A power distribution station equipment defect discrimination method, device and medium

Abstract

The application relates to the technical field of equipment intelligent diagnosis, and discloses a power distribution station equipment defect discrimination method, equipment and medium, which belong to the technical field of equipment intelligent diagnosis. Image-text-time sequence triple data is subjected to feature extraction, and a three-mode contrast learning framework is used for joint optimization to generate a multi-mode joint embedding vector, a preliminary defect type and a confidence degree. A power distribution station equipment topology graph is constructed, a structured discrimination result is introduced into the power distribution station equipment topology graph, a neighbor state is aggregated by using a graph neural network, and a defect risk score is obtained. In combination with the domain features of the current power distribution station, the defect risk score is subjected to feature distribution alignment, a defect judgment list is obtained, the power distribution station equipment is overhauled, a structured defect report is generated, and persistent storage is carried out. Through the neighbor state aggregation driven by the three-mode contrast learning and the graph neural network, the application realizes defect discrimination with high consistency and high generalization.

Description

Technical Field

[0001] This invention relates to the field of intelligent equipment diagnostic technology, and in particular to a method, equipment and medium for identifying defects in power distribution station equipment. Background Technology

[0002] In the field of equipment condition monitoring and defect identification in substations, with the development of multi-source sensing technology in recent years, some methods have begun to integrate images and operational data. Through feature stitching or shallow fusion strategies, preliminary multimodal analysis is achieved to improve accuracy. These methods have certain practicality under standardized operating conditions and are widely used in intelligent operation and maintenance systems.

[0003] However, conventional methods have limitations in modal fusion depth and contextual correlation modeling: First, multimodal data are often processed independently, lacking cross-modal semantic alignment mechanisms, making it difficult to effectively capture the inherent correlation between images, text logs and time-series signals; Second, the defect identification process rarely considers the topological relationship of equipment in the distribution network and the differences in inter-domain distribution, resulting in a lack of systematicness and generalization ability in risk assessment. Summary of the Invention

[0004] The purpose of this invention is to overcome the shortcomings of existing technologies and provide a method for identifying defects in substation equipment. By employing a trimodal contrastive learning framework for joint optimization, cross-modal semantic alignment of visual, textual, and temporal vectors is achieved, effectively integrating the deep semantics of heterogeneous data and improving the accuracy of preliminary defect type identification. By constructing a substation equipment topology map and using a graph neural network for neighbor state aggregation, the equipment is placed in the system-level electrical and physical association for risk transmission modeling. This allows the defect risk score to not only reflect the local state but also integrate into the global operating context, improving the reliability, generalization ability, and operation and maintenance decision support level of defect identification.

[0005] Another objective of this invention is to provide a computer device and a computer-readable storage medium for executing the method for identifying defects in power distribution equipment.

[0006] Firstly, the technical solution to achieve the above objective is: a method for identifying defects in substation equipment, comprising the following steps:

[0007] S1, real-time acquisition of data from substation equipment and preprocessing to obtain image-text-time series triplet data;

[0008] S2 extracts features from image-text-time series triplet data and uses a trimodal contrastive learning framework for joint optimization to generate multimodal joint embedding vectors, preliminary defect types, and confidence levels.

[0009] S3, perform consistency verification and correction on the multimodal joint embedding vector, preliminary defect type and confidence level to obtain the corrected defect confidence level and credibility flag;

[0010] S4, based on the corrected confidence level and credibility indicator, identifies equipment defects according to the current operating conditions and historical structured defect reports, performs multi-source cross-validation, and generates structured discrimination results;

[0011] S5. Construct a topology map of the substation equipment based on the substation equipment data, import the structured discrimination results into the substation equipment topology map, and use a graph neural network to aggregate neighbor states to obtain a defect risk score.

[0012] S6 combines the domain characteristics of the current substation to perform feature distribution alignment of the defect risk score, obtains a defect judgment list, performs maintenance on the substation equipment, generates a structured defect report, and stores it persistently.

[0013] The aforementioned method for identifying defects in substation equipment includes real-time data collection of substation equipment data such as regional weather, equipment service life, real-time load data, substation location information, current substation domain characteristics, and substation equipment ledger parameters.

[0014] The preprocessing includes data cleaning and timestamp alignment.

[0015] The aforementioned method for identifying defects in substation equipment includes the following steps: feature extraction from image-text-time series triplet data, followed by joint optimization using a three-modal contrastive learning framework to generate a multimodal joint embedding vector, preliminary defect type, and confidence level.

[0016] Feature extraction is performed on image-text-temporal triplet data to obtain visual vectors, text vectors, and temporal vectors;

[0017] A trimodal contrastive learning framework is used to optimize the cross-modal semantic consistency of visual vectors, text vectors, and temporal vectors, generating multimodal joint embedding vectors.

[0018] Defect category matching is performed based on multimodal joint embedding vectors to obtain preliminary defect types and confidence levels.

[0019] The aforementioned method for identifying defects in substation equipment includes the following steps for performing consistency verification and correction on the multimodal joint embedding vector, preliminary defect type, and confidence level to obtain the corrected defect confidence level and credibility flag:

[0020] By combining the ledger parameters of the substation equipment, the physical rationality of the multimodal joint embedding vector is initially screened, and the prior matching degree between the equipment operating status and the defect type is obtained.

[0021] The prior matching degree is used to assess the physical consistency of the initial defect confidence level, and a defect credibility indicator is obtained.

[0022] The initial defect confidence level is dynamically corrected based on the defect confidence indicator to obtain the corrected defect confidence level and confidence indicator.

[0023] The above-mentioned method for identifying defects in substation equipment includes the following steps: Based on the corrected confidence level and credibility indicator, equipment defects are identified according to the current operating conditions and historical structured defect reports, and multi-source cross-validation is performed to generate structured identification results.

[0024] The results are dynamically adjusted based on the corrected defect confidence level and credibility indicator to obtain dynamic discrimination results that adapt to the current working conditions.

[0025] The dynamic discrimination results are compared with historical structured defect reports using both semantic and numerical methods to obtain a comprehensive judgment after multi-source cross-validation.

[0026] The dynamic discrimination results and comprehensive judgments are integrated to generate structured discrimination results.

[0027] The above-mentioned method for identifying defects in substation equipment includes the following steps: constructing a substation equipment topology map based on substation equipment data, importing the structured identification results into the substation equipment topology map, and using a graph neural network to aggregate neighbor states to obtain a defect risk score.

[0028] Based on the equipment ledger data of the power distribution station, an original topology graph is constructed with equipment as nodes and electrical paths and physical proximity as edges;

[0029] The structured discrimination results are mapped to the corresponding nodes in the original topology diagram to obtain a device state diagram with defect attributes;

[0030] A graph neural network is used to perform multiple rounds of neighbor information transmission and aggregation on the device state graph with defect attributes to obtain a comprehensive defect risk score.

[0031] The aforementioned method for identifying defects in substation equipment includes a step of aligning the defect risk score with the domain characteristics of the current substation to obtain a defect list. The specific steps are as follows:

[0032] Based on the domain characteristics of the current substation, the defect risk score is subjected to unsupervised feature alignment processing to obtain a domain-invariant corrected risk score.

[0033] The corrected risk score is compared with the preset defect judgment threshold, and the judgment boundary is dynamically adjusted in combination with the equipment type and operation stage to obtain a defect judgment list.

[0034] The aforementioned method for identifying defects in substation equipment includes the following steps for inspecting the substation equipment, generating a structured defect report, and persistently storing it.

[0035] Targeted repairs were carried out based on the defect assessment list, and on-site handling records were obtained;

[0036] The on-site handling records, defect judgment list, multimodal raw data, and intermediate judgment results are integrated in a structured manner to form a structured defect report;

[0037] Structured defect reports are written to the database via a secure interface.

[0038] In a second aspect, the present invention provides a computer device including a memory and a processor, wherein the memory stores a computer program, and the computer program, when executed by the processor, implements any step of the substation equipment defect identification method as described in the first aspect of the present invention.

[0039] Thirdly, the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements any step of the substation equipment defect identification method as described in the first aspect of the present invention.

[0040] The beneficial effects of this invention are as follows: By employing a trimodal contrastive learning framework for joint optimization, cross-modal semantic alignment of visual, textual, and temporal vectors is achieved, effectively integrating the deep semantics of heterogeneous data and improving the accuracy of preliminary defect type identification; by constructing a topology map of substation equipment and using graph neural networks for neighbor state aggregation, the equipment is placed in the system-level electrical and physical association for risk transmission modeling, so that the defect risk score not only reflects the local state but also integrates into the global operating context, improving the credibility, generalization ability, and operation and maintenance decision support level of defect identification. Attached Figure Description

[0041] Figure 1 This is a flowchart of the substation equipment defect identification method of the present invention.

[0042] Figure 2 This is a flowchart for multi-source data acquisition and preprocessing.

[0043] Figure 3 This is a flowchart for defect identification.

[0044] Figure 4 This is a flowchart for generating risk scores using a graph neural network. Detailed Implementation

[0045] To enable those skilled in the art to better understand the technical solution of the present invention, its specific embodiments will be described in detail below with reference to the accompanying drawings.

[0046] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0047] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0048] Please see Figure 1 , Figure 2 , Figure 3 and Figure 4 One embodiment of the present invention provides a method for identifying defects in substation equipment, comprising the following steps:

[0049] S1. Real-time acquisition and preprocessing of data from substation equipment to obtain image-text-time series triplet data;

[0050] Real-time data collection for substation equipment includes regional weather, equipment age, real-time load data, substation location information, current substation domain characteristics, and substation equipment ledger parameters;

[0051] Furthermore, sensors and monitoring devices deployed at the substation site synchronously acquire regional weather, equipment service life, real-time load data, substation location information, current substation domain characteristics, and substation equipment ledger parameters. The current substation domain characteristics include ambient temperature and humidity, geographical location, and power supply area attributes. The substation equipment ledger parameters cover equipment model, commissioning time, and rated parameters. Infrared and visible light images are simultaneously captured by infrared thermal imagers and visible light cameras during inspections. On-site text logs are derived from operation records and anomaly descriptions entered by maintenance personnel. Time-series operation data is continuously collected by current transformers, voltage transformers, and status monitoring terminals.

[0052] Preprocessing includes data cleaning and timestamp alignment;

[0053] Furthermore, data cleaning is performed on the real-time collected data from the substation equipment. Specifically, the Laplace variance method is used to remove blurry frames from infrared and visible light images and correct formatting errors in the field text logs. Median filtering is used to filter out abrupt changes in the time-series operational data. Based on a unified time reference, the acquisition times of the infrared and visible light images, the timestamps of the field text logs, and the timestamps of the time-series operational data are aligned. Linear interpolation is used to ensure a one-to-one correspondence between different modal data in the time dimension. The cleaned and time-aligned infrared and visible light images, field text logs, and time-series operational data are then combined into image-text-time-series triplet data.

[0054] S2. Feature extraction is performed on the image-text-time series triplet data, and joint optimization is performed using a three-modal contrastive learning framework to generate multimodal joint embedding vectors, preliminary defect types, and confidence levels.

[0055] Feature extraction is performed on image-text-temporal triplet data to obtain visual vectors, text vectors, and temporal vectors;

[0056] Furthermore, a weighted average method is used to grayscale each frame of the infrared and visible light images, and a min-max normalization method is used for normalization. Scale-invariant feature transformation is used to extract each frame of the processed infrared and visible light images, and principal component analysis is used to reduce the dimensionality of each frame. The dimensionality-reduced frames are then aggregated into fixed-length visual vectors. For the on-site text logs, word segmentation and stop word removal are performed. The weighted word vectors of all words in the text are summed and normalized to form a text vector. For the time-series operational data, a sliding window segmentation is used for multi-dimensional time series such as current, voltage, and equipment status. The mean, variance, maximum, minimum, and rate of change within each window are calculated as statistical features. The statistical features of each window are concatenated in chronological order and compressed using principal component analysis to generate a time-series vector.

[0057] It should be noted that the scale-invariant feature transformation method refers to detecting potential key points through the difference of Gaussian function, locating the position, scale and orientation of each key point, extracting the gradient orientation distribution in the neighborhood around the key point, and generating each frame of image.

[0058] A trimodal contrastive learning framework is used to optimize the cross-modal semantic consistency of visual vectors, text vectors, and temporal vectors, generating multimodal joint embedding vectors.

[0059] Furthermore, the trimodal contrastive learning framework uses visual vectors, text vectors, and temporal vectors as three input modalities. Following the basic principles of contrastive learning, the training process is organized as follows: Semantic consistency between different modalities within the same image-text-temporal triplet data is used as a supervision signal. Linear projection is used to map different modalities within the image-text-temporal triplet data, obtaining embedding vectors of the same dimension. Cosine similarity is then used to measure the embedding vectors, generating a multimodal joint embedding vector, expressed as:

[0060] ;

[0061] in, For multimodal joint embedding vectors, For visual vectors, For text vectors, The magnitude of the visual vector. Indicates the length of the text vector;

[0062] It should be noted that during the data preprocessing stage, the acquisition times of infrared and visible light images, the timestamps of on-site text logs, and the timestamps of time-series running data have been aligned based on a unified time reference. Linear interpolation is used to make different modal data correspond one-to-one in the time dimension, thereby ensuring the semantic consistency of images, text descriptions, and running status from the same moment. No additional annotation is required, and they can be directly used as positive sample supervision signals in contrastive learning.

[0063] Defect category matching is performed based on multimodal joint embedding vectors to obtain preliminary defect types and confidence levels;

[0064] Furthermore, cosine similarity is calculated for each of the multimodal joint embedding vectors and the various defect prototype vectors in the pre-built defect category prototype library. The defect category prototype library is formed by clustering multimodal joint embedding vectors obtained from historical image-text-time series triplet data through the same feature extraction and three-modal contrastive learning framework, and taking the cluster center according to defect type. The defect category with the highest cosine similarity is selected as the initial defect type (e.g., joint overheating, partial discharge, and insulation degradation), and the cosine similarity value is normalized and used as the corresponding confidence level.

[0065] S3. Perform consistency verification and correction on the multimodal joint embedding vector, preliminary defect type and confidence level to obtain the corrected defect confidence level and credibility flag.

[0066] By combining the ledger parameters of the substation equipment, the physical rationality of the multimodal joint embedding vector is initially screened, and the prior matching degree between the equipment operating status and the defect type is obtained.

[0067] Furthermore, the equipment model, commissioning time, and rated parameters are extracted from the equipment ledger parameters of the substation. Based on the equipment model, the typical operating range and common defect type knowledge base of the corresponding equipment under normal operating conditions are queried. The knowledge base is formed by summarizing historical defect reports. The initial defect type is compared with the common defect types corresponding to the equipment model (e.g., the common defect types corresponding to the "KYN28-12" switchgear are {joint overheating, mechanical jamming, insulation degradation}) for set intersection judgment. If there is an intersection, it is considered physically reasonable, and the prior matching degree is set to 1. If the initial defect type does not belong to the common defect types of the equipment model (e.g., joint overheating, mechanical jamming, and insulation degradation), the matching score between the multimodal joint embedding vector and the defect category prototype vector is weighted and adjusted by the cosine similarity measurement method to obtain the prior matching degree between the equipment operating status and the defect type (the instance value range is 0-1).

[0068] The prior matching degree is used to assess the physical consistency of the initial defect confidence level, and a defect credibility indicator is obtained.

[0069] Furthermore, a weighted summation method is used to weight the prior matching degree between the equipment operating status and the defect type and the preliminary defect confidence degree to obtain a weighted confidence value for physical consistency. If the weighted confidence value for physical consistency is greater than or equal to the preliminary defect confidence degree, the defect credibility flag is set to credible; if the weighted confidence value for physical consistency is less than the preliminary defect confidence degree, the defect credibility flag is set to uncredible.

[0070] The initial defect confidence level is dynamically corrected based on the defect credibility indicator to obtain the corrected defect confidence level and credibility indicator;

[0071] Furthermore, when the defect confidence flag is credible, the weighted confidence value of physical consistency is used as the corrected defect confidence level, and the defect confidence flag remains credible; when the defect confidence flag is unreliable, the initial defect confidence level and the weighted confidence value of equipment operating status and physical consistency are attenuated by a linear attenuation function, and the attenuated result is used as the corrected defect confidence level, and the defect confidence flag remains unreliable.

[0072] S4. Based on the corrected confidence level and credibility indicator, identify equipment defects according to the current operating conditions and historical structured defect reports, and perform multi-source cross-validation to generate structured discrimination results;

[0073] The results are dynamically adjusted based on the corrected defect confidence level and credibility indicator to obtain dynamic discrimination results that adapt to the current working conditions.

[0074] Furthermore, environmental temperature and humidity, geographical location, and power supply area attributes are extracted from the current substation's domain characteristics. The environmental temperature and humidity are compared with the temperature limits in the equipment's rated parameters. If the environmental temperature and humidity exceed the historical operating range, the historical load fluctuation pattern of the area (e.g., industrial area, commercial area, and residential area) is retrieved based on the power supply area attributes. If the load change reflected by the current time-series operating data exceeds the range of the historical load fluctuation pattern (e.g., the historical operating temperature of the residential area is -10℃ to 40℃, and the current ambient temperature is 48°C), the corrected defect confidence is further weighted and adjusted. The confidence level corrected by the operating condition factors is combined with the preliminary defect type through the field binding method to obtain a dynamic discrimination result that includes the equipment's unique identifier, the preliminary defect type, the corrected defect confidence level, and the defect credibility indicator.

[0075] It should be noted that the historical operating range and regional historical load fluctuation patterns are determined through historical defect reports;

[0076] The dynamic discrimination results are compared with historical structured defect reports using both semantic and numerical methods to obtain a comprehensive judgment after multi-source cross-validation.

[0077] Furthermore, for the preliminary defect types in the dynamic discrimination results, records of the same type of defect under similar operating stages and the same equipment model are retrieved from historical structured defect reports. The text descriptions of the defect types are weighted by cosine similarity based on word frequency-inverse document frequency to complete semantic comparison. At the same time, the time-series operation data statistical features corresponding to the occurrence of the defects are extracted from the historical structured defect reports and Euclidean distance is calculated with the time-series operation data statistical features on which the current dynamic discrimination results are based to complete numerical comparison and obtain a comprehensive judgment after multi-source cross-validation.

[0078] It should be noted that historical structured defect reports are obtained by retrieving persistently stored structured defect reports from the database where historical structured defect reports are stored.

[0079] The dynamic discrimination results and comprehensive judgments are integrated to generate structured discrimination results;

[0080] Furthermore, the preliminary defect type, corrected defect confidence level, and defect credibility indicator in the dynamic discrimination results are aligned with the comprehensive judgment at the field level. The defect type, final defect confidence level, defect credibility indicator, domain characteristics of the current substation, ledger parameters of the substation equipment, corresponding image-text-time series triplet data identifiers, and cross-validation conclusions are organized into a structured discrimination result in the order of fields.

[0081] S5. Construct a topology map of the substation equipment based on the substation equipment data, import the structured discrimination results into the substation equipment topology map, and use a graph neural network to aggregate neighbor states to obtain a defect risk score.

[0082] Based on the equipment ledger data of the power distribution station, an original topology graph is constructed with equipment as nodes and electrical paths and physical proximity as edges;

[0083] Furthermore, equipment model, unique identifier, and installation location information are extracted from the equipment ledger parameters of the substation equipment, and each equipment is treated as a node in the diagram. Based on the substation primary wiring diagram and equipment connection records, if two devices are directly connected in the main electrical circuit or form an electrical path through components such as busbars, switches, and cables, this electrical path is treated as an electrical connection edge. At the same time, the physical distance between devices is calculated based on the equipment installation location information. If the physical path between two devices is less than the adjacent spacing of devices within the substation, this physical path is treated as a physical proximity edge. All nodes and their electrical connection edges are combined with physical proximity edges to form the original topology diagram.

[0084] The structured discrimination results are mapped to the corresponding nodes in the original topology diagram to obtain a device state diagram with defect attributes;

[0085] Furthermore, the unique equipment identifier, defect type, final defect confidence level, and defect credibility flag are extracted from the structured discrimination results. The corresponding node is located in the original topology diagram based on the unique equipment identifier. The defect type is encoded into a category label using one-hot encoding, and the final defect confidence level is used as a numerical attribute of the node. The defect credibility flag is converted into a Boolean state attribute using a threshold determination method, and all are attached to the node. For nodes not appearing in the structured discrimination results in the original topology diagram, a default "no defect" category label, a zero confidence level, and a credibility flag indicating "no anomaly" are assigned. After assigning values ​​to all node attributes, a device state diagram with defect attributes is formed.

[0086] A graph neural network is used to perform multi-round neighbor information transmission and aggregation on the device state diagram with defect attributes to obtain a comprehensive defect risk score.

[0087] Furthermore, the defect type encoding, final defect confidence, and defect credibility flag of each node in the device status graph with defect attributes are converted into initial node feature vectors. A graph convolutional network is used as the specific implementation of the graph neural network. In each round of aggregation, each node collects the feature vectors of its neighboring nodes and performs a weighted summation. After multiple iterations, the feature vector of each node integrates the defect information of multiple neighboring nodes. The aggregated feature vector of each node is input into a fully connected layer, and the comprehensive defect risk score corresponding to the node is output.

[0088] S6. Align the defect risk score with the feature distribution based on the current domain characteristics of the substation, obtain a defect judgment list, inspect the substation equipment, generate a structured defect report and store it persistently.

[0089] Based on the domain characteristics of the current substation, the defect risk score is subjected to unsupervised feature alignment processing to obtain a domain-invariant corrected risk score.

[0090] Furthermore, environmental temperature and humidity, geographical location, and power supply area attributes are extracted from the domain characteristics of the current substation. Based on the geographical location, the historical defect risk score distribution of similar substations is retrieved. The maximum mean difference method is used to calculate the distribution difference between the current substation defect risk score distribution and the historical defect risk score distribution. The current defect risk score is corrected by linear transformation so that the corrected score distribution is aligned with the historical score distribution of the same domain. The corrected risk score with the domain unchanged is output.

[0091] The corrected risk score is compared with the preset defect judgment threshold, and the judgment boundary is dynamically adjusted in combination with the equipment type and operation stage to obtain a defect judgment list.

[0092] Furthermore, the equipment model and commissioning time are extracted from the equipment ledger parameters of the substation equipment. The equipment type is determined based on the equipment model, and the current operating stage of the equipment is calculated based on the commissioning time. The defect judgment benchmark value corresponding to the equipment type and operating stage is queried. The benchmark value is extracted from historical defect reports, and the range of defect judgment threshold is determined by quantile statistics (exemplary range: 0.65-0.85). The corrected risk score is compared with the corresponding defect judgment threshold. If the corrected risk score is higher than the defect judgment threshold, the substation equipment is included in the defect judgment list. The resulting defect judgment list includes the equipment's unique identifier, corrected risk score, equipment type, operating stage, and judgment basis.

[0093] Targeted repairs were carried out based on the defect assessment list, and on-site handling records were obtained;

[0094] Furthermore, based on the unique equipment identifier, corrected risk score, equipment type, operating stage, and judgment criteria listed in the defect judgment list, maintenance personnel are organized to conduct on-site inspections and maintenance operations on the corresponding equipment. During the maintenance process, maintenance personnel determine the maintenance items based on the equipment type and judgment criteria, including infrared temperature measurement verification, joint tightening, insulation testing, or component replacement. After the maintenance is completed, the actual handling measures, processing results, information on replaced components, retest data, and information on the handling personnel are entered into the on-site handling record. The on-site handling record is collected through a mobile terminal or on-site industrial control terminal to ensure that each record is traceable to the corresponding defect judgment result.

[0095] The on-site handling records, defect judgment list, multimodal raw data, and intermediate judgment results are integrated in a structured manner to form a structured defect report;

[0096] Furthermore, using the unique equipment identifier as the association key, the handling measures, processing results, and retest data in the on-site handling records are aligned with the corrected risk score, equipment type, operating stage, and judgment basis in the defect judgment list. Simultaneously, the image-text-time series triplet data, multimodal joint embedding vectors, preliminary defect type and confidence level, corrected defect confidence level and credibility indicator, structured discrimination results, and comprehensive defect risk score, etc., corresponding to the equipment are associated. These intermediate discrimination results are then organized according to a unified field order and data format into a complete record containing basic equipment information, defect identification process, risk assessment basis, maintenance execution status, and verification results; generating a structured defect report.

[0097] Structured defect reports are written to the database via a secure interface;

[0098] Furthermore, symmetric encryption is used to encrypt the structured defect report at the field level, and an authentication interface based on the transport layer security protocol is used to establish a connection with the database. Before writing, the unique device identifier, timestamp, and operator information in the structured defect report are digitally signed to ensure data integrity and traceability. Each field of the structured defect report is mapped to the corresponding column in the defect management table in the database through parameterized querying to avoid injection risks. After receiving the data, the database executes a transaction commit and returns a write status code, completing the persistent storage of the structured defect report.

[0099] This embodiment also provides a computer device applicable to the method for identifying defects in power distribution equipment, comprising: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to implement the method for identifying defects in power distribution equipment as proposed in the above embodiment.

[0100] The computer device can be a terminal, comprising a processor, memory, communication interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, carrier networks, NFC (Near Field Communication), or other technologies. The display screen can be an LCD screen or an e-ink screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad on the computer device's casing, or an external keyboard, touchpad, or mouse.

[0101] This embodiment also provides a storage medium storing a computer program, which, when executed by a processor, implements the method for identifying defects in substation equipment as proposed in the above embodiments. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Red-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0102] In summary, this invention achieves cross-modal semantic alignment of visual, textual, and temporal vectors through joint optimization using a three-modal contrastive learning framework, effectively integrating the deep semantics of heterogeneous data and improving the accuracy of preliminary defect type identification. Furthermore, by constructing a topology map of substation equipment and employing graph neural networks for neighbor state aggregation, the equipment is placed within the system-level electrical and physical relationships for risk transmission modeling. This ensures that defect risk scoring not only reflects local conditions but also incorporates the global operational context, thereby improving the reliability, generalization ability, and operational decision support level of defect identification.

[0103] Those skilled in the art should recognize that the above embodiments are merely illustrative of the present invention and are not intended to limit the present invention. Any variations or modifications to the above embodiments that are within the spirit and essence of the present invention will fall within the scope of the claims of the present invention.

Claims

1. A method for identifying defects in substation equipment, characterized in that, The process includes the following steps: real-time acquisition and preprocessing of data from power distribution station equipment to obtain image-text-time series triplet data; Feature extraction is performed on image-text-time series triplet data, and joint optimization is carried out using a trimodal contrastive learning framework to generate multimodal joint embedding vectors, preliminary defect types, and confidence scores. Consistency verification and correction are performed on the multimodal joint embedding vectors, preliminary defect types, and confidence scores to obtain corrected defect confidence scores and credibility indicators. Based on the revised confidence level and credibility indicator, equipment defects are identified according to the current operating conditions and historical structured defect reports, and multi-source cross-validation is performed to generate structured discrimination results. A topology map of the substation equipment is constructed based on the substation equipment data. The structured discrimination results are imported into the substation equipment topology map and a graph neural network is used to aggregate neighbor states to obtain a defect risk score. By aligning the feature distribution of the defect risk score with the domain characteristics of the current substation, a defect judgment list is obtained, the substation equipment is inspected, a structured defect report is generated and persistently stored. The process involves feature extraction from image-text-time series triplet data, followed by joint optimization using a trimodal contrastive learning framework to generate multimodal joint embedding vectors, preliminary defect types, and confidence levels. The specific steps are as follows: Feature extraction is performed on image-text-temporal triplet data to obtain visual vectors, text vectors, and temporal vectors; A trimodal contrastive learning framework is used to optimize the cross-modal semantic consistency of visual vectors, text vectors, and temporal vectors, generating multimodal joint embedding vectors. Defect category matching is performed based on multimodal joint embedding vectors to obtain preliminary defect types and confidence levels; The process of performing consistency verification and correction on the multimodal joint embedding vector, preliminary defect type, and confidence level to obtain the corrected defect confidence level and trustworthiness flag involves the following steps: By combining the ledger parameters of the substation equipment, the physical rationality of the multimodal joint embedding vector is initially screened, and the prior matching degree between the equipment operating status and the defect type is obtained. The prior matching degree is used to assess the physical consistency of the initial defect confidence level, and a defect credibility indicator is obtained. Specifically, a weighted summation method is used to weight the prior matching degree between the equipment operating status and the defect type and the preliminary defect confidence level to obtain a weighted confidence value for physical consistency; if the weighted confidence value for physical consistency is greater than or equal to the preliminary defect confidence level, the defect credibility flag is set to credible. If the weighted confidence value of physical consistency is less than the initial defect confidence level, the defect credibility flag is set to unreliable. The initial defect confidence level is dynamically corrected based on the defect credibility indicator to obtain the corrected defect confidence level and credibility indicator. When the defect credibility indicator is credible, the physical consistency weighted confidence value is used as the corrected defect confidence level, and the defect credibility indicator remains credible. When the defect confidence flag is unreliable, the initial defect confidence level and the weighted confidence value of the equipment operating status and physical consistency are attenuated by a linear attenuation function. The attenuated result is used as the corrected defect confidence level, and the defect confidence flag remains unreliable.

2. The method for identifying defects in substation equipment as described in claim 1, characterized in that, The real-time data collected from the substation equipment includes regional weather, equipment age, real-time load data, substation location information, current substation domain characteristics, and substation equipment ledger parameters. The preprocessing includes data cleaning and timestamp alignment.

3. The method for identifying defects in substation equipment as described in claim 1, characterized in that, The process involves identifying equipment defects based on the corrected confidence level and credibility indicator, according to the current operating conditions and historical structured defect reports, and performing multi-source cross-validation to generate structured discrimination results. The specific steps are as follows: The results are dynamically adjusted based on the corrected defect confidence level and credibility indicator to obtain dynamic discrimination results that adapt to the current working conditions. The dynamic discrimination results are compared with historical structured defect reports using both semantic and numerical methods to obtain a comprehensive judgment after multi-source cross-validation. The dynamic discrimination results and comprehensive judgments are integrated to generate structured discrimination results.

4. The method for identifying defects in substation equipment as described in claim 3, characterized in that, The steps for constructing a substation equipment topology map based on substation equipment data, importing the structured discrimination results into the substation equipment topology map, and using a graph neural network to aggregate neighbor states to obtain a defect risk score are as follows: Based on the equipment ledger data of the power distribution station, an original topology graph is constructed with equipment as nodes and electrical paths and physical proximity as edges; The structured discrimination results are mapped to the corresponding nodes in the original topology diagram to obtain a device state diagram with defect attributes; A graph neural network is used to perform multiple rounds of neighbor information transmission and aggregation on the device state graph with defect attributes to obtain a comprehensive defect risk score.

5. The method for identifying defects in substation equipment as described in claim 4, characterized in that, The defect risk score is aligned with the feature distribution based on the current substation's domain characteristics to obtain a defect judgment list. The specific steps are as follows: Based on the domain characteristics of the current substation, the defect risk score is subjected to unsupervised feature alignment processing to obtain a domain-invariant corrected risk score. The corrected risk score is compared with the preset defect judgment threshold, and the judgment boundary is dynamically adjusted in combination with the equipment type and operation stage to obtain a defect judgment list.

6. The method for identifying defects in substation equipment as described in claim 5, characterized in that, The specific steps for inspecting and repairing the substation equipment, generating structured defect reports, and persistently storing them are as follows: Targeted repairs were carried out based on the defect assessment list, and on-site handling records were obtained; The on-site handling records, defect judgment list, multimodal raw data, and intermediate judgment results are integrated in a structured manner to form a structured defect report; Structured defect reports are written to the database via a secure interface.

7. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the substation equipment defect identification method according to any one of claims 1 to 6.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the substation equipment defect identification method according to any one of claims 1 to 6.

Images