Intelligent panoramic monitoring management system and method for new energy box transformer

Abstract

The present invention relates to the field of renewable energy power generation technology, and discloses an intelligent panoramic monitoring management system and method for new energy box transformers. The method comprises: obtaining an original monitoring data set of a target box transformer and performing data cleaning; extracting key features from the cleaned data and generating a weighted feature set; constructing a box transformer fault prediction model to obtain a fault risk assessment value; performing graded early warnings on potential faults based on the assessment values; constructing a box transformer fault diagnosis knowledge graph, and performing reasoning and tracing in combination with the knowledge graph when providing high-risk or medium-risk early warnings; the present invention realizes comprehensive monitoring of the box transformer operating status and intelligent diagnosis of faults through multi-source data fusion and knowledge graph reasoning, thereby improving the operating reliability and maintenance efficiency of the new energy box transformer.

Description

Technical Field

[0001] The present invention relates to the field of renewable energy power generation technology, and more specifically, to an intelligent panoramic monitoring management system and method for renewable energy box transformers. Background Art

[0002] As an important part of the new energy power generation system, the operation status of the new energy box transformer directly affects the power generation efficiency and safety of the new energy power station. The traditional new energy box transformer monitoring and management system has some shortcomings and cannot meet the growing demand for intelligent and panoramic monitoring.

[0003] The patent application with publication number CN116613884A discloses a comprehensive monitoring and management system for new energy box transformers, which is used to solve the problem that the integration of various electrical equipment in the box transformer is low and the booster station cannot remotely manage and monitor the new energy electric field box transformer. The system realizes the high and low voltage side protection control of the new energy box transformer and the collection and control of environmental data through the cooperation of CPU plug-in, AC plug-in, input plug-in, output plug-in, environmental control main plug-in and environmental control auxiliary plug-in, so that the booster station can remotely manage and automatically monitor the new energy electric field box transformer. However, the system mainly focuses on the collection and control of the electrical quantity of the box transformer, lacks the fusion analysis of the multi-source heterogeneous data of the box transformer, and it is difficult to comprehensively evaluate the operating status and potential risks of the box transformer.

[0004] The patent application with publication number CN117578732A discloses a new energy box transformer monitoring and management system for special environments. The system predicts the equipment operation status and environmental status of the day based on the box transformer's historical operation data and environmental data, and determines whether the box transformer has abnormal conditions by comparing the predicted value with the actual value. This data-driven prediction method can, to a certain extent, warn of potential abnormal risks of the box transformer, but the prediction time scale is short and lacks the ability to analyze the causes and intelligently diagnose the detected abnormal conditions, so its practical value is limited.

[0005] In summary, most of the existing new energy box-type transformer monitoring and management systems use a single data source and lack the ability to integrate and analyze multi-source heterogeneous monitoring data, making it difficult to accurately and comprehensively evaluate the operating status of the box-type transformer; at the same time, there is often a lack of effective diagnostic reasoning and traceability analysis methods for detected box-type transformer faults. Summary of the invention

[0006] In order to overcome the above-mentioned defects of the prior art, the present invention provides an intelligent panoramic monitoring management system and method for a new energy box transformer to improve the operation reliability and maintenance efficiency of the new energy box transformer.

[0007] To achieve the above object, the present invention provides the following technical solutions:

[0008] Intelligent panoramic monitoring and management method for new energy box-type substations, including:

[0009] Obtain the original monitoring data set of the target box-type substation, clean the original monitoring data set to obtain the cleaned monitoring data set, extract the key features reflecting the operation status of the box-type substation from the cleaned monitoring data set to obtain the original feature set; perform importance analysis on each feature in the original feature set, and assign weights to each feature in the original feature set according to the analysis results to generate a weighted feature set;

[0010] Construct a box-type substation fault prediction model, and obtain a fault risk assessment value Q according to the weighted feature set and the box-type substation fault prediction model; based on the fault risk assessment value, perform hierarchical early warning on potential faults to generate fault early warning information; 0 < Q < 1; The hierarchical early warning of potential faults includes early warning of high-risk level, early warning of medium-risk level, early warning of low-risk level, and early warning of no-risk level;

[0011] Construct a box-type substation fault diagnosis knowledge graph, and when the high-risk level or medium-risk level is early warned, reason and trace the fault early warning information in combination with the box-type substation fault diagnosis knowledge graph.

[0012] Further, the obtaining of the original monitoring data set of the target box-type substation includes:

[0013] Collect the time-series change data of the internal temperature of the box-type substation; collect the time-series change data of the internal air humidity of the box-type substation; collect the time-series pulse signals of partial discharge on the surface of the insulating parts; collect the time-series concentration data of characteristic fault gases; collect the time-series waveform data of the core grounding current; summarize all the data to form the original monitoring data set.

[0014] Further, the generating of the weighted feature set includes:

[0015] Calculate the Pearson correlation coefficient between each feature in the original feature set and the fault state, select the features with Pearson correlation coefficients higher than the preset correlation threshold as candidate features, and generate a candidate feature subset;

[0016] Calculate the information gain of each candidate feature in the candidate feature subset with respect to the fault state, select the candidate features with information gains higher than the preset gain threshold as the optimal features, and generate an optimal feature subset;

[0017] Use the Gini index method to calculate the Gini index of each optimal feature in the optimal feature subset, and after normalizing the Gini index of each optimal feature, use it as the weight coefficient;

[0018] Assign the weight coefficient to each feature in the optimal feature subset to generate a weighted feature set.

[0019] Further, the hierarchical early warning of potential faults to generate fault early warning information includes:

[0020] According to the fault risk assessment value output by the box-type transformer fault prediction model, three risk thresholds are set: high risk threshold H, medium risk threshold M and low risk threshold L, where 0 <L<M<H<1;

[0021] When Q>H, a high risk level is warned; when M≤Q≤H, a medium risk level is warned; when L<Q<M, a low risk level is warned; when Q≤L, a no risk level is warned;

[0022] When a high risk level or a medium risk level is warned, an alarm signal is triggered and fault warning information is generated. The fault warning information includes the fault risk level and fault symptom characteristics; the fault symptom characteristics include numerical symptom characteristics and discrete symptom characteristics.

[0023] Furthermore, the construction of the box-type transformer fault diagnosis knowledge graph includes:

[0024] Define core concepts to form a concept layer; the core concepts include the concept of box-type transformer components, the concept of materials, and the concept of faults;

[0025] Extracting semantic relations between core concepts to form a relational layer; the semantic relations include causal relations, concurrent relations and co-location relations; the causal relations are presented as causal chains, the co-location relations are presented as co-location relation families, and the concurrent relations are presented as concurrent probability networks;

[0026] The core concepts of the concept layer are used as nodes, and the semantic relations of the relationship layer are used as edges to construct the topological structure of the knowledge graph for box-type transformer fault diagnosis.

[0027] Typical features of each fault concept are mined, marked as attribute features, and filled into the topological structure of the transformer fault diagnosis knowledge graph as the attributes of the fault nodes corresponding to the fault concepts; the attribute features of the fault concepts include numerical attribute features and discrete attribute features.

[0028] Furthermore, the reasoning and tracing of the fault warning information in combination with the box-type transformer fault diagnosis knowledge graph includes:

[0029] Calculate the comprehensive matching degree between the fault symptom characteristics in the fault warning information and the attribute characteristics of each fault node in the box-type transformer fault diagnosis knowledge graph, and select the top N nodes with the highest comprehensive matching degree as candidate fault nodes;

[0030] Extract the one-hop neighbor nodes of the candidate fault node in the transformer fault diagnosis knowledge graph to form a semantic subgraph of the candidate fault;

[0031] Based on the causal chain of the box-type transformer fault diagnosis knowledge graph, the causal chain in the semantic subgraph is generated. By analyzing the causal chain in the semantic subgraph, the cause of the candidate fault can be traced back.

[0032] Based on the co-location relationship family and concurrent probability network of the box transformer fault diagnosis knowledge graph, count the number of co-location nodes and the occurrence probabilities of concurrent nodes in the semantic subgraph, and discriminate the location, severity, and associated faults of the candidate faults;

[0033] Integrate the causes, locations, severities, and associated faults of the candidate faults, and output a fault diagnosis report based on the box transformer fault diagnosis knowledge graph.

[0034] Furthermore, the calculation of the comprehensive matching degree between each fault symptom feature in the fault warning information and each fault node attribute feature in the box transformer fault diagnosis knowledge graph includes:

[0035] Calculate the similarity DE between each numerical symptom feature in the fault warning information and each numerical attribute feature of each fault node in the box transformer fault diagnosis knowledge graph;

[0036] Based on the weighted Jaccard coefficient, calculate the similarity DB between each discrete symptom feature in the fault warning information and each discrete attribute feature of each fault node in the box transformer fault diagnosis knowledge graph;

[0037] Perform a weighted average on the similarity DE and the similarity DB to obtain the comprehensive matching degree MP between the fault warning information and each fault node.

[0038] For a new energy box transformer intelligent panoramic monitoring and management system, which is used to implement the above-mentioned new energy box transformer intelligent panoramic monitoring and management method, the system includes:

[0039] Data processing module: used to obtain the original monitoring data set of the target box transformer, clean the original monitoring data set to obtain the cleaned monitoring data set, extract the key features reflecting the operation status of the box transformer from the cleaned monitoring data set to obtain the original feature set; perform importance analysis on each feature in the original feature set, and assign weights to each feature in the original feature set according to the analysis results to generate a weighted feature set;

[0040] Fault prediction module: used to construct a box transformer fault prediction model, and obtain a fault risk assessment value Q according to the weighted feature set and the box transformer fault prediction model; based on the fault risk assessment value, perform hierarchical early warning on potential faults to generate fault warning information; 0 < Q < 1; the hierarchical early warning of potential faults includes early warning of high-risk level, early warning of medium-risk level, early warning of low-risk level, and early warning of no-risk level;

[0041] Fault tracing module: used to construct a box transformer fault diagnosis knowledge graph, and when the high-risk level or medium-risk level is early warned, perform reasoning and tracing on the fault warning information in combination with the box transformer fault diagnosis knowledge graph.

[0042] An electronic device includes a memory, a central processing unit, and a computer program stored in the memory and executable on the central processing unit. When the central processing unit executes the computer program, the above-mentioned intelligent panoramic monitoring management method for new energy box transformers is implemented.

[0043] A computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed, implements the above-mentioned intelligent panoramic monitoring management method for new energy box transformers.

[0044] Compared with the prior art, the present invention has the following beneficial effects:

[0045] The present invention realizes a comprehensive perception of the operating status of the box transformer by collecting multi-source heterogeneous data such as temperature, humidity, partial discharge, characteristic gas concentration, core grounding current, etc. inside the box transformer. This multi-dimensional data collection and fusion greatly improves the accuracy and comprehensiveness of fault diagnosis, and can capture complex fault modes that are difficult to detect with a single sensor. The original features are screened and weighted using multiple algorithms such as Pearson correlation coefficient, information gain, and Gini index. This multi-level feature engineering process effectively extracts the most representative fault features, reduces data redundancy, and improves the efficiency and accuracy of subsequent fault prediction models. By setting three risk thresholds of high, medium, and low, a refined hierarchical warning of potential faults is achieved. This hierarchical warning mechanism can not only detect high-risk faults in a timely manner, but also provide early warnings for medium and low risk hidden dangers, providing more decision-making space and prevention time for operation and maintenance personnel. The knowledge graph for box transformer fault diagnosis is constructed, combining the experience of domain experts with the data-driven model. When warning of high or medium risks, the knowledge graph is used for reasoning and tracing, which can quickly locate the cause of the fault, determine the location and severity of the fault, and greatly improve the efficiency and explainability of fault diagnosis. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0047] Figure 1 It is a principle flow chart of the intelligent panoramic monitoring management method for new energy box transformers in the present invention;

[0048] Figure 2 A flow chart of a method for obtaining an original monitoring data set of a target box-type transformer in the intelligent panoramic monitoring management method for a new energy box-type transformer of the present invention;

[0049] Figure 3 A flowchart of a method for extracting key features reflecting the operating status of a box-type transformer from a cleaned monitoring data set in a method for intelligent panoramic monitoring management of a new energy box-type transformer of the present invention;

[0050] Figure 4 A flow chart of a method for generating a weighted feature set in a method for intelligent panoramic monitoring management of a new energy box transformer of the present invention;

[0051] Figure 5 A flow chart of a method for constructing a knowledge graph for fault diagnosis of a new energy box-type transformer in the method for intelligent panoramic monitoring and management of the new energy box-type transformer of the present invention;

[0052] Figure 6 A flow chart of a method for calculating the comprehensive matching degree of each fault symptom feature in the fault warning information and each fault node attribute feature in the box transformer fault diagnosis knowledge map in the intelligent panoramic monitoring management method for new energy box transformer of the present invention;

[0053] Figure 7 This is a functional module diagram for the intelligent panoramic monitoring and management system for new energy box transformers in the present invention. DETAILED DESCRIPTION

[0054] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0055] Example 1

[0056] See also Figure 1 As shown, this embodiment provides a method for intelligent panoramic monitoring and management of new energy box transformers, including:

[0057] Step S1000, obtaining the original monitoring data set of the target box-type transformer, performing data cleaning on the original monitoring data set to obtain a cleaned monitoring data set, extracting key features reflecting the operating status of the box-type transformer from the cleaned monitoring data set to obtain an original feature set; performing importance analysis on each feature in the original feature set, assigning weights to each feature in the original feature set according to the analysis results, and generating a weighted feature set;

[0058] Further, step S1000 includes:

[0059] Step S1100, obtaining an original monitoring data set of a target box-type transformer;

[0060] Furthermore, if Figure 2 As shown, step S1100 includes:

[0061] Step S1110, arranging optical fiber temperature sensors at key locations such as the inner wall of the oil tank, the iron core, the winding, and the casing of the box-type transformer body to collect the time series variation data of the internal temperature of the box-type transformer;

[0062] Step S1120, arranging a humidity sensor in the internal space of the box-type transformer body to collect time-series variation data of the air humidity inside the box-type transformer;

[0063] Step S1130, arranging ultrasonic partial discharge sensors on the surfaces of insulating parts such as high-voltage bushings and insulating partitions to collect timing pulse signals of partial discharge on the surfaces of the insulating parts;

[0064] Step S1140, a dissolved gas analyzer is built into the transformer oil pillow to collect concentration time series data of characteristic fault gases such as hydrogen, carbon monoxide, carbon dioxide, and methane in the oil;

[0065] Step S1150, connecting a Hall current sensor in series between the lower leg of the core and the ground to collect timing waveform data of the core grounding current;

[0066] Step S1160 , aggregate the multi-source heterogeneous time series data collected by each sensor, remove outliers, unify timestamps, and form an original monitoring data set.

[0067] Specifically, step S1100 aims to fully perceive the operating status of the box transformer, and arrange multi-source heterogeneous sensors to collect key indicator data reflecting the health level of the box transformer. Temperature is an important indicator for judging the insulation aging and heat dissipation abnormality of the box transformer. The optical fiber temperature measurement technology is used to accurately measure the temperature of the inner wall of the oil tank, the iron core, the winding, the casing and other parts, and timely discover local hot spots. Humidity reflects the sealing performance of the box transformer and the moisture absorption degree of the oil-paper insulation system. The capacitive humidity sensor can continuously monitor the relative humidity changes of the air inside the box transformer. Partial discharge is a typical sign of insulation defects in high-voltage equipment. The ultrasonic method can capture the pulse discharge signal on the insulation surface. Dissolved gas in oil is a gas byproduct of local overheating and discharge inside. The dissolved gas analyzer quantitatively detects characteristic gases in oil through optical, chromatographic and other methods. The core grounding current is the circulating current caused by multi-point grounding of the core. The Hall sensor measures the grounding loop current at the bottom of the core to judge the insulation state of the core. Multi-source heterogeneous sensors provide monitoring data in multiple dimensions such as the internal environment, insulation state, and local defects of the box transformer, forming a three-dimensional perception of the operating state of the box transformer. In order to eliminate the data misalignment caused by different sensor sampling frequencies and clock drift, it is necessary to synchronize the sensor data, align them to the substation clock, and remove obvious outliers to form a time-continuous and reliable original monitoring data set. The original monitoring data set collects data within the time period Δt1.

[0068] Step S1100 obtains original monitoring data that fully reflects the internal and external environment and local status of the box-type transformer by arranging sensors at key locations and multiple physical fields. This is the basis for evaluating the health level of the box-type transformer. Advanced sensing technologies such as optical fiber temperature measurement, capacitive hygrometer, and ultrasonic detection ensure the accuracy and reliability of data acquisition. Dissolved gas analyzers and core grounding current sensors provide sensitive indicator data for internal local defects. By time alignment and anomaly elimination of heterogeneous data, the timeliness and consistency of the data are improved, laying a high-quality data foundation for subsequent analysis. The systematic workflow from sensor layout to data set construction has formed a complete data acquisition system for panoramic perception of the operating status of the box-type transformer, realizing seamless integration from measurement points to data.

[0069] Step S1200, performing data cleaning on the original monitoring data set to obtain a cleaned monitoring data set;

[0070] Specifically, step S1200 aims to improve the quality of monitoring data and provide reliable input for feature extraction. By combining statistical methods and expert knowledge, the normal value range of each monitoring quantity is set, and data points beyond this range are judged as abnormal values; abnormal values ​​are usually introduced by factors such as sensor failure and external interference, which affects the representativeness of the data. Based on statistical methods such as the 3σ principle, the statistical upper and lower limits of normal data can be set, and data points beyond this threshold are regarded as abnormal values ​​and removed. Based on experience, domain experts can also directly identify abnormal data that obviously deviates from the common sense range. Traverse the original monitoring data set, detect and remove abnormal points in the time series data of each monitoring quantity according to the set abnormal value criterion, and obtain the monitoring data after removing the abnormal values. Select the time series data of each monitoring quantity after removing the abnormal values, and use the moving average method to smooth and denoise the monitoring quantities of stable and slow-changing processes such as temperature and humidity, and use the wavelet threshold denoising method to denoise the monitoring quantities of non-stationary pulse processes such as partial discharge and core grounding current, and obtain the smoothed and denoised monitoring data; smoothing and denoising are to eliminate random noise and highlight the real change trend of the measured quantity. Different measurement objects have different signal characteristics, and targeted noise reduction methods are required. Slow-changing processes such as temperature and humidity are generally stable signals with high-frequency random noise. The moving average method suppresses noise by averaging adjacent data points. However, partial discharge pulses and core grounding current spikes are non-stationary processes with transient pulse components. Wavelet decomposition can effectively separate the scale characteristics and noise of the pulses. The wavelet coefficient threshold is set for the noise component, and the soft threshold method filters out the noise while retaining the pulse characteristics.

[0071] The nearest neighbor interpolation method is used to interpolate the smoothed and denoised monitoring data, estimate the missing values ​​in each time series monitoring quantity, and obtain a complete and continuous monitoring data set after cleaning; interpolation completion mainly solves the problem of missing data caused by equipment maintenance, communication interruption, etc. Although linear interpolation is simple, it will introduce unrealistic linear trends. Although spline interpolation is smoother, it is prone to large fluctuations when there are many missing data points. The nearest neighbor interpolation directly estimates the missing values ​​with valid data points adjacent to the missing points, avoiding additional assumptions and is more robust and reliable.

[0072] Step S1200 systematically improves the quality of the original monitoring data from removing outliers, smoothing and denoising to interpolating missing values, which not only eliminates interference factors but also highlights the true characteristics of the measurement signal. It is an indispensable key link in data preprocessing. Outlier discrimination makes full use of statistical characteristics and expert experience, dynamically sets discrimination thresholds, and adapts to the differences in box-type transformer equipment and operating conditions. Classification adopts smoothing and wavelet denoising methods to suppress noise in a targeted manner according to the characteristics of different physical process signals and improve the signal-to-noise ratio. The idea of ​​nearest neighbor interpolation is simple and effective, and it avoids the risk of overfitting through direct estimation of neighborhood effective values. Step S1200 performs a comprehensive cleaning of the original heterogeneous data from anomalies to noise to missing values, greatly improving the data quality and laying a solid data foundation for subsequent feature extraction and mining. High-quality monitoring data is like high-purity industrial raw materials, which is a necessary prerequisite for intelligent monitoring and management.

[0073] Step S1300, extracting key features reflecting the operating status of the box-type transformer from the cleaned monitoring data set to obtain an original feature set;

[0074] Furthermore, if Figure 3 As shown, step S1300 includes:

[0075] Step S1310, based on the time domain analysis method, extracting statistical features reflecting the numerical distribution characteristics of the monitored value, including mean, standard deviation, peak-to-peak value, root mean square value, waveform factor, kurtosis, skewness, etc.;

[0076] Step S1320, based on the frequency domain analysis method, extracting the spectrum characteristics reflecting the frequency distribution characteristics of the monitored quantity, including the main frequency, higher harmonic components, spectrum peak, frequency corresponding to the spectrum peak, frequency band energy, frequency band spectrum entropy, etc.;

[0077] Step S1330, based on the time-frequency domain analysis method, extracting time-frequency features reflecting the time-frequency joint distribution characteristics of the monitoring quantity, including wavelet coefficient energy, Hilbert envelope, S-transform ridge, WVD distribution ridge, etc.;

[0078] Step S1340, based on the morphological analysis method, extracting morphological features reflecting the geometric morphological characteristics of the monitored quantity, including rise time, peak time, fall time, duration, area, asymmetry, etc.;

[0079] Step S1350 , summarizing the various features extracted in steps S1310 - S1340 to obtain an original feature set reflecting the operating status of the box-type transformer.

[0080] Specifically, step S1300 aims to extract the key feature quantity that best characterizes the equipment operation status and fault signs from the time series monitoring data. The time domain characteristics reflect the amplitude statistical characteristics of the signal. The mean and standard deviation characterize the central tendency and discreteness of the signal, which is effective for stable signals such as temperature and humidity. The peak-to-peak value, root mean square value, and waveform factor highlight the severity of the impact vibration, which is suitable for pulse signals such as partial discharge and core grounding current. The frequency domain characteristics reflect the frequency component distribution of the signal. The fundamental frequency component corresponds to the periodic characteristics of the signal, the harmonic component reflects the degree of waveform distortion, and the spectrum entropy measures the degree of chaos of the spectrum. The fault signs of slowly varying signals such as gas in oil and insulation resistance are often manifested as abnormal ultra-low frequency components. Time-frequency joint analysis can reveal the evolution law of the signal frequency component over time, which is suitable for capturing the sign information of intermittent faults and transient processes. Features such as wavelet energy and Hilbert envelope can quantify the degree of local energy aggregation, and WVD and S transforms intuitively show the time-frequency ridge structure. Geometric morphological features focus on the morphological parameters of pulse signals. Rise time, duration, etc. can quantitatively characterize the steepness and oscillation period of pulses. Different types of features describe the external manifestations of the operating status of the box-type transformer from different angles, complementing each other and providing multiple perspectives for fully exploring the inherent laws of monitoring data.

[0081] Step S1300 extracts the features that characterize the state of the box transformer in an all-round way based on time series analysis, signal processing and other methods, and realizes the knowledge extraction from the surface to the inside and from the appearance to the essence. Multi-domain and multi-scale features constitute the basic semantic elements that characterize the health state of the box transformer, reduce the redundancy of the original data, and highlight the key influencing factors of fault prediction and diagnosis. The selection of feature extraction methods follows the principles of pertinence, diversity, and complementarity. The time domain statistical features reveal the overall distribution of the signal, the frequency domain features mine the intrinsic frequency components of the signal, the time and frequency domain features capture the dynamic evolution of the spectrum over time, and the geometric morphological features characterize the local structural details of the pulse signal. The unique features are extracted in different physical fields and different time scales, and a multi-pronged approach is used to provide a three-dimensional and systematic quantitative indicator for a comprehensive grasp of the box transformer state. Redundant features with exactly the same values ​​are removed to further highlight the scene pertinence, avoid excessive invalid calculations, and reduce the difficulty of subsequent modeling. The high-quality feature set is like the essence of extraction, which is the concentration of monitoring big data and conveys the most essential and valuable status information.

[0082] Step S1400 , performing importance analysis on each feature in the original feature set, assigning weights to each feature in the original feature set according to the analysis results, and generating a weighted feature set.

[0083] Furthermore, if Figure 4 As shown, step S1400 includes:

[0084] Step S1410, calculating the Pearson correlation coefficient between each feature in the original feature set and the fault state, selecting features whose Pearson correlation coefficient is higher than a preset correlation threshold as candidate features, and generating a candidate feature subset;

[0085] Step S1420, calculating the information gain of each candidate feature in the candidate feature subset for the fault state, selecting the candidate feature whose information gain is higher than a preset gain threshold as the optimal feature, and generating the optimal feature subset;

[0086] Step S1430, using the Gini index method to calculate the Gini index of each optimal feature in the optimal feature subset, and normalizing the Gini index of each optimal feature as a weight coefficient;

[0087] Step S1440, assigning the weight coefficient obtained in step S1430 to each feature in the optimal feature subset to generate a weighted feature set.

[0088] Specifically, the Pearson correlation coefficient is an indicator to measure the degree of linear correlation between a feature and a fault state. The larger the absolute value of the coefficient, the stronger the correlation, and the greater the contribution of the feature to the discrimination of the fault state. By setting the correlation threshold, a group of features that are highly linearly correlated with the fault state are screened out as candidate features. Information gain reflects the degree to which the feature eliminates the uncertainty of the fault state. When the feature takes different values, the greater the difference in the distribution of the fault state categories, the higher the information gain of the feature, and the stronger the ability to discriminate the fault state. By comparing the information gain of the candidate features, a group of features with the strongest ability to distinguish the fault state are further selected as the optimal feature subset.

[0089] First, based on the optimal feature subset, these features are used as input to train and generate a decision tree model. In the decision tree, each internal node represents classification according to the value of a feature. The Gini index method is used to calculate the importance of each feature, and the Gini index of each feature is normalized as the feature weight; the decision tree model is a classification learner based on a tree structure. The internal nodes represent classification according to the feature values, and the Gini index measures the purity of the node, that is, the degree of confusion of the sample represented by the node in the fault state. During the training process, the decision tree tends to select the feature with the smallest Gini index as the split attribute. At this time, the child node after the split is the purest in the fault state. Therefore, the Gini index corresponding to the feature close to the root in the decision tree structure is often the smallest, and the contribution to the classification result is the greatest. It is the most important feature. The frequency and level of each feature as a split attribute in the decision tree are counted, and converted into the corresponding Gini index, and then normalized to obtain the weight coefficient representing the importance of each feature.

[0090] In the weighted feature set, the higher the weight of the feature, the greater its contribution to fault prediction, and the higher the importance it should be given in subsequent modeling. Fault status refers to various types of faults of box transformer equipment, such as local overheating, local discharge, insulation aging and other faults. Pearson correlation coefficient and information gain are both common knowledge widely used in the fields of data analysis, statistics and machine learning. Therefore, the specific calculation method will not be repeated in this embodiment.

[0091] Step S1400 uses a method that combines correlation, information theory, and tree models to evaluate the importance of each feature from multiple angles, forming a progressive feature selection scheme of "rough screening-fine selection-weighting". First, redundant features are preliminarily screened out through linear correlation with the fault state; then, the most discriminative feature subset is further selected through the information gain criterion; finally, the importance differences of each feature are quantified in combination with the decision tree model to generate scientific and reasonable feature weights. This strategy overcomes the limitations of a single feature selection criterion, and can comprehensively consider multiple factors such as linear correlation, nonlinear discriminability, and model contribution of features. While removing redundancy, selecting subsets, and determining weights, it achieves the complementarity and fusion of feature importance evaluation indicators, so that the resulting weighted feature set can retain the key information of fault diagnosis to the greatest extent.

[0092] Exemplarily, an example is given to illustrate the working principle of this method: Assuming that through Pearson correlation coefficient analysis, it is found that the absolute value of the correlation coefficient between 10 features such as temperature, gas in oil, partial discharge and the fault state is large, exceeding the preset correlation threshold, then they are used as candidate features. Further, by calculating the information gain of the candidate features, it is found that the information gain of temperature and gas in oil is the highest, exceeding the preset gain threshold, then these two features are selected as the optimal feature subset, which contributes the most to the discrimination of the fault state. On this basis, the optimal feature subset is used as the input attribute of the decision tree to train a decision tree. By analyzing the split structure of the decision tree, it is found that the temperature feature appears in the first and second layer split nodes closest to the root node, and the corresponding Gini index is very small, which is the most critical classification basis; while the gas in oil feature appears in the third layer split node, and the Gini index is slightly larger than the temperature feature, which is the second most important classification basis. It can be seen from this that the weight of the temperature feature should be greater than that of the gas in oil. Assuming that the weight coefficient converted from the Gini index of the temperature feature is 0.6 and the weight coefficient of the gas in oil is 0.4, the generated weighted feature set includes the temperature feature with a weight of 0.6 and the gas in oil feature with a weight of 0.4. This example vividly illustrates that by analyzing the correlation between the candidate features and the fault state, the information gain and the contribution in the classification model, the weight coefficients of the two key features of temperature and gas in oil are screened and determined, and they are combined into a weighted feature set, highlighting the importance difference of each feature.

[0093] Step S1400 automatically identifies the most valuable key features for fault prediction by quantitatively analyzing the correlation degree and contribution of each feature to the fault state, and uses weighted coefficients to reasonably distinguish the importance differences of each feature. This not only reduces the feature dimension of subsequent modeling and reduces the training burden, but also highlights the role of decisive features and avoids the interference of secondary features, thereby improving the generalization performance of the fault prediction model. In addition, the qualitative feature screening criteria are quantified into continuous weight coefficients, which has better interpretability. The weighted feature set intuitively presents the importance ranking of each feature, which is convenient for decision-making reference in engineering applications. The objective, automatic and efficient feature selection method gives the features clearer semantics, allowing the data to better serve fault prediction, which is a key step in the knowledge extraction link. The proposed method has strong universality and good adaptability. It can highly intelligently determine the optimal feature subset and weight coefficient for feature data generated under different types of equipment and different working conditions, and has broad application prospects.

[0094] Step S2000: construct a box-type transformer fault prediction model, and obtain a fault risk assessment value Q based on the weighted feature set and the box-type transformer fault prediction model; based on the fault risk assessment value, perform graded warnings on potential faults and generate fault warning information; <Q<1;

[0095] Furthermore, step S2000 includes:

[0096] Step S2100: construct a box-type transformer failure prediction model, and obtain a failure risk assessment value Q based on the weighted feature set and the box-type transformer failure prediction model; <Q<1;

[0097] Specifically, the long short-term memory neural network (LSTM) is selected as the machine learning model for fault prediction. The weighted feature set is divided into a training set, a validation set, and a test set. The training set is used for model parameter learning, the validation set is used for hyperparameter optimization, and the test set is used for model performance evaluation. The grid search method is used to optimize the hyperparameters of the LSTM model, including the number of hidden layers, the number of hidden units, the learning rate, the regularization coefficient, etc., and the optimal hyperparameter combination is determined by combining cross-validation. Based on the optimal hyperparameter combination, the LSTM model is trained, and the prediction accuracy on the training set and the validation set is evaluated. The model is continuously iterated and optimized to obtain a box-type transformer fault prediction model with stable performance. The box-type transformer fault prediction model is used to predict the fault risk of the box-type transformer. The prediction results are mapped to the (0,1) interval to obtain a normalized fault risk assessment value. This assessment value quantitatively describes the risk level of box-type transformer failure. The closer the fault risk assessment value is to 1, the greater the predicted fault risk is; the closer the assessment value is to 0, the smaller the predicted fault risk is. The fault risk assessment value is an important output indicator given by the intelligent monitoring system of the box-type transformer. It quantitatively assesses the risk of failure of the box-type transformer equipment in the current state, which helps managers to intuitively understand the health level of the box-type transformer and take preventive maintenance measures in a timely manner. This assessment value is based on a comprehensive analysis of the multi-source monitoring data and characteristics of the box-type transformer, and can comprehensively consider various factors that affect the box-type transformer failure, and has a high reference value.

[0098] The time period of the weighted feature set is also Δt1, because the weighted feature set is obtained from the original monitoring data set collected within the time period Δt1 after cleaning and feature extraction, so the time range covered by the weighted feature set is also Δt1. The output of the box transformer fault prediction model is the fault risk assessment value in the future. Specifically, assuming that the current time is t, the input of the model is the weighted feature set in the time period [t-Δt1, t], then the model output is the probability of various types of faults occurring in the future time period [t, t+Δt2] (Δt2 is the prediction lead time).

[0099] The selection of LSTM as a prediction model is mainly based on the following considerations: LSTM is a deep learning model for modeling time series data, which is particularly suitable for processing long-sequence, multi-feature time-varying data. The operating status monitoring data of the box transformer is a typical time series data with a long time span and high data dimension. The use of LSTM can fully explore the temporal correlation of historical state changes. LSTM introduces three gated units, namely the input gate, the forget gate, and the output gate, to control the writing, forgetting, and output of information in the long-term and short-term memory units, overcomes the long-term dependency problem and gradient vanishing problem faced by the classic RNN model, and can model sequence dependencies with longer time scales. The weighted fault features of the historical moments are input into the LSTM, and the network adaptively learns the long-term trend and short-term disturbance of the feature changes, integrates the information of each time step at the last moment, and outputs the predicted evaluation value of the fault risk at the next moment.

[0100] Selecting appropriate hyperparameters is crucial to the performance of LSTM networks. Grid search exhaustively enumerates all combinations in the parameter space and selects the best performing group in combination with cross-validation to improve the generalization performance of the model. The number of hidden layers and hidden units determines the complexity of the network. Too few layers make it difficult to learn the inherent laws of the data, while too many layers make it easy to overfit. The learning rate controls the step size of parameter updates. If it is too small, the convergence will be slow, while if it is too large, it will oscillate around the optimal solution. The regularization coefficient controls the intensity of the model complexity penalty to prevent overfitting. Use sufficient historical running data to train the LSTM model, and use the early stopping method to prevent overfitting, that is, stop training if the model performance does not improve after multiple consecutive rounds (epochs). Continuously evaluate the prediction accuracy of the model on the training set and validation set, continuously iterate and optimize the model, and finally obtain a robust fault prediction model.

[0101] The optimized LSTM prediction model can be used to realize the forward-looking judgment of box transformer faults, learn the precursors of faults from massive monitoring data, and timely warn of potential risks, so that operation and maintenance personnel can switch from passive response to active prevention. The LSTM model depicts the gradual process of fault occurrence by modeling the long time series of historical states, avoiding the limitation of the traditional threshold judgment of "dealing with the facts". Grid search combined with cross-validation does not rely on prior assumptions, and adaptively optimizes from the parameter space, reducing the cost of manual trial and error. The training based on big data gives the prediction model stronger robustness and generalization, and improves the extrapolation performance of the model. In short, step S2100 provides a new paradigm of data-driven and adaptive optimization for box transformer fault prediction, grasps the whole process of equipment from health to decline to failure, and makes prediction and warning more forward-looking and explainable.

[0102] Step S2200: Based on the fault risk assessment value, potential faults are graded and warned, and fault warning information is generated.

[0103] Further, step S2200 includes:

[0104] Step S2210: for the fault risk assessment value output by the box-type transformer fault prediction model, three risk thresholds are set: a high risk threshold H, a medium risk threshold M, and a low risk threshold L, where 0 <L<M<H<1;

[0105] Step S2220, when Q>H, a high risk level is warned; when M≤Q≤H, a medium risk level is warned; when L<Q<M, a low risk level is warned; when Q≤L, a no risk level is warned;

[0106] Step S2230, when the warning is a high risk level or a medium risk level, an alarm signal is triggered to generate fault warning information, the fault warning information includes the fault risk level and fault symptom characteristics; the fault symptom characteristics include numerical symptom characteristics and discrete symptom characteristics.

[0107] Specifically, step S2200 aims to discretize the continuous risk assessment value output by the LSTM model into an intuitive risk level, so that operation and maintenance personnel can quickly interpret the warning information and take targeted operation and maintenance control measures in a timely manner. The setting of high, medium and low risk thresholds needs to take into account the sensitivity and false alarm rate of the warning: if the threshold is set too high, the warning will be delayed and the best time for maintenance will be missed; if the threshold is set too low, it will be too sensitive and false alarms will be issued frequently, increasing the maintenance burden. The reasonable division of risk thresholds requires statistical analysis and accumulated experience in a large number of historical fault cases. Tools such as ROC curves and Cost curves can be used to select the best threshold combination that takes into account both recall rate and precision rate.

[0108] The fault risk classification and quantification into four levels: high, medium, low, and no, can provide a fine-grained auxiliary reference for maintenance decisions. High risk generally means that the equipment has shown significant signs of failure and is very likely to have malfunctions in the short term. On-site inspections should be arranged as soon as possible. If necessary, load reduction or even shutdown for maintenance can be considered to avoid accidents caused by the equipment running with "diseases". Medium risk means that although the equipment has no obvious defects, it has deteriorated to a certain extent internally. Its subsequent status changes should be closely monitored, and preventive tests and maintenance should be carried out at the right time. Low risk reflects the overall health of the equipment, and the performance degradation is not obvious, but it is still necessary to continuously monitor its status and accumulate degradation data. No risk means that all indicators of the equipment are within the normal range, and maintenance resources can be prioritized for high-risk equipment.

[0109] Step S2230 aims to intuitively present the abnormal performance of the transformer at the time of warning through the fault symptom characteristics, and provide a basis and clues for further diagnosis by the operation and maintenance personnel. Fault symptom characteristics refer to physical quantities that reflect the typical fault characteristics of the transformer, such as partial discharge, overheating, and insulation breakdown. These physical quantities are at a stable level under normal conditions. Once an abnormal deviation occurs, it often indicates that some potential fault is occurring. According to the fault mechanism and equipment structure, faults of different natures in different parts have different symptoms. By extracting multiple features of multi-sensor data in the time domain, frequency domain, and time-frequency domain, the external manifestations of fault symptoms can be characterized. Taking the core multi-point grounding fault as an example, due to poor core compression, damaged insulating gaskets, etc., the core clamps are connected to the ground in multiple places, forming a closed loop current between the core, iron yoke, and oil tank. Under normal circumstances, the iron core is well insulated and the grounding current is extremely small. Once the iron core is grounded at multiple points, the temperature of the insulating oil in the local area will rise, accelerating the aging of the insulation. At the same time, the alternating magnetic flux flowing through the iron core will induce current in the multi-point grounding loop, causing the iron core grounding current to rise significantly. Local overheating points will cause the insulating paper to decompose, releasing characteristic gases such as hydrogen, causing the dissolved gas in the oil to exceed the standard. The above comprehensive symptoms reflect the typical symptoms of multi-point grounding of the iron core and are the key criteria for identifying this fault.

[0110] The fault warning information is organized in the form of a triple (physical quantity, feature, value), which clearly and intuitively reflects the abnormal performance that triggers the warning. Physical quantities represent the monitored objects, such as gas in oil and winding temperature; features represent statistics extracted from physical quantities, such as mean and peak values; and values ​​give the actual value of the feature at the warning time. The triple form is standardized and unified, easy to store, transmit and visualize. Each triple describes local symptoms, and the combination of triples outlines the overall "portrait" of the fault. Characterizing fault symptoms by multiple features of multiple physical quantities can fully reflect the diversity of fault signs and avoid misjudgment and missed judgment caused by a single feature. Different monitoring quantities verify each other, different features complement each other, and a comprehensive multi-dimensional depiction of the state of the box transformer. Faults are often a combination of multiple symptoms, and it is difficult for a single feature to accurately warn. However, the joint analysis of multiple features can capture mild early signs and prevent the problem from taking shape. The feature value provides a quantitative evaluation, which provides a reference for judging the severity and estimating the development trend. Fault warning information forms a bridge from state perception to auxiliary decision-making. The early warning information intuitively displays the key physical quantities, abnormal characteristics and quantitative indicators that trigger the early warning, providing a basis for further diagnosis of the cause of the fault and formulation of maintenance plans. The early warning information can be pushed to the mobile app or work group for operation and maintenance personnel to review at any time, greatly improving the availability of early warnings.

[0111] For example, if the high-risk warning information contains the triplet (partial discharge pulse, amplitude, 500pC), (partial discharge pulse, repetition rate, 1000 times / s), (gas content, acetylene, 5ppm), etc., the operation and maintenance personnel can preliminarily judge that it is insulation damage caused by partial discharge of the high-voltage bushing. The large amplitude and high repetition rate of the partial discharge pulse indicate that the insulation defect is large in size and highly active. At the same time, the excessive acetylene indicates that the insulator has suffered thermal damage. Based on this, further live testing can be arranged to accurately locate the discharge source and formulate a maintenance plan. Compared with a single risk level, the fault symptom characteristics provide richer auxiliary information, which is conducive to shortening the diagnosis time and improving the maintenance efficiency.

[0112] The numerical symptom characteristics in the fault warning information refer to the indicators reflecting the equipment status extracted from the online monitoring data and expressed as continuous numerical values. For example:

[0113] Dissolved gas concentration in oil: reflects the content of combustible gas in the oil, such as hydrogen concentration is 150ppm, carbon monoxide concentration is 50ppm, etc. The higher the value, the more serious the overheating or discharge phenomenon in the oil.

[0114] Core grounding current: reflects the amplitude of circulating current caused by multi-point grounding of the core, such as 0.2A, 0.5A, etc. The larger the value, the more serious the core insulation defect.

[0115] Winding temperature: reflects the real-time temperature of various parts of the transformer, such as 65℃, 85℃, etc. The higher the value, the more serious the winding overheating.

[0116] Insulation resistance: reflects the resistance value of high-voltage insulation, such as 50MΩ, 100MΩ, etc. The lower the value, the more obvious the decline in insulation performance.

[0117] Discrete symptom features in fault warning information refer to indicators that reflect the status of equipment and are extracted from online monitoring data and expressed as discrete values ​​or symbols. For example:

[0118] Partial discharge frequency: reflects the activity level of partial discharge, and can be "frequent", "intermittent", "sparse", etc. The more frequent the discharge, the more serious the insulation defect.

[0119] Discharge type: reflects the type characteristics of local discharge, and can be "pinpoint discharge", "floating potential discharge", "surface discharge", etc. Different types correspond to different defects.

[0120] Oil flow state: reflects the amount of bubbles entrained in the oil. Possible values ​​include "bubbles", "aerosols", "laminar flow", etc. The more bubbles there are, the greater the possibility of overheating or discharge.

[0121] Dancing discharge: reflects the phenomenon that the insulating oil oscillates violently under the action of the electric field. The possible values ​​are "violent", "mild", "none", etc. The more violent the dancing, the higher the risk of liquid insulation failure.

[0122] In summary, step S2200 visualizes the risk assessment value output by LSTM through multiple risk thresholds, internalizes the industry's many years of operation and maintenance experience summary into early warning criteria, and forms a "data-information-knowledge" argumentation and decision-making chain. The intelligent perception of the model is organically integrated with the expert's experience and judgment, so as to achieve machine early warning, human-machine diagnosis, and expert argumentation, make full use of the perspective of digital twin technology, but not completely rely on it, so as to make professional and prudent risk assessment and maintenance decisions, and achieve a balance between efficiency and reliability. The hierarchical early warning provides a digital gripper for differentiated and refined state maintenance strategies, and promotes the transformation of maintenance from post-fault handling to proactive mode of pre-prevention, real-time monitoring, and precise analysis. The application of fault symptom characteristics makes the early warning explainable, traceable, and responsive. The path from data to decision-making is clear and transparent, avoiding blind emergency response. Multi-dimensional feature characterization makes the early warning criteria objective, quantitative, and comprehensive, reducing subjective assumptions. The early warning information is concise and clear, highlighting the key points, which is easy for non-professionals to understand and helps scientific policy implementation.

[0123] Step S3000, construct a box-type transformer fault diagnosis knowledge graph, and when warning of a high risk level or a medium risk level, reason and trace the fault warning information in combination with the box-type transformer fault diagnosis knowledge graph.

[0124] Furthermore, step S3000 includes:

[0125] Step S3100, constructing a knowledge graph for box-type transformer fault diagnosis;

[0126] Furthermore, if Figure 5 As shown, step S3100 includes:

[0127] Step S3110, defining core concepts to form a concept layer; the core concepts include the concept of box-type transformer components, the concept of materials, and the concept of faults;

[0128] Step S3120, extracting semantic relations between core concepts to form a relational layer; the semantic relations include causal relations, concurrent relations and co-location relations; the causal relations are presented as causal chains, the co-location relations are presented as co-location relation families, and the concurrent relations are presented as concurrent probability networks;

[0129] Step S3130, using the core concepts of the concept layer as nodes and the semantic relationships of the relationship layer as edges to construct a topological structure of the box-type transformer fault diagnosis knowledge graph;

[0130] Step S3140, mining typical features of each fault concept, marking them as attribute features, and filling them into the topological structure of the box-type transformer fault diagnosis knowledge graph as the attributes of the fault nodes corresponding to the fault concepts; the attribute features of the fault concepts include numerical attribute features and discrete attribute features.

[0131] Specifically, step S3100 aims to build a structured and semantic knowledge base to support the fault diagnosis reasoning of box transformers. The knowledge graph is a knowledge base that represents domain knowledge in a graphical way, which is composed of concept nodes and relationship edges. The construction process of the box transformer fault diagnosis knowledge graph is divided into four steps: concept extraction, relationship extraction, topology construction, and attribute filling. Step S3110 is based on structured data sources such as power equipment standards, specifications, and patent documents, and adopts an ontology construction method to define core concepts such as box transformer components, materials, and faults to form the concept layer of the ontology knowledge base. Ontology is a formal description of domain knowledge. It uses a terminology system with clear definitions and rigorous logic to standardize the connotation and extension of concepts and construct a category tree to describe the hierarchical relationship of concepts. The concept of box transformer components refers to the various components that constitute the box transformer, such as iron core, winding, oil tank, bushing, respirator, etc., which are the carriers of faults. The concept of materials refers to the various materials that make up the box transformer components, such as silicon steel sheets, copper wires, insulating paper, transformer oil, etc. The performance of the materials directly affects the health status of the components. The fault concept covers various types of faults that may occur during the operation of the box-type transformer, such as multi-point grounding of the core, short circuit between winding turns, partial discharge of the bushing, etc. The fault concept is directly related to the concepts of components and materials.

[0132] Step S3120 further extracts semantic relationships between concepts based on the concept layer. Based on unstructured data sources such as expert experience, fault cases, and online questions and answers, natural language processing techniques such as conditional random fields are used to extract semantic relationships such as causal, concurrent, and co-location between concepts to form a relational layer of the ontology knowledge base. The relational layer depicts the intricate horizontal connections between concepts, revealing the causal mechanisms, propagation paths, and association patterns of faults. Causal relationships depict the triggering and causing mechanisms between faults, have clear directionality, and are often presented in the form of multiple causal chains. Co-location relationships reveal non-causal relationships such as parallel, progressive, and inclusive relationships between faults, reflecting different granularities and levels of knowledge. Concurrency relationships describe the strength of association between faults from a probability perspective. The higher the probability of concurrency between fault nodes, the stronger the degree of association.

[0133] The co-location family is essentially a data structure that represents the co-location relationship between concepts in the knowledge graph. In a co-location family, co-location nodes are clustered together, indicating that they belong to the same parent concept at the ontology level, reflecting their similarities in position, function, etc. The concurrent probability network is a probability graph that represents the concurrent relationship between concepts in the knowledge graph. In the concurrent probability network, each edge is accompanied by a probability value, which indicates the probability that the faults represented by the two connected nodes occur at the same time. The larger the probability value, the stronger the concurrency between the two events.

[0134] For example, taking the oil-immersed box transformer as an example, there is a causal chain between the core multi-point grounding fault and the winding inter-turn short circuit fault: core multi-point grounding → core local overheating → insulation oil deterioration → insulation strength reduction → winding inter-turn short circuit. This causal chain clearly depicts the causal evolution mechanism between the two faults. In terms of the same-position relationship family, although there is no obvious causal logic between the three, the excessive dissolved gas in the oil, the local discharge of the high-voltage bushing, and the abnormal core grounding current, they reflect the different manifestations of the insulation aging of the box transformer to a certain extent and have a category relationship. The concurrent probability network measures the correlation strength between faults in the form of probability values. For example, the concurrent probability of core multi-point grounding and winding inter-turn short circuit is as high as 0.8, while the concurrent probability of core multi-point grounding and porcelain bushing cracking is only 0.1, so the former's correlation characteristics should be focused on.

[0135] Step S3130 uses the core concepts of the concept layer as nodes and the semantic relationships of the relationship layer as edges to construct the topological structure of the knowledge graph for fault diagnosis of the box transformer. Among them, the fault concept, component concept, and material concept constitute three types of nodes, corresponding to the diagnosis target, fault carrier, and influencing factor, respectively. The causal chain is reflected as a directed edge, with the start node pointing to the result node; the co-location relationship is reflected as an undirected edge, connecting concepts of the same level or related concepts; the concurrent probability is reflected as a weighted edge, and the weight corresponds to the concurrent probability value. The topological structure intuitively shows the hierarchical distribution and connection of the concept nodes, and specifies the rules for fault diagnosis reasoning. For example, the central node of the knowledge graph is the key fault mode of partial discharge. The left branch revolves around the cause of partial discharge, which can be gradually traced back to the root causes such as insulation aging and moisture; the right branch reveals the consequences of partial discharge, including insulation breakdown, equipment fire, etc.; the upper branch classifies partial discharge according to the fault location, forming a co-location relationship family; the lower branch lists the concurrent faults of partial discharge such as multi-point grounding of the iron core. This form of graph organization is intuitive and clear, which is convenient for quickly retrieving the causes and consequences of the fault.

[0136] Step S3140 mines the typical features of each fault concept and annotates them as attributes to the nodes to complete the attribute filling of the knowledge graph. Fault characteristics include both cause phenomena and result representations. The cause characteristics of partial discharge include decreased insulation resistance, increased dielectric loss factor, etc., and the result characteristics include pulse current, ultrasonic signals, etc. The partial discharge of high-frequency cabinets is accompanied by high-frequency current pulses, and the neutral point grounding current flickers frequently; the core grounding current increases significantly when the core is grounded at multiple points; the short circuit between winding turns causes the coil current to change sharply; the excessive dissolved gas in the oil is a sensitive signal of local overheating. These symptomatic features give the knowledge graph real-time status perception capabilities, which can be used to make early warnings of faults.

[0137] Typical features can be divided into numerical quantitative indicators, such as winding temperature, partial discharge amplitude, core current, etc., and can also be divided into discrete semantic descriptions, such as discharge type, oil color spectrum, vibration form, etc. Attribute features further clarify the external manifestations of each fault concept, making it easier to quickly determine the fault type from online monitoring data.

[0138] For example, taking the winding turn-to-turn short circuit fault as an example, its numerical attribute features include winding temperature exceeding 95°C, top oil temperature exceeding 80°C, FeC hydrocarbon concentration exceeding 15μL / L, partial discharge amplitude between 10 and 30dBm, etc.; discrete attribute features include high acetylene content in oil, local overheating of the transformer body, and discharge sound can be heard at the bushing, etc. These attribute features replace the abstract fault concept and give the knowledge graph a more intuitive and measurable semantic connotation.

[0139] In summary, step S3100 adopts a bottom-up, point-to-surface construction strategy. After concept extraction, relationship extraction, topology construction, and attribute filling, a knowledge graph of box transformer fault diagnosis with clear semantics and complete structure is finally formed. The knowledge graph is like a roadmap for fault diagnosis, which clarifies the internal connection of fault concepts, defines the external representation of faults, and points out the direction for fault diagnosis reasoning. The causal chain reveals the logic of fault occurrence, the co-location relationship reveals the hierarchical structure of faults, the concurrent probability reveals the correlation strength of faults, and the attribute characteristics reveal the measurable elements of faults. In the graph, nodes are centered on concepts and edges are linked by relationships, forming a semantically fused and logically self-consistent knowledge network. This structured and semantic knowledge representation form transforms scattered and unstructured fault diagnosis experience into interconnected and computable formal knowledge, laying a knowledge foundation for the construction of intelligent diagnosis systems. Reasoning based on knowledge graphs can use the connectivity of graphs and the diversity of relationships to flexibly deal with the diagnosis of multi-cause and multi-effect combined faults, and has strong adaptability. In short, the knowledge graph of box-type transformer fault diagnosis is a comprehensive and systematic knowledge base, which can be regarded as the brain of intelligent diagnosis. It is of great value in improving diagnostic efficiency, standardizing diagnostic processes, and accumulating diagnostic knowledge.

[0140] Step S3200, calculating the comprehensive matching degree between each fault symptom feature in the fault warning information and each fault node attribute feature in the box-type transformer fault diagnosis knowledge graph, and selecting the top N nodes with the highest comprehensive matching degree as candidate fault nodes;

[0141] Furthermore, if Figure 6 As shown, step S3200 includes:

[0142] Step S3210, calculating the similarity DE between each numerical symptom feature in the fault warning information and the numerical attribute feature of each fault node in the box-type transformer fault diagnosis knowledge graph;

[0143] ;

[0144] in:

[0145] Indicates Fault warning information;

[0146] represents the first Faulty nodes;

[0147] express and similarity;

[0148] express No. Numerical symptom feature values;

[0149] express No. Numeric attribute feature values;

[0150] Indicates The value range of the feature is calculated as follows: ; and They represent all the fault nodes in the knowledge graph of box-type transformer fault diagnosis in the first The maximum and minimum values ​​of the features;

[0151] Indicates The weight coefficient of each feature is obtained by normalizing the Gini index in step S1400;

[0152] Indicates the total number of numeric symptom features.

[0153] This formula calculates the degree of difference between the numerical features of the fault warning information and each fault node in the knowledge graph. For each feature dimension, the absolute value of the difference between the two feature values ​​is divided by the value range of the feature to measure the local difference, and the weighted average of the local differences of all feature dimensions is obtained to obtain the comprehensive difference DE. The feature difference is divided by the value range to achieve normalization, eliminate the dimension effect, and make the difference values ​​of different features comparable. The smaller the difference value, the closer the fault warning information is to the numerical features of the fault node, and the greater the similarity. The introduction of reflects the difference in importance of different features. Features with greater weights have a higher proportion in similarity calculation. The weighted average method comprehensively considers the contribution of each feature to obtain the overall similarity. The beneficial effect of this formula is to evaluate the proximity between fault warning information and candidate faults in the knowledge graph with a unified quantitative indicator, providing a comparable numerical basis for subsequent semantic matching. The numerical feature similarity is inversely proportional to the matching degree of the fault node. The smaller the DE, the higher the matching degree, and the more likely the fault node is to be the cause of the current fault warning.

[0154] If some features of the fault warning information are significantly different from the corresponding features of the knowledge graph fault node, that is, If the value is large, their DE values ​​will become larger and the similarity will decrease. On the contrary, if the eigenvalues ​​of the two are close, that is, If the value is smaller, the DE value will be smaller and the similarity will increase. In short, this formula quantitatively describes the similarity between the fault warning information and the fault node of the knowledge graph by calculating the weighted normalized feature difference, which is a key link in the semantic matching process. The formula makes full use of the numerical feature information, takes into account the difference in feature importance, is concise and efficient, and has strong interpretability.

[0155] Step S3220, calculating the similarity DB between each discrete symptom feature in the fault warning information and the discrete attribute feature of each fault node in the box-type transformer fault diagnosis knowledge graph based on the weighted Jaccard coefficient;

[0156] Step S3230, performing weighted average of the similarity DE and the similarity DB to obtain the comprehensive matching degree MP between the fault warning information and each fault node;

[0157] Step S3240 , sorting from high to low according to the comprehensive matching degree MP, and selecting the first N fault nodes with the highest matching degree as candidate fault nodes.

[0158] Specifically, step S3200 adopts the idea of ​​divide and conquer, adopts different similarity metrics for different types of features, and evaluates the matching degree between the warning information and the graph nodes from local to overall. The key innovation is to introduce feature weights to highlight the influence of key features. The similarity of discrete features is calculated with the help of weighted Jaccard coefficient. The Jaccard coefficient measures the degree of overlap between two sets, which is equal to the number of elements in the intersection divided by the number of elements in the union, and its value range is [0,1]. The similarity of each feature is weighted averaged to obtain a comprehensive matching degree, that is, from the division (local similarity of each feature) to the combination (overall similarity of the node). The matching metric quantifies the overall proximity between the warning information and the graph nodes in multiple feature dimensions. The higher the matching degree, the closer the association between the two. In order to further narrow the scope of investigation, the first N nodes with the highest matching degree are selected as the key suspected objects, that is, candidate faults. The selection of the candidate number N needs to balance the recall rate and the precision rate. The larger N is, the higher the recall rate is, but the precision rate may decrease, and the amount of calculation will increase accordingly. It can be flexibly configured according to the timeliness requirements of online diagnosis.

[0159] In summary, step S3200 proposes a comprehensive similarity evaluation method that integrates numerical features, discrete features, unifies feature weights, and aggregates in layers and steps. Numerical features focus on comparing the degree of numerical deviation, while discrete features focus on comparing the overlap of values. On the basis of local feature similarity, the overall similarity is evaluated by the weighted average method, and the most suspicious candidate faults are dynamically screened out in combination with the top N sorting idea. This method can compare the correlation strength between fault warning information and knowledge graph nodes from multiple perspectives combining the overall and local aspects. The weighting mechanism and top N strategy improve the pertinence and efficiency of matching. The multi-granularity and multi-level similarity calculation model can organically unify heterogeneous features of structured and unstructured, continuous and discrete types, and realize the fusion correlation analysis of cross-modal data. It effectively integrates quantitative data and qualitative knowledge, cleverly combines fault diagnosis experience with online monitoring data, and greatly reduces the subsequent reasoning search space.

[0160] Step S3300, extracting one-hop neighbor nodes of the candidate fault node in the box-type transformer fault diagnosis knowledge graph to form a semantic subgraph of the candidate fault;

[0161] Specifically, step S3300 aims to extract the local subgraph that is most closely related to the candidate fault from the entire box-type transformer fault diagnosis knowledge graph. The box-type transformer fault diagnosis knowledge graph is a huge heterogeneous information network, containing hundreds of nodes and multiple types of edges. Direct analysis on the entire graph will introduce a large amount of irrelevant information, affecting the pertinence and efficiency of diagnosis. Therefore, it is necessary to extract a semantically highly relevant subgraph as the key analysis object around the candidate fault. One-hop neighbor nodes refer to nodes that are directly connected to the candidate fault node in the graph. They represent the semantically directly related concepts of the candidate fault and are the contextual semantic environment that needs to be focused on to understand the candidate fault. Taking the candidate fault node as the central node, extracting all its one-hop neighbor nodes, plus the edges between them, constitutes a microcosm of the specific domain knowledge around the candidate fault, that is, the semantic subgraph of the candidate fault. The semantic subgraph reveals important semantic information such as the cause context, location clues, severity, and associated risks of the candidate fault, and is an indispensable knowledge basis for fault tracing analysis.

[0162] For example, it is assumed that two candidate faults are generated based on the real-time online evaluation of the operating conditions of the box transformer, namely, partial discharge of the winding and poor grounding of the core. On this basis, step S3300 extracts one-hop neighbor nodes from the knowledge graph for the two candidate faults respectively. Taking partial discharge of the winding as an example, its one-hop neighbors may include nodes such as aging of the winding insulation, high-order harmonics, and increased acetylene content in the oil. These nodes are directly related to partial discharge of the winding in terms of business semantics, implying knowledge about the causes, detection methods, and hazards of partial discharge. Similarly, the one-hop neighbors of poor core grounding may include nodes such as multi-point grounding, uneven core stacking, and loose grounding bolts, reflecting the background knowledge of the causes, mechanisms, and consequences of core grounding faults. By organizing the one-hop neighbor nodes and their edges of the two candidate faults respectively, two fault-specific semantic subgraphs are obtained, providing a focused knowledge index for subsequent diagnosis.

[0163] This automatic indexing capability of domain-specific knowledge highlights the semantic unification advantage of the knowledge graph. In traditional diagnostic systems, knowledge is often scattered in databases, documents, and rule bases with different structures, making it difficult to quickly retrieve and analyze semantic associations of knowledge in specific scenarios. The knowledge graph is stored in the form of a graph database, with interconnected nodes. It can easily extract the local subgraph of any node in the semantic network, and realize the rapid positioning and semantic indexing of diagnostic knowledge. Combining the knowledge graph with real-time fault cases and dynamically generating case-specific semantic subgraphs avoids the over-generalization of the knowledge base without losing scalability, and strikes a balance between the pertinence and applicability of diagnosis. This not only greatly narrows the scope of knowledge sources for subsequent analysis, but also lays the foundation for semantic diagnostic reasoning by constructing a specific semantic network that integrates cases and knowledge.

[0164] Step S3400: Generate a causal chain in a semantic subgraph based on the causal chain of the box-type transformer fault diagnosis knowledge graph, and trace the cause of the candidate fault by analyzing the causal chain in the semantic subgraph;

[0165] Specifically, step S3400 traces the cause of the candidate fault layer by layer based on the causal chain contained in the semantic subgraph. The occurrence of box transformer failure is often the result of the chain reaction of multiple factors. Clarifying the causal logic chain of the fault and reasoning layer by layer to the root cause of the fault is the key to diagnosis. The knowledge graph uses causal relationships to link the fault symptom nodes with the cause nodes, and depicts the causal reasoning path of the fault. The semantic subgraph refines the local causal network of the candidate fault, making the traceability analysis more focused and efficient. Therefore, by searching for the causal chain path starting from the candidate fault node on the semantic subgraph, the cause mechanism of the fault can be traced back layer by layer.

[0166] For example, for the semantic subgraph of winding partial discharge, starting from the winding partial discharge node, there are directed causal edges leading to neighboring nodes such as winding insulation aging and higher harmonics. This reveals that winding insulation aging and higher harmonics are the direct causes of winding partial discharge. In order to avoid stopping here, it is necessary to continue to trace back along the causal edge. Upstream of the winding insulation aging node, there may be nodes such as multiple load fluctuations and moisture intrusion, and there is also a causal relationship between these nodes and winding insulation aging. It can be seen that frequent load fluctuations and external moisture intrusion are the deep-seated reasons for insulation aging, which in turn causes partial discharge. Similarly, by tracing the high-order harmonic nodes through the causal chain, higher-level inducements such as improper transformer operating parameters and core saturation can be found. Similar causal tracing is performed on the semantic subgraph of poor core grounding. Multi-point grounding can be traced back to loose bolts and cracked bushings, and uneven grounding can be traced back to the root causes of core stacking deviation and silicon steel sheet oxidation. By recursively tracing along the causal edges and piecing them together into a causal chain that reflects the cause of the fault, a global causal picture of the occurrence mechanism of the candidate fault can be formed.

[0167] This causal chain analysis mode, which combines knowledge graphs with case semantic subgraphs, can fully utilize the semantic connectivity of graphs and form a powerful intelligent analysis capability for fault tracing. Knowledge is abstracted into nodes and edges in the graph, making the logic of complex cross-domain fault causes clear and orderly.

[0168] Step S3500: Based on the co-location relationship family and concurrent probability network of the box-type transformer fault diagnosis knowledge graph, the number of co-location nodes and the probability of concurrent node occurrence in the semantic subgraph are counted to determine the location, severity and associated faults of the candidate fault;

[0169] Specifically, step S3500 aims to use the co-location relationship and concurrent probability in the knowledge graph to infer the key attributes of the candidate fault. The co-location relationship family refers to a cluster of nodes in the knowledge graph that represent the same or similar concepts. For example, "partial discharge", "surface discharge" and "suspended discharge" are all co-location nodes of discharge phenomena. The number of co-location nodes of each candidate fault node in the statistical semantic subgraph can be used to determine the location of the fault. The candidate fault with the largest number of co-location nodes has a relatively clear location and the greatest possibility.

[0170] Exemplarily, continuing with the example in steps S3300-S3400, the co-location nodes of "increased core grounding current" in the semantic subgraph include "multiple-point grounding of the core", "aging of the core insulation", and "loosening of the core bolts", indicating that the fault may be located in the transformer core. The co-location nodes of "overheating of the high-voltage bushing" include "discharge of porcelain bushing" and "overheating of bushing joint", and the location points to the high-voltage bushing. The co-location nodes of "partial discharge of winding" include "short circuit between winding turns" and "deformation of winding", and the location points to the transformer winding. It can be preliminarily determined by the number of co-location nodes that the fault involves multiple parts of the core, high-voltage bushing, and winding, but the possibility of core failure is higher.

[0171] The concurrent probability network is a directed probability graph that represents the concurrent relationship between nodes in the knowledge graph. For example, if node A points to node B, it means that when fault A occurs, fault B will inevitably be triggered, and the probability of occurrence is marked. By counting the probability of occurrence of candidate fault nodes in the associated nodes in the concurrent probability network, the severity of the fault and the associated faults can be determined. Candidate faults associated with nodes with high occurrence probability are often serious faults and may trigger a series of derivative and associated faults.

[0172] For example, in the concurrent probability network of the semantic subgraph in the above example, "multiple-point grounding of the core" points to "core insulation breakdown", the latter of which has a probability of 0.8 and is a serious fault. "Loosening of the core bolts" points to "deformation of the core frame", the latter of which has a probability of 0.6 and is also a serious derivative fault. "Overheating of the high-voltage bushing" points to "bursting of the porcelain bushing", which has a probability of only 0.1 and the consequences of the fault are relatively minor. Although partial discharge of the winding can cause "short circuit between winding turns", the probability of occurrence is only 0.2, which is not enough to cause serious faults. Comprehensive analysis shows that the core fault represented by "multiple-point grounding of the core" is the most serious and must be taken seriously, followed by partial overheating of the high-voltage bushing, and winding discharge is relatively minor.

[0173] The analysis method of step S3500 makes full use of the semantic association information of the knowledge graph, infers the fault location through the statistics of the co-located nodes, and infers the fault level through the concurrent probability network, giving the semantic intelligence of the knowledge graph to the candidate fault, so that it has semantic attributes such as location, severity, and correlation, which greatly releases the application value of the graph knowledge. In particular, the concurrent relationship and the probability of occurrence are combined, and the impact of the fault is quantitatively evaluated through the conduction effect of the causal chain, which avoids subjective experience judgment and improves the objectivity of fault diagnosis. A vivid analogy is to compare the fault to a disease, such as core fault is like heart disease, high-voltage bushing fault is like stomach disease, and winding fault is like skin disease. Through the reasoning analysis of the knowledge graph, it can be preliminarily concluded that heart disease is the most critical, stomach disease needs timely treatment, and skin disease is currently not a big deal, which provides an objective basis for subsequent maintenance strategies.

[0174] Step S3600: Comprehensively consider the cause, location, severity and associated faults of the candidate faults, and output a fault diagnosis report based on the box-type transformer fault diagnosis knowledge graph.

[0175] Specifically, step S3600, based on the above analysis, integrates all semantic attributes of the candidate faults to form a comprehensive and targeted fault diagnosis report. The report contains key information such as the cause of the fault, the location of the fault, the severity level, and the associated risks, and answers a series of diagnostic questions such as "what fault occurred", "where did the fault occur", "how serious is the fault", and "what other problems will the fault cause", forming a panoramic and three-dimensional explanation of the fault. The diagnostic report adopts a causal logic framework, explains the development of the fault in the form of a causal chain, and reveals the deep mechanism, which is easy for maintenance personnel to intuitively understand and accept.

[0176] For example, the comprehensive diagnostic report for the above case is as follows: There are multiple local faults in the transformer body, and the root cause is the circulating current caused by multi-point grounding of the core. The long-term impact of large current causes the core insulation to gradually age, and the mechanical stress of loose bolts may eventually cause serious faults such as core breakdown and deformation, which must be taken seriously. The problem of local overheating of the high-voltage bushing should not be ignored. If left unchecked, the porcelain insulation may be broken down, and timely inspections must be carried out to eliminate hidden dangers. The partial discharge of the winding is not serious at present, and it will not induce inter-turn short circuits for the time being, so it can be included in the medium- and long-term maintenance plan. It is recommended to check for defects in multi-point grounding of the core, replace aging insulation, and tighten bolts; perform infrared temperature measurement on the high-voltage bushing to find hot spots and eliminate safety hazards; perform ultrasonic testing on the winding to promptly discover the source of partial discharge and formulate a maintenance plan.

[0177] Step S3000 realizes in-depth diagnosis and decision support for high-risk and medium-risk early warnings by constructing a knowledge graph for transformer substation fault diagnosis and conducting reasoning and tracing based on fault early warning information. Its beneficial effects are as follows: First, it efficiently matches and locates faults. By calculating the similarity between the fault symptom features and the fault node attribute features in the knowledge graph, it quickly screens the most likely candidate faults, avoiding the inefficiency of full-graph search and accurately matching the fault types. Second, it traces and analyzes the fault causes. It constructs a local semantic subgraph and traces back along the causal chain in reverse to reveal the deep causes of the faults and comprehensively sort out different causal paths. In addition, by inferring the risks of associated faults, it analyzes the co-located nodes and concurrent nodes in the semantic subgraph, estimates the occurrence probability of potential associated faults, and reveals the risk of accident expansion. The association analysis of the knowledge graph starts from the point and spreads to the surface, effectively preventing problems before they occur. Third, it assists in judging the severity of the faults. By combining factors such as the length of the causal chain, co-located nodes, and concurrent probability, it makes inferences about the severity of the candidate faults and makes scientific judgments by fusing evidence from multiple dimensions. Finally, the output diagnostic report covers the fault causes, locations, severities, and associated risks, providing clear decision support for fault handling. Through these steps, the knowledge graph not only associates discrete early warning information but also reveals the full picture of the faults. In particular, it conducts in-depth diagnosis of high-risk and medium-risk early warnings, and the reasoning with both breadth and depth provides practical decision support for emergency response. Compared with shallow fault code matching, the reasoning based on multiple semantics such as causality, co-location, and concurrency is more in line with human thinking, and the conclusions are interpretable, facilitating engineers to quickly understand and apply them.

[0178] Embodiment 2

[0179] On the basis of Embodiment 1, this embodiment provides a new energy transformer substation intelligent panoramic monitoring and management system, as Figure 7 shown, including:

[0180] Data processing module: It is used to obtain the original monitoring data set of the target transformer substation, clean the original monitoring data set to obtain the cleaned monitoring data set, extract the key features reflecting the operation status of the transformer substation from the cleaned monitoring data set to obtain the original feature set; conduct importance analysis on each feature in the original feature set, and assign weights to each feature in the original feature set according to the analysis results to generate a weighted feature set;

[0181] Fault prediction module: It is used to construct a transformer substation fault prediction model, and obtain a fault risk assessment value Q according to the weighted feature set and the transformer substation fault prediction model; based on the fault risk assessment value, conduct hierarchical early warning on potential faults to generate fault early warning information; 0 < Q < 1; The hierarchical early warning of potential faults includes early warning of high-risk level, early warning of medium-risk level, early warning of low-risk level, and early warning of no-risk level;

[0182] Fault tracing module: used to construct a knowledge graph for fault diagnosis of box-type transformers. When warning of a high risk level or a medium risk level, the fault warning information is reasoned and traced in combination with the knowledge graph for fault diagnosis of box-type transformers.

[0183] In the data processing module, the step of obtaining the original monitoring data set of the target box-type transformer includes:

[0184] Step S1110, arranging optical fiber temperature sensors at key locations such as the inner wall of the oil tank, the iron core, the winding, and the casing of the box-type transformer body to collect the time series variation data of the internal temperature of the box-type transformer;

[0185] Step S1120, arranging a humidity sensor in the internal space of the box-type transformer body to collect time-series variation data of the air humidity inside the box-type transformer;

[0186] Step S1130, arranging ultrasonic partial discharge sensors on the surfaces of insulating parts such as high-voltage bushings and insulating partitions to collect timing pulse signals of partial discharge on the surfaces of the insulating parts;

[0187] Step S1140, a dissolved gas analyzer is built into the transformer oil pillow to collect concentration time series data of characteristic fault gases such as hydrogen, carbon monoxide, carbon dioxide, and methane in the oil;

[0188] Step S1150, connecting a Hall current sensor in series between the lower leg of the core and the ground to collect timing waveform data of the core grounding current;

[0189] Step S1160 , aggregate the multi-source heterogeneous time series data collected by each sensor, remove outliers, unify timestamps, and form an original monitoring data set.

[0190] In the data processing module, the key features reflecting the operating status of the box-type transformer are extracted from the cleaned monitoring data set, and the original feature set obtained includes:

[0191] Step S1310, based on the time domain analysis method, extracting statistical features reflecting the numerical distribution characteristics of the monitored value, including mean, standard deviation, peak-to-peak value, root mean square value, waveform factor, kurtosis, skewness, etc.;

[0192] Step S1320, based on the frequency domain analysis method, extracting the spectrum characteristics reflecting the frequency distribution characteristics of the monitored quantity, including the main frequency, higher harmonic components, spectrum peak, frequency corresponding to the spectrum peak, frequency band energy, frequency band spectrum entropy, etc.;

[0193] Step S1330, based on the time-frequency domain analysis method, extracting time-frequency features reflecting the time-frequency joint distribution characteristics of the monitoring quantity, including wavelet coefficient energy, Hilbert envelope, S-transform ridge, WVD distribution ridge, etc.;

[0194] Step S1340, based on the morphological analysis method, extracting morphological features reflecting the geometric morphological characteristics of the monitored quantity, including rise time, peak time, fall time, duration, area, asymmetry, etc.;

[0195] Step S1350 , summarizing the various features extracted in steps S1310 - S1340 to obtain an original feature set reflecting the operating status of the box-type transformer.

[0196] In the data processing module, the step of assigning weights to the features in the original feature set to generate a weighted feature set includes:

[0197] Step S1410, calculating the Pearson correlation coefficient between each feature in the original feature set and the fault state, selecting features whose Pearson correlation coefficient is higher than a preset correlation threshold as candidate features, and generating a candidate feature subset;

[0198] Step S1420, calculating the information gain of each candidate feature in the candidate feature subset for the fault state, selecting the candidate feature whose information gain is higher than a preset gain threshold as the optimal feature, and generating the optimal feature subset;

[0199] Step S1430, using the Gini index method to calculate the Gini index of each optimal feature in the optimal feature subset, and normalizing the Gini index of each optimal feature as a weight coefficient;

[0200] Step S1440, assigning the weight coefficient obtained in step S1430 to each feature in the optimal feature subset to generate a weighted feature set.

[0201] In the fault prediction module, the potential faults are graded and warned based on the fault risk assessment value, and the fault warning information is generated, including:

[0202] Step S2210: for the fault risk assessment value output by the box-type transformer fault prediction model, three risk thresholds are set: a high risk threshold H, a medium risk threshold M, and a low risk threshold L, where 0 <L<M<H<1;

[0203] Step S2220, when Q>H, a high risk level is warned; when M≤Q≤H, a medium risk level is warned; when L<Q<M, a low risk level is warned; when Q≤L, a no risk level is warned;

[0204] Step S2230, when the warning is a high risk level or a medium risk level, an alarm signal is triggered to generate fault warning information, the fault warning information includes the fault risk level and fault symptom characteristics; the fault symptom characteristics include numerical symptom characteristics and discrete symptom characteristics.

[0205] In the fault tracing module, the construction of the box-type transformer fault diagnosis knowledge graph includes:

[0206] Step S3110, defining core concepts to form a concept layer; the core concepts include the concept of box-type transformer components, the concept of materials, and the concept of faults;

[0207] Step S3120, extracting semantic relations between core concepts to form a relational layer; the semantic relations include causal relations, concurrent relations and co-location relations; the causal relations are presented as causal chains, the co-location relations are presented as co-location relation families, and the concurrent relations are presented as concurrent probability networks;

[0208] Step S3130, using the core concepts of the concept layer as nodes and the semantic relationships of the relationship layer as edges to construct a topological structure of the box-type transformer fault diagnosis knowledge graph;

[0209] Step S3140, mining typical features of each fault concept, marking them as attribute features, and filling them into the topological structure of the box-type transformer fault diagnosis knowledge graph as the attributes of the fault nodes corresponding to the fault concepts; the attribute features of the fault concepts include numerical attribute features and discrete attribute features.

[0210] In the fault tracing module, the reasoning and tracing of the fault warning information in combination with the box-type transformer fault diagnosis knowledge graph includes:

[0211] Step S3200, calculating the comprehensive matching degree between each fault symptom feature in the fault warning information and each fault node attribute feature in the box-type transformer fault diagnosis knowledge graph, and selecting the top N nodes with the highest comprehensive matching degree as candidate fault nodes;

[0212] Step S3300, extracting one-hop neighbor nodes of the candidate fault node in the box-type transformer fault diagnosis knowledge graph to form a semantic subgraph of the candidate fault;

[0213] Step S3400: Generate a causal chain in a semantic subgraph based on the causal chain of the box-type transformer fault diagnosis knowledge graph, and trace the cause of the candidate fault by analyzing the causal chain in the semantic subgraph;

[0214] Step S3500: Based on the co-location relationship family and concurrent probability network of the box-type transformer fault diagnosis knowledge graph, the number of co-location nodes and the probability of concurrent node occurrence in the semantic subgraph are counted to determine the location, severity and associated faults of the candidate fault;

[0215] Step S3600: Comprehensively consider the cause, location, severity and associated faults of the candidate faults, and output a fault diagnosis report based on the box-type transformer fault diagnosis knowledge graph.

[0216] The step S3200 includes:

[0217] Step S3210, calculate the similarity DE between each numerical symptom feature in the fault warning information and each numerical attribute feature of each fault node in the box transformer fault diagnosis knowledge graph;

[0218] Step S3220, calculate the similarity DB between each discrete symptom feature in the fault warning information and each discrete attribute feature of each fault node in the box transformer fault diagnosis knowledge graph based on the weighted Jaccard coefficient;

[0219] Step S3230, perform weighted averaging on the similarity DE and the similarity DB to obtain the comprehensive matching degree MP between the fault warning information and each fault node;

[0220] Step S3240, sort according to the comprehensive matching degree MP from high to low, and select the top N fault nodes with the highest matching degree as candidate fault nodes.

[0221] Embodiment 3

[0222] This embodiment discloses an electronic device, which may include one or more processors and one or more memories. Among them, computer-readable code is stored in the memory, and when the computer-readable code is run by one or more processors, it can execute the method for intelligent panoramic monitoring and management of new energy box transformers as described above.

[0223] The method or system according to the embodiment of the present application can also be implemented by means of the architecture of an electronic device. The electronic device may include a bus, one or more CPUs, a read-only memory (ROM), a random access memory (RAM), a communication port connected to a network, input / output components, a hard disk, etc. The storage device in the electronic device, such as ROM or hard disk, can store the method for intelligent panoramic monitoring and management of new energy box transformers provided by the present application. The method for intelligent panoramic monitoring and management of new energy box transformers may include, for example: obtaining the original monitoring data set of the target box transformer, performing data cleaning on the original monitoring data set to obtain the cleaned monitoring data set, extracting key features reflecting the operation state of the box transformer from the cleaned monitoring data set to obtain the original feature set; performing importance analysis on each feature in the original feature set, assigning weights to each feature in the original feature set according to the analysis results to generate a weighted feature set; constructing a box transformer fault prediction model, and obtaining a fault risk assessment value Q according to the weighted feature set and the box transformer fault prediction model; based on the fault risk assessment value, performing hierarchical early warning on potential faults to generate fault warning information; 0 < Q < 1; the hierarchical early warning of potential faults includes warning high-risk level, warning medium-risk level, warning low-risk level, and warning risk-free level; constructing a box transformer fault diagnosis knowledge graph, and when warning high-risk level or warning medium-risk level, reasoning and tracing the fault warning information in combination with the box transformer fault diagnosis knowledge graph.

[0224] Further, the electronic device may further include a user interface. Of course, the architecture disclosed in the present invention is only exemplary. When implementing different devices, one or more components in the electronic device disclosed in the present invention may be omitted according to actual needs.

[0225] Embodiment 4

[0226] This embodiment discloses a computer-readable storage medium, on which computer-readable instructions are stored. When the computer-readable instructions are run by a processor, the method for intelligent panoramic monitoring and management of a new energy box substation according to the embodiments of the present application can be executed. The storage medium includes, but is not limited to, for example, volatile memory and / or non-volatile memory. Volatile memory may include, for example, random access memory (RAM) and cache memory. Non-volatile memory may include, for example, read-only memory (ROM), hard disk, flash memory, etc.

[0227] In addition, according to the embodiments of the present application, the process described above with reference to the flowchart can be implemented as a computer software program. For example, the present application provides a non-transitory machine-readable storage medium, which stores machine-readable instructions that can be run by a processor to execute instructions corresponding to the method steps provided by the present application, such as: obtaining the original monitoring data set of the target box substation, performing data cleaning on the original monitoring data set to obtain the cleaned monitoring data set, extracting key features reflecting the operating state of the box substation from the cleaned monitoring data set to obtain the original feature set; performing importance analysis on each feature in the original feature set, assigning weights to each feature in the original feature set according to the analysis results to generate a weighted feature set; constructing a box substation fault prediction model, and obtaining a fault risk assessment value Q according to the weighted feature set and the box substation fault prediction model; based on the fault risk assessment value, performing hierarchical early warning on potential faults to generate fault early warning information; 0 < Q < 1; the performing hierarchical early warning on potential faults includes warning high-risk level, warning medium-risk level, warning low-risk level, and warning risk-free level; constructing a box substation fault diagnosis knowledge graph, and when warning high-risk level or warning medium-risk level, reasoning and tracing the fault early warning information in combination with the box substation fault diagnosis knowledge graph. When this computer program is executed by a central processing unit (CPU), the above functions defined in the method of the present application are executed.

[0228] The methods, systems, and devices of the present application may be implemented in many ways. For example, the methods, systems, and devices of the present application may be implemented by software, hardware, firmware, or any combination of software, hardware, and firmware. The above order of steps for the method is for illustration only, and the steps of the method of the present application are not limited to the order specifically described above unless otherwise specified. In addition, in some embodiments, the present application may also be implemented as programs recorded in a recording medium, which include machine-readable instructions for implementing the method according to the present application. Therefore, the present application also covers recording media storing programs for executing the method according to the present application.

[0229] In addition, the parts of the above-mentioned technical solutions provided in the embodiments of the present application that are consistent with the implementation principles of the corresponding technical solutions in the prior art are not described in detail to avoid excessive redundancy.

[0230] The specific implementation modes as described above further describe the purpose, technical solutions and beneficial effects of the present invention in detail. It should be understood that the above description is only a specific implementation mode of the present invention and is not intended to limit the present invention. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A method for intelligent panoramic monitoring and management of new energy box-type transformers, characterized in that: The method includes: Obtain the original monitoring data set of the target box substation, clean the original monitoring data set to obtain the cleaned monitoring data set, extract the key features reflecting the operation status of the box substation from the cleaned monitoring data set to obtain the original feature set; conduct importance analysis on each feature in the original feature set, assign weights to each feature in the original feature set according to the analysis results, and generate a weighted feature set; Construct a box substation fault prediction model, and obtain a fault risk assessment value Q according to the weighted feature set and the box substation fault prediction model; based on the fault risk assessment value, conduct hierarchical early warning on potential faults to generate fault early warning information; 0 < Q < 1; the hierarchical early warning of potential faults includes warning of high-risk level, warning of medium-risk level, warning of low-risk level, and warning of no-risk level; Construct a box substation fault diagnosis knowledge graph, and when warning of high-risk level or medium-risk level, conduct reasoning and tracing of the fault early warning information in combination with the box substation fault diagnosis knowledge graph; The generation of the weighted feature set includes: Calculate the Pearson correlation coefficient between each feature in the original feature set and the fault state, select the features with Pearson correlation coefficients higher than the preset correlation threshold as candidate features, and generate a candidate feature subset; Calculate the information gain of each candidate feature in the candidate feature subset with respect to the fault state, select the candidate features with information gain higher than the preset gain threshold as optimal features, and generate an optimal feature subset; Use the Gini index method to calculate the Gini index of each optimal feature in the optimal feature subset, and after normalizing the Gini index of each optimal feature, use it as the weight coefficient; Assign the weight coefficient to each feature in the optimal feature subset to generate a weighted feature set; The construction of the box substation fault diagnosis knowledge graph includes: Define core concepts to form a concept layer; the core concepts include box substation component concepts, material concepts, and fault concepts; Extract the semantic relationships between core concepts to form a relationship layer; the semantic relationships include causal relationships, concurrent relationships, and同位 relationships; the causal relationships are presented as causal chains, the同位 relationships are presented as同位 relationship families, and the concurrent relationships are presented as concurrent probability networks; Use the core concepts in the concept layer as nodes and the semantic relationships in the relationship layer as edges to construct the topological structure of the box substation fault diagnosis knowledge graph; Mine the typical features of each fault concept, mark them as attribute features, and fill them into the topological structure of the box substation fault diagnosis knowledge graph as the attributes of the fault nodes corresponding to the fault concepts; the attribute features of the fault concepts include numerical attribute features and discrete attribute features; The combination of the box substation fault diagnosis knowledge graph for reasoning and tracing of the fault early warning information includes: Calculate the comprehensive matching degree between each fault symptom feature in the fault early warning information and the attribute features of each fault node in the box substation fault diagnosis knowledge graph, and screen out the top N nodes with the highest comprehensive matching degree as candidate fault nodes; Extract the one-hop neighbor nodes of the candidate fault nodes in the box substation fault diagnosis knowledge graph to form a semantic subgraph of the candidate fault; Based on the causal chain of the box substation fault diagnosis knowledge graph, generate the causal chain in the semantic subgraph, and trace the cause of the candidate fault by analyzing the causal chain in the semantic subgraph; Based on the co-location relationship family and concurrent probability network of the box substation fault diagnosis knowledge graph, count the number of co-location nodes and the occurrence probability of concurrent nodes in the statistical semantic subgraph, and distinguish the location, severity, and associated faults of the candidate faults; Integrate the causes, locations, severities, and associated faults of the candidate faults, and output a fault diagnosis report based on the box substation fault diagnosis knowledge graph.

2. The intelligent panoramic monitoring management method for new energy box-type transformer according to claim 1 is characterized in that: The acquisition of the original monitoring data set of the target box substation includes: Collect the time-series change data of the internal temperature of the box substation; collect the time-series change data of the internal air humidity of the box substation; collect the time-series pulse signals of the partial discharge on the surface of the insulating parts; collect the time-series concentration data of characteristic fault gases; collect the time-series waveform data of the iron core grounding current; summarize all the data to form the original monitoring data set.

3. The intelligent panoramic monitoring management method for new energy box-type transformer according to claim 1 is characterized in that: The hierarchical early warning of potential faults and the generation of fault early warning information include: For the fault risk assessment value output by the box substation fault prediction model, set three risk thresholds: high risk threshold H, medium risk threshold M, and low risk threshold L, where 0 < L < M < H < 1; When Q > H, warn of the high risk level; when M ≤ Q ≤ H, warn of the medium risk level; when L < Q < M, warn of the low risk level; when Q ≤ L, warn of the no risk level; When warning of the high risk level or the medium risk level, trigger an alarm signal and generate fault early warning information, where the fault early warning information includes the fault risk level and the fault symptom characteristics; the fault symptom characteristics include numerical symptom characteristics and discrete symptom characteristics.

4. The intelligent panoramic monitoring management method for new energy box-type transformer according to claim 1 is characterized in that: The calculation of the comprehensive matching degree between each fault symptom characteristic in the fault early warning information and each fault node attribute characteristic in the box substation fault diagnosis knowledge graph includes: Calculate the similarity DE between each numerical symptom characteristic in the fault early warning information and each numerical attribute characteristic of each fault node in the box substation fault diagnosis knowledge graph; Calculate the similarity DB between each discrete symptom characteristic in the fault early warning information and each discrete attribute characteristic of each fault node in the box substation fault diagnosis knowledge graph based on the weighted Jaccard coefficient; Perform a weighted average on the similarity DE and the similarity DB to obtain the comprehensive matching degree MP between the fault early warning information and each fault node.

5. An intelligent panoramic monitoring and management system for a new energy box transformer, which is used to implement the intelligent panoramic monitoring and management method for a new energy box transformer as described in any one of claims 1 to 4, characterized in that: The system includes: Data processing module: used to obtain the original monitoring data set of the target box substation, clean the original monitoring data set to obtain the cleaned monitoring data set, extract the key features reflecting the operation status of the box substation from the cleaned monitoring data set to obtain the original feature set; perform importance analysis on each feature in the original feature set, and assign weights to each feature in the original feature set according to the analysis results to generate a weighted feature set; Fault prediction module: used to construct a box substation fault prediction model, obtain the fault risk assessment value Q according to the weighted feature set and the box substation fault prediction model; based on the fault risk assessment value, perform hierarchical early warning on potential faults and generate fault early warning information; 0 < Q < 1; the hierarchical early warning of potential faults includes warning of the high risk level, warning of the medium risk level, warning of the low risk level, and warning of the no risk level; Fault tracing module: used to construct a knowledge graph for fault diagnosis of box-type transformers. When warning of a high risk level or a medium risk level, the fault warning information is reasoned and traced in combination with the knowledge graph for fault diagnosis of box-type transformers.

Images