A transformer fault report multi-label risk assessment method based on big data

Abstract

This invention proposes a multi-label risk assessment method for transformer fault reports based on big data, comprising: acquiring transformer fault data to be assessed; associating and aligning the transformer fault data, constructing a set of devices in the same group, and generating feature vectors; performing multi-source encoding on the feature vectors to obtain sample fusion representation vectors; constructing a label relationship graph and obtaining a label prior embedding matrix; using the label prior embedding matrix to perform label-level interactive attention decoding on text semantic features and sample fusion representation vectors to obtain the occurrence probability corresponding to the risk label; calculating the label risk score of each risk label and the comprehensive risk score of the transformer corresponding to the sample to be assessed based on the occurrence probability, trend information, group anomaly information, and information corresponding to each risk label, and outputting the multi-label risk assessment result according to a preset threshold; this invention adopts a label-level interactive attention mechanism to extract textual evidence and sample feature evidence for different risk labels respectively, enabling different labels to focus on different fault semantic fragments, effectively improving the recognition accuracy and label discrimination ability in multi-label concurrent scenarios.

Description

Technical Field

[0001] This invention relates to the fields of power equipment condition monitoring, artificial intelligence and data mining technology, and in particular to a multi-label risk assessment method for transformer fault reports based on big data. Background Technology

[0002] Transformers are critical primary equipment in power transmission and transformation systems. Their insulation systems, conductive circuits, cooling systems, core structures, and on-load tap changers gradually deteriorate under the combined effects of long-term electrical, thermal, mechanical, and environmental stresses. To ensure power supply safety, operation and maintenance units typically accumulate a large amount of data related to the transformer's operating status. This data includes structured and time-series data collected by online monitoring devices, testing and inspection devices, and dispatch automation systems, as well as natural language text such as fault reports, test conclusions, and handling records generated by maintenance, operation, and testing personnel.

[0003] In existing operation and maintenance scenarios, data related to the status of a single transformer is often scattered across multiple business systems. Due to differences in time granularity, field format, naming conventions, and sampling periods among different systems, if data cannot be uniformly aligned around the same device and the same event, it is easy to mix data from different time periods and operating conditions, leading to distorted evaluation results.

[0004] In existing technologies, one type of method mainly relies on single indicator thresholds, rule templates, or keyword matching to judge transformer risk. This type of method has the following drawbacks: First, fault report texts contain a large number of professional abbreviations, synonym substitutions, and semantic omissions, making it difficult to reliably identify fault semantic clues using only keyword matching; Second, transformers of different voltage levels, capacity specifications, cooling methods, and manufacturers exhibit natural differences in indicator distribution under normal operating conditions. Directly comparing the original monitoring values ​​in a uniform manner can easily misjudge differences in equipment specifications as abnormal risks; Third, the actual fault risk of transformers often exhibits multi-label concurrent characteristics, while traditional single-label classification methods cannot effectively characterize the dependencies between labels; Fourth, low-frequency but high-risk fault labels have fewer samples in historical data, and class imbalance can cause the model to favor high-frequency labels, thereby reducing the ability to identify rare high-risk faults.

[0005] Therefore, there is an urgent need to propose a multi-label risk assessment method and device for transformer fault reports that is oriented towards actual transformer equipment operation scenarios, can uniformly process multi-source heterogeneous big data, and simultaneously output multiple risk labels, their risk ranking, and comprehensive risk levels. Summary of the Invention

[0006] To address the problems existing in the prior art, this invention proposes a multi-label risk assessment method for transformer fault reports based on big data. The method includes: acquiring transformer fault data to be assessed; inputting the transformer fault data to be assessed into a trained multi-label risk assessment model for transformer fault reports to obtain multi-label risk assessment results.

[0007] Training the multi-label risk assessment model for transformer fault reports includes:

[0008] S1. Obtain the training sample set, which includes fault report text, online monitoring time series data, test and detection data, maintenance data, environmental operation data, risk labels, and severity information corresponding to the risk labels;

[0009] S2. Align and correlate fault report text, online monitoring time series data, test and inspection data, maintenance data, and environmental operation data. Construct a group of equipment based on equipment samples with the same or similar operating attributes as the transformer corresponding to the target sample, and generate structured feature vectors, time series feature matrices, and group deviation feature vectors. Obtain trend information based on the aligned online monitoring time series data.

[0010] S3. Perform multi-source encoding on the fault report text, structured feature vector, temporal feature matrix, and group bias feature vector to obtain the sample fusion representation vector;

[0011] S4. Construct a label relationship graph based on risk labels and historical co-occurrence information between labels, and obtain the label prior embedding matrix;

[0012] S5. Using the label prior embedding matrix, perform label-level interactive attention decoding on the text semantic features and sample fusion representation vector corresponding to the fault report text to obtain the occurrence probability of each risk label.

[0013] S6. Based on the occurrence probability, trend information, group bias feature vector and severity information corresponding to each risk label, calculate the label risk score of each risk label and the comprehensive risk score of the transformer corresponding to the sample to be evaluated, and output the multi-label risk assessment results according to the preset threshold.

[0014] S7. Construct a model loss function based on the multi-label risk assessment results, adjust the model loss function, adjust the model parameters, and complete the model training when the loss function converges.

[0015] The beneficial effects of this invention are:

[0016] 1. This invention takes equipment identification, event timestamps, and preset time windows as its core, and maps fault report text, online monitoring time series data, test and detection data, maintenance data, and environmental operation data into sample-level objects in a unified manner. This fundamentally changes the problems in the prior art where data from different systems are fragmented, time granularity is inconsistent, and it is difficult to conduct joint analysis around the same equipment event. This enables multi-source heterogeneous data to be processed and utilized in a unified manner under the same technical framework.

[0017] 2. This invention does not use fixed empirical thresholds to uniformly judge the original monitoring indicators of all transformers. Instead, it introduces a group of equipment and a robust baseline mechanism, using equipment with the same or similar operating attributes to form a comparison benchmark. It also uses the median and absolute median difference to construct a group reference center and discrete range, thereby effectively reducing the interference of equipment specification differences, manufacturing differences and local abnormal sampling points on the risk assessment results, and making the degree of abnormal deviation between similar equipment more accurately quantified.

[0018] 3. This invention not only processes fault report text in natural language form, but also processes monitoring trends, test results, maintenance records and environmental operation information. Through a multi-source feature fusion mechanism, it encodes the "abnormal phenomena described in the text", the "dynamic changes reflected in the monitoring", and the "objective state recorded in the test and maintenance" into the same feature space, thereby enhancing the overall characterization ability of complex fault mechanisms.

[0019] 4. This invention addresses the characteristics of transformer risks, which often involve the coexistence, mutual induction, and co-evolution of multiple labels. It constructs a priori label relationship graph and employs a label-level interactive attention mechanism to extract textual evidence and sample feature evidence for different risk labels. This enables different labels to focus on different fault semantic fragments, while also incorporating historical co-occurrence patterns and expert physical correlation knowledge, thereby effectively improving the recognition accuracy and label differentiation capabilities in multi-label concurrent scenarios.

[0020] 5. This invention improves the weight of low-frequency labels during training by using class imbalance weighted loss, suppresses probability outputs that significantly contradict historical patterns by using label relevance consistency constraints, and enhances the consistency between prediction results and true risk frequencies by combining probability calibration constraints, thereby improving the identification stability, output reliability, and engineering usability of high-risk labels with few samples.

[0021] 6. After outputting the probability of label occurrence, this invention further integrates trend coefficient, group anomaly coefficient and severity coefficient to generate label risk score and comprehensive risk score. This makes the model output no longer limited to simple binary classification results or abstract probability values, but can form risk label ranking, comprehensive risk level and its corresponding evidence source, providing more direct decision-making basis for operation and maintenance personnel to formulate maintenance priorities, arrange test verification order and configure operation and maintenance resources.

[0022] 7. This invention not only provides a method flow, but also provides a device structure, an electronic device implementation method, and a computer-readable storage medium implementation method, which can be applied to various deployment scenarios such as main station servers, edge intelligent terminals, or cloud-edge collaborative operation and maintenance platforms, and is convenient for implementation in existing transformer condition monitoring and intelligent operation and maintenance systems. Attached Figure Description

[0023] To more clearly illustrate the technical solutions in the embodiments of this application, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0024] Figure 1 This is a schematic diagram of the overall method flow according to an embodiment of the present invention;

[0025] Figure 2 This is a schematic diagram of the model structure for multi-source fusion, label relationship graph prior, and label-level interactive attention decoding in an embodiment of the present invention;

[0026] Figure 3 This is a schematic diagram illustrating the process of calculating label risk score, comprehensive risk score, and risk level determination in an embodiment of the present invention.

[0027] Figure 4 This is a schematic diagram of the device structure according to an embodiment of the present invention. Detailed Implementation

[0028] 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, and 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.

[0029] In this embodiment, the present invention can be deployed in a transformer condition monitoring and intelligent operation and maintenance platform. The platform may include edge acquisition nodes, an operation and maintenance master station, and a model server. The edge acquisition nodes are used to access data streams from the online dissolved gas monitoring device, the winding hot spot temperature monitoring device, the partial discharge monitoring device, and the load acquisition device; the operation and maintenance master station is used to receive fault reports, test reports, maintenance work orders, and inspection records; the model server is used to complete sample construction, model training, risk inference, and result output. The processor in the model server may be at least one of a CPU, GPU, and NPU, and the memory may be RAM, ROM, solid-state drive, disk array, or a combination thereof.

[0030] For ease of explanation, the large data sample set of transformer fault reports will be denoted as:

[0031]

[0032] in, This represents the current monitoring indicator value of the target sample. This represents the median of the set of devices in the same group on this metric. Indicates the absolute median difference. It is a smoothing constant. Indicates the first The fault report text of each sample, This indicates online monitoring time-series data. This represents the test and detection data. This indicates maintenance and repair data. This represents environmental operation data. express Each risk label corresponds to a historically confirmed multi-label vector. It should be noted that the historically confirmed multi-label vector preferably originates from historical samples that have undergone closed-loop processing, i.e., engineering confirmation conclusions formed after experimental verification, disassembly inspection, maintenance confirmation, or defect elimination, to ensure the reliability of the training labels.

[0033] A multi-label risk assessment method for transformer fault reports based on big data, such as Figures 1-4 As shown, the process includes: acquiring transformer fault data to be evaluated; inputting the transformer fault data to be evaluated into the trained transformer fault report multi-label risk assessment model to obtain the multi-label risk assessment results.

[0034] Training the multi-label risk assessment model for transformer fault reports includes:

[0035] S101. Obtain the training sample set, which includes fault report text, online monitoring time series data, test and detection data, maintenance data, environmental operation data, risk labels, and severity information corresponding to the risk labels.

[0036] S102. Align and correlate fault report text, online monitoring time series data, test and inspection data, maintenance data, and environmental operation data. Construct a group of equipment based on equipment samples with the same or similar operating attributes as the transformer corresponding to the target sample, and generate structured feature vectors, time series feature matrices, and group deviation feature vectors. Obtain trend information based on the aligned online monitoring time series data.

[0037] S103. Perform multi-source encoding on the fault report text, structured feature vector, temporal feature matrix, and group bias feature vector to obtain the sample fusion representation vector;

[0038] S104. Construct a label relationship graph based on risk labels and historical co-occurrence information between labels, and obtain the label prior embedding matrix;

[0039] S105. Using the label prior embedding matrix, perform label-level interactive attention decoding on the text semantic features and sample fusion representation vector corresponding to the fault report text to obtain the occurrence probability of each risk label.

[0040] S106. Based on the occurrence probability, trend information, group bias feature vector and severity information corresponding to each risk label, calculate the label risk score of each risk label and the comprehensive risk score of the transformer corresponding to the sample to be evaluated, and output the multi-label risk assessment results according to the preset threshold.

[0041] S107. Construct a model loss function based on the multi-label risk assessment results, adjust the model loss function, adjust the model parameters, and complete the model training when the loss function converges.

[0042] In this embodiment, the trend information is obtained from the previously acquired and aligned online monitoring time-series data. Specifically, key monitoring indicators, such as dissolved gas in oil, partial discharge, top oil temperature, winding hot spot temperature, load rate, and insulation resistance, are extracted from the most recent preset time window. The slope of these indicators along the abnormal direction is calculated, and then they are aggregated according to the importance weight of the indicators to obtain the trend coefficient. It reflects whether the transformer's operating condition is continuously deteriorating.

[0043] Group anomaly information is obtained from the generated group deviation feature vector. A robust benchmark has already been established using equipment of the same voltage level, capacity range, and cooling method, and the deviation value of the target transformer relative to this group of equipment has been calculated. In S6, these deviation values ​​are truncated, normalized, and weighted to obtain the group anomaly coefficient. This coefficient reflects whether the target transformer is significantly abnormal compared to similar equipment.

[0044] Severity information consists of pre-configured parameters for each risk label, determined based on transformer operation and maintenance procedures, historical fault handling records, and expert ratings, and stored in a risk parameter table. For example, risks such as winding insulation aging, bushing discharge, and on-load tap changer malfunction have more severe consequences and therefore higher severity coefficients. It reflects the degree of impact on equipment safety, power outage losses, and the urgency of maintenance should such risks occur.

[0045] In this embodiment, the method may include the following steps:

[0046] S1. Obtain a large sample of transformer fault reports and perform unified alignment.

[0047] S11. Obtain raw data from multiple sources.

[0048] The processor retrieves raw data from the online monitoring system, test management system, equipment management system, and environmental database via a data interface, and writes the raw data into a sample buffer. The raw data includes at least the following:

[0049] (1) Fault report text, such as defect phenomenon description, fault analysis opinions, test diagnosis conclusions, handling suggestions and retest records;

[0050] (2) Online monitoring of time-series data, such as the content of hydrogen, methane, acetylene, total hydrocarbons, carbon monoxide and carbon dioxide in oil, top oil temperature, winding hot spot temperature, load rate, partial discharge amplitude, vibration value and water content;

[0051] (3) Test data, such as dielectric loss factor, insulation resistance, DC resistance, turns ratio test results, on-load tap changer operating characteristics and winding frequency response test results;

[0052] (4) Maintenance data, such as power outage records, component replacement records, tap changer maintenance records, inspection records and defect work orders;

[0053] (5) Environmental operation data, such as ambient temperature, humidity, pollution level, altitude, salt spray level and thunderstorm day information.

[0054] S12. Create an event-level time window.

[0055] The processor uses the fault report generation time or alarm trigger time as the event anchor point, centered around the event moment. Establish a preset time window:

[0056]

[0057] in, Indicates the duration of the backtracking before the event. Indicates the duration of the supplementary time after the event. Preferably, Available for 7, 15, or 30 days. 0 days, 1 day, or 3 days can be selected to adapt to different sampling cycles and operation and maintenance systems.

[0058] S13, Perform multi-source data association and alignment.

[0059] Based on the device's unique identifier, ledger identifier, event timestamp, and the preset time window, the processor uniformly associates fault report text, online monitoring time series data, test and detection data, maintenance data, and environmental operation data, so that various types of data form a single sample object to be evaluated around the same transformer, the same risk event, and the same time range.

[0060] S14. Complete data preprocessing.

[0061] For online monitoring of high-frequency sequences, the processor can further perform resampling, denoising, missing data imputation, and abnormal timestamp correction to obtain a time-series feature matrix:

[0062]

[0063] in, Indicates the number of sampling points within the time window. This indicates the number of monitoring indicators. For test data, maintenance data, and environmental operation data, the processor performs field standardization, category mapping, and numerical normalization to form a structured feature vector.

[0064] Through steps S11 to S14, the original multi-source data scattered across different business systems can be uniformly mapped into sample-level objects oriented towards a single device and a single event, laying the data foundation for subsequent modeling.

[0065] S2. Construct a set of devices in the same group and generate group bias characteristics.

[0066] S21. Filter the set of devices in the same group.

[0067] Because transformers of different voltage levels, capacities, cooling methods, manufacturers, and operating areas exhibit natural distribution differences under normal operating conditions, this invention does not directly use fixed thresholds to evaluate individual monitoring values. Instead, it constructs a set of similar devices around the target sample. The processor selects samples from the historical sample library that are the same as or similar to the target transformer in terms of voltage level, capacity range, cooling method, manufacturer, operating area, and years of operation.

[0068] S22. Establish a robust baseline.

[0069] After obtaining the same group of devices Then, the processor retrieves the first from the set. The historical value sequences of each monitoring indicator are used as the robust central value, and the absolute median difference is used as the robust dispersion to construct the robust baseline of the target monitoring indicator.

[0070] S23. Calculate the population bias value.

[0071] For the The first sample For each monitoring indicator, the processor calculates the robustness deviation value according to the following formula:

[0072]

[0073] S24. Form a group bias feature vector.

[0074] The processor concatenates the robust deviation values ​​corresponding to all monitoring indicators in a preset order to form a group deviation feature vector. It is used to characterize the degree of abnormal deviation of the target transformer relative to a group of similar equipment.

[0075] Through steps S21 to S24, the present invention converts the original monitoring values, which are originally difficult to compare directly, into a standardized deviation representation relative to the same set of devices, thereby significantly enhancing the comparability between different devices.

[0076] S3. Perform multi-source encoding on fault report text, structured features, and temporal features.

[0077] S31. Perform fault report text standardization processing.

[0078] The processor first calls a terminology normalization dictionary to standardize the fault report text, mapping synonyms, technical abbreviations, and common abbreviations to a unified expression. For example, "on-load tap changer," "voltage regulator," and "OLTC" are mapped to the same term, and "light gas action" and "light gas alarm" are mapped to the same event category.

[0079] S32. Perform text semantic encoding.

[0080] After normalization, the fault report text is represented as a sequence of terms. And input it into the text semantic encoder to obtain the text semantic feature matrix:

[0081]

[0082] in, Indicates the first In the nth sample The contextual semantic representation of each word element.

[0083] S33. Perform numerical feature encoding.

[0084] The processor concatenates the structured feature vector with the population bias feature vector and converts it into a numerical representation vector through linear mapping, multilayer perceptron mapping, or normalization mapping. The structured feature vector may include equipment ledger information, quantitative values ​​of test results, number of maintenance operations, number of component replacements, and environmental indicator statistics.

[0085] S34. Perform temporal feature encoding.

[0086] The processor will use the timing feature matrix Inputting the data into a time-series coding network extracts the changing trends and dynamic evolution patterns of key monitoring indicators, resulting in a trend representation vector. The temporal coding network may be a one-dimensional convolutional network, a gated recurrent network, a temporal Transformer network, or a combination thereof.

[0087] S35. Perform multi-source feature fusion.

[0088] After encoding text, numerical, and time-series data, the processor will convert the text semantic feature matrix... Global text representation and numerical representation vectors obtained through pooling processing Trend representation vector The samples are concatenated and mapped to obtain the fused representation vector:

[0089]

[0090] in, Can be adopted Vector, average pooling vector, or attention pooling vector.

[0091] Through steps S31 to S35, the present invention maps fault report text, structured status information and dynamic time sequence information into the same feature space, providing a unified sample expression basis for subsequent risk label decoding.

[0092] S4. Construct a prior set of risk labels, label severity coefficients, and label relationship graphs.

[0093] S41. Set the risk label set and severity coefficient.

[0094] In this embodiment, the risk label set can be set based on the enterprise's operation and maintenance procedures, historical defect database, and expert experience. As a preferred embodiment, the following risk labels can be set: winding insulation aging risk, on-load tap changer abnormality risk, bushing and partial discharge risk, core multi-point grounding risk, local overheating risk, moisture deterioration risk, cooling system abnormality risk, protection and secondary circuit abnormality risk, residual risk after external short circuit impact, and accessory and seal failure risk.

[0095] Severity coefficients for each risk label Risk parameters can be preset based on the severity of the accident consequences, power outage losses, urgency of maintenance, and expert ratings, and stored in the risk parameter table.

[0096] S42, Encode label description text.

[0097] The processor reads the label description text of each risk label and inputs it into the label text encoder to obtain the initial label embedding. .

[0098] S43. Construct a label relationship matrix.

[0099] The processor constructs a label relationship matrix based on the pairwise co-occurrence frequency of labels in historical samples and prior expert knowledge, in the following form:

[0100]

[0101] in, Indicates label With tags The number of times they co-occur. and Each represents the number of samples. This indicates the expert's pre-defined correlation strength.

[0102] S44, Update the prior embedding of the tag.

[0103] After obtaining the label relation matrix, the processor adds self-connects and normalizes it, then performs graph convolutional propagation to update the initial label embeddings, resulting in the label prior embedding matrix:

[0104]

[0105] in, .

[0106] Through steps S41 to S44, the present invention can uniformly encode the tag semantic information, historical statistical relationships and expert physical priors into the tag representation, making the subsequent decoding process more consistent with the actual association rules of transformer fault tags.

[0107] S5. Utilize a tag-level interactive attention mechanism to output the probability of occurrence of each risk tag.

[0108] S51. Construct label-level attention weights.

[0109] In this step, the processor extracts relevant semantic fragments from the text semantic feature matrix for each risk label. Specifically, this is done using label prior embedding. As a query vector, it is represented by lexical units in the text semantic feature matrix. Using these as keys and values, calculate the attention weight of each term for the corresponding risk tag:

[0110]

[0111] S52. Generate tag-specific context vectors.

[0112] The processor uses the aforementioned attention weights to perform a weighted summation of the representations of each lexical unit, resulting in a tag-specific context vector:

[0113]

[0114] S53, Output the probability of occurrence of each risk label.

[0115] The processor will use the tag-specific context vector Sample fusion representation vector and corresponding tag embedding Perform concatenation mapping and then transform using an activation function to obtain the occurrence probability of each risk label:

[0116]

[0117] Through steps S51 to S53, different risk labels can focus on different semantic evidence in the fault report, and at the same time, comprehensively call the multi-source sample fusion features and the prior semantics of the labels to achieve refined decoding in multi-label concurrent scenarios.

[0118] S6. Construct the training loss and optimize the model.

[0119] S61. Construct class-imbalanced weighted loss.

[0120] To mitigate the problem of low-frequency high-risk labels being overwhelmed by high-frequency labels, the processor constructs category weights based on the number of valid samples for each risk label:

[0121]

[0122] Based on this, construct a multi-label classification loss:

[0123]

[0124] S62. Construct label-related consistency constraints.

[0125] To ensure that the prediction results are reasonably consistent with the adjacency structure in the label relationship graph, the processor constructs consistency constraints based on the normalized label relationship matrix:

[0126]

[0127] S63. Construct probabilistic calibration constraints and form the total loss.

[0128] The processor further introduces probabilistic calibration loss. This is then combined with the multi-label classification loss and consistency constraint to form the total loss function:

[0129]

[0130] S64. Complete parameter training.

[0131] The processor iteratively updates the parameters of the text semantic encoder, temporal coding network, numerical mapping layer, label graph construction module and label-level interactive attention decoding module based on the total loss function until the preset convergence condition is met, thereby obtaining the trained multi-label risk assessment model for transformer fault reports.

[0132] S7. Calculate the trend coefficient, the population anomaly coefficient, the label risk score, and the comprehensive risk score.

[0133] S71. Calculate the trend coefficient.

[0134] The processor reads the time series of key monitoring indicators within the most recent preset time window, calculates the slope of change along a preset anomaly direction, and then aggregates them according to the importance weight of the indicators to obtain the trend coefficient:

[0135]

[0136] S72. Calculate the population anomaly coefficient.

[0137] The processor truncates and normalizes the deviation values ​​in the population deviation feature vector, and aggregates them according to importance weights to obtain the population anomaly coefficient:

[0138]

[0139] S73. Calculate the label risk score.

[0140] The processor is for the first Each risk label is weighted and fused with its corresponding probability of occurrence, severity coefficient, trend coefficient, and population anomaly coefficient to obtain a label risk score:

[0141]

[0142] S74. Calculate the comprehensive risk score.

[0143] The processor performs non-linear aggregation based on the risk scores of multiple tags to obtain a comprehensive risk score:

[0144]

[0145] Through steps S71 to S74, the present invention unifies the probability of risk occurrence, the degree of trend deterioration, and the severity of serious consequences into the same risk scoring system, making the output results more consistent with the handling logic in equipment operation and maintenance.

[0146] S8. Perform label threshold calibration and output the evaluation results.

[0147] S81, Perform label-level threshold calibration.

[0148] Since the distribution and business costs of different risk labels vary, this invention preferably sets thresholds for each label separately. The processor traverses the candidate threshold set on the validation set and selects the label-level threshold that optimizes the target evaluation metric.

[0149]

[0150] S82, Output the final set of risk labels.

[0151] During the model application phase, the processor compares the occurrence probability of each label of the sample to be evaluated with its label-level threshold one by one, and outputs the final risk label set and risk ranking results.

[0152] S83. Output the overall risk level and explanatory information.

[0153] The processor further calculates based on the comprehensive risk score. The mapping outputs a comprehensive risk level, and can combine the attention weight hot zone, trend contribution value and group anomaly contribution value corresponding to each risk label to generate explanatory information to help operation and maintenance personnel understand the cause of risk triggering and formulate handling strategies.

[0154] Corresponding to the above method, the present invention also provides a multi-label risk assessment device for transformer fault reports based on big data. The device may include: a data acquisition module, a data alignment and baseline construction module, a multi-source feature encoding module, a label graph prior construction module, a label-level interactive attention decoding module, a risk score output module, and a training and optimization module. Each module can be implemented in software or in a combination of software and hardware.

[0155] Furthermore, the present invention can also be implemented by an electronic device. The electronic device may include a processor, a memory, a network interface, and a display interface. The memory stores a computer program for implementing the above method, and when the processor executes the computer program, it can complete multi-source data acquisition, sample construction, model training, risk inference, and result output.

[0156] Those skilled in the art will understand that the above method steps can be implemented in whole or in part by software, hardware, or a combination of software and hardware. The corresponding software program can be stored in a non-transitory computer-readable storage medium, including but not limited to ROM, RAM, disk, optical disk, flash memory, solid-state drive, or other media capable of storing program code. When the program is executed by a processor, it can complete the method steps in the above embodiments.

[0157] The above-described embodiments further illustrate the purpose, technical solution, and advantages of the present invention. It should be understood that the above-described embodiments are merely preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made to the present invention within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A multi-label risk assessment method for transformer fault reports based on big data, characterized in that, include: Obtain fault data for the transformer to be evaluated; The transformer fault data to be evaluated is input into the trained transformer fault report multi-label risk assessment model to obtain the multi-label risk assessment results. Training the multi-label risk assessment model for transformer fault reports includes: S1. Obtain the training sample set, which includes fault report text, online monitoring time series data, test and detection data, maintenance data, environmental operation data, risk labels, and severity information corresponding to the risk labels; S2. Align and correlate fault report text, online monitoring time series data, test and inspection data, maintenance data, and environmental operation data. Construct a group of equipment based on equipment samples with the same or similar operating attributes as the transformer corresponding to the target sample, and generate structured feature vectors, time series feature matrices, and group deviation feature vectors. Obtain trend information based on the aligned online monitoring time series data. S3. Perform multi-source encoding on the fault report text, structured feature vector, temporal feature matrix, and group bias feature vector to obtain the sample fusion representation vector; S4. Construct a label relationship graph based on risk labels and historical co-occurrence information between labels, and obtain the label prior embedding matrix; S5. Using the label prior embedding matrix, perform label-level interactive attention decoding on the text semantic features and sample fusion representation vector corresponding to the fault report text to obtain the occurrence probability of each risk label. S6. Based on the occurrence probability, trend information, group bias feature vector and severity information corresponding to each risk label, calculate the label risk score of each risk label and the comprehensive risk score of the transformer corresponding to the sample to be evaluated, and output the multi-label risk assessment results according to the preset threshold. S7. Construct a model loss function based on the multi-label risk assessment results, adjust the model loss function, adjust the model parameters, and complete the model training when the loss function converges.

2. The method for multi-label risk assessment of transformer fault reports based on big data as described in claim 1, characterized in that, Constructing a set of devices within the same group and generating a group deviation feature vector includes: grouping samples according to voltage level, capacity range, cooling method, manufacturer, operating region, and years of operation to obtain a set of devices within the same group corresponding to the target sample; for the first... The first sample The robust deviation value is calculated for each monitoring indicator; the robust deviation values ​​corresponding to each monitoring indicator are spliced ​​together in a preset order to form the group deviation feature vector.

3. The method for multi-label risk assessment of transformer fault reports based on big data as described in claim 1, characterized in that, Multi-source encoding of fault report text, structured feature vectors, temporal feature matrix, and group bias feature vectors includes: term normalization, word segmentation, and lexicalization of the fault report text, inputting it into a text semantic encoder, and outputting a text semantic feature matrix; concatenating the structured feature vectors and group bias feature vectors and performing numerical mapping to obtain a numerical representation vector. The temporal feature matrix is ​​input into a temporal coding network to extract the trend representation vector. The global text representation and numerical representation vectors are obtained by pooling the text semantic feature matrix. Trend representation vector Perform concatenation mapping to obtain the sample fusion representation vector.

4. The method for multi-label risk assessment of transformer fault reports based on big data as described in claim 1, characterized in that, Constructing the tag relationship graph includes: encoding the tag description text of each risk tag to obtain the initial tag embedding. A label relationship matrix is ​​constructed based on the co-occurrence frequency of labels in historical samples, the occurrence frequency of each label, and the label association strength preset by experts; a label relationship graph is constructed based on the label relationship matrix.

5. The method for multi-label risk assessment of transformer fault reports based on big data as described in claim 1, characterized in that, Constructing the label relation matrix includes: adding self-connects to the label relation matrix, performing degree normalization, and updating the initial label embeddings through at least one layer of graph convolution propagation to obtain the label prior embedding matrix. The update process satisfies the following: ； ； in, It is the identity matrix. for The degree matrix, For the first Layer graph convolution parameters, It is a non-linear activation function.

6. The method for multi-label risk assessment of transformer fault reports based on big data as described in claim 1, characterized in that, Performing tag-level interactive attention decoding includes: embedding each tag in the tag prior embedding matrix. As a query vector, the lexical representation in the text semantic feature matrix As keys and values, the attention weights of the corresponding risk labels to each word are calculated; the word representations are weighted and summed according to the attention weights to obtain the label-specific context vectors corresponding to each risk label; the label-specific context vectors are concatenated and mapped with the sample fusion representation vectors and the corresponding label embeddings, and the occurrence probabilities corresponding to each risk label are obtained by activation function transformation.

7. The method for multi-label risk assessment of transformer fault reports based on big data as described in claim 6, characterized in that, The probability of occurrence for each risk label is as follows: ； ； in, Represents the value projection matrix, and They represent the first The classification parameters corresponding to each risk label This is the tag-specific context vector corresponding to the risk tag. Pay attention to the weights.

8. The method for multi-label risk assessment of transformer fault reports based on big data as described in claim 1, characterized in that, The model's loss function is: ； in, Indicates the first The number of positive samples for each risk label. For category weighting parameters, This is the focusing factor.

9. The method for multi-label risk assessment of transformer fault reports based on big data as described in claim 1, characterized in that, Before calculating the label risk score and the overall risk score for each risk label, the process also includes: aggregating the changes of each key monitoring indicator along a preset abnormal direction within the most recent preset time window according to the importance weight of each key monitoring indicator to obtain a trend coefficient. The deviation values ​​of each indicator in the group deviation feature vector are truncated and normalized, and then aggregated according to the aforementioned importance weights to obtain the group anomaly coefficient. .

10. The method for multi-label risk assessment of transformer fault reports based on big data as described in claim 1, characterized in that, The calculation of the label risk score for each risk label and the overall risk score includes: for the first... Each risk label is assigned a risk score. The probability of occurrence corresponding to the risk label is weighted and fused with the severity coefficient, trend coefficient, and group anomaly coefficient corresponding to the risk label to obtain the label risk score. Multiple label risk scores are nonlinearly aggregated to obtain a comprehensive risk score. The comprehensive risk score is mapped according to a preset level range to output the comprehensive risk level.

Images