Intelligent diagnosis method for surge arresters based on leakage current characteristics and knowledge enhancement model

Abstract

The application discloses a lightning arrester intelligent diagnosis method based on leakage current characteristics and a knowledge enhancement model, comprising the following steps: firstly, full current, bus voltage, action frequency, environmental parameters and historical operation data of a metal oxide lightning arrester are synchronously collected, and phase reference data required for resistive current extraction is obtained; then, the leakage current signal is preprocessed, and a fast Fourier transform spectrum analysis method is used to extract the resistive current component; then, a multi-dimensional state feature vector is constructed by using the resistive current, harmonics, phase, trend and environmental correction information; subsequently, the multi-dimensional state feature vector is input into a knowledge enhancement Transformer diagnosis model to obtain lightning arrester state recognition results and degradation trend prediction results; finally, fusion reasoning is carried out in combination with a lightning arrester domain knowledge base to output a state level, an abnormal reason, a risk level and operation and maintenance suggestions. The application does not depend on image acquisition, and improves diagnosis accuracy and interpretability.

Description

Technical Field

[0001] This invention relates to an intelligent diagnostic method for surge arresters based on leakage current characteristics and a knowledge enhancement model, belonging to the fields of leakage current signal analysis and artificial intelligence diagnostic technology. Background Technology

[0002] Currently, metal oxide surge arresters are widely used for overvoltage protection in power transmission and transformation systems, and their operating status directly affects the safety of power equipment. During long-term operation, surge arresters are susceptible to factors such as operating voltage, ambient humidity, surface contamination, and aging of valve plates, which can lead to increased resistive current, decreased insulation performance, or deterioration. Therefore, online monitoring and condition diagnosis of their leakage current are necessary.

[0003] Existing patents and systems related to online monitoring of surge arresters mostly acquire the total current signal through a leakage current acquisition unit, and calculate parameters such as resistive current, harmonic components, and number of trips by combining them with bus voltage or phase reference signals. The monitoring results are then uploaded to a backend platform for display and alarm. Some methods decompose the capacitive and resistive components in the total current, or use the fundamental wave method, third harmonic method, or FFT spectrum analysis method to extract resistive current characteristics, and then determine whether the surge arrester is abnormal based on a preset threshold.

[0004] However, existing methods still primarily rely on single characteristics or fixed thresholds for judgment, making it difficult to fully utilize multi-source information such as total current, resistive current, harmonics, phase, trends, historical alarms, and environmental parameters. The leakage current characteristics of surge arresters vary under different voltage levels, operating years, and environmental conditions, making fixed thresholds prone to false alarms or missed alarms. Furthermore, existing methods lack sufficient explanation for long-term leakage current trends and anomalies, making it difficult to accurately distinguish between different degradation modes such as moisture, valve plate aging, and surface contamination, and also hindering the development of specific operation and maintenance recommendations.

[0005] Therefore, there is an urgent need to propose an intelligent diagnostic method based on the multidimensional characteristics of surge arrester leakage current. Through the construction of multidimensional state features, time-series trend modeling, and knowledge-enhanced diagnosis, this method can realize surge arrester state identification, degradation trend prediction, anomaly cause explanation, and operation and maintenance suggestions, thereby improving the utilization rate of online monitoring data and the reliability of diagnosis. Summary of the Invention

[0006] To address the shortcomings of existing online surge arrester monitoring methods, such as insufficient utilization of monitoring data, weak ability to identify abnormal patterns, inadequate prediction of degradation trends, and poor interpretability of diagnostic results, this invention provides an intelligent surge arrester diagnostic method based on leakage current characteristics and a knowledge enhancement model. This method can achieve multi-dimensional state feature extraction, intelligent state identification, degradation trend prediction, and interpretable operation and maintenance decision output for surge arresters.

[0007] The intelligent diagnostic method for surge arresters based on leakage current characteristics and knowledge enhancement models includes the following steps: Step S1: Through the online monitoring terminal and leakage current digital acquisition device of the surge arrester, the total current, bus voltage, number of operations, historical operating data and environmental parameters of the metal oxide surge arrester are collected simultaneously, and the resistive current measurement value or the phase reference data required for resistive current extraction is obtained. Step S2: The leakage current signal acquired in step S1 is subjected to outlier removal, filtering, normalization and power frequency interference suppression. Based on the phase relationship between the bus voltage and the leakage current, the resistive current component is extracted using the fast Fourier transform spectrum analysis method. Step S3: Based on the resistive current component extracted in step S2 and the multi-source data collected in step S1, construct a multi-dimensional state feature vector that includes current amplitude characteristics, resistive current ratio, harmonic characteristics, phase characteristics, trend characteristics, historical alarm characteristics and environmental correction factors, and dynamically correct the state judgment threshold according to the surge arrester voltage level, historical operating baseline, environmental parameters and surface contamination status. Step S4: Input the multi-dimensional state feature vector constructed in step S3 into the knowledge-enhanced Transformer diagnostic model, integrate the pre-constructed surge arrester domain knowledge information, obtain the state recognition result, model confidence, model uncertainty and degradation trend prediction result, and calculate the comprehensive risk score by combining the corrected state judgment threshold to determine the surge arrester state level. Step S5: Integrate the output of Step S4 with the surge arrester domain knowledge base to output the cause of the anomaly, key impact characteristics, risk level and operation and maintenance suggestions, and review the diagnostic results that have model-rule conflicts or insufficient model reliability.

[0008] Furthermore, in step S1, the original monitoring data is obtained from the online monitoring terminal of the surge arrester, the digital acquisition device of leakage current, the digital acquisition result of the pointer-type leakage current meter, or the historical database of the existing monitoring system; the historical operating data includes historical detection data, historical alarm records, and maintenance records.

[0009] Furthermore, in step S2, the outlier removal adopts the Laida criterion, the filtering adopts a Butterworth low-pass filter, the power frequency interference suppression adopts an adaptive notch filter, and the normalization adopts the Z-score standardization method.

[0010] Further, in step S3, the multidimensional state feature vector includes the total current amplitude feature, fundamental resistive current, resistive current ratio, third harmonic content, total harmonic distortion rate, fundamental phase difference, resistive current trend change rate, historical alarm features, and environmental correction factor; the dynamic correction of the state judgment threshold is based on voltage level, operating voltage fluctuation, ambient temperature, ambient humidity, and surface contamination status, and is performed by multiplying the correction coefficients corresponding to each influencing factor.

[0011] Further, in step S4, the knowledge-enhanced Transformer diagnostic model includes a feature embedding layer, a temporal encoding layer, a knowledge enhancement layer, and an output layer; wherein, the temporal encoding layer uses a multi-head self-attention mechanism to extract the temporal correlation of state features within multiple consecutive sampling periods; the knowledge enhancement layer converts lightning arrester domain knowledge information into rule activation vectors, case similarity vectors, and knowledge graph embedding vectors, and performs gating fusion with temporal features.

[0012] Furthermore, in step S4, the comprehensive risk score comprehensively considers the degree of abnormal deviation of multidimensional state features relative to the corrected threshold, the degree of insufficient model confidence, model uncertainty, and the prediction results of degradation trend, and determines the normal, attention, abnormal, or severely degraded state level according to the interval of the comprehensive risk score; when the model output is inconsistent with the level corresponding to the comprehensive risk score or the model reliability is insufficient, a review mark is generated.

[0013] Furthermore, in step S5, the surge arrester domain knowledge base includes a rule base, a fault case base, a maintenance record base, and a knowledge graph; the knowledge base is verified and updated using historical fault cases and maintenance results to obtain an updated knowledge base; the output result of step S4 is input into the updated knowledge base for fusion reasoning to obtain a knowledge fusion diagnostic result.

[0014] Furthermore, in step S5, when there is a conflict between the model and the rules or the model's reliability is insufficient, the diagnostic task is sent to the manual review module, where professionals review the data by combining the original waveform, spectrum analysis, multidimensional features, and model output. If the review fails, the review conclusion is used as the final diagnostic result and marked as an error correction sample for use in knowledge base updates or model retraining.

[0015] Further, in step S5, the abnormal causes and key impact features are determined by calculating the normalized abnormal deviation and contribution of each state feature, and matching them with rules and fault cases in the knowledge base; the operation and maintenance suggestions are generated based on the final state diagnosis results: normal state output normal inspection, attention state output shorten inspection cycle, abnormal state output power outage detection, and severely degraded state output shutdown inspection or replacement.

[0016] Further, in step S5, the verification and updating process of the surge arrester domain knowledge base includes: inputting the state characteristics in historical operating data into the knowledge base for reasoning, and comparing the reasoned fault causes, risk levels, and maintenance measures with actual fault records and repair results; when the reasoning results are consistent with the actual records, retaining the corresponding knowledge entities and relationships; when there are deviations, generating candidate correction information for corresponding rule weights, knowledge entity relationships, and state classification rules, and using the candidate correction information, after being reviewed by professionals and verified using historical fault cases, to update the surge arrester domain knowledge base.

[0017] Compared with the prior art, the beneficial effects of the present invention are: (1) This invention performs intelligent diagnosis based on the leakage current time series data of surge arresters, and integrates multi-dimensional features such as total current, resistive current, harmonic content, phase difference and trend change rate, overcoming the problem that traditional methods rely only on a single threshold for judgment and have poor adaptability.

[0018] (2) This invention combines leakage current characteristic analysis with intelligent model, which can identify different abnormal modes such as moisture and valve plate aging, and predict the deterioration trend of surge arrester, thereby improving diagnostic accuracy and early warning capability.

[0019] (3) This invention introduces a knowledge base in the field of surge arresters to interpret the model output results, and can provide the cause of the anomaly, key impact features and operation and maintenance suggestions, thereby improving the interpretability of the diagnostic results and the engineering application value. Attached Figure Description

[0020] Figure 1 This is an overall flowchart of the intelligent diagnosis method for surge arresters based on leakage current characteristics and knowledge enhancement model proposed in this invention; Figure 2 This is a flowchart of the leakage current signal preprocessing and resistive current extraction in this invention; Figure 3 This is a flowchart of the multidimensional state feature construction and knowledge enhancement Transformer diagnostic model identification in this invention; Figure 4 This is a flowchart of the knowledge enhancement interpretation and operation and maintenance decision output in this invention; Figure 5 This is a schematic diagram of the intelligent diagnostic simulation platform for surge arresters and the visualization interface for diagnostic results in this invention. Detailed Implementation

[0021] The technical solution of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are merely 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.

[0022] The present invention adopts the following technical solution, combined with Figure 1 To complete the detection and evaluation, the present invention proposes an intelligent diagnostic method for surge arresters based on leakage current characteristics and a knowledge enhancement model, which includes the following steps: Step S1: Through the online monitoring terminal and leakage current digital acquisition device of the surge arrester, the total current, bus voltage, number of operations, historical operating data and environmental parameters of the metal oxide surge arrester are collected simultaneously, and the resistive current measurement value or the phase reference data required for resistive current extraction is obtained. Step S2: The leakage current signal acquired in step S1 is subjected to outlier removal, filtering, normalization and power frequency interference suppression. Based on the phase relationship between the bus voltage and the leakage current, the resistive current component is extracted using the Fast Fourier Transform (FFT) spectrum analysis method. Step S3: Based on the resistive current component extracted in step S2 and the multi-source data collected in step S1, construct a multi-dimensional state feature vector that includes current amplitude characteristics, resistive current ratio, harmonic characteristics, phase characteristics, trend characteristics, historical alarm characteristics and environmental correction factors, and dynamically correct the state judgment threshold according to the surge arrester voltage level, historical operating baseline, environmental parameters and surface contamination status. Step S4: Input the multi-dimensional state feature vector constructed in step S3 into the knowledge-enhanced Transformer diagnostic model, integrate the pre-constructed surge arrester domain knowledge information, obtain the state recognition result, model confidence, model uncertainty and degradation trend prediction result, and calculate the comprehensive risk score by combining the corrected state judgment threshold to determine the surge arrester state level. Step S5: Integrate the output of Step S4 with the surge arrester domain knowledge base to output the cause of the anomaly, key impact characteristics, risk level and operation and maintenance suggestions, and review the diagnostic results that have model-rule conflicts or insufficient model reliability.

[0023] Preferably, step S1 includes: The raw monitoring data is obtained from the online monitoring terminal of the surge arrester, the digital acquisition device for leakage current, the digital acquisition results of the pointer-type leakage current meter, or the historical database of an existing monitoring system; the raw monitoring data includes total current, bus voltage, number of operations, historical operating data, and environmental parameters, and the historical operating data includes historical detection data, historical alarm records, and maintenance records; wherein, the total current is denoted as... The bus voltage is denoted as The number of actions is recorded as The historical operating data is recorded as The environmental parameters include ambient temperature. Ambient humidity and surface dirt .

[0024] Combination Figure 2 Preferably, step S2 includes: Step S2-1: Process the full current signal acquired in step S1 Signal preprocessing is performed to obtain the preprocessed full current signal. The signal preprocessing includes outlier removal, low-pass filtering, power frequency interference suppression, and normalization. Outlier removal uses the Raida criterion, identifying and removing sampling points that satisfy the following formula:

[0025] in, Indicates the first One full current sample value, This represents the average current within the sampling window. This represents the standard deviation of the total current within the sampling window.

[0026] After outlier removal, the full current signal is filtered by a Butterworth low-pass filter to remove high-frequency noise, an adaptive notch filter to suppress power frequency interference, and normalized using the Z-score normalization method. The normalization formula is as follows:

[0027] in, Represents the normalized i-th One full current sample value.

[0028] Step S2-2: Based on the full current signal after outlier removal, filtering, and power frequency interference suppression. The bus voltage is denoted as The phase relationship was determined, and the resistive current component was extracted using Fast Fourier Transform (FFT) spectral analysis. The normalization process was used for scale unification before inputting the multidimensional state features into the model, and was not used to change the physical quantity values ​​of the resistive current component. The preprocessed full current signal was then analyzed. The bus voltage is denoted as Perform FFT spectral analysis to obtain the first RMS value of total current under subharmonics Identify the phase of the bus voltage and the phase of the total current, and calculate the phase difference between the bus voltage and the total current. Then calculate the first according to the following formula. Subharmonic resistive current component:

[0029] in, Indicates the first Subharmonic resistive current component Indicates the first RMS value of total current of the second harmonic Indicates the first The phase difference between the bus voltage and the total current under subharmonic conditions.

[0030] when At that time, the fundamental resistive current is obtained. ;when At that time, the third harmonic resistive current is obtained. Step S2 outputs the preprocessed full current signal. Fundamental resistive current Third harmonic resistive current Harmonic resistive current components and phase difference This is used in step S3 to construct a multidimensional state feature vector.

[0031] Preferably, step S3 includes: Step S3-1: The preprocessed full current signal output in step S2 Fundamental resistive current Third harmonic resistive current Harmonic resistive current components and phase difference Construct multidimensional state feature vectors The multidimensional state feature vector This includes current amplitude characteristics, resistive current ratio characteristics, third harmonic content characteristics, total harmonic distortion rate characteristics, phase characteristics, trend change rate characteristics, historical alarm characteristics, and environmental correction factors.

[0032] Among them, the characteristic of resistive current proportion Calculate according to the following formula:

[0033] in, Indicates the proportion of resistive current. Indicates the fundamental resistive current. This represents the preprocessed full current signal.

[0034] Third harmonic content characteristics Calculate according to the following formula:

[0035] in, Indicates the third harmonic content. This represents the third harmonic resistive current. This represents the fundamental resistive current.

[0036] Total Harmonic Distortion Characteristics Calculate according to the following formula:

[0037] in, Indicates the total harmonic distortion rate. Indicates the first Subharmonic resistive current component This indicates the highest harmonic order involved in the calculation.

[0038] Trend change rate characteristics Calculate according to the following formula:

[0039] in, This represents the rate of change of the resistive current trend. This represents the fundamental resistive current at the current sampling moment. This represents the fundamental resistive current at the previous sampling time.

[0040] The multidimensional state feature vector Represented as:

[0041] in, Represents a multidimensional state feature vector. The full current amplitude characteristics extracted from the preprocessed leakage current signal. Indicates the fundamental resistive current. Indicates the proportion of resistive current. Indicates the third harmonic content. Indicates the total harmonic distortion rate. This represents the phase difference between the bus voltage and the leakage current under the fundamental frequency. This represents the rate of change of the resistive current trend. Indicates historical alarm characteristics. Represents the environmental correction factor; step S3 outputs the multidimensional state feature vector. This is used as input for the intelligent diagnostic model in step S4.

[0042] Step S3-2: Based on the surge arrester voltage level, historical operating baseline, operating voltage fluctuation, and ambient temperature Ambient humidity and surface dirt The resistive current anomaly threshold, resistive current proportion threshold, third harmonic content threshold, total harmonic distortion rate threshold, phase drift threshold, and trend change rate threshold are dynamically corrected to obtain the corrected state judgment threshold vector. .

[0043] The initial judgment thresholds for each state characteristic are determined based on historical operating data of surge arresters confirmed to be in normal condition through testing or maintenance. The historical samples of normal conditions are grouped according to voltage level, operating voltage fluctuation range, ambient temperature range, ambient humidity range, and surface contamination level, and the 95th percentile value of the corresponding state characteristic under reference operating conditions is used as the initial judgment threshold. .

[0044] For the The state feature and the first For each type of influencing factor, a corresponding correction coefficient is determined based on the ratio of the 95th percentile of the current operating condition sample to that of the reference operating condition sample. The correction coefficient is calculated using the following formula:

[0045] in, Indicates the first Each state feature This indicates one of the influencing factors, including voltage level and operating voltage fluctuations, ambient temperature, ambient humidity, or surface contamination. This indicates the reference operating conditions corresponding to the influencing factors. This represents the 95th percentile of the sample. Represents the amplitude limiting function. and These represent the lower and upper limits of the correction factor, respectively.

[0046] In a preferred embodiment, the range of values ​​for each correction coefficient is as follows:

[0047] in, Indicates the first Voltage level and operating voltage fluctuation correction coefficient corresponding to each state characteristic This represents the ambient temperature correction factor. Indicates the environmental humidity correction factor. This represents the surface contamination state correction coefficient. When the number of historical samples of normal state under the corresponding operating condition is lower than the preset minimum sample number, the corresponding correction coefficient is set to 1, or the correction coefficient corresponding to the group of adjacent operating conditions closest to the current operating condition is used.

[0048] No. The corrected state judgment threshold for each state feature Calculate according to the following formula:

[0049] in, Indicates the first The initial judgment threshold for each state feature, and These represent the lower and upper limits of the threshold determined according to the surge arrester operation procedures, historical normal operating range, and test results, respectively.

[0050] The corrected state judgment threshold vector is represented as follows: ;in, This represents the number of state features involved in the state determination. Step S3 outputs a multi-dimensional state feature vector. and the corrected state judgment threshold vector This is used in step S4 to identify the operating status of the surge arrester and calculate the comprehensive risk score.

[0051] Combination Figure 3 Preferably, step S4 includes: Step S4-1: Input the multidimensional state feature vector \(F\) output in step S3 into the knowledge-enhanced Transformer diagnostic model, perform feature mapping on the multidimensional state feature vector, and obtain a high-dimensional state feature representation that the model can recognize. .

[0052] The feature embedding module is used to embed multidimensional state feature vectors The current amplitude characteristics, resistive current ratio characteristics, harmonic characteristics, phase characteristics, trend change characteristics, historical alarm characteristics, and environmental correction factors included in the model are uniformly encoded to reduce the dimensional differences between different types of state characteristics and improve the model's ability to represent the operating state characteristics of surge arresters.

[0053] The feature embedding process is represented as follows: ;in, This represents the high-dimensional state feature representation after feature embedding. This represents the multidimensional state feature vector constructed in step S3. Represents the feature mapping matrix, This represents the bias parameter.

[0054] When using a time-series diagnostic model, continuous The multidimensional state feature vectors of each sampling period constitute the temporal input sequence: ;in, Indicates the current time The corresponding time-series input sequence, This represents the multidimensional state feature vector at the current moment. This indicates the number of sampling periods involved in the time series modeling.

[0055] Step S4-1 outputs a high-dimensional feature representation. and timing input sequence This is used in step S4-2 for temporal feature extraction and knowledge enhancement fusion.

[0056] Step S4-2: Represent the high-dimensional state features output from step S4-1. and timing input sequence The timing encoding module of the knowledge-enhanced Transformer diagnostic model extracts the temporal variation pattern of leakage current characteristics within consecutive sampling periods through a self-attention mechanism, thereby obtaining a deep feature representation containing time-related information. .

[0057] The timing coding module includes a multi-head self-attention layer and a feedforward neural network layer. The multi-head self-attention layer is used to calculate the correlation weights between state features of different sampling periods, and to capture the correlation between resistive current amplitude changes, harmonic content changes, phase drift and trend changes. The feedforward neural network layer is used to perform nonlinear mapping on the attention calculation results to enhance the model's ability to extract features of the arrester degradation process.

[0058] The self-attention calculation process is expressed as follows: ; in, , and These represent the query matrix, key matrix, and value matrix, respectively. This represents the dimension of the key vector.

[0059] For the multi-head self-attention mechanism, its expression is: ; in, This represents the feature representation after time-series coding. Indicates the first Each attention head output result This indicates the output mapping matrix.

[0060] Step S4-2 Outputs the temporal feature representation This is used in step S4-3 for knowledge enhancement fusion and state feature optimization.

[0061] Step S4-3: Represent the deep temporal features output in step S4-2 Input knowledge enhancement module, and combine with surge arrester domain knowledge base The domain knowledge information is fused to obtain a knowledge-enhanced fusion feature representation. .

[0062] The knowledge base in the field of surge arresters It includes a rule base, a fault case base, a maintenance record base, and a knowledge graph. The rule base stores surge arrester operating procedures, leakage current anomaly criteria, and state classification rules. The fault case base stores historical fault types, abnormal state characteristics, fault causes, and corresponding risk levels. The maintenance record base stores power outage test results, maintenance measures, and maintenance conclusions. The knowledge graph is used to establish the correlation between state characteristics, anomaly thresholds, fault causes, risk levels, and operation and maintenance measures.

[0063] To enable the domain knowledge information of the surge arrester to participate in the numerical calculation of the knowledge-enhanced Transformer diagnostic model, the domain knowledge information is structured and encoded. Specifically, the leakage current anomaly criterion and state classification rules are converted into rule data units, which include state feature names, comparison relationships, judgment thresholds, applicable operating conditions, state levels, and rule weights.

[0064] Based on the current multidimensional state feature vector of the surge arrester The system determines whether the state feature satisfies the corresponding rule and converts the degree of rule satisfaction into a rule activation value. All rule activation values ​​corresponding to each rule are then arranged according to a preset rule number to form a rule activation vector. .

[0065] Furthermore, the current multi-dimensional state feature vector of the surge arrester is... The current state features are standardized with historical fault feature vectors in the fault case library. The similarity between the current state features and historical fault cases is calculated. Several historical fault cases with high similarity are selected, and their similarity, fault type, and state level are encoded to form a case similarity vector. .

[0066] Simultaneously, based on the abnormal features in the current surge arrester status characteristics that exceed the corrected status judgment threshold, the fault causes, risk levels, and operation and maintenance measures associated with the abnormal features are retrieved from the knowledge graph. Furthermore, the knowledge entities and the relationships between entities are converted into knowledge graph embedding vectors using a knowledge graph embedding method. .

[0067] Activate the rule vector Case similarity vector and knowledge graph embedding vectors By concatenating and mapping, a domain knowledge feature representation consistent with the temporal feature representation dimension is obtained. , represented as:

[0068] in, Representation of domain knowledge features Represents the knowledge feature mapping matrix. Indicates the bias parameter. This represents the concatenation result of the rule activation vector, case similarity vector, and knowledge graph embedding vector.

[0069] The knowledge enhancement module is based on deep temporal features. Domain knowledge feature representation Calculate the knowledge fusion weights, and then use these weights to selectively fuse domain knowledge features. The fusion process is expressed as follows:

[0070] in, Indicates the knowledge fusion weight, and Represents a trainable mapping matrix. Indicates the bias parameter. This represents the activation function. This indicates element-wise multiplication. This represents the enhanced feature representation after incorporating domain knowledge.

[0071] By using the above method, the domain knowledge of surge arresters, which was originally stored in the form of text, rules and cases, is converted into rule activation vectors, case similarity vectors and knowledge graph embedding vectors. This enables the domain knowledge to be subjected to unified-dimensional mathematical operations with the temporal features extracted by the Transformer model, thereby improving the model's ability to identify and explain abnormal states of surge arresters.

[0072] The knowledge-enhanced Transformer diagnostic model is trained using offline supervised learning. Training data includes online surge arrester monitoring data, power outage test data, historical fault records, and maintenance results data. Corresponding status labels are determined based on the test results, fault records, and maintenance conclusions.

[0073] The status labels include normal status, warning status, abnormal status, and severe deterioration status. For samples with multiple status judgment results, the final status label is determined by professionals with experience in the operation, maintenance, or testing of surge arresters.

[0074] In one implementation, leakage current characteristics, resistive current characteristics, harmonic characteristics, phase characteristics, and environmental parameters from multiple consecutive sampling periods are constructed into time-series training samples. The training set, validation set, and test set are divided according to the device number to avoid data leakage caused by adjacent time data of the same surge arrester entering different datasets at the same time.

[0075] During model training, a joint loss function is constructed using state classification loss, trend prediction loss, and knowledge consistency loss.

[0076] in, Represents the state classification loss. Indicates the predicted loss based on the deterioration trend. This represents the consistency loss between the model output and the domain rules. This indicates the weight of the corresponding loss term.

[0077] After training, the knowledge-enhanced fusion feature representation is used. Input the model's output layer to obtain state classification probabilities, state recognition results, and degradation trend prediction results:

[0078] in, This represents the state classification probability vector. This indicates the state recognition result. This indicates the predicted degradation trend. .

[0079] The model confidence level is calculated based on the state classification probability vector. and model uncertainty :

[0080] in, This represents the maximum probability predicted by the model, used to characterize the reliability of the model's identification of the current arrester state. This represents the model uncertainty calculated based on the state classification probability distribution, and is used to reflect the degree of uncertainty in the model's ability to distinguish between different state categories.

[0081] Through the above process, step S4-3 outputs the state classification probability vector. Status recognition results Model confidence Model uncertainty and the results of the degradation trend prediction The above output results are then compared with the multidimensional state feature vector obtained in step S3. and the corrected state judgment threshold The common input step S4-4 is used to calculate the comprehensive risk score of the surge arrester and determine the final operating status level.

[0082] Step S4-4: Based on the state recognition result output in step S4-3 Model confidence Model uncertainty Deterioration trend prediction results Combined with the multidimensional state feature vector constructed in step S3 and the corrected state judgment threshold Calculate the comprehensive risk score of the surge arrester .

[0083] The comprehensive risk score takes into account the degree of abnormality in state characteristics, model reliability, and equipment degradation trend. Its calculation formula is as follows:

[0084] in, This indicates the degree of abnormal deviation of the multidimensional state features from the corrected state judgment threshold. Indicates the model confidence level. Indicates model uncertainty. This indicates the predicted degradation trend. , , , This represents the corresponding weighting coefficient.

[0085] The weighting coefficients satisfy:

[0086] Based on comprehensive risk score The operating status level of the surge arrester is determined by the location of the interval. When the risk level is below the first risk threshold, it is considered a normal state; when... When the risk level is between the first and second risk thresholds, it is determined to be in a state of alert; when... When the risk level is between the second and third risk thresholds, it is determined to be an abnormal state; when... When the value exceeds the third risk threshold, it is determined to be in a severely deteriorated state.

[0087] When the comprehensive risk score The corresponding operational state level and the state recognition results output by the knowledge-enhanced Transformer diagnostic model If there is a discrepancy, mark it as pending review. Set to 1; when the two are consistent, retain the verification identifier determined in step S4-3. constant.

[0088] Step S4-4 Outputs the surge arrester status identification result Comprehensive risk score Model confidence Model uncertainty and the results of the degradation trend prediction and pending review mark The output results and the multidimensional state feature vector constructed in step S3 are then combined. As input for the knowledge fusion reasoning in step S5. Combined with Figure 4 Preferably, step S5 includes: Step S5-1: Call the lightning arrester domain knowledge base that was pre-built before training the knowledge-enhanced Transformer diagnostic model or conducting online diagnostics. .

[0089] The knowledge base in the field of surge arresters By collecting surge arrester operation procedures, leakage current anomaly criteria, state classification rules, historical fault cases, maintenance records, and test results, the collected data is cleaned, standardized, and subjected to knowledge extraction processing. State characteristics, anomaly thresholds, fault types, fault causes, risk levels, and operation and maintenance measures knowledge entities are extracted, and the relationships between these knowledge entities are established to construct the system.

[0090] The knowledge base in the field of surge arresters Including rule base Fault Case Library Maintenance record library and knowledge graph It is represented as:

[0091] in, This represents the rule base, which stores surge arrester operating procedures, leakage current anomaly criteria, and status classification rules. This represents a fault case library, used to store historical fault types, anomaly characteristics, fault causes, and risk levels. This refers to the maintenance record library, used to store maintenance measures, test results, and maintenance effects. A knowledge graph is used to establish relationships between state characteristics, anomaly thresholds, fault causes, risk levels, and operational measures.

[0092] Step S5-1 Output surge arrester domain knowledge base This is used for knowledge verification and fusion reasoning in step S5-2.

[0093] Step S5-2: Utilize historical fault cases and maintenance results to access the surge arrester domain knowledge base called in Step S5-1. Verification is performed by inputting the status characteristics from historical operating data into the surge arrester domain knowledge base for reasoning, and comparing the reasoned fault causes, risk levels, and maintenance measures with actual fault records and repair results.

[0094] When the reasoning result of the knowledge base matches the actual fault records and maintenance results, the corresponding knowledge entities and relationships are retained, and the verified surge arrester domain knowledge base is recorded as follows: When there is a discrepancy between the knowledge base reasoning results and the actual fault records and maintenance results, candidate correction information is generated for the corresponding rule weights, knowledge entity associations, and state classification rules. This candidate correction information is reviewed by professionals and verified using historical fault cases and corresponding maintenance results. The updated surge arrester knowledge base is then recorded as follows: .

[0095] The state recognition result output in step S4 Comprehensive risk score Model confidence Model uncertainty Degradation trend prediction results Combined with the multidimensional state feature vector constructed in step S3 Input the updated lightning arrester domain knowledge base Perform fusion reasoning to obtain knowledge fusion diagnostic results. The fusion reasoning process is expressed as follows:

[0096] in, This indicates the results of the knowledge fusion diagnosis. This indicates the arrester status identification result. This represents the overall risk score. Indicates the model confidence level. Indicates the model uncertainty. This indicates the predicted degradation trend. Represents a multidimensional state feature vector. This indicates the updated knowledge base for surge arresters. This represents a knowledge fusion inference function.

[0097] Step S5-2: Output the knowledge fusion diagnostic results. This is used in step S5-3 to generate an interpretable diagnostic conclusion.

[0098] Step S5-3: Read the check flag output in step S4 When pending review At that time, the state recognition result will be... This is determined to be the final diagnostic result. The knowledge fusion diagnostic result (G) is determined as the final knowledge fusion diagnostic result. When the pending verification flag (B=1) is displayed, the corresponding diagnostic task will be sent to the manual verification module, where it will be verified by professionals with experience in surge arrester operation and maintenance, testing and inspection, or condition evaluation.

[0099] The manual review module displays the original leakage current waveform, bus voltage waveform, spectrum analysis results, and multi-dimensional state feature vector to the reviewers. Corrected state judgment threshold, state classification probability vector Status recognition results Comprehensive risk score Model confidence Model uncertainty Deterioration trend prediction results The abnormal criteria for the hit, the rule judgment results, similar historical fault cases, and historical maintenance records.

[0100] When the manual review conclusion and the status recognition result If they match, the review is deemed successful, and the status identification result is entered. This is determined to be the final diagnostic result. And integrate knowledge with diagnostic results This was determined as the final knowledge fusion diagnostic result. When the manual review conclusion and the status identification result are different If there is no consistency, the review is deemed unsuccessful, and the manual review conclusion is determined as the final diagnostic result. And the corresponding diagnostic samples are marked as error correction samples.

[0101] When the final state diagnosis result With state recognition results In case of inconsistency, the final diagnosis result shall prevail. Replacement status recognition result Re-enter the updated knowledge base for surge arresters. Perform fusion reasoning to obtain the final knowledge fusion diagnostic result. Its expression is:

[0102] in, This represents the final knowledge fusion diagnostic result obtained based on the final state diagnostic result. This indicates the final diagnostic result determined after manual review. Record the surge arrester equipment number, data acquisition time, original model output, comprehensive risk score, model confidence, model uncertainty, hit rule, reviewer, review time, review basis, review conclusion, and operation and maintenance handling results to form a review record; after review, write the error correction sample and review record into the fault case library or model training sample library to generate candidate update information for rule weights, knowledge entity association relationships, and state classification rules, or to use for offline retraining of the knowledge-enhanced Transformer diagnostic model.

[0103] Step S5-3 Output the final state diagnosis result And the final knowledge fusion diagnostic results This is used in step S5-4 to generate an interpretable diagnostic conclusion.

[0104] Step S5-4: Based on the knowledge fusion diagnostic results obtained in step S5-3 The output surge arrester can interpret diagnostic conclusions, which include status level, cause of anomaly, key impact characteristics, risk level, and operation and maintenance recommendations.

[0105] Among them, the key influencing features are based on the multidimensional state feature vector. The degree of abnormal deviation of each state characteristic and its influence weight are jointly determined. Contribution of each state feature Calculate according to the following formula:

[0106] in, Indicates the first The influence weights corresponding to each state feature Indicates the first The influence weights corresponding to each state feature Indicates the first The normalized anomaly deviation of each state feature relative to the corresponding corrected state judgment threshold. This indicates the number of state features involved in the diagnosis; when none of the state features exceed the corresponding state judgment threshold, the contribution of each state feature is set to 0.

[0107] The normalized outlier deviation Determined based on the direction of abnormal changes in state characteristics; for state characteristics whose values ​​increase, indicating abnormalities, the following formula is used for calculation:

[0108] in, Indicates the first The current feature value of each state feature. Indicates the first The corrected state judgment threshold corresponding to each state feature This represents a preset constant used to prevent the denominator from being zero; for state characteristics such as phase drift that may indicate abnormality in both positive and negative directions, the normalized abnormal deviation is calculated based on its absolute deviation from the normal operating baseline.

[0109] Based on contribution The state features are sorted, and those with a contribution higher than a preset contribution threshold or ranking within the top preset number are selected as key influencing features. These key influencing features are then compared with the updated surge arrester domain knowledge base. The abnormal criteria in the rule base, the historical fault patterns in the fault case base, and the relationships in the knowledge graph are matched to determine the fault cause that leads to the abnormal state of the surge arrester.

[0110] Based on the state recognition results Final knowledge fusion diagnostic results Comprehensive risk score Model confidence Model uncertainty Deterioration trend prediction results Based on the cause of the failure, corresponding risk levels and operation and maintenance recommendations are generated.

[0111] To verify the feasibility of the intelligent arrester diagnosis method based on leakage current characteristics and knowledge enhancement model described in this invention, in a specific embodiment, an intelligent arrester diagnosis simulation platform is constructed, and the leakage current signal and operating status data of the arrester generated by simulation are used to verify the leakage current preprocessing, multi-dimensional state feature construction, dynamic threshold correction, knowledge enhancement Transformer diagnosis, comprehensive risk assessment, knowledge fusion reasoning and interpretable diagnostic output process described in this invention.

[0112] like Figure 5As shown, the intelligent diagnostic simulation platform for surge arresters includes a simulation condition setting area, a surge arrester simulation scene display area, a multi-dimensional state feature display area, a time-domain waveform display area, a spectrum analysis area, a model diagnostic output area, a knowledge fusion and rule matching area, and an interpretable diagnostic conclusion output area.

[0113] This embodiment uses a metal oxide surge arrester as the simulated diagnostic object. Based on the arrester's voltage level, equipment model, operating environment, and surface contamination status, corresponding simulation conditions are set. The leakage current signal sampling frequency, sampling duration, timing input length, number of simulated samples, and the structural and training parameters of the knowledge-enhanced Transformer diagnostic model are also set. The specific settings for the simulation data and model parameters are shown in Table 1.

[0114]

[0115] During the simulation, based on the differences in leakage current amplitude, resistive current component, harmonic content, phase characteristics, and trends under normal, warning, abnormal, and severely degraded states, corresponding simulated leakage current signals are generated. These simulated leakage current signals undergo outlier removal, filtering, normalization, and power frequency interference suppression. Furthermore, spectral analysis and phase difference calculations are performed in conjunction with the simulated bus voltage signal to extract multi-dimensional state characteristics, including total current amplitude characteristics, fundamental resistive current component, third harmonic resistive current component, resistive current proportion, third harmonic content, total harmonic distortion rate, fundamental phase difference, and resistive current trend change rate.

[0116] Based on the voltage level, operating voltage fluctuation, ambient temperature, ambient humidity, and surface contamination status of the simulated surge arrester, the state judgment thresholds corresponding to each state characteristic are dynamically corrected, and the corrected state judgment thresholds are used to calculate the degree of abnormal deviation of multidimensional state characteristics.

[0117] The constructed multidimensional state feature vectors are input into the knowledge-enhanced Transformer diagnostic model to obtain the classification probabilities, state identification results, model confidence, model uncertainty, and degradation trend prediction results for four states: normal, attentive, abnormal, and severely deteriorated. Furthermore, a comprehensive risk score is calculated based on the degree of deviation of the state features from the normal state, and the corresponding operating state level is determined according to the interval of the comprehensive risk score. The model diagnostic results under different simulation conditions are shown in Table 2.

[0118]

[0119] As shown in Table 2, under the four simulated operating conditions—normal, attentive, abnormal, and severely degraded—the knowledge-enhanced Transformer diagnostic model can output state recognition results corresponding to the preset simulated operating conditions. Specifically, under the abnormal state simulated operating condition, when the model uncertainty exceeds the preset uncertainty threshold, the system sets the pending verification flag to 1 and sends the corresponding diagnostic task to the manual verification module; under other simulated operating conditions where the verification condition is not triggered, the pending verification flag is set to 0.

[0120] Furthermore, the multidimensional state characteristics under the simulated abnormal operating conditions were matched with the leakage current anomaly criteria, state classification rules, historical fault cases, and knowledge graph associations in the surge arrester knowledge base. The knowledge fusion reasoning results show that the simulated operating conditions hit the anomaly criteria such as increased fundamental resistive current, increased resistive current ratio, and increased third harmonic content, and have a high degree of similarity with fault cases of surface contamination and internal moisture.

[0121] Based on the degree of abnormal deviation and contribution of each state characteristic, the key influencing features leading to the abnormal state of the simulated surge arrester are determined as the fundamental resistive current component, the resistive current ratio, and the resistive current trend change rate. Combining the relationships in the rule base, fault case base, and knowledge graph, the potential cause of the abnormality is determined to be surface contamination or internal moisture risk. Maintenance recommendations are output, including shortening the detection cycle, continuously tracking leakage current changes, and conducting surface cleaning, insulation condition inspection, and leakage current comparison tests when power outage conditions are available.

[0122] Figure 5 The simulation results show the operating parameters, multidimensional state characteristics, time-domain waveforms, spectrum analysis results, model diagnostic output, knowledge fusion and rule matching results, and interpretable diagnostic conclusions. Simulation results demonstrate that this invention can complete surge arrester leakage current feature extraction, state identification, comprehensive risk assessment, verification triggering, anomaly cause analysis, and operation and maintenance decision output.

[0123] It should be noted that the leakage current data, operating status, and diagnostic results used in this embodiment are all simulated data generated by the simulation platform. They are used to illustrate the specific execution process of the present invention and the data transmission relationship between each functional module, and do not represent the actual measured data or actual maintenance results of a specific substation surge arrester.

Claims

1. A method for intelligent diagnosis of surge arresters based on leakage current characteristics and a knowledge enhancement model, characterized in that, Includes the following steps: Step S1: Through the online monitoring terminal and leakage current digital acquisition device of the surge arrester, the total current, bus voltage, number of operations, historical operating data and environmental parameters of the metal oxide surge arrester are collected simultaneously, and the resistive current measurement value or the phase reference data required for resistive current extraction is obtained. Step S2: The leakage current signal acquired in step S1 is subjected to outlier removal, filtering, normalization and power frequency interference suppression. Based on the phase relationship between the bus voltage and the leakage current, the resistive current component is extracted using the fast Fourier transform spectrum analysis method. Step S3: Based on the resistive current component extracted in step S2 and the multi-source data collected in step S1, construct a multi-dimensional state feature vector that includes current amplitude characteristics, resistive current ratio, harmonic characteristics, phase characteristics, trend characteristics, historical alarm characteristics and environmental correction factors, and dynamically correct the state judgment threshold according to the surge arrester voltage level, historical operating baseline, environmental parameters and surface contamination status. Step S4: Input the multi-dimensional state feature vector constructed in step S3 into the knowledge-enhanced Transformer diagnostic model, integrate the pre-constructed surge arrester domain knowledge information, obtain the state recognition result, model confidence, model uncertainty and degradation trend prediction result, and calculate the comprehensive risk score by combining the corrected state judgment threshold to determine the surge arrester state level. Step S5: Integrate the output of Step S4 with the surge arrester domain knowledge base to output the cause of the anomaly, key impact characteristics, risk level and operation and maintenance suggestions, and review the diagnostic results that have model-rule conflicts or insufficient model reliability.

2. The intelligent diagnostic method for surge arresters based on leakage current characteristics and knowledge enhancement model according to claim 1, characterized in that, In step S1, the original monitoring data is obtained from the online monitoring terminal of the surge arrester, the digital acquisition device of leakage current, the digital acquisition result of the pointer-type leakage current meter, or the historical database of the existing monitoring system; the historical operating data includes historical detection data, historical alarm records and maintenance records.

3. The intelligent diagnostic method for surge arresters based on leakage current characteristics and knowledge enhancement model according to claim 1, characterized in that, In step S2, the outlier removal adopts the Laida criterion, the filtering adopts the Butterworth low-pass filter, the power frequency interference suppression adopts the adaptive notch filter, and the normalization adopts the Z-score standardization method.

4. The intelligent diagnostic method for surge arresters based on leakage current characteristics and knowledge enhancement model according to claim 1, characterized in that, In step S3, the multidimensional state feature vector includes the total current amplitude feature, fundamental resistive current, resistive current ratio, third harmonic content, total harmonic distortion rate, fundamental phase difference, resistive current trend change rate, historical alarm features, and environmental correction factor; the dynamic correction of the state judgment threshold is based on voltage level, operating voltage fluctuation, ambient temperature, ambient humidity, and surface contamination status, and is performed by multiplying the correction coefficients corresponding to each influencing factor.

5. The intelligent diagnostic method for surge arresters based on leakage current characteristics and knowledge enhancement model according to claim 1, characterized in that, In step S4, the knowledge-enhanced Transformer diagnostic model includes a feature embedding layer, a temporal coding layer, a knowledge enhancement layer, and an output layer. The temporal coding layer uses a multi-head self-attention mechanism to extract the temporal correlation of state features within multiple consecutive sampling periods. The knowledge enhancement layer converts lightning arrester domain knowledge information into rule activation vectors, case similarity vectors, and knowledge graph embedding vectors, and performs gating fusion with the temporal features.

6. The intelligent diagnostic method for surge arresters based on leakage current characteristics and knowledge enhancement model according to claim 1, characterized in that, In step S4, the comprehensive risk score comprehensively considers the degree of abnormal deviation of multidimensional state features from the corrected threshold, the degree of insufficient model confidence, model uncertainty, and the prediction results of degradation trend, and determines the normal, attention, abnormal, or severely degraded state level according to the interval of the comprehensive risk score; when the model output is inconsistent with the level corresponding to the comprehensive risk score or the model reliability is insufficient, a review mark is generated.

7. The intelligent diagnostic method for surge arresters based on leakage current characteristics and knowledge enhancement model according to claim 1, characterized in that, In step S5, the surge arrester domain knowledge base includes a rule base, a fault case base, a maintenance record base, and a knowledge graph; the knowledge base is verified and updated using historical fault cases and maintenance results to obtain an updated knowledge base; the output of step S4 is input into the updated knowledge base for fusion reasoning to obtain a knowledge fusion diagnostic result.

8. The intelligent diagnostic method for surge arresters based on leakage current characteristics and knowledge enhancement model according to claim 1 or 7, characterized in that, In step S5, when there is a conflict between the model and the rules or the model is not reliable enough, the diagnostic task is sent to the manual review module, where professionals review the model by combining the original waveform, spectrum analysis, multidimensional features and model output. If the review fails, the review conclusion will be used as the final diagnosis result and marked as an error correction sample for use in knowledge base updates or model retraining.

9. The intelligent diagnostic method for surge arresters based on leakage current characteristics and knowledge enhancement model according to claim 1, characterized in that, In step S5, the abnormal causes and key impact features are determined by calculating the normalized abnormal deviation and contribution of each state feature, and matching them with rules and fault cases in the knowledge base; the operation and maintenance suggestions are generated based on the final state diagnosis results: normal state output normal inspection, attention state output shorten inspection cycle, abnormal state output power outage detection, and severely degraded state output shutdown inspection or replacement.

10. The intelligent diagnostic method for surge arresters based on leakage current characteristics and knowledge enhancement model according to claim 1, characterized in that, In step S5, the verification and updating process of the surge arrester domain knowledge base includes: inputting the state characteristics in historical operating data into the knowledge base for reasoning, and comparing the reasoned fault causes, risk levels, and operation and maintenance measures with actual fault records and maintenance results; when the reasoning results are consistent with the actual records, retaining the corresponding knowledge entities and relationships; when there are deviations, generating candidate correction information for corresponding rule weights, knowledge entity relationships, and state classification rules, and using the candidate correction information, after being reviewed by professionals and verified using historical fault cases, to update the surge arrester domain knowledge base.

Images