CN-121980473-A - Transformer fault feature fusion diagnosis method and system based on multi-mode time sequence data reconstruction

Abstract

The invention provides a transformer fault feature fusion diagnosis method and system based on multi-mode time sequence data reconstruction, which relate to the technical field of power system transformer fault diagnosis, and comprise the steps of firstly, respectively capturing a dissolved gas time sequence waveform, a tank wall vibration time sequence waveform and a sleeve heat map time sequence waveform in oil, inputting a pre-constructed transformer running situation reconstruction network to perform cross-mode physical field space-time alignment processing, generating a transformer multi-physical field evolution track taking a fault energy injection point as an initial reference, and performing topology tracing and tracking processing of fault characteristic clues on the fault characteristic clues to obtain a global fault characteristic topology chain of the transformer, inputting a fault mode semantic analysis network, generating a comprehensive diagnosis descriptor of the transformer, fusing fault type semantic labels and fault space positioning coordinates, extracting a local evolution track fragment according to the comprehensive diagnosis descriptor, generating a maintenance decision triggering instruction containing the residual life consumption rate of the fault, and sending the maintenance decision triggering instruction to a transformer on-line monitoring master station system to ensure safe and stable operation of the transformer.

Inventors

• WANG YANG
• LIAO PENGZHAN
• DING CHENG
• WANG LEI
• WANG WENLIANG
• ZHAO XINHUI
• GUO YUAN
• ZHAO GUANGMING
• YANG BING
• TIAN YULIN
• JIANG YIHAN
• XIE YALONG
• WANG LIANGYU
• WU PEIJUN
• ZHANG HUASONG
• XUE QINGXI

Assignees

• 四川盐源华电新能源有限公司
• 华电科工股份有限公司
• 四川鸿华睿橙智能电气有限公司

Dates

Publication Date

20260505

Application Date

20260403

Claims (10)

1. 1. The transformer fault feature fusion diagnosis method based on multi-mode time sequence data reconstruction is characterized by comprising the following steps of: Acquiring a dissolved gas time sequence waveform, a tank wall vibration time sequence waveform and a sleeve heat map time sequence waveform of oil respectively captured by an oil chromatograph monitor, a vibration acceleration sensor and an infrared thermal imager of the transformer in a continuous operation period; Inputting the dissolved gas time sequence waveform, the tank wall vibration time sequence waveform and the sleeve heat map time sequence waveform in the oil into a pre-constructed transformer running situation reconstruction network to perform cross-mode physical field space-time alignment processing, and generating a transformer multi-physical field evolution track taking a fault energy injection point as an initial reference; Performing topology tracing tracking processing of fault characteristic clues on the transformer multi-physical-field evolution track to obtain a transformer global fault characteristic topology chain comprising fault starting energy release point coordinates, fault path physical-field distortion point sequences and fault ending energy dissipation zone boundaries; Inputting the global fault characteristic topological chain of the transformer into a fault mode semantic analysis network to carry out space-time coupling verification processing on fault symptoms and focus areas, and generating a transformer comprehensive diagnosis descriptor fusing fault type semantic tags and fault space positioning coordinates; And extracting a corresponding local evolution track segment from the transformer multi-physical-field evolution track according to the fault space positioning coordinates in the transformer comprehensive diagnosis descriptor, generating a maintenance decision trigger instruction containing the fault residual life consumption rate based on the local evolution track segment, and transmitting the maintenance decision trigger instruction to a transformer online monitoring master station system.
2. 2. The method for diagnosing transformer fault characteristics based on multi-mode time series data reconstruction according to claim 1, wherein the step of inputting the oil dissolved gas time series waveform, the tank wall vibration time series waveform and the sleeve heat map time series waveform into a pre-constructed transformer operation situation reconstruction network to perform cross-mode physical field space-time alignment processing, and generating a transformer multi-physical field evolution track taking a fault energy injection point as an initial reference comprises the steps of: performing gas concentration curve inflection point detection processing on the time sequence waveform of the dissolved gas in the oil to obtain a plurality of gas concentration inflection points corresponding to the time sequence waveform of the dissolved gas in the oil and characteristic gas component proportion vectors associated with each gas concentration inflection point; capturing peak points of vibration waveform envelope curves of the tank wall vibration time sequence waveforms to obtain a plurality of vibration amplitude peak values corresponding to the tank wall vibration time sequence waveforms and vibration main frequency offset direction identifiers associated with each vibration amplitude peak value time; Performing thermal image pixel temperature gradient field calculation processing on the sleeve thermal map time sequence waveform to obtain a plurality of temperature field distortion moments corresponding to the sleeve thermal map time sequence waveform and a temperature anomaly region space coordinate set associated with each temperature field distortion moment; inputting the gas concentration inflection point moment, the vibration amplitude peak value moment and the temperature field distortion moment into a time-associated mapping layer of a transformer operation situation reconstruction network to perform cross-mode mutation event time sequence arrangement processing, and identifying the gas concentration inflection point moment appearing first on a time axis as a candidate fault energy injection point; Identifying a vibration amplitude peak time with the shortest time interval with the candidate fault energy injection point on a time axis as a concomitant vibration response point of the candidate fault energy injection point, and simultaneously identifying a temperature field distortion time with the shortest time interval with the candidate fault energy injection point on the time axis as a concomitant thermal field response point of the candidate fault energy injection point; inputting the characteristic gas component proportion vector corresponding to the candidate fault energy injection point into a physical field coupling verification layer of a transformer operation situation reconstruction network, judging whether the characteristic gas component proportion vector points to a discharge fault type or not, and outputting a first verification Boolean value; Inputting a vibration main frequency offset direction identifier corresponding to the accompanying vibration response point into the physical field coupling verification layer, judging whether the vibration main frequency offset direction identifier points to a winding radial deformation mode or not, and outputting a second verification Boolean value; Inputting the temperature abnormal region space coordinate set corresponding to the accompanying thermal field response point into the physical field coupling verification layer, judging whether the temperature abnormal region space coordinate set is positioned in a preset high field intensity region range in the transformer or not, and outputting a third verification Boolean value; According to the logical AND operation results of the first verification Boolean value, the second verification Boolean value and the third verification Boolean value, whether the candidate fault energy injection point is a real fault energy injection point is confirmed, and if the logical AND operation result is true, the candidate fault energy injection point is confirmed to be the real fault energy injection point; And carrying out physical field type classification and recombination processing on all gas concentration inflection points and associated characteristic gas component proportion vectors, all vibration amplitude peak values and associated vibration main frequency offset direction identifiers and all temperature field distortion moments and associated temperature anomaly region space coordinate sets after the fault energy injection points are confirmed by taking the actual fault energy injection moments corresponding to the fault energy injection points as initial references according to a time increment sequence, so as to generate a transformer multi-physical field evolution track taking the fault energy injection points as initial references, the physical field distortion event occurrence time as a time axis and the physical field distortion types as classification axes.
3. 3. The method for transformer fault feature fusion diagnosis based on multimode time series data reconstruction according to claim 1, wherein the performing a topology tracing process of fault feature clues on the transformer multi-physical-field evolution track to obtain a transformer global fault feature topology chain including a fault starting energy release point coordinate, a fault path physical-field distortion point sequence and a fault ending energy dissipation zone boundary comprises: Extracting real fault energy injection time corresponding to a fault energy injection point from the transformer multi-physical-field evolution track as a starting node of topology tracing and tracking, marking the real fault energy injection time as a root node of a transformer global fault characteristic topology chain, and endowing a root node type identifier and a root node space coordinate with the root node type identifier; Searching a first gas concentration inflection point moment adjacent to the real fault energy injection moment in the transformer multi-physical-field evolution track, judging whether a gas distortion intensity value corresponding to the first gas concentration inflection point moment exceeds a preset gas response threshold, and if so, marking the first gas concentration inflection point moment as a first-stage gas child node of a root node in a transformer global fault characteristic topological chain; searching a first vibration amplitude peak value time adjacent to the real fault energy injection time in the multi-physical-field evolution track of the transformer, judging whether a vibration distortion intensity value corresponding to the first vibration amplitude peak value time exceeds a preset vibration response threshold value, and if so, marking the first vibration amplitude peak value time as a primary vibrator node of a root node in a global fault characteristic topological chain of the transformer; Searching a first temperature field distortion moment adjacent to the real fault energy injection moment in the multi-physical-field evolution track of the transformer, judging whether a thermal field distortion intensity value corresponding to the first temperature field distortion moment exceeds a preset thermal field response threshold, and marking the first temperature field distortion moment as a primary thermal field sub-node of a root node in a global fault characteristic topological chain of the transformer if the thermal field distortion intensity value exceeds the preset thermal field response threshold; Taking the corresponding moment of each primary sub-node as a time reference, searching the next gas concentration inflection point moment, the next vibration amplitude peak moment and the next temperature field distortion moment which meet the conditions that the distortion intensity values of the corresponding type of physical fields are adjacent in time and present monotonic increment on a subsequent time axis respectively aiming at each physical field type in the multi-physical-field evolution track of the transformer, and marking the searched moments as secondary sub-nodes of the corresponding primary sub-nodes respectively; Taking the corresponding moment of each secondary sub-node as a time reference, searching the next gas concentration inflection point moment, the next vibration amplitude peak moment and the next temperature field distortion moment which meet the conditions of time adjacency and monotonically increasing the distortion intensity value of the type of the physical field per se on a subsequent time axis in the multi-physical-field evolution track of the transformer for each physical field type, and marking the searched moments as three-level sub-nodes of the corresponding secondary sub-nodes respectively; Continuously searching downwards step by step until the next physical field distortion event moment meeting the condition of adjacent time and increasing the distortion intensity of the corresponding physical field per se cannot be searched, and marking all the physical field distortion event moments searched in the last stage as leaf nodes of a transformer global fault characteristic topological chain; extracting a root node space coordinate corresponding to a root node of a global fault characteristic topological chain of the transformer as a fault starting energy release point coordinate, and extracting physical field distortion event occurrence time offset corresponding to intermediate nodes of all non-root nodes and non-leaf nodes in the global fault characteristic topological chain of the transformer and space coordinates thereof as a fault path physical field distortion point sequence; and extracting space coordinates corresponding to all leaf nodes in the global fault characteristic topological chain of the transformer, performing boundary convex hull calculation processing to obtain a minimum convex polygon surrounding the space coordinates of all leaf nodes as a fault ending energy dissipation area boundary, performing topology association coding processing to the fault starting energy release point coordinates, the fault path physical field distortion point sequence and the fault ending energy dissipation area boundary, and generating the global fault characteristic topological chain of the transformer.
4. 4. The method for transformer fault feature fusion diagnosis based on multimode time series data reconstruction according to claim 1, wherein the step of inputting the transformer global fault feature topological chain into a fault mode semantic analysis network to perform space-time coupling verification processing of fault symptoms and focus areas and generating a transformer comprehensive diagnosis descriptor fusing fault type semantic tags and fault space positioning coordinates comprises the following steps: inputting the global fault characteristic topological chain of the transformer into a topological structure coding layer of a fault mode semantic analysis network, and performing adjacency matrix coding processing on the connection relation between a root node and each level of sub-nodes to obtain a topological connection relation coding vector; inputting the global fault characteristic topological chain of the transformer into a topological structure coding layer of a fault mode semantic analysis network, and performing independent-heat coding treatment on the physical field distortion type corresponding to each node to obtain a physical field type distribution coding vector of each node; inputting the global fault characteristic topological chain of the transformer into a topological structure coding layer of a fault mode semantic analysis network, and carrying out normalized mapping treatment on the distortion intensity grade parameters corresponding to each node to obtain distortion intensity evolution trend coding vectors of each node; Performing splicing fusion processing on the topological connection relation coding vector, the physical field type distribution coding vector and the distortion strength evolution trend coding vector to obtain a topological structure embedded vector representing a tree-shaped topological structure, inputting the topological structure embedded vector into a fault type matching layer of a fault mode semantic analysis network, performing similarity comparison processing on the topological structure embedded vector and each pre-stored fault mode standard topological structure embedded vector to obtain a structural similarity score sequence, and taking a fault mode corresponding to the structural similarity score sequence as a candidate fault mode and outputting a corresponding fault type semantic label; Inputting the fault initial energy release point coordinates in the global fault characteristic topological chain of the transformer into a space positioning reference generation layer of a fault mode semantic analysis network, establishing a three-dimensional space coordinate system inside the transformer by taking the fault initial energy release point coordinates as a space origin, and simultaneously converting each distorted point coordinate in a fault path physical field distorted point sequence into a polar coordinate representation form relative to the space origin to obtain a fault path point polar coordinate set containing azimuth angle parameters, pitch angle parameters and radial distance parameters; Inputting the polar coordinate set of the fault path points into a focus region demarcation layer of a fault mode semantic analysis network to perform spatial clustering processing of polar coordinate points, identifying a spatial sector corresponding to a main release direction of fault energy according to the distribution density degree of azimuth angle parameters and pitch angle parameters of each polar coordinate point, identifying the extension length of a fault influence region in the radial direction according to the distribution range of radial distance parameters of each polar coordinate point, performing combined coding processing on the spatial sector and the radial extension length, and generating fault spatial positioning coordinates taking the coordinates of a fault initial energy release point as an origin point; And carrying out association binding processing on fault type semantic tags corresponding to the screened candidate fault modes and the fault space positioning coordinates to generate preliminary diagnosis descriptors which are corresponding to each candidate fault mode and contain the fault type semantic tags and the fault space positioning coordinates, packaging and combining all the preliminary diagnosis descriptors, and outputting the preliminary diagnosis descriptors as transformer comprehensive diagnosis descriptors fused with the fault type semantic tags and the fault space positioning coordinates.
5. 5. The method for transformer fault feature fusion diagnosis based on multimode time series data reconstruction according to claim 4, wherein the step of inputting the fault starting energy release point coordinates in the transformer global fault feature topological chain into the space positioning reference generation layer of the fault mode semantic analysis network, using the fault starting energy release point coordinates as a space origin to establish a three-dimensional space coordinate system inside the transformer, and simultaneously converting each distortion point coordinate in the fault path physical field distortion point sequence into a polar coordinate representation form relative to the space origin to obtain a fault path point polar coordinate set comprising an azimuth angle parameter, a pitch angle parameter and a radial distance parameter comprises the steps of: Reading fault starting energy release point coordinates from the global fault characteristic topological chain of the transformer, wherein the fault starting energy release point coordinates are vectors containing three dimensional values and respectively represent positions under a preset global three-dimensional space coordinate system of the transformer; Establishing a new three-dimensional space coordinate system inside the transformer by taking the fault initial energy release point coordinate as a space coordinate system origin, wherein three coordinate axis directions of the new coordinate system are kept parallel and consistent with the coordinate axis directions of a global three-dimensional space coordinate system of the transformer; Reading a fault path physical field distortion point sequence from the transformer global fault characteristic topological chain, wherein the fault path physical field distortion point sequence comprises a plurality of physical field distortion points, and each physical field distortion point comprises a distortion point coordinate of the physical field distortion point in a transformer global three-dimensional space coordinate system; For each distorted point coordinate in the distorted point sequence of the physical field of the fault path, calculating a difference vector obtained by subtracting the coordinate vector of the initial energy release point of the fault from the coordinate vector of the distorted point, wherein the difference vector is used as a position vector of the distorted point relative to the origin of the space coordinate system; Decomposing the calculated position vector into a component along the X-axis direction of the new coordinate system, a component along the Y-axis direction of the new coordinate system and a component along the Z-axis direction of the new coordinate system; Calculating azimuth parameters of the distortion point relative to the origin of the space coordinate system according to the projection of the position vector on the X-Y plane, wherein the azimuth parameters are defined as angles from the positive direction of the X axis of the new coordinate system to the position where the projection vector is positioned; Calculating a pitch angle parameter of the distortion point relative to the origin of the space coordinate system according to the included angle between the position vector and the positive direction of the Z axis, wherein the pitch angle parameter is defined as the complementary angle of the included angle between the position vector and the positive direction of the Z axis; Calculating the modular length of the position vector to obtain a radial distance parameter of the distortion point relative to the origin of the space coordinate system; Combining azimuth angle parameters, pitch angle parameters and radial distance parameters corresponding to the distortion point into a three-dimensional parameter set, wherein the three-dimensional parameter set represents the coordinates of the distortion point in a newly built polar coordinate system taking the coordinates of the fault initial energy release point as an origin; repeatedly executing the steps of calculating a difference vector obtained by subtracting the coordinate vector of the fault starting energy release point from the coordinate vector of each distortion point in the physical field distortion point sequence of the fault path to combining the azimuth parameter, the pitch parameter and the radial distance parameter corresponding to the distortion point into a three-dimensional parameter set, and obtaining three-dimensional parameter sets corresponding to all the distortion points; and arranging the three-dimensional parameter sets corresponding to all the distortion points according to the sequence of the distortion points in the sequence to form a fault path point polar coordinate set containing azimuth angle parameters, pitch angle parameters and radial distance parameters.
6. 6. The method for transformer fault feature fusion diagnosis based on multi-mode time series data reconstruction according to claim 4, wherein the step of inputting the fault path point polar coordinate set into a focus region demarcation layer of a fault mode semantic analysis network to perform spatial clustering processing of polar coordinate points, identifying a spatial sector corresponding to a main release direction of fault energy according to distribution density of azimuth angle parameters and pitch angle parameters of each polar coordinate point, identifying an extension length of a fault influence region in a radial direction according to a distribution range of radial distance parameters of each polar coordinate point, performing combined coding processing on the spatial sector and the radial extension length, and generating fault spatial positioning coordinates with a fault start energy release point coordinate as an origin, comprising: Extracting azimuth angle parameters and pitch angle parameters corresponding to each fault path point from the fault path point polar coordinate set, and mapping the azimuth angle parameters and the pitch angle parameters of each fault path point onto a unit spherical surface taking a fault starting energy release point coordinate as a spherical center to obtain a direction point coordinate on the unit spherical surface corresponding to each fault path point; carrying out density-based spatial clustering on all direction point coordinates on a unit sphere, setting a clustering scanning radius parameter and a minimum neighborhood point parameter, and identifying a plurality of dense areas with the distribution density of the direction point coordinates exceeding a preset density threshold value, wherein each dense area corresponds to a group of fault path points with similar spatial directions; Calculating an arithmetic average value of azimuth parameters corresponding to all direction point coordinates in each dense area as a central azimuth parameter of the dense area, calculating an arithmetic average value of pitch angle parameters corresponding to all direction point coordinates in each dense area as a central pitch angle parameter of the dense area, and combining the central azimuth parameter and the central pitch angle parameter of each dense area as a space sector center point corresponding to the dense area; Taking the coordinates of a fault starting energy release point as a starting point, taking the central direction of a space sector corresponding to each dense area as a direction ray, respectively constructing conical space areas taking each direction ray as an axis and taking a preset cone angle as a half cone angle, and marking each conical space area as a space sector corresponding to the dense area; screening out all fault path points of which the space positions respectively belong to the inside of each space sector from the fault path point polar coordinate set, extracting radial distance parameters of the fault path points corresponding to each space sector, and respectively carrying out maximum value search processing on the radial distance parameters corresponding to each space sector to obtain the radial extension length of the fault influence area corresponding to each space sector; And carrying out combined coding processing on the central direction of the space sector of each space sector and the radial extension length of the corresponding fault influence area to generate a plurality of fault space positioning coordinates which are described by taking the initial fault energy release point coordinates as an origin, the central direction of each space sector as an azimuth, and the radial extension length of the corresponding fault influence area as a distance.
7. 7. The method for transformer fault feature fusion diagnosis based on multimode time series data reconstruction according to claim 1, wherein the extracting a corresponding local evolution track segment from the transformer multi-physical-field evolution track according to the fault space positioning coordinates in the transformer comprehensive diagnosis descriptor, generating a maintenance decision trigger instruction including a fault remaining life consumption rate based on the local evolution track segment, comprises: analyzing each primary diagnosis descriptor in the comprehensive diagnosis descriptors of the transformer, extracting fault space positioning coordinates corresponding to each primary diagnosis descriptor and associated fault type semantic tags, and respectively converting the fault space positioning coordinates into corresponding target space region ranges in a three-dimensional space coordinate system inside the transformer; Taking each target space region range as a space screening condition, respectively searching all physical field distortion event moments of which the space positions fall into the corresponding target space region range in the multi-physical-field evolution track of the transformer, and arranging and processing the searched physical field distortion event moments according to time sequence to obtain a local evolution track segment corresponding to each preliminary diagnosis descriptor; Respectively carrying out time sequence analysis processing on the distortion intensity of the physical field on each local evolution track segment, respectively calculating the intensity increment rate of the distortion intensity of each physical field type contained in the local evolution track segment along with the time change, carrying out normalization processing on the intensity increment rate of each physical field type, carrying out fusion calculation to obtain a comprehensive evolution activity index, and marking the local evolution track segment with the comprehensive evolution activity index exceeding a preset threshold as an active evolution track segment; The method comprises the steps of respectively extracting distortion intensity values of various physical field types contained in each active evolution track segment at the moment of a first distortion event and a last distortion event according to each active evolution track segment, respectively carrying out dimensionless treatment on the intensity values of the physical fields, weighting and fusing to generate a comprehensive current fault severity standard value and a comprehensive fault starting severity standard value; Dividing the accumulated increment of the fault severity corresponding to each active evolution track segment by the total time span of the active evolution track segment to obtain the average acceleration of the fault severity corresponding to each active evolution track segment, multiplying the inverse of the average acceleration of each fault severity by the difference between a preset dimensionless comprehensive limit threshold of the fault severity and a comprehensive current fault severity reference value to obtain the predicted value of the residual life consumption rate of the fault corresponding to each active evolution track segment; Determining corresponding maintenance decision emergency degree grades according to the numerical intervals of the fault remaining life consumption rate predicted values of each dimensionless class; And carrying out instruction encoding processing on fault type semantic tags and fault space positioning coordinates, which are associated with the corresponding active evolution track segments, of each maintenance decision emergency level, generating maintenance decision triggering instructions comprising a plurality of dimensionless fault residual life consumption rate predicted values, a plurality of emergency levels, a plurality of fault type semantic tags and a plurality of fault space positioning coordinates, and sending the maintenance decision triggering instructions to a transformer online monitoring master station system.
8. 8. The transformer fault feature fusion diagnosis method based on multi-mode time sequence data reconstruction according to claim 2, wherein the pre-constructed transformer operation situation reconstruction network is obtained by the following network training method: obtaining a transformer historical fault case sample set, wherein the transformer historical fault case sample set comprises a plurality of groups of transformer historical fault case samples with known real fault energy injection time and sample fault type labels, and the transformer historical fault case samples comprise a dissolved gas time sequence waveform sample in oil, a tank wall vibration time sequence waveform sample and a sleeve heat map time sequence waveform sample; Performing gas concentration curve inflection point detection processing on each group of dissolved gas time sequence waveform samples in the oil to obtain a historical gas concentration inflection point moment set and an associated historical characteristic gas component proportion vector set corresponding to each group of dissolved gas time sequence waveform samples in the oil; Executing vibration waveform envelope curve peak point capturing processing on each group of the box wall vibration time sequence waveform samples to obtain a historical vibration amplitude peak value moment set corresponding to each group of the box wall vibration time sequence waveform samples and an associated historical vibration main frequency offset direction identifier set; Performing thermal image pixel temperature gradient field calculation processing on each group of sleeve thermal map time sequence waveform samples to obtain a historical temperature field distortion moment set and an associated historical temperature abnormal region space coordinate set corresponding to each group of sleeve thermal map time sequence waveform samples; Initializing network parameters of a transformer operation situation reconstruction network, wherein the transformer operation situation reconstruction network comprises a time-associated mapping sub-network and a physical field coupling verification sub-network; Inputting the historical gas concentration inflection point moment set, the historical vibration amplitude peak moment set and the historical temperature field distortion moment set into the time-associated mapping sub-network, and outputting predicted candidate fault energy injection moment, predicted accompanying vibration response moment and predicted accompanying thermal field response moment corresponding to the transformer historical fault case sample by the time-associated mapping sub-network; Inputting the historical characteristic gas component proportion vector associated with the predicted candidate fault energy injection time, the historical vibration dominant frequency offset direction identifier associated with the predicted accompanying vibration response time and the historical temperature abnormal region space coordinate set associated with the predicted accompanying thermal field response time into the physical field coupling verification sub-network, wherein the physical field coupling verification sub-network outputs a predicted first verification Boolean value, a predicted second verification Boolean value and a predicted third verification Boolean value; Obtaining a predicted fault energy injection point confirmation signal output by the transformer operation situation reconstruction network according to the logic and operation results of the predicted first verification Boolean value, the predicted second verification Boolean value and the predicted third verification Boolean value; calculating a first time alignment loss between the predicted candidate fault energy injection time and a known real fault energy injection time of the transformer historical fault case sample; calculating a first Boolean value matching loss between the predicted first verification Boolean value and a theoretical first verification Boolean value corresponding to the sample fault type label; calculating a second Boolean value matching loss between the predicted second verification Boolean value and a theoretical second verification Boolean value corresponding to the sample fault type label; calculating a third Boolean value matching loss between the predicted third verification Boolean value and a theoretical third verification Boolean value corresponding to the sample fault type label; Carrying out weighted summation on the first time alignment loss, the first Boolean value matching loss, the second Boolean value matching loss and the third Boolean value matching loss to obtain a total training loss value corresponding to the transformer historical fault case sample; Calculating the gradient of the total training loss value to all network parameters of the transformer operation situation reconstruction network by adopting a gradient back propagation algorithm; updating network parameters of the transformer operation situation reconstruction network according to the calculated gradient; And traversing all samples in the transformer historical fault case sample set, and repeatedly executing the step of executing gas concentration curve inflection point detection processing on each group of dissolved gas time sequence waveform samples in the oil to the step of updating the network parameters of the transformer operation situation reconstruction network according to the calculated gradient until the total training loss value converges on an independent verification set to obtain the pre-constructed transformer operation situation reconstruction network after training is completed.
9. 9. The method for transformer fault signature fusion diagnosis based on multimode time series data reconstruction of claim 3, further comprising a diagnostic conclusion dynamic reliability assessment and knowledge base enhancement step performed after said sending said maintenance decision trigger instruction to a transformer on-line monitoring master station system, said diagnostic conclusion dynamic reliability assessment and knowledge base enhancement step comprising: continuously acquiring a dissolved gas time sequence waveform, a subsequent tank wall vibration time sequence waveform and a subsequent sleeve heat map time sequence waveform of the transformer in a preset observation time window after the maintenance decision triggering instruction is sent; Executing the steps of inputting the dissolved gas time sequence waveform, the tank wall vibration time sequence waveform and the sleeve heat map time sequence waveform in the oil into a pre-constructed transformer running situation reconstruction network to perform cross-mode physical field space-time alignment processing on the dissolved gas time sequence waveform, the tank wall vibration time sequence waveform and the sleeve heat map time sequence waveform in the oil, and generating a subsequent transformer multi-physical field evolution track; Executing the topology tracing tracking processing of fault characteristic clues on the multi-physical-field evolution tracks of the subsequent transformers to generate global fault characteristic topology chains of the subsequent transformers; Performing topology structure evolution comparison analysis on the subsequent transformer global fault characteristic topological chain and an original transformer global fault characteristic topological chain according to which the maintenance decision triggering instruction is generated, and calculating the similarity of the topology structure and the attribute offset of the nodes; judging whether a predicted fault development path in the maintenance decision triggering instruction is consistent with a fault evolution trend of a follow-up actual observation or not according to the topological structure similarity and the node attribute offset, and generating a diagnosis conclusion validity verification label; Extracting an original transformer global fault characteristic topological chain and an original transformer comprehensive diagnosis descriptor corresponding to a current diagnosis case, and storing the original transformer global fault characteristic topological chain, the original transformer comprehensive diagnosis descriptor and the diagnosis conclusion validity verification tag in an associated manner to form a diagnosis case record with a verification tag; If the diagnosis conclusion validity verification tag indicates that the diagnosis conclusion is valid, converting an original transformer global fault characteristic topological chain in the diagnosis case record with the verification tag into a topological structure embedded vector, and adding the topological structure embedded vector as a new positive sample into a fault mode standard topological structure embedded vector library of the fault mode semantic analysis network; if the diagnosis conclusion validity verification tag indicates that the diagnosis conclusion is in doubt, triggering a diagnosis process backtracking analysis, wherein the diagnosis process backtracking analysis comprises the steps of verifying the time alignment precision of physical field distortion events in an original multi-physical field evolution track of the transformer output by the transformer operation situation reconstruction network and rechecking the matching threshold setting of the fault mode semantic analysis network; Generating a diagnosis uncertainty analysis report according to the result of the retrospective analysis of the diagnosis process, wherein the diagnosis uncertainty analysis report is used for guiding and adjusting the preset gas response threshold, the preset vibration response threshold or the preset thermal field response threshold; integrating the diagnosis conclusion validity verification tag, the knowledge base update state and the diagnosis uncertainty analysis report, generating a reliability assessment and knowledge enhancement report of the diagnosis, and transmitting the reliability assessment and knowledge enhancement report to a transformer online monitoring master station system.
10. 10. The utility model provides a transformer fault feature fuses diagnostic system based on multimode time sequence data reconstruction which characterized in that includes: A processor; a machine-readable storage medium storing machine-executable instructions for the processor; Wherein the processor is configured to perform the transformer fault signature fusion diagnostic method based on the reconstruction of multimodal temporal data of any one of claims 1 to 9 via execution of the machine executable instructions.

Description

Transformer fault feature fusion diagnosis method and system based on multi-mode time sequence data reconstruction Technical Field The invention relates to the technical field of power system transformer fault diagnosis, in particular to a transformer fault feature fusion diagnosis method and system based on multi-mode time sequence data reconstruction. Background In an electrical power system, the operational state of a transformer directly affects the stability and reliability of the grid. Traditional transformer fault diagnosis methods mainly rely on a single monitoring means, such as only analyzing the components and contents of dissolved gas in oil by an oil chromatograph monitor, or only detecting the vibration condition of the tank wall by using a vibration acceleration sensor, or only observing the temperature distribution of the sleeve by using an infrared thermal imager. However, the occurrence of transformer faults is often the result of the combined action of multiple factors, and the information acquired by a single monitoring means has limitations. Although the oil chromatographic analysis can reflect the thermal decomposition condition of the insulating material in the transformer, the detection capability of the oil chromatographic analysis on mechanical faults is limited, the vibration analysis is sensitive to the mechanical faults, the type and the position of the faults are difficult to accurately judge, the infrared thermal imaging can find out the local overheating phenomenon, and the electric and mechanical faults in the transformer can not be comprehensively reflected. In addition, most of the existing fault diagnosis methods only simply analyze and process the monitoring data, and lack effective fusion and deep mining of multi-mode time sequence data. The data acquired by different monitoring means have differences in time and space, so that unified analysis and judgment are difficult to perform, the accuracy and timeliness of fault diagnosis are affected, and the high requirements of a modern power system on transformer fault diagnosis cannot be met. Disclosure of Invention In view of the above-mentioned problems, in combination with the first aspect of the present invention, the present invention provides a transformer fault feature fusion diagnosis method based on multi-mode time series data reconstruction, the method comprising: Acquiring a dissolved gas time sequence waveform, a tank wall vibration time sequence waveform and a sleeve heat map time sequence waveform of oil respectively captured by an oil chromatograph monitor, a vibration acceleration sensor and an infrared thermal imager of the transformer in a continuous operation period; Inputting the dissolved gas time sequence waveform, the tank wall vibration time sequence waveform and the sleeve heat map time sequence waveform in the oil into a pre-constructed transformer running situation reconstruction network to perform cross-mode physical field space-time alignment processing, and generating a transformer multi-physical field evolution track taking a fault energy injection point as an initial reference; Performing topology tracing tracking processing of fault characteristic clues on the transformer multi-physical-field evolution track to obtain a transformer global fault characteristic topology chain comprising fault starting energy release point coordinates, fault path physical-field distortion point sequences and fault ending energy dissipation zone boundaries; Inputting the global fault characteristic topological chain of the transformer into a fault mode semantic analysis network to carry out space-time coupling verification processing on fault symptoms and focus areas, and generating a transformer comprehensive diagnosis descriptor fusing fault type semantic tags and fault space positioning coordinates; And extracting a corresponding local evolution track segment from the transformer multi-physical-field evolution track according to the fault space positioning coordinates in the transformer comprehensive diagnosis descriptor, generating a maintenance decision trigger instruction containing the fault residual life consumption rate based on the local evolution track segment, and transmitting the maintenance decision trigger instruction to a transformer online monitoring master station system. In still another aspect, the present invention further provides a transformer fault feature fusion diagnosis system based on multi-mode time series data reconstruction, including: The transformer fault feature fusion diagnosis method based on multi-mode time sequence data reconstruction comprises a processor, a machine-readable storage medium, a machine-executable instruction of the processor, wherein the processor is configured to execute the transformer fault feature fusion diagnosis method based on multi-mode time sequence data reconstruction by executing the machine-executable instruction. In still another aspect, the prese