equipment characteristic parameters of the low-voltage meter box with historical fault handling data, outputting an appropriate maintenance plan through a decision logic model. This ensures timely and accurate repair of faulty equipment, effectively reducing the duration of power outages and economic losses caused by burn-out faults.

[0080] In this embodiment, a test platform for burning out a single-phase energy meter with a multi-meter box was built based on actual user conditions. The communication protocol of this single-phase energy meter conforms to DL / T 645—2007 Multifunctional Energy Meter Communication Protocol and its filing documents (hereinafter referred to as the 07 meter protocol). The operation of this energy meter is mainly completed by two major functional systems: the first is the energy measurement unit, which is composed of large-scale application-specific integrated circuits and peripheral circuits. It converts current and voltage into pulses proportional to active power to realize energy metering and can measure the voltage, current, power and power factor of the power grid in real time; the second is the data processing unit, which uses a dedicated low-power microprocessor to complete functions such as energy metering, data storage, communication, display and load switch control.

[0081] Using 220V AC mains power as the power source for a single-phase electricity meter, an adjustable-power heater and various nonlinear loads were used to simulate the variability and complexity of actual electricity consumption. A communication connection was established with the meter's built-in RS-485 interface, with the establishment and termination of the communication link controlled by information frames sent from the master station. On the host computer, serial port software, based on the 07 meter protocol, communicated with the electricity meter to obtain meter operating data (current, voltage, instantaneous active power, power factor, phase angle, voltage harmonic content, etc.) and various event records (current loss event records, voltage loss event records, overload event records, etc.). The electricity meter operating data obtained through communication can be used to extract fault characteristic values ​​in subsequent meter burn-in experiments and to establish a multi-dimensional risk factor assessment system.

[0082] refer to Figure 1 In this embodiment, a method for detecting burn-out faults in low-voltage meter boxes is provided, including the following steps:

[0083] Acquire time-series data of low-voltage meter boxes, and use the trained PidanGAE-U model to detect abnormal behavior, identify the status of low-voltage meter boxes in the distribution area, and generate preliminary risk assessment results.

[0084] Acquire multi-dimensional historical and real-time data, classify and preliminarily assess identified faults, and construct a risk factor assessment system;

[0085] Based on the risk factor assessment system, the preliminary risk assessment results are corrected by risk factors. A comprehensive weight evaluation algorithm is introduced to quantify the impact of each corrected factor and finally generate the risk level of burn-out failure.

[0086] refer to Figure 2In this embodiment, the PidanGAE-U model integrates U-Net and the U-GAEchasm architecture of GAE as the backbone network, and combines the PidanGAF texture focusing module. The model adopts the PidanGAF feature extraction mechanism to transform the time-series data such as current and voltage of the low-voltage meter box into GAF images: its unique texture mapping capability not only fully preserves the temporal correlation and fluctuation characteristics of the data, but also encodes abstract numerical changes into a resolvable texture pattern similar to the pattern of a preserved egg.

[0087] In this embodiment, the PidanGAE-U model maps the time-series data X of the low-pressure meter box to a GAF image using the Gramian angular field algorithm. The time-series data X is first normalized numerically and mapped to obtain the angle sequence θ; the original time-series data is then mapped to... Interval:

[0088] ;

[0089] in, Let X be the i-th element of the time series data. The normalized value. , These represent the maximum and minimum values ​​of the time series data, respectively.

[0090] Mapping normalized values ​​to angles using the inverse cosine function:

[0091] ;

[0092] in, This represents the data value after normalization. For normalized data values The angle value is obtained by performing an inverse cosine operation; the angle feature matrix of the GAF image is generated by the correlation operation of the angle sequence.

[0093] ;

[0094] ;

[0095] in, ), These represent the elements in the angular feature matrices of the Gramian Angular Summation Field and the Gramian Angular Difference Field images, respectively. This represents the i-th angle. This represents the j-th angle.

[0096] Because the model incorporates a preserved egg texture focusing module, when data anomalies occur, the original angular features undergo abrupt changes. The resulting texture angles exhibit a wavy property similar to the pine flower pattern in a preserved egg. By utilizing the radial characteristics from the center to the edge, the radial texture spreading outward from the abnormal center is enhanced, much like the pine flower pattern spreading from the yolk to the egg white in a preserved egg.

[0097] Texture angle mapping:

[0098] ;

[0099] in, It is the angle value after texture angle mapping. The angle value of the Gram angle field. , It is a coefficient parameter, L is related to the texture structure. To take the maximum value of i and j.

[0100] Anomalies can cause a sharp increase in the rate of change of texture angles. By calculating and processing the ratio of this rate to the global average rate of change, the weight of the anomaly region can be amplified.

[0101] ;

[0102] in, It amplifies the weight of abnormal regions. It is an activation function. It is a coefficient. The absolute value of the gradient, It is the mean of the gradient.

[0103] In abnormal situations, the periodic changes in angle are more pronounced. By combining the original image features with weighted processing, the unique alternating light and dark texture of preserved egg yolk is simulated, making the yolk features in abnormal areas clearer and more prominent in the final encoded features.

[0104] .

[0105] in, It is the final result obtained after relevant calculations. This is the basic input data. It is used to amplify the weights of abnormal regions. These are coefficient parameters used to adjust the contribution of the cosine term. It is the angle value obtained after texture angle mapping.

[0106] The backbone network adopts the U-GAEchasm model, which integrates the U-Net and GAE architectures. It uses a three-stage fault structure of "symmetric encoding-probabilistic bridging-progressive decoding" and introduces a cross-scale feature gap fusion mechanism. By combining the texture focusing sampling of preserved egg to simulate the texture boundary characteristics in the pattern of preserved egg and variational attention gating, it retains the ability of U-Net to capture subtle features and gives GAE the advantage of probabilistic modeling of normal pattern distribution.

[0107] Its core structure consists of a U-shaped symmetric encoder, a variational discontinuous bridging layer, and a progressive decoder: the encoder extracts texture features from GAF images through multi-scale convolutions, and retains detailed information with skip connections; the variational discontinuous bridging layer maps features to a mean-variance distribution, achieving a "discontinuity crossing" from deterministic features to probability space through random sampling; the decoder reconstructs the image based on the sampled features and selectively fuses the cross-layer features of the encoder through attention gating. Simultaneously, residual variational connections and normalization mechanisms are introduced to balance reconstruction accuracy and distribution learning stability, enabling the model to maintain high-fidelity reconstruction of normal textures while remaining highly sensitive to the "discontinuous differences" in abnormal preserved egg textures.

[0108] In this embodiment, the GAF image feature G is first obtained as a latent mean matrix through eggshell texture encoding and convolution transformation. Latent variance matrix Together with the skip connection feature matrix F, these three matrices have dimensions [L, C, S], where L is the image side length, C is the number of channels, and S is the scale level;

[0109] These features were fed into the variational fault:

[0110] ;

[0111] Where z is the final value of the variable; It is the mean of a normal distribution; It is the standard deviation of the normal distribution; These are standard normal distribution sample values;

[0112] After cross-scale fracture fusion:

[0113] ;

[0114] in, This represents the feature matrix at the s-th scale. This represents attention gating at scale S. This represents the element-wise product; thus, a potential vector that fuses multi-scale features is obtained.

[0115] The output of the variational fault is added to the original skip features via a residual variational connection, and then the GLU is focused through the eggshell texture:

[0116] ;

[0117] Through convolution (Conv) and activation functions In addition, it incorporates angle-related cosine operations to process feature F. This enhances the discriminative power of texture patterns and the stability of the model, outputting an enhanced feature matrix.

[0118] Layer normalization is performed, and the normalized features are then fed into the progressive decoder.

[0119] ;

[0120] in, , These are learnable parameters. Through deconvolution... Linear transformation Activation function and fusion features .

[0121] After the PidanGAE-U model completes feature extraction and preliminary anomaly detection of the input data, its output results, while capturing the anomaly representation of the data's inherent texture patterns, are limited by the model-driven decision logic, making it difficult to fully incorporate the coupling effects of multiple sources of interference in real-world scenarios. To achieve scenario adaptability and decision robustness in anomaly diagnosis, a deep fusion analysis based on a multi-dimensional risk factor system constructed on an experimental platform is necessary.

[0122] The study focuses on risk factors that cause electricity meter burnout, such as poor contact, operation for 7 years or more, operation during high-temperature weather, and months with high electricity consumption. It primarily extracts historical load curve data of smart meters, including voltage, current, electricity consumption, live and neutral wire current, and power, as well as dismantled and sorted inspection data. Through data analysis, phenomenon analysis, and mechanism analysis, the study conducts case verification on an experimental platform. It summarizes and identifies key characteristics of burnt metering equipment, such as abnormal voltage, abnormal live and neutral wire current ratio, and abnormal electricity consumption fluctuations, establishing a correlation with the main causes of electricity meter burnout. Furthermore, it demonstrates that different risk factors have varying degrees of impact on electricity meter burnout.

[0123] In this embodiment, multi-dimensional historical and real-time data are acquired to classify and preliminarily assess the identified faults, and a risk factor assessment system is constructed, as follows:

[0124] An experimental platform simulating an actual distribution transformer area was built to reproduce fault scenarios. Multi-dimensional parameters such as voltage, current, and contact resistance were collected simultaneously to form an experimental database of fault scenarios. Field data was acquired through intelligent monitoring terminals. After outlier removal and missing value completion by edge computing units, the data was spatiotemporally aligned with the experimental data to construct a dual-source dataset of experimental and field data.

[0125] Using the occurrence of centralized burnout faults in dual-source datasets as the objective variable, and the degree of poor contact, duration of current overload, equipment operating years, ambient temperature, and deviation of the live-neutral current ratio as input features (i.e., initial risk factors), an XGBoost model is constructed:

[0126] The XGBoost model employs 5-fold cross-validation, using AUC (Area Under ROC Curve) as the evaluation metric to optimize hyperparameters. After training, the model outputs the importance scores of each feature, which are then normalized to obtain initial weights.

[0127] ;

[0128] in, Let be the initial weight of the i-th initial risk factor. denoted as the overall importance score of the i-th risk factor.

[0129] In this embodiment, based on the risk factor assessment system, the preliminary risk assessment results are revised according to risk factors, as follows:

[0130] Based on the preliminary risk assessment results output by the PidanGAE-U model, the abnormal texture types include radial pine flower textures, dense alternating light and dark textures, and blurred and diffuse textures, which correspond one-to-one with burn-off fault causes such as poor contact, current overload, and equipment aging. Abnormal time periods are also identified by timestamps. Mark the start time of the anomaly. Mark the end time and severity of anomalies: quantified by combining model reconstruction error and anomaly texture weights;

[0131] The preliminary risk assessment results are spatiotemporally aligned with the real-time data of each initial risk factor collected by the on-site monitoring terminal (matching with the station area number and collection timestamp as key fields) to form an abnormal event-risk factor data pair. This data pair is then used as the analysis unit to construct a correlation calculation model.

[0132] Using the monitoring data sequence of a certain initial risk factor within an abnormal period as the independent variable:

[0133] ;

[0134] in, This is the start timestamp of the anomaly. This is the timestamp of the abnormal end. This is the real-time monitoring value of the risk factor at time t.

[0135] Based on the sequence of abnormal intensity during this abnormal period:

[0136] ;

[0137] in, The anomaly intensity value at time t is obtained by coupling calculation using anomaly degree quantification parameters:

[0138] ;

[0139] in, The PidanGAE-U model reconstruction error at time t reflects the model's deviation from the reconstruction of the GAF image (converted from current and voltage time series data). The larger the deviation, the higher the degree of anomaly. The abnormal texture weight at time t reflects the saliency of the texture in abnormal regions of the GAF image; a larger weight indicates a more prominent abnormal region. The calculation formulas for the above two are as follows:

[0140] Model reconstruction error calculation:

[0141] ;

[0142] in The original GAF image at time t is the first pixel value, The GAF image reconstructed for the model Pixel values, L is the side length of the GAF image, and C is the number of image channels;

[0143] Abnormal texture weight calculation:

[0144] ;

[0145] in, The weight of the preserved egg texture at time t is used to amplify the weight ratio of abnormal regions in the GAF image.

[0146] In this embodiment, the correlation calculation model uses anomaly-factor data pairs as the analysis unit, and the specific calculation formula is as follows:

[0147] ;

[0148] Based on the absolute value of the correlation coefficient, the correlation between risk factors and abnormal events is divided into three categories: no correlation ∈ [0, 0.3); weak correlation ∈ [0.3, 0.7]; and strong correlation ∈ (0.7, 1). This classification serves as the basis for weight adjustment. The weight of strongly correlated factors is increased by 1.2-1.5 times according to the correlation strength, the weight of uncorrelated factors is reduced to 80%-95% of the original weight, and the weight of weakly correlated factors is finely adjusted by ±10%. After adjustment, normalization ensures that the weight sums to 1.

[0149] By sorting the burnout risk indices of each low-pressure meter box obtained from the risk assessment in descending order, a priority ranking from high risk to low risk can be formed, intuitively identifying the meter boxes that require key attention. The ranked risk results and corresponding cause information are fed back to the low-pressure meter box burnout fault detection and early warning module, and differentiated processing is performed based on the risk score.

[0150] 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.

[0151] 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 processor, 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 and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0152] 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.

[0153] 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.

[0154] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

Claims

1. A low-voltage transformer area meter box burnout fault detection method, characterized in that, The method comprises the following steps: Obtain the time series data of the low-voltage meter box, and detect abnormal behaviors by using the trained PidanGAE-U model, identify the state of the low-voltage transformer area meter box, and generate a preliminary risk assessment result; Obtain multi-dimensional historical and real-time data, classify and preliminarily evaluate the identified faults, and build a risk factor evaluation system; Based on the risk factor evaluation system, the preliminary risk assessment result is corrected by introducing a comprehensive weight evaluation algorithm to quantitatively calculate the influence degree of each corrected factor, and finally the risk level of the burning failure is generated.

2. The low-voltage transformer area metering cabinet burnout fault detection method according to claim 1, characterized in that, The PidanGAE-U model fuses the U-GAEchasm of U-Net and GAE architecture as the backbone network, and combines the pidan texture focusing module PidanGAF.

3. The low-voltage transformer area metering cabinet burnout fault detection method according to claim 2, characterized in that, The PidanGAE-U model maps the time series data X of the low-pressure meter box into a GAF image through a Gramian angle field algorithm. The time series data X is first mapped into an angle sequence θ through numerical normalization and angle mapping; the original time series data is mapped into Interval: ； wherein, is the i-th element of the time series data X, is the normalized value, , respectively represent the maximum and minimum values of the time series data. The normalized value is mapped to an angle by using the inverse cosine function: ； wherein, represents a data value after normalization processing, is an angle value obtained by performing an inverse cosine operation on the normalized data value . The angle feature matrix of the GAF image is generated by the correlation operation of the angle sequence: ； ； wherein ), denote the elements of the angle feature matrix of the Gram angle and field, Gram angle difference field image, respectively, denotes the i-th angle, denotes the j-th angle; Texture angle mapping: ； wherein is the angle value after texture angle mapping, is the angle value of the Gram angle field, , is a coefficient parameter, L is related to the texture structure, is the maximum value of i and j; Abnormalities will sharply increase the rate of change of the texture angle, and by calculating the proportion of the global average change rate and processing, the weight of the abnormal area is amplified: ； wherein, is a weight amplifying abnormal regions, is an activation function, is a coefficient, is an absolute value of a gradient of, is a mean of gradients Under abnormal conditions, the periodic change of the angle is more obvious, and after combining the original image features and weight processing, the unique light and dark texture of the pidan is simulated, so that the characteristics of the pidan are more clear and prominent in the final encoding features: ； wherein, is the final result after correlation operation, is the basic input data, is the weight for amplifying the abnormal area, is the coefficient parameter for adjusting the contribution degree of the cosine term, is the angle value after texture angle mapping.

4. The low-voltage transformer area metering cabinet burnout fault detection method according to claim 3, characterized in that, The GAF image feature G is first obtained by a potential mean matrix through egg texture coding and convolution transformation , a potential variance matrix , and a skip connection feature matrix F, the dimensions of the three matrices being [L, C, S], wherein L is the image side length, C is the channel number, and S is the scale level; These features are sent to the variational layer: ； where z is the final variable value; is the mean of the normal distribution; is the standard deviation of the normal distribution; is the standard normal distribution sampling value; After cross-scale crack fusion: ； wherein, represents the feature matrix at the s-th scale, represents the attention gate at scale S, represents the element-wise product; thereby obtaining the latent vector that fuses the multi-scale features; The output of the variational layer is added to the original jump feature through the residual variational connection, and then the pidan texture focusing GLU is used: ； by a convolution (Conv), an activation function and a cosine operation in connection with an angle-dependent cosine operation to process the features F; Enhance the discrimination ability and model stability of the texture pattern, and output the enhanced feature matrix; Layer normalization is performed, and the normalized features are sent to the progressive decoder: ； wherein, , are learnable parameters; denotes deconvolution, ReLU as activation function; is a fused feature.

5. The low-voltage transformer area metering cabinet burnout fault detection method according to claim 1, characterized in that, The multi-dimensional historical and real-time data are obtained, the identified faults are classified, divided and preliminarily evaluated, and the risk factor evaluation system is built, as follows: An experimental platform simulating the actual distribution area is built, the fault scene is reproduced, and the voltage, current and contact resistance multi-dimensional parameters are synchronously collected to form an experimental database of the fault scene; the field data are obtained through intelligent monitoring terminals, and after the abnormal values are removed and the missing values are completed by the edge computing unit, the data are time and space aligned with the experimental data to build an experimental-field dual-source dataset; Taking whether the burning failure occurs in the dual-source dataset as the target variable, and taking the contact failure degree, current overload time, equipment operation time, environmental temperature and zero-fire line current ratio deviation as input features, an XGBoost model is built: The XGBoost model uses cross-validation to optimize the hyperparameters with AUC as the evaluation index; after the model training is completed, the importance scores of the features are output, and the initial weights are obtained by normalization processing: ； where k is a variable for indexing, starting from 1 to the total n risk factors. is the initial weight for the i-th initial risk factor, is the comprehensive importance score for the i-th risk factor.

6. The low-voltage transformer area metering cabinet burnout fault detection method according to claim 1, characterized in that, Based on the risk factor evaluation system, the preliminary risk assessment result is corrected, as follows: According to the preliminary risk assessment result output by the PidanGAE-U model, including the abnormal texture type, abnormal period and abnormal degree; The preliminary risk assessment result is matched with the initial risk factor real-time data collected by the field monitoring terminal in time and space to form an abnormal event-risk factor data pair, and the data pair is taken as an analysis unit to construct a correlation degree calculation model: The monitoring data sequence of an initial risk factor in an abnormal period is taken as an independent variable: ； wherein, is an abnormal start timestamp, is an abnormal end timestamp, is a real-time monitoring value of the risk factor at time t; The abnormal intensity sequence in the abnormal period is taken as an independent variable: ； wherein is the abnormal intensity value for time t, calculated by coupling the abnormality degree quantification parameter: ； wherein, is the PidanGAE-U model reconstruction error for time t; is the abnormal texture weight for time t; the calculation formulas for both of the above are as follows: Model reconstruction error calculation: ； wherein is the original GAF image at time t, is the pixel value of the original GAF image at time t, is the pixel value of the model reconstructed GAF image at time t, L is the side length of the GAF image, and C is the number of channels of the image. Abnormal texture weight calculation: ； wherein, is the texture weight of the preserved egg at time t, and is used to amplify the weight proportion of the abnormal area in the GAF image.

7. The low-voltage transformer area metering cabinet burnout fault detection method according to claim 6, characterized in that, The correlation degree calculation model takes the abnormal-event factor data pair as an analysis unit, and the specific calculation formula is as follows: ； Based on the absolute value of the correlation coefficient The correlation between the risk factor and the abnormal event is divided into three categories: no correlation, weak correlation, and strong correlation.

8. A low-voltage transformer area metering cabinet burning failure detection system, characterized in that, The low-voltage transformer area meter box burning failure detection method comprises a processor, a memory and a computer program stored in the memory, and the processor executes the computer program to specifically execute the steps in any one of claims 1-7.

Images