A transformer DGA diagnosis method based on mechanism diagram gated dual-flow network

Abstract

The application discloses a transformer DGA diagnosis method based on a mechanism diagram gate dual-flow network, and comprises the following steps: inputting concentration flow and rate flow into the dual-flow network; constructing a fault stage mechanism diagram; extracting concentration response features and gas production rate response features of a gas correlation edge; determining a gate coefficient of the gas correlation edge, and fusing the concentration response features and the gas production rate response features of the gas correlation edge to obtain a gate fusion feature of the gas correlation edge; and determining a fault diagnosis result of the transformer. According to the application, the fault judgment basis is expanded from the static content of a single gas to the cooperative evolution relationship of gas combination in the concentration accumulation level and the fault development speed, and the stability and reliability of the transformer DGA fault diagnosis under complex working conditions, noise samples and boundary samples are improved.

Description

Technical Field

[0001] This invention relates to the field of transformer fault diagnosis technology, and in particular to a transformer DGA diagnosis method based on a mechanism graph-gated two-current network. Background Technology

[0002] Power transformers are critical equipment in the power grid, and their operating status directly affects the reliability of power supply. Under thermal and electrical stress, the insulating oil and insulating paper inside the transformer can decompose, producing hydrogen gas. methane Acetylene ethylene ethane Characteristic gases, such as those present in transformer oil, exhibit varying gas generation mechanisms, gas combination relationships, and gas production rate changes corresponding to different fault types and development stages. Dissolved gas analysis (DGA) detects the content of various characteristic gases in transformer oil to identify latent faults such as overheating and discharge within the transformer, making it an important means of transformer condition monitoring. With the development of artificial intelligence technology, DGA intelligent diagnostic methods based on machine learning and deep learning are gradually replacing traditional ratio methods and region partitioning methods. Graph neural networks, temporal networks, and gating fusion mechanisms have also been introduced into the field of transformer DGA fault diagnosis, achieving automatic fault category identification by mining the nonlinear features contained in gas concentration data.

[0003] However, existing intelligent diagnostic methods typically employ simple splicing or global weighting at the overall feature level when fusing gas concentration and gas production rate information, without constraining the consistency of the two types of information at the gas correlation level. In actual detection environments, factors such as instantaneous sensor interference and oil sample acquisition errors can cause isolated spikes in the gas production rate of individual gases. Detection noise and short-term abnormal fluctuations can also cause non-fault-related changes in gas concentration. Since existing fusion methods cannot determine whether the concentration response and gas production rate response support each other on the same gas correlation, the aforementioned isolated spikes, detection noise, and short-term abnormal fluctuations are easily mistaken as valid evidence of fault development and used in fault identification, resulting in insufficient stability of transformer DGA fault diagnosis results under complex operating conditions, noise samples, and boundary samples. Summary of the Invention

[0004] This invention provides a transformer DGA diagnostic method based on a mechanism graph-gated two-stream network to solve the above-mentioned problems.

[0005] This invention provides a transformer DGA diagnostic method based on a mechanism graph-gated two-stream network, comprising: acquiring concentration data of multiple characteristic gases in transformer oil at multiple sampling times; determining the gas generation rate of each characteristic gas based on the concentration change of each characteristic gas between adjacent sampling times; constructing a concentration stream using the concentration data of each characteristic gas and a rate stream using the gas generation rate of each characteristic gas, and inputting the concentration stream and the rate stream into a two-stream network; constructing a fault stage mechanism graph using multiple characteristic gases as graph nodes through the two-stream network; wherein, the gas association edges in the fault stage mechanism graph are used to characterize the association relationship between corresponding characteristic gases that conforms to the fault stage mechanism; and using the two-stream network, for the fault stage mechanism... For each gas-related edge in the fault mechanism diagram, the concentration response feature of the gas-related edge is extracted based on the concentration data of the gas at both ends of the edge, and the gas production rate response feature is extracted based on the gas production rate at both ends of the edge. Using the dual-flow network, for each gas-related edge in the fault mechanism diagram, the gating coefficient of the gas-related edge is determined based on the consistency between its concentration response feature and gas production rate response feature. The concentration response feature and gas production rate response feature of the gas-related edge are then fused based on the gating coefficient to obtain the gating fusion feature of the gas-related edge. Based on the gating fusion feature of each gas-related edge, the fault diagnosis result of the transformer is determined.

[0006] In the above scheme, gas concentration data and gas production rate are organized into concentration streams and rate streams, respectively, and input into a dual-stream network. Dual-stream response features are extracted using gas-related edges in the fault stage mechanism diagram as units, and gating fusion is performed based on consistency. This expands the fault discrimination criterion from the static content of a single gas to the co-evolutionary relationship between gas combinations in two dimensions: concentration accumulation level and fault development speed. This improves the stability and reliability of transformer DGA fault diagnosis under complex operating conditions, noise samples, and boundary samples. Furthermore, the fault stage mechanism diagram is constructed based on the fault stage mechanism, and the gas-related edges represent the correlations between corresponding characteristic gases that conform to the fault stage mechanism. This approach directly correlates the extracted features with the gas generation mechanisms in the development of faults such as overheating and discharge. The network's extraction and fusion of concentration response and gas production rate have clear physical meanings, improving the interpretability of diagnostic results. Furthermore, the gating coefficient is determined based on the consistency between the concentration response features and the gas production rate response features on the same gas-related edge. When the two types of responses support each other, the feature contribution of the gas-related edge is preserved. When the gas production rate shows an isolated spike, or when the gas concentration shows detection noise or short-term abnormal fluctuations, the feature contribution of the gas-related edge is suppressed. This avoids non-fault fluctuations being mistakenly used as evidence of fault development for judgment, reducing the risk of misjudgment.

[0007] Furthermore, the step of constructing a fault stage mechanism diagram using multiple characteristic gases as graph nodes and based on the fault stage mechanism includes: constructing an overheating mechanism sub-graph based on the gas generation mechanism of the overheating fault stage; constructing a discharge mechanism sub-graph based on the gas generation mechanism of the discharge fault stage; constructing an overheating-discharge transition mechanism sub-graph based on the gas generation mechanism of the overheating-discharge transition stage; determining the combination weights of the overheating mechanism sub-graph, the discharge mechanism sub-graph, and the overheating-discharge transition mechanism sub-graph based on the concentration data and gas production rate of each characteristic gas; and weighting the overheating mechanism sub-graph, the discharge mechanism sub-graph, and the overheating-discharge transition mechanism sub-graph according to the combination weights to obtain the fault stage mechanism diagram.

[0008] In the above scheme, corresponding mechanism subgraphs are constructed based on the gas generation mechanisms of the three fault stages: overheating, discharge, and overheating-discharge transition. The mechanism subgraphs are then weighted and combined according to the combined weights determined by the current concentration data and gas production rate. This allows the fault stage mechanism graph to adaptively adjust the correlation between gases as the fault stage characteristics presented by the samples change. This avoids the problem of mismatch in correlation between fixed empirical graphs or statistical correlation graphs when the fault stage changes. It improves the applicability of the mechanism graph to samples of overheating, discharge, and their boundary regions, and provides a graph structure foundation that matches the current fault stage for subsequent edge-level feature extraction and gating fusion.

[0009] Further, determining the combined weights of the superheating mechanism sub-graph, the discharge mechanism sub-graph, and the superheating-discharge transition mechanism sub-graph based on the concentration data and gas production rate of each characteristic gas includes: determining the current concentration and concentration change of each characteristic gas based on the concentration data of each characteristic gas at multiple sampling times; determining the average gas production rate and gas production rate change trend of each characteristic gas based on the gas production rate of each characteristic gas between adjacent sampling times; determining the key gas ratio based on the current concentration of each characteristic gas; and determining the combined weights of the superheating mechanism sub-graph, the discharge mechanism sub-graph, and the superheating-discharge transition mechanism sub-graph based on the current concentration, the concentration change, the average gas production rate, the gas production rate change trend, and the key gas ratio of each characteristic gas.

[0010] In the above scheme, the combined weight is determined by comprehensively considering the current concentration, concentration change, average gas production rate, gas production rate trend, and key gas ratio. The combined weight comprehensively characterizes the current fault stage of the sample from multiple dimensions, including gas accumulation level, concentration evolution direction, fault development speed, and gas combination ratio. This allows the combined weight to accurately reflect the degree of proximity of the sample to the three types of fault stages: overheating, discharge, and overheating-discharge transition. It avoids misjudgment of the stage caused by relying solely on static concentration or a single ratio criterion, and improves the matching accuracy between the fault stage mechanism diagram and the actual fault stage of the sample.

[0011] Furthermore, the method further includes: for each gas-related edge in the fault stage mechanism diagram, extracting a stage progression response term for the gas-related edge based on the concentration change and gas production rate of the gas at both ends of the gas-related edge and the preset fault stage progression direction corresponding to the gas-related edge; wherein, the stage progression response term is used to characterize whether the change relationship of the gas at both ends of the gas-related edge conforms to the fault development law under the corresponding fault stage. The step of determining the gating coefficient of the gas-related edge based on the degree of consistency between the concentration response characteristics and the gas production rate response characteristics of the gas-related edge includes: determining the gating coefficient of the gas-related edge based on the degree of consistency between the concentration response characteristics and the gas production rate response characteristics of the gas-related edge and the stage progressive response term of the gas-related edge.

[0012] In the above scheme, in addition to the consistency between the concentration response and the gas production rate response, a stage-progressive response term is introduced when determining the gating coefficient. This allows the gating coefficient to simultaneously reflect whether the change relationship of the gases at both ends of the gas-related edge conforms to the fault development law under the corresponding fault stage. Even if the concentration response and the gas production rate response on a certain gas-related edge are consistent in value, as long as their change relationship contradicts the progressive law of the fault stage, the contribution of the gas-related edge to the diagnostic result will still be reduced by the gating coefficient. This excludes the cooperative fluctuations caused by atypical fault evolution from the valid fault evidence and improves the diagnostic accuracy of adjacent fault level samples and atypical evolution samples.

[0013] Furthermore, the method for setting the preset fault stage progression direction includes: for the gas-related edge corresponding to the overheating fault stage, the preset fault stage progression direction is the direction in which the concentration of the thermal fault-related gas increases as the fault develops; for the gas-related edge corresponding to the discharge fault stage, the preset fault stage progression direction is the direction in which the concentration of the discharge-related gas increases as the fault develops; for the gas-related edge corresponding to the overheating-discharge transition stage, the preset fault stage progression direction is the direction in which the concentration of the discharge-related gas increases as the fault develops. The step of extracting the stage progression response term of the gas-related edge based on the concentration change and gas production rate of the gas at both ends of the gas-related edge and the preset fault stage progression direction corresponding to the gas-related edge includes: determining that the stage progression response term takes a positive value when the change relationship of the gas at both ends of the gas-related edge conforms to the preset fault stage progression direction; and determining that the stage progression response term takes a negative value when the change relationship of the gas at both ends of the gas-related edge is opposite to the preset fault stage progression direction.

[0014] In the above scheme, the gas-related edges of the three fault stages of overheating, discharge, and overheating-discharge transition are preset with progressive directions according to the gas enhancement direction under the corresponding fault mechanism. The positive and negative values ​​of the stage progressive response terms are determined based on whether the change relationship of the gases at both ends matches the progressive direction. This makes the discrimination rule of the stage progressive response terms directly derived from the objective mechanism that thermal fault gases are gradually generated as the temperature rises and discharge characteristic gases are rapidly accumulated as the discharge energy increases. Evolutions that conform to the progressive law obtain positive responses, while evolutions that do not conform to the progressive law obtain negative responses. This ensures that the directionality of stage progressive discrimination is strictly consistent with the gas generation sequence of various faults, and improves the ability to identify the progression of thermal fault levels, the development of discharge faults, and the evolution direction of thermoelectric composite boundary samples.

[0015] Further, determining the gating coefficient of the gas-related edge based on the degree of consistency between the concentration response characteristics and the gas production rate response characteristics of the gas-related edge includes: determining the difference between the concentration response characteristics and the gas production rate response characteristics of the gas-related edge; wherein the difference is used to characterize the degree of inconsistency between the concentration response characteristics and the gas production rate response characteristics; determining the gating coefficient of the gas-related edge based on the difference; wherein the value of the gating coefficient decreases as the difference increases, so as to suppress the influence of the gating fusion characteristics of the gas-related edge on the fault diagnosis result when the concentration response characteristics and the gas production rate response characteristics are inconsistent.

[0016] In the above scheme, the value of the gating coefficient decreases as the difference between the concentration response characteristics and the gas production rate response characteristics on the same gas correlation edge increases, so that the degree of gating adjustment and the degree of inconsistency between the two types of responses form a quantitative correspondence. The greater the difference, the more likely the gas correlation edge is to correspond to isolated rate peaks, detection noise, or short-term fluctuations with inconsistent mechanisms. The influence of its gating fusion characteristics on the fault diagnosis results is suppressed to a greater extent. Meanwhile, gas correlation edges with smaller differences are retained and strengthened as effective fault evidence that the concentration accumulation and fault development speed are mutually corroborated. The diagnosis results are mainly supported by the mechanism-consistent gas correlation relationship, which improves the stability of the diagnosis results under noise samples and complex operating conditions.

[0017] Further, determining the transformer fault diagnosis result based on the gating fusion features of each gas-related edge includes: determining the edge weight of each gas-related edge in the fault stage mechanism diagram; wherein the edge weight is used to characterize the degree of matching between the corresponding gas-related edge and the current fault stage, and the larger the edge weight, the higher the degree of matching between the gas-related edge and the current fault stage; weighting and aggregating the gating fusion features of each gas-related edge according to the edge weight of each gas-related edge to obtain fault discrimination features; and determining the transformer fault diagnosis result based on the fault discrimination features.

[0018] In the above scheme, the gated fusion features of each gas-related edge are weighted and aggregated according to their matching degree with the current fault stage to form fault discrimination features. The larger the edge weight, the more the gas correlation represented by the gas-related edge matches the current fault stage, and the larger its proportion in the fault discrimination features. This makes the fault discrimination features mainly composed of gas correlations that match the current fault stage, while edge features that are weakly correlated with the current fault stage only play an auxiliary role. The fault discrimination criteria at the sample level thus maintain a consistent emphasis structure with the fault stage mechanism, avoiding the dilution of key fault information caused by the indiscriminate aggregation of each gas-related edge, and improving the pertinence of fault classification.

[0019] Furthermore, the step of obtaining the concentration data of multiple characteristic gases in transformer oil at multiple sampling times includes: obtaining the original concentration sequence of each characteristic gas in transformer oil; performing missing value completion and abnormal spike suppression on the original concentration sequence to obtain a preprocessed concentration sequence; moving the time window along the chronological order of the sampling times and extracting the concentration data within each time window to obtain the concentration data of each characteristic gas at multiple consecutive sampling times.

[0020] In the above scheme, the original concentration sequence is filled with missing values ​​and suppressed for abnormal spikes, and then truncated by a moving time window into concentration data at multiple consecutive sampling times. This eliminates the damage to the integrity of the concentration sequence caused by sampling interruptions and data spikes, so that the concentration flow input to the dual-stream network and the gas production rate calculated by it are based on continuous and smooth data. The window truncation method also ensures that the trend of gas production rate change and the concentration evolution relationship can be stably extracted within a fixed time range, providing data quality assurance for the reliability of subsequent mechanism side response features.

[0021] Further, determining the gas production rate of each characteristic gas based on the concentration change of each characteristic gas between adjacent sampling times includes: determining the absolute gas production rate of each characteristic gas based on the concentration change of each characteristic gas between adjacent sampling times; wherein the absolute gas production rate is used to characterize the increase in gas concentration per unit time; and determining the relative gas production rate of each characteristic gas based on the concentration data of each characteristic gas at the later sampling time and the concentration data at the previous sampling time in two adjacent sampling times; wherein the relative gas production rate is used to characterize the magnitude of change in gas concentration relative to the concentration value at the previous sampling time.

[0022] In the above scheme, the gas production rate includes both the absolute gas production rate, which characterizes the increase in gas concentration per unit time, and the relative gas production rate, which characterizes the change in gas concentration relative to the previous sampling time. This allows the rate stream to reflect both the absolute speed of gas generation and the relative change of gas concentration relative to its own baseline concentration. The rapid relative growth of low-concentration gas and the stable absolute growth of high-concentration gas can both be effectively characterized by the rate stream. This avoids the problem of some gas fault development information being masked by using only a single rate form, and improves the completeness of the rate stream in depicting the fault development dynamics of characteristic gases of different concentration levels.

[0023] Furthermore, the method further includes: constructing training samples; wherein the training samples are used for end-to-end training of the two-stream network; The construction of training samples includes: using the concentration data of real samples of dissolved gas in transformer oil as the concentration data of the fault evolution sequence at the final sampling time; generating concentration data of multiple sampling times before the final sampling time according to the evolution law of gas concentration with fault development, and determining the corresponding gas production rate according to the concentration change between adjacent sampling times to obtain the simulation time series sample; and determining the fault category label of the real sample as the fault category label of the simulation time series sample.

[0024] In the above scheme, the concentration data of real samples is used as the endpoint of the fault evolution sequence, and the concentration data and gas production rate of the previous sampling time are generated according to the evolution law of gas concentration with fault development. Under the condition that the publicly available DGA data only contains static samples and does not have continuous online monitoring time series, it provides simulated training samples with time series dimension for end-to-end training of dual-stream network. Moreover, the fault category label of the simulated time series sample directly adopts the fault category label of the real sample. The reliability of the label is not affected by the time series generation process. This enables the concentration stream processing, rate stream processing, mechanism graph construction, mechanism side response feature extraction and gating fusion to be jointly optimized on the time series data with accurate labels, thus ensuring the effectiveness of dual-stream network training.

[0025] Other features and advantages of the present invention will be set forth in the following description. Attached Figure Description

[0026] Figure 1 This is a flowchart illustrating the transformer DGA diagnostic method based on a mechanism graph-gated two-stream network provided in an embodiment of the present invention. Figure 2 This is a flowchart illustrating a transformer DGA diagnostic method based on a mechanism graph-gated dual-stream network in a specific application scenario provided by an embodiment of the present invention. Figure 3 This is a schematic diagram of the concentration-rate dual-stream input and fault stage mechanism diagram generation process provided in an embodiment of the present invention; Figure 4 This is a schematic diagram of the feature extraction process based on the fault stage mechanism side response provided in an embodiment of the present invention; Figure 5 This is a schematic diagram of the mechanism-based edge response consistency gating fusion and fault classification process provided in an embodiment of the present invention; Figure 6 The simulation evolution curves of five DGA gases for typical high-temperature overheating samples provided in this embodiment of the invention; Figure 7 The simulation evolution curves of five DGA gases for typical low-energy discharge samples provided in the embodiments of the present invention; Figure 8 The simulation evolution curves of five DGA gases for typical high-energy discharge samples provided in the embodiments of the present invention are shown. Figure 9 The final fault classification confusion matrix of the transformer DGA diagnostic method based on mechanistic graph gated two-stream network provided in the embodiments of the present invention on the simulated DGA time series dataset. Detailed Implementation

[0027] The present application will now be described in further detail with reference to the accompanying drawings.

[0028] Existing diagnostic methods for dissolved gases in oil can be broadly categorized into two types based on their technological evolution: traditional diagnostic methods based on gas ratios or regional divisions, and data-driven intelligent diagnostic methods. The implementation methods and inherent limitations of each type are described below: (1) Traditional diagnostic methods are represented by the three-ratio method, the characteristic gas method, and the Duval triangle method (also known as the David's triangle method). The basic idea is to pre-set several gas ratio codes, gas content thresholds, or gas ratio coordinate regions based on the fault mechanism and operating experience, and then compare the gas detection value of the sample to be tested with the preset rules, and determine the fault type according to the code combination or region it falls into. The discrimination basis of this method is a fixed threshold, a fixed ratio, and a fixed region boundary. However, in the actual operation of transformers, gas generation is affected by a variety of factors such as oil temperature, load, oil quality, and equipment structure. The gas combination of the same fault type presents different on different equipment. When the gas ratio of the sample is exactly located at the code boundary or region boundary, the fixed discrimination rule cannot make a stable distinction between similar samples on both sides of the boundary. It is easy to make misjudgments when facing complex operating conditions, boundary samples, and multiple fault superposition situations.

[0029] (2) Data-driven intelligent diagnostic methods employ models such as support vector machines, backpropagation neural networks, and random forests to learn the nonlinear mapping relationship between gas content and fault category from labeled samples, overcoming the limitations of traditional methods that rely on manually set rules. In recent years, graph neural networks have been introduced into this field. Their approach is to construct a gas relationship graph using characteristic gases as graph nodes and mine the correlation features between gases through information propagation on the graph. Temporal networks are used to process concentration sequences obtained from continuous monitoring to capture the changing patterns of gas content over time. However, the gas relationship graphs in existing graph neural network schemes are mostly calculated from the statistical correlation between samples or fitted by learnable adjacency matrix parameters during training. The graph structure itself does not carry prior information about the gas generation mechanism under overheating, discharge, and overheating-discharge transition stages, and the correlation between gases cannot be adaptively adjusted according to the fault stage of the sample.

[0030] At the object level of feature extraction, most existing solutions use the features of individual gas nodes or the global features of the entire sample as modeling units, extracting only the concentration or concentration change of each gas. The cooperative evolution information contained in the gas-related edges is not directly characterized. Specifically, the cooperative relationship of related gases in the direction of concentration accumulation, the dynamic correlation in the change of gas production rate, and the development law of gas combination as the fault stage progresses cannot be expressed by single-node features or global features, resulting in insufficient basis for distinguishing adjacent fault levels, boundary samples, and complex operating condition samples.

[0031] At the level of fusion of concentration and gas production rate information, existing dual-flow structures typically extract concentration and rate features separately and then simply stitch them together at the overall feature level, or use global gating to uniformly weight the two types of features. This type of fusion method applies to the entire sample feature set, failing to pinpoint specific gas correlations or distinguish whether the concentration response and gas production rate response are consistent for the same gas correlation. When transient sensor interference or oil sample acquisition errors cause isolated spikes in the gas production rate of individual gases, or when detection noise and short-term abnormal fluctuations cause non-fault-related changes in gas concentration, existing fusion methods still treat these fluctuations as valid evidence for judgment. They fail to suppress their contribution based on the corroborating relationship between concentration accumulation and fault development speed, ultimately leading to insufficient stability of diagnostic results under complex operating conditions, noisy samples, and boundary samples.

[0032] In view of this, embodiments of the present invention provide a transformer DGA diagnostic method based on a mechanism graph-gated dual-flow network. This method organizes gas concentration data and gas generation rate into concentration streams and rate streams, respectively, and inputs them into a dual-flow network. It extracts dual-flow response features using gas-related edges in the fault stage mechanism graph as units and performs gated fusion based on consistency. This expands the fault discrimination criterion from the static content of a single gas to the co-evolutionary relationship between gas combinations in two dimensions: concentration accumulation level and fault development speed. This improves the stability and reliability of transformer DGA fault diagnosis under complex operating conditions, noise samples, and boundary samples. Furthermore, the fault stage mechanism graph is constructed based on the fault stage mechanism, where gas-related edges represent corresponding characteristic gases. The correlation between the entities conforms to the fault stage mechanism, so that the feature extraction object directly corresponds to the gas generation mechanism in the fault development process such as overheating and discharge. The network's extraction and fusion process of concentration response and gas production rate has a clear physical meaning, which improves the interpretability of the diagnostic results. On the other hand, the gating coefficient is determined according to the consistency between the concentration response feature and the gas production rate response feature on the same gas correlation edge. When the two types of responses support each other, the feature contribution of the gas correlation edge is preserved. When the gas production rate has an isolated spike, or the gas concentration has detection noise or short-term abnormal fluctuations, the feature contribution of the gas correlation edge is suppressed. This avoids non-fault fluctuations being mistakenly regarded as evidence of fault development and participates in the discrimination, thus reducing the risk of misjudgment.

[0033] The following section introduces the application scenarios of the transformer DGA diagnostic method based on the mechanism graph-gated two-stream network: The aforementioned transformer DGA diagnostic method based on mechanism graph-gated dual-current networks is applied to the condition monitoring and latent fault diagnosis scenarios of oil-immersed power transformers. In this scenario, the detection data of dissolved gases in transformer oil are obtained through online monitoring devices or periodic oil sample analysis. The detection targets cover the concentration values ​​of characteristic gases such as hydrogen, methane, acetylene, ethylene, and ethane at multiple sampling times. The power grid operation and maintenance department needs to perform real-time or near-real-time analysis on the continuously collected DGA data in order to identify the fault type and fault stage in a timely manner as latent faults such as overheating and discharge develop from their early stages to their severity, providing a basis for transformer maintenance decisions and operation mode adjustments. Due to the instantaneous interference of sensors, oil sample collection errors, and non-fault gas changes caused by load fluctuations in the field detection environment, the collected data is often accompanied by isolated rate spikes, detection noise, and short-term abnormal fluctuations. Furthermore, there are a large number of boundary samples in the actual samples that are at the junction of adjacent fault levels and have multiple fault mechanisms superimposed, which places high demands on the anti-interference ability and fault stage identification ability of the diagnostic method. The transformer DGA diagnostic method based on mechanism graph-gated dual-flow network provided in this invention is precisely designed to accurately diagnose the operating states of transformers, such as normal state, low-temperature overheating, medium-temperature overheating, high-temperature overheating, low-energy discharge, and high-energy discharge, by performing dual-flow modeling of concentration flow and velocity flow on continuously sampled DGA data, extracting gas-related edge features under fault stage mechanism constraints, and performing mechanism consistency gating fusion.

[0034] like Figure 1 As shown, this embodiment of the invention provides a transformer DGA diagnostic method based on a mechanism graph-gated two-stream network, including: Step S110: Obtain the concentration data of various characteristic gases in transformer oil at multiple sampling times.

[0035] Optionally, step S110 includes: obtaining the original concentration sequence of each characteristic gas in the transformer oil; performing missing value completion and abnormal spike suppression on the original concentration sequence to obtain the preprocessed concentration sequence; moving the time window along the chronological order of sampling time and extracting the concentration data within each time window to obtain the concentration data of each characteristic gas at multiple consecutive sampling times.

[0036] The concentration of dissolved gases in transformer oil can be obtained through continuous sampling of the transformer during operation using an online monitoring device, or through periodic oil sampling and chromatographic analysis. The original concentration sequence refers to a one-dimensional concentration data sequence formed by arranging the same characteristic gas in chronological order of sampling time. The main characteristic gas mentioned above includes hydrogen. methane Acetylene ethylene ethane Five types of characteristic gases are identified, each corresponding to an independent original concentration sequence. In online monitoring, the time interval between adjacent sampling times is determined by the sampling cycle of the monitoring device; in periodic analysis, the time interval between adjacent sampling times is determined by the analysis cycle. The concentration values ​​at each sampling time collectively reflect the evolution of the characteristic gas content over time.

[0037] During actual data acquisition, communication interruptions, temporary faults, or missed detections in laboratory tests can cause missing concentration values ​​at certain sampling times. Factors such as instantaneous sensor interference, oil sample contamination, or environmental electromagnetic interference can cause sudden spikes in concentration values ​​at certain sampling times, deviating from the normal range of variation; these spikes are called abnormal glitches. Missing value completion is used to estimate and fill in the concentration values ​​at missing sampling times. Available completion methods include linear interpolation of concentration values ​​from adjacent sampling times, filling with the average of concentration values ​​from several neighboring sampling times, or extrapolating based on the previous concentration trend of the gas. The completed concentration sequence is temporally complete. Abnormal glitch suppression is used to identify and correct sudden values ​​that deviate from the normal fluctuation range. Available suppression methods include replacing sampling values ​​exceeding a preset amplitude range with concentration values ​​from neighboring sampling times, or smoothing the concentration sequence using a sliding median filter. The preprocessed concentration sequence obtained after missing value completion and abnormal glitch suppression meets the requirements for subsequent time series analysis in terms of both temporal dimension and numerical reliability.

[0038] The preprocessed concentration sequence is transformed into several window samples by a sliding time window. The sliding time window refers to a fixed-length truncation interval that moves sequentially along the sampling time sequence. In this embodiment, each time window contains concentration data from 10 consecutive sampling times. Adjacent time windows move one sampling point backward; that is, after each truncation, the starting position of the window is shifted backward by one sampling time before the next truncation. Since only one sampling point moves between adjacent windows, the actual time interval between adjacent windows is equal to the sampling interval of the original DGA data, thus ensuring that the window samples are uniformly distributed on the time axis and can continuously track the evolution of gas concentration. To ensure that a single time window can reflect the trend of gas concentration change and the change law of gas production rate, the number of effective sampling points within the window should be no less than 6. Effective sampling points refer to sampling points within the window that have true detection values ​​or reliable concentration values ​​after completion. When the number of effective sampling points within a window is less than 6, the data within that window is insufficient to support a reliable estimate of the change trend, and therefore, no sample is generated for that time window. After being captured, each window corresponds to a set of concentration data of five characteristic gases at 10 consecutive sampling times, which serves as the data unit for subsequently determining the gas production rate and constructing the concentration stream and rate stream.

[0039] Step S120: Determine the gas production rate of each characteristic gas based on the concentration change of each characteristic gas between adjacent sampling times.

[0040] Understandably, gas concentration reflects the cumulative effect of the pyrolysis products of insulating oil and insulating paper under the long-term influence of fault energy. The concentration level only indicates the current level of internal transformer deterioration, not the rate at which it is progressing. In the actual fault evolution of a transformer, the gas generation rate varies significantly depending on the fault type and stage of development. For example, the generation rate of thermal fault-related gases gradually accelerates as the fault progresses from low-temperature overheating to high-temperature overheating. In discharge faults, especially high-energy discharges, characteristic discharge gases such as acetylene accumulate rapidly in a short period. Relying solely on the static concentration at a single sampling point is insufficient to distinguish between three different states—slow development, stable maintenance, and accelerated deterioration—where the concentration is at the same level. It also makes it difficult to identify the short-term abrupt changes characteristic of discharge faults. This invention determines the gas generation rate of each characteristic gas by measuring the concentration change between adjacent sampling times. This provides dynamic information, beyond the gas accumulation degradation level represented by concentration, to depict the speed of fault development and the degree of short-term abrupt changes. This allows subsequent diagnosis to determine both the current stage of the fault based on the degree of gas accumulation and the direction of fault evolution based on the rate of gas generation. It provides another dimension of data support for constructing fault stage mechanism diagrams, extracting mechanism-side response features, and determining the consistency between concentration and rate. An optional scheme for determining the gas generation rate of the aforementioned characteristic gases is described below: Optionally, step S120 includes: determining the absolute gas production rate of each characteristic gas based on the concentration change of each characteristic gas between adjacent sampling times; wherein the absolute gas production rate is used to characterize the increase in gas concentration per unit time; and determining the relative gas production rate of each characteristic gas based on the concentration data of each characteristic gas at the later sampling time and the concentration data at the previous sampling time in two adjacent sampling times; wherein the relative gas production rate is used to characterize the magnitude of change in gas concentration relative to the concentration value at the previous sampling time.

[0041] The absolute gas production rate mentioned above refers to the increase in the concentration of a certain characteristic gas per unit time. It is calculated by dividing the concentration change between adjacent sampling times by the time interval between those adjacent sampling times. The physical meaning of the result is the rate at which the gas concentration increases over time. In applications involving periodic oil sample testing, the time interval is usually measured in days or months, and the absolute gas production rate is correspondingly expressed as the daily or monthly concentration increase. In applications involving continuous online monitoring, the time interval is taken as the sampling cycle of the monitoring device, and the absolute gas production rate represents the rate of concentration increase between two adjacent samplings. The absolute gas production rate directly characterizes the gas production speed with the absolute increase in concentration; a larger value indicates that more gas is generated per unit time, and the faster the fault energy acts on the insulating medium.

[0042] The relative gas production rate mentioned above refers to the change in gas concentration relative to the concentration value at the previous sampling time. It is calculated by dividing the concentration change between two adjacent sampling times by the concentration value at the previous sampling time, and then dividing by the corresponding time interval when a time correlation is needed. The result reflects the relative increase in gas concentration from its existing level. Unlike the absolute gas production rate, the relative gas production rate does not directly focus on the absolute increase in concentration, but rather on the proportional relationship between the increase and the baseline value. For gases such as methane, which already have a certain background concentration during normal transformer operation, even if the absolute increase is large, the relative change may be relatively gradual. For gases such as acetylene, which are present in extremely low concentrations under normal conditions, once a fault occurs, their concentration relative to the original baseline value can increase several times or even tens of times. This drastic change is fully reflected in the relative gas production rate.

[0043] The absolute and relative gas production rates exhibit complementary sensitivity to different gas concentration levels. The absolute gas production rate tends to highlight the growth of high-concentration gases, while rapid changes in low-concentration gases are less pronounced in absolute values ​​and easily masked. The relative gas production rate, using the concentration baseline of each gas as a reference, can accurately measure both the rapid relative growth of low-concentration gases and the slow relative changes of high-concentration gases, aligning with their respective baseline values. The rate flow, comprised of both absolute and relative gas production rates, can simultaneously reflect the absolute speed and relative drasticness of fault development. This allows subsequent steps to characterize the fault development speed and the degree of short-term abrupt changes without relying on a single gas concentration level, providing a complete data foundation for calculating the gas production rate response in the mechanistic side-response characteristics.

[0044] Step S130: Concentration stream is constructed using the concentration data of each characteristic gas, and velocity stream is constructed using the gas production rate of each characteristic gas. The concentration stream and velocity stream are then input into the dual-stream network.

[0045] The aforementioned concentration stream refers to an input data stream organized from the concentration sequences of various characteristic gases within the current time window. Its data content covers the concentration values ​​of characteristic gases such as hydrogen, methane, acetylene, ethylene, and ethane at each consecutive sampling time within the window. Since the accumulation of gas concentration is the result of the continuous decomposition of insulating oil and insulating paper under thermal and electrical stress, the concentration stream carries information about the level of degradation already achieved within the transformer. Higher concentration values ​​for each gas in the concentration stream indicate a longer duration or greater intensity of the corresponding gas generation process; therefore, the concentration stream is used to characterize the level of cumulative gas degradation. In terms of data organization, the concentration stream can be arranged as a two-dimensional matrix with sampling times as rows and characteristic gases as columns. Each row in the matrix corresponds to the concentration values ​​of all characteristic gases at a given sampling time, and each column corresponds to the concentration sequence of one characteristic gas within the window.

[0046] The aforementioned rate stream refers to the input data stream organized by the gas production rates of each characteristic gas. Its data content covers the values ​​of absolute and relative gas production rates between adjacent sampling times within a window. The gas production rate characterizes the instantaneous speed of gas generation rather than its cumulative result. The rate stream shows an upward trend when the fault is in an accelerated development phase, fluctuates around a lower level when the fault is in a stable maintenance phase, and short-term abrupt changes caused by discharge-type faults will form obvious pulse-like changes in the rate stream. Therefore, the rate stream is used to characterize the speed of fault development and the degree of short-term abrupt changes. The data organization of the rate stream is consistent with that of the concentration stream, i.e., a two-dimensional matrix is ​​constructed with the time interval between sampling times as rows and the characteristic gas as columns, thus ensuring a point-to-point correspondence between the concentration stream and the rate stream in the time dimension. This provides an aligned data basis for subsequent comparison of the consistency between the concentration response and the gas production rate response on the same gas-related edge.

[0047] A dual-stream network refers to a network structure with two parallel feature processing channels that can fuse the outputs of the two channels. One channel receives the concentration stream, and the other receives the rate stream. After entering the network through their respective channels, the concentration and rate streams participate in the extraction of concentration response features and gas production rate response features on gas-related edges in subsequent steps. The features extracted by the two channels are weighted and combined according to the principle of mechanistic consistency in the gating fusion stage. The reason for adopting a dual-stream structure in this embodiment of the invention is that the physical meaning, numerical dimensions, and variation characteristics of concentration and rate information are different. If the two are directly mixed at the input, the correspondence between the concentration accumulation level and the fault development rate will be destroyed. By carrying the data separately through two channels and fusing them under mechanistic constraints, the network can learn the evolution laws of the two types of information separately, and can also use the relationship between whether the two support each other to determine the reliability of fault evidence during the fusion stage.

[0048] Step S140: Using a dual-flow network, a fault stage mechanism graph is constructed based on the fault stage mechanism, with multiple characteristic gases as graph nodes; wherein, the gas association edges in the fault stage mechanism graph are used to characterize the association relationship between the corresponding characteristic gases that conforms to the fault stage mechanism.

[0049] It is understandable that the generation correlations between characteristic gases are not fixed in the different stages of a transformer's internal fault development from normal to overheating and discharge, but rather exhibit regular changes as the fault stage evolves. In the process of overheating faults progressing from low to high temperatures, methane, ethylene, and other thermal fault-related gases successively become the main products, and their concentration increase synergistically changes with different temperature ranges. In discharge faults, the co-generation relationship between hydrogen and acetylene shows different intensities depending on the discharge energy level. In the composite region at the boundary between high-temperature overheating and strong discharge, transitional coupling correlations also appear between thermal fault gases and discharge characteristic gases. If a fixed empirical correlation graph is used to describe the gas relationships under all fault states, the edges in the graph can only reflect the correlation pattern of a certain typical stage. When the sample is in other stages or transitional regions, the graph structure mismatches with the actual gas generation relationships. If a correlation graph obtained from sample statistical correlation or fully data-driven learning is used, the edges in the graph only reflect the strength of the correlation at the data level and cannot reflect the causal mechanism of gas generation at each stage. Statistical correlation may also misjudge synchronous fluctuations of unrelated gases as mechanistic correlations. In view of this, embodiments of the present invention construct a fault stage mechanism diagram. This mechanism diagram determines the associated edges between gas nodes based on the gas generation mechanisms of stages such as overheating, discharge, and overheating-discharge transition. Each gas associated edge in the diagram represents the association relationship between corresponding feature gases that conforms to the fault stage mechanism. On the one hand, this provides a mechanistically reliable extraction object for subsequent extraction of concentration response and gas production rate response using gas associated edges as units, so that feature extraction is no longer targeted at isolated gases or fully connected gas combinations. On the other hand, it constrains the information propagation range between gases by the fault mechanism, and gas associations unrelated to the current fault stage will not participate in feature organization, thereby providing a graph structure basis adapted to the fault stage for fault diagnosis. An optional construction scheme for the fault stage mechanism diagram is described below: Optionally, step S140 includes: using multiple characteristic gases as graph nodes, constructing an overheating mechanism subgraph based on the gas generation mechanism of the overheating fault stage; constructing a discharge mechanism subgraph based on the gas generation mechanism of the discharge fault stage; constructing an overheating-discharge transition mechanism subgraph based on the gas generation mechanism of the overheating-discharge transition stage; determining the combination weights of the overheating mechanism subgraph, the discharge mechanism subgraph, and the overheating-discharge transition mechanism subgraph based on the concentration data and gas production rate of each characteristic gas; and weighting the overheating mechanism subgraph, the discharge mechanism subgraph, and the overheating-discharge transition mechanism subgraph according to the combination weights to obtain the fault stage mechanism graph.

[0050] The above overheating mechanism sub-diagram is denoted as follows: This subgraph uses five characteristic gases—hydrogen, methane, acetylene, ethylene, and ethane—as nodes, with edges set according to the cracking mechanism of insulating oil and insulating paper under thermal stress. The types and proportions of gases generated during low-temperature, medium-temperature, and high-temperature overheating of the insulating medium exhibit stage-specific differences. At low temperatures, methane generation is dominant; as temperature increases, ethylene generation gradually strengthens and becomes dominant at medium and high temperatures; ethane generation is related to the overall thermal degradation degree of the oil. The overheating mechanism subgraph characterizes the stage-specific relationships between gases such as methane and ethylene during low-temperature, medium-temperature, and high-temperature overheating. Edges in the subgraph connect gas nodes that exhibit sequential generation or synergistic growth relationships during the development of thermal faults, thus reflecting the evolution path of the thermal fault gases within the graph structure.

[0051] The above discharge mechanism sub-diagram is denoted as follows: This subgraph uses the same five gas nodes, and its edges are set according to the cracking mechanism of the insulating medium under electrical stress. The energy level of the discharge fault determines the gas generation characteristics. During low-energy discharge, the insulating oil undergoes local ionization, with hydrogen generation being the main process. During high-energy discharge, the high temperature of the arc causes a large amount of acetylene to be generated. The co-generation of hydrogen and acetylene constitutes the typical gas correlation characteristics of discharge faults. The discharge mechanism subgraph is used to characterize the correlation of gases such as hydrogen and acetylene in low-energy and high-energy discharge processes. The edges in the subgraph connect gas nodes that change synergistically during the discharge development process.

[0052] The above overheating-discharge transition mechanism is denoted as follows: This subgraph addresses the complex situation at the boundary between high-temperature overheating and strong discharge, and is used to characterize the transitional relationships between gases such as hydrogen, ethylene, and acetylene in the samples of high-temperature overheating, strong discharge, and thermoelectric composite boundary. Its edges simultaneously cover the gas relationships on both the thermal fault evolution path and the discharge evolution path, so as to depict the mixed characteristics of gas combinations when the two mechanisms are superimposed.

[0053] It is understandable that the overheating mechanism sub-diagram, discharge mechanism sub-diagram, and overheating-discharge transition mechanism sub-diagram described above depict the gas correlation relationships under a single fault mechanism or a typical transition mechanism. However, the samples to be diagnosed in actual transformer operation do not always correspond exactly to a certain pure fault stage. The samples may be in the middle of the development from low-temperature overheating to medium-temperature overheating, or they may be in the transition region from thermal fault to thermoelectric combined fault. Their gas correlation relationships simultaneously have the characteristics of multiple stages. If the samples are directly and rigidly classified into a certain fault stage according to preset rules and the corresponding mechanism sub-diagram is selected separately, any deviation in stage identification will lead to errors in the edge connection structure of the entire mechanism diagram, and the rigid stage division cannot achieve a continuous transition between adjacent stages. The aforementioned combined weights are assigned based on the proximity of each of the three subgraphs to the current sample state, with a sum of weights of 1. This ensures that the final generated fault stage mechanism diagram is a weighted superposition of the three subgraphs according to their proximity. The more pronounced the overheating characteristics of the sample, the greater the contribution of the overheating mechanism subgraph to the combination. When the sample is in the boundary region, the contribution of the transition mechanism subgraph increases accordingly. Therefore, the gas correlations of different fault stages can be integrated into the same mechanism diagram in a continuous proportion that matches the sample state. By determining the combined weights, the fault stage mechanism diagram transforms from a static, single-stage structure to a structure that adaptively generates according to the current sample state. The matching between the diagram structure and the actual fault stage of the sample no longer relies on manually preset stage discrimination rules, but is driven by the sample's own concentration data and gas production rate. An optional scheme for determining the combined weights is described below: Optionally, determining the combined weights of the superheating mechanism sub-graph, the discharge mechanism sub-graph, and the superheating-discharge transition mechanism sub-graph based on the concentration data and gas production rate of each characteristic gas includes: determining the current concentration and concentration change of each characteristic gas based on the concentration data of each characteristic gas at multiple sampling times; determining the average gas production rate and gas production rate change trend of each characteristic gas based on the gas production rate of each characteristic gas between adjacent sampling times; determining the key gas ratio based on the current concentration of each characteristic gas; and determining the combined weights of the superheating mechanism sub-graph, the discharge mechanism sub-graph, and the superheating-discharge transition mechanism sub-graph based on the current concentration, concentration change, average gas production rate, gas production rate change trend, and key gas ratio of each characteristic gas.

[0054] The current concentration refers to the concentration value of each characteristic gas at the latest sampling time within the current time window, which directly reflects the cumulative level of each gas at the current time of the sample. The concentration change refers to the degree of change in the concentration of each characteristic gas within the current time window. It can be represented by the concentration difference between the latest and earliest sampling times within the window, or by the normalized result of the concentration change at adjacent sampling times within the window. The current concentration and concentration change characterize the cumulative evolution state of the gas from both static and dynamic perspectives. Samples with excessively high concentrations but slowing changes, and samples with moderate concentrations but continuously increasing changes, correspond to different fault development trends. Combining these two factors provides a cumulative dimension for judging the fault stage.

[0055] The average gas production rate refers to the average gas production rate of each characteristic gas within the current time window across multiple adjacent sampling times. The absolute gas production rate is measured by the concentration increase per unit time, while the relative gas production rate is measured by the proportion of change in concentration relative to a baseline value. The average gas production rate reflects the overall level of gas generation rate within the window. The gas production rate trend refers to the overall direction of the gas production rate sequence within the current time window. It is determined by performing a linear fit for each characteristic gas, using the sampling order of each time interval within the window as the independent variable and the corresponding gas production rate as the dependent variable. The slope of the fitted line represents the rate change trend of that gas. A positive slope indicates an overall increase in the gas production rate, suggesting the fault is in an accelerated development phase; a negative slope indicates an overall decrease in the gas production rate, suggesting the fault is slowing down; a slope close to zero indicates a generally stable gas production rate, suggesting the fault is in a relatively stable state. The above linear fit can be achieved using the least squares method, with the slope calculated as the ratio of the covariance of each time interval number to the gas production rate value to the variance of the time interval number. The average gas production rate and the trend of gas production rate change work together. The former describes the current speed of fault development, while the latter describes the direction of acceleration or deceleration of fault development. Together, they provide a basis for the speed dimension for fault stage judgment.

[0056] The critical gas ratio refers to the ratio between the current concentrations of two characteristic gases that indicate the type of fault. Optional critical gas ratios include the acetylene to ethylene ratio, the methane to hydrogen ratio, and the ethylene to ethane ratio. The acetylene to ethylene ratio reflects the relative strength of discharge energy and the degree of thermal fault; the methane to hydrogen ratio reflects the relative relationship between low-energy discharge and thermal fault in oil; and the ethylene to ethane ratio reflects the temperature level of the thermal fault. These ratios are consistent with the gas combinations used in traditional three-ratio diagnostics, allowing the incorporation of gas ratio patterns accumulated through long-term operational experience into the characterization of sample states. The critical gas ratio is calculated directly from the current concentration, without relying on time-related information. Its indication of fault type is complementary to dynamic information such as concentration changes and gas production rates.

[0057] The current concentration, concentration change, average gas production rate, gas production rate trend, and key gas ratios mentioned above are normalized and then concatenated to form the current sample state vector. Normalization is used to eliminate differences in dimensions and numerical ranges between different physical quantities, preventing concentration terms with larger numerical ranges from overshadowing ratio terms with smaller numerical ranges in the state vector. Based on the current sample state vector... Adaptive weights for generating three mechanistic subgraphs: ; In the formula, and These are learnable parameters; , and These represent the combined weights of the overheating mechanism subgraph, the discharge mechanism subgraph, and the overheating-discharge transition mechanism subgraph, respectively, and their sum is 1. If the current sample better matches the characteristics of an overheating fault, then... Larger; if the current sample better matches the characteristics of a discharge fault, then Larger; if the current sample is located in the boundary region between high temperature overheating and strong discharge, then The parameters for generating the combined weights are relatively large. The parameters are updated as the network is trained end-to-end, enabling the weight allocation rules to be learned automatically from labeled samples, without relying on manually set stage discrimination thresholds.

[0058] Based on this, the adjacency matrix of the fault stage mechanism corresponding to the current time window is obtained: ; In the formula, This represents the adjacency matrix of the fault stage mechanism, adaptively generated from the current sample state. This is the sequence number of the current time window. This adjacency matrix is ​​neither a fixed empirical graph nor a gas relationship graph simply obtained from sample statistical correlation. Instead, it is dynamically generated by combining mechanism subgraphs for different fault stages based on the current DGA sample state. In this way, embodiments of the present invention can adaptively adjust the information propagation relationships between different gases as the fault stage changes.

[0059] Step S150: Using a dual-stream network, for each gas-related edge in the fault stage mechanism diagram, extract the concentration response features of the gas-related edge based on the gas concentration data at both ends of the gas-related edge, and extract the gas production rate response features of the gas-related edge based on the gas production rate at both ends of the gas-related edge.

[0060] Understandably, each gas-related edge in the aforementioned fault stage mechanism diagram represents a correlation between the gases at both ends that conforms to the fault stage mechanism. Whether this correlation truly exists in the current sample needs to be determined through two types of independent yet mutually corroborating evidence. The concentration response characteristic reflects the synergistic relationship between the associated gases in terms of concentration accumulation and change trends. When the fault develops according to the mechanism, the concentrations of the associated gases will show synergistic changes in the same direction and with related trends. The concentration response characteristic is precisely a quantitative characterization of this synergistic relationship. The gas production rate response characteristic reflects the dynamic correlation between the associated gases in terms of the fault development speed. When the fault is in a real development state, the gas production rate of the associated gases will show synchronous growth or slowdown. The gas production rate response characteristic is precisely a quantitative characterization of this dynamic correlation. If only concentration response features are extracted, the aforementioned dual-stream network can only obtain static results of gas accumulation, and cannot distinguish whether the coordinated changes in concentration originate from the continuous development of a fault or a historically residual stable state. If only gas production rate response features are extracted, the aforementioned dual-stream network can only obtain the instantaneous speed of gas generation; isolated rate spikes, detection noise, and short-term abnormal fluctuations will also appear as rate changes, making it impossible to determine whether they are supported by concentration accumulation. This embodiment of the invention, by determining two types of response features, ensures that each gas-related edge simultaneously possesses dual evidence in both the accumulation and velocity dimensions. Subsequent steps can then compare whether the two types of responses on the same related edge are consistent, thereby distinguishing between fault evolution evidence consistent with the mechanism and abnormal fluctuations inconsistent with the mechanism, providing a discriminant for mechanism consistency gating fusion.

[0061] The following describes one possible extraction scheme for the above concentration response features: The aforementioned concentration response feature refers to a feature vector calculated based on the concentration data of the gases at both ends of the gas-related edges in the fault stage mechanism diagram within the current time window. It characterizes the synergistic relationship between the related gases in terms of concentration accumulation and change trends. Unlike methods that extract concentration features for individual gas nodes, each component of the concentration response feature involves both gases at the ends of the related edge. The feature describes the relationship between gases rather than the gases themselves, which is the basic unit of feature extraction under the constraints of the fault stage mechanism. For gases... With gas The mechanism between them, and its concentration response characteristics are expressed as: ; In the formula, Represents gas With gas The concentration response characteristics of the mechanism between them , They represent gases With gas The normalized change in concentration within the current time window. Represents a symbolic function. Used to indicate whether the directions of gas concentration change at both ends are consistent; This indicates the correlation between the gas concentration sequences at both ends.

[0062] Normalized concentration change refers to the dimensionless quantity obtained by normalizing the concentration change of a gas within the current time window according to the concentration range of that gas. Normalization eliminates the influence of differences in concentration magnitude between different gases, allowing the changes in concentration of gases with high baseline concentrations, such as hydrogen and methane, to be compared and combined with the changes in concentration of gases with low baseline concentrations, such as acetylene, on the same scale. The first component of the above concentration response characteristic... This is the sum of the absolute values ​​of the normalized concentration changes of the gases at both ends, reflecting the overall intensity of the concentration changes at both ends of the associated edge. A larger value indicates a more dramatic concentration evolution of the gas combination within the window. The second component... The sign of the product of the changes in gas concentrations at both ends is denoted by . A positive product indicates that the concentrations of both gases are changing in the same direction, and the sign function is positive. A negative product indicates that the concentrations of the two gases are changing in opposite directions, and the sign function is negative. This component characterizes the consistency of the direction of change in the related gases. The third component... This is the absolute value of the difference between the normalized concentration changes of the two gases, reflecting the degree of difference in the intensity of the concentration changes of the two gases. The smaller the difference, the closer the change ranges of the two gases are, and the stronger the synergy. (Fourth component) The correlation coefficient of the concentration sequences of the two gases within the current time window can be calculated using the Pearson correlation coefficient. This component characterizes the similarity of the changes in the concentrations of the two gases over time from the perspective of sequence morphology. All four components mentioned above are dimensionless quantities.

[0063] The four components described above jointly describe the co-evolution relationship of associated gases from four aspects: intensity of change, consistency of change direction, difference of change amplitude, and sequence correlation. Their working principle is as follows: when a fault develops according to a corresponding mechanism, the two gases related to the mechanism should not only have large absolute intensity of concentration changes, but also consistent change directions, similar change amplitudes, and highly correlated concentration sequence shapes. All four components simultaneously exhibit value shapes consistent with the fault mechanism. Conversely, when the concentration changes of the two gases are opposite in direction, have significantly different amplitudes, or have unrelated sequence shapes, even if the concentration change of one gas is significant, the combination of the four components will exhibit a shape that does not conform to the mechanistic co-evolution relationship. Therefore, the concentration response characteristics transform the mechanistic judgment of "whether the gas combination co-evolves" into a computable vector form, providing a concentration dimension as a basis for subsequent consistency comparison with the gas production rate response characteristics on the same associated edge.

[0064] The following describes an optional extraction scheme for the above-mentioned gas production rate response characteristics: The aforementioned gas production rate response feature refers to a feature vector calculated based on the gas production rate sequence of the gases at both ends of the same gas-related edge within the current time window, using the same gas-related edge as the extraction object. It characterizes the dynamic correlation of related gases in terms of fault development speed. The concentration response feature answers the question of whether the cumulative changes of related gases are coordinated, while the gas production rate response feature answers the question of whether the generation rates of related gases are coordinated; these two correspond to the accumulation dimension and the velocity dimension in the fault evolution process, respectively. For gases... With gas The mechanism boundary between them, and its gas production rate response characteristics are expressed as: ; In the formula, Represents gas With gas The gas production rate response characteristics of the mechanism between them; , They represent gases With gas The average gas production rate within the current time window; , The trends of the gas production rate sequences at both ends are respectively represented. Represents a symbolic function. Used to indicate whether the gas production rate changes in the same direction at both ends. This indicates the synchronicity of the gas production rate sequences at both ends.

[0065] The average gas production rate refers to the average gas production rate between adjacent sampling times within the current time window. Its dimensions are consistent with the adopted gas production rate form: when using absolute gas production rate, it is measured as the concentration increase per unit time; when using relative gas production rate, it is measured as the proportion of change in concentration relative to the baseline value. The trend of the gas production rate sequence is determined by the slope of the linear fit of the gas production rate sequence within the window. That is, a linear fit is performed with the sampling order of each time interval within the window as the independent variable and the corresponding gas production rate as the dependent variable. The resulting slope represents the trend. The dimension of the slope is the ratio of the gas production rate unit to the sampling interval. A positive slope indicates that the overall gas production rate is increasing, a negative slope indicates that it is decreasing, and a slope close to zero indicates overall stability. The first component of the above gas production rate response characteristics... This is the sum of the average gas production rates at both ends of the associated edge, reflecting the overall level of gas generation rate at both ends. A larger value indicates a faster fault development rate corresponding to this gas combination. The second component... This represents the sign of the product of the changing trends in the gas production rates at both ends. A positive product indicates that the production rates of both gases are changing in the same direction, and the evolution of the fault development speed is consistent. A negative product indicates that the production rate of one gas is increasing while the other is decreasing, and the dynamic evolution directions are diverging. The third component... This is the absolute value of the difference between the average gas production rates at both ends, reflecting the degree of difference in the production rates of the two gases. (Fourth component) The correlation coefficient of the gas production rate sequence of the two gases within the current time window can be calculated using the Pearson correlation coefficient, which is a dimensionless quantity. This component characterizes the degree of synchronous fluctuation of the gas production rates of the two gases over time from the perspective of sequence morphology.

[0066] The gas production rate response feature and the concentration response feature maintain a corresponding component structure; both are four-dimensional vectors, and the fourth component uses a serial correlation coefficient. This structural correspondence provides a basis for dimensional alignment in subsequent calculations of the consistency differences between the two types of response features on the same correlation edge. From the perspective of working principle, in the actual fault development process, the two gases related to the mechanism not only accumulate synergistically in concentration, but their gas production rates should also exhibit a dynamic correlation in the same direction and synchronously. Although isolated rate spikes and short-term fluctuations caused by detection noise may cause significant values ​​in the gas production rate response component of a certain gas, they cannot make the directional consistency component and the synchronous component on the correlation edge simultaneously exhibit a form consistent with the mechanism. The gas production rate response feature, composed of four components, thus transforms the mechanistic correlation at the fault development speed level into a computable vector form, which, together with the concentration response feature, constitutes two pieces of evidence for mechanism consistency judgment.

[0067] Step S160: Using a dual-stream network, for each gas-related edge in the fault stage mechanism diagram, determine the gating coefficient of the gas-related edge based on the consistency between the concentration response characteristics and the gas production rate response characteristics of the gas-related edge, and fuse the concentration response characteristics and the gas production rate response characteristics of the gas-related edge based on the gating coefficient to obtain the gating fusion characteristics of the gas-related edge.

[0068] Understandably, while the concentration response features and gas production rate response features extracted from each gas association edge characterize the synergistic relationship between associated gases in terms of accumulation level and fault development speed, the reference value of the two types of features for fault diagnosis is not always equal. Their reliability depends on whether the two types of responses on the same association edge corroborate each other. When the concentration response and gas production rate response of a certain association edge show a consistent pattern, it indicates that the gas combination exhibits synergistic evolution consistent with the fault mechanism in both the dimensions of concentration accumulation and generation rate. The features of this association edge are reliable fault evidence and should be given a significant contribution in the fusion. When the two types of responses differ significantly, it indicates that the change in the rate dimension is not supported by the concentration accumulation level. This association edge is more likely to correspond to isolated rate spikes, detection noise, or short-term abnormal fluctuations. If it is indiscriminately included in the fusion with normal association edges, non-fault fluctuations will be treated as evidence of fault development and affect the diagnostic results. The aforementioned gating coefficient is precisely the adjustment factor that meets this discrimination requirement. It transforms the consistency between concentration response characteristics and gas production rate response characteristics into weights with values ​​between zero and one. By weighting and fusing the two types of response characteristics through the gating coefficient, the correlation edges with high consistency receive a coefficient close to one, thus preserving and enhancing their rate response; the correlation edges with low consistency receive a smaller coefficient, thus compressing their rate response. This achieves edge-level evidence screening based on mechanistic consistency, avoiding the shortcomings of the overall weighting method in locating the source of anomalies, and ensuring that the gating fusion characteristics entering the fault discrimination stage mainly consist of reliable fault evidence supported by mutual concentration and rate. The following describes an optional scheme for determining the aforementioned gating coefficient: Optionally, the above-mentioned transformer DGA diagnostic method based on mechanism graph gated dual-current network further includes: for each gas-related edge in the fault stage mechanism graph, extracting the stage progressive response term of the gas-related edge according to the concentration change and gas production rate of the gas at both ends of the gas-related edge and the preset fault stage progression direction corresponding to the gas-related edge; wherein, the stage progressive response term is used to characterize whether the change relationship of the gas at both ends of the gas-related edge conforms to the fault development law under the corresponding fault stage. The above-mentioned determination of the gating coefficient of the gas-related edge based on the degree of consistency between the concentration response characteristics and the gas production rate response characteristics of the gas-related edge includes: determining the gating coefficient of the gas-related edge based on the degree of consistency between the concentration response characteristics and the gas production rate response characteristics of the gas-related edge and the stage progressive response term of the gas-related edge.

[0069] The stage-progressive response term refers to a scalar calculated for each gas-related edge in the fault stage mechanism diagram, based on the changes in gas concentration and gas production rate at both ends, and the preset fault stage progression direction of that edge. It characterizes whether the changes in the gases at both ends conform to the gas generation sequence and fault development law under the corresponding fault stage. Gas generation corresponding to different fault stages exhibits a clear progressive law: as low-temperature overheating progresses to high-temperature overheating, the dominant position of gases such as methane and ethane gradually gives way to ethylene; as a discharge fault develops, the dominant position of hydrogen gradually gives way to acetylene. The stage-progressive response term transforms this mechanistic-level progressive law into a numerical quantity that can be examined one by one for each related edge. For gases... With gas The mechanism between them, the stage-progressive response term is represented as: ; In the formula, Represents gas With gas The progressive response items between stages, , They represent gases With gas Normalized concentration change within the current time window , They represent gases With gas The normalized average gas production rate within the current time window This represents the fault stage progression direction coefficient corresponding to the edge of this mechanism. It represents the balance coefficient between changes in concentration and changes in gas production rate.

[0070] Optionally, the above-mentioned preset fault stage progression direction is set in the following ways: for the gas-related edge corresponding to the overheating fault stage, the preset fault stage progression direction is the direction in which the concentration of the thermal fault-related gas increases as the fault develops; for the gas-related edge corresponding to the discharge fault stage, the preset fault stage progression direction is the direction in which the concentration of the discharge-related gas increases as the fault develops; for the gas-related edge corresponding to the overheating-discharge transition stage, the preset fault stage progression direction is the direction in which the concentration of the discharge-related gas increases as the fault develops. The above-mentioned extraction of the stage progressive response term of the gas-related edge based on the concentration change and gas production rate of the gas at both ends of the gas-related edge and the preset fault stage progression direction corresponding to the gas-related edge includes: when the change relationship of the gas at both ends of the gas-related edge conforms to the preset fault stage progression direction, the stage progressive response term is determined to take a positive value; when the change relationship of the gas at both ends of the gas-related edge is opposite to the preset fault stage progression direction, the stage progressive response term is determined to take a negative value.

[0071] The aforementioned preset fault stage progression direction is based on the objective mechanism of the generation sequence of characteristic gases during the development of various faults. For the gas-related edge corresponding to the overheating fault stage, the progression direction is set as the direction in which the concentration of thermal fault-related gases increases with the development of the fault. The underlying mechanism is that the degree of thermal decomposition of insulating oil and insulating paper deepens with increasing temperature. In the low-temperature overheating stage, methane is the main product. As the overheating temperature develops towards medium and high temperatures, the generation ratio of methane and ethane gradually decreases while the generation ratio of ethylene continues to increase. The direction of the thermal fault-related gas increasing from methane and ethane to ethylene thus characterizes the progression law of overheating faults evolving from low to high temperatures. For the gas-related edge corresponding to the discharge fault stage, the progression direction is set as the direction in which the concentration of discharge-related gases increases with the development of the fault. The underlying mechanism is that the discharge energy level determines the type of decomposition products. Low-energy discharge is mainly dominated by hydrogen produced by local ionization in the oil. After the discharge energy increases, the high temperature of the arc causes a large amount of acetylene to be generated. The direction of the hydrogen increasing towards acetylene characterizes the progression law of discharge faults evolving from low to high energy. For the gas-related edge corresponding to the overheat-discharge transition stage, the progression direction is also set as the direction in which the concentration of discharge-related gas increases as the fault develops. The mechanism is that the enhancement of the discharge component in the thermoelectric recombination region is a sign that the fault nature is changing to a more severe direction. The direction of the enhancement of methane and hydrogen to acetylene describes the progressive law of the transition from thermal fault to discharge fault.

[0072] The result of setting the progression direction is applied to each gas-related edge through the failure stage progression direction coefficient. This coefficient is a dimensionless constant pre-set based on the failure stage mechanism and is not used as a network training parameter. For mechanism edges that conform to the preset failure stage progression direction, the coefficient is 1; for mechanism edges whose calculation direction is opposite to the preset progression direction, the coefficient is -1. In the calculation of the stage progression response term, the difference between the normalized concentration changes of the gases at both ends and the difference between the normalized average gas production rate are summed after adjustment by the balance coefficient, and then multiplied by the failure stage progression direction coefficient. When the change direction and intensity of the gases at both ends conform to the progression law of the corresponding failure stage after correction by the direction coefficient, the stage progression response term takes a larger positive value; when the change relationship of the gases at both ends is opposite to the preset progression direction, the negative correction of the direction coefficient makes the stage progression response term take a negative value; when the change relationship does not conform to the stage progression law, the stage progression response term takes a smaller value or is negative. The positive or negative value of the progressive response term directly corresponds to whether the change relationship is consistent with or contradicts the progressive law, and the magnitude of the value corresponds to the degree of consistency or contradiction.

[0073] The working principle of pre-setting progressive directions according to fault stages lies in using prior knowledge of the mechanism to determine the evolution direction of gas-related edges. Different types of faults have different characteristic gas combinations, and therefore different progressive directions. If a uniform progressive direction is used for all related edges, the evolution of thermal fault gases and discharge gases will be measured by the same standard, and the progressive discrimination will lose its fault type specificity. After setting the stages, the progressive direction of each related edge strictly corresponds to the fault stage mechanism to which the edge belongs. The change relationship on the overheating edge is tested according to the temperature evolution law of thermal faults, the change relationship on the discharge edge is tested according to the energy evolution law of discharge, and the change relationship on the transition edge is tested according to the thermoelectric recombination transformation law. Since the progressive direction coefficient and the balance coefficient are preset constants and do not participate in training, the progressive discrimination rule remains stable throughout the entire network training and application process and will not drift due to differences in the distribution of training samples. Therefore, the adjustment of the gating coefficient by the stage progressive response term always follows the objective law of the fault mechanism.

[0074] Mechanism-based edge response features refer to a unified edge-level feature representation formed by combining the concentration response features, gas production rate response features, and stage-based progressive response terms of the same gas-related edge. This signifies that the object of feature extraction has formally changed from a single gas node to a gas-related edge. For gases... With gas The mechanism edge between them, and the mechanism edge response characteristics are represented as follows: ; In the formula, Represents gas With gas Mechanistic response characteristics of the inter-mechanistic edge. This indicates the concentration response characteristics of the gas at both ends of the edge. This indicates the gas production rate response characteristics at both ends of the edge. This represents the stage-progressive response term of this edge. Both the concentration response feature and the gas production rate response feature are four-dimensional vectors, and the stage-progressive response term is a scalar. The combined mechanism edge response feature fully characterizes the co-evolution state of the associated gases in three dimensions: concentration accumulation, fault development speed, and stage-progressive law.

[0075] The extraction focus of mechanism edge response features varies depending on the fault stage type to which the associated edge belongs. For overheating mechanism edges, the emphasis is on extracting the stage-specific evolution responses between gases related to thermal faults, such as methane, ethane, and ethylene. The coordinated changes in the concentration and gas production rates of these gases reflect the process of thermal fault development from low to high temperature. For discharge mechanism edges, the emphasis is on extracting the growth responses between discharge-related gases, such as hydrogen and acetylene. The rapid coordinated growth of these gases reflects the energy development process of discharge faults. For overheating-discharge transition mechanism edges, the emphasis is on extracting the thermoelectric composite transition responses between gases such as ethylene, acetylene, and hydrogen. The correlation changes between these gases reflect the composite evolution process of the superimposed region of thermal and discharge faults. Through this approach, feature extraction is no longer simply about extracting the concentration or gas production rate characteristics of each gas itself. Instead, it uses the gas-related edges in the fault stage mechanism diagram as the feature extraction objects to obtain mechanism edge response features that can reflect the evolution law of gas combination.

[0076] During the aggregation process, the contribution of each mechanism edge response feature is determined based on the edge weights of the gas-related edges in the current fault stage mechanism diagram. For gas-related edges with higher weights, their mechanism edge response features contribute more to the aggregation process; conversely, for edges with lower weights, their response feature contributions decrease accordingly. Aggregating the mechanism edge response features of each gas-related edge in the current fault stage mechanism diagram yields the concentration-gas production rate mechanism edge feature representation for the current sample. ; In the formula, This represents the set of gas-related edges in the current fault stage mechanism diagram; Adjacency matrix representing the fault phase mechanism Middle gas With gas The corresponding edge weights (see the following description for the method of determining edge weights) are used to characterize the degree of matching between the gas-related edge and the current fault stage; This represents the concentration-gas production rate mechanism edge characteristics after aggregation.

[0077] The concentration-gas production rate mechanism edge feature representation is the sample-level output of the feature extraction stage. This feature representation includes both the gas cumulative degradation information in the concentration stream and the fault development dynamic information in the gas production rate stream, and preserves the fault stage mechanism correlation between different gases. Compared to directly extracting node features from the concentration stream and the gas production rate stream separately, or simply concatenating the two types of features, the above feature extraction process can directly characterize the fault evolution response on the gas correlation edge during the feature extraction stage. This allows subsequent fault classification to be based on gas combination relationships rather than changes in a single gas, thereby improving the ability to identify adjacent fault levels, boundary samples, and complex operating condition samples.

[0078] Optionally, determining the gating coefficient of the gas-related edge based on the degree of consistency between the concentration response characteristics and the gas production rate response characteristics of the gas-related edge includes: determining the difference between the concentration response characteristics and the gas production rate response characteristics of the gas-related edge; wherein the difference is used to characterize the degree of inconsistency between the concentration response characteristics and the gas production rate response characteristics; determining the gating coefficient of the gas-related edge based on the difference; wherein the value of the gating coefficient decreases as the difference increases, so as to suppress the influence of the gating fusion characteristics of the gas-related edge on the fault diagnosis results when the concentration response characteristics and the gas production rate response characteristics are inconsistent.

[0079] The aforementioned consistency difference refers to the difference between the concentration response characteristics and the gas production rate response characteristics on the same gas-related edge, and it is used to characterize the degree of inconsistency between the two types of response characteristics. For gases With gas The consistency difference between the mechanistic edges is represented as follows: ; In the formula, This indicates the degree of inconsistency between the concentration response characteristics and the gas production rate response characteristics at the edge of this mechanism. , These represent the concentration-side response characteristics and gas production rate-side response characteristics of this mechanism, respectively. The L2 norm is used to represent the difference between concentration response features and gas production rate response features. Since both are four-dimensional vectors with the same dimension, the L2 norm of the difference between the vectors can be directly calculated, yielding a dimensionless non-negative scalar. The reason for using the L2 norm to measure the difference is that the component structures of the two types of response features correspond to each other. Each component characterizes the cooperative relationship of the associated gases from four aspects: the intensity of change, the consistency of the direction of change, the difference in the magnitude of change, and the sequence correlation. The L2 norm of the vector difference can synthesize the deviations from the four aspects into a single value; inconsistency in any aspect will increase this value. The smaller the value, the more consistent the concentration response and gas production rate response are with the same mechanism. The larger the value, the more likely the mechanism may have short-term abnormal fluctuations, detection noise, or inconsistent changes in the mechanism.

[0080] Gating coefficient The calculation formula is: ; In the formula, Represents gas With gas The gating coefficient of the mechanism edge between them has a value range from 0 to 1; , These represent the concentration-side response characteristics and gas production rate-side response characteristics of this mechanism, respectively. The degree of inconsistency between the concentration response characteristics and the gas production rate response characteristics at the edge of the mechanism is indicated by the L2 norm of the difference between the concentration response characteristics and the gas production rate response characteristics. Since both types of response characteristics are four-dimensional vectors with the same dimension, the L2 norm difference can be calculated directly. and These are learnable parameters that are updated as the network trains. It is a non-negative adjustment coefficient; This represents the Sigmoid function, used to constrain the gating coefficient between zero and one. This is a consistency difference negative feedback term, used to suppress the contribution of the mechanism edge based on the degree of inconsistency between the two types of responses.

[0081] In the above-mentioned gating coefficient In the calculation formula, consistency differences participate in the calculation of the gating coefficient in the form of negative feedback, and the formula for calculating the gating coefficient includes a consistency difference negative feedback term. ,in This is a non-negative adjustment coefficient used to control the strength of negative feedback. The larger the value, the stronger the inhibitory effect of consistency differences on the gating coefficient. The negative feedback term causes the gating coefficient to decrease monotonically as consistency differences increase. When the negative feedback term is close to zero, it has almost no effect on the gating coefficient, and the correlation edge supported by the concentration response and the gas production rate response obtains a large gating coefficient; when As the negative feedback term increases, its absolute value increases synchronously, and the gating coefficient is correspondingly reduced. The gating coefficient is constrained to between 0 and 1 by the Sigmoid function. When multiplied as a scalar with the gas production rate side response feature, it means that each component in the gas production rate side response feature is multiplied by the same gating coefficient. During fusion, the concentration side response feature is retained and the gas production rate side response feature weighted by the gating coefficient is superimposed on it.

[0082] In addition, consistency differences We will examine, from a numerical perspective, whether the concentration response and the gas production rate response on the same associated edge support each other. The smaller the value, the more consistent the two types of responses; stage-progressive response term From a mechanistic perspective, examine whether the changing relationship of this associated edge conforms to the progressive pattern of the fault stage. When Smaller and This indicates that when the edge follows a progressive pattern, the gating coefficient increases, and the contribution of this mechanism's edge feature to the classification result increases; when Larger or This indicates that when the edge does not conform to the progressive law, the gating coefficient decreases, and the influence of the edge characteristics of this mechanism is suppressed. After introducing the stage progressive response term, the gating discrimination no longer depends solely on the numerical consistency of the two types of responses. Even if the concentration response and gas production rate response of a certain associated edge are similar in numerical form, as long as their evolution direction contradicts the progressive law of the fault stage, the gating coefficient will still be lowered accordingly. The cooperative fluctuations caused by atypical fault evolution are thus excluded from valid fault evidence.

[0083] The aforementioned gated fusion feature refers to the feature vector obtained by fusing the concentration response feature and gas production rate response feature of the same gas-related edge according to a gating coefficient. It carries comprehensive fault evidence for that edge after mechanistic consistency screening. For gases... With gas The mechanism between them, the gated fusion feature is represented as: ; In the formula, Represents gas With gas Gated fusion features of the mechanism edges between them This indicates the concentration-side response characteristics of the mechanism. This indicates the gas production rate and response characteristics of the mechanism. This represents the gating coefficient of the mechanism side. Since the gating coefficient is a scalar ranging from 0 to 1, when it is multiplied by the gas production rate side response feature, it means that each feature component in the gas production rate side response feature is multiplied by the same gating coefficient. The fusion result maintains the same dimension as the two types of response features, which are both dimensionless four-dimensional vectors.

[0084] The fusion process employs a method of retaining concentration-based side-response features and then weighting and superimposing gas production rate-based side-response features. This design is based on the different roles of the two types of response features in the fault evidence system. Concentration-based side-response features reflect the deterioration results caused by long-term accumulation of associated gases. After preprocessing, the impact of data noise on the concentration sequence is relatively limited, and concentration-based side-response features are fully retained as the basic term for fusion. Gas production rate-based side-response features reflect the dynamic speed of fault development and are more sensitive to instantaneous sensor interference and short-term fluctuations. Their reliability varies with the sample operating conditions, therefore they are weighted by the gating coefficient as the adjusted term. When the gating coefficient is close to one, the gas production rate-based side-response features are superimposed on the concentration-based side-response features with near-complete strength, and the fused features simultaneously contain fault information in both the accumulation and velocity dimensions. When the gating coefficient is close to zero, the contribution of the gas production rate-based side-response features is compressed, and the fused features are mainly composed of concentration-based side-response features, excluding suspicious rate fluctuations from the fusion results.

[0085] Understandably, the aforementioned gating coefficient allows for on-demand adjustment of evidence strength, rather than simple selection. When the gating coefficient is at an intermediate value, the gas production rate edge response features participate in the fusion in a corresponding proportion. Rate evidence from associated edges is neither fully adopted nor completely discarded; its participation level strictly corresponds to the consistency between the concentration and rate responses of that edge. This continuous adjustment method ensures that the gating fusion features output by each gas-related edge naturally carry information about the reliability of that edge's evidence, eliminating the need for secondary judgment on the usability of each edge's features during subsequent aggregation. Furthermore, fusion is performed edge-by-edge on a per-edge basis. Abnormal fluctuations on a particular edge only affect its own fusion features and do not spread to other associated edges through the fusion process. The independence between the fusion features of each edge provides clear data boundaries for weighted aggregation based on edge weights to form fault discrimination features.

[0086] Step S170: Determine the fault diagnosis result of the transformer based on the gating fusion characteristics of each gas-related edge.

[0087] Optionally, step S170 includes: determining the edge weight of each gas-related edge in the fault stage mechanism diagram; wherein the edge weight is used to characterize the degree of matching between the corresponding gas-related edge and the current fault stage, and the larger the edge weight, the higher the degree of matching between the gas-related edge and the current fault stage; weighted aggregation of the gated fusion features of each gas-related edge according to the edge weight of each gas-related edge to obtain fault discrimination features; and determining the fault diagnosis result of the transformer based on the fault discrimination features.

[0088] The meaning of the edge weights mentioned above is consistent with the convergence in the feature extraction stage. Specifically, the values ​​of the matrix elements corresponding to the gas-related edges in the fault stage mechanism adjacency matrix A_t are used to characterize the degree of matching between the gas-related edge and the current fault stage. A larger edge weight indicates a higher degree of matching between the gas-related edge and the current fault stage. The edge weights are dimensionless. Since the gated fusion features and the mechanism edge response features are on the same fault stage mechanism graph, the set of gas-related edges used for convergence is also [missing information]. Nodes outside the set do not participate in the aggregation. It is important to distinguish that the aggregation object in the feature extraction stage is the mechanism edge response feature that has not been screened for consistency, while the aggregation object here is the gated fusion feature after the gating coefficient has been adjusted according to the mechanism consistency. The latter has already completed the suppression of noise and abnormal fluctuations before entering the aggregation stage.

[0089] Fault discrimination features refer to the sample-level feature representation obtained by weighted aggregation of the gated fusion features of each gas-related edge according to the edge weight. The calculation formula is as follows: ; In the formula, This represents the fault discrimination features of the current sample. This represents the set of gas-related edges in the current fault stage mechanism diagram. The gas in the adjacency matrix representing the fault stage mechanism With gas The corresponding edge weights, Represents gas With gas Gated fusion features of the mechanistic edges between gas-related edges. Edges with higher weights indicate a better match between the gas correlation and the current fault stage, and their gated fusion features contribute more to the fault discrimination features; conversely, edges with lower weights contribute less. The dimensions of the fault discrimination features are consistent with those of the gated fusion features, being dimensionless vectors. The aggregation can also be normalized by summing the weights of each edge after weighted summation, ensuring that the numerical range of the fault discrimination features does not fluctuate with the number of related edges.

[0090] The fault diagnosis result is determined by a fault classifier based on fault discrimination features. The fault classifier consists of a fully connected layer and a Softmax output layer. The fully connected layer maps the fault discrimination features to the discrimination scores corresponding to each fault category. The Softmax output layer converts the discrimination scores of each category into category probabilities. The probability of a category is equal to the ratio of the exponentially calculated discrimination score of that category to the sum of the exponentially calculated discrimination scores of all categories. The fault categories include six types: normal state, low-temperature overheating, medium-temperature overheating, high-temperature overheating, low-energy discharge, and high-energy discharge. The category corresponding to the highest probability among all categories is the final diagnostic result for the current DGA sample. From the perspective of the composition of the discrimination criteria, the main components of the fault discrimination features come from gas association edges that have a high degree of matching with the current fault stage and have passed the consistency test. The discrimination of the classifier is actually dominated by the gas combination relationship that conforms to the mechanism of the current fault stage and the mutual support of concentration and rate. The edge features that are weakly associated with the current fault stage only play a supplementary role. Therefore, the sample-level discrimination criteria are consistent with the fault stage mechanism. The distinction between adjacent fault level samples and boundary samples is concentrated on the association edges that truly reflect the stage differences, and the relevance and interpretability of the classification results are guaranteed.

[0091] Optionally, the above-mentioned transformer DGA diagnostic method based on mechanism graph gated two-stream network further includes: constructing training samples; wherein the training samples are used to perform end-to-end training of the two-stream network; The above-mentioned construction of training samples includes: using the concentration data of real samples of dissolved gas in transformer oil as the concentration data of the fault evolution sequence at the final sampling time; generating concentration data of multiple sampling times before the final sampling time according to the evolution law of gas concentration with fault development, and determining the corresponding gas production rate according to the concentration change between adjacent sampling times to obtain the simulation time series sample; and determining the fault category label of the real sample as the fault category label of the simulation time series sample.

[0092] The construction of the above training samples is based on real samples from the publicly available DGA dataset. Since the publicly available dataset does not contain real continuous online monitoring timestamps, it is impossible to directly obtain training data with a time-series dimension. This embodiment of the invention uses real DGA samples as the endpoint of the fault evolution sequence, generates concentration data for multiple sampling times prior to the endpoint sampling time according to the gradual evolution of gas concentration, and calculates the corresponding gas production rate characteristics based on the concentration changes between adjacent sampling times, forming simulated time-series samples. The category labels of the generated time-series samples retain the fault type labels of the corresponding original publicly available samples. The time-series generation algorithm is only used to expand the time dimension of the samples, forming concentration sequences and gas production rate sequences, without changing the original fault categories. Therefore, the reliability of the labels is not affected by the time-series generation process.

[0093] The network training employs an end-to-end approach, jointly training the concentration stream input, gas production rate stream input, fault stage mechanism diagram generation, mechanism side-response feature extraction, gating fusion, and fault classifier as a whole. For each training sample, the network outputs the predicted probability of each preset fault category. The training loss uses class-weighted cross-entropy with an added regularization term, expressed as: ; In the formula, Indicates training loss; Indicates the number of training samples; Indicates the first The true fault category of each sample; Indicates a two-stream network for the first... The predicted probability of the true fault category of each sample; This represents the category weights, used to reduce the impact of imbalanced sample sizes across different fault categories. Indicates the trainable parameters of the network; Represents the regularization coefficient; This represents the squared L2 norm of the trainable parameters of the network. The category-weighted cross-entropy term assigns larger category weights to fault classes with fewer samples, causing the network to bear greater losses for misclassification of minority class samples. The category weights are dimensionless. The regularization term suppresses overfitting by constraining the parameter size. The regularization coefficient is dimensionless, and its value controls the strength of the regularization constraint.

[0094] During training, the Adam optimizer is used to update network parameters. The initial learning rate is determined based on the validation set results. When the validation set loss does not decrease for several consecutive rounds, the learning rate is reduced to decrease the magnitude of parameter updates in the later stages of training and stabilize network convergence. Training stops when the validation set diagnostic index no longer improves for several consecutive rounds, resulting in the final network used for fault classification. The parameter updates cover feature mapping parameters, gating fusion parameters, and classifier parameters in the network. All of these parameters are updated through backpropagation. Pre-defined mechanism parameters, such as the fault stage progression direction coefficient, are not involved in training, and their values ​​remain fixed throughout the training process. The distinction between pre-defined mechanism parameters and trainable parameters ensures that the fault mechanism prior is always injected into the network as an invariant constraint. The part of the network that learns from the data is limited to the numerical relationship between feature mappings and fusion weights. The discrimination rules at the mechanism level will not shift due to differences in the distribution of training samples.

[0095] To facilitate understanding of the working principle of the transformer DGA diagnostic method based on the mechanistic graph-gated two-stream network described above, this invention also provides specific application examples of this method in a certain application scenario. For example... Figure 2 As shown, in this application scenario, the transformer DGA diagnostic method based on the mechanism graph-gated two-stream network mainly includes: Step 1: Construct a mechanism diagram of concentration-rate dual-stream input and fault stages; This step is used to construct a concentration-rate dual-stream input and generate a fault phase mechanism diagram that matches the current DGA sample state. The diagram construction process is as follows: Figure 3 As shown, it specifically includes: First, obtain dissolved gas detection data in transformer oil, the detection data including... , , , , Concentration values ​​of five key characteristic gases at multiple sampling times. The original concentration sequence is processed by imputation of missing values, suppression of abnormal spikes, and truncation within a sliding time window to obtain the concentration sequence within the current time window, which serves as the input to the concentration stream. The sliding time window uses a fixed length truncation method. Each time window contains concentration data from 10 consecutive sampling times, with adjacent time windows shifting forward by one sampling point. Since only one sampling point is shifted between adjacent windows, the actual time interval between adjacent windows is equal to the sampling interval of the original DGA data. To ensure that the time window reflects the gas concentration change trend and the gas production rate change pattern, the number of effective sampling points within the window is no less than 6; if the number of effective sampling points is less than 6, no sample is generated for that time window.

[0096] Based on this, the gas production rate is calculated according to the change in gas concentration between adjacent sampling times, forming the rate stream input. The gas production rate includes absolute and relative rates, where the absolute rate characterizes the increase in gas concentration per unit time, and the relative rate characterizes the change in gas concentration relative to the baseline value at the previous sampling time. The concentration stream characterizes the level of cumulative gas degradation, and the rate stream characterizes the speed of fault development and the degree of short-term abrupt changes.

[0097] Furthermore, with , , , , Five characteristic gases are used as graph nodes. Based on the gas generation mechanisms under different fault stages, sub-graphs for overheating mechanism, discharge mechanism, and overheating-discharge transition mechanism are constructed. The overheating mechanism sub-graph is denoted as... , used to characterize , , The phased correlation of gases during low-temperature, medium-temperature, and high-temperature superheating processes; the discharge mechanism sub-diagram is denoted as follows. This is used to characterize the correlation between gases such as H2 and C2H2 in low-energy and high-energy discharge processes; the overheating-discharge transition mechanism sub-diagram is denoted as... Used to characterize samples with high-temperature overheating, strong discharge, and thermoelectric recombination boundaries. , , Transitional correlations between gases.

[0098] To enable the mechanism diagram to adapt to the current sample state, the current concentration, concentration change, average gas production rate, rate change trend, and key gas ratio are extracted from the current time window to form the current sample state vector. The rate change trend is determined by the fitting slope of the gas production rate sequence within the current time window. For each characteristic gas, a linear fit is performed with the sampling order of each time interval within the window as the independent variable and the corresponding gas production rate as the dependent variable. The resulting slope represents the rate change trend of that gas; a positive slope indicates an overall increase in gas production rate, a negative slope indicates an overall decrease in gas production rate, and a slope close to zero indicates a generally stable gas production rate. Key gas ratios include... / , / , / The above features are normalized and then concatenated to determine whether the current sample is closer to an overheating fault, a discharge fault, or an overheating-discharge transition state.

[0099] Based on the current sample state vector Adaptive weights for generating three mechanistic subgraphs: Based on this, the adjacency matrix of the fault stage mechanism corresponding to the current time window is obtained: .

[0100] Step 2: Extraction of concentration-gas production rate features based on the side response of the fault stage mechanism; The purpose of this step is to extract features from the concentration flow and gas production rate flow based on the fault stage mechanism diagram generated in Step 1, focusing on gas correlation relationships. Unlike extracting features from individual gas nodes, this step uses gas correlation edges in the fault stage mechanism diagram as feature extraction units to characterize whether correlated gases exhibit co-evolutionary characteristics consistent with overheating, discharge, or overheating-discharge transition fault mechanisms. The feature extraction process is as follows: Figure 4 As shown, it mainly includes: For the gas-related edge in the current fault stage mechanism diagram, based on the concentration sequence and gas production rate sequence of the gas at both ends of the edge within the current time window, the concentration response feature, gas production rate response feature, and stage progressive response term are extracted respectively.

[0101] For gases With gas The mechanism between them, and its concentration response characteristics are expressed as: Its gas production rate response characteristics are expressed as follows: .

[0102] For gases With gas Mechanism between them, constructing progressive response terms in the construction phase: The aforementioned fault stage progression direction coefficients As a constant pre-set based on the fault stage mechanism, it is not used as a network training parameter. For mechanism edges that conform to the preset fault stage progression direction, =1, for mechanism edges whose calculation direction is opposite to the preset progressive direction, take =-1, where the overheating mechanism is based on , Towards Enhanced direction setting, discharge mechanism according to Towards Enhanced directional setting, overheat-discharge transition mechanism according to , Towards Enhanced direction setting; the balance coefficient These are preset hyperparameters, not used as network training parameters, and their values ​​range from 0 to 1. In this embodiment... A value of 0.35 is used to retain the dominant role of concentration change in the staged progressive response term, while introducing a supplementary characterization of the fault development rate by the change in gas production rate; when the direction and intensity of gas change at both ends of this mechanism are... When the correction conforms to the progressive pattern of the corresponding fault stage, Take the larger positive value; when the changes in both do not conform to the progressive law of stages, Smaller or negative.

[0103] Based on the concentration response characteristics, gas production rate response characteristics, and stage-progressive response terms of the gases at both ends of the associated edge, the mechanism edge response characteristics are extracted: .

[0104] For the overheating mechanism, the key points are extracted. , , The phased evolution response among gases related to isothermal faults; for the discharge mechanism boundary, the focus is on extracting... , The growth response between gases related to discharge is emphasized; for the overheating-discharge transition mechanism, the key extraction is... , , Thermoelectric composite transition response between gases. Using the above method, instead of simply extracting the concentration characteristics or gas production rate characteristics of each gas, the gas-related edges in the fault stage mechanism diagram are used as feature extraction objects to obtain mechanism edge response characteristics that reflect the evolution law of gas combination.

[0105] During the aggregation process, the contribution of each mechanism edge response feature is determined based on the weight of the gas-related edges in the current fault stage mechanism diagram. For gas-related edges with higher weights, their mechanism edge response features contribute more to the aggregation process; for edges with lower weights, their response feature contribution decreases accordingly. Furthermore, the mechanism edge response features of each gas-related edge in the current fault stage mechanism diagram are aggregated to obtain the concentration-gas production rate mechanism edge feature representation of the current sample: , .

[0106] Compared to directly extracting node features from concentration flow and gas production rate flow separately, or simply concatenating the two types of features, this method can directly characterize the fault evolution response on gas-related edges during the feature extraction stage. This allows subsequent fault classification to be based on gas combination relationships rather than single gas changes, thereby improving the ability to identify adjacent fault levels, boundary samples, and complex operating condition samples.

[0107] Step 3: Gated fusion and fault classification based on mechanism-based edge response consistency; This step is used to adjust the contribution of different gas-related edges to fault classification based on the consistency between the concentration response and the gas production rate response at the mechanistic edge, thereby suppressing the effects of detection noise, isolated rate spikes, and mechanistic inconsistency fluctuations. The specific process is as follows: Figure 5 As shown, it mainly includes: Based on the mechanistic edge response features extracted in the above steps, the concentration edge response features and gas production rate edge response features corresponding to each gas association edge are obtained. The concentration edge response features characterize the cumulative degradation relationship of the associated gases within the current time window, while the gas production rate edge response features characterize the dynamic development relationship of the associated gases within the current time window. If the concentration response and gas production rate response on the same gas association edge match, it indicates that the gas combination is more likely to conform to the evolution pattern of the current fault stage; if the two differ significantly, it may correspond to short-term abnormal fluctuations, detection noise, or atypical fault evolution.

[0108] For gases With gas The consistency difference between the concentration-side response characteristics and the gas production rate-side response characteristics of the mechanism-related edge is calculated: In this application scenario, and Both are 4-dimensional vectors, and since they have the same dimensions, the difference in L2 can be calculated directly. The smaller the value, the more consistent the concentration response and gas production rate response are with each other along the same mechanism. The larger the value, the more likely the mechanism may have short-term abnormal fluctuations, detection noise, or inconsistent changes in the mechanism.

[0109] Based on the existing dual-flow gating fusion, a mechanism-edge consistency difference negative feedback module is added, and the gating object is adjusted from overall characteristics to gas-related edges in the fault stage mechanism diagram. Specifically, based on the concentration edge response characteristics... Gas production rate side response characteristics and the differences in the consistency of their mechanisms. Generate the gating coefficients corresponding to the edge of this mechanism: .because As a scalar, its side response characteristics are related to the gas production rate. When multiplying, it means that the product will be multiplied. Each feature component is multiplied by the same gating coefficient. In the formula This is a consistency difference negative feedback term, used to suppress the contribution of this mechanism based on the degree of inconsistency between the concentration response and the gas production rate response. When Smaller and This indicates that when the edge conforms to the progressive law of fault stages, the gating coefficient increases, and the contribution of the edge feature of this mechanism to the classification result increases; when Larger or This indicates that when the edge does not conform to the progressive rule, the gating coefficient decreases, and the influence of the edge feature of this mechanism is suppressed.

[0110] Subsequently, based on the gating coefficients of each mechanism edge, the corresponding concentration edge response characteristics and gas production rate edge response characteristics are fused to obtain the gating fusion characteristics of each gas-related edge: The above fusion process indicates that the concentration-side response characteristics are preserved. And based on the scalar gating coefficient Side response characteristics of gas production rate After overall weighting, the features are superimposed. Furthermore, the gated fusion features of each gas-related edge are weighted and aggregated according to the edge weights in the current fault stage mechanism graph to form the fault discrimination features of the current sample. The higher the edge weight, the better the gas correlation matches the current fault stage, and the greater the contribution of its gating fusion feature to the fault discrimination feature; the lower the edge weight, the lower its contribution.

[0111] During the fault classification stage, the fault discrimination features of the current sample are... The input consists of a fault classifier comprised of a fully connected layer and a Softmax output layer. The fully connected layer maps fault discrimination features to discrimination scores corresponding to each fault category, while the Softmax output layer converts the discrimination scores of each category into category probabilities. The fault categories include normal state, low-temperature overheating, medium-temperature overheating, high-temperature overheating, low-energy discharge, and high-energy discharge. The category with the highest probability is taken as the final diagnostic result for the current DGA sample.

[0112] During the training phase, time window samples constructed from the publicly available DGA dataset are used as training samples. Since this embodiment uses publicly available static DGA samples to generate simulation time-series data, the category labels of the generated time-series samples retain the fault type labels of the corresponding original publicly available samples; the time-series generation algorithm is only used to expand the time dimension of the samples, forming concentration sequences and gas production rate sequences, without changing the original fault categories. During training, the concentration stream input, gas production rate stream input, fault stage mechanism diagram generation, mechanism side-response feature extraction, gating fusion, and fault classifier are treated as a whole for end-to-end training. For each training sample, the network outputs the predicted probability of each preset fault category, and uses category-weighted cross-entropy as the training loss. .

[0113] During training, the Adam optimizer is used to update network parameters, and the initial learning rate is determined based on the validation set results. The learning rate is reduced when the validation set loss fails to decrease for several consecutive rounds; training stops when the validation set diagnostic index no longer improves for several consecutive rounds, resulting in the final network used for fault classification. Except for preset mechanism parameters such as the fault stage progression direction coefficient, the feature mapping parameters, gating fusion parameters, and classifier parameters in the network are all updated through backpropagation.

[0114] By adjusting the fusion contribution of concentration response and gas production rate response on a fault stage mechanism basis, the above methods enhance effective fault evidence when the two are consistent and suppress the influence of abnormal fluctuations when they are inconsistent, thereby improving the diagnostic stability and interpretability of boundary samples, adjacent fault level samples and complex operating condition samples.

[0115] To verify the effectiveness of the above-mentioned transformer DGA diagnostic method based on the mechanism graph-gated two-current network, the embodiments of the present invention also conducted simulation verification, and the simulation results are as follows: Figure 6 , Figure 7 and Figure 8 As shown, the simulation results are analyzed as follows: exist Figure 6 In the high-temperature overheated sample shown, the ethane concentration increased continuously from a low initial level to approximately 500 ppm, while the methane and hydrogen concentrations rose to approximately 140 ppm and 95 ppm, respectively. The ethylene concentration remained at a low level, and the acetylene concentration remained almost constant throughout the evolution. This gas evolution pattern is consistent with the gas generation mechanism of thermal faults, namely, the insulating oil under continuous thermal stress mainly produces stable alkanes without the generation of discharge characteristic gases, and a clear concentration gradient with distinct primary and secondary components exists among the five gases. The transformer DGA diagnostic method based on the mechanism graph-gated dual-current network described above provides a high-temperature overheating discrimination result for this sample, with a discrimination confidence level of 41.33%. This indicates that the concentration response extracted from the overheating mechanism edge in the mechanism graph is mutually supportive with the gas generation rate response, and the synergistic growth of thermal fault gases is identified as valid fault evidence.

[0116] exist Figure 7In the low-energy discharge sample shown, the hydrogen concentration continuously increased from approximately 10 ppm to approximately 95 ppm, exhibiting an accelerated growth trend. The acetylene concentration rose from near zero to approximately 33 ppm in the later stages of evolution, while the concentrations of methane, ethane, and ethylene remained at low levels with gradual increases. This evolutionary pattern is consistent with the mechanistic characteristics of low-energy discharge, namely, local ionization primarily generates hydrogen with a small amount of acetylene, and the concentration levels of discharge-related gases and thermal fault-related gases are clearly distinguished. The transformer DGA diagnostic method based on the mechanistic graph-gated dual-current network (GMGA) provided a low-energy discharge discrimination result for this sample with a discrimination confidence level of 41.84%, indicating that the side-response of the discharge mechanism between hydrogen and acetylene was effectively captured in both the concentration accumulation and rate increase dimensions.

[0117] exist Figure 8 In the high-energy discharge sample shown, the concentrations of hydrogen and acetylene rapidly increased to approximately 215 ppm and 185 ppm, respectively, while the ethylene concentration increased synchronously to approximately 105 ppm, and the increases in methane and ethane concentrations were limited. Compared to the low-energy discharge sample, the increase in acetylene in this sample was close to that of hydrogen, reflecting the mechanism characteristic of the large-scale generation of acetylene due to the high temperature of the arc after the discharge energy increases. The rapid increase of hydrogen and acetylene together constitutes the typical gas combination of high-energy discharge. The transformer DGA diagnostic method based on the mechanism graph-gated two-current network described above provides a high-energy discharge discrimination result for this sample, with a discrimination confidence level of 40.47%, indicating that the stage-progressive response term on the discharge mechanism edge correctly identified the progressive direction of hydrogen to acetylene enhancement.

[0118] comprehensive Figure 6 , Figure 7 and Figure 8 It can be seen that there are significant differences in the gas growth patterns corresponding to different fault types. Thermal faults are mainly characterized by the continuous accumulation of alkane gases, while discharge faults are mainly characterized by the rapid growth of hydrogen and acetylene. Furthermore, different discharge energy levels lead to different ratios of hydrogen and acetylene. These differences are reflected in both the concentration accumulation level and the gas production rate, verifying the ability of concentration time-series characteristics and gas production rate characteristics to characterize the fault evolution process and providing data support for the construction of the concentration-rate dual-stream input in this invention. The discrimination confidence levels of the three types of samples are all around 40%, not close to 100%. This is because adjacent fault levels in the six-class classification task have similar gas compositions, and the discrimination probabilities are reasonably dispersed among similar categories. However, the highest probability category is consistent with the actual fault type, indicating that the model can still provide correct discrimination in multi-class environments with similar boundaries. This demonstrates the supporting role of fault stage mechanism diagram constraints and mechanism edge response consistency gating in diagnostic accuracy.

[0119] Furthermore, this embodiment of the invention also compares the diagnostic performance of the transformer DGA diagnostic method based on the mechanism graph-gated two-stream network with that of various comparative models on the simulated DGA time series dataset. The comparison results are shown in Table 1. The evaluation indicators include accuracy, precision, recall, and F1 score. The comparative models are arranged in a progressive order of technical composition as follows: static DGA method, concentration-flow model, concentration-rate two-stream model, mechanism graph two-stream network, and mechanism graph-gated fusion model. The method of this invention achieves the highest values ​​in all four indicators.

[0120] Table 1 Performance Comparison Results

[0121] The static DGA method described above only uses gas concentration information at a single moment for discrimination, achieving an accuracy of 70.43%, precision of 62.22%, recall of 65.73%, and F1 score of 61.32%. All four metrics are the lowest among all models, indicating that static concentration information cannot distinguish samples with similar concentration values ​​but different fault development stages. After introducing time-series concentration information, the accuracy of the concentration-flow model improves to 88.91%, 82.94%, 87.13%, and 84.33%, respectively, a significant improvement over the static DGA method. This demonstrates that the evolution trend of gas concentration over time includes fault development stage information not present in the static method. Furthermore, by adding the gas production rate flow, the accuracy, precision, recall, and F1 score of the concentration-rate dual-flow model reach 93.72%, 89.92%, 90.85%, and 90.31%, respectively. The accuracy of each metric is approximately 5 percentage points higher than that of the concentration-flow model, validating that the gas production rate effectively complements concentration features in characterizing fault development speed and short-term abrupt changes.

[0122] After further introducing a fault stage mechanism graph into the two-stream structure, the various indicators of the mechanism graph two-stream network continued to improve to 95.01%, 91.79%, 92.45%, and 92.10%, respectively. This indicates that using gas-related edges as feature extraction units and propagating information according to fault stage mechanism constraints is more effective in reflecting the essential laws of fault evolution than overall feature extraction that does not distinguish gas-related relationships. The accuracy of the mechanism graph gating fusion model is 94.82%, and the F1 score is 91.55%, slightly lower than that of the mechanism graph two-stream network. This is because gating without introducing mechanism consistency constraints is a general weighted adjustment, which has limited ability to distinguish between noise samples and boundary samples. This invention introduces a mechanism-side response consistency gating based on the mechanism diagram and the two-flow structure. The gating coefficient is adjusted by negative feedback through the consistency difference between the concentration response and the gas production rate response. The evolution direction is then verified by combining the stage-progressive response term. Finally, an accuracy of 95.56%, precision of 93.02%, recall of 93.73%, and F1 score of 93.36% are achieved, all of which are the highest values ​​among all models.

[0123] Combination Figure 6 , Figure 7 and Figure 8 It can be seen that the gas growth patterns of different fault types exhibit distinguishable forms in both concentration accumulation and gas production rate dimensions. The performance comparison results quantitatively confirm the correspondence between technical components and diagnostic performance: the introduction of concentration time-series information solves the problem that static information cannot reflect the fault development stage; the addition of gas production rate flow supplements the characterization of fault development speed; the fault stage mechanism graph enables feature extraction to be implemented on gas correlation edges that conform to the mechanism; and mechanism consistency gating suppresses the interference of isolated spikes and abnormal fluctuations. The step-by-step superposition of various technical components brings about a stable improvement in performance, and the method of this invention still has an accuracy improvement of about 0.5 percentage points compared with the mechanism graph dual-stream network that only lacks consistency gating. This indicates that the mechanism edge response consistency gating plays an actual discriminative gain on noisy samples and boundary samples, and the overall diagnostic performance of the method provided by the embodiments of this invention is effectively verified.

[0124] Figure 9The final fault classification confusion matrix of the method provided in this embodiment of the invention is shown on a simulated DGA time-series dataset. Rows in the matrix correspond to the true categories, columns to the predicted categories, and cell values ​​represent the number of samples of that true category predicted as belonging to the corresponding category and the percentage of rows. From the overall distribution of the confusion matrix, samples from all six categories are consistently classified as their respective true categories, with diagonal elements dominating each row. A total of 185 normal state samples were correctly identified, achieving a 100% recognition rate. No instances of misclassifying normal states as faults were observed, indicating that the model has a good ability to distinguish between stable fluctuations in gas concentration and fault-related increases in normal samples, and will not generate false alarms due to normal fluctuations in field applications. The recognition rate for high-temperature overheating samples was 96.4%, the highest among the fault categories. The recognition rates for low-temperature and medium-temperature overheating were 93.3% and 90.9%, respectively. The recognition rates for low-energy discharge and high-energy discharge were 92.1% and 89.6%, respectively, with the recognition rates for each fault category remaining at a high level. The distribution of misclassifications shows that almost all misclassifications occurred between adjacent fault levels: both misclassifications of low-temperature overheating fell into medium-temperature overheating; one misclassification of medium-temperature overheating fell into low-temperature overheating and two into high-temperature overheating; two misclassifications of high-temperature overheating fell into both low-temperature and medium-temperature overheating; five misclassifications of low-energy discharge fell into high-energy discharge; and eight misclassifications of high-energy discharge fell into low-energy discharge. Adjacent fault levels exhibit continuity in characteristic gas composition and evolution trends. There are no abrupt gas combination boundaries between low-temperature, medium-temperature, and high-temperature overheating, and the transition between low-energy and high-energy discharge is also gradual. The concentration of misclassifications in adjacent levels precisely demonstrates that the model's discrimination logic conforms to the objective laws of fault evolution. It is particularly noteworthy that there were almost no cross-type misjudgments between thermal faults and discharge faults. Only one case of high-temperature overheating was misjudged as high-energy discharge in all samples, accounting for 0.7%. This indicates that the fault stage mechanism diagram effectively plays its role in distinguishing between the overheating mechanism edge and the discharge mechanism edge, and the model will not confuse the fundamental nature of the fault due to similar absolute values ​​of gas concentration.

[0125] The results of the overall confusion matrix show that the method provided by the embodiments of the present invention has a good recognition effect on the main categories such as normal state, high temperature overheating, low energy discharge and high energy discharge. Among them, all normal state samples are correctly identified, and the recognition rate of each type of fault remains above 89%. A small number of misclassifications mainly occur between adjacent thermal fault levels such as low temperature overheating, medium temperature overheating and high temperature overheating, as well as between low energy discharge and high energy discharge. The misclassification pattern is consistent with the similarity of adjacent fault categories in characteristic gas composition and evolution trend, indicating that the classification results of the method provided by the embodiments of the present invention have good engineering rationality and interpretability.

[0126] The above description is merely an embodiment of this application and is not intended to limit the scope of protection of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of protection of this application.

Claims

1. A transformer DGA diagnostic method based on a mechanistic graph-gated two-current network, characterized in that, The method includes: Acquire concentration data of multiple characteristic gases in transformer oil at multiple sampling times; The gas production rate of each characteristic gas is determined based on the concentration change of each characteristic gas between adjacent sampling times. A concentration stream is constructed using the concentration data of each characteristic gas, and a rate stream is constructed using the gas production rate of each characteristic gas. The concentration stream and the rate stream are then input into a dual-stream network. Using the dual-flow network, a fault stage mechanism diagram is constructed based on the fault stage mechanism, with multiple characteristic gases as graph nodes. The gas association edges in the fault stage mechanism diagram are used to characterize the association relationships between corresponding characteristic gases that conform to the fault stage mechanism. Through the dual-stream network, for each gas-related edge in the fault stage mechanism diagram, the concentration response features of the gas-related edge are extracted based on the concentration data of the gas at both ends of the gas-related edge, and the gas production rate response features of the gas-related edge are extracted based on the gas production rate of the gas at both ends of the gas-related edge. Through the dual-stream network, for each gas-related edge in the fault stage mechanism diagram, the gating coefficient of the gas-related edge is determined according to the consistency between the concentration response characteristics and the gas production rate response characteristics of the gas-related edge. The concentration response characteristics and the gas production rate response characteristics of the gas-related edge are then fused according to the gating coefficient to obtain the gating fusion characteristics of the gas-related edge. The fault diagnosis result of the transformer is determined based on the gating fusion characteristics of each gas-associated edge.

2. The transformer DGA diagnostic method based on mechanistic graph-gated two-stream network according to claim 1, characterized in that, The method of constructing a fault stage mechanism diagram using multiple characteristic gases as graph nodes and based on the fault stage mechanism includes: Using multiple characteristic gases as graph nodes, an overheating mechanism subgraph is constructed based on the gas generation mechanism during the overheating fault stage; A discharge mechanism sub-diagram is constructed based on the gas generation mechanism during the discharge fault stage; A sub-diagram of the overheating-discharge transition mechanism is constructed based on the gas generation mechanism in the overheating-discharge transition stage; Based on the concentration data and gas production rate of each characteristic gas, the combined weights of the superheating mechanism subgraph, the discharge mechanism subgraph, and the superheating-discharge transition mechanism subgraph are determined. The overheating mechanism sub-diagram, the discharge mechanism sub-diagram, and the overheating-discharge transition mechanism sub-diagram are weighted and combined according to the combined weights to obtain the fault stage mechanism diagram.

3. The transformer DGA diagnostic method based on mechanistic graph-gated two-stream network according to claim 2, characterized in that, The step of determining the combined weights of the superheating mechanism subgraph, the discharge mechanism subgraph, and the superheating-discharge transition mechanism subgraph based on the concentration data and gas production rate of each characteristic gas includes: Based on the concentration data of each characteristic gas at multiple sampling times, determine the current concentration and concentration change of each characteristic gas; Based on the gas production rate of each characteristic gas between adjacent sampling times, determine the average gas production rate and the trend of gas production rate change of each characteristic gas; Determine the key gas ratio based on the current concentration of each characteristic gas; Based on the current concentration of each characteristic gas, the amount of concentration change, the average gas production rate, the trend of gas production rate change, and the ratio of key gases, the combined weights of the superheating mechanism sub-graph, the discharge mechanism sub-graph, and the superheating-discharge transition mechanism sub-graph are determined.

4. The transformer DGA diagnostic method based on mechanistic graph-gated two-stream network according to claim 1, characterized in that, The method further includes: For each gas-related edge in the fault stage mechanism diagram, a stage progression response term is extracted based on the concentration change and gas production rate of the gas at both ends of the gas-related edge and the preset fault stage progression direction corresponding to the gas-related edge; wherein, the stage progression response term is used to characterize whether the change relationship of the gas at both ends of the gas-related edge conforms to the fault development law under the corresponding fault stage. The step of determining the gating coefficient of the gas-related edge based on the consistency between the concentration response characteristics and the gas production rate response characteristics of the gas-related edge includes: The gating coefficient of the gas-related edge is determined based on the degree of consistency between the concentration response characteristics and the gas production rate response characteristics of the gas-related edge and the stage progressive response term of the gas-related edge.

5. The transformer DGA diagnostic method based on mechanistic graph-gated two-stream network according to claim 4, characterized in that, The method for setting the preset fault stage progression direction includes: For the gas-related edge corresponding to the overheating fault stage, the preset fault stage progression direction is the direction in which the concentration of the heat fault-related gas increases as the fault develops; For the gas-related edge corresponding to the discharge fault stage, the preset fault stage progression direction is the direction in which the concentration of the discharge-related gas increases as the fault develops; For the gas-associated edge corresponding to the overheat-discharge transition stage, the preset fault stage progression direction is the direction in which the concentration of discharge-related gas increases as the fault develops; The step of extracting the stage progression response term of the gas-related edge based on the concentration change and gas production rate of the gas at both ends of the gas-related edge and the preset fault stage progression direction corresponding to the gas-related edge includes: When the change relationship of the gas at both ends of the gas-associated edge conforms to the preset fault stage progression direction, the stage progression response term is determined to take a positive value. When the change relationship of the gas at both ends of the gas-associated edge is opposite to the progression direction of the preset fault stage, the stage progression response term is determined to take a negative value.

6. The transformer DGA diagnostic method based on mechanistic graph-gated two-stream network according to claim 1, characterized in that, The step of determining the gating coefficient of the gas-related edge based on the consistency between the concentration response characteristics and the gas production rate response characteristics of the gas-related edge includes: Determine the difference between the concentration response characteristics and the gas production rate response characteristics of the gas associated edge; wherein the difference is used to characterize the degree of inconsistency between the concentration response characteristics and the gas production rate response characteristics; The gating coefficient of the gas-related edge is determined based on the difference; wherein the value of the gating coefficient decreases as the difference increases, so as to suppress the influence of the gating fusion feature of the gas-related edge on the fault diagnosis result when the concentration response feature and the gas production rate response feature are inconsistent.

7. The transformer DGA diagnostic method based on a mechanistic graph-gated two-stream network according to any one of claims 1 to 6, characterized in that, The step of determining the transformer fault diagnosis result based on the gating fusion characteristics of each gas-related edge includes: Determine the edge weight of each gas-related edge in the fault stage mechanism diagram; wherein, the edge weight is used to characterize the degree of matching between the corresponding gas-related edge and the current fault stage, and the larger the edge weight, the higher the degree of matching between the gas-related edge and the current fault stage; According to the edge weight of each gas-related edge, the gated fusion features of each gas-related edge are weighted and aggregated to obtain fault discrimination features; Based on the fault discrimination characteristics, the fault diagnosis result of the transformer is determined.

8. The transformer DGA diagnostic method based on a mechanistic graph-gated two-stream network according to any one of claims 1 to 6, characterized in that, The acquisition of concentration data of multiple characteristic gases in transformer oil at multiple sampling times includes: Obtain the original concentration sequences of each characteristic gas in transformer oil; The original concentration sequence is filled with missing values ​​and abnormal spikes are suppressed to obtain the preprocessed concentration sequence; The time windows are moved sequentially along the sampling times, and the concentration data within each time window is captured to obtain the concentration data of each characteristic gas at multiple consecutive sampling times.

9. The transformer DGA diagnostic method based on a mechanistic graph-gated two-stream network according to any one of claims 1 to 6, characterized in that, The step of determining the gas production rate of each characteristic gas based on the concentration change of each characteristic gas between adjacent sampling times includes: The absolute gas production rate of each characteristic gas is determined based on the concentration change of each characteristic gas between adjacent sampling times; wherein, the absolute gas production rate is used to characterize the amount of gas concentration increase per unit time. Based on the concentration data of each characteristic gas at the later sampling time and the concentration data at the previous sampling time in two adjacent sampling times, the relative gas production rate of each characteristic gas is determined; wherein, the relative gas production rate is used to characterize the change in gas concentration relative to the concentration value at the previous sampling time.

10. The transformer DGA diagnostic method based on a mechanistic graph-gated two-stream network according to any one of claims 1 to 6, characterized in that, The method further includes: Construct training samples; wherein the training samples are used for end-to-end training of the two-stream network; The construction of training samples includes: The concentration data of dissolved gases in transformer oil from real samples were used as the concentration data of the fault evolution sequence at the final sampling time. According to the evolution law of gas concentration with fault development, the concentration data of multiple sampling times before the endpoint sampling time are generated, and the corresponding gas production rate is determined according to the concentration change between adjacent sampling times to obtain the simulation time series sample. The fault category label of the real sample is determined as the fault category label of the simulated time series sample.

Images