Transformer assembly fault diagnosis method based on machine learning

Abstract

The invention relates to the technical field of transformers, in particular to a transformer component fault diagnosis method based on machine learning, which comprises the following steps: analyzing medium data acquired by a transformer component, comparing a dielectric loss factor sequence with offset characteristics, training and generating an offset time sequence characteristic set, screening abnormal time periods and aggregating key discharge characteristics, and performing fault diagnosis. And obtaining an abnormal clustering signal group, concluding a trend evolution sequence, performing layered optimization on a path structure, and outputting fault link positioning data. According to the method, various parameters such as dielectric loss, discharge, temperature rise, gas and vibration are integrated, abnormal distribution is actively screened, key node trend change is captured in real time, historical and real-time characteristics in different operation scenes are fused, and efficient analysis of complex link evolution and implicit anomalies is realized. And the fault evolution relationship among the nodes is output in a structured manner, direct application of a diagnosis result in full-link traceability and multi-node trend research and judgment is supported, and the fault diagnosis pertinence and the hidden danger recognition capability are effectively improved.

Description

Technical Field

[0001] The present invention relates to the technical field of transformers, and in particular to a transformer component fault diagnosis method based on machine learning. Background Art

[0002] The field of transformer technology involves the design, manufacture, operation and maintenance of transformer equipment in the process of electric energy conversion, transmission and distribution. The core issues in this field include performance optimization and reliability improvement of components such as transformer cores, coils, insulation structures, and cooling systems. It also covers the monitoring of transformer operating conditions, the formulation of maintenance strategies, and the research and application of fault diagnosis and prevention measures. Among them, traditional transformer fault diagnosis refers to the identification of abnormal conditions of transformer components based on physical quantity change detection methods such as thermal imaging detection, oil chromatography analysis, and acoustic emission monitoring. For component-level fault identification problems such as local overheating of the transformer core, aging of coil insulation, and failure of the cooling system, gas component analysis in oil, partial discharge detection, core vibration signal extraction, temperature rise testing and other methods are usually used to diagnose and preliminarily determine faults in different transformer components.

[0003] In the process of daily equipment inspection and operation management, existing technologies rely on a single indicator as the basis for faults, making it difficult to form global trend monitoring and link tracing judgment. Faced with multiple types of anomalies or complex evolution phenomena across nodes, it is easy to miss local linkage hidden dangers. In actual work, if only oil-gas analysis or single-point temperature rise testing is used to evaluate equipment health, there is a lack of insight into the interaction of multiple links and the synchronous changes in trends. This limitation is particularly obvious when there are multiple working conditions interacting, long-term load fluctuations, or the accumulation of early hidden faults, which affects the timely discovery and accurate positioning of fault hazards. Summary of the Invention

[0004] The purpose of the present invention is to solve the shortcomings of the prior art and propose a transformer component fault diagnosis method based on machine learning.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: a transformer component fault diagnosis method based on machine learning, comprising the following steps:

[0006] S1: Based on the transformer components, the collected dielectric data is analyzed, the dielectric loss factor sequences of different components are compared, the offset characteristics of the signals at the same frequency point are determined, and the signal differences under various working conditions are integrated to obtain the offset time series feature set;

[0007] S2: Based on the offset time series feature set, filter the offset concentration period, calculate the change in the discharge amplitude of the transformer component, analyze the abnormal mutation performance, determine the aggregation of the signal in the feature space, and obtain the abnormal cluster signal group;

[0008] S3: Based on the abnormal cluster signal group, calculate the transformer component temperature rise monitoring information, compare the gas composition changes, determine the vibration signal center frequency changes, summarize the dynamic evolution of key features, and combine historical data to obtain a trend evolution sequence;

[0009] S4: According to the trend evolution sequence, determine the trend change, analyze the arrangement order of transformer components in the feature space, compare the trunk path feature trajectories, optimize the path hierarchical structure, summarize the hierarchical results, and obtain the path trunk hierarchical structure.

[0010] The improvements of the present invention are that the offset timing feature set includes insulation performance indicators, spectrum change characteristics, and working condition correlation factors; the abnormal clustering signal group includes discharge amplitude identification, abnormal classification number, and signal cluster characteristics; the trend evolution sequence includes temperature rise trajectory parameters, gas change patterns, and vibration response characteristics; and the path trunk hierarchical structure includes node hierarchical relationships, link status sets, and hierarchical index parameters.

[0011] The present invention is improved in that the steps of obtaining the offset timing feature set are specifically as follows:

[0012] S111: Based on the transformer component, the collected dielectric data is analyzed, and the dissipation factor sequences of the same frequency point at different time points are compared to determine the offset amplitude of the dissipation factor in the time dimension. The variation range of each frequency point sequence is calculated, and the frequency point trajectory with offset is identified to obtain the offset trajectory set.

[0013] S112: Based on the offset trajectory set, compare the performance of the dielectric loss factor under various operating conditions, analyze the impact of operating conditions on the changes in frequency point trajectories, identify key characteristic trajectories of offset performance under different operating conditions, and integrate the analyzed signal features to obtain an offset timing feature set.

[0014] The present invention is improved in that the step of obtaining the abnormal cluster signal group is specifically as follows:

[0015] S211: Based on the offset time series feature set, analyze the insulation performance parameters and spectrum change signals, optimize the continuous change trend of each feature within the different time periods, compare the fluctuation range under the time series, screen the sudden change section of the characteristic curve, determine the correlation between the signal amplitude increase and the discharge monitoring data, and obtain the discharge sudden change amplitude sequence;

[0016] S212: Based on the discharge mutation amplitude sequence, compare the spatial distribution of signal mutations in each segment, analyze the combination relationship between amplitude change and duration period, calculate the spatial combination difference, identify abnormal signal points of the structure, and obtain abnormal discharge cluster points;

[0017] S213: Based on the abnormal discharge cluster points, feature points with large signal amplitude changes, long duration periods, and high spatial density are screened, the aggregation relationships of the feature points are compared, and the spatial structure between representative points is analyzed to obtain an abnormal cluster signal group.

[0018] The present invention is improved in that the steps of obtaining the trend evolution sequence are specifically as follows:

[0019] S311: Based on the abnormal cluster signal group, calculate the discharge intensity change trend of each monitoring node in the target period, compare the gas composition change and temperature rise of each node in the same period, optimize the trend characteristics of each monitoring point, and obtain the discharge trend coefficient;

[0020] S312: Determine the distribution characteristics of the discharge trend coefficient among the structures, analyze the gas concentration change and temperature rise amplitude of each monitoring point, identify the offset of the center frequency of the vibration signal of each node, and obtain the structural evolution trend degree;

[0021] S313: Analyze the temperature rise trajectory, gas concentration change, and frequency offset of each node in the same monitoring period according to the structural evolution trend, compare the time series changes in the historical state data, and obtain a trend evolution sequence.

[0022] The present invention is improved in that the step of obtaining the hierarchical structure of the path backbone is specifically as follows:

[0023] S411: determining the change trend of each characteristic parameter of the transformer component during the operation cycle according to the trend evolution sequence, comparing the trend differences of each parameter in different time periods, and obtaining a component change trajectory indicator;

[0024] S412: Based on the component change trajectory indicator, analyzing the spatial distribution of characteristic points of each node of the transformer under temperature rise, gas and vibration parameters, identifying the order and arrangement differences between each node, and obtaining the component characteristic arrangement structure;

[0025] S413: According to the component feature arrangement structure, determine the hierarchical relationship of each node in the path, analyze the spatial coordinate difference, signal direction change and distribution between the nodes, and obtain the path trunk hierarchical structure.

[0026] The present invention is improved in that the steps further include:

[0027] S5: Based on the hierarchical structure of the path backbone, each path segment is analyzed, diagnostic features of temperature rise changes are compared, correlation characteristics between gas component rate increase and vibration frequency shift are identified, node order is adjusted, and fault link location data is obtained;

[0028] The fault link location data includes abnormal node number, fault link mapping information, and diagnosis index set.

[0029] The present invention is improved in that the steps of acquiring the fault link location data are specifically as follows:

[0030] S511: Based on the hierarchical structure of the path backbone, the temperature rise monitoring sequence of each path segment and the gas composition change sequence are compared to determine the characteristic distribution of the temperature rise change process of each path segment. Combined with the actual arrangement of the nodes in the hierarchical structure, the correlation performance between the diagnostic features of the path segments is optimized to obtain the temperature rise gas correlation data;

[0031] S512: Based on the temperature-rising gas correlation data, analyzing the synchronization performance between the gas component change sequence and the vibration frequency offset sequence, screening the behavior of the synchronization between the gas component rate increase and the vibration main frequency offset, and calculating the spatial distribution of the synchronization behavior between the path segment nodes to obtain a synchronization feature aggregation group;

[0032] S513: Based on the synchronization feature aggregation group, optimize the node connection order in the path backbone structure, identify the node paths with related characteristic performance, analyze the abnormal node numbers and link mapping information of each key path segment, and obtain the fault link location data.

[0033] Compared with the prior art, the advantages and positive effects of the present invention are:

[0034] In the present invention, multiple parameters such as dielectric loss, discharge, temperature rise, gas and vibration are integrated to actively screen abnormal distributions, capture trend changes of key nodes in real time, and integrate historical and real-time features under different operating scenarios to achieve efficient analysis of complex link evolution and hidden anomalies, and output the fault evolution relationship between nodes in a structured manner. This supports the direct application of diagnostic results in full-link tracing and multi-node trend analysis, effectively improving the targeted fault diagnosis and hidden danger identification capabilities, and improving the accuracy and foresight of fault diagnosis. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] Figure 1 It is a flow chart of the main steps of the present invention;

[0036] Figure 2 This is a flow chart for obtaining the offset timing feature set in the present invention;

[0037] Figure 3 This is a flowchart for obtaining an abnormal cluster signal group in the present invention;

[0038] Figure 4 A flowchart for obtaining a trend evolution sequence in the present invention;

[0039] Figure 5 This is a flowchart for obtaining the hierarchical structure of the path backbone in the present invention;

[0040] Figure 6 This is a flow chart for obtaining fault link location data in the present invention. DETAILED DESCRIPTION

[0041] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0042] In the description of the present invention, it should be understood that the terms "length", "width", "up", "down", "front", "back", "left", "right", "vertical", "horizontal", "top", "bottom", "inside", "outside", etc., indicating directions or positional relationships, are based on the directions or positional relationships shown in the accompanying drawings and are only for the convenience of describing the present invention and simplifying the description. They do not indicate or imply that the devices or elements referred to must have a specific direction, be constructed and operate in a specific direction, and therefore should not be understood as limiting the present invention. In addition, in the description of the present invention, the meaning of "plurality" is two or more, unless otherwise clearly and specifically defined.

[0043] Example

[0044] See also Figure 1 The present invention provides a technical solution: a transformer component fault diagnosis method based on machine learning, comprising the following steps:

[0045] S1: Based on transformer components, the collected dielectric data is analyzed, the dielectric loss factor sequences of different transformer components are compared, the offset characteristics of the same frequency point signals in the time dimension are determined, the performance differences of the collected signals under various operating conditions are optimized, the analyzed signal features are integrated and input into the machine learning process, the dataset is trained, and the offset time series feature set is obtained;

[0046] S2: Based on the offset time series feature set, the corresponding offset concentration period is selected, the discharge amplitude change of the transformer component is calculated, the abnormal mutation performance during the component operation is analyzed, the aggregation pattern of its signal in the feature space is determined, and the spatial distribution between the feature points is compared. The key abnormal discharge features are aggregated to obtain the abnormal cluster signal group;

[0047] S3: Based on the abnormal clustered signal group, calculate the temperature rise monitoring information of the transformer component, compare the changes in the gas composition of the component, determine the changing trend of the center frequency of the vibration signal, summarize the dynamic evolution of the key characteristics of the component, and combine it with historical status data to obtain the trend evolution sequence;

[0048] S4: According to the trend evolution sequence, determine the trend change it shows, analyze the arrangement order of transformer components in the feature space, compare the characteristic trajectories of the trunk path, optimize the hierarchical structure of each component path, summarize the hierarchical results, and obtain the path trunk hierarchical structure;

[0049] S5: Based on the hierarchical structure of the path backbone, analyze the involved path segments, compare the diagnostic features during the temperature rise process, identify the correlation characteristics between the rate increase of gas components and the vibration frequency offset of the transformer components, adjust the connection order between the path nodes, and obtain the fault link location data.

[0050] The offset time series feature set includes insulation performance indicators, spectrum change characteristics, and working condition correlation factors. The abnormal clustering signal group includes discharge amplitude identification, abnormal classification number, and signal cluster characteristics. The trend evolution sequence includes temperature rise trajectory parameters, gas change pattern, and vibration response characteristics. The path backbone hierarchical structure includes node hierarchical relationships, link status sets, and hierarchical index parameters. The fault link location data includes abnormal node numbers, fault link mapping information, and diagnostic index sets.

[0051] The transformer components in S1 refer to the insulation components (such as the main insulation paper layer, end insulation pads, and lead insulation parts). The analyzed dielectric data mainly involves the dielectric loss factor of the insulation system, focusing on insulation aging, local loss, and changes in material properties. It is the core collection object for judging dielectric loss changes and insulation status; the transformer components in S2 refer to coils, leads, lead ends, and local discharge monitoring related components, focusing on discharge amplitude, abnormal mutations and other characteristics. Related signals usually come from areas prone to local discharge such as windings (coils), lead ends, and lead-out devices. They are key monitoring areas for electrical anomalies; the transformer components in S3 refer to components related to temperature rise, gas and vibration monitoring points such as iron cores, oil channels, cooling systems, and seals. The analysis is of comprehensive conditions such as temperature rise, gas composition, and vibration signals. The status involves the transformer's internal core, oil channel temperature points, gas online monitoring interface, and vibration sensing unit, reflecting the overall operation and damage trends; the transformer components in S4 refer to key operating components (such as functional units, nodes / ports, path nodes, etc. in the main circuit), focusing on all major monitoring objects and their associated paths. It is a comprehensive judgment of the trends and trunk path structures among all key components, reflecting the global fault chain relationship; the transformer components in S5 refer to all component nodes participating in the path link (such as components / nodes with significantly abnormal indicators such as temperature rise, gas, and vibration), targeting actions such as path layering, node sorting, and link positioning, and actually involving components with all previous abnormal indicators. It integrates all the aforementioned monitoring points and nodes for tracing and locating the fault chain trunk.

[0052] In S1, the dissipation factor sequence refers to the continuous measurement data sequence of the dissipation factor (tgδ) of the transformer insulation medium at different acquisition times or frequency points, reflecting the changes in insulation status over time and frequency; the same-frequency point signal refers to the dissipation factor signal obtained at different time nodes under the same detection frequency, corresponding to the data trajectory of a specific frequency point, which facilitates the comparison of insulation aging trends; performance difference refers to the difference in the value or change trend of the dissipation factor or other insulation characteristics under different operating conditions (such as load, voltage, temperature changes, etc.), reflecting the changes in insulation characteristics under the influence of operating conditions.

[0053] In S2, the offset concentration period refers to the specific time interval when signals such as the dielectric dissipation factor show abnormal drift or drastic changes, which usually indicates deterioration of insulation performance or a sign of fault; the discharge amplitude change refers to the change in discharge intensity or amplitude in the electrical signal obtained through local discharge detection, which can indicate local insulation breakdown, discharge and other abnormalities; the abnormal mutation manifestation refers to the phenomenon of drastic changes, sudden increases and decreases in monitoring signals such as discharge, current, and voltage in a short period of time, which is a signal characteristic of local abnormalities or incipient faults; the aggregation rule refers to the similarity distribution and classification rules of multiple abnormal points (such as discharge mutation points) in the feature space (such as amplitude, duration, shape and other dimensions), which is used to identify homologous abnormalities or similar faults; the key abnormal discharge feature refers to the representative discharge signal feature obtained through the above clustering and screening (such as the abnormal point with the largest amplitude, longest duration, and special shape), which can serve as an important input for subsequent fault judgment.

[0054] In S3, gas composition changes refer to the changes in the concentration of various characteristic gases (such as H2, C2H2, CH4, etc.) in transformer oil over time, which is an important parameter for oil-gas analysis (DGA) fault diagnosis; vibration signals refer to the mechanical vibration data collected by vibration monitoring sensors at transformer structural parts, housings, coils, etc., which are used to reflect abnormalities such as mechanical looseness, partial discharge or internal impact; the dynamic evolution of key characteristics refers to the changing trajectories and trends of multiple characteristic parameters such as temperature rise, gas composition, vibration, etc. during operation, which is a full record of the evolution of equipment health status; historical status data refers to various types of transformer operation characteristic data collected and archived at different time nodes and working conditions, which are used for horizontal or vertical trend comparison and machine learning modeling training.

[0055] In S4, the trunk path refers to the key fault propagation path between components, which is sorted out based on the relationships such as signal transmission, fault impact link, and physical connection. It is the main chain of traceability diagnosis; the characteristic trajectory refers to the change path of the characteristic parameters (such as temperature rise, gas, vibration, etc.) of each node along the trunk path, which is used to analyze the evolution and transmission process of the fault signal; each component path refers to the multiple signal flow or fault propagation routes formed between multiple functional components inside the transformer, which comprehensively reflects the complex physical and electrical link relationships inside the equipment.

[0056] In S5, gas composition refers to the distribution and changes of various characteristic gases (H2, C2H2, CO, CH4, etc.) in oil at different path nodes, which is the basis for locating fault points and judging degradation trends; vibration frequency refers to the dominant frequency changes of mechanical vibration signals collected by component nodes, reflecting the existence of looseness, abnormal impact, internal electromagnetic faults, etc.; path nodes refer to specific component nodes in the trunk hierarchical structure (such as iron core, coil, lead, cooler, etc.), which are important observation and judgment locations on the fault link.

[0057] See also Figure 2 , the specific steps for obtaining the offset time series feature set are:

[0058] S111: Based on the transformer component, the collected dielectric data is analyzed, and the dissipation factor sequences of the same frequency point at different time points are compared to determine the offset amplitude of the dissipation factor in the time dimension. The variation range of each frequency point sequence is calculated, and the frequency point trajectory with offset is identified to obtain the offset trajectory set.

[0059] Based on the insulation components in the transformer, such as the main insulation paper layer and the lead insulation part, the dielectric data collected at different time points are first extracted. This data is mainly the measurement value of the dielectric loss factor, which is generally derived from periodic detection or online monitoring systems. Each detection is performed at multiple frequency points (for example, 50Hz, 100Hz, 200Hz, etc.). A time series is established for each frequency point. For example, at the 50Hz frequency point, the dielectric loss factors measured in January, February, and March are 0.0082, 0.0091, and 0.0113, respectively. The measured values ​​need to be classified by frequency point, and a two-dimensional sequence structure of frequency point-time is established. The difference calculation between two adjacent time nodes is performed on the sequence, and the results are 0.0009 and 0.0022. The offset amplitude of the dielectric loss factor is determined. Using a set offset judgment threshold, such as 0.0015, the change values ​​of each frequency point are compared. If the change value of any interval exceeds the threshold, such as 0.0022>0.0015, the frequency point is marked as an offset point within this time period. Next, all frequency points are traversed and the above comparison is repeated in the time series of each frequency point to determine whether there is an abnormal offset. In each frequency point time series, if there are three consecutive rising trends in the dielectric loss factor, and the three increments are 0.0010, 0.0013, and 0.0015 respectively, the frequency point is recorded as a continuous offset trajectory. All frequency points with offset records and their change values ​​are combined into a data set, namely the offset trajectory set, which is used for subsequent working condition difference analysis and signal feature extraction.

[0060] S112: Based on the offset trajectory set, compare the performance of the dielectric loss factor under various operating conditions, analyze the impact of operating conditions on the changes in frequency point trajectories, identify key characteristic trajectories of offset performance under different operating conditions, and integrate the analyzed signal features to obtain an offset time series feature set;

[0061] The offset trajectory set is correlated with the operating condition data of the transformer in the corresponding time period. The operating condition data generally includes variables such as operating current, voltage, oil temperature and ambient temperature. For example, in the offset trajectory, it is found that the dielectric loss factor of the frequency point 200Hz increases from 0.0078 to 0.0104 between April and May. At the same time, the transformer load current increases from 480A to 760A and the oil temperature increases from 62℃ to 76℃ during this period, indicating that the trajectory offset is related to high load and high temperature conditions. All offset frequency point trajectories are grouped and compared. For example, the dielectric loss factor sequence of the same frequency point under different operating conditions (such as low load and low temperature and high load and high temperature) is selected, and its average value and fluctuation range are compared. If the average dielectric loss factor is 0.0 under high load, the average dielectric loss factor is 0.0 under high load. 0.098, and 0.0081 under low load, with a difference of 0.0017. The standard deviation under high load conditions is 0.0009, which is significantly higher than 0.0004 under low load. It can be determined that this frequency point shows significant differences under different working conditions. The trajectories with obvious change trends are further screened out. For example, the increase amplitudes of three consecutive time points are 0.0012, 0.0016, and 0.0019, respectively, which are continuous upward trend trajectories. The data items such as the frequency point number, working condition type, change trend type, maximum amplitude, mean, and standard deviation of such key trajectories are recorded and encapsulated into structured feature records. All such records are integrated to form an offset time series feature set for subsequent clustering analysis and diagnostic path construction.

[0062] See also Figure 3 ,The specific steps for obtaining the abnormal clustering signal group are:

[0063] S211: Based on the offset time series feature set, analyze the insulation performance parameters and spectrum change signals, optimize the continuous change trend of each feature within the different time periods, compare the fluctuation range under the time series, screen the sudden change section of the characteristic curve, determine the correlation between the signal amplitude increase and the discharge monitoring data, and obtain the discharge sudden change amplitude sequence;

[0064] Extract the insulation performance parameters and their corresponding spectrum response signal values ​​from each characteristic curve. The insulation performance parameters refer to the dielectric loss factor values ​​at the acquisition time point, while the spectrum change signal refers to the amplitude information of different frequency components extracted from partial discharge detection or dielectric response testing. During the execution process, it is necessary to parse the insulation medium signal at each frequency point one by one, and extract the dielectric loss factor sequence corresponding to the frequency points of 100Hz, 200Hz, and 400Hz. For example, at 100Hz, the time series values ​​are 0.0085, 0.0090, 0.0093, and 0.0120, and at 200Hz they are 0.0079, 0.0081, 0.0084, and 0.0099. After arranging the sequence, calculate the first-order difference change value of each curve in chronological order to obtain a continuous change amplitude sequence. , then sum the absolute values ​​of the differences between two adjacent time points to determine whether there is an obvious fluctuation segment. If two consecutive differences in any segment are greater than 0.0015, it is determined that there is a trend mutation, and each mutation point is marked. Subsequently, the amplitude response of the spectrum signal in the mutation segment is compared to determine whether the energy of the main frequency component in the frequency domain in the segment increases. If the spectrum amplitude value in the mutation segment increases from 0.42 to 0.78, and the monitored discharge intensity voltage amplitude increases from 2.1mV to 5.3mV, then the mutation segment is recorded as a region with a significant increase in signal amplitude. According to the position of the mutation in each characteristic curve, the mutation start time point, mutation end time point, mutation duration, starting amplitude value, peak amplitude, and corresponding discharge amplitude value need to be extracted, and a mutation amplitude sequence is constructed.

[0065] S212: Based on the discharge mutation amplitude sequence, compare the spatial distribution of signal mutations in each segment and analyze the combined relationship between amplitude change and duration period using the formula:

[0066]

[0067] Calculate the spatial combination difference KG, identify the abnormal signal points of the structure, and obtain the abnormal discharge cluster points, where M i is the amplitude change of the i-th signal point, P i is the duration of the i-th signal point, Q i is the interval parameter of the i-th signal point in the spatial distribution, R i is the number of time superpositions of the i-th signal point, and n is the total number of signal points;

[0068] Determine whether each signal point belongs to the same type of discharge area structure, analyze its projection relationship in the winding arrangement and sensor array coordinates, and extract the amplitude change M corresponding to each signal point i , duration P i , spatial interval Q i Number of times superimposed with time R i, after normalization, unified weighting is used to enter the combined operation model, the amplitude change M i The duration period P is obtained by detecting the peak amplitude change in the partial discharge waveform. i is the continuous time length of the mutation segment, and the spatial interval Q i The time superposition times R is obtained by converting the three-dimensional distance of the point in the device coordinates. i The cumulative number of times an anomaly is repeatedly triggered within a sliding time window. Taking four typical signal points as an example, the parameter values ​​are as follows (normalized values ​​are shown in brackets):

[0069] Signal point 1: M1 = 3.5 (0.538), P1 = 2.0 (0.5), Q1 = 1.2 (0.4), R1 = 2 (0.333);

[0070] Signal point 2: M2 = 4.2 (0.646), P2 = 2.5 (0.625), Q2 = 1.5 (0.5), R2 = 3 (0.5);

[0071] Signal point 3: M3=2.8(0.431), P3=1.8(0.45), Q3=1.1(0.366), R3=2(0.333);

[0072] Signal point 4: M4 = 5.1 (0.785), P4 = 3.0 (0.75), Q4 = 1.3 (0.433), R4 = 4 (0.666);

[0073] Substituting into the formula, the calculation is as follows:

[0074] Molecular part:

[0075] (0.538·0.5-0.4·0.333)=0.269-0.133=0.136;

[0076] (0.646·0.625-0.5·0.5)=0.404-0.25=0.154;

[0077] (0.431·0.45-0.366·0.333)=0.194-0.122=0.072;

[0078] (0.785 0.75-0.433 0.666) = 0.589-0.288 = 0.301;

[0079] The sum of the molecules is:

[0080] 0.136+0.154+0.072+0.301=0.663;

[0081] Denominator:

[0082] (0.538·0.5+0.4·0.333)=0.269+0.133=0.402;

[0083] (0.646·0.625+0.5·0.5)=0.404+0.25=0.654;

[0084] (0.431 0.45 + 0.366 0.333) = 0.194 + 0.122 = 0.316;

[0085] (0.785 0.75 + 0.433 0.666) = 0.589 + 0.288 = 0.877;

[0086] The sum of the denominators is:

[0087] 0.402+0.654+0.316+0.877=2.249;

[0088] The combination difference is obtained as:

[0089]

[0090] The results show that in the signal mutation behavior, the difference between structural factors and signal characteristics is 0.2948, which is lower than the lower limit of the benchmark judgment of combined aggregation difference of 0.35. Therefore, it can be considered that although the current signal point has a mutation phenomenon, it does not have a significant clustering relationship in terms of spatial concentration characteristics. Subsequently, only points whose KG exceeds the set threshold are selected for abnormal point aggregation extraction to construct an abnormal discharge cluster point set. The formula can form a two-way comparison mechanism for spatial aggregation by introducing the product of the normalized discharge signal intensity and time characteristics, as well as the composite parameters of equipment structure distribution and time series repeatability. This makes spatial anomaly identification not only based on structural density, but also takes into account the temporal activity of signal mutations, thereby improving the stability and discrimination ability of fault signal aggregation judgment.

[0091] S213: Based on the abnormal discharge clustering points, feature points with large signal amplitude changes, long duration periods, and high spatial density are screened, the aggregation relationships of the feature points are compared, and the spatial structure between the representative points is analyzed to obtain an abnormal cluster signal group.

[0092] Spatial combination difference refers to a quantitative indicator used to measure the degree of combination difference of a group of signal points in the spatial dimension structure during the transformer fault diagnosis process. It is used to determine whether different signal points (usually discharge mutation signals) show aggregation rules or discrete characteristics in the space composed of multidimensional parameters.

[0093] First, traverse each identified discharge amplitude curve sequence and extract the feature points whose amplitude changes are greater than the set judgment threshold. The amplitude change threshold can be set with reference to twice the average discharge amplitude. For example, if the current background average discharge amplitude is 1.8mV, the judgment threshold is set to 3.6mV. If the amplitude of a feature point rises to 4.1mV, it is considered that its amplitude change meets the conditions. Further extract the mutation duration period of the curve where the feature point is located. If the discharge amplitude continues to be greater than the judgment threshold within five consecutive sampling cycles (5 seconds per cycle), mark the point as a feature point with a long duration period. Then obtain the spatial coordinate information between the feature point and other abnormal points. For example, record the three-dimensional position of its sensor node in the transformer shell coordinate system as X=3.5m, Y=1.2m, Z= 2.0m, by calculating the number of abnormal points within a radius of 1 meter around each feature point, if more than 15 abnormal points are detected within this radius, the spatial density of the point meets the high-density judgment condition, and all feature points with large amplitude changes, long duration periods, and high spatial density are screened according to the above conditions to form a feature point set. All points in the set are compared according to their positional relationship, the Euclidean distance between point pairs is obtained, and their average spacing is calculated. If the average spacing is less than 1.2 meters, it is considered that there is a spatial aggregation relationship between the points, and then the spatial clustering center is established according to the coordinate mean of each group of aggregation points. The connection relationship between the centers is topologically reconstructed, numbered according to the degree of aggregation, and the coordinates of the representative points of each cluster, the number of abnormal points contained, and the spatial extension range are identified, and the output is the abnormal clustering signal group.

[0094] See also Figure 4 , the specific steps for obtaining the trend evolution sequence are:

[0095] S311: Based on the abnormal cluster signal group, calculate the discharge intensity change trend of each monitoring node in the target period, compare the gas composition change and temperature rise of each node in the same period, optimize the trend characteristics of each monitoring point, and obtain the discharge trend coefficient;

[0096] The discharge amplitude data of each corresponding monitoring node in the transformer during the time period when the abnormal signal appears are retrieved in turn, and the discharge intensity sequence of the node in the selected time interval is constructed. For example, the discharge intensity of node A at 0 minutes, 5 minutes, 10 minutes, and 15 minutes is recorded as 1.8mV, 2.5mV, 3.4mV, and 5.2mV. The amplitude change difference sequence of the node in adjacent sampling time intervals is calculated to obtain the discharge change trend of the node in the target time period. By traversing all monitoring nodes with abnormal clustering signals, a discharge change curve of each node is formed. Then, the gas component change data collected by the node in the same time period are called to analyze the gas concentration. For example, the hydrogen concentration is 152ppm, 180ppm, 235ppm, and 312ppm respectively. At the same time, the temperature rise data is extracted, such as the oil temperature measuring point temperature is 64.1℃, 65.3℃, and 66. .7℃, 68.4℃, and perform time trend extraction operations on these three types of data respectively. The difference amplitude of each time point is normalized and then sorted. By comparing whether the discharge amplitude trend grows synchronously with the gas concentration trend and the temperature rise trend, the coupling characteristics of the abnormal signal evolution of the node are judged. Then the same process is performed on all nodes. The discharge change rate, gas growth rate, and temperature rise rate are calculated according to the trend rising slope of each node, and a trend comparison matrix is ​​constructed. Nodes with strong trend consistency are given a higher trend weight coefficient. The trend weight coefficient range is set to 0 to 1. If the discharge slope of a node is 0.24, the gas slope is 0.21, and the temperature rise slope is 0.22, and the variance of the three is within 0.0002, then a weight coefficient of 0.92 is given. Finally, the trend level is formed according to the sorting of the weight coefficients of each node, and the discharge trend coefficient of each node is output.

[0097] S312: Determine the distribution characteristics of the discharge trend coefficient among various structures, analyze the gas concentration changes and temperature rise amplitudes at each monitoring point, and identify the offset of the center frequency of the vibration signal at each node using the formula:

[0098]

[0099] Get the structural evolution trend EG Δ , among which, VG ch is the current gas component concentration at the hth monitoring point, VG bh is the reference gas component concentration at the hth monitoring point, TG max,h Indicates the highest temperature during the temperature rise monitoring process at the hth monitoring point, TG min,h Indicates the lowest temperature during the temperature rise monitoring process at the hth monitoring point, GQ h is the discharge trend coefficient corresponding to the hth monitoring point, n EG is the number of gas and temperature rise monitoring points, Gf cj is the current center frequency of the jth vibration monitoring point, Gf 0jis the reference center frequency of the jth vibration monitoring point, m EG is the number of vibration monitoring points;

[0100] According to the discharge trend coefficients of the three monitoring nodes P1, P2, and P3, the corresponding gas concentration changes and temperature rise spans are compared to obtain the current gas concentration VG of each node. ch , reference gas concentration VG bh , the highest temperature rise point TG max,h , lowest temperature rise point TG min,h and discharge trend coefficient GQ h , substitute into the calculation formula, and further introduce the current center frequency Gf corresponding to the vibration monitoring points V1 and V2 cj and reference center frequency Gf 0j , among which, the gas concentration at P1 node increases from 100μL / L to 125μL / L, the temperature rise increases from 65℃ to 78℃, the discharge trend coefficient is 0.82, and the normalized values ​​are 0.25, 0.52, and 0.41 respectively; the gas concentration at P2 node increases from 140μL / L to 150μL / L, the temperature rise increases from 70℃ to 85℃, the discharge trend coefficient is 0.95, and the normalized values ​​are 0.10, 0.63, and 0.48 respectively; the gas concentration at P3 node increases from 100μL / L to 125μL / L, the temperature rise increases from 65℃ to 78℃, the discharge trend coefficient is 0.82, and the normalized values ​​are 0.25, 0.52, and 0.41 respectively; The temperature rises from 76°C to 92°C, the discharge trend coefficient is 1.10, and the normalized values ​​are 0.10, 0.73, and 0.52 respectively. The current center frequency of the vibration monitoring point V1 is 62.5Hz, the reference frequency is 60.0Hz, the current frequency of V2 is 68.0Hz, and the reference frequency is 64.0Hz. The normalized frequency offsets are 0.18 and 0.28 respectively. Substitute them into the formula for calculation.

[0101] The molecular part is:

[0102] (0.25 0.52 0.41)+(0.10 0.63 0.48)+(0.10 0.73 0.52)

[0103] =0.0533+0.0302+0.03796=0.12146;

[0104] The denominator is:

[0105]

[0106] The structural evolution trend is:

[0107]

[0108] The results show that the structural evolution trend EG Δ≈0.1153 reflects the structural response intensity of multiple monitoring nodes in the current period under the combined effects of gas concentration changes, temperature rise span, and discharge trend. After normalization of the vibration center frequency offset, this value reflects the composite evolution level among multiple physical characteristic indicators. The higher the value, the stronger the synchronous fluctuation of multiple indicators in the structural level response. The lower the value, the relatively weaker the coupling between the current monitoring characteristics. The formula realizes cross-dimensional integrated expression in structural trend judgment by simultaneously integrating the discharge trend, gas concentration, and temperature rise amplitude, and normalizing them in the frequency perturbation space.

[0109] S313: Based on the structural evolution trend, analyze the temperature rise trajectory, gas concentration change, and frequency offset of each node within the same monitoring period, compare the time series changes in the historical state data, and obtain the trend evolution sequence.

[0110] The structural evolution trend is an indicator used to measure the comprehensive state change trend of the internal structural components of the transformer during operation. It mainly reflects the degree of response change of the structural components under the combined action of multiple characteristic parameters. Based on the joint performance of gas concentration changes, temperature rise amplitude and discharge trend at the structural monitoring points, combined with the degree of vibration frequency deviation, the operating state evolution trend of the transformer structural components is comprehensively quantified.

[0111] Extract the temperature rise data series, gas concentration data series, and vibration frequency data series of all monitoring nodes in the same period. The temperature rise data comes from thermistors distributed near the iron core, winding, and oil channel. For example, the oil temperature of node B in the continuous monitoring period is 67.2℃, 68.5℃, 69.3℃, and 70.9℃, respectively. The gas concentration increases from 15ppm C2H2 to 22ppm, 29ppm, and 40ppm. The frequency offset changes from 128Hz to 131Hz, 134Hz, and 139Hz. First, calculate the growth rate and average change rate of the temperature rise series. The temperature rise judgment threshold is set at 2°C per cycle. If the temperature rise is greater than this threshold for two consecutive cycles, it is marked as a temperature anomaly segment. At the same time, the gas growth is analyzed to see whether it exceeds the benchmark value of 6ppm per day. If the continuous growth rate is 7ppm or 8ppm, the anomaly condition is met. The points in the frequency offset sequence with each change greater than 3Hz are marked as mutations. The above temperature rise, gas and frequency anomaly points are further marked on the same time axis to analyze whether the three types of anomalies overlap at certain time nodes. If the overlap rate of the three anomaly time nodes exceeds 70%, it is constructed as a trend coupling segment. Then, the historical status data is called to search for the trend performance of the same type of components under the same operating conditions, and the monitoring records from three months ago are extracted as a comparison benchmark. For example, the historical temperature rise sequence is 64.2°C, 64.8°C, 65.6°C, and 66.1°C, the gas concentration is stable in the range of 12ppm to 18ppm, and the vibration frequency is maintained between 127Hz and 129Hz. The historical and current values ​​are differentially processed and normalized, and then superimposed to form a comparison sequence. If the temperature rise difference exceeds 5°C, the gas concentration increases by more than 15ppm, and the frequency offset exceeds 8Hz, the sequence is determined to be a trend evolution mutation segment. Finally, the node sequence with significant changes in multiple trend parameters is integrated into a trend evolution sequence.

[0112] See also Figure 5 ,The specific steps for obtaining the path backbone hierarchical structure are:

[0113] S411: judging the change trend of each characteristic parameter of the transformer component during the operation cycle according to the trend evolution sequence, comparing the trend difference of each parameter in different time periods, and obtaining the component change trajectory index;

[0114] For each transformer component, the time series data of temperature rise, gas concentration and vibration frequency are extracted within the specified operating cycle. Each type of parameter is indexed by time to form an independent sequence structure. For example, the temperature records of a component at 0 minutes, 10 minutes, 20 minutes and 30 minutes are 63.2℃, 64.5℃, 66.1℃ and 67.8℃, the acetylene concentration in the gas is 12ppm, 17ppm, 23ppm and 31ppm, and the vibration frequency changes are 122Hz, 124Hz, 127Hz and 131Hz. The single-point change calculation in the continuous time period is performed on the sequence in turn, and the difference between each two adjacent points is recorded as an increase sequence to determine whether it meets the set change trend standard. The temperature rise trend judgment standard is set to be no less than 1.2℃ every 10 minutes. If the current three-segment increase is 1.3℃, 1.6℃ and 1.7℃ respectively, it is determined that the parameter is in a continuous upward trend. The acetylene concentration trend judgment benchmark is set to an increase of more than 4ppm every 10 minutes. If the increase in each segment is 5ppm, 6ppm, 8ppm, it also meets the trend continuity requirement. For the vibration frequency, 3Hz is set as the mutation judgment benchmark. If the frequency increase is 2Hz, 3Hz, or 4Hz, at least the latter two segments meet the mutation standard. Subsequently, the trend categories of various parameters are classified and identified. For example, the continuous temperature rise, the continuous increase in gas concentration, and the vibration frequency mutation are marked as T1, T2, and T3 type features respectively. The three trends in the same time period are concurrently recorded as a joint trend event. At the same time, trend types are compared in different time periods, and the mean difference, maximum increase, and trend direction retention length of each parameter change in the operation cycle are extracted. The trend stability index is obtained by counting the number of trend direction change points. For example, if the temperature rise change direction changes no more than 2 times in 60 minutes of operation, the trend is marked as a unidirectional stable trend. Finally, a parameter change trend list is formed according to the above operations, and by integrating the change trajectories of the three types of parameters in different time periods, a sequence index characterizing the change path of the component operation state is constructed and output as a component change trajectory index.

[0115] S412: Based on the component change trajectory indicators, analyze the spatial distribution of characteristic points of each node of the transformer under temperature rise, gas and vibration parameters, identify the order and arrangement differences between each node, and obtain the component characteristic arrangement structure;

[0116] The characteristic value peaks of temperature rise, gas concentration and vibration frequency are extracted from all transformer nodes. The extraction method is to locate the maximum value change rate and the maximum value point in the time series of each parameter as the characteristic point. For example, the temperature of a node jumps to 69.5℃, the gas concentration reaches 35ppm, and the vibration frequency rises to 137Hz in the 40th minute of operation. The time point is recorded as the three types of characteristic points of the node. In the spatial structure, each node has corresponding specific coordinate information. According to the structural layout diagram of the transformer equipment, it is recorded that node A is at 2.5 meters on the X axis, 1.2 meters on the Y axis, and 3.0 meters on the Z axis, and node B is at 3.0 meters on the X axis, 1.6 meters on the Y axis, and 3.1 meters on the Z axis. Next, the spatial distance between each pair of nodes is calculated, and the change order of the characteristic parameters at the corresponding time points is compared. If node A has a gas peak in the 30th minute, node B experiences a sudden change in vibration frequency at the 35th minute. The order of its changes in the time dimension is judged to be a sequence relationship. Combined with its relative spatial position, it is determined whether it is arranged along the main axis of the structure. If the straight line connection direction between the nodes is consistent and the time difference between the parameters is less than 10 minutes, it is determined that there is a signal propagation path relationship. The arrangement order of the nodes in the path and the spatial connection direction are recorded as a permutation chain structure. Subsequently, all node combinations are processed in sequence. By extracting the order of parameter changes of each node relative to its previous and next nodes, as well as the degree of overlap of feature points in the three types of parameters, the permutation consistency score is calculated. The score threshold is set to 0.8. If the proportion of times the previous and next order in a node sequence remains consistent is greater than 80%, the sequence is marked as a highly consistent permutation chain. All highly consistent permutation chains are classified and numbered to construct a node group permutation structure information set.

[0117] S413: Arrange the structure based on component features, determine the hierarchical relationship of each node in the path, analyze the spatial coordinate difference between nodes, signal direction changes and distribution, and use the formula:

[0118]

[0119] The hierarchical structure of the path backbone is obtained, where HB represents the hierarchical coordination level of the path backbone, reflecting the hierarchical coordination status of each node of the transformer component path backbone. is the directional difference of the signal strength of the zth node in the θ direction, which indicates the degree of feature change of a specific node in the specified direction. DB z is the spatial coordinate difference between the zth node and its adjacent nodes, reflecting the physical distribution distance of the nodes in the structural path, ΔTB z is the offset of the zth node in the trend cycle, reflecting the dynamic evolution of node characteristics over time, n HB is the total number of nodes in the trunk path, used to determine the range of the summation, BR v is the spatial density distribution of the vth node, indicating the clustering of the node in the backbone path distribution. The hierarchical coordinate offset of the vth node in the α direction reflects the position change of the node in different hierarchical structures.

[0120] The path trunk hierarchical coordination level is an indicator that measures whether multiple key nodes in the trunk path have consistency and reasonable hierarchical relationships in terms of space, time and feature evolution. It can help identify potential abnormal combinations, incorrect hierarchical arrangements or disordered transmission of fault signals in the path structure, and is a key parameter for building fault chain tracing logic.

[0121] The nodes within the transformer component's trunk path are sorted by number, and their relative hierarchical position in the structural map is determined for each node. The main signal propagation direction of each node is then determined based on the path direction. The relative distribution of nodes within the path segment and the direction of change of the corresponding characteristics are analyzed. For each node, three key indicators are obtained: signal strength direction difference, spatial coordinate difference, and trend cycle offset. Assuming that there are four nodes within the component's trunk path, namely N1, N2, N3, and N4, the original monitoring data are as follows: the signal strength direction difference of node N1 is 6.5, the spatial coordinate difference is 2.4, and the trend cycle offset is 1.3; N2 is 5.2, 3.1, and 1.1; N3 is 7.1, 2.8, and 1.4; and N4 is 6.8, 3.0, and 1. 2. After normalizing all the original data, the normalized signal strength direction differences are 0.76, 0.58, 0.85, and 0.82, the normalized coordinate differences are 0.60, 0.78, 0.70, and 0.75, and the normalized trend cycle offsets are 0.68, 0.57, 0.73, and 0.63. The spatial density distribution and hierarchical coordinate offset of each node are continued to be called, where N1 is 0.90 and 3.0, N2 is 0.85 and 2.8, N3 is 0.92 and 3.2, and N4 is 0.88 and 3.1. The normalized density distribution is 0.72, 0.65, 0.75, and 0.70, and the hierarchical coordinate offsets are 0.81, 0.75, 0.85, and 0.82. Substituting into the formula:

[0122] Molecular part:

[0123] N1:

[0124] N2:

[0125] N3:

[0126] N4:

[0127] The sum of the numerators is:

[0128] -0.147+(-0.386)+(-0.161)+(-0.159)=-0.853;

[0129] The denominator is calculated as follows:

[0130] N1: 0.72·0.81=0.5832;

[0131] N2: 0.65·0.75=0.4875;

[0132] N3: 0.75·0.85=0.6375;

[0133] N4: 0.70·0.82=0.574;

[0134] The sum of the denominators is:

[0135] 0.5832+0.4875+0.6375+0.574=2.2822;

[0136] get:

[0137]

[0138] The results show that there are collective deviations in the signal direction, spatial position and evolution trend of multiple nodes in the current trunk path. p =-0.374 is a negative number, reflecting that the coupling coordination of the nodes in the trunk path in the spatial and trend dimensions is insufficient. It is necessary to further adjust the node order or reconstruct the path structure to obtain the hierarchical structure of the path trunk. The numerical result is in the negative interval, further indicating that the structure of the path is not highly consistent. It should be optimized by adjusting the component order and reorganizing the path segments to construct a path with clear hierarchy and continuous structure. Therefore, this numerical result is directly used to identify the node pair combinations that need to be adjusted in the trunk path. As an intermediate judgment basis, it can be compared with the structural stability reference interval (for example, the reference interval set is [0.2, 1.5]) to identify the distribution imbalance area.

[0139] See also Figure 6 ,The steps for obtaining fault link location data are as follows:

[0140] S511: Based on the hierarchical structure of the path backbone, the temperature rise monitoring sequence of each path segment and the gas composition change sequence are compared to determine the characteristic distribution of the temperature rise change process of each path segment. Combined with the actual arrangement of the nodes in the hierarchical structure, the correlation performance between the diagnostic features of the path segments is optimized to obtain the temperature rise gas correlation data;

[0141] The temperature rise monitoring sequence and gas composition change sequence corresponding to the nodes involved in each path segment are extracted. The temperature rise monitoring sequence is provided by the temperature sensor at the node location. For example, the temperature sequences corresponding to nodes 1 to 3 in path segment A are 67.2°C, 68.5°C, 70.4°C, and 71.3°C, respectively. The gas composition change sequence records the characteristic gas concentration measured at each time point. For example, the concentrations of acetylene in the same time period are 12ppm, 17ppm, 25ppm, and 31ppm, respectively. The continuous time difference in the temperature rise curve of each path segment is calculated to obtain the temperature rise slope of each segment to determine whether there is a sudden increase trend. If the temperature rise amplitude of a path segment exceeds 1.5°C in two consecutive sampling cycles, it is marked as a high-temperature rise path segment. The gas composition curve is processed in the same way. If the acetylene concentration growth values ​​in three consecutive cycles are 5ppm, 8ppm, and 6ppm, respectively, and exceed the baseline rate of 5ppm, it is recorded as a gas abnormality segment. , then determine whether the temperature rise and gas change trends in the path segment overlap. If the change directions of the slopes of the two sequences are consistent and the peak time difference does not exceed 10 minutes, it is considered that there is trend synchronization between the two. Combined with the actual arrangement of the nodes of the path segment in the hierarchical structure of the path trunk, such as node 1 in the upper layer, node 2 in the middle layer, and node 3 in the lower layer, record their vertical arrangement numbers in the structure. If the heat change develops sequentially from the upper layer to the lower layer, and the gas concentration first rises at the lower layer node and then transmits to the middle layer and upper layer nodes, it is determined that the temperature rise and gas change in the path segment have a hierarchical offset characteristic. On this basis, a parameter trend matching matrix is ​​constructed for each path segment, and the consistency score is calculated according to the time point and intensity change level of the node corresponding parameter change. If the score is higher than 0.85, it is recorded as a strongly coupled path segment. The above data are summarized and integrated into a set of path segment-level parameter comparison data, which are output as temperature rise gas correlation data.

[0142] S512: Based on the temperature-rise gas correlation data, analyze the synchronization performance between the gas component change sequence and the vibration frequency offset sequence, select the behavior of the synchronization between the gas component rate increase and the vibration main frequency offset, and calculate the spatial distribution of the synchronization behavior between the path segment nodes to obtain the synchronization feature aggregation group;

[0143] The gas composition change sequence and vibration frequency offset sequence of the key nodes in each path segment are extracted, and aligned and compared in the time dimension to determine whether there is synchronization between the two. For example, in path segment B, the acetylene concentration of node X increases from 14ppm to 21ppm in the 20th minute, to 28ppm in the 25th minute, and to 33ppm in the 30th minute. The corresponding vibration main frequency shifts from 129Hz to 133Hz, 136Hz, and 139Hz, respectively. First, calculate the increase rate of the two adjacent sequences in the same time period, record the gas change rate as 7ppm every 5 minutes, and the vibration offset rate as 3Hz every 5 minutes, and determine whether they exceed their respective reference thresholds at the same time. The reference thresholds are set as 5ppm per 5 minutes for gas rate and 2Hz per 5 minutes for vibration frequency. If the two types of data are in two or more consecutive time periods, If the threshold is exceeded at the same time, synchronization behavior is determined to have occurred. The continuity of the synchronization behavior is also calculated. If the continuous time is greater than 15 minutes, it is marked as continuous synchronization behavior. The set of all nodes in the path segment where synchronization behavior occurs is further counted, and the distribution density of the nodes in the spatial coordinate system is analyzed. The spatial coordinates are read according to the physical location of the node. For example, the X, Y, and Z nodes are located at (2.0m, 1.2m, 3.0m), (2.4m, 1.3m, 3.2m), and (2.7m, 1.4m, 3.3m), respectively. A spherical area with a radius of 1.0 meter is constructed with each node as the center. The number of synchronous behavior nodes in each area is counted. If a node contains two other synchronous nodes within its radius, the node is a dense aggregation center. Nodes that meet the requirements of synchronization rate, duration, and spatial density are grouped into synchronization feature aggregation groups.

[0144] S513: Based on the synchronization feature aggregation group, the node connection sequence in the path backbone structure is optimized, node paths with associated characteristic performance are identified, and abnormal node numbers and link mapping information of each key path segment are analyzed to obtain fault link location data;

[0145] Read the connection order of the nodes in each aggregation group in the path backbone structure. If the current order is node X→Y→Z, and X, Y, and Z all belong to the same synchronization aggregation group, then it is necessary to re-evaluate the rationality of their connection and analyze whether there is a reverse propagation of synchronization behavior, that is, node Z generates a synchronization event before X in time. If this situation is repeated in three or more path segments, and the time difference is less than 10 minutes, it is considered that the original connection order is misarranged. Record all node pairs with reverse synchronization time characteristics, extract their connection direction, time difference, spatial distance and other indicators, and re-sort the connection order in the path segment in chronological order. , update the node connection relationship to Z→Y→X, and correct the backbone structure path diagram. Then, mark the node where all synchronization feature anomalies first appear in the corrected path diagram, record it as the abnormal node number, and establish a mapping path between the node and the downstream propagation node. If the anomaly first appears at node Z and propagates to X through Y, then the link mapping information is recorded as ZYX. At the same time, assign a number to each link, and summarize the mutation time, mutation amplitude, synchronization strength and other information of the corresponding parameters of each node in the link. Build a fault signal transmission path model, and output all mapped links and abnormal node numbers as fault link location data.

[0146] The above are merely preferred embodiments of the present invention and do not limit the present invention in any other form. Any technician familiar with the profession may use the technical content disclosed above to change or modify it into an equivalent embodiment with equivalent changes and apply it to other fields. However, any simple modification, equivalent change and modification made to the above embodiment based on the technical essence of the present invention without departing from the content of the technical solution of the present invention shall still fall within the scope of protection of the technical solution of the present invention.

Claims

1. A transformer component fault diagnosis method based on machine learning, characterized in that: The following steps are involved: S1: Based on the transformer components, the collected dielectric data is analyzed, the dielectric loss factor sequences of different components are compared, the offset characteristics of the signals at the same frequency point are determined, and the signal differences under various working conditions are integrated to obtain the offset time series feature set; S2: Based on the offset time series feature set, filter the offset concentration period, calculate the change in the discharge amplitude of the transformer component, analyze the abnormal mutation performance, determine the aggregation of the signal in the feature space, and obtain the abnormal cluster signal group; S3: Based on the abnormal cluster signal group, calculate the transformer component temperature rise monitoring information, compare the gas composition changes, determine the vibration signal center frequency changes, summarize the dynamic evolution of key features, and combine historical data to obtain a trend evolution sequence; S4: According to the trend evolution sequence, determine the trend change, analyze the arrangement order of transformer components in the feature space, compare the trunk path feature trajectories, optimize the path hierarchical structure, summarize the hierarchical results, and obtain the path trunk hierarchical structure.

2. The transformer component fault diagnosis method based on machine learning according to claim 1 is characterized in that: The offset timing feature set includes insulation performance indicators, spectrum change characteristics, and working condition correlation factors; the abnormal clustering signal group includes discharge amplitude identification, abnormal classification number, and signal cluster characteristics; the trend evolution sequence includes temperature rise trajectory parameters, gas change pattern, and vibration response characteristics; the path backbone hierarchical structure includes node hierarchical relationship, link state set, and hierarchical index parameters.

3. The transformer component fault diagnosis method based on machine learning according to claim 1 is characterized in that: The steps for obtaining the offset timing feature set are specifically as follows: S111: Based on the transformer component, the collected dielectric data is analyzed, and the dissipation factor sequences of the same frequency point at different time points are compared to determine the offset amplitude of the dissipation factor in the time dimension. The variation range of each frequency point sequence is calculated, and the frequency point trajectory with offset is identified to obtain the offset trajectory set. S112: Based on the offset trajectory set, compare the performance of the dielectric loss factor under various operating conditions, analyze the impact of operating conditions on the changes in frequency point trajectories, identify key characteristic trajectories of offset performance under different operating conditions, and integrate the analyzed signal features to obtain an offset timing feature set.

4. The transformer component fault diagnosis method based on machine learning according to claim 1, characterized in that: The steps for obtaining the abnormal cluster signal group are specifically as follows: S211: Based on the offset time series feature set, analyze the insulation performance parameters and spectrum change signals, optimize the continuous change trend of each feature within the different time periods, compare the fluctuation range under the time series, screen the sudden change section of the characteristic curve, determine the correlation between the signal amplitude increase and the discharge monitoring data, and obtain the discharge sudden change amplitude sequence; S212: Based on the discharge mutation amplitude sequence, compare the spatial distribution of signal mutations in each segment, analyze the combination relationship between amplitude change and duration period, calculate the spatial combination difference, identify abnormal signal points of the structure, and obtain abnormal discharge cluster points; S213: Based on the abnormal discharge cluster points, feature points with large signal amplitude changes, long duration periods, and high spatial density are screened, the aggregation relationships of the feature points are compared, and the spatial structure between representative points is analyzed to obtain an abnormal cluster signal group.

5. The transformer component fault diagnosis method based on machine learning according to claim 1, characterized in that: The steps for obtaining the trend evolution sequence are specifically as follows: S311: Based on the abnormal cluster signal group, calculate the discharge intensity change trend of each monitoring node in the target period, compare the gas composition change and temperature rise of each node in the same period, optimize the trend characteristics of each monitoring point, and obtain the discharge trend coefficient; S312: Determine the distribution characteristics of the discharge trend coefficient among the structures, analyze the gas concentration change and temperature rise amplitude of each monitoring point, identify the offset of the center frequency of the vibration signal of each node, and obtain the structural evolution trend degree; S313: Analyze the temperature rise trajectory, gas concentration change, and frequency offset of each node in the same monitoring period according to the structural evolution trend, compare the time series changes in the historical state data, and obtain a trend evolution sequence.

6. The transformer component fault diagnosis method based on machine learning according to claim 1, characterized in that: The steps for obtaining the path backbone hierarchical structure are specifically as follows: S411: determining the change trend of each characteristic parameter of the transformer component during the operation cycle according to the trend evolution sequence, comparing the trend differences of each parameter in different time periods, and obtaining a component change trajectory indicator; S412: Based on the component change trajectory indicator, analyzing the spatial distribution of characteristic points of each node of the transformer under temperature rise, gas and vibration parameters, identifying the order and arrangement differences between each node, and obtaining the component characteristic arrangement structure; S413: According to the component feature arrangement structure, determine the hierarchical relationship of each node in the path, analyze the spatial coordinate difference, signal direction change and distribution between the nodes, and obtain the path trunk hierarchical structure.

7. The transformer component fault diagnosis method based on machine learning according to claim 1, characterized in that: The steps also include: S5: Based on the hierarchical structure of the path backbone, each path segment is analyzed, diagnostic features of temperature rise changes are compared, correlation characteristics between gas component rate increase and vibration frequency shift are identified, node order is adjusted, and fault link location data is obtained; The fault link location data includes abnormal node number, fault link mapping information, and diagnosis index set.

8. The transformer component fault diagnosis method based on machine learning according to claim 7 is characterized in that: The steps for obtaining the fault link location data are specifically as follows: S511: Based on the hierarchical structure of the path backbone, the temperature rise monitoring sequence of each path segment and the gas composition change sequence are compared to determine the characteristic distribution of the temperature rise change process of each path segment. Combined with the actual arrangement of the nodes in the hierarchical structure, the correlation performance between the diagnostic features of the path segments is optimized to obtain the temperature rise gas correlation data; S512: Based on the temperature-rising gas correlation data, analyzing the synchronization performance between the gas component change sequence and the vibration frequency offset sequence, screening the behavior of the synchronization between the gas component rate increase and the vibration main frequency offset, and calculating the spatial distribution of the synchronization behavior between the path segment nodes to obtain a synchronization feature aggregation group; S513: Based on the synchronization feature aggregation group, optimize the node connection order in the path backbone structure, identify the node paths with related characteristic performance, analyze the abnormal node numbers and link mapping information of each key path segment, and obtain the fault link location data.

Images