Transformer potential fault mode identification method and device based on reverse derivation

Abstract

The invention is suitable for the field of transformer fault analysis, and provides a transformer potential fault mode identification method and device based on reverse derivation, and the method comprises the steps: generating a multi-field coupled feature puzzle model through real-time collection of four-dimensional physical quantity spatio-temporal data of temperature, vibration, oil chromatography, partial discharge and the like, and carrying out the topological matching of the feature puzzle model and a historical health file, and reversely deducing a fault coupling path for the abnormal vacancy region through a constraint satisfaction algorithm, reconstructing a missing feature block, decoupling and extracting interface abnormal features, dynamically optimizing model topology in combination with an equipment aging factor, and finally quantifying the full-life-cycle risk accumulation intensity and positioning a composite fault source. The method breaks through the limitation of a traditional single-parameter threshold value, achieves the early recognition and precise traceability of the multi-field coupling fault through a reverse derivation mechanism of the feature jigsaw blocks, is suitable for the health management of the whole life cycle of the transformer, and can remarkably improve the timeliness and accuracy of potential fault early warning.

Description

Technical Field

[0001] The present invention belongs to the field of transformer fault analysis, and particularly relates to a method and device for identifying potential fault modes of transformers based on reverse derivation. Background Technique

[0002] As a core device of power equipment, the early identification of potential faults in transformers is crucial for the safe operation of the power grid. The current industry has developed from single-parameter monitoring to multi-physical-field (temperature, vibration, oil chromatography, partial discharge, etc.) fusion monitoring, collecting multi-dimensional data in real time through a sensor array, and realizing state assessment with the help of machine learning algorithms. However, when dealing with the early signs of multi-field coupling faults, the existing technologies still face challenges such as insufficient depth of feature correlation analysis, poor adaptability of the model to equipment aging, and difficulty in quantifying gradual risks.

[0003] The existing technologies mainly adopt a single physical quantity threshold alarm mechanism or a static pattern matching method based on historical data. For example, judging faults by setting the threshold of the characteristic gas concentration in oil chromatography, or classifying a single vibration spectrum using a neural network. Some solutions attempt to construct a multi-parameter correlation model, but mostly rely on fixed topology rules, do not achieve dynamic deduction based on the coupling mechanism of physical fields, and lack the ability of reverse derivation for missing features.

[0004] The existing technologies cannot effectively capture the early signs of multi-physical-field coupling anomalies. Single-parameter threshold alarms are prone to false negatives or false positives; static topology models are difficult to adapt to the changes in the interaction relationships of physical fields during the aging process of equipment; there is a lack of quantification means for the risk accumulation process of gradual faults and it is impossible to achieve trend prediction throughout the life cycle; in the face of compound faults, it is difficult to locate the true fault source, and diagnostic deviations are often caused by apparent anomalies. Summary of the Invention

[0005] The purpose of the present invention is to provide a method and device for identifying potential fault modes of transformers based on reverse derivation, aiming to solve the technical problems existing in the prior art as determined in the background technique.

[0006] The present invention is implemented as follows. A method for identifying potential fault modes of transformers based on reverse derivation, the method includes: Collecting transformer physical quantity monitoring data in real time and parsing it into a spatio-temporally correlated standardized feature module. Each standardized feature module represents the correlation pattern of a specific physical quantity in the spatio-temporal dimension, and generating a feature puzzle model by topologically connecting all standardized feature modules according to the multi-physical-field coupling rules; Topologically matching the real-time feature puzzle model with the historical health file of the transformer. When detecting the absence of feature puzzle pieces or deviation from the established topological structure, marking the abnormal vacant area and triggering the reverse filling engine; Deduce the fault coupling path that leads to the abnormal vacancy area through the constraint satisfaction algorithm, and reversely generate the potential form of the missing feature puzzle piece; Decouple the multi-physical field signals of the reversely generated missing feature puzzle piece, extract the abnormal features of the interface effect, and feedback them to the feature puzzle model; Dynamically adjust the topological connection rules of the feature puzzle model according to the equipment aging index, and control the self-organization and optimization of the feature puzzle model; Establish an association mapping channel between the current missing feature puzzle piece and the historical operation fluctuation, and quantify the intensity of the gradual risk accumulation in the whole life cycle of the transformer; Based on the reverse puzzle verification mechanism, output the positioning result of the composite fault source.

[0007] As a further solution of the present invention, the real-time acquisition of the transformer physical quantity monitoring data and the generation of the feature puzzle model specifically include: Synchronously collect the spatio-temporal sequence data of four-dimensional physical quantities of temperature, vibration, oil chromatography, and partial discharge amount, and generate a multi-dimensional monitoring time series matrix; Extract the frequency domain energy entropy of each physical quantity as the basic feature vector through wavelet packet decomposition, and calculate the spatio-temporal correlation degree between physical quantities in combination with mutual information entropy; Set the correlation degree threshold, and encode the combination of basic feature vectors with a spatio-temporal correlation degree higher than the correlation degree threshold into a standardized puzzle piece; Establish the connection rules of the puzzle pieces based on the physical field coupling mechanism, and assemble all the standardized puzzle pieces according to the puzzle piece connection rules to form a feature puzzle model.

[0008] As a further solution of the present invention, the marking of the abnormal vacancy area and the triggering of the reverse filling engine specifically include: Construct a topological graph of the historical health record based on the graph convolutional network, where the nodes represent the feature puzzle pieces and the edge weights represent the association strength between modules; Obtain the real-time feature puzzle model, and calculate the structural similarity index between the feature puzzle model and the topological graph : ; Among them, represents the node set of the real-time feature puzzle model, represents the node set of the historical graph, represents the edge set of the historical graph, represents the historical edge weight, represents the corresponding edge weight in the real-time feature puzzle model, is the weight coefficient; When < 0.85 and the area of the vacancy area > 5% of the total area of the graph, it is determined as an abnormal vacancy area; Generate an inverse filling task instruction set that includes the position coordinates of abnormal vacancy areas and the associated constraints of adjacent modules.

[0009] As a further aspect of the present invention, the potential forms of the missing feature puzzle pieces generated inversely specifically include: Establish a directed graph of fault propagation centered on the abnormal vacancy area, with nodes being fault transformer components and edges being fault transfer probabilities; Adopt the Monte Carlo constraint satisfaction algorithm to iteratively generate candidate filling schemes under the condition of satisfying the physical field boundary conditions of adjacent standardized puzzle pieces; Verify the energy conservation of candidate paths through multi-physics field coupling simulation, screen the filling scheme set with the highest fault probability, and generate reconstruction parameters of the missing module based on the filling scheme set to generate reconstructed standardized puzzle pieces.

[0010] As a further aspect of the present invention, the multi-physics field signal decoupling of the inversely generated missing feature puzzle pieces, extracting abnormal interface effect features and feeding them back to the feature puzzle model specifically includes: Decompose the reconstructed standardized puzzle pieces to generate latent fault feature vectors; Inject the latent fault feature vectors as new dimensions into the corresponding standardized puzzle pieces of the feature puzzle model and update the model feature descriptors.

[0011] As a further aspect of the present invention, the dynamic adjustment of the topological connection rules of the feature puzzle model to control the self-organization and optimization of the feature puzzle model specifically includes: Quantify the equipment aging factor α =∫(operation years, load rate, fault history weight); Dynamically adjust the topological rules of the feature puzzle model according to the value of the quantified equipment aging factor; Based on the reinforcement learning mechanism, optimize the topological update strategy with the fault recognition accuracy as the reward function, and the reward function: R = 1 / (false alarm rate + miss rate).

[0012] As a further aspect of the present invention, the quantification of the cumulative intensity of gradual risk in the whole life cycle of the transformer specifically includes: Construct a time-scale mapping function , and align the evolution rates of the current missing feature puzzle pieces with the feature puzzle pieces in the historical health record; Calculate the distribution similarity between the current abnormal pattern and the historical fluctuation pattern through the Wasserstein distance, where the current abnormal pattern refers to the physical field parameter distribution of the inversely reconstructed feature puzzle pieces; Generate a risk accumulation path map and output the risk accumulation intensity of the whole life cycle : ; ; Wherein, represents the risk acceleration, represents the set of physical field parameters of the feature puzzle pieces obtained by reverse reconstruction, represents the health status reference parameter corresponding to the spatial position in the historical health record, represents the operation time of the relay device, is the current time.

[0013] As a further solution of the present invention, the output of the composite fault source localization result specifically includes: Reverse-embed the feature puzzle pieces obtained by reverse reconstruction into the original feature puzzle model and verify the physical field continuity; Compare the deviation degree between the fault probability judged by the single physical quantity threshold and the fault probability deduced by multi-field coupling : ; Wherein, represents the fault probability judged by the single physical quantity threshold, represents the fault probability deduced by multi-field coupling; When the deviation degree > 40%, it is determined that there is an apparent fault, and the composite fault source localization result and the multi-field coupling contribution weight are output.

[0014] Another object of the present invention is to provide a transformer potential fault mode recognition device based on reverse derivation, and the device includes: A topology connection module, configured to collect transformer physical quantity monitoring data in real time and parse it into a spatio-temporally correlated standardized feature module. Each standardized feature module represents the correlation pattern of a specific physical quantity in the spatio-temporal dimension, and generates a feature puzzle model by topologically connecting all the standardized feature modules according to the multi-physical field coupling rule; An abnormal vacancy area marking module, configured to perform topological matching between the real-time feature puzzle model and the historical health record of the transformer. When it detects that a feature puzzle piece is missing or deviates from the established topological structure, it marks the abnormal vacancy area and triggers the reverse filling engine; A missing feature generation module, configured to deduce the fault coupling path that causes the abnormal vacancy area to appear through a constraint satisfaction algorithm, and reversely generate the potential form of the missing feature puzzle piece; An abnormal feature extraction module, configured to perform multi-physical field signal decoupling on the reversely generated missing feature puzzle piece, extract the interface effect abnormal feature and feed it back to the feature puzzle model; A topology rule adjustment module, configured to dynamically adjust the topology connection rule of the feature puzzle model according to the device aging index, and control the self-organization and optimization of the feature puzzle model; The associated mapping channel establishment module is used to establish the associated mapping channel between the currently missing feature puzzle pieces and the historical operation fluctuations, and quantify the intensity of the gradual risk accumulation in the whole life cycle of the transformer; The result output module is used to output the composite fault source location result based on the reverse puzzle verification mechanism.

[0015] The beneficial effects of the present invention are as follows: Through the closed-loop mechanism of "feature puzzle model construction - abnormal area marking - reverse derivation of missing blocks - multi-field decoupling feedback - dynamic topology optimization", this solution converts the multi-physical field monitoring data of the transformer into standardized feature puzzle pieces with spatio-temporal correlation. When missing blocks are found through the topological matching between the real-time model and the historical health record, the fault coupling path is deduced reversely through the constraint satisfaction algorithm, the potential fault form is reconstructed, and the abnormal features of the interface effect are decoupled and extracted.

[0016] This solution breaks through the limitations of traditional single-parameter analysis and realizes the early identification of multi-field coupling faults: through the topological association of feature puzzle pieces, the early signs of sudden changes in the inter-field association strength when a single parameter does not exceed the threshold can be captured; the reverse derivation mechanism can dynamically generate the potential forms of missing features and quantify the probability of fault propagation; the model topology is dynamically adjusted in combination with the equipment aging factor, enabling the diagnostic ability to adaptively evolve with the equipment life cycle; finally, through the reverse puzzle verification mechanism, the composite fault source is accurately located, solving the deficiencies of the existing technology in multi-field coupling anomaly identification, aging adaptability, and fault tracing, extending the lead time of early fault warning, and significantly improving the operation reliability and condition-based maintenance efficiency of the transformer. Description of the Drawings

[0017] Figure 1 It is the flowchart of the method for identifying potential fault modes of a transformer based on reverse derivation provided by an embodiment of the present invention; Figure 2 It is the flowchart of the real-time acquisition of transformer physical quantity monitoring data and the generation of the feature puzzle model provided by an embodiment of the present invention; Figure 3 It is the flowchart of marking the abnormal vacant area and triggering the reverse filling engine provided by an embodiment of the present invention; Figure 4 It is the flowchart of reversely generating the potential form of the missing feature puzzle piece provided by an embodiment of the present invention; Figure 5 It is the flowchart of extracting the abnormal features of the interface effect and feeding them back to the feature puzzle model provided by an embodiment of the present invention; Figure 6 It is the flowchart of dynamically adjusting the topological connection rules of the feature puzzle model to control the self-organization and optimization of the feature puzzle model provided by an embodiment of the present invention; Figure 7A flowchart for quantifying the cumulative intensity of gradual risks in the whole life cycle of a transformer provided by an embodiment of the present invention; Figure 8 A flowchart for outputting the location result of a composite fault source provided by an embodiment of the present invention; Figure 9 A structural block diagram of a transformer potential fault mode recognition device based on reverse derivation provided by an embodiment of the present invention. Detailed implementation manners

[0018] In order to make the objectives, technical solutions and advantages of the present invention clearer and more understandable, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0019] Figure 1 A flowchart of a method for recognizing potential fault modes of a transformer based on reverse derivation provided by an embodiment of the present invention, as Figure 1 shown, the method includes: S100, collecting transformer physical quantity monitoring data in real time and parsing it into a spatio-temporal correlated standardized feature module. Each standardized feature module represents the correlation pattern of a specific physical quantity in the spatio-temporal dimension, and all standardized feature modules are generated into a feature jigsaw model according to the topological connection rule of multi-physical field coupling; This step synchronously collects spatio-temporal sequence data of four-dimensional physical quantities such as temperature field distribution (such as three-dimensional temperature gradients of windings, iron cores, and oil), mechanical vibration spectrum (covering high-frequency abnormal noises and low-frequency resonance signals under different working conditions), oil chromatographic component concentration (including real-time volume fractions of characteristic gases such as hydrogen, acetylene, and carbon monoxide), and partial discharge time-domain waveforms through a sensor array deployed at key parts of the transformer, and generates a multi-dimensional monitoring time series matrix containing timestamps, spatial coordinates, and physical quantity amplitudes after preliminary preprocessing by an edge computing node.

[0020] In the feature parsing stage, wavelet packet decomposition is used to perform multi-scale decomposition of each physical quantity in the time-frequency domain, and the energy entropy of each frequency band is extracted as a basic feature vector representing the operating state of the device. At the same time, the correlation degree of different physical quantities in the spatio-temporal dimension is quantified through mutual information entropy. For example, when there is a strong correlation between the sudden change in acetylene concentration in the oil chromatogram and the partial discharge pulse on the time axis, it can be determined that there is a potential discharge decomposition coupling relationship between the two.

[0021] Taking a preset correlation degree threshold as a screening condition, the feature vector combinations with strong coupling relationships are encoded into standardized jigsaw blocks. Each jigsaw block not only contains the feature parameters of a single physical quantity but also embeds a cross-physical field correlation weight matrix.

[0022] Based on the coupling mechanism of electromagnetic field - thermal field - mechanical field - chemical field, connection rules for puzzle pieces are established. For example, the thermoelectric coupling rule stipulates the non - linear mapping relationship between the winding temperature rise and the partial discharge quantity. Thus, all standardized puzzle pieces are topologically assembled into a dynamic characteristic puzzle model, which visually presents the spatio - temporal coupling relationship of the multi - physical fields of the transformer in the form of a graph structure. The nodes are characteristic puzzle pieces, and the edges are the interaction intensities of the physical fields.

[0023] This step breaks through the limitations of traditional single - parameter threshold alarms through spatio - temporal fusion modeling of multi - physical fields, realizing the holographic characterization and dynamic correlation analysis of equipment status. Specifically, four - dimensional physical quantities are synchronously collected and a spatio - temporal correlation feature module is constructed, which can capture early fault signs when the abnormality of a single parameter has not yet become prominent. For example, when there is a weak partial discharge inside the transformer, it may first be manifested as a coordinated small change in the high - frequency components in the vibration spectrum and the hydrogen content in the oil chromatogram. Such weak cross - physical - field correlation signals can be effectively identified through mutual information entropy calculation, while traditional single - parameter monitoring often misses detections because the alarm threshold is not reached.

[0024] In addition, the characteristic puzzle model constructs topological connections based on the physical - field coupling mechanism, transforming the abstract equipment status into a visual graph structure. This not only facilitates engineers to intuitively understand the interaction relationships of various physical quantities but also provides a structured carrier for the reverse - derivation fault propagation path. For example, when the temperature - vibration coupling block shows an abnormality, the model can quickly locate the possible fault - source area based on the preset thermo - mechanical coupling rule.

[0025] As Figure 2 shown, the real - time acquisition of transformer physical quantity monitoring data and the generation of the characteristic puzzle model specifically include: S110, synchronously collect the spatio - temporal sequence data of four - dimensional physical quantities of temperature, vibration, oil chromatogram, and partial discharge quantity to generate a multi - dimensional monitoring time - series matrix; S120, extract the frequency - domain energy entropy of each physical quantity as a basic feature vector through wavelet packet decomposition, and calculate the spatio - temporal correlation degree across physical quantities in combination with mutual information entropy; S130, set a correlation - degree threshold, and encode the combination of basic feature vectors with a spatio - temporal correlation degree higher than the correlation - degree threshold into standardized puzzle pieces; S140, establish connection rules for puzzle pieces based on the physical - field coupling mechanism, and assemble all standardized puzzle pieces according to the puzzle - piece connection rules to form a characteristic puzzle model.

[0026] S200, perform topological matching between the real - time characteristic puzzle model and the historical health record of the transformer. When it is detected that a characteristic puzzle piece is missing or deviates from the established topological structure, mark the abnormal vacancy area and trigger the reverse - filling engine; This step first constructs a topological map of the health record based on historical operation data, and uses a graph convolutional network (GCN) to perform deep learning on the spatio-temporal correlation patterns of historical feature puzzle pieces. Here, the nodes correspond to standardized feature modules (such as the temperature-gas chromatography coupling block), and the edge weights are dynamically assigned through the average mutual information entropy between modules in historical data, representing the long-term coupling strength of physical fields such as mechanical vibration and partial discharge.

[0027] In the real-time matching stage, through the graph isomorphism algorithm, the current feature puzzle model is aligned with the historical topological map in terms of node alignment and edge weight comparison. When calculating the structural similarity index SSI, the weight coefficient ω can be dynamically adjusted according to the device type (such as for oil-immersed transformers ω = 0.6 to emphasize node matching, and for dry-type transformers ω = 0.4 to emphasize edge association). When SSI < 0.85 and the proportion of the vacant area exceeds 5%, the device not only marks the coordinates of the abnormal area, but also extracts the physical field constraint conditions of adjacent modules (such as the temperature module requires that the gas generation rate gradient of the adjacent gas chromatography module ≤ 0.5 ppm / °C), and generates a reverse filling task package containing spatial position, association constraints, and historical coupling probability.

[0028] When constructing the topological map, incorporating historical operation fluctuation data can capture early fault signs where a single parameter does not exceed the threshold but the associated structure has shifted. For example, when the transformer core is slightly loose, the change in the high-frequency component of the vibration spectrum may not have reached the threshold yet, but the edge weight associated with the temperature module will decrease significantly due to abnormal mechanical-thermal coupling. This structural change can be identified in advance through SSI calculation.

[0029] In addition, the SSI index takes into account both node existence and edge association strength, and can effectively distinguish normal operation fluctuations from structural abnormalities in the initial stage of a fault. For example, the temperature increase caused by a sudden load change will maintain the integrity of the topological structure, while the abnormalities caused by inter-turn short circuits will be accompanied by node deletion and sudden changes in edge weights, and the SSI value can suddenly drop from the normal 0.92 to 0.78.

[0030] Such as Figure 3 shown, marking the abnormal vacant area and triggering the reverse filling engine specifically includes: S210, constructing a topological map of the historical health record based on the graph convolutional network, where the nodes represent feature puzzle pieces and the edge weights represent the association strength between modules; S220, obtaining the real-time feature puzzle model and calculating the structural similarity index between the feature puzzle model and the topological map : ; Among them, represents the set of nodes of the real-time feature puzzle model, represents the set of nodes of the historical map, Represents the set of edges in the historical graph. Represents the weights of historical edges. Represents the corresponding edge weights in the real-time feature puzzle model. Is the weight coefficient. S230, when the structural similarity index < 0.85 and the area of the vacant region > 5% of the total area of the graph, it is determined as an abnormal vacant region. S240, generate a reverse filling task instruction set including the position coordinates of the abnormal vacant region and the associated constraints of adjacent modules.

[0031] S300, deduce the fault coupling path that leads to the abnormal vacant region through the constraint satisfaction algorithm, and reversely generate the potential form of the missing feature puzzle pieces. This step takes the abnormal vacant region as the topological center, constructs a directed graph of fault propagation including key components such as transformer windings, iron cores, and conservators. The edge weights in the graph are assigned by the conditional probability matrix trained from historical fault data (for example, the transfer probability from winding overheating to insulation aging is 0.72).

[0032] When using the Monte Carlo constraint satisfaction algorithm, first randomly sample the physical field boundary conditions of adjacent standardized puzzle pieces (such as the field strength gradient ≤ 1.2 kV / mm for the adjacent discharge module required by the temperature module), generate more than 1000 sets of candidate filling schemes, and each set of schemes includes a combination of faulty components, parameter distributions, and propagation paths.

[0033] Subsequently, through the multi-physical field coupling simulation platform, verify the energy conservation of the candidate schemes. When simulating the turn-to-turn short circuit fault, it is necessary to verify the energy conversion balance between the winding temperature field distribution and the gas production rate of oil chromatography (the error needs to be < 5%). Finally, screen out the top 10% of the scheme sets with the highest fault probabilities, and generate standardized puzzle pieces containing temperature-discharge-chromatography coupling features based on their reconstructed parameters. For example, a non-linear mapping function (y = 0.3x² + 1.5x) of the discharge amount and acetylene concentration is embedded in the reconstructed block.

[0034] Combined with historical fault probability data, the directed graph of fault propagation can quantify the likelihood of the chain reaction of faults in different components. For example, when detecting an abnormal vacancy in the oil temperature, the algorithm can simultaneously deduce potential fault paths such as winding overheating (probability 0.65) and abnormal hysteresis loss of the iron core (probability 0.28), rather than a single attribution.

[0035] The combination of the Monte Carlo algorithm and physical field simulation makes the generated candidate schemes satisfy both randomness and physical laws, avoiding the rigid problems of traditional rule-based reasoning.

[0036] Such as Figure 4 shown, the specific potential form of reversely generating the missing feature puzzle pieces includes: S310. Establish a fault propagation directed graph centered on the abnormal vacancy area, with the nodes being the faulty transformer components and the edges being the fault transfer probabilities. S320. Adopt the Monte Carlo constraint satisfaction algorithm to iteratively generate candidate filling schemes under the condition of satisfying the physical field boundary conditions of adjacent standardized puzzle pieces. S330. Verify the energy conservation of the candidate paths through multi-physical field coupling simulation, screen the set of filling schemes with the highest fault probabilities, and generate the reconstruction parameters of the missing module based on the set of filling schemes to generate the reconstructed standardized puzzle pieces.

[0037] S400. Decouple the multi-physical field signals of the reversely generated missing feature puzzle pieces, extract the abnormal features of the interface effect, and feedback them to the feature puzzle model. In this step, the tensor decomposition algorithm is used to couple and separate the temperature-vibration-oil chromatography-discharge four-dimensional physical field signals contained in the reconstructed standardized puzzle pieces. For example, the mixed signal is decomposed into independent heat conduction components, mechanical vibration components, gas diffusion components, and electromagnetic discharge components, and then the abnormal feature parameters of the interface effect between each field are extracted, such as the gas generation rate gradient at the temperature-oil chromatography interface and the energy transfer efficiency attenuation coefficient at the vibration-discharge interface.

[0038] When generating the latent fault feature vector, the blind source separation technology is used to strip out the hidden features representing the interface failure from the coupled signals. When the insulation of the transformer winding deteriorates, it can be found that there is an abnormal coupled signal with a characteristic frequency of 10 kHz at the interface between the temperature field and the discharge field after decoupling, and this signal is easily submerged by noise in traditional single-field analysis.

[0039] Subsequently, these feature vectors containing abnormal interface effects are injected into the corresponding modules of the feature puzzle model as new dimensions, and the feature descriptors of the model are updated synchronously. For example, an "insulation interface deterioration index" attribute is added to the temperature-discharge coupling block, and this attribute is dynamically calculated through the non-linear mapping relationship between the discharge energy and the temperature gradient obtained by decoupling.

[0040] The signal decoupling process in this step can decompose the composite signal generated by multi-field interaction into independent components, so as to capture the subtle abnormal changes at the interface. When there is a slight looseness at the mechanical interface between the transformer core and the winding, the decoupled vibration signal will show an energy concentration phenomenon in a specific frequency band (such as 200 - 500 Hz), and this hidden feature is difficult to identify by traditional time-domain analysis.

[0041] In addition, feeding the decoupled latent features back to the model enables the feature puzzle model to have dynamic learning ability. After decoupling and discovering the abnormal generation rate of C2H2 at the interface between oil chromatography and temperature, the model automatically updates the association rules of this coupling block, adjusting the mapping relationship between C2H2 concentration and temperature from linear y = 0.2x to non-linear y = 0.2x + 0.05x², so as to more accurately identify similar faults in subsequent monitoring.

[0042] In the practical application of power devices, when the method has slight abnormalities in the main transformer oil chromatography data (acetylene concentration of 22 ppm), by decoupling the gas generation rate characteristics at the temperature-chromatography interface, it discovers the non-linear coupling abnormality between winding partial discharge and oil temperature rise, and locates the micro-breakage fault of the winding turn insulation 56 hours earlier than traditional single-parameter analysis, avoiding sudden tripping accidents caused by the spread of faults, which reflects the key role of multi-physical field decoupling in early fault mechanism identification and model adaptive optimization.

[0043] As Figure 5 shown, the multi-physical field signal decoupling of the reverse-generated missing feature puzzle blocks, extracting the abnormal interface effect features and feeding them back to the feature puzzle model specifically includes: S410, decomposing the reconstructed standardized puzzle block to generate latent fault feature vectors; S420, injecting the latent fault feature vectors as new dimensions into the corresponding standardized puzzle blocks of the feature puzzle model and updating the model feature descriptors.

[0044] S500, dynamically adjusting the topological connection rules of the feature puzzle model according to the equipment aging index to control the self-organization and optimization of the feature puzzle model; This step constructs a multi-dimensional aging assessment system, quantifying parameters such as the operation years (the aging weight of transformers over 15 years increases by 20%), the long-term load rate (the aging acceleration coefficient is 1.5 when exceeding 80% of the rated value), and the number of historical faults (the aging factor is superimposed by 0.1 for each internal fault) through integral operations into the equipment aging factor α , when α > 0.7, triggering the dynamic adjustment mechanism of topological rules: Expanding the association radius of the temperature-oil chromatography puzzle block from the original d1 to 1.3d1, enabling the model to capture the delayed association between temperature and gas generation caused by the decrease in heat conduction efficiency in aging equipment; At the same time, expanding the angular constraint of the mechanical-chemical coupling block from θ ∈[30°, 60°] to [20°, 70°] to adapt to the vibration mode offset of the aging iron core due to the loosening of silicon steel sheets.

[0045] In the reinforcement learning optimization stage, guided by the fault recognition accuracy rate, the false alarm rate and the missed alarm rate data are fed back in real time to iteratively adjust the topology update strategy. For example, when it is found that the missed alarm rate of a certain type of aging fault continuously exceeds 15%, the associated weight threshold of the corresponding coupling block is automatically increased, making the model more sensitive to aging-related anomalies.

[0046] The aging factor quantization mechanism converts the gradual change process of the device health state into computable model parameters, enabling the topology rules to automatically evolve with the degree of device aging. When the thermal-chromatography correlation of a transformer weakens due to the deterioration of insulating oil, the model can α trigger rule adjustment by the

[0047] value (0.75), reducing the association threshold of the temperature-oil chromatography block from 0.6 to 0.5, so as to continue to effectively capture the weak coupling anomalies between the oil temperature and the gas concentration.

[0048] For example Figure 6 As shown, the topological connection rules of the dynamic adjustment feature puzzle model are controlled to self-organize and optimize the feature puzzle model, specifically including: S510, quantifying the device aging factor α =∫(operation years, load rate, fault history weight); S520, dynamically adjusting the topological rules of the feature puzzle model according to the value of the quantified device aging factor: When α >0.7, allow the association radius of the puzzle block to expand by 30%, and the association threshold τ is reduced by 15%; When α >0.7, the maximum distance between the thermal-electric coupling blocks is relaxed from d1 to 1.3d1; The angle constraint of the mechanical-chemical coupling block is extended to θ ∈[20°,70°]; S530, based on the reinforcement learning mechanism, optimize the topology update strategy with the fault recognition accuracy rate as the reward function, and the reward function: R = 1 / (false alarm rate + missed alarm rate).

[0049] S600, establish the association mapping channel between the currently missing feature puzzle block and the historical operation fluctuations, and quantify the intensity of the gradual risk accumulation in the whole life cycle of the transformer; This step constructs a time-scale mapping function to non-linearly align the parameters of the currently reverse-generated feature puzzle pieces with the corresponding modules in the historical health records in the time dimension, solving the problem of inconsistent data rates at different operating stages of the device (such as the difference in temperature change rates between the initial operation stage and the aging stage).

[0050] When calculating the distribution similarity between the current abnormal pattern and the historical fluctuation pattern using the Wasserstein distance, not only the mean deviation of the physical quantity parameters is considered, but also the overall shift of the probability density function is quantified. When the distribution of the hydrogen concentration in the oil chromatogram changes from unimodal to bimodal, this distance can sensitively capture the potential fault pattern transition. In the risk accumulation calculation, the risk acceleration is characterized by the second derivative a(t), and the inflection point of the risk mutation can be identified (for example, when a(t) rises from 0.05 to 0.12, it indicates that the fault development rate accelerates). Then, through the integral operation F, a risk accumulation path diagram from the commissioning time t0 to the current time t is generated. This diagram superimposes the evolution trajectories of multi-physical field parameters such as temperature and discharge, intuitively presenting the development trend of gradual faults.

[0051] The time-scale mapping function solves the problem of comparability of data with different operating years. For example, the vibration data of transformers in operation for 5 years and 15 years are unified to the same time scale through logarithmic transformation, enabling the historical fluctuation pattern to accurately map the current anomaly; The Wasserstein distance is more suitable for analyzing the fault characteristics of asymmetric distributions compared to the Euclidean distance. In addition, the introduction of risk acceleration enables the device to capture the non-linear characteristics of fault development, issuing a replacement warning earlier than traditional life assessments, and avoiding sudden failures caused by insulation aging.

[0052] Such as Figure 7 As shown, the quantification of the intensity of gradual risk accumulation in the whole life cycle of the transformer specifically includes: S610, constructing a time-scale mapping function , aligning the evolution rates of the currently missing feature puzzle pieces with the feature puzzle pieces in the historical health records; S620, calculating the distribution similarity between the current abnormal pattern and the historical fluctuation pattern using the Wasserstein distance, where the current abnormal pattern refers to the distribution of the physical field parameters of the reverse-reconstructed feature puzzle pieces; S630, generating a risk accumulation path diagram and outputting the intensity of risk accumulation in the whole life cycle : ; ; Among them, represents the risk acceleration, represents the set of physical field parameters of the reverse-reconstructed feature puzzle pieces, Represents the health status reference parameters corresponding to the spatial positions in the historical health records, Represents the operation time of the relay device, Is the current time.

[0053] S700, based on the reverse jigsaw verification mechanism, outputs the composite fault source location result.

[0054] This step reversely embeds the feature jigsaw blocks of the reverse reconstruction (such as the module containing temperature-discharge-chromatography coupling features) into the abnormal vacancy areas of the original feature jigsaw model, and verifies the physical field continuity after embedding through the finite element simulation platform (such as the temperature field gradient change rate needs to be <0.3 °C / mm, and the discharge field strength distribution needs to satisfy the Poisson equation), ensuring the self-consistency of the field interaction logic between the reconstructed module and the existing model.

[0055] In the deviation calculation stage, the single physical quantity threshold judgment engine (such as judging the fault probability only based on the 30 ppm threshold of the acetylene concentration in the oil chromatography) and the multi-field coupling deduction engine (synthesizing parameters such as the temperature rise rate and the degree of vibration spectrum distortion) are synchronously called. When the deviation degree Δ of the fault probabilities of the two is >40%, it is determined that there is an apparent fault (such as the oil temperature rise may be caused by both winding overheating and abnormal cooling device). At this time, the output composite fault source location result will include the contribution weight matrix of each physical field (such as 45% contribution of the temperature field, 35% contribution of the discharge field, and 20% contribution of the chromatography field), and highlight the key path of fault propagation through the topological graph (such as the thermal-electric-mechanical coupling chain of winding → oil gap → iron core).

[0056] This step breaks through the one-sidedness of traditional single-parameter diagnosis, and realizes the accurate traceability of composite faults and the identification of apparent faults through multi-field coupling deviation analysis.

[0057] The reverse embedding verification mechanism can test the physical rationality of the reverse reconstruction result, and avoid misjudgment caused by algorithm errors. For example, after decoupling reconstruction, it is found through verification that the energy conservation deviation between the temperature field and the discharge field reaches 8%, thus overthrowing the initial reconstruction plan and re-deducting, and finally locating the real fault source of abnormal hysteresis loss of the iron core.

[0058] The introduction of the deviation degree Δ effectively solves the false alarm problem of single physical quantity alarm. For example, when the cooler fan of the transformer is abnormal, the single parameter may only show that the oil temperature rises (fault probability 60%), while the multi-field coupling deduction finds that there is no abnormality in the fan speed-related frequency band in the vibration spectrum (fault probability 25%), Δ = 58% > 40%, and the device determines that there is an apparent fault and further investigation is needed.

[0059] Such as Figure 8 As shown, the output of the composite fault source location result specifically includes: S710, inversely embed the feature puzzle pieces obtained from inverse reconstruction into the original feature puzzle model, and verify the physical field continuity; S720, compare the deviation degree between the fault probability judged by single physical quantity threshold and the fault probability deduced by multi-field coupling : ; wherein, represents the fault probability judged by single physical quantity threshold, represents the fault probability deduced by multi-field coupling; S730, when the deviation degree > 40%, determine that there is an apparent fault, and output the composite fault source location result and the multi-field coupling contribution weight.

[0060] Figure 9 is the structural block diagram of the transformer potential fault mode recognition device provided by the embodiment of the present invention. As Figure 9 shown, the device includes: The topology connection module 100 is used to collect the transformer physical quantity monitoring data in real time and parse it into a spatio-temporally correlated standardized feature module. Each standardized feature module represents the correlation pattern of a specific physical quantity in the spatio-temporal dimension, and all standardized feature modules are generated into a feature puzzle model according to the multi-field coupling rule topology connection; The abnormal vacancy area marking module 200 is used to topologically match the real-time feature puzzle model with the historical health file of the transformer. When it detects that a feature puzzle piece is missing or deviates from the established topology structure, it marks the abnormal vacancy area and triggers the inverse filling engine; The missing feature generation module 300 is used to deduce the fault coupling path that causes the abnormal vacancy area through the constraint satisfaction algorithm, and inversely generate the potential form of the missing feature puzzle piece; The abnormal feature extraction module 400 is used to decouple the multi-field signals of the inversely generated missing feature puzzle piece, extract the interface effect abnormal features and feedback them to the feature puzzle model; The topology rule adjustment module 500 is used to dynamically adjust the topology connection rule of the feature puzzle model according to the equipment aging index, and control the self-organization optimization of the feature puzzle model; The correlation mapping channel establishment module 600 is used to establish the correlation mapping channel between the current missing feature puzzle piece and the historical operation fluctuation, and quantify the intensity of the gradual risk accumulation in the whole life cycle of the transformer; The result output module 700 is used to output the composite fault source location result based on the inverse puzzle verification mechanism.

[0061] The technical features of the above-described embodiments may be combined arbitrarily. For the sake of brevity of description, not all possible combinations of the technical features in the above-described embodiments are described. However, as long as there is no contradiction in the combination of these technical features, it should be considered as the scope described in this specification.

[0062] The above-described embodiments merely represent several implementation manners of the present invention. The description thereof is relatively specific and detailed, but it should not be construed as a limitation on the scope of the patent for the present invention. It should be noted that for those of ordinary skill in the art, without departing from the concept of the present invention, several modifications and improvements can still be made, and these all belong to the protection scope of the present invention. Therefore, the protection scope of the patent for the present invention shall be subject to the appended claims.

[0063] The foregoing is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent replacements, and improvements made within the spirit and principle of the present invention shall be included within the protection scope of the present invention.

Claims

1. A method for identifying potential fault modes of a transformer based on reverse derivation, characterized in that, The method includes: Collecting real-time monitoring data of transformer physical quantities and parsing them into a standardized feature module with spatio-temporal correlation. Each standardized feature module characterizes the correlation pattern of a specific physical quantity in the spatio-temporal dimension, and generating a feature jigsaw model by topologically connecting all standardized feature modules according to the multi-physical field coupling rule; Topologically matching the real-time feature jigsaw model with the historical health record of the transformer. When missing feature jigsaw pieces or deviation from the established topological structure are detected, mark the abnormal vacancy area and trigger the reverse filling engine; Deducing the fault coupling path that leads to the abnormal vacancy area through a constraint satisfaction algorithm, and reversely generating the potential form of the missing feature jigsaw piece; Performing multi-physical field signal decoupling on the reversely generated missing feature jigsaw piece, extracting abnormal interface effect features and feeding them back to the feature jigsaw model; Dynamically adjusting the topological connection rule of the feature jigsaw model according to the equipment aging index, and controlling the self-organization and optimization of the feature jigsaw model; Establishing an association mapping channel between the currently missing feature jigsaw piece and the historical operation fluctuation, and quantifying the intensity of the cumulative gradual risk in the whole life cycle of the transformer; Based on the reverse jigsaw verification mechanism, outputting the positioning result of the composite fault source.

2. The method according to claim 1, wherein The real-time collection of transformer physical quantity monitoring data and the generation of the feature jigsaw model specifically include: Synchronously collecting the spatio-temporal sequence data of four-dimensional physical quantities of temperature, vibration, oil chromatography, and partial discharge amount, and generating a multi-dimensional monitoring time series matrix; Extracting the frequency domain energy entropy of each physical quantity as the basic feature vector through wavelet packet decomposition, and calculating the spatio-temporal correlation degree across physical quantities in combination with mutual information entropy; Setting a correlation degree threshold, and encoding the combination of basic feature vectors with a spatio-temporal correlation degree higher than the correlation degree threshold into standardized jigsaw pieces; Establishing the connection rule of jigsaw pieces based on the physical field coupling mechanism, and assembling all standardized jigsaw pieces according to the connection rule of jigsaw pieces to form a feature jigsaw model.

3. The method according to claim 2, characterized in that The marking of the abnormal vacancy area and triggering the reverse filling engine specifically includes: Constructing a topological map of the historical health record based on a graph convolutional network, where the nodes represent feature jigsaw pieces and the edge weights represent the association strength between modules; Obtain a real-time feature puzzle model and calculate the structural similarity index between the feature puzzle model and the topological map : ； Among them, represents the set of nodes of the real-time feature puzzle model, represents the set of nodes of the historical graph, represents the set of edges of the historical graph, represents the historical edge weight, represents the corresponding edge weight in the real-time feature puzzle model, is the weight coefficient; When <0.85 and the area of the vacant area > 5% of the total area of the atlas, it is determined as an abnormal vacant area; Generating a reverse filling task instruction set including the position coordinates of the abnormal vacancy area and the association constraints of adjacent modules.

4. The method according to claim 3, characterized in that, The reverse generation of the potential form of the missing feature jigsaw piece specifically includes: Establishing a directed graph of fault propagation centered on the abnormal vacancy area, where the nodes are the components of the faulty transformer and the edges are the fault transfer probabilities; Using the Monte Carlo constraint satisfaction algorithm to iteratively generate candidate filling schemes under the condition of satisfying the physical field boundary conditions of adjacent standardized jigsaw pieces; Verifying the energy conservation of the candidate paths through multi-physical field coupling simulation, screening the set of filling schemes with the highest fault probability, and generating the reconstruction parameters of the missing module based on the set of filling schemes to generate a reconstructed standardized jigsaw piece.

5. The method according to claim 4, wherein The multi-physical field signal decoupling of the reversely generated missing feature jigsaw piece, extracting abnormal interface effect features and feeding them back to the feature jigsaw model specifically includes: Decomposing the reconstructed standardized jigsaw piece to generate a latent fault feature vector; Injecting the latent fault feature vector as a new dimension into the corresponding standardized jigsaw piece of the feature jigsaw model, and updating the model feature descriptor.

6. The method according to claim 5, characterized in that Dynamically adjust the topological connection rules of the feature puzzle model to control the self-organization and optimization of the feature puzzle model, specifically including: Quantify the equipment aging factor α = ∫(Years of operation, Load factor, Fault history weight); Dynamically adjust the topological rules of the feature puzzle model according to the value of the quantization device aging factor; Based on the reinforcement learning mechanism, optimize the topology update strategy with the fault recognition accuracy as the reward function, and the reward function: R = 1 / (false alarm rate + miss rate).

7. The method according to claim 6, wherein Quantify the gradual risk accumulation intensity in the whole life cycle of the quantization transformer, specifically including: Construct a time scale mapping function , and align the evolution rates of the currently missing feature puzzle pieces with the feature puzzle pieces in the historical health records; Calculate the distribution similarity between the current abnormal pattern and the historical fluctuation pattern through the Wasserstein distance, where the current abnormal pattern refers to the physical field parameter distribution of the reversely reconstructed feature puzzle block; Generate a risk accumulation path diagram and output the risk accumulation intensity throughout the life cycle : ； ； Among them, represents the risk acceleration, represents the set of physical field parameters of the feature puzzle pieces for reverse reconstruction, represents the health status reference parameters at the corresponding spatial position in the historical health record, represents the operation time of the relay device, is the current time.

8. The method according to claim 7, wherein Output the composite fault source location result, specifically including: Reverse embed the reversely reconstructed feature puzzle block into the original feature puzzle model and verify the physical field continuity; Deviation degree between the failure probability judged by single physical quantity threshold and the failure probability deduced by multi-field coupling : ； Among them, represents the failure probability of single physical quantity threshold judgment, represents the failure probability of multi-field coupling deduction; When the deviation degree > 40%, it is determined that there is an apparent fault, and the composite fault source location result and the multi-field coupling contribution weight are output.

9. Transformer potential fault mode recognition device based on reverse derivation, characterized in that The device includes: A topological connection module, which is used to collect the transformer physical quantity monitoring data in real time and parse it into a spatio-temporally correlated standardized feature module. Each standardized feature module represents the correlation pattern of a specific physical quantity in the spatio-temporal dimension, and generates a feature puzzle model for all standardized feature modules according to the multi-physical field coupling rule topology connection; An abnormal vacancy area marking module, which is used to perform topological matching between the real-time feature puzzle model and the historical health record of the transformer. When it detects that a feature puzzle block is missing or deviates from the established topological structure, it marks the abnormal vacancy area and triggers the reverse filling engine; A missing feature generation module, which is used to deduce the fault coupling path that causes the abnormal vacancy area to appear through the constraint satisfaction algorithm, and reversely generate the potential form of the missing feature puzzle block; An abnormal feature extraction module, which is used to decouple the multi-physical field signals of the reversely generated missing feature puzzle block, extract the interface effect abnormal features and feedback them to the feature puzzle model; A topological rule adjustment module, which is used to dynamically adjust the topological connection rules of the feature puzzle model according to the device aging index, and control the self-organization and optimization of the feature puzzle model; An association mapping channel establishment module, which is used to establish an association mapping channel between the current missing feature puzzle block and the historical operation fluctuation, and quantify the gradual risk accumulation intensity in the whole life cycle of the transformer; A result output module, which is used to output the composite fault source location result based on the reverse puzzle verification mechanism.

Images