Power distribution cabinet fault positioning method and system

Abstract

The invention relates to the technical field of power distribution cabinets, in particular to a power distribution cabinet fault positioning method and system. The method comprises the steps of generating a multi-source monitoring sequence by synchronously collecting sensing data of a plurality of monitoring points deployed according to an electrical topological relation in a power distribution cabinet; extracting and fusing target features from the multi-source monitoring sequence based on a feature template corresponding to a monitoring point associated component, and generating a comprehensive feature vector; dynamically fusing and analyzing the comprehensive feature vector and the current operation condition parameters through a preset diagnosis model to generate a diagnosis data set; calculating the probability that each component is a fault source in combination with the fault positioning knowledge graph and the diagnosis data set corresponding to the power distribution cabinet, and generating a component fault probability data set; and judging the component exceeding a preset fault threshold value as a fault component, and generating corresponding fault alarm data. According to the invention, through collaborative innovation of multi-source data synchronous acquisition, feature template precise fusion, working condition dynamic adaptation analysis and knowledge graph assisted positioning, the accuracy and reliability of fault positioning are effectively improved.

Description

Technical Field

[0001] This invention relates to the field of power distribution cabinet technology, and in particular to a method and system for locating faults in power distribution cabinets. Background Technology

[0002] With the continuous expansion of urban power grids and the constant improvement of their intelligence level, distribution cabinets, as key nodes in power distribution and control, directly affect power supply quality due to their operational reliability. Currently, distribution cabinets are widely deployed in various substations, industrial and mining enterprises, and commercial buildings, and are numerous and widely distributed.

[0003] To ensure power continuity, rapid and accurate fault location within the distribution cabinet is crucial for timely early warning and proactive maintenance. Existing fault location methods primarily rely on installing single sensors, such as temperature or partial discharge sensors, at key locations within the cabinet to collect specific physical quantity data and upload it to the monitoring center. When the data exceeds a preset fixed threshold, the system triggers an alarm and roughly indicates the abnormal area.

[0004] However, in actual operation, the above-mentioned location methods have significant drawbacks, resulting in low accuracy in fault location. First, relying on a single characteristic parameter makes it difficult to comprehensively and early capture complex or latent faults, easily leading to misjudgments or missed diagnoses. Second, using fixed static thresholds cannot adapt to equipment aging curves, load fluctuations, and environmental changes, making it difficult to balance alarm sensitivity and accuracy under complex operating conditions. Third, existing methods lack intelligent analysis of the correlation between fault type and location, typically only able to locate anomalies to a certain cabinet or a general area, unable to accurately locate the smallest replaceable unit such as a specific circuit breaker, busbar connection point, or cable terminal, making on-site troubleshooting difficult for maintenance personnel and prolonging power outage time. Summary of the Invention

[0005] This invention provides a method and system for locating faults in power distribution cabinets, which solves the technical problem that existing power distribution cabinet fault location methods lack the ability to comprehensively perceive and intelligently analyze multi-dimensional and time-varying fault characteristics, resulting in low location accuracy.

[0006] The first aspect of this invention provides a method for locating faults in a power distribution cabinet, comprising:

[0007] Simultaneously collect sensor data from multiple monitoring points deployed in the power distribution cabinet according to electrical topology to generate a multi-source monitoring sequence;

[0008] Based on the feature templates corresponding to the components associated with the monitoring points, target features are extracted and fused from the multi-source monitoring sequence to generate a comprehensive feature vector.

[0009] A diagnostic dataset is generated by dynamically fusing the comprehensive feature vector and current operating condition parameters using a preset diagnostic model.

[0010] Based on the fault location knowledge graph corresponding to the power distribution cabinet and the diagnostic dataset, the probability that each component in the power distribution cabinet is a fault source is calculated, and a component fault probability dataset is generated.

[0011] Components in the component failure probability dataset that exceed a preset failure threshold are identified as faulty components, and corresponding fault alarm data is generated for the faulty components.

[0012] Optionally, the step of synchronously collecting sensor data from multiple monitoring points deployed according to electrical topology within the distribution cabinet and generating a multi-source monitoring sequence includes:

[0013] Based on the electrical wiring diagram of the distribution cabinet, physical nodes that characterize the key electrical connection status are used as monitoring points;

[0014] Temperature, partial discharge, and vibration data are collected synchronously at the monitoring points to generate a time-aligned multi-source monitoring sequence.

[0015] Optionally, the step of extracting and fusing target features from the multi-source monitoring sequence based on the feature template corresponding to the monitoring point associated component to generate a comprehensive feature vector includes:

[0016] According to the electrical topology of the distribution cabinet, the key stress types of each physical component in the distribution cabinet are matched from the preset component stress rule library, and the typical fault modes associated with the key stress types are determined based on the historical fault database.

[0017] Based on a preset fault mechanism and feature mapping relationship table, features corresponding to each typical fault mode are extracted from the multi-source monitoring sequence to generate target features corresponding to the typical fault modes.

[0018] Based on the occurrence probability of the typical fault modes in the historical fault database, and combined with the historical representation contribution of the target feature, the fusion weight coefficient of the target feature is calculated.

[0019] A comprehensive feature vector is generated by weighting all the target features and their corresponding fusion weight coefficients.

[0020] Optionally, the preset diagnostic model includes multiple sub-diagnostic models optimized for different fault types or different component areas; the step of generating a diagnostic dataset by dynamically fusing and analyzing the comprehensive feature vector and current operating condition parameters using the preset diagnostic model includes:

[0021] The current operating condition parameters are input into the pre-trained weight allocation function to calculate the dynamic decision weights of each sub-diagnostic model.

[0022] Each of the sub-diagnostic models performs forward reasoning on the comprehensive feature vector to generate candidate fault types, component regions, and initial confidence levels corresponding to the sub-diagnostic models.

[0023] The initial confidence level is multiplied by the corresponding dynamic decision weight to generate a weighted confidence level.

[0024] The weighted confidence scores are normalized using the softmax function to generate normalized confidence probabilities.

[0025] Select the normalized confidence probability that is higher than the preset reporting threshold to generate the target confidence probability;

[0026] A diagnostic dataset is constructed using the candidate fault type, component region, and target confidence probability corresponding to the target confidence probability.

[0027] Optionally, the step of calculating the probability that each component in the distribution cabinet is a fault source based on the fault location knowledge graph corresponding to the distribution cabinet and the diagnostic dataset, and generating a component fault probability dataset, includes:

[0028] Using each physical component in the power distribution cabinet as a node and the electrical connection relationship between the components as an edge, a priori fault probability based on historical statistics is configured for each node, and a conditional probability based on historical fault propagation statistics is configured for each edge, thereby constructing a fault location knowledge graph.

[0029] The diagnostic dataset is used as observational evidence and mapped to the corresponding nodes and fault types in the knowledge graph.

[0030] Based on the observed evidence, the prior fault probability of each node, and the conditional probability of each edge, iterative probability propagation calculations are performed on the topology of the knowledge graph to update the posterior probability of each node as a fault source.

[0031] A component failure probability dataset is constructed using each node and its updated posterior probability.

[0032] Optionally, the step of identifying components in the component failure probability dataset that exceed a preset failure threshold as faulty components and generating fault alarm data corresponding to the faulty components includes:

[0033] Components whose posterior probability in the component failure probability dataset exceeds a first preset threshold are selected as first-level faulty components.

[0034] Components whose posterior probability in the component failure probability dataset exceeds a second preset threshold but does not exceed a first preset threshold are selected as secondary failure components; wherein, the first preset threshold is higher than the second preset threshold;

[0035] Extract typical fault causes and maintenance measures of similar components corresponding to the first-level faulty components from the historical maintenance database, and generate high-priority alarm information;

[0036] Based on the component failure probability data generated by the secondary faulty component within a preset historical period, the failure probability change trend characteristics corresponding to the secondary faulty component are determined.

[0037] The fault probability change trend characteristics are formatted to generate a trend judgment conclusion that includes trend description and risk level;

[0038] Using the trend analysis conclusions and the component identifiers and current failure probabilities corresponding to the secondary faulty components, observation-level early warning information is constructed;

[0039] The high-priority alarm information and the observation-level early warning information are combined to generate fault alarm data.

[0040] A second aspect of the present invention provides a power distribution cabinet fault location system, comprising:

[0041] The multi-source synchronous acquisition module is used to synchronously acquire sensor data from multiple monitoring points deployed in the power distribution cabinet according to the electrical topology, and generate a multi-source monitoring sequence.

[0042] The mechanism feature fusion module is used to extract and fuse target features from the multi-source monitoring sequence based on the feature templates corresponding to the monitoring point associated components, and generate a comprehensive feature vector.

[0043] The adaptive diagnostic module is used to dynamically fuse and analyze the comprehensive feature vector and current operating condition parameters through a preset diagnostic model to generate a diagnostic dataset.

[0044] The probabilistic localization reasoning module is used to calculate the probability that each component in the power distribution cabinet is a fault source based on the fault localization knowledge graph corresponding to the power distribution cabinet and the diagnostic dataset, and generate a component fault probability dataset.

[0045] The hierarchical intelligent alarm module is used to identify components in the component failure probability dataset that exceed a preset failure threshold as faulty components and generate corresponding fault alarm data for the faulty components.

[0046] A third aspect of the present invention provides an electronic device, including a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor performs the steps of the power distribution cabinet fault location method described above.

[0047] The fourth aspect of the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed, implements the power distribution cabinet fault location method as described above.

[0048] The fifth aspect of the present invention provides a computer program product, the computer program product comprising a computer program stored on a non-transitory computer-readable storage medium, the computer program comprising program instructions, wherein when the program instructions are executed by a computer, the computer performs the power distribution cabinet fault location method as described above.

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

[0050] This invention deploys monitoring points according to the electrical topology of the distribution cabinet and synchronously collects sensor data to generate a multi-source monitoring sequence. Based on the feature templates of the components associated with the monitoring points, target features are extracted and fused from the multi-source monitoring sequence to form a comprehensive feature vector. Using a preset diagnostic model, the comprehensive feature vector is dynamically fused and analyzed with current operating parameters to generate a diagnostic dataset. Combining the fault location knowledge graph corresponding to the distribution cabinet and the diagnostic dataset, the probability of each component being a fault source is calculated, generating a component fault probability dataset. Finally, components exceeding a preset fault threshold are identified as faulty components, and corresponding fault alarm data is generated. This invention specifically addresses the technical problem of low location accuracy in existing distribution cabinet fault location methods due to a lack of comprehensive perception and intelligent analysis capabilities for multi-dimensional, time-varying fault features.

[0051] Its beneficial effects are reflected in the following aspects: Deploying monitoring points according to electrical topology and collecting data synchronously ensures the relevance and consistency of multi-dimensional fault-related data, avoiding the blindness and fragmentation of data collection, and providing a foundation for the comprehensive perception of multi-dimensional fault characteristics; Extracting and fusing target features based on feature templates achieves accurate screening and effective integration of multi-dimensional fault characteristics, avoiding interference from irrelevant information and improving the identification of fault characteristics; The preset diagnostic model, combined with comprehensive feature vectors and current operating condition parameters, performs dynamic fusion analysis, enabling the fault analysis process to adapt to changes in equipment operating status, effectively capturing the patterns of time-varying fault characteristics, and overcoming... This approach overcomes the limitations of traditional methods in handling dynamically changing fault characteristics. By combining a fault location knowledge graph with a diagnostic dataset to calculate the probability of component fault sources, it fully utilizes the structural correlation characteristics and fault-related clues of the distribution cabinet, making fault source tracing more logical and accurate, and avoiding location deviations caused by isolated judgments. By setting a fault threshold to determine faulty components and generating alarm data, the reliability of the location results is further verified. The entire process forms a complete link from multi-dimensional data collection, feature integration, dynamic analysis to precise location, comprehensively improving the ability to perceive and intelligently analyze multi-dimensional and time-varying fault characteristics, thereby improving location accuracy. Attached Figure Description

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

[0053] Figure 1 A flowchart illustrating the steps of a power distribution cabinet fault location method provided in an embodiment of the present invention;

[0054] Figure 2 A structural block diagram of a power distribution cabinet fault location system provided in an embodiment of the present invention;

[0055] Figure 3 This is a structural block diagram of a computer device provided in an embodiment of the present invention. Detailed Implementation

[0056] This invention provides a method and system for locating faults in power distribution cabinets, which addresses the technical problem that existing methods for locating faults in power distribution cabinets lack the ability to comprehensively perceive and intelligently analyze multi-dimensional and time-varying fault characteristics, resulting in low location accuracy.

[0057] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0058] Please see Figure 1 , Figure 1 The flowchart illustrates the steps of a power distribution cabinet fault location method provided in this embodiment of the invention.

[0059] The present invention provides a method for locating faults in a power distribution cabinet, comprising:

[0060] Step 101: Synchronously collect sensor data from multiple monitoring points deployed in the distribution cabinet according to the electrical topology, and generate a multi-source monitoring sequence.

[0061] Further, step 101 may include the following sub-steps:

[0062] S11. Based on the electrical wiring diagram of the distribution cabinet, physical nodes that characterize the key electrical connection status are used as monitoring points.

[0063] In this embodiment of the invention, standardized electrical topology drawings of the original manufacturer of the distribution cabinet are obtained. These drawings must include core information such as component models, wiring methods, and rated parameters. Maintenance personnel use these drawings to disassemble the main circuit and control circuit of the distribution cabinet one by one, clarifying the connection relationships, installation locations, and functional divisions of each component, forming a component topology association list. This list provides the basis for subsequent selection of monitoring points and data mapping. Based on the dual principles of high-incidence fault areas and core electrical connection points, monitoring points are selected from the component topology association list. Specific selection targets include 12 typical nodes such as main switch contacts, busbar segment connections, circuit breaker inlet and outlet terminals, transformer secondary winding terminals, and cable joints. The selection is based on historical fault statistics. These nodes cover 92% of the fault locations in the distribution cabinet, ensuring that the collected data accurately captures fault-related signals.

[0064] Each selected monitoring point is assigned a unique physical identifier, consisting of three parts: component name, component number, and topological location (e.g., the main switch K1 incoming line terminal). This identifier will serve as a core metadata field, directly linked to nodes in the subsequent fault location knowledge graph, ensuring that every piece of sensor data collected can be accurately mapped to its corresponding physical component, avoiding data ambiguity. With the distribution cabinet de-energized, sensors are installed at the locations corresponding to their identifiers. Installation locations must avoid exposed high-voltage parts and maintain a safe distance of at least 15 cm from live parts, complying with GB 7251.1 2022 Safety Specifications for Low-Voltage Switchgear and Controlgear Assemblies. After deployment, the consistency between the sensor installation locations and the monitoring point identifiers is verified to ensure that each sensor corresponds to a unique monitoring point, providing physical assurance for the accuracy of data collection.

[0065] S12. Simultaneously collect temperature, partial discharge, and vibration data at the monitoring points to generate a time-aligned multi-source monitoring sequence.

[0066] In this embodiment of the invention, data with different sampling frequencies are timestamped based on the lowest frequency, and interpolation or downsampling methods are used to ensure the consistency of the data sequence in the time dimension. Metadata is associated with each data point in the multi-source monitoring sequence. This metadata includes monitoring point identifiers used to map the data points to specific physical components in the electrical topology of the distribution cabinet. A three-parameter synchronous acquisition terminal is used, which integrates a temperature sensor, an ultrasonic partial discharge sensor, and a triaxial vibration sensor. These three sensors are integrated into the same acquisition unit to ensure the synchronization of acquisition triggering. The parameters of each sensor are set according to the following standards to ensure data validity:

[0067] Temperature data: PT100 platinum resistance sensor, measurement range -20°C. Up to 150 The measurement accuracy is The sampling frequency is 1 Hz. This parameter setting is based on the normal operating temperature range of the distribution cabinet: -5°C. Up to 85 The maximum temperature rise during a malfunction can reach 120°C. A 1 Hz sampling frequency can capture slow temperature change trends while avoiding data redundancy, ensuring that the generated temperature data reflects the changes in the thermal state of the components.

[0068] Partial discharge data: An ultrasonic sensor is used, with a detection frequency range of 20 kHz to 200 kHz and a sensitivity of no more than 5 picocuts. The sampling frequency is 10 kHz. This frequency range can effectively capture ultrasonic signals of partial discharge generated by insulation aging and loose contacts inside the distribution cabinet. The sensitivity of 5 picocuts meets the testing requirements for medium-voltage equipment in the DL / T 846 2018 Guidelines for Partial Discharge Measurement of High Voltage Equipment, ensuring that the generated partial discharge data can accurately reflect the relevant characteristics of insulation faults.

[0069] Vibration data: A MEMS triaxial accelerometer was used, with a measurement range of [missing information]. The sample size is 5g, with a sampling frequency of 100 Hz and a resolution of 1 milligram. These parameters are designed for low-frequency vibrations (10 Hz to 50 Hz) caused by switch operations and loose components inside the distribution cabinet. The 100 Hz sampling frequency can completely reproduce the vibration waveform characteristics, ensuring that the generated vibration data can capture signals related to mechanical faults.

[0070] A dual mechanism of hard synchronization triggering and timestamp calibration is employed to ensure the time consistency of temperature, partial discharge, and vibration data. The specific process is as follows: The acquisition terminal obtains standard time through a unified GPS timing module. All sensors initiate acquisition under the same trigger signal, avoiding data asynchrony caused by differences in start time; the time synchronization error is no greater than 1 millisecond. Each sensor continuously acquires corresponding data according to set parameters: the temperature sensor generates one temperature data point every 1 second, the partial discharge sensor generates one partial discharge data point every 0.1 seconds, and the vibration sensor generates one vibration data point every 0.01 seconds. Raw signals are stored in real-time during data acquisition without any filtering or modification to ensure data integrity. Each frame of acquired data is appended with a timestamp accurate to the millisecond level, in the format YYYY MM DD HH MM SSmmm. The timestamp originates from the same source as the acquisition trigger signal, ensuring a consistent time base for the three types of data. After acquisition, the three types of data are matched point-by-point based on the timestamp, integrating the temperature, partial discharge, and vibration data corresponding to the same timestamp into a single data record, forming a multi-dimensional data set with consistent time dimensions.

[0071] Each multi-dimensional dataset corresponds to a set of exclusive metadata, containing four core fields: monitoring point identifier, sensor type identifier, acquisition terminal number, and data quality identifier. The monitoring point identifier is the unique identifier registered in S11, such as the main switch K1 incoming line terminal, directly associated with the physical component corresponding to the multi-dimensional dataset. The sensor type identifier uses T to represent temperature sensors, PD to represent partial discharge sensors, and V to represent vibration sensors, clearly indicating the data source for each type of data in the multi-dimensional dataset. The acquisition terminal number distinguishes multiple acquisition terminals, such as acquisition terminal 01, suitable for multi-terminal deployment scenarios in large distribution cabinets. The data quality identifier uses 0 to represent normal data, 1 to represent suspicious data, and 2 to represent invalid data, facilitating rapid filtering of qualified multi-dimensional datasets during subsequent data preprocessing.

[0072] The generated metadata fields are linked and bound to the corresponding multi-dimensional data sets one by one. Each multi-dimensional data set has a newly added metadata attribute. This metadata, as supplementary information to the data unit, is stored along with the multi-dimensional data set, enabling a direct mapping between the multi-dimensional data set and monitoring points / physical components. This resolves the issue of ambiguous data attribution and ensures that each data unit is traceable to its specific collection location and device. All multi-dimensional data sets with bound metadata are sorted according to a dual rule: first, by timestamp from earliest to latest to ensure temporal continuity; then, within the same timestamp, by monitoring point identifier order to facilitate rapid location of contemporaneous data from different components. The sorted sequence forms a complete multi-source monitoring sequence, which, with time as its axis, sequentially contains the multi-dimensional data sets and corresponding metadata for each monitoring point at different times.

[0073] Step 102: Based on the feature templates corresponding to the components associated with the monitoring points, extract and fuse target features from the multi-source monitoring sequences to generate a comprehensive feature vector.

[0074] Furthermore, step 102 may include the following sub-steps:

[0075] S21. According to the electrical topology of the distribution cabinet, match the key stress types of each physical component in the distribution cabinet from the preset component stress rule library, and determine the typical fault modes associated with the key stress types based on the historical fault database.

[0076] In this embodiment of the invention, the electrical topology is the connection relationship of the main circuit and control circuit of the power distribution cabinet and the list of component topology associations analyzed in step 101. It is the core basis for matching key stress types and ensures that the stress type corresponds accurately to the actual working scenario of the component.

[0077] The preset component stress rule library is a structured database built based on distribution cabinet design specifications, industry operation and maintenance standards, and component material characteristics. The library stores mapping rules between various component types and key stress types, stored in a four-dimensional structure of "component type - topology location - operating parameter range - key stress type". For example, the key stress types corresponding to the main switch component (component type) - incoming terminal (topology location) - rated current 100A (operating parameter range) are electrical stress and thermal stress; the key stress types corresponding to the busbar segment connection (component type) - middle section (topology location) - rated voltage 10kV (operating parameter range) are electrical stress and mechanical stress.

[0078] The critical stress type matching process is as follows: using the unique identifier of each physical component in the electrical topology (such as the incoming terminal of the main switch K1) as an index, the preset component stress rule library is searched to match all critical stress types corresponding to the component, ensuring that the stress types cover the main failure causes of the component.

[0079] The historical fault database contains fault records of 300 similar distribution cabinets over 5 years. Each record includes core fields such as fault component identification, fault occurrence time, operating condition parameters, key stress exceeding standards, fault manifestation, and fault cause analysis. A total of 1,200 valid records of 20 typical fault types have been collected.

[0080] For each key stress type identified, all fault records caused by exceeding the limit for that stress type are retrieved from the historical fault database. Fault cause clustering analysis (using the K-means algorithm, with the number of clusters preset to 5-8 based on the fault type distribution) is then performed to extract typical fault modes strongly correlated with that stress type. For example, typical fault modes corresponding to excessive electrical stress are contact erosion and insulation breakdown; typical fault modes corresponding to excessive thermal stress are component deformation and insulation aging; and typical fault modes corresponding to excessive mechanical stress are loose joints and mechanical jamming. Ultimately, each key stress type for a physical component corresponds to 1-3 typical fault modes, ensuring the relevance and representativeness of the fault modes.

[0081] S22. Based on the preset fault mechanism and feature mapping relationship table, extract the features corresponding to each typical fault mode from the multi-source monitoring sequence and generate the target features corresponding to the typical fault modes.

[0082] In this embodiment of the invention, a preset fault mechanism and feature mapping table serves as the core bridge connecting the essence of the fault with data characteristics. The table stores mapping rules using a six-dimensional structure: "typical fault mode - fault mechanism - multi-source data type - feature extraction dimension - extraction method - feature threshold range." This table is constructed based on research findings on electrical equipment fault mechanisms, historical fault data feature analysis, and industry standards. For example, the typical fault mode "contact erosion" (fault mechanism: excessive current leading to increased contact resistance and localized heating) corresponds to multi-source data types such as temperature data and partial discharge data. The feature extraction dimensions are the rate of temperature rise and the average amplitude of the partial discharge signal. The extraction methods are sliding window statistics and peak detection. The feature threshold range is a temperature rise rate > 0.5℃ / min and an average partial discharge amplitude > 300mV.

[0083] First, using typical failure modes as an index, a pre-defined table of failure mechanisms and feature mappings is queried to determine the corresponding multi-source data types, feature extraction dimensions, extraction methods, and feature threshold ranges. Second, based on the monitoring point identifiers in the metadata, target multi-source data for the corresponding component of the typical failure mode is selected from the multi-source monitoring sequence, such as extracting temperature data and partial discharge data corresponding to "contact ablation". Next, the target multi-source data is feature extracted using the extraction method specified in the mapping table: (1) Temperature data feature extraction: For the "component deformation" fault mode, the maximum temperature and temperature fluctuation variance are calculated using a 10-minute sliding window, and the extraction method is arithmetic statistics; (2) Partial discharge data feature extraction: For the "insulation aging" fault mode, the peak value, pulse frequency, and phase distribution entropy of the partial discharge signal are extracted. The peak value and pulse frequency are detected using the threshold detection method (the threshold is 200mV), and the phase distribution entropy is calculated using the information entropy method; (3) Vibration data feature extraction: For the "joint loosening" fault mode, the root mean square value, peak factor, and fundamental frequency amplitude of the vibration signal are extracted using a combination of time domain analysis and frequency domain analysis. The fundamental frequency amplitude is obtained through fast Fourier transform. Finally, the extracted features are validated, and abnormal features that exceed the feature threshold range in the mapping table are removed (e.g., when the partial discharge peak value is <200mV, it is determined to be an invalid feature and is not included in the target features). The valid feature combinations are retained to generate the target features corresponding to the typical fault mode. Each typical failure mode corresponds to 2-4 target features, ensuring that the features can comprehensively characterize the failure mechanism.

[0084] S23. Based on the occurrence probability of typical failure modes in the historical failure database and the historical representation contribution of the target features, calculate the fusion weight coefficient of the target features.

[0085] In this embodiment of the invention, the probability of occurrence corresponding to a typical fault mode is calculated as follows: In the historical fault database, the ratio of the number of fault records for that typical fault mode to the total number of fault records is calculated using the following formula: ,in, For the first The probability of occurrence of a typical failure mode; This represents the number of fault records for this mode. This represents the total number of fault records in the historical fault database. For example, if the number of fault records for the contact erosion fault mode is 180, and the total number of fault records is 1200, then its occurrence probability is... =180 / 1200=0.15.

[0086] The historical representation contribution of a target feature refers to the degree to which that feature contributes to the identification of typical fault modes in historical fault diagnosis. It is calculated using a feature importance evaluation algorithm (random forest algorithm). The specific process is as follows: using fault data and normal operating condition data of the typical fault mode from the historical fault database as samples, and using whether the fault occurred as the label, a random forest classification model is constructed. After the model is trained, the Gini coefficient importance value of each target feature is extracted, and after normalization, it is taken as the historical representation contribution of that target feature, denoted as . (No. The first of the typical failure modes The historical representation contribution of each target feature), the normalization formula is: ,in, For the first The first mode The importance value of each target feature's Gini coefficient; This is the sum of the Gini coefficient importance values ​​of all target features in this model. The formula for calculating the fusion weight coefficient is:

[0087] ;

[0088] in, For the first The first of the typical failure modes The fusion weight coefficients of each target feature; For the first Probability of occurrence of typical failure modes , This represents the number of fault records for this mode. This represents the total number of fault records. For the first The first of the typical failure modes The historical representation contribution of each target feature, after normalization. , For the mode, the The importance value of each target feature's Gini coefficient; This represents the total number of typical failure modes. This represents the number of target features under a single typical fault mode. The fusion weight coefficient ranges from 0 to 1, and the sum of the weight coefficients of all target features is 1, ensuring the rationality of the weighted calculation.

[0089] S24. Use all target features and corresponding fusion weight coefficients to perform weighted calculations to generate a comprehensive feature vector.

[0090] In this embodiment of the invention, all target features are valid feature sets after validity verification under each typical fault mode. They are uniformly numbered and sorted according to the priority of typical fault modes and the order of feature extraction to form a feature sequence. The sorting rule is as follows: target features corresponding to typical fault modes with high occurrence probability are sorted first, and under the same mode, they are sorted from high to low according to the contribution of historical representation, to ensure that important features occupy the front position in the vector, so that the subsequent diagnostic model can quickly focus on key information.

[0091] The weighted feature value is obtained by multiplying the feature value of each target feature with its corresponding fusion weight coefficient. The weighted feature values ​​of all target features are then arranged sequentially according to a preset feature sequence to form a one-dimensional numerical vector. The vector dimension is the same as the total number of target features (36 dimensions in this embodiment). For example, if the weighted feature values ​​of feature sequences F1 to F36 are 0.48, 0.24, ..., 0.03, then the comprehensive feature vector is [0.48, 0.24, ..., 0.03]. This comprehensive feature vector, through the fusion weight coefficient, distinguishes the importance of different features, retaining key information strongly correlated with the fault while eliminating interference from redundant features, resulting in a streamlined dimension and a focus on the essence of the fault.

[0092] Step 103: Using a preset diagnostic model, a dynamic fusion analysis is performed using comprehensive feature vectors and current operating condition parameters to generate a diagnostic dataset.

[0093] Furthermore, the preset diagnostic model includes multiple sub-diagnostic models optimized for different fault types or different component regions. Step 103 may include the following sub-steps:

[0094] S31. Input the current operating condition parameters into the pre-trained weight allocation function to calculate the dynamic decision weights of each sub-diagnostic model.

[0095] In this embodiment of the invention, the current operating condition parameters are a set of key parameters that reflect the real-time operating status of the distribution cabinet, specifically including load rate, ambient temperature, running time, input voltage fluctuation amplitude, and three-phase current imbalance. These parameters are collected in real time by the operating condition monitoring sensors built into the distribution cabinet, and the sampling frequency is consistent with the multi-source monitoring sequence to ensure data time synchronization.

[0096] The pre-trained weight allocation function is a multivariate linear weighted function constructed based on historical fault databases and operating condition-model adaptation experiments. Its core function is to dynamically adjust the decision importance of each sub-diagnostic model according to changes in operating conditions. The pre-training process of the weight allocation function is as follows: using operating condition parameters from historical fault data as input, and the diagnostic accuracy of each sub-diagnostic model under the corresponding operating condition as the label, the function parameters are optimized using a gradient descent algorithm to maximize the overall diagnostic accuracy of the output weights. The formula for calculating the dynamic decision weights is:

[0097] ;

[0098] in, For the first The dynamic decision weights of each sub-diagnostic model range from 0 to 1, and the sum of the weights of all sub-models is 1. For the first The basic weight coefficients of each sub-model are set according to the inherent diagnostic accuracy of the model, such as the short-circuit fault sub-model. ; For the first The sub-model under current operating conditions parameters The following adaptability functions, such as load factor Corresponding compatibility , To achieve the optimal load rate for this model, The maximum load rate threshold; For the first The compensation coefficients of each sub-model are used to balance the basic weights of each model, satisfying... ; In this embodiment, the total number of sub-diagnostic models is [number]. These correspond to short-circuit faults, overheating faults, insulation aging faults, mechanical faults, and poor contact faults, respectively. The dynamic decision weights calculated using this formula allow sub-diagnostic models adapted to the current operating conditions to receive higher weights. For example, when the load rate is >80%, the short-circuit fault sub-model... The dynamic decision weight can be increased from the basic 0.25 to 0.4 to ensure the effectiveness of adaptive diagnosis under working conditions.

[0099] S32. By performing forward reasoning on the comprehensive feature vector through each sub-diagnostic model, candidate fault types, component regions, and initial confidence levels corresponding to the sub-diagnostic models are generated.

[0100] In this embodiment of the invention, the preset diagnostic model includes five independent sub-diagnostic models. Each sub-diagnostic model is optimized for a specific fault type or component area, specifically: Sub-model 1 (short circuit fault - main circuit area), Sub-model 2 (overheating fault - heat-generating component area), Sub-model 3 (insulation aging fault - insulation component area), Sub-model 4 (mechanical fault - mechanical transmission component area), and Sub-model 5 (poor contact fault - connection component area). Each sub-diagnostic model is a deep learning model trained based on historical fault data. Sub-models 1-3 use a CNN network, and sub-models 4-5 use a random forest model. The model structure has been optimized for operating conditions to ensure stable inference capabilities under different operating conditions.

[0101] The comprehensive feature vector is the 36-dimensional weighted feature vector generated in step 102. The forward inference process is executed according to the following logic:

[0102] (1) Filter the comprehensive feature vector according to the feature dimensions corresponding to the sub-model (e.g., sub-model 1 focuses on current fluctuation and partial discharge related features, and filters the first 12 features in the vector).

[0103] (2) After the feature vector is input into the sub-model, the inference result is output after network layer operation (CNN model includes convolutional layer, pooling layer and fully connected layer, random forest model includes decision tree ensemble operation);

[0104] (3) The reasoning results contain three types of core information: candidate fault type (consistent with the fault type corresponding to the sub-model, such as the output of "short circuit fault" of sub-model 1), component area (the specific component range corresponding to the sub-model optimization, such as the output of "main switch-bus area" of sub-model 1), and initial confidence (reflecting the reliability of the reasoning results, with a value range of 0-1, calculated by the softmax function of the model output layer).

[0105] S33. Multiply the initial confidence level by the corresponding dynamic decision weight to generate the weighted confidence level.

[0106] In this embodiment of the invention, the calculation of weighted confidence aims to combine the adaptability to the working conditions with the reliability of model inference, highlighting the weight of the inference results of the sub-models that are adapted to the current working conditions. The formula for calculating weighted confidence is:

[0107] ;

[0108] in, For the first Weighted post-confidence of individual diagnostic models; For the first Initial confidence level of the individual diagnostic model; For the first Dynamic decision weights for individual diagnostic models.

[0109] For example, sub-model 1 has an initial confidence level of 0.8 and a dynamic decision weight of 0.4. Its weighted confidence level is... Sub-model 2 has an initial confidence level of 0.85 and a dynamic decision weight of 0.2. Its weighted confidence level is... This calculation preserves the reliability information of the model inference while incorporating operating condition adaptability weights, making the final confidence level more closely match the real-time operating status of the equipment.

[0110] S34. Use the softmax function to normalize the weighted confidence scores and generate normalized confidence probabilities.

[0111] In this embodiment of the invention, the core function of the softmax function is to transform the weighted confidence scores of each sub-model into a probability distribution, ensuring that the sum of the normalized confidence probabilities of all sub-models is 1, thus facilitating the subsequent screening of highly reliable diagnostic results. The formula for calculating the normalized confidence probability is:

[0112] ;

[0113] in, For the first Normalized confidence probabilities of individual diagnostic models; For the first Weighted post-confidence of individual diagnostic models; The total number of sub-diagnostic models; It is a natural constant with a value of approximately 2.718.

[0114] For example, the weighted confidence scores of the five sub-models are 0.32, 0.17, 0.21, 0.15, and 0.15, respectively. After calculation using the softmax function, their normalized confidence probabilities are 0.34, 0.18, 0.20, 0.14, and 0.14, respectively, summing to 1. This normalization process amplifies the difference between high-weighted and low-weighted confidence scores, making reliable results more prominent.

[0115] S35. Select a normalized confidence probability that is higher than the preset reporting threshold and generate the target confidence probability.

[0116] In this embodiment of the invention, the preset reporting threshold is a confidence screening threshold set based on historical diagnostic data statistics, used to exclude low-reliability diagnostic results and ensure the credibility of the output results.

[0117] The preset reporting threshold is 0.5. This value is determined by statistically analyzing the correlation between the normalized confidence probability and the actual fault in historical fault data. When the normalized confidence probability is ≥0.5, the matching accuracy of the diagnostic results and the actual fault reaches 92%; when it is below 0.5, the matching accuracy is only 35%. Therefore, 0.5 is set as the screening threshold.

[0118] The process of generating the target confidence probability is as follows: the normalized confidence probability of each sub-model is compared with the preset reporting threshold (0.5), and a normalized confidence probability greater than or equal to 0.5 is selected as the target confidence probability; if the normalized confidence probability of all sub-models is less than 0.5, it is determined that there is no clear fault diagnosis result, and the target confidence probability is an empty set. For example, under a certain working condition, the normalized confidence probability of sub-model 1 is 0.52, and the other sub-models are all less than 0.5, then only the normalized confidence probability of sub-model 1 is selected as the target confidence probability.

[0119] S36. Construct a diagnostic dataset using the candidate fault type, component region, and target confidence probability corresponding to the target confidence probability.

[0120] In this embodiment of the invention, the target confidence probability is associated with its corresponding candidate fault type and component region to form a triplet data of candidate fault type-component region-target confidence probability. All triplet data are sorted from high to low according to the target confidence probability to form a complete diagnostic dataset.

[0121] Step 104: Based on the fault location knowledge graph and diagnostic dataset corresponding to the power distribution cabinet, calculate the probability that each component in the power distribution cabinet is a fault source, and generate a component fault probability dataset.

[0122] Furthermore, step 104 may include the following sub-steps:

[0123] S41. Using each physical component in the distribution cabinet as a node and the electrical connection relationship between the components as an edge, configure the prior fault probability based on historical statistics for each node and the conditional probability based on historical fault propagation statistics for each edge to construct a fault location knowledge graph.

[0124] In this embodiment of the invention, the physical components in the distribution cabinet are the monitoring point associated components determined in step 101, including 12 types of core components such as main switch, busbar, circuit breaker, transformer, and cable joint. Each physical component corresponds to a unique node in the knowledge graph, and the node identifier is consistent with the monitoring point identifier to ensure the continuity of data traceability.

[0125] 1. Knowledge Graph Topology Construction

[0126] (1) Node definition: Each node stores the basic attributes of the physical component, including component name, model, rated parameters, installation location, etc. The core associated attribute is "prior failure probability";

[0127] (2) Edge definition: The electrical connection relationship between components is taken as the edge, such as "main switch K1 incoming terminal → bus middle section → circuit breaker K2 outgoing terminal". The direction of the edge is consistent with the current flow direction, which represents the possible propagation path of the fault. The core attribute of the edge is "conditional probability".

[0128] (3) Topology verification: After the construction is completed, check the node connection relationship with the electrical wiring diagram of the distribution cabinet to ensure that the edge association is consistent with the actual electrical topology and avoid probability propagation deviation due to topology errors.

[0129] 2. Prior Fault Probability Configuration

[0130] Prior failure probability refers to the statistical probability of a physical component failing without any observational evidence. It is calculated based on a historical failure database. The calculation formula is:

[0131] ;

[0132] in, For the first The prior failure probability of each node (physical component); This represents the total number of fault records for this component in the historical fault database. For example, the number of faults at the input terminal of the main switch K1 is 36. This refers to the average service life of the component, such as the average service life of the incoming line terminal of the main switch K1 being 8 years. This refers to the average number of times a component operates per year. For example, the average number of times a main switch component operates per year is 365.

[0133] For example: the input terminal of main switch K1 Then its prior fault probability That is, 1.23%. The statistical prior failure probability is based on 5 years of operation and maintenance data from 300 similar distribution cabinets to ensure the representativeness and reliability of the data.

[0134] 3. Conditional probability configuration

[0135] Conditional probability refers to the probability that when a component (parent node) fails, its associated components (child nodes) will also fail due to fault propagation. It is also derived from statistical analysis of fault propagation records in a historical fault database. The calculation formula is:

[0136] ;

[0137] in, In the first Under the condition that the first node (parent component) fails, the second... The conditional probability of a node (sub-component) failing; For the historical fault database, due to the first The failure propagated from one component, leading to the first... The number of times a component failure is recorded, such as the number of times a failure at the input end of the main switch K1 caused a failure in the middle section of the busbar, which was 12 times. For the first Total number of fault records for each component.

[0138] For example: Main switch K1 input terminal With the middle section of the busbar of =12、 =36, then the conditional probability That is, 33.3%. The conditional probability of all edges is calculated according to this logic to ensure the quantitative accuracy of the failure propagation probability.

[0139] S42. Use the diagnostic dataset as observational evidence and map it to the corresponding nodes and fault types in the knowledge graph.

[0140] In this embodiment of the invention, the mapping types include: (1) Component region → node mapping: The component region in the diagnostic dataset corresponds to one or more nodes in the knowledge graph, such as the main switch-bus region corresponding to the two nodes of the main switch K1 incoming end and the bus middle section. The mapping is based on the physical range of the component region and the topological association of the node identifier; (2) Candidate fault type → node fault attribute mapping: Each node predefines 3-5 possible fault type attributes (such as the fault type attributes of the main switch K1 incoming end include short circuit fault, poor contact fault, and overheating fault). The candidate fault type in the diagnostic dataset directly matches the fault type attribute corresponding to the node; (3) Target confidence probability → evidence weight mapping: The target confidence probability in the diagnostic dataset is used as the weight of the observed evidence to characterize the reliability of the evidence. The weight value is consistent with the target confidence probability (such as the target confidence probability of 0.52 corresponds to the evidence weight of 0.52).

[0141] The mapping process is as follows: traverse each triplet in the diagnostic dataset, extract candidate fault types, component regions, and target confidence probabilities; filter target nodes in the knowledge graph based on component regions, such as filtering the main switch K1 incoming terminal and bus mid-section nodes when the component region is the main switch-bus region; match the corresponding fault type attribute for each target node, such as activating the short-circuit fault attribute for the target node when the candidate fault type is a short-circuit fault; use the target confidence probability as the observation evidence weight of the corresponding fault type attribute of the node, store it in the node attributes of the knowledge graph, and form an association record of node-fault type-observation evidence weight.

[0142] For example, in the diagnostic dataset, the triplet short-circuit fault-main switch-bus region has a weight of 0.52. After mapping, the short-circuit fault attribute of the main switch K1 incoming node in the knowledge graph is associated with the observation evidence weight of 0.52, and the short-circuit fault attribute of the bus mid-section node is also associated with the observation evidence weight of 0.52, thus completing the accurate mapping of the observation evidence.

[0143] S43. Based on observational evidence, the prior fault probability of each node, and the conditional probability of each edge, perform iterative probability propagation calculations on the topological structure of the knowledge graph to update the posterior probability of each node as a fault source.

[0144] In this embodiment of the invention, the probability propagation calculation employs the Bayesian network belief propagation algorithm. Its core principle is to correct the prior fault probability of each node based on observed evidence, and then propagate the fault probability in the knowledge graph through iterative calculation, ultimately obtaining the posterior probability of each node as a fault source. The specific iterative process is as follows:

[0145] 1. Initial probability settings: (1) Nodes not associated with observation evidence: the initial probability is equal to their prior failure probability; (2) Nodes associated with observation evidence: the initial probability is the product of their prior failure probability and the weight of the observation evidence.

[0146] 2. Iterative Probability Propagation Calculation: Probability propagation proceeds in the direction of "parent node → child node". The probability of the child node is updated in each iteration. The iteration process is as follows:

[0147] (1) Calculation of single-step propagation probability

[0148] The propagation probability of a child node is the weighted sum of the current probabilities of all parent nodes and the conditional probabilities of their corresponding edges, calculated using the following formula:

[0149] ;

[0150] in, For the first After the nth iteration The probability of each child node; For the first After -1 iterations, the th The probability of each parent node; parent node To child nodes The conditional probability; child node The total number of parent nodes; The attenuation coefficient, with a value of 0.95, is used to characterize the probability attenuation during fault propagation and is set based on historical fault propagation intensity statistics.

[0151] (2) Iteration termination condition

[0152] The iteration terminates when the probability change of all nodes is less than the preset convergence threshold, which is set to 0.001. The formula for calculating the probability change is:

[0153] ;

[0154] in, This represents the maximum probability change across all nodes. For the first After the nth iteration The probability of each node; For the first After the nth iteration The probability of each node.

[0155] when When the value is less than 0.001, the iteration is considered to have converged, and the probability propagation calculation is stopped.

[0156] 3. Posterior probability determination

[0157] After the iteration converges, the final probability of each node is the posterior probability of that node as a fault source. This probability includes the node's own prior fault characteristics, as well as the constraints of observational evidence and the influence of fault propagation, thus quantifying the possibility of the node as a fault source.

[0158] S44. Construct a component failure probability dataset using each node and its updated posterior probability.

[0159] In this embodiment of the invention, the component failure probability dataset is a structured collection of probability information. Its core function is to integrate the posterior probabilities of all nodes, providing clear and standardized input data for subsequent hierarchical alarms and ensuring the accuracy of alarm judgment. The construction process of the component failure probability dataset is as follows: each data item contains three core fields: node identifier (physical component identifier), posterior probability, and corresponding fault type. The corresponding fault type is the fault type attribute with the highest posterior probability for that node (e.g., short circuit fault at the input terminal of main switch K1). All data items are sorted from high to low posterior probability to facilitate quick location of high-probability fault sources. After construction, the posterior probability value range (0-1) of all data items is verified, and abnormal data exceeding the range is removed to ensure the validity of the dataset.

[0160] Step 105: Identify components in the component failure probability dataset that exceed the preset failure threshold as faulty components, and generate corresponding fault alarm data for the faulty components.

[0161] Furthermore, step 105 may include the following sub-steps:

[0162] S51. Select components whose posterior probability in the component failure probability dataset exceeds the first preset threshold as first-level fault components.

[0163] In this embodiment of the invention, the first preset threshold is a high-risk judgment standard set based on historical fault handling data statistics, used to screen components that have clearly failed and require emergency handling. The first preset threshold is set at 90%, determined based on: based on 1200 fault records from 300 similar distribution cabinets, using 50% cross-validation statistics, when the posterior probability of a component exceeds 90%, the matching accuracy with the actual fault reaches 98%, and the false positive rate is ≤1.2% and the false negative rate is ≤0.8%; the fault handling response time requirement corresponding to this threshold is ≤2 hours, which can avoid the risk of distribution cabinet downtime due to delayed fault handling.

[0164] The selection process for Level 1 faulty components is as follows: Traverse all data items in the component fault probability dataset, extract the posterior probability of each component, compare it with a first preset threshold (90%), and select components with a posterior probability > 90% as Level 1 faulty components. For example, the posterior probability of the main switch K1 incoming line terminal is 92%, exceeding the first preset threshold, and is therefore determined to be a Level 1 faulty component; the posterior probability of the busbar mid-section is 85%, not exceeding the threshold, and is not included in the Level 1 faulty component list.

[0165] S52. Select components whose posterior probability in the component failure probability dataset exceeds the second preset threshold but does not exceed the first preset threshold as secondary failure components; wherein, the first preset threshold is higher than the second preset threshold.

[0166] In this embodiment of the invention, the second preset threshold is a criterion for screening components with potential fault risks that require continuous monitoring. The threshold value is 60%, lower than the first preset threshold (90%). The second preset threshold of 60% is determined based on the following: through historical data statistics (5 years of operation and maintenance data from 300 devices), components with a posterior probability between 60% and 90% have a 75% probability of actually having potential fault hazards, of which 40% have a probability of developing into an obvious fault within one month. The false positive rate corresponding to this threshold is ≤5%, which can balance monitoring costs and fault early warning sensitivity.

[0167] The selection process for secondary fault components is as follows: traverse all data items in the component fault probability dataset, extract the posterior probability of each component, and select components with a posterior probability > 60% and ≤ 90% as secondary fault components. For example, the posterior probability of the busbar mid-section is 85%, which meets the condition of exceeding 60% but not exceeding 90%, and is therefore determined to be a secondary fault component; the posterior probability of the circuit breaker K2 outgoing terminal is 55%, which does not exceed the second preset threshold, and is not included in the secondary fault component list.

[0168] S53. Extract typical fault causes and maintenance measures of similar components corresponding to first-level faulty components from the historical maintenance database, and generate high-priority alarm information.

[0169] In this embodiment of the invention, the historical maintenance database is a structured database built based on five years of operation and maintenance data from 300 similar distribution cabinets. The stored content includes associated records of component type, fault type, typical fault causes, maintenance steps, required tools, and precautions. It contains 800 valid maintenance records across 12 component categories and 20 fault types, ensuring the relevance and operability of maintenance information. The generation process for high-priority alarm information is as follows:

[0170] 1. Extract the core information of the primary faulty component, including component identification (such as the incoming terminal of main switch K1) and corresponding fault type (such as short circuit fault).

[0171] 2. Using component type + fault type as a combined index, search the historical maintenance database to extract typical fault causes for similar components (e.g., typical causes of main switch short circuit faults: contact oxidation and adhesion, and insulation damage at the input terminal).

[0172] 3. Simultaneously extract the corresponding maintenance measures, including specific maintenance steps (e.g., 1. Disconnect the main power supply of the distribution cabinet; 2. Disassemble the main switch casing; 3. Clean the oxide layer of the contacts or replace the contacts; 4. Check the insulation layer of the incoming line, wrap the damaged parts with insulation or replace the wires; 5. Test the power after assembly), required tools (e.g., insulated wrench, multimeter, insulating tape, new contact assembly), and safety precautions (e.g., before maintenance, confirm that the main power supply is completely disconnected, test for voltage before operation; avoid metal tools touching live parts).

[0173] 4. Integrate the identification of first-level faulty components, post-fault probability, typical fault causes, maintenance measures, and safety precautions to form structured high-priority alarm information, ensuring that maintenance personnel can directly carry out emergency response based on the information.

[0174] S54. Based on the component failure probability data generated within a preset historical period for secondary faulty components, determine the failure probability change trend characteristics corresponding to the secondary faulty components.

[0175] In this embodiment of the invention, the core information of the secondary fault component includes the component identifier, the current posterior probability, and the corresponding fault type, which is the object of trend analysis. The preset historical period is set to 3 months. This period length has been tested and can effectively reflect the changing pattern of the fault probability, which can avoid misjudgment caused by short-term fluctuations and capture the development trend of potential faults in a timely manner.

[0176] The generated component failure probability data refers to the posterior probability data generated by the secondary failure component in each data acquisition cycle (consistent with the sampling frequency of the multi-source monitoring sequence, 1 time / second) within the past 3 months, forming a time-continuous probability sequence (such as "2024-01-01 00:00:00-0.62; 2024-01-02 00:00:00-0.65; ...; 2024-03-31 23:59:59-0.85").

[0177] The process for determining the trend characteristics of failure probability changes is as follows:

[0178] 1. Extract all posterior probability data of secondary faulty components within a preset historical period (3 months) and arrange them in order of timestamps to form a probability sequence;

[0179] 2. A linear regression algorithm is used to fit the probability sequence and calculate the trend slope, which reflects the rate of change of the failure probability.

[0180] 3. By combining the trend slope and the fluctuation of the probability sequence, determine the trend characteristic type, including three core types:

[0181] (1) Rapid upward trend: The trend slope is >0.1 / month (that is, the failure probability increases by more than 10 percentage points per month), and the probability fluctuation variance is <0.005 (small fluctuation and stable rise).

[0182] (2) Slow upward trend: 0.03 / month ≤ trend slope ≤ 0.1 / month (the probability of failure increases by 3-10 percentage points per month), and the variance of fluctuation is <0.01;

[0183] (3) Stable trend: |Trend slope| < 0.03 / month (the monthly failure probability changes by no more than 3 percentage points), and fluctuation variance < 0.01.

[0184] For example, after fitting the probability sequence of the middle section of the busbar of the secondary fault component within 3 months, the trend slope is 0.08 / month and the fluctuation variance is 0.003, indicating that its fault probability change trend is a slow upward trend.

[0185] S55. Format the characteristics of the failure probability change trend to generate a trend analysis conclusion that includes trend description and risk level.

[0186] In this embodiment of the invention, the trend characteristics of fault probability change are the core trend information of secondary faulty components. The purpose of formatting is to transform abstract trend data into intuitive and easy-to-understand judgment conclusions, providing core content for observation-level early warning information. The generation rules for trend judgment conclusions are as follows:

[0187] 1. Trend Description: Based on the trend feature type, standardized language is used to describe it, such as rapid increase in failure probability, slow increase in failure probability, and basically stable failure probability.

[0188] 2. Risk Level: Based on a comprehensive assessment of the trend slope and the current posterior probability, it is divided into three levels: high, medium, and low.

[0189] High risk: Rapidly rising trend, and the current posterior probability is ≥80% (easily develops into a level 1 failure in the short term);

[0190] Medium risk: slow upward trend, or rapid upward trend but current posterior probability <80% (may develop into a level 1 failure in the medium term).

[0191] Low risk: Stable trend, and the current posterior probability is ≤70% (no obvious fault risk in the short term, but continuous monitoring is required).

[0192] After formatting, the structure of the trend analysis conclusion is trend description + risk level, such as: slow increase in failure probability + medium risk, rapid increase in failure probability + high risk, and basically stable failure probability + low risk, to ensure that maintenance personnel can quickly grasp the development trend of potential failures.

[0193] S56. Using trend analysis conclusions and the component identifiers and current failure probabilities of secondary faulty components, construct observation-level early warning information.

[0194] In this embodiment of the invention, the observation-level early warning information is monitoring information generated for potential risks of secondary faulty components. Its core function is to remind maintenance personnel to pay attention to changes in component status and formulate preventative maintenance plans. The construction process of the observation-level early warning information is as follows:

[0195] 1. Extract the core information of the secondary faulty component, including component identification and current posterior probability;

[0196] 2. Incorporate trend analysis conclusions;

[0197] 3. Supplementary monitoring recommendations (generated based on historical data, such as recommending to collect fault probability data once a week, with a focus on monitoring changes in bus temperature).

[0198] 4. Integrate the above information to form structured observation-level early warning information, such as: Component identification: busbar mid-section; Current fault probability: 85%; Trend analysis conclusion: Fault probability is slowly increasing + medium risk; Monitoring suggestion: Collect fault probability data once a week, focusing on monitoring changes in busbar temperature to ensure that the information is complete and operable.

[0199] S57. Summarize high-priority alarm information and observation-level early warning information to generate fault alarm data.

[0200] In this embodiment of the invention, fault alarm data is a comprehensive data set integrating emergency fault handling information and potential fault monitoring information. Its core function is to provide maintenance personnel with comprehensive and hierarchical fault handling guidance, thereby improving maintenance efficiency. The process of summarizing fault alarm data is as follows: high-priority alarm information and observation-level early warning information are classified and organized. High-priority alarm information is sorted from high to low posterior probability (to facilitate prioritizing the handling of high-risk faults), and observation-level early warning information is sorted from high to low risk level (to facilitate focusing on high-risk potential faults). A unified alarm number (format: alarm-date-serial number, such as alarm-20240520-001) and generation time (accurate to the second) are added to both types of information for easy traceability and management. The classified and sorted high-priority alarm information, observation-level early warning information, and basic identification information are integrated to form complete fault alarm data.

[0201] Please see Figure 2 , Figure 2This is a structural block diagram of a power distribution cabinet fault location system provided in an embodiment of the present invention.

[0202] The present invention provides a power distribution cabinet fault location system, comprising:

[0203] The multi-source synchronous acquisition module 201 is used to synchronously acquire sensor data from multiple monitoring points deployed in the power distribution cabinet according to the electrical topology, and generate a multi-source monitoring sequence.

[0204] The mechanism feature fusion module 202 is used to extract and fuse target features from multi-source monitoring sequences based on feature templates corresponding to the components associated with monitoring points, and generate a comprehensive feature vector.

[0205] The adaptive diagnostic module 203 is used to generate a diagnostic dataset by dynamically fusing and analyzing comprehensive feature vectors and current operating condition parameters through a preset diagnostic model.

[0206] The probabilistic localization reasoning module 204 is used to calculate the probability that each component in the power distribution cabinet is a fault source based on the fault localization knowledge graph and diagnostic dataset corresponding to the power distribution cabinet, and generate a component fault probability dataset.

[0207] The hierarchical intelligent alarm module 205 is used to identify components whose failure probability data exceeds a preset failure threshold as faulty components and generate corresponding fault alarm data for the faulty components.

[0208] Furthermore, the multi-source synchronous acquisition module 201 can perform the following steps:

[0209] Based on the electrical wiring diagram of the distribution cabinet, physical nodes that characterize the key electrical connection status are used as monitoring points;

[0210] Temperature, partial discharge, and vibration data are collected simultaneously at monitoring points to generate time-aligned multi-source monitoring sequences.

[0211] Furthermore, the mechanism feature fusion module 202 can perform the following steps:

[0212] According to the electrical topology of the distribution cabinet, the key stress types of each physical component in the distribution cabinet are matched from the preset component stress rule library, and the typical failure modes associated with the key stress types are determined based on the historical failure database.

[0213] Based on a pre-defined fault mechanism and feature mapping relationship table, features corresponding to each typical fault mode are extracted from multi-source monitoring sequences to generate target features corresponding to typical fault modes.

[0214] Based on the occurrence probability of typical failure modes in the historical failure database, and combined with the historical representation contribution of the target features, the fusion weight coefficient of the target features is calculated.

[0215] A comprehensive feature vector is generated by weighting all target features and their corresponding fusion weight coefficients.

[0216] Furthermore, the preset diagnostic model includes multiple sub-diagnostic models optimized for different fault types or different component regions; the operating condition adaptive diagnostic module 203 can perform the following steps:

[0217] The current operating condition parameters are input into the pre-trained weight allocation function to calculate the dynamic decision weights of each sub-diagnostic model.

[0218] Each sub-diagnostic model performs forward inference on the comprehensive feature vector to generate candidate fault types, component regions, and initial confidence levels corresponding to the sub-diagnostic models.

[0219] The initial confidence level is multiplied by the corresponding dynamic decision weight to generate a weighted confidence level.

[0220] The softmax function is used to normalize the weighted confidence scores to generate normalized confidence probabilities.

[0221] Select a normalized confidence probability that is higher than the preset reporting threshold to generate a target confidence probability;

[0222] A diagnostic dataset is constructed using candidate fault types, component regions, and target confidence probabilities corresponding to the target confidence probabilities.

[0223] Furthermore, the probabilistic localization reasoning module 204 can perform the following steps:

[0224] Using each physical component in the distribution cabinet as a node and the electrical connection relationship between the components as an edge, a priori fault probability based on historical statistics is configured for each node, and a conditional probability based on historical fault propagation statistics is configured for each edge, thus constructing a fault location knowledge graph.

[0225] The diagnostic dataset is used as observational evidence and mapped to the corresponding nodes and fault types in the knowledge graph.

[0226] Based on observational evidence, the prior failure probability of each node, and the conditional probability of each edge, iterative probability propagation calculations are performed on the topological structure of the knowledge graph to update the posterior probability of each node as a failure source.

[0227] A component failure probability dataset is constructed using each node and its updated posterior probability.

[0228] Furthermore, the hierarchical intelligent alarm module 205 can perform the following steps:

[0229] Components whose posterior probability in the component failure probability dataset exceeds a first preset threshold are selected as first-level faulty components.

[0230] Components whose posterior probability in the component failure probability dataset exceeds a second preset threshold but does not exceed a first preset threshold are selected as secondary failure components; wherein, the first preset threshold is higher than the second preset threshold.

[0231] Extract typical fault causes and maintenance measures of similar components corresponding to the first-level faulty components from the historical maintenance database, and generate high-priority alarm information;

[0232] Based on the component failure probability data generated by the secondary faulty component within a preset historical period, determine the failure probability change trend characteristics of the secondary faulty component.

[0233] The fault probability change trend characteristics are formatted to generate trend judgment conclusions that include trend descriptions and risk levels;

[0234] Observation-level early warning information is constructed by using trend analysis conclusions, component identifiers corresponding to secondary faulty components, and their current fault probabilities.

[0235] It aggregates high-priority alarm information and observation-level early warning information to generate fault alarm data.

[0236] Please see Figure 3 , Figure 3 This is a structural block diagram of an electronic device provided in an embodiment of the present invention.

[0237] An electronic device according to an embodiment of the present invention includes: a memory 301 and a processor 302. The memory 301 stores a computer program. When the computer program is executed by the processor 302, the processor 302 performs the power distribution cabinet fault location method as described in any of the above embodiments.

[0238] Memory 301 may be an electronic memory such as flash memory, EEPROM (Electrically Erasable Programmable Read-Only Memory), EPROM, hard disk, or ROM. Memory 301 has storage space 303 for program code 313 for performing any of the method steps described above. For example, storage space 303 for program code may include various program codes 313 for implementing the various steps in the methods described above. These program codes may be read from or written to one or more computer program products. These computer program products include program code carriers such as hard disks, CDs, memory cards, or floppy disks. The program code may be compressed, for example, in a suitable form. When run by a computing processing device, this code causes the computing processing device to perform the various steps in the methods described above. These program codes may be read from or written to one or more computer program products. These computer program products include program code carriers such as hard disks, CDs, memory cards, or floppy disks. The program code may be compressed, for example, in a suitable form. When these codes are run by a computing processing device, the device causes it to perform the various steps in the power distribution cabinet fault location method described above.

[0239] This invention also provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the power distribution cabinet fault location method as described in any of the above embodiments.

[0240] This invention also provides a computer program product, which includes a computer program stored on a non-transitory computer-readable storage medium. The computer program includes program instructions, wherein when the program instructions are executed by a computer, the computer performs the power distribution cabinet fault location method as described in any of the above embodiments.

[0241] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0242] In the several embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces, or indirect coupling or communication connection between apparatuses or units, and may be electrical, mechanical, or other forms.

[0243] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0244] Furthermore, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0245] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0246] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for locating faults in a power distribution cabinet, characterized in that, include: Simultaneously collect sensor data from multiple monitoring points deployed in the power distribution cabinet according to electrical topology to generate a multi-source monitoring sequence; Based on the feature templates corresponding to the components associated with the monitoring points, target features are extracted and fused from the multi-source monitoring sequence to generate a comprehensive feature vector. A diagnostic dataset is generated by dynamically fusing the comprehensive feature vector and current operating condition parameters using a preset diagnostic model. Based on the fault location knowledge graph corresponding to the power distribution cabinet and the diagnostic dataset, the probability that each component in the power distribution cabinet is a fault source is calculated, and a component fault probability dataset is generated. Components in the component failure probability dataset that exceed a preset failure threshold are identified as faulty components, and corresponding fault alarm data is generated for the faulty components.

2. The method for locating faults in a power distribution cabinet according to claim 1, characterized in that, The step of synchronously acquiring sensor data from multiple monitoring points deployed according to electrical topology within the distribution cabinet and generating a multi-source monitoring sequence includes: Based on the electrical wiring diagram of the distribution cabinet, physical nodes that characterize the key electrical connection status are used as monitoring points; Temperature, partial discharge, and vibration data are collected synchronously at the monitoring points to generate a time-aligned multi-source monitoring sequence.

3. The method for locating faults in a power distribution cabinet according to claim 1, characterized in that, The step of extracting and fusing target features from the multi-source monitoring sequence based on the feature template corresponding to the monitoring point associated component to generate a comprehensive feature vector includes: According to the electrical topology of the distribution cabinet, the key stress types of each physical component in the distribution cabinet are matched from the preset component stress rule library, and the typical fault modes associated with the key stress types are determined based on the historical fault database. Based on a preset fault mechanism and feature mapping relationship table, features corresponding to each typical fault mode are extracted from the multi-source monitoring sequence to generate target features corresponding to the typical fault modes. Based on the occurrence probability of the typical fault modes in the historical fault database, and combined with the historical representation contribution of the target feature, the fusion weight coefficient of the target feature is calculated. A comprehensive feature vector is generated by weighting all the target features and their corresponding fusion weight coefficients.

4. The method for locating faults in a power distribution cabinet according to claim 1, characterized in that, The preset diagnostic model includes multiple sub-diagnostic models optimized for different fault types or different component areas; the step of generating a diagnostic dataset by dynamically fusing the comprehensive feature vector and current operating condition parameters through the preset diagnostic model includes: The current operating condition parameters are input into the pre-trained weight allocation function to calculate the dynamic decision weights of each sub-diagnostic model. Each of the sub-diagnostic models performs forward reasoning on the comprehensive feature vector to generate candidate fault types, component regions, and initial confidence levels corresponding to the sub-diagnostic models. The initial confidence level is multiplied by the corresponding dynamic decision weight to generate a weighted confidence level. The weighted confidence scores are normalized using the softmax function to generate normalized confidence probabilities. Select the normalized confidence probability that is higher than the preset reporting threshold to generate the target confidence probability; A diagnostic dataset is constructed using the candidate fault type, component region, and target confidence probability corresponding to the target confidence probability.

5. The method for locating faults in a power distribution cabinet according to claim 1, characterized in that, The step of calculating the probability that each component in the power distribution cabinet is a fault source based on the fault location knowledge graph corresponding to the power distribution cabinet and the diagnostic dataset, and generating a component fault probability dataset, includes: Using each physical component in the power distribution cabinet as a node and the electrical connection relationship between the components as an edge, a priori fault probability based on historical statistics is configured for each node, and a conditional probability based on historical fault propagation statistics is configured for each edge, thereby constructing a fault location knowledge graph. The diagnostic dataset is used as observational evidence and mapped to the corresponding nodes and fault types in the knowledge graph. Based on the observed evidence, the prior fault probability of each node, and the conditional probability of each edge, iterative probability propagation calculations are performed on the topology of the knowledge graph to update the posterior probability of each node as a fault source. A component failure probability dataset is constructed using each node and its updated posterior probability.

6. The method for locating faults in a power distribution cabinet according to claim 5, characterized in that, The step of identifying components in the component failure probability dataset that exceed a preset failure threshold as faulty components and generating corresponding fault alarm data for the faulty components includes: Components whose posterior probability in the component failure probability dataset exceeds a first preset threshold are selected as first-level faulty components. Components whose posterior probability in the component failure probability dataset exceeds a second preset threshold but does not exceed a first preset threshold are selected as secondary failure components; wherein, the first preset threshold is higher than the second preset threshold; Extract typical fault causes and maintenance measures of similar components corresponding to the first-level faulty components from the historical maintenance database, and generate high-priority alarm information; Based on the component failure probability data generated by the secondary faulty component within a preset historical period, the failure probability change trend characteristics corresponding to the secondary faulty component are determined. The fault probability change trend characteristics are formatted to generate a trend judgment conclusion that includes trend description and risk level; Using the trend analysis conclusions and the component identifiers and current failure probabilities corresponding to the secondary faulty components, observation-level early warning information is constructed; The high-priority alarm information and the observation-level early warning information are combined to generate fault alarm data.

7. A power distribution cabinet fault location system, characterized in that, include: The multi-source synchronous acquisition module is used to synchronously acquire sensor data from multiple monitoring points deployed in the power distribution cabinet according to the electrical topology, and generate a multi-source monitoring sequence. The mechanism feature fusion module is used to extract and fuse target features from the multi-source monitoring sequence based on the feature templates corresponding to the monitoring point associated components, and generate a comprehensive feature vector. The adaptive diagnostic module is used to dynamically fuse and analyze the comprehensive feature vector and current operating condition parameters through a preset diagnostic model to generate a diagnostic dataset. The probabilistic localization reasoning module is used to calculate the probability that each component in the power distribution cabinet is a fault source based on the fault localization knowledge graph corresponding to the power distribution cabinet and the diagnostic dataset, and generate a component fault probability dataset. The hierarchical intelligent alarm module is used to identify components in the component failure probability dataset that exceed a preset failure threshold as faulty components and generate corresponding fault alarm data for the faulty components.

8. An electronic device, characterized in that, The device includes a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor performs the steps of the power distribution cabinet fault location method as described in any one of claims 1-6.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed, it implements the power distribution cabinet fault location method as described in any one of claims 1-6.

10. A computer program product, characterized in that, The computer program product includes a computer program stored on a non-transitory computer-readable storage medium, the computer program including program instructions, wherein when the program instructions are executed by a computer, the computer performs the power distribution cabinet fault location method as described in any one of claims 1-6.

Images