Transformer fault feature fusion diagnosis method and system based on multi-modal time series data reconstruction

Abstract

The application provides a transformer fault feature fusion diagnosis method and system based on multi-modal time series data reconstruction, and relates to the technical field of power system transformer fault diagnosis. First, the time series waveforms of dissolved gas in oil, the time series waveforms of box wall vibration and the time series waveforms of sleeve heat map are captured respectively, and are input into a pre-constructed transformer operation situation reconstruction network for cross-modal physical field space-time alignment processing, to generate a transformer multi-physical field evolution track with a fault energy injection point as a starting reference. Topological source tracking processing of fault feature clues is performed on the track, to obtain a transformer global fault feature topology chain. The chain is input into a fault mode semantic analysis network, to generate a transformer comprehensive diagnosis descriptor that fuses fault type semantic labels and fault space positioning coordinates. According to the descriptor, a local evolution track segment is extracted, a maintenance decision trigger instruction containing a fault residual life consumption rate is generated, and is sent to a transformer online monitoring master station system, to ensure safe and stable operation of the transformer.

Description

Technical Field

[0001] This invention relates to the field of power system transformer fault diagnosis technology, and more specifically, to a transformer fault feature fusion diagnosis method and system based on multimodal time-series data reconstruction. Background Technology

[0002] In power systems, the operating status of transformers directly affects the stability and reliability of the power grid. Traditional transformer fault diagnosis methods mainly rely on a single monitoring method, such as analyzing the composition and content of dissolved gases in oil using an oil chromatography monitor, detecting tank wall vibration using a vibration acceleration sensor, or observing bushing temperature distribution using an infrared thermal imager.

[0003] However, transformer failures are often the result of multiple factors working together, and the information obtained by a single monitoring method has limitations. While oil chromatography can reflect the thermal decomposition of the insulation materials inside the transformer, its ability to detect mechanical faults is limited; vibration analysis is more sensitive to mechanical faults, but it is difficult to accurately determine the type and location of the fault; infrared thermal imaging can detect localized overheating, but it cannot comprehensively reflect the electrical and mechanical faults inside the transformer.

[0004] Furthermore, most existing fault diagnosis methods only perform simple analysis and processing of monitoring data, lacking effective fusion and in-depth mining of multimodal time-series data. Data obtained by different monitoring methods differ in time and space, making unified analysis and judgment difficult, which affects the accuracy and timeliness of fault diagnosis and fails to meet the high requirements of modern power systems for transformer fault diagnosis. Summary of the Invention

[0005] In view of the aforementioned problems, and in conjunction with the first aspect of the present invention, the present invention provides a transformer fault feature fusion diagnosis method based on multimodal time series data reconstruction, the method comprising:

[0006] The time-series waveforms of dissolved gas in oil, tank wall vibration, and bushing thermal image were captured by an oil chromatograph, a vibration acceleration sensor, and an infrared thermal imager, respectively, during the continuous operation of the transformer.

[0007] The time-series waveforms of dissolved gas in oil, tank wall vibration, and bushing thermal image are input into a pre-constructed transformer operation status reconstruction network for cross-modal physical field spatiotemporal alignment processing to generate a transformer multi-physics field evolution trajectory with the fault energy injection point as the starting reference.

[0008] The multi-physics field evolution trajectory of the transformer is subjected to topological source tracing processing of fault feature clues to obtain a transformer global fault feature topology chain containing the coordinates of the fault initiation energy release point, the sequence of physical field distortion points of the fault path, and the boundary of the fault termination energy dissipation zone.

[0009] The transformer global fault feature topology chain is input into the fault mode semantic parsing network for spatiotemporal coupling verification of fault symptoms and lesion areas, generating a transformer comprehensive diagnostic descriptor that integrates fault type semantic labels and fault spatial location coordinates.

[0010] Based on the fault spatial location coordinates in the transformer comprehensive diagnostic descriptor, the corresponding local evolution trajectory fragments are extracted from the transformer multiphysics field evolution trajectory. Based on the local evolution trajectory fragments, a maintenance decision triggering command containing the fault remaining life consumption rate is generated, and the maintenance decision triggering command is sent to the transformer online monitoring master station system.

[0011] Furthermore, this invention also provides a transformer fault feature fusion diagnostic system based on multimodal time-series data reconstruction, comprising:

[0012] A processor; a machine-readable storage medium for storing machine-executable instructions of the processor; wherein the processor is configured to execute the above-described transformer fault feature fusion diagnosis method based on multimodal time-series data reconstruction by executing the machine-executable instructions.

[0013] In another aspect, the present invention also provides a computer program product, the computer program product including machine-executable instructions stored in a computer-readable storage medium, wherein the processor of the transformer fault feature fusion diagnosis system based on multimodal time-series data reconstruction reads the machine-executable instructions from the computer-readable storage medium, and the processor executes the machine-executable instructions, causing the transformer fault feature fusion diagnosis system based on multimodal time-series data reconstruction to perform the above-described transformer fault feature fusion diagnosis method based on multimodal time-series data reconstruction.

[0014] Based on the above, firstly, the time-series waveforms of dissolved gas in oil, tank wall vibration, and bushing thermal images captured by various monitoring devices during continuous operation of the transformer are obtained, comprehensively collecting information on different physical fields of the transformer. Next, a pre-constructed transformer operation status reconstruction network is used to perform cross-modal physical field spatiotemporal alignment processing on the multi-mode time-series waveforms, generating a multi-physical field evolution trajectory of the transformer with the fault energy injection point as the starting reference. This effectively solves the problem of temporal and spatial differences between different monitoring data. Then, topological tracing processing of fault feature clues is performed on the multi-physical field evolution trajectory to obtain a transformer-wide fault feature topology chain, clarifying the starting, development, and termination positions of the fault. Finally, the global fault feature topology chain is input into a fault mode semantic parsing network for spatiotemporal coupling verification processing of fault symptoms and lesion areas, generating a transformer comprehensive diagnostic descriptor that integrates fault type semantic labels and fault spatial location coordinates, further improving the accuracy and reliability of fault diagnosis. Finally, maintenance decision triggering instructions are generated based on the comprehensive diagnostic descriptor and sent to the transformer online monitoring master station system, realizing the organic combination of fault diagnosis and maintenance decision-making. This enables timely and effective maintenance measures to ensure the safe and stable operation of the transformer and improve the reliability and stability of the power system. Attached Figure Description

[0015] Figure 1 This is a flowchart illustrating the transformer fault feature fusion diagnosis method based on multimodal time-series data reconstruction provided in this embodiment of the invention.

[0016] Figure 2 This is a schematic diagram of exemplary hardware and software components of the transformer fault feature fusion diagnostic system based on multimodal time-series data reconstruction provided in this embodiment of the invention. Detailed Implementation

[0017] Figure 1 This is a flowchart illustrating a transformer fault feature fusion diagnosis method based on multimodal time-series data reconstruction provided by an embodiment of the present invention, which will be described in detail below.

[0018] Step S110: Obtain the time-series waveforms of dissolved gas in oil, tank wall vibration, and bushing thermal image captured by an oil chromatograph, vibration acceleration sensor, and infrared thermal imager respectively during continuous operation of the transformer.

[0019] In this embodiment, a specific oil-immersed power transformer under operational monitoring is taken as an example. The multimodal time-series waveform can be captured by sensors fixedly installed on the transformer body, or it can be synchronously collected and transmitted by multiple sensors integrated on a self-contained intelligent robot performing autonomous inspections within the station. When the self-contained intelligent robot is used for inspection, its onboard oil chromatography monitor, multi-channel vibration acceleration sensor array, and infrared thermal imager are synchronously triggered after the robot arrives at the preset monitoring point, collecting data with a unified clock reference. The oil chromatography monitor extracts samples from the transformer oil at fixed sampling time intervals, such as once every 30 minutes, and performs on-site gas chromatography analysis to generate a continuous time series waveform of dissolved gases in the oil. This waveform data is a multivariate time series, with each sampling point recording the concentration of a set of characteristic gases, including hydrogen concentration, acetylene concentration, ethylene concentration, ethane concentration, methane concentration, carbon monoxide concentration, and carbon dioxide concentration, forming a seven-dimensional gas concentration vector. The time-series waveform of dissolved gas in the oil is represented as a matrix of shape [T_dga, 7], where T_dga represents the total number of sampling time points.

[0020] Multiple vibration acceleration sensors mounted on the transformer tank wall, or contact sensor arrays carried by the robotic arm during robot inspections, are deployed at points on the high-voltage, medium-voltage, and low-voltage sides of the tank wall, or sequentially contacted for measurements. They continuously capture the time-series waveform of the tank wall vibration caused by winding vibration or core magnetostriction at a sampling frequency of 10 kHz per second. Each sensor captures a time-varying vibration amplitude sequence. Through multi-channel synchronous acquisition, the readings of all sensors at the same time reference constitute a multi-channel tank wall vibration time-series waveform. This waveform data is organized into a matrix of shape [T_vib, C_vib], where T_vib is the total number of vibration sampling time points, C_vib is the total number of vibration sensor channels (C_vib is 12), and each element in the matrix represents the vibration acceleration value of the corresponding channel at the corresponding time.

[0021] Multiple fixed infrared thermal imagers deployed around the transformer, or mobile thermal imagers carried by robotic gimbals, are aimed at key parts of the transformer, such as bushings and tank walls, from different angles, continuously capturing the timing waveform of bushing thermal images at a fixed frame rate of one frame every 10 minutes. Each frame of the thermal image is a two-dimensional pixel matrix with a resolution of 240. 320, each pixel records the temperature value at that point. The entire casing thermal image time-series waveform is a three-dimensional data tensor, represented by the shape [T_th, 240, 320], where T_th is the total number of thermal image acquisition frames. These three types of data are synchronized and marked by the data acquisition unit with precise GPS timestamps in fixed installation scenarios; in robot inspection scenarios, they are synchronized and marked by the robot control system by integrating its own positioning and timing information, and finally packaged into a single raw data packet containing multimodal time-series data, which is the output of step S110.

[0022] Step S120: Input the time-series waveform of dissolved gas in oil, the time-series waveform of tank wall vibration, and the time-series waveform of bushing thermal image into the pre-constructed transformer operation status reconstruction network for cross-modal physical field spatiotemporal alignment processing to generate a transformer multi-physics field evolution trajectory with the fault energy injection point as the starting reference.

[0023] The transformer operation status reconstruction network is a deep neural network model. Its core function is to align and fuse time-series waveforms with different physical fields, sampling rates, and data formats in both time and space dimensions to reconstruct the evolution paths of each physical field after a fault occurs. This process includes the following specific sub-steps:

[0024] Step S121: Perform gas concentration curve inflection point detection processing on the time-series waveform of dissolved gas in the oil to obtain multiple gas concentration inflection point times corresponding to the time-series waveform of dissolved gas in the oil and the characteristic gas component ratio vector associated with each gas concentration inflection point time.

[0025] For the time-series waveform of dissolved gases in oil with shape [T_dga, 7], the first-order and second-order difference sequences are calculated independently for each gas component. Taking the hydrogen concentration sequence as an example, the difference in concentration between adjacent time points is calculated to obtain the first-order difference sequence, and then the difference of the first-order difference sequence is calculated to obtain the second-order difference sequence. When the first-order difference value of a time point changes from positive to negative or from negative to positive, and the absolute value of the second-order difference value exceeds a dynamic threshold estimated based on the noise level of historical data, that time point is marked as the inflection point of the concentration curve of that gas component. After performing the above operation on all 7 gas components, all inflection point times are merged and deduplicated to obtain multiple gas concentration inflection point times. For each inflection point time, the concentration values ​​of each gas component at that time are extracted to form a seven-dimensional characteristic gas component proportion vector, which includes the concentration values ​​of hydrogen, acetylene, ethylene, ethane, methane, carbon monoxide, and carbon dioxide.

[0026] Step S122: Perform vibration waveform envelope peak point capture processing on the vibration time sequence waveform of the box wall to obtain multiple vibration amplitude peak times corresponding to the vibration time sequence waveform of the box wall and the vibration main frequency offset direction identifier associated with each vibration amplitude peak time.

[0027] For a box wall vibration time-series waveform of shape [T_vib, 12], a Hilbert transform is first applied to the vibration signal of each channel to extract the modulus of its analytic signal, obtaining the instantaneous amplitude envelope of the vibration signal for that channel. The average of the envelopes of all 12 channels is calculated to obtain a composite envelope. Local maxima are then searched on this composite envelope. When the amplitude of the envelope at a point is greater than the amplitudes of its immediate and adjacent points, and this amplitude exceeds an adaptive threshold, the time corresponding to that point is marked as a peak vibration amplitude moment. For each peak moment, a Fast Fourier Transform is performed on the original multi-channel vibration signal within a short time window of 100 milliseconds before and after that moment to obtain the vibration spectrum. Analyzing this spectrum, the frequency component with the highest energy is identified as the dominant frequency. The dominant frequency is compared with the transformer's normal operating fundamental frequency of 50 Hz and its harmonics. If the dominant frequency shifts below the fundamental frequency, the vibration dominant frequency shift direction identifier is assigned a negative shift identifier; if the dominant frequency shifts above the fundamental frequency, it is assigned a positive shift identifier; if there is no significant shift, it is assigned a no-shift identifier. Each vibration amplitude peak moment is associated with a discrete vibration dominant frequency shift direction identifier.

[0028] Step S123: Perform temperature gradient field calculation on the heat map pixel points of the heat map time-series waveform to obtain multiple temperature field distortion moments corresponding to the heat map time-series waveform and a set of spatial coordinates of the temperature anomaly region associated with each temperature field distortion moment.

[0029] For a heat map time-series waveform of shape [T_th, 240, 320], the temperature gradient field is calculated for each frame of the heat map. For each pixel in the image, its gray-level difference in the horizontal and vertical directions is calculated to obtain the gradient vector of that point. The divergence and curl of the entire gradient field are calculated to identify abnormal changes in temperature distribution. When the root mean square statistic of the gradient field divergence or curl of a heat map frame changes at a rate exceeding a preset abrupt change threshold compared to the previous frame, the corresponding moment of that frame is marked as a temperature field distortion moment. For each heat map frame corresponding to a distortion moment, an image segmentation algorithm based on region growing is applied to mark connected pixel regions with temperatures significantly higher than the average temperature of their surrounding neighborhoods as temperature anomaly regions. The actual spatial coordinates of all pixels in this region in the transformer's three-dimensional spatial model are extracted to form a set of spatial coordinates for temperature anomaly regions.

[0030] Step S124: Input the gas concentration inflection point, vibration amplitude peak time, and temperature field distortion time into the time correlation mapping layer of the transformer operation status reconstruction network for cross-modal mutation event time sequence processing, and identify the gas concentration inflection point that appears first on the time axis as the candidate fault energy injection point.

[0031] The time-related mapping layer first receives a list of gas concentration inflection point times, a list of vibration amplitude peak times, and a list of temperature field distortion times, and sorts all the times in chronological order. From the sorted event sequence, the earliest time of type gas concentration inflection point is selected and recorded as a candidate fault energy injection point T_candidate.

[0032] Step S125: Identify the peak vibration amplitude moment with the shortest time interval to the candidate fault energy injection point on the time axis as the accompanying vibration response point of the candidate fault energy injection point, and at the same time identify the temperature field distortion moment with the shortest time interval to the candidate fault energy injection point on the time axis as the accompanying thermal field response point of the candidate fault energy injection point.

[0033] This layer calculates the absolute value of the time difference between the candidate fault energy injection point T_candidate and the peak times of all vibration amplitudes, finds the vibration time corresponding to the smallest absolute value, and marks it as the associated vibration response point T_vib_asso. Simultaneously, it calculates the absolute value of the time difference between T_candidate and the peak times of all temperature field distortions, finds the temperature time corresponding to the smallest absolute value, and marks it as the associated thermal field response point T_th_asso.

[0034] Step S126: Input the characteristic gas component ratio vector corresponding to the candidate fault energy injection point into the physical field coupling verification layer of the transformer operation status reconstruction network, determine whether the characteristic gas component ratio vector points to the discharge fault type and output the first verification Boolean value.

[0035] The physical field coupling verification layer internally incorporates a classifier based on the support vector machine (SVM) algorithm. This classifier takes a seven-dimensional feature gas component proportion vector V_gas_candidate as input. Internally, the classifier maps the input vector to a high-dimensional space using a kernel function, performs an inner product operation with the trained support vectors, and outputs a probability value P_discharge indicating whether the fault belongs to the discharge fault category. P_discharge is compared with a preset decision threshold Th_gas, which is statistically determined based on the classification boundary between discharge and non-discharge faults in historical fault samples. If P_discharge is greater than Th_gas, the fault is determined to be a discharge fault, and the first verification Boolean value B_gas is true; otherwise, B_gas is false.

[0036] Step S127: Input the vibration dominant frequency offset direction identifier corresponding to the accompanying vibration response point into the physical field coupling verification layer, determine whether the vibration dominant frequency offset direction identifier points to the winding radial deformation mode and output the second verification Boolean value.

[0037] The physical field coupling verification layer maintains a lookup table LUT_vib, which maps vibration dominant frequency offset direction identifiers to preset fault modes. In LUT_vib, the entry corresponding to a negative offset identifier is the winding radial deformation mode. Using the vibration dominant frequency offset direction identifier F_vib_asso corresponding to the accompanying vibration response point as an index, the corresponding mode is searched in LUT_vib. If the search result is a winding radial deformation mode, the output second verification Boolean value B_vib is true; otherwise, B_vib is false.

[0038] Step S128: Input the set of spatial coordinates of the temperature anomaly region corresponding to the accompanying thermal field response point into the physical field coupling verification layer, determine whether the set of spatial coordinates of the temperature anomaly region is located within the preset high field strength region inside the transformer, and output the third verification Boolean value.

[0039] The physical field coupling verification layer pre-stores spatial coordinate boundary data for a pre-defined high-field-strength region within the transformer. This boundary is described by a closed polyhedron, such as a spatial region R_intense enclosed by multiple triangular facets. This layer calculates the center point coordinate P_center of all coordinate points in the spatial coordinate set L_th_asso of the temperature anomaly region by taking the arithmetic mean of the X, Y, and Z coordinates of all points in the set. Then, a ray casting method is used to determine whether P_center is located inside the spatial region R_intense: a ray is emitted from P_center in any direction, and the number of intersections between the ray and the boundary surface of R_intense is calculated. If the number of intersections is odd, the point is inside; if it is even, the point is outside. If P_center is inside R_intense, the output third verification Boolean value B_th is true; otherwise, B_th is false.

[0040] Step S129: Based on the logical AND operation result of the first verification Boolean value, the second verification Boolean value, and the third verification Boolean value, confirm whether the candidate fault energy injection point is a real fault energy injection point. If the logical AND operation result is true, then confirm that the candidate fault energy injection point is a real fault energy injection point.

[0041] A logical AND operation is performed on the first verification Boolean value B_gas, the second verification Boolean value B_vib, and the third verification Boolean value B_th to obtain the result B_and, i.e., B_and = B_gas AND B_vib AND B_th. B_and is true only when B_gas, B_vib, and B_th are all true. In this case, the candidate fault energy injection point T_candidate is confirmed as the true fault energy injection point T_true. If B_and is false, it indicates that the candidate fault energy injection point T_candidate does not meet the physical field coupling verification condition. The system will backtrack to step S124, select the next earliest time from the remaining list of gas concentration inflection point times as the new candidate fault energy injection point, and repeat the verification process from steps S125 to S129 until the true fault energy injection point T_true that satisfies B_and is true is found.

[0042] Step S1210: Taking the actual fault energy injection time corresponding to the confirmed fault energy injection point as the starting reference, the physical field type classification and recombination processing is performed on all gas concentration inflection point times and their associated characteristic gas component ratio vectors, all vibration amplitude peak times and their associated vibration dominant frequency offset direction identifiers, and all temperature field distortion times and their associated temperature anomaly region spatial coordinate sets after the fault energy injection time in ascending time order. This generates a transformer multi-physics evolution trajectory with the fault energy injection point as the starting reference, the physical field distortion event occurrence time as the time axis, and the physical field distortion type as the classification axis.

[0043] Specifically, taking the actual fault energy injection moment T_true as the time zero point, an empty data structure Traj is created to store the evolution trajectory. Traj contains three lists: Gas_list, Vib_list, and Thermal_list. All gas concentration inflection points T_gas_i and their characteristic gas component proportion vectors V_gas_i that are later than T_true are stored in Gas_list in ascending order of time, with each element being a tuple (T_gas_i, V_gas_i). All vibration amplitude peak moments T_vib_j and their vibration dominant frequency offset direction identifier F_vib_j that are later than T_true are stored in Vib_list in ascending order of time, with each element being a tuple (T_vib_j, F_vib_j). All temperature field distortion moments T_th_k and their temperature anomaly region spatial coordinate set L_th_k that are later than T_true are stored in Thermal_list in ascending order of time, with each element being a tuple (T_th_k, L_th_k). The resulting transformer multiphysics evolution trajectory Traj is a composite data structure containing three branch lists: Gas_list, Vib_list, and Thermal_list. These three lists together describe the evolution sequence of each physics distortion event over time, starting from T_true.

[0044] Step S130: Perform topological source tracing processing on the multi-physics evolution trajectory of the transformer to obtain a transformer global fault feature topology chain containing the coordinates of the fault initiation energy release point, the sequence of physical field distortion points along the fault path, and the boundary of the fault termination energy dissipation zone.

[0045] This step aims to further organize the parallel multiphysics evolution trajectory Traj generated in step S1210 into a tree-like topological chain with hierarchical structure and spatial properties, to reveal the complete path of the fault from its inception to its development and termination. This process includes the following specific sub-steps:

[0046] Step S131: Extract the actual fault energy injection time corresponding to the fault energy injection point from the multiphysics field evolution trajectory of the transformer as the starting node of topology tracing, mark the actual fault energy injection time as the root node of the transformer global fault feature topology chain, and assign the root node type identifier and root node spatial coordinates.

[0047] The actual fault energy injection moment T_true is set as the root node. This root node is assigned a type identifier Type_root, for example, set to 0. The spatial coordinates P_root of the root node are not directly given by a single sensor, but are inferred by synthesizing the preliminary responses of various physical fields at time T_true. For example, taking the vibration signal corresponding to the accompanying vibration response point T_vib_asso, using the time difference of arrival localization method, and utilizing the time difference of the first vibration wave received by multiple vibration sensors, a hyperboloid equation system is constructed and solved to deduce the spatial coordinates of the vibration source. These coordinates are used as the spatial coordinates P_root of the root node. P_root is a three-dimensional vector containing the X-axis coordinate component X_root, the Y-axis coordinate component Y_root, and the Z-axis coordinate component Z_root.

[0048] Step S132: Search for the first gas concentration inflection point time adjacent to the actual fault energy injection time in the multiphysics field evolution trajectory of the transformer, and determine whether the gas distortion intensity value corresponding to the first gas concentration inflection point time exceeds the preset gas response threshold. If it exceeds the threshold, mark the first gas concentration inflection point time as the first-level gas sub-node of the root node in the transformer global fault feature topology chain.

[0049] From Gas_list, find the first gas concentration inflection point T_gas_1 that is temporally closest to and later than T_true, and its corresponding characteristic gas component proportion vector V_gas_1. Calculate the gas distortion intensity value S_gas_1 corresponding to this moment. This gas distortion intensity value is the modulus of V_gas_1, i.e., S_gas_1 equals the square root of the sum of the squares of the components in V_gas_1. Compare S_gas_1 with a preset gas response threshold G_th. If S_gas_1 is greater than G_th, create a child node Node_gas_1 of the root node Root in the topology chain. The node type Type_gas is set to 1, the node time is T_gas_1, and the node attributes include the characteristic gas component proportion vector V_gas_1 and the distortion intensity value S_gas_1.

[0050] Step S133: Search for the first vibration amplitude peak moment adjacent to the actual fault energy injection moment in the multiphysics field evolution trajectory of the transformer, and determine whether the vibration distortion intensity value corresponding to the first vibration amplitude peak moment exceeds the preset vibration response threshold. If it exceeds the threshold, mark the first vibration amplitude peak moment as the first-level vibration sub-node of the root node in the transformer global fault feature topology chain.

[0051] From Vib_list, find the first vibration amplitude peak time T_vib_1 that is temporally closest to and later than T_true, and its associated vibration dominant frequency offset direction identifier F_vib_1. Calculate the vibration distortion intensity value S_vib_1 corresponding to this time, which can be the amplitude A_vib_1 of the vibration envelope at that time. Compare A_vib_1 with a preset vibration response threshold V_th. If A_vib_1 is greater than V_th, create a child node Node_vib_1 of the root node Root in the topology chain. The node type Type_vib is set to 2, the node time is T_vib_1, and the node attributes include the vibration dominant frequency offset direction identifier F_vib_1 and the distortion intensity value S_vib_1.

[0052] Step S134: Search for the first temperature field distortion moment adjacent to the actual fault energy injection moment in the multiphysics field evolution trajectory of the transformer, and determine whether the thermal field distortion intensity value corresponding to the first temperature field distortion moment exceeds the preset thermal field response threshold. If it exceeds the threshold, mark the first temperature field distortion moment as the first-level thermal field sub-node of the root node in the transformer global fault feature topology chain.

[0053] From the Thermal_list, find the first temperature field distortion moment T_th_1 that is temporally closest to and later than T_true, and its associated set of spatial coordinates of the temperature anomaly region L_th_1. Calculate the corresponding thermal field distortion intensity value S_th_1, which can be the difference ΔT_1 between the average temperature of the temperature anomaly region and the historical average temperature of that region during normal transformer operation. Compare this difference ΔT_1 with a preset thermal field response threshold T_threshold. If ΔT_1 is greater than T_threshold, create a child node Node_th_1 of the root node Root in the topology chain. The node type Type_th is set to 3, the node moment is T_th_1, and the node attributes include the set of spatial coordinates of the temperature anomaly region L_th_1 and the distortion intensity value S_th_1.

[0054] Step S135: Using the time corresponding to each first-level sub-node as the time reference, in the multi-physics field evolution trajectory of the transformer, for each type of physical field, search for the next gas concentration inflection point, the next vibration amplitude peak time, and the next temperature field distortion time on the subsequent time axis that satisfy the conditions of temporal adjacency and the monotonically increasing distortion intensity value of the physical field of that type. Mark each searched time as the second-level sub-node of the corresponding first-level sub-node.

[0055] For the primary gas sub-node Node_gas_1, using its time T_gas_1 as a reference, the next inflection point time T_gas_2 is searched in Gas_list. The criteria include: temporal contiguity (T_gas_2 is the minimum value in Gas_list greater than T_gas_1); and monotonically increasing gas distortion intensity (V_gas_2's modulus S_gas_2 is greater than S_gas_1). If these conditions are met, a sub-node of Node_gas_1 is created, namely the secondary gas sub-node Node_gas_2. Similarly, for the primary vibration sub-node Node_vib_1, T_vib_2 is searched in Vib_list, requiring its corresponding envelope amplitude A_vib_2 to be greater than A_vib_1, and the secondary vibration sub-node Node_vib_2 is created. For the primary thermal sub-node Node_th_1, T_th_2 is searched in Thermal_list, requiring its corresponding temperature difference ΔT_2 to be greater than ΔT_1, and the secondary thermal sub-node Node_th_2 is created.

[0056] Step S136: Using the time corresponding to each secondary sub-node as the time reference, in the multi-physics field evolution trajectory of the transformer, for each type of physical field, search for the next gas concentration inflection point, the next vibration amplitude peak time, and the next temperature field distortion time on the subsequent time axis that satisfy the condition of temporal adjacency and the distortion intensity value of the physical field of that type showing a monotonically increasing condition. Mark each searched time as the tertiary sub-node of the corresponding secondary sub-node.

[0057] Repeat the same logic as step S135. Using the time T_gas_2 of the secondary gas sub-node Node_gas_2 as a reference, search for T_gas_3 in Gas_list, requiring S_gas_3 to be greater than S_gas_2, and create the tertiary gas sub-node Node_gas_3. Using the time T_vib_2 of the secondary vibration sub-node Node_vib_2 as a reference, search for T_vib_3, requiring A_vib_3 to be greater than A_vib_2, and create the tertiary vibration sub-node Node_vib_3. Using the time T_th_2 of the secondary thermal field sub-node Node_th_2 as a reference, search for T_th_3, requiring ΔT_3 to be greater than ΔT_2, and create the tertiary thermal field sub-node Node_th_3.

[0058] Step S137: Continue searching downwards level by level until the next physical field distortion event time that satisfies the condition of temporal proximity and increasing distortion intensity of the corresponding physical field itself cannot be found. Mark all physical field distortion event times found in the last level search as leaf nodes of the transformer global fault feature topology chain.

[0059] This process is a recursive depth-first search. For each newly created child node of each type, it attempts to find the next event with greater distortion intensity in its respective physics branch list (Gas_list, Vib_list, or Thermal_list), using that node as a reference. When no subsequent events satisfying temporal proximity and increasing distortion intensity can be found for a node on a certain branch, that node becomes a leaf node on that branch. Ultimately, the entire topology chain consists of a root node (Root), several intermediate nodes, and multiple leaf nodes, forming a tree structure.

[0060] Step S138: Extract the spatial coordinates of the root node corresponding to the root node of the transformer global fault feature topology chain as the coordinates of the fault initiation energy release point. At the same time, extract the physical field distortion event occurrence time offset and its spatial coordinates corresponding to all intermediate nodes that are neither root nodes nor leaf nodes in the transformer global fault feature topology chain as the fault path physical field distortion point sequence.

[0061] The spatial coordinates P_root of the root node are directly used as the coordinates of the fault initiation energy release point Coord_start. For each intermediate node in the tree, its time offset relative to the root node is first calculated. For example, for a node at time T_gas_2, its time offset Offset_gas_2 is T_gas_2 minus T_true. Simultaneously, spatial coordinates are assigned to this intermediate node. For gas nodes, their spatial coordinates P_gas_i can be obtained by inversely calculating the concentration distribution at multiple times using a gas diffusion model, for example, by solving the source term position in the diffusion equation using the least squares method. For vibration nodes, their spatial coordinates P_vib_j can be obtained by locating the vibration signal arrival time difference. For thermal field nodes, their spatial coordinates P_th_k are taken as the center point of the spatial coordinate set L_th_k of the temperature anomaly region. All intermediate nodes are sorted according to their time offsets, forming a sequence Seq_path. Each element in the sequence is a tuple containing a three-dimensional vector of the time offset and the spatial coordinates P_i.

[0062] Step S139: Extract the spatial coordinates corresponding to all leaf nodes in the transformer global fault feature topology chain and perform boundary convex hull calculation to obtain the smallest convex polygon that surrounds the spatial coordinates of all leaf nodes as the boundary of the fault termination energy dissipation zone. Perform topological association encoding on the coordinates of the fault initiation energy release point, the sequence of physical field distortion points of the fault path, and the boundary of the fault termination energy dissipation zone to generate the transformer global fault feature topology chain.

[0063] Collect the spatial coordinates of all leaf nodes in the tree to form a three-dimensional point set, Points_leaf. Apply the Andrew algorithm to calculate the convex hull of this point set: first, sort the points by their X-coordinates, and if X is the same, sort by their Y-coordinates; then construct the upper and lower convex hulls separately, and finally merge them. In three-dimensional space, the convex hull (ConvexHull) is a convex polyhedron enclosed by multiple triangular facets, which represents the boundary of the fault termination energy dissipation region. Finally, associate and encode the coordinates of the fault initiation energy release point (Coord_start), the fault path physical field distortion point sequence (Seq_path), and the fault termination energy dissipation region boundary (ConvexHull), for example, by combining them into a structure called TopoChain, which is the complete transformer global fault feature topology chain.

[0064] Step S140: Input the transformer global fault feature topology chain into the fault mode semantic parsing network to perform spatiotemporal coupling verification processing of fault symptoms and lesion areas, and generate a transformer comprehensive diagnostic descriptor that integrates fault type semantic labels and fault spatial location coordinates.

[0065] This fault mode semantic parsing network is a deep model combining graph neural networks and spatial transformation networks, designed to parse the fault type and precise spatial location from the topology chain TopoChain. The process includes the following specific sub-steps:

[0066] Step S141: Input the transformer global fault feature topology chain into the topology encoding layer of the fault mode semantic parsing network, and perform adjacency matrix encoding on the connection relationship between the root node and each level of child nodes to obtain the topology connection relationship encoding vector.

[0067] The topology encoding layer first converts the tree-like topology chain, TopoChain, into graph-structured data. An adjacency matrix A is created, with a size of N multiplied by N, where N is the total number of nodes in the topology chain. Elements A_ij in the matrix are assigned a value of 1 if a parent-child connection exists between node i and node j, and a value of 0 otherwise. The adjacency matrix A is then flattened into a topology connection encoding vector V_adj, with a dimension of N squared.

[0068] Step S142: Input the transformer global fault feature topology chain into the topology coding layer of the fault mode semantic parsing network, perform one-hot coding on the physical field distortion type corresponding to each node, and obtain the physical field type distribution coding vector of each node.

[0069] For each node, a one-hot encoded vector of length 4 is generated to represent its type. For example, [1, 0, 0, 0] represents the root node, [0, 1, 0, 0] represents the gas node, [0, 0, 1, 0] represents the vibration node, and [0, 0, 0, 1] represents the thermal field node. The one-hot encoded vectors of all nodes are concatenated into a long vector according to the fixed order of the nodes in the topological chain, such as the depth-first traversal order. This vector is the physical field type distribution encoding vector V_type of each node, and its dimension is N times 4.

[0070] Step S143: Input the transformer global fault feature topology chain into the topology coding layer of the fault mode semantic parsing network, and perform normalization mapping on the distortion intensity level parameter corresponding to each node to obtain the distortion intensity evolution trend coding vector of each node.

[0071] For each non-root node, its distortion intensity value, such as the modulus S_gas for gas nodes, the amplitude A_vib for vibration nodes, and the temperature difference ΔT for thermal nodes, is normalized. The normalization method is to divide the distortion intensity value S_i of that node by the maximum distortion intensity value S_type_max of all nodes in the same type of physical field, resulting in a normalized intensity value S_i_norm between 0 and 1. For the root node, its normalized intensity value is set to 0. These normalized intensity values ​​are then arranged in the same node order as in step S142 to form a distortion intensity evolution trend encoding vector V_strength, with dimension N.

[0072] Step S144: The topology connection relationship encoding vector, the physical field type distribution encoding vector, and the distortion intensity evolution trend encoding vector are spliced ​​and fused to obtain a topology embedding vector representing the tree topology. The topology embedding vector is input into the fault type matching layer of the fault mode semantic parsing network and compared with the pre-stored standard topology embedding vectors of each fault mode to obtain a structural similarity score sequence. The fault mode corresponding to the structural similarity score sequence is used as a candidate fault mode and the corresponding fault type semantic label is output.

[0073] V_adj, V_type, and V_strength are concatenated along the feature dimension to form a comprehensive topological embedding vector V_embedding, whose dimension is N squared plus 4N plus N. The fault type matching layer internally maintains a fault mode standard library Lib, which stores standard topological embedding vectors corresponding to standard fault modes extracted from a large number of historical fault cases and confirmed by experts. Examples include the standard embedding vector V_short_discharge for inter-turn short-circuit faults accompanied by partial discharge, V_buckle_overheat for winding axial instability accompanied by overheating, and V_shield_rattle for loose magnetic shielding accompanied by vibration and abnormal noise. The cosine similarity Sim_i between V_embedding and each standard embedding vector V_std_i in the library Lib is calculated as Sim_i = (V_embedding·V_std_i) / (|V_embedding|×|V_std_i|). A similarity score sequence S_sim is obtained, such as [Sim_short_discharge, Sim_buckle_overheat, Sim_shield_rattle]. The standard pattern with the highest score is selected as the candidate fault pattern, and the corresponding fault type semantic label Label_fault is output, such as "Inter-turn short circuit in the middle of the high-voltage winding accompanied by partial discharge and inducing a surge of dissolved acetylene gas in the oil".

[0074] Step S145: Input the coordinates of the fault initiation energy release point in the transformer global fault feature topology chain into the spatial positioning reference generation layer of the fault mode semantic parsing network, establish a three-dimensional spatial coordinate system inside the transformer with the coordinates of the fault initiation energy release point as the spatial origin, and convert the coordinates of each distortion point in the fault path physical field distortion point sequence into a polar coordinate representation relative to the spatial origin to obtain a set of polar coordinates of fault path points including azimuth angle parameters, pitch angle parameters, and radial distance parameters.

[0075] This step comprises a series of sub-steps to achieve precise spatial coordinate transformation.

[0076] Step S1451: Read the coordinates of the fault initiation energy release point from the transformer global fault feature topology chain. The coordinates of the fault initiation energy release point are a vector containing three-dimensional values, which represent the positions in the preset transformer global three-dimensional spatial coordinate system.

[0077] Read Coord_start from the topology chain TopoChain and denote it as P0. P0 is equal to X0, Y0, and Z0.

[0078] Step S1452: Using the coordinates of the fault initiation energy release point as the origin of the spatial coordinate system, establish a new three-dimensional spatial coordinate system inside the transformer. The three coordinate axes of the new coordinate system are parallel to the coordinate axes of the transformer's global three-dimensional spatial coordinate system.

[0079] This means that the new coordinate system O' simply shifts the origin to P0, and its X', Y', and Z' axes are parallel to and in the same direction as the X, Y, and Z axes of the original global coordinate system, respectively.

[0080] Step S1453: Read the sequence of physical field distortion points of the fault path from the topology chain of the global fault features of the transformer. The sequence of physical field distortion points of the fault path contains multiple physical field distortion points, and each physical field distortion point contains its distortion point coordinates in the global three-dimensional spatial coordinate system of the transformer.

[0081] Read Seq_path from the topology chain TopoChain. For the i-th point in the sequence, its distorted point coordinates are labeled P_i, where P_i equals X_i, Y_i, and Z_i.

[0082] Step S1454: For the coordinates of each distorted point in the sequence of distorted points in the physical field of the fault path, calculate the difference vector obtained by subtracting the coordinate vector of the fault initiation energy release point from the distorted point coordinate vector. This difference vector is used as the position vector of the distorted point relative to the origin of the spatial coordinate system.

[0083] For each P_i, calculate the difference vector D_i, which is equal to P_i minus P0, i.e., D_i equals (X_i-X0), (Y_i-Y0), (Z_i-Z0).

[0084] Step S1455: Decompose the calculated position vector into components along the X-axis of the new coordinate system, components along the Y-axis of the new coordinate system, and components along the Z-axis of the new coordinate system.

[0085] These three components are D_i itself, that is, D_ix equals (X_i-X0), D_iy equals (Y_i-Y0), and D_iz equals (Z_i-Z0).

[0086] Step S1456: Calculate the azimuth parameter of the distorted point relative to the origin of the spatial coordinate system based on the projection of the position vector on the XY plane. The azimuth parameter is defined as the angle traversed by rotating counterclockwise from the positive X-axis of the new coordinate system to the position of the projection vector.

[0087] Calculate the length of the projection vector L_xy = √(D_ix² + D_iy²). Calculate the azimuth Phi_i = arctan2(D_iy, D_ix). The return value of the function arctan2 ranges from -π to π. For ease of subsequent processing, it is converted to a range of 0 to 2π: if the return value of arctan2 is less than 0, then Phi_i is equal to the return value plus 2π; otherwise, Phi_i is equal to the return value.

[0088] Step S1457: Calculate the pitch angle parameter of the distorted point relative to the origin of the spatial coordinate system based on the angle between the position vector itself and the positive direction of the Z-axis. The pitch angle parameter is defined as the complementary angle between the position vector and the positive direction of the Z-axis.

[0089] Calculate the magnitude of the position vector R_i = √(D_ix² + D_iy² + D_iz²). Calculate the pitch angle Theta_i = arccos(D_iz / R_i). The value of Theta_i ranges from 0 to π.

[0090] Step S1458: Calculate the magnitude of the position vector to obtain the radial distance parameter of the distorted point relative to the origin of the spatial coordinate system.

[0091] The radial distance parameter is R_i calculated in step S1457.

[0092] Step S1459: Combine the azimuth, pitch, and radial distance parameters corresponding to the distortion point into a three-dimensional parameter set, which represents the coordinates of the distortion point in the newly established polar coordinate system with the coordinates of the fault initiation energy release point as the origin.

[0093] We obtain the polar coordinate representation of point P_i (Phi_i, Theta_i, R_i).

[0094] Step S14510: Repeat the step of calculating the difference vector obtained by subtracting the coordinate vector of the fault initiation energy release point from the coordinate vector of the distorted point in the sequence of distorted points in the physical field of the fault path until the step of combining the azimuth parameter, pitch parameter and radial distance parameter corresponding to the distorted point into a three-dimensional parameter set, to obtain the three-dimensional parameter set corresponding to all distorted points.

[0095] For all points P_i in Seq_path, execute steps S1454 to S1459 sequentially to obtain a series of polar coordinate points.

[0096] Step S14511: Arrange the three-dimensional parameter sets corresponding to all distortion points according to the order of the distortion points in the sequence to form a set of polar coordinates of the fault path points containing azimuth, pitch and radial distance parameters.

[0097] Arrange the obtained polar coordinate points in their original order to form the polar coordinate set PolarSet of the fault path points. PolarSet is equal to the sequence [(Phi_1, Theta_1, R_1), (Phi_2, Theta_2, R_2), ...].

[0098] Step S146: Input the set of polar coordinates of the fault path points into the lesion area delineation layer of the fault mode semantic parsing network for spatial clustering of polar coordinate points. Based on the distribution density of the azimuth and pitch angle parameters of each polar coordinate point, identify the spatial sector corresponding to the main release direction of fault energy. At the same time, based on the distribution range of the radial distance parameter of each polar coordinate point, identify the radial extension length of the fault-affected area. Combine the spatial sector and the radial extension length for encoding to generate fault spatial positioning coordinates with the coordinates of the fault initiation energy release point as the origin.

[0099] This step comprises a series of sub-steps to achieve accurate characterization of the fault space.

[0100] Step S1461: Extract the azimuth and pitch parameters corresponding to each fault path point from the set of polar coordinates of the fault path points, and map the azimuth and pitch parameters of each fault path point onto a unit sphere centered on the coordinates of the fault initiation energy release point, to obtain the coordinates of the direction point on the unit sphere corresponding to each fault path point.

[0101] Extract the azimuth angle Phi_i and elevation angle Theta_i of each point i from the PolarSet. Calculate its Cartesian coordinates U_i on the unit sphere, where U_ix = sin(Theta_i) × cos(Phi_i), U_iy = sin(Theta_i) × sin(Phi_i), and U_iz = cos(Theta_i). This gives the direction point coordinates U_i.

[0102] Step S1462: Perform density-based spatial clustering on the coordinates of all directional points on the unit sphere, set the clustering scan radius parameter and the minimum number of neighboring points parameter, identify multiple dense regions where the distribution density of directional point coordinates exceeds the preset density threshold, and each dense region corresponds to a set of fault path points with similar spatial orientations.

[0103] The DBSCAN algorithm is used to cluster the point set {U_i}. A scan radius parameter R_eps is set, for example, to a value of 0.1, and a minimum number of neighboring points parameter MinPts is set, for example, to a value of 3. Starting from an unvisited point U_i, the algorithm finds all neighboring points whose distance is within R_eps. Euclidean distance is used for distance calculation. If the number of neighboring points is greater than or equal to MinPts, then U_i and its neighbors form a cluster core, and this cluster is recursively expanded. Finally, several clusters Cluster_k are identified, and the points within each cluster are spatially similar fault path points.

[0104] Step S1463: Calculate the arithmetic mean of the azimuth parameters corresponding to the coordinates of all directions within each dense region as the center azimuth parameter of that dense region, calculate the arithmetic mean of the elevation parameters corresponding to the coordinates of all directions within each dense region as the center elevation parameter of that dense region, and combine the center azimuth parameter and center elevation parameter of each dense region as the center pointing direction of the space sector corresponding to that dense region.

[0105] For the k-th cluster, Cluster_k, it contains a set of point indices. Calculate the average of the raw azimuth parameters corresponding to these points, Phi_ck, where Phi_ck equals the sum of all Phi_i within the cluster divided by the number of points in the cluster. Calculate the average of the raw pitch parameters corresponding to these points, Theta_ck, where Theta_ck equals the sum of all Theta_i within the cluster divided by the number of points in the cluster. Then the spatial sector center points to (Phi_ck, Theta_ck).

[0106] Step S1464: Starting from the coordinates of the fault initiation energy release point, and taking the direction of the center of the spatial sector corresponding to each dense region as the directional ray, construct a cone-shaped spatial region with each directional ray as the axis and a preset cone angle as the half cone angle, and mark each cone-shaped spatial region as the spatial sector corresponding to the dense region.

[0107] With P0 as the vertex and (Phi_ck, Theta_ck) as the axis, construct a cone with a semi-cone angle of Alpha. Alpha is a configurable hyperparameter, for example, a value of 15 degrees (i.e., π / 12 radians). The spatial region inside this cone is the spatial sector Sector_k corresponding to the k-th dense region.

[0108] Step S1465: Select all fault path points whose spatial locations belong to each spatial sector from the set of polar coordinates of the fault path points, extract the radial distance parameter of the fault path point corresponding to each spatial sector, perform maximum value search processing on the radial distance parameter corresponding to each spatial sector, and obtain the radial extension length of the fault influence area corresponding to each spatial sector.

[0109] For a spatial sector Sector_k, iterate through all points in PolarSet and determine whether each point is located inside the cone Sector_k. The determination method is as follows: calculate the angle Delta between the direction vector corresponding to point (Phi_i, Theta_i, R_i) and the axis direction (Phi_ck, Theta_ck). If Delta is less than Alpha, the point is considered to be inside the sector. Collect the radial distances R_i of all points within sector Sector_k and take the maximum value R_max_k. R_max_k is the radial extension length of the fault-affected area corresponding to this sector.

[0110] Step S1466: Combine the spatial sector and the radial extension length for encoding to generate fault spatial positioning coordinates with the fault initiation energy release point coordinates as the origin.

[0111] The center direction (Phi_ck, Theta_ck) of each spatial sector Sector_k is combined with its radial extension length R_max_k to form a location coordinate Loc_k, which can be represented as a triple (Phi_ck, Theta_ck, R_max_k). The location coordinates of all sectors form a list Locs, which is the fault space location coordinate.

[0112] Step S147: Associate and bind the fault type semantic tags corresponding to the selected candidate fault modes with the fault spatial positioning coordinates to generate a preliminary diagnostic descriptor for each candidate fault mode that includes fault type semantic tags and fault spatial positioning coordinates. After packaging and combining all the preliminary diagnostic descriptors, output the transformer comprehensive diagnostic descriptor that integrates fault type semantic tags and fault spatial positioning coordinates.

[0113] For each candidate fault mode selected in step S144, its fault type semantic label Label_fault is associated with the fault spatial location coordinate list Locs generated in step S1466. A preliminary diagnostic descriptor Diag_prelim is generated, for example, a structure containing the fields fault_type and location_coords. Since there may be multiple spatial sectors, location_coords itself is a list, i.e., Locs. The preliminary diagnostic descriptors corresponding to all candidate fault modes are combined into a list Diag_list, which is the final transformer comprehensive diagnostic descriptor Final_Diag.

[0114] Step S150: Extract the corresponding local evolution trajectory segment from the transformer multiphysics field evolution trajectory according to the fault spatial location coordinates in the transformer comprehensive diagnostic descriptor, generate a maintenance decision triggering command containing the fault remaining life consumption rate based on the local evolution trajectory segment, and send the maintenance decision triggering command to the transformer online monitoring master station system.

[0115] This step aims to translate macroscopic diagnostic results into specific, actionable maintenance decisions. This process includes the following specific sub-steps:

[0116] Step S151: Parse each preliminary diagnostic descriptor in the transformer comprehensive diagnostic descriptor, extract the fault spatial location coordinates and associated fault type semantic tags corresponding to each preliminary diagnostic descriptor, and convert the fault spatial location coordinates into the corresponding target spatial region range in the transformer's internal three-dimensional spatial coordinate system.

[0117] Iterate through each preliminary diagnostic descriptor Diag_prelim in Final_Diag. Take, for example, a location coordinate Loc_k(Phi_ck, Theta_ck, R_max_k) from its fault spatial location coordinate list Locs. First, convert its polar coordinate representation back to the direction vector Dir_k relative to P0. The three components of Dir_k are sin(Theta_ck) multiplied by cos(Phi_ck), sin(Theta_ck) multiplied by sin(Phi_ck), and cos(Theta_ck). Then, calculate the endpoint coordinate P_end_k, which is the distance reached by extending R_max_k along the Dir_k direction from P0. P_end_k equals P0 plus Dir_k multiplied by R_max_k. To obtain a spatial region, construct a cylindrical spatial region Region_k with a preset radius R_neighbor, centered on the line segment from P0 to P_end_k. The value of R_neighbor can be set based on the typical dimensions of the transformer's internal structure, for example, 0.1 meters. Region_k is the target spatial region range corresponding to this positioning coordinate.

[0118] Step S152: Using the range of each target spatial region as the spatial filtering condition, search for all physical field distortion event moments in the multi-physics field evolution trajectory of the transformer where the spatial position falls within the corresponding target spatial region range. Arrange the searched physical field distortion event moments in chronological order to obtain the local evolution trajectory segment corresponding to each preliminary diagnostic descriptor.

[0119] For a region_k, traverse each physical field distortion event in the transformer's multiphysics evolution trajectory Traj. For gas events, take their spatial coordinates P_gas_i; for vibration events, take their spatial coordinates P_vib_j; for thermal events, take their spatial coordinates P_th_k. Determine whether the spatial coordinates are located inside the cylindrical region_k. The determination method is: calculate the distance d from point P to line segment P0P_end_k, and the parameter t of the projection point of point P on the line segment relative to P0. If t is between 0 and 1 and d is less than R_neighbor, then the point is considered to fall within the region. Sort all events falling within Region_k according to their occurrence time to form a subset Traj_local_k. This subset is the local evolution trajectory segment corresponding to the fault's spatial location coordinates.

[0120] Step S153: Perform time series analysis on the physical field distortion intensity of each local evolution trajectory segment. For each type of physical field contained in the local evolution trajectory segment, calculate the intensity growth rate of the distortion intensity over time. Normalize the intensity growth rate of each type of physical field and merge them to obtain a comprehensive evolution activity index. Mark the local evolution trajectory segments whose comprehensive evolution activity index exceeds a preset threshold as active evolution trajectory segments.

[0121] Taking the local segment Traj_local_k as an example, it contains a series of events ordered by time. Assume it includes gas events and vibration events. For gas events, with their occurrence time t as the independent variable and the gas distortion intensity value S_gas as the dependent variable, a linear regression line S_gas=a_gas is fitted using the least squares method. The slope a_gas of t+b_gas is the rate of increase of gas distortion intensity Rate_gas. Similarly, for vibration events, the slope Rate_vib is fitted. Rate_gas and Rate_vib are normalized, for example, by dividing them by the maximum rate of change Rate_gas_max and Rate_vib_max obtained statistically from historical data for their respective physical field types, to obtain normalized rates Rate_gas_norm and Rate_vib_norm. The comprehensive evolutionary activity index A_act is calculated as a weighted sum of Rate_gas_norm and Rate_vib_norm. The weighting coefficients w_gas and w_vib can be preset according to the sensitivity of the physical field to fault development, and their sum is 1. If A_act is greater than the preset activity threshold A_th, for example, 0.7, then the local segment Traj_local_k is marked as an active evolutionary trajectory segment.

[0122] Step S154: For each active evolution trajectory segment, extract the distortion intensity values ​​corresponding to the various physical field types contained therein at the first and last distortion events. Perform dimensionless processing on the intensity values ​​of the different physical fields, and then weightedly fuse them to generate a comprehensive current fault severity baseline value and a comprehensive fault initial severity baseline value. Calculate the difference between the comprehensive current fault severity baseline value and the comprehensive fault initial severity baseline value for each active evolution trajectory segment as the cumulative increment of fault severity.

[0123] For the active segment Traj_local_k, its first event time is denoted as T_start, and its last event time as T_end. The gas distortion intensity S_gas_start and vibration distortion intensity S_vib_start are extracted at time T_start. The gas distortion intensity S_gas_end and vibration distortion intensity S_vib_end are extracted at time T_end. Dimensionless processing can be performed using the same normalization method as in step S153, i.e., dividing by the maximum possible distortion intensity S_gas_limit and S_vib_limit for their respective physical field types, mapping the intensity values ​​to between 0 and 1, obtaining S_gas_start_norm, S_vib_start_norm, S_gas_end_norm, and S_vib_end_norm. Then, the comprehensive initial severity Severity_start is calculated, which equals w_gas multiplied by S_gas_start_norm plus w_vib multiplied by S_vib_start_norm. Similarly, the overall current severity level Severity_end is calculated, which equals w_gas multiplied by S_gas_end_norm plus w_vib multiplied by S_vib_end_norm. The cumulative severity increment Delta_Severity equals Severity_end minus Severity_start.

[0124] Step S155: Divide the cumulative increment of fault severity corresponding to each active evolution trajectory segment by the total time span of the active evolution trajectory segment to obtain the average growth rate of fault severity corresponding to each active evolution trajectory segment. Multiply the reciprocal of the average growth rate of fault severity by the difference between the preset, dimensionless comprehensive fault severity threshold and the comprehensive current fault severity benchmark value to obtain the predicted value of the remaining lifetime consumption rate of fault corresponding to each active evolution trajectory segment.

[0125] Calculate the total time span Delta_T of the segment Traj_local_k, where Delta_T equals T_end minus T_start. The average rate of increase in fault severity, V_avg, equals Delta_Severity divided by Delta_T. Let the preset, dimensionless comprehensive threshold for fault severity be Severity_limit, which represents the critical severity at which a device failure leads to failure; for example, a value of 1.5. Then the predicted rate of remaining lifetime consumption, RUL_rate, equals (Severity_limit minus Severity_end) divided by V_avg. RUL_rate is a time-based value, such as days, indicating how much time is expected to be needed to reach the critical state at the current rate of deterioration.

[0126] Step S156: Determine the corresponding maintenance decision urgency level based on the numerical range to which each dimensionless fault remaining life consumption rate prediction value belongs.

[0127] A pre-defined urgency level mapping rule is established. For example, two key thresholds are set: an urgency threshold Th_urgent, for example, with a value of 7 days; and a warning threshold Th_warning, for example, with a value of 30 days. For each active evolution trajectory segment, the predicted rate of remaining fault life consumption RUL_rate_k is compared with the above thresholds. If RUL_rate_k is less than Th_urgent, the maintenance decision urgency level Level_k corresponding to that segment is assigned the value "urgency," with a corresponding code of 3. If RUL_rate_k is greater than or equal to Th_urgent and less than Th_warning, Level_k is assigned the value "warning," with a corresponding code of 2. If RUL_rate_k is greater than or equal to Th_warning, Level_k is assigned the value "routine monitoring," with a corresponding code of 1. In this way, each active evolution trajectory segment is associated with a discrete maintenance decision urgency level code.

[0128] Step S157: Perform instruction encoding processing on the fault type semantic label and fault spatial location coordinates associated with each maintenance decision urgency level and the corresponding active evolution trajectory segment to generate a maintenance decision trigger instruction containing multiple dimensionless fault remaining life consumption rate prediction values, multiple urgency levels, multiple fault type semantic labels and multiple fault spatial location coordinates, and send the maintenance decision trigger instruction to the transformer online monitoring master station system.

[0129] For each evolutionary trajectory segment marked as active, all corresponding diagnostic and predictive information is integrated and encoded. Taking an active segment as an example, its associated information includes: the fault type semantic label Label_fault obtained from step S144, such as "inter-turn short circuit in the middle of the high-voltage winding accompanied by partial discharge and inducing a surge of dissolved acetylene gas in the oil"; the fault spatial location coordinate list Locs generated from step S1466, which contains one or more triples (Phi_ck, Theta_ck, R_max_k); the fault remaining lifetime consumption rate prediction value RUL_rate_k calculated from step S155; and the maintenance decision urgency level Level_k determined from step S156. The above information is combined into a structured maintenance suggestion unit Unit_k, for example, in key-value pair format: {"fault_type": "Inter-turn short circuit in the middle of the high-voltage winding accompanied by partial discharge and inducing a surge of dissolved acetylene gas in the oil", "location": [{"azimuth": Phi_c1, "elevation": Theta_c1, "radius": R_max1}], "rul_rate": RUL_rate_k, "urgency_level": 3}.

[0130] All active segments' corresponding maintenance suggestion units (Unit_k) are aggregated into a list (Units). Then, a maintenance decision trigger command (Command) is generated. This command is a data packet containing a metadata header and the aforementioned Units list. The metadata header includes the command generation timestamp (Time_stamp), the command identifier (Command_id), and the targeted transformer equipment identifier (Transformer_id). Finally, the Command is sent to the transformer online monitoring master station system via a dedicated communication network within the power system. Upon receiving the command, the master station system parses it and displays it on the monitoring interface.

[0131] Step S210: Obtain a set of historical transformer fault case samples. The set of historical transformer fault case samples includes multiple sets of known real fault energy injection times and sample fault type labels. The historical transformer fault case samples include time-series waveform samples of dissolved gas in oil, time-series waveform samples of tank wall vibration, and time-series waveform samples of bushing thermal images.

[0132] To train the pre-built transformer operation status reconstruction network described in step S120, a dedicated training dataset is required. A large number of transformer fault cases are collected from the power company's equipment fault history database. Each case sample contains a complete set of monitoring data recorded from a period before the fault occurred to a period after the fault occurred. Specifically, each sample includes a time-series waveform sample of dissolved gas in oil, whose data format is consistent with that described in step S110, and is a matrix of [T_dga_sample, 7]; a time-series waveform sample of tank wall vibration, which is a matrix of [T_vib_sample, 12]; and a time-series waveform sample of bushing thermal image, which is a three-dimensional tensor of [T_th_sample, 240, 320]. In addition, each sample is accompanied by a precise fault energy injection time T_true_label confirmed by post-fault analysis, and a sample fault type label Label_sample annotated by experts, such as "inter-turn short circuit in high-voltage winding accompanied by partial discharge and development into arc", "local overheating caused by multiple grounding of iron core and accompanied by abnormal vibration", "tap switch switching failure leading to contact burnout and generation of characteristic gas", etc. The above samples constitute the transformer historical fault case sample set Dataset.

[0133] Step S220: Perform gas concentration curve inflection point detection processing on each group of dissolved gas time-series waveform samples in the oil to obtain the set of historical gas concentration inflection point times and the associated set of historical characteristic gas component proportion vectors corresponding to each group of dissolved gas time-series waveform samples in the oil.

[0134] For each sample in the Dataset, the same processing as step S121 is performed on its dissolved gas time-series waveform sample in oil. Specifically, for a matrix of shape [T_dga_sample, 7], first and second-order differences are calculated for each gas component, and inflection points are detected using zero-crossing points of the differences and dynamic thresholds. All detected inflection point moments are recorded as a historical gas concentration inflection point moment set, GasTimes_sample. Simultaneously, for each inflection point moment, a seven-dimensional gas concentration vector is extracted to form a historical characteristic gas component proportion vector set, GasVectors_sample.

[0135] Step S230: Perform vibration waveform envelope peak point capture processing on each group of the box wall vibration time sequence waveform samples to obtain the set of historical vibration amplitude peak times and the associated set of historical vibration main frequency offset direction identifiers corresponding to each group of box wall vibration time sequence waveform samples.

[0136] For each sample in the Dataset, perform the same processing as step S122 on its time-series waveform sample of the box wall vibration. That is, calculate the Hilbert transform envelope of the signal for each channel, find the peak value of the envelope after synthesis, and analyze the spectrum near the peak time to determine the dominant frequency offset direction. Record the obtained peak vibration amplitude time as the historical vibration amplitude peak time set VibTimes_sample, and record the vibration dominant frequency offset direction identifier associated with each time as the historical vibration dominant frequency offset direction identifier set VibFlags_sample.

[0137] Step S240: Perform temperature gradient field calculation processing on each group of the casing heat map time-series waveform samples to obtain the set of historical temperature field distortion moments and the associated set of historical temperature anomaly region spatial coordinates corresponding to each group of casing heat map time-series waveform samples.

[0138] For each sample in the Dataset, the same processing as step S123 is performed on its tubing heatmap time-series waveform sample. Specifically, the divergence and curl of the temperature gradient field are calculated for each heatmap frame, distortion moments are detected using their rate of change, and the distorted frames are segmented to extract the spatial coordinates of the abnormal regions. The detected temperature field distortion moments are recorded as the historical temperature field distortion moment set ThermalTimes_sample, and the set of spatial coordinates of the temperature abnormal regions associated with each moment is recorded as the historical temperature abnormal region spatial coordinate set ThermalRegions_sample.

[0139] Step S250: Initialize the network parameters of the transformer operation status reconstruction network, which includes a time correlation mapping subnetwork and a physical field coupling verification subnetwork.

[0140] A transformer operation status reconstruction network model is constructed. This network consists of two main sub-networks: a time-correlation mapping sub-network and a physical field coupling verification sub-network. The time-correlation mapping sub-network can adopt a sequence-to-sequence model based on the Transformer architecture. Its encoder part encodes the time sequence of each physical field event, and its decoder part outputs the candidate fault energy injection points and their accompanying response points. The physical field coupling verification sub-network can adopt a multilayer perceptron. Its input layer receives the feature vectors from steps S126, S127, and S128, the hidden layer uses the ReLU activation function, and the output layer is a three-node binary classification layer that outputs the probabilities of three verification Boolean values. Before training, all weight parameters and bias parameters of the network need to be initialized. The initialization method can be Xavier initialization, which randomly samples initial values ​​from a uniform distribution according to the number of input and output neurons to ensure good signal flow during forward and backward propagation.

[0141] Step S260: Input the set of historical gas concentration inflection point times, the set of historical vibration amplitude peak times, and the set of historical temperature field distortion times into the time correlation mapping sub-network. The time correlation mapping sub-network outputs the predicted candidate fault energy injection time, the predicted accompanying vibration response time, and the predicted accompanying thermal field response time corresponding to the transformer historical fault case sample.

[0142] The three time-series lists (GasTimes_sample, VibTimes_sample, and ThermalTimes_sample) of a sample are fed into a temporal correlation mapping sub-network. This sub-network first transforms discrete time values ​​into continuous feature vectors through an embedding layer, and then learns the temporal correlation patterns of different physical field events using a Transformer encoder-decoder structure. The network's output layer is a linear layer with three output heads, each outputting a scalar value. The first output head outputs the predicted candidate fault energy injection time T_pred_candidate, the second output head outputs the predicted accompanying vibration response time T_pred_vib_asso, and the third output head outputs the predicted accompanying thermal field response time T_pred_th_asso.

[0143] Step S270: Input the historical characteristic gas component ratio 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 set of historical temperature anomaly region spatial coordinates associated with the predicted accompanying thermal field response time into the physical field coupling verification sub-network. The physical field coupling verification sub-network outputs the predicted first verification Boolean value, the predicted second verification Boolean value, and the predicted third verification Boolean value.

[0144] Based on the predicted times T_pred_candidate, T_pred_vib_asso, and T_pred_th_asso output in step S260, the corresponding features are searched or interpolated from the results obtained in steps S220, S230, and S240. Specifically, the feature gas component ratio vector V_gas_pred closest to the time of T_pred_candidate is extracted from GasVectors_sample; the vibration dominant frequency offset direction identifier F_vib_pred closest to the time of T_pred_vib_asso is extracted from VibFlags_sample; and the set of spatial coordinates of the temperature anomaly region L_th_pred closest to the time of T_pred_th_asso is extracted from ThermalRegions_sample. V_gas_pred (a 7-dimensional vector), F_vib_pred (a discrete identifier that needs to be converted to one-hot encoding, e.g., [1, 0] represents a negative offset, [0, 1] represents a positive offset), and L_th_pred (a set of spatial points, whose center coordinates and the number of points can be calculated as features to form a feature vector, for example, 4-dimensional) are concatenated into a comprehensive feature vector V_input. This V_input is then fed into the physical field coupled verification sub-network (multilayer perceptron). After forward propagation, the values ​​of the three nodes in the output layer are activated by the Sigmoid function, yielding the probabilities P_b1, P_b2, and P_b3 for predicting the first verification Boolean value, respectively. By setting a threshold, e.g., 0.5, if P_b1 is greater than 0.5, the predicted first verification Boolean value B1_pred is true; otherwise, it is false. Similarly, B2_pred and B3_pred are obtained.

[0145] Step S280: Based on the logical AND operation results of the predicted first verification Boolean value, the predicted second verification Boolean value, and the predicted third verification Boolean value, obtain the predicted fault energy injection point confirmation signal output by the transformer operation status reconstruction network.

[0146] Perform a logical AND operation on B1_pred, B2_pred, and B3_pred outputs from step S270 to obtain the final predicted fault energy injection point confirmation signal, Confirm_pred. That is, Confirm_pred = B1_pred AND B2_pred AND B3_pred. Confirm_pred is a Boolean value indicating whether the network considers the candidate point output from step S260 to be a real fault energy injection point.

[0147] Step S290: Calculate the first time alignment loss between the predicted candidate fault energy injection time and the known actual fault energy injection time of the transformer's historical fault case sample.

[0148] A regression loss function, such as smoothed L1 loss, is used to calculate the difference between T_pred_candidate and the known true fault energy injection time T_true_label in the sample labels. The loss L_time is calculated as follows: if the absolute difference d = |T_pred_candidate - T_true_label| is less than a certain threshold δ, for example, 1.0, then L_time = 0.5 × d² / δ; otherwise, L_time = d - 0.5 × δ. This loss function encourages the network to predict the time as close as possible to the true time.

[0149] Step S2100: Calculate the first Boolean value matching loss between the predicted first verification Boolean value and the theoretical first verification Boolean value corresponding to the fault type label of the sample.

[0150] First, the theoretical first verification Boolean value needs to be determined based on the sample fault type label Label_sample. For example, a mapping table is established where if Label_sample belongs to a discharge fault (such as "inter-turn short circuit in high-voltage winding accompanied by partial discharge and development into arc" or "surface discharge inside bushing"), then the theoretical first verification Boolean value B1_true is true; otherwise, it is false. Then, the binary cross-entropy loss function is used to calculate the loss between the predicted probability P_b1 and the theoretical true value B1_true: L_b1 = -[B1_true × log(P_b1) + (1 - B1_true) × log(1 - P_b1)].

[0151] Step S2110: Calculate the second Boolean value matching loss between the predicted second verification Boolean value and the theoretical second verification Boolean value corresponding to the sample fault type label.

[0152] Similarly, the theoretical second verification Boolean value B2_true is determined based on Label_sample. If the fault type involves radial deformation of the winding, then B2_true is true; otherwise, it is false. The binary cross-entropy loss L_b2 between the predicted probability P_b2 and B2_true is calculated.

[0153] Step S2120: Calculate the third Boolean value matching loss between the predicted third verification Boolean value and the theoretical third verification Boolean value corresponding to the fault type label of the sample.

[0154] Similarly, the third theoretical verification Boolean value B3_true is determined based on Label_sample. If the fault occurs in a high field strength region, then B3_true is true; otherwise, it is false. The binary cross-entropy loss L_b3 between the predicted probability P_b3 and B3_true is calculated.

[0155] Step S2130: The first time alignment loss, the first Boolean value matching loss, the second Boolean value matching loss, and the third Boolean value matching loss are weighted and summed to obtain the total training loss value corresponding to the historical fault case sample of the transformer.

[0156] Different weighting coefficients are set to balance the importance of each loss. For example, let the weight λ_time of the time alignment loss be 1.0, and the weights λ_b1, λ_b2, and λ_b3 of the three Boolean value matching losses all be 0.5. Then the total loss L_total for this sample is equal to λ_time. L_time+λ_b1 L_b1+λ_b2 L_b2+λ_b3 L_b3.

[0157] Step S2140: Calculate the gradient of the total training loss with respect to all network parameters of the transformer operation state reconstruction network using the gradient backpropagation algorithm.

[0158] Based on L_total calculated in step S2130, the partial derivatives of L_total with respect to each trainable parameter (weight W and bias b) in the network are calculated in reverse, starting from the output layer. These gradients indicate the direction and magnitude of adjustment required for each parameter to reduce the loss.

[0159] Step S2150: Update the network parameters of the transformer operation status reconstruction network according to the calculated gradient.

[0160] An optimization algorithm, such as the Adam optimizer, is used to update the network parameters using the gradient calculated in step S2140. The update rule is: New parameters = Old parameters - Learning rate. Gradient. The learning rate is a hyperparameter, for example, set to 0.001. Through continuous iterative updates, the network's performance gradually improves.

[0161] Step S2160: Traverse all samples in the transformer historical fault case sample set, and repeatedly perform the step of performing gas concentration curve inflection point detection processing on each group of dissolved gas time-series waveform samples in the oil until the network parameters of the transformer operation status reconstruction network are updated according to the calculated gradient, until the total training loss value converges on the independent validation set.

[0162] The entire dataset is divided into training, validation, and test sets. Steps S220 to S2150 are repeated using the training set to complete one training epoch. After each epoch, the average total loss is calculated using the validation set. Training continues for multiple epochs. When the average total loss on the validation set stops decreasing or begins to increase, it indicates that the model has converged, and training is stopped. At this point, the network parameters are saved, resulting in the pre-built transformer operation state reconstruction network after training.

[0163] Step S310: Within the preset observation time window after sending the maintenance decision trigger command, continuously collect the subsequent dissolved gas in the transformer oil timing waveform, the subsequent tank wall vibration timing waveform, and the subsequent bushing thermal image timing waveform.

[0164] Sending the maintenance decision trigger command (Command) to the master station system in step S157 does not signify the end of the diagnostic process. The system will initiate a follow-up verification process. An observation time window is set, for example, 7 days from the time the command is sent. During this period, the latest monitoring data of the transformer continues to be collected in real time as described in step S110, generating subsequent dissolved gas in oil time-series waveforms (Data_dga_post), subsequent tank wall vibration time-series waveforms (Data_vib_post), and subsequent bushing thermal image time-series waveforms (Data_th_post).

[0165] Step S320: Perform the step of inputting the time-series waveforms of dissolved gas in oil, vibration of the tank wall, and heat map of the bushing into the pre-constructed transformer operation status reconstruction network for cross-modal physical field spatiotemporal alignment processing on the subsequent time-series waveforms of dissolved gas in oil, vibration of the tank wall, and heat map of the bushing, to generate the subsequent multi-physics field evolution trajectory of the transformer.

[0166] The collected subsequent data, Data_dga_post, Data_vib_post, and Data_th_post, are used as input and fed into the already trained transformer operation status reconstruction network. Step S120 and all its sub-steps S121 to S1210 are then executed. The network will output the subsequent transformer multiphysics evolution trajectory Traj_post, based on the new fault energy injection points that may appear within this subsequent time window.

[0167] Step S330: Perform the step of performing topological source tracing processing on the multiphysics evolution trajectory of the transformer for the subsequent transformer, and generate the global fault feature topology chain of the subsequent transformer.

[0168] Using the subsequent evolution trajectory Traj_post generated in step S320 as input, step S130 and all its sub-steps S131 to S139 are executed completely. By performing topological source tracing on the events in Traj_post, a new tree structure is constructed, namely the subsequent transformer global fault feature topology chain TopoChain_post.

[0169] Step S340: Perform a topology evolution comparison analysis between the subsequent transformer global fault feature topology chain and the original transformer global fault feature topology chain on which the maintenance decision trigger command is based, and calculate the topology similarity and node attribute offset.

[0170] The original transformer global fault feature topology chain, TopoChain_orig, is retrieved from the system when generating the Command instruction. A graph similarity metric is used to compare TopoChain_post and TopoChain_orig. First, the topological similarity Sim_graph is calculated. This can be done using a graph kernel function, such as calculating the length of the longest common subsequence of two tree structures, or using network embedding methods to obtain the embedding vectors of the two graphs and then calculating the cosine similarity. Second, the node attribute offset Offset_attr is calculated. For key nodes in TopoChain_orig (such as the root node and major intermediate nodes), the temporally closest corresponding nodes are found in TopoChain_post, and the relative percentage change in their distortion intensity values ​​is calculated. This average is then applied to all nodes to obtain the average attribute offset Offset_mean.

[0171] Step S350: Based on the topological similarity and node attribute offset, determine whether the fault development path predicted in the maintenance decision triggering instruction is consistent with the subsequent actual observed fault evolution trend, and generate a diagnostic conclusion validity verification label.

[0172] Define the judgment rules. If the topological similarity Sim_graph is greater than the preset structural similarity threshold Th_sim (e.g., 0.8), and the average attribute offset Offset_mean is less than the preset offset threshold Th_offset (e.g., 0.2), then the fault development path predicted by the original diagnostic conclusion is considered to be highly consistent with the subsequent observed evolution trend. In this case, the diagnostic conclusion validity verification label Label_valid is generated as "valid". Otherwise, the diagnostic conclusion is considered to deviate from the actual situation, and the label Label_valid is generated as "questionable".

[0173] Step S360: Extract the original transformer global fault feature topology chain and the original transformer comprehensive diagnostic descriptor corresponding to the current diagnostic case, and associate and store the original transformer global fault feature topology chain, the original transformer comprehensive diagnostic descriptor and the diagnostic conclusion validity verification label to form a diagnostic case record with verification label.

[0174] The original topology chain TopoChain_orig used in this diagnosis process, the generated original comprehensive diagnostic descriptor Final_Diag, and the validation label Label_valid obtained in step S350 are stored as a whole in a dedicated diagnostic case knowledge base. This forms a diagnostic case record Case_record with a validation label, whose structure is {"topo_chain": TopoChain_orig, "diagnosis": Final_Diag, "validation_label": Label_valid, "timestamp": Time_stamp}.

[0175] Step S370: If the diagnostic conclusion validity verification label indicates that the diagnostic conclusion is valid, then the original transformer global fault feature topology chain in the diagnostic case record with the verification label is converted into a topology embedding vector and added as a new positive sample to the fault mode standard topology embedding vector library of the fault mode semantic parsing network.

[0176] If Label_valid is "valid", it indicates that the fault mode represented by TopoChain_orig is accurate and reliable. TopoChain_orig is fed into the topology encoding layer of the fault mode semantic parsing network (i.e., the forward propagation process described in steps S141 to S144), but without passing through the subsequent matching layer; only its output topology embedding vector V_embedding_valid is extracted. This V_embedding_valid is associated with its fault type semantic label (extracted from Final_Diag, such as "inter-turn short circuit in the middle of the high-voltage winding accompanied by partial discharge and inducing a surge of dissolved acetylene gas in the oil"), and added as a new standard mode sample to the fault mode standard library Lib of the fault mode semantic parsing network. This achieves dynamic expansion and self-reinforcement of the knowledge base.

[0177] Step S380: If the diagnostic conclusion validity verification label indicates that the diagnostic conclusion is questionable, then a diagnostic process backtracking analysis is triggered. The diagnostic process backtracking analysis includes verifying the time alignment accuracy of physical field distortion events in the original transformer multi-physics field evolution trajectory output by the transformer operation status reconstruction network, and reviewing the matching threshold setting of the fault mode semantic parsing network.

[0178] If Label_valid is "Questionable," a diagnostic process backtracking analysis is initiated. First, the temporal alignment accuracy of events in the original evolutionary trajectory Traj_orig is verified. The event times in Traj_orig are compared with the more accurate fault evolution start times derived from subsequent data Traj_post, and the temporal alignment error is calculated. Second, the similarity matching threshold used by the fault mode semantic parsing network in step S144 is reviewed. It is checked whether the difference between the highest similarity score and other candidate scores is too small when generating Final_Diag, leading to the selection of an incorrect mode.

[0179] Step S390: Generate a diagnostic uncertainty analysis report based on the results of the diagnostic process backtracking analysis. The diagnostic uncertainty analysis report is used to guide the adjustment of the preset gas response threshold, preset vibration response threshold, or preset thermal field response threshold.

[0180] The analysis results from step S380 are compiled into a diagnostic uncertainty analysis report, Report_analysis. This report identifies specific steps that may lead to diagnostic bias. For example, if a large time alignment error is found, the report might suggest adjusting the time window size used in steps S124 to S129 for searching for accompanying response points, or adjusting the response thresholds G_th, V_th, and T_threshold used in steps S132 to S134 for determining child nodes. If ambiguous pattern matching is found, the report might suggest adjusting the similarity threshold in step S144, or adding new standard samples to refine the pattern library.

[0181] Step S3100: Integrate the validity verification tag of the diagnostic conclusion, the knowledge base update status, and the diagnostic uncertainty analysis report to generate a credibility assessment and knowledge enhancement report for this diagnosis, and send the credibility assessment and knowledge enhancement report to the transformer online monitoring master station system.

[0182] Finally, the validity verification label (Label_valid) of the diagnostic conclusions, the update status of the knowledge base (e.g., "new sample added" or "not updated"), and the diagnostic uncertainty analysis report (Report_analysis) are integrated to generate a final credibility assessment and knowledge enhancement report (Final_report). This report is then sent to the transformer online monitoring master station system via the communication network for review by senior management and system maintenance engineers. This forms a complete closed-loop diagnostic system with self-learning and self-optimization capabilities.

[0183] For example, the method further includes:

[0184] Step S410: The physical field coupling verification layer pre-stores a standard characteristic gas component ratio vector template corresponding to the discharge fault type, a standard vibration main frequency offset direction identifier template corresponding to the winding radial deformation mode, and a spatial coordinate boundary template of the preset high field strength region inside the transformer.

[0185] Three types of template data are pre-stored within the physical field coupling verification layer. The first type is the standard template for discharge faults, which is not a single vector but a vector space. For example, it stores the characteristic gas component ratio vectors under multiple typical discharge fault cases, forming a template library Template_gas_discharge={V_gas_sample1, V_gas_sample2, ...}, where each vector V_gas_sample corresponds to a specific discharge type, such as "partial discharge of bubbles in oil", "surface discharge of solid insulation", "discharge of metal particles", etc. The second type is the standard template for the radial deformation mode of windings, which stores a standard vibration dominant frequency offset direction identifier, i.e., a negative offset identifier. The third type is the spatial coordinate boundary template for the preset high field strength region, which, as mentioned earlier, is the boundary coordinate data of a spatial region R_intense described by a polyhedron.

[0186] Step S420: Calculate the vector cosine similarity between the characteristic gas component ratio vector corresponding to the candidate fault energy injection point and the standard characteristic gas component ratio vector template to obtain the gas mode matching degree.

[0187] Calculate the cosine similarity between the feature gas component proportion vector V_gas_candidate of the candidate point and each standard vector V_gas_sample in the template library Template_gas_discharge, and take the maximum value as the gas mode matching degree Match_gas. That is, Match_gas=max(cos_sim(V_gas_candidate,V_gas_sample)forallsamples).

[0188] Step S430: Calculate the sign matching degree between the vibration dominant frequency offset direction identifier corresponding to the accompanying vibration response point and the standard vibration dominant frequency offset direction identifier template to obtain the vibration mode matching degree.

[0189] The vibration dominant frequency offset direction identifier F_vib_asso corresponding to the accompanying vibration response point is compared with the standard identifier (negative offset identifier). If F_vib_asso is also a negative offset identifier, the vibration mode matching degree Match_vib is assigned a value of 1; otherwise, Match_vib is assigned a value of 0.

[0190] Step S440: Calculate the spatial overlap between the set of spatial coordinates of the temperature anomaly region corresponding to the accompanying thermal field response point and the spatial coordinate boundary template of the preset high field strength region inside the transformer, and obtain the thermal field mode matching degree.

[0191] Calculate the spatial overlap between the temperature anomaly region's spatial coordinate set L_th_asso and the preset high-field-strength region R_intense. The intersection-union ratio (IUGR) can be used to measure this. First, calculate the intersection volume V_overlap of the spatial voxels (or the convex hull of the point set) occupied by L_th_asso and R_intense. Then, calculate the volume V_thermal of L_th_asso itself. The spatial overlap is Match_th = V_overlap / V_thermal. This spatial overlap, between 0 and 1, indicates the proportion of the anomaly region that falls within the high-field-strength region.

[0192] Step S450: The physical field coupling verification layer is pre-set with a physical field interference coefficient matrix. The physical field interference coefficient matrix defines the causal support strength coefficients between each pair of gas mode, vibration mode and thermal mode when characterizing the same fault.

[0193] The physical field coupling verification layer maintains a 3 The physical field interference coefficient matrix M_interfere is 3. The rows and columns of this matrix correspond to the three modes: gas, vibration, and thermal. The off-diagonal elements M_ij represent the causal support strength of mode i for mode j. For example, M_gas_vib represents the causal support of gas anomalies for vibration anomalies, which can be preset to 0.8; M_gas_th represents the causal support of gas anomalies for thermal anomalies, which is preset to 0.6; and M_vib_th represents the causal support of vibration anomalies for thermal anomalies, which is preset to 0.7. The diagonal elements can be set to 1.

[0194] Step S460: Based on the physical field interference coefficient matrix, multiply the gas mode matching degree, vibration mode matching degree, and thermal field mode matching degree pairwise and sum them by weight to calculate the causal support degree of the gas mode to the vibration mode, the causal support degree of the gas mode to the thermal field mode, and the causal support degree of the vibration mode to the thermal field mode.

[0195] Calculate the causal support C_gas_vib of the gas mode to the vibration mode, where C_gas_vib = Match_gas Match_vib M_gas_vib. Calculate the causal support C_gas_th of the gas mode to the thermal mode, where C_gas_th = Match_gas. Match_th M_gas_th. Calculate the causal support C_vib_th of the vibration mode to the thermal mode, where C_vib_th = Match_vib. Match_th M_vib_th.

[0196] Step S470: Compare the causal support of the gas mode to the vibration mode with a preset gas vibration causal support threshold. If it exceeds the gas vibration causal support threshold, generate a first intermediate verification Boolean value; otherwise, generate the inverse of the first intermediate verification Boolean value.

[0197] Set the gas vibration causal support threshold Th_gv, for example, a value of 0.5. Compare C_gas_vib with Th_gv. If C_gas_vib is greater than Th_gv, then the first intermediate verification boolean value B_inter1 is true; otherwise, B_inter1 is false.

[0198] Step S480: Compare the causal support of the gas mode to the thermal field mode with a preset gas thermal field causal support threshold. If it exceeds the gas thermal field causal support threshold, generate a second intermediate verification Boolean value; otherwise, generate the inverse of the second intermediate verification Boolean value.

[0199] Set the causal support threshold Th_gt for the gas thermal field, for example, a value of 0.4. Compare C_gas_th with Th_gt. If C_gas_th is greater than Th_gt, then the second intermediate verification boolean value B_inter2 is true; otherwise, B_inter2 is false.

[0200] Step S490: Compare the causal support of the vibration mode to the thermal field mode with a preset vibration thermal field causal support threshold. If it exceeds the vibration thermal field causal support threshold, generate a third intermediate verification Boolean value; otherwise, generate the inverse of the third intermediate verification Boolean value.

[0201] Set the vibrational thermal field causal support threshold Th_vt, for example, a value of 0.45. Compare C_vib_th with Th_vt. If C_vib_th is greater than Th_vt, then the third intermediate verification boolean value B_inter3 is true; otherwise, B_inter3 is false.

[0202] Step S4100: Perform majority voting logic operation on the first intermediate verification Boolean value, the second intermediate verification Boolean value and the third intermediate verification Boolean value, and use the voting result as the final assignment output of the first verification Boolean value, the second verification Boolean value and the third verification Boolean value to complete the physical field coupling verification.

[0203] A majority vote is performed on B_inter1, B_inter2, and B_inter3. That is, the number of true values ​​for these three Boolean values ​​is counted. If the number of true values ​​is greater than or equal to 2 (i.e., more than half), the majority vote result Maj_vote is true; otherwise, Maj_vote is false. Finally, the first verification Boolean value B_gas, the second verification Boolean value B_vib, and the third verification Boolean value B_th, which were originally to be output in steps S126, S127, and S128, are all uniformly assigned the same value as this majority vote result Maj_vote. That is, B_gas = Maj_vote, B_vib = Maj_vote, and B_th = Maj_vote. This utilizes the mutual verification relationship between physical fields, improves the robustness of single-stage verification, and completes coupled verification based on the physical field interference mechanism.

[0204] In one exemplary embodiment, a transformer fault feature fusion diagnostic system based on multimodal time-series data reconstruction is provided. This system can be a terminal, server, etc., and its internal structure diagram can be as follows: Figure 2 As shown, the transformer fault feature fusion diagnostic system based on multimodal time-series data reconstruction includes a processor, memory, input / output interface, communication interface, display unit, and input device. The processor, memory, and input / output interface are connected via a system bus, and the communication interface, display unit, and input device are also connected to the system bus via the input / output interface. The processor provides computational and control capabilities. The memory includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores the operating system and computer programs. The internal memory provides the environment for the operation of the operating system and computer programs in the non-volatile storage medium. The input / output interface is used for exchanging information between the processor and external devices. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, mobile cellular networks, near-field communication, or other technologies. When the computer program is executed by the processor, it implements a transformer fault feature fusion diagnostic method based on multimodal time-series data reconstruction. The display unit is used to form a visually visible image and can be a display screen, projection device, or virtual reality imaging device. The display screen can be an LCD screen or an e-ink screen. The input device can be a touch layer covering the display screen, or a button, trackball, or touchpad set on the casing of the transformer fault feature fusion diagnostic system based on multimodal time-series data reconstruction, or an external keyboard, touchpad, or mouse, etc.

[0205] The aforementioned technical solution can be fully implemented by an embodied intelligent robot. Specifically, the embodied intelligent robot can serve as a mobile sensing and computing platform, integrating an oil chromatography monitor, a multi-channel vibration acceleration sensor array, and an infrared thermal imager. After autonomously navigating to the preset monitoring point of the transformer, it synchronously triggers multimodal sensors through a unified clock reference to accurately collect the time-series waveforms of dissolved gas in the oil, tank wall vibration, and bushing thermal images. Utilizing an onboard edge computing unit, the embodied intelligent robot first detects inflection points on the gas concentration curve, captures the envelope peak of the vibration waveform, and calculates the temperature gradient field of the thermal image sequence, extracting the abrupt event moments of each physical field and their associated feature vectors. Subsequently, it inputs the three types of event moments into a pre-trained transformer operating status reconstruction network. Through a time-related mapping layer and a physical field coupling verification layer, it identifies the earliest real fault energy injection point that meets the gas-vibration-thermal field coupling verification conditions, and uses this moment as a reference to generate a transformer multi-physics field evolution trajectory containing three branch lists: gas, vibration, and thermal field. Building upon this foundation, the embodied intelligent robot further performs topological tracing, organizing the evolution trajectory into a tree-like topological chain with the fault energy injection point as the root node and each physical field distortion event as a child node. It then extracts the coordinates of the fault initiation energy release point, the sequence of physical field distortion points along the fault path, and the boundary of the fault termination energy dissipation zone. Subsequently, the robot inputs the topological chain into a fault mode semantic parsing network. Through graph neural network encoding and polar coordinate space clustering, it generates a comprehensive transformer diagnostic descriptor that integrates fault type semantic labels and spatial location coordinates. Finally, based on the fault spatial location coordinates in the diagnostic results, the embodied intelligent robot extracts local segments from the evolution trajectory to predict the remaining lifespan consumption rate. It automatically generates maintenance decision trigger commands containing urgency level, fault type, and spatial location, and sends them to the online monitoring master station system via a dedicated power network. This completes the entire autonomous operation process from data acquisition and fusion diagnosis to decision command generation, making the embodied intelligent robot a mobile intelligent diagnostic terminal for transformer health status.

[0206] It should be noted that, in order to simplify the description of the present invention and thus help to understand one or more embodiments of the invention, multiple features may sometimes be grouped into one embodiment, drawing or description thereof in the foregoing description of the embodiments of the present invention.

Claims

1. A transformer fault feature fusion diagnosis method based on multimodal time-series data reconstruction, characterized in that, The method includes: The time-series waveforms of dissolved gas in oil, tank wall vibration, and bushing thermal image were captured by an oil chromatograph, a vibration acceleration sensor, and an infrared thermal imager, respectively, during the continuous operation of the transformer. The time-series waveforms of dissolved gas in oil, tank wall vibration, and bushing thermal imaging are input into a pre-constructed transformer operation status reconstruction network for cross-modal physical field spatiotemporal alignment processing. This generates a transformer multi-physics field evolution trajectory with the fault energy injection point as the starting reference. The transformer operation status reconstruction network is a deep neural network model used to align and fuse time-series waveforms of different physical fields, different sampling rates, and different data formats in the time and space dimensions to reconstruct the evolution path of each physical field after the fault occurs. The multi-physics field evolution trajectory of the transformer is subjected to topological source tracing processing of fault feature clues to obtain a transformer global fault feature topology chain containing the coordinates of the fault initiation energy release point, the sequence of physical field distortion points of the fault path, and the boundary of the fault termination energy dissipation zone. The transformer global fault feature topology chain is input into the fault mode semantic parsing network for spatiotemporal coupling verification of fault symptoms and lesion areas, generating a transformer comprehensive diagnostic descriptor that integrates fault type semantic labels and fault spatial location coordinates. This fault mode semantic parsing network is a deep model that combines graph neural networks and spatial transformation networks, aiming to parse the fault type and precise spatial location from the transformer global fault feature topology chain. Based on the fault spatial location coordinates in the transformer comprehensive diagnostic descriptor, the corresponding local evolution trajectory fragments are extracted from the transformer multiphysics field evolution trajectory. Based on the local evolution trajectory fragments, a maintenance decision triggering command containing the fault remaining life consumption rate is generated, and the maintenance decision triggering command is sent to the transformer online monitoring master station system. The process involves inputting the time-series waveforms of dissolved gas in oil, tank wall vibration, and bushing thermal imaging into a pre-constructed transformer operation status reconstruction network for cross-modal physical field spatiotemporal alignment processing. This generates a transformer multi-physics field evolution trajectory starting from the fault energy injection point, including: The time-series waveform of dissolved gas in oil is processed by gas concentration curve inflection point detection to obtain multiple gas concentration inflection point times corresponding to the time-series waveform of dissolved gas in oil and the characteristic gas component ratio vector associated with each gas concentration inflection point time. The vibration waveform envelope peak point capture processing is performed on the vibration time sequence waveform of the box wall to obtain multiple vibration amplitude peak times corresponding to the vibration time sequence waveform of the box wall and the vibration main frequency offset direction identifier associated with each vibration amplitude peak time; The heat map time-series waveform of the casing is processed by the temperature gradient field calculation of the heat map pixels to obtain multiple temperature field distortion moments corresponding to the heat map time-series waveform of the casing and the set of spatial coordinates of the temperature anomaly region associated with each temperature field distortion moment. The gas concentration inflection point, vibration amplitude peak time, and temperature field distortion time are input into the time correlation mapping layer of the transformer operation status reconstruction network for cross-modal mutation event time sequence processing, and the gas concentration inflection point time that appears first on the time axis is identified as the candidate fault energy injection point. The peak vibration amplitude moment with the shortest time interval to the candidate fault energy injection point on the time axis is identified as the accompanying vibration response point of the candidate fault energy injection point. At the same time, the temperature field distortion moment with the shortest time interval to the candidate fault energy injection point on the time axis is identified as the accompanying thermal field response point of the candidate fault energy injection point. The characteristic gas component ratio vector corresponding to the candidate fault energy injection point is input into the physical field coupling verification layer of the transformer operation status reconstruction network to determine whether the characteristic gas component ratio vector points to the discharge fault type and output the first verification Boolean value. Input the vibration dominant frequency offset direction identifier corresponding to the accompanying vibration response point into the physical field coupling verification layer, determine whether the vibration dominant frequency offset direction identifier points to the winding radial deformation mode and output the second verification Boolean value; Input the set of spatial coordinates of the temperature anomaly region corresponding to the accompanying thermal field response point into the physical field coupling verification layer, determine whether the set of spatial coordinates of the temperature anomaly region is located within the preset high field strength region inside the transformer and output the third verification Boolean value. Based on the logical AND operation result of the first verification Boolean value, the second verification Boolean value, and the third verification Boolean value, it is confirmed whether the candidate fault energy injection point is a real fault energy injection point. If the logical AND operation result is true, the candidate fault energy injection point is confirmed to be a real fault energy injection point. Taking the actual fault energy injection time corresponding to the confirmed fault energy injection point as the starting benchmark, the physical field type classification and recombination processing is performed on all gas concentration inflection points and their associated characteristic gas component ratio vectors, all vibration amplitude peak times and their associated vibration dominant frequency offset direction identifiers, and all temperature field distortion times and their associated temperature anomaly region spatial coordinate sets after the fault energy injection time in ascending order of time. This generates a transformer multi-physics evolution trajectory with the fault energy injection point as the starting benchmark, the occurrence time of physical field distortion events as the time axis, and the physical field distortion type as the classification axis.

2. The transformer fault feature fusion diagnosis method based on multimodal time series data reconstruction according to claim 1, characterized in that, The topological source tracing processing of the multi-physics evolution trajectory of the transformer, which involves fault feature clues, yields a transformer-wide fault feature topological chain containing the coordinates of the fault initiation energy release point, the sequence of physical field distortion points along the fault path, and the boundary of the fault termination energy dissipation region. The actual fault energy injection time corresponding to the fault energy injection point is extracted from the multiphysics field evolution trajectory of the transformer and used as the starting node for topology tracing. The actual fault energy injection time is marked as the root node of the transformer global fault feature topology chain and assigned a root node type identifier and root node spatial coordinates. Search for the first gas concentration inflection point time adjacent to the actual fault energy injection time in the multiphysics field evolution trajectory of the transformer. Determine whether the gas distortion intensity value corresponding to the first gas concentration inflection point time exceeds the preset gas response threshold. If it does, mark the first gas concentration inflection point time as the first-level gas child node of the root node in the transformer global fault feature topology chain. Search for the first vibration amplitude peak moment adjacent to the actual fault energy injection moment in the multiphysics field evolution trajectory of the transformer. Determine whether the vibration distortion intensity value corresponding to the first vibration amplitude peak moment exceeds the preset vibration response threshold. If it does, mark the first vibration amplitude peak moment as the first-level vibration sub-node of the root node in the transformer global fault feature topology chain. Search for the first temperature field distortion moment adjacent to the actual fault energy injection moment in the multiphysics field evolution trajectory of the transformer. Determine whether the thermal field distortion intensity value corresponding to the first temperature field distortion moment exceeds the preset thermal field response threshold. If it exceeds the threshold, mark the first temperature field distortion moment as the first-level thermal field sub-node of the root node in the transformer global fault feature topology chain. Using the time corresponding to each first-level sub-node as the time reference, in the multi-physics field evolution trajectory of the transformer, for each type of physical field, search for the next gas concentration inflection point, the next vibration amplitude peak time, and the next temperature field distortion time on the subsequent time axis that satisfy the condition of temporal proximity and the monotonically increasing distortion intensity value of the corresponding type of physical field. Mark each searched time as the second-level sub-node of the corresponding first-level sub-node. Using the time corresponding to each secondary sub-node as the time reference, in the multi-physics field evolution trajectory of the transformer, for each type of physical field, search for the next gas concentration inflection point, the next vibration amplitude peak time, and the next temperature field distortion time on the subsequent time axis that satisfy the conditions of temporal adjacency and the monotonically increasing distortion intensity value of the physical field itself. Mark each searched time as the tertiary sub-node of the corresponding secondary sub-node. Continue searching downwards level by level until the next physical field distortion event time that satisfies the condition of temporal proximity and increasing distortion intensity of the corresponding physical field itself can no longer be found. Mark all physical field distortion event times found in the last level search as leaf nodes of the transformer global fault feature topology chain. The spatial coordinates of the root node corresponding to the root node of the transformer global fault feature topology chain are extracted as the coordinates of the fault initiation energy release point. At the same time, the physical field distortion event occurrence time offset and its spatial coordinates corresponding to all intermediate nodes that are neither root nodes nor leaf nodes in the transformer global fault feature topology chain are extracted as the fault path physical field distortion point sequence. The spatial coordinates corresponding to all leaf nodes in the transformer global fault feature topology chain are extracted and the boundary convex hull calculation is performed to obtain the smallest convex polygon that surrounds the spatial coordinates of all leaf nodes as the boundary of the fault termination energy dissipation zone. The coordinates of the fault initiation energy release point, the sequence of physical field distortion points of the fault path, and the boundary of the fault termination energy dissipation zone are then subjected to topological association encoding to generate the transformer global fault feature topology chain.

3. The transformer fault feature fusion diagnosis method based on multimodal time series data reconstruction according to claim 1, characterized in that, The transformer global fault feature topology chain is input into the fault mode semantic parsing network for spatiotemporal coupling verification of fault symptoms and lesion areas, generating a comprehensive transformer diagnostic descriptor that integrates fault type semantic labels and fault spatial location coordinates, including: The transformer global fault feature topology chain is input into the topology coding layer of the fault mode semantic parsing network, and the connection relationship between the root node and each level of child nodes is processed by adjacency matrix coding to obtain the topology connection relationship coding vector. The transformer global fault feature topology chain is input into the topology coding layer of the fault mode semantic parsing network, and the physical field distortion type corresponding to each node is processed by one-hot coding to obtain the physical field type distribution coding vector of each node. The transformer global fault feature topology chain is input into the topology coding layer of the fault mode semantic parsing network, and the distortion intensity level parameter corresponding to each node is normalized and mapped to obtain the distortion intensity evolution trend coding vector of each node. The topological connection relationship encoding vector, the physical field type distribution encoding vector, and the distortion intensity evolution trend encoding vector are concatenated and fused to obtain a topological structure embedding vector representing the tree topological structure. The topological structure embedding vector is input into the fault type matching layer of the fault mode semantic parsing network and compared with the pre-stored standard topological structure embedding vectors of each fault mode to obtain a structural similarity score sequence. The fault mode corresponding to the structural similarity score sequence is used as a candidate fault mode and the corresponding fault type semantic label is output. The coordinates of the fault initiation energy release point in the transformer global fault feature topology chain are input into the spatial positioning reference generation layer of the fault mode semantic parsing network. The three-dimensional spatial coordinate system inside the transformer is established with the coordinates of the fault initiation energy release point as the spatial origin. At the same time, the coordinates of each distortion point in the fault path physical field distortion point sequence are converted into polar coordinates relative to the spatial origin, resulting in a set of polar coordinates of fault path points including azimuth angle parameters, pitch angle parameters, and radial distance parameters. The set of polar coordinates of the fault path points is input into the lesion area delineation layer of the fault mode semantic parsing network for spatial clustering of polar coordinate points. Based on the distribution density of the azimuth and pitch angle parameters of each polar coordinate point, the spatial sector corresponding to the main release direction of fault energy is identified. At the same time, based on the distribution range of the radial distance parameter of each polar coordinate point, the radial extension length of the fault-affected area is identified. The spatial sector and the radial extension length are combined and encoded to generate fault spatial positioning coordinates with the coordinates of the fault initiation energy release point as the origin. The fault type semantic tags corresponding to the selected candidate fault modes are associated and bound with the fault spatial location coordinates to generate a preliminary diagnostic descriptor for each candidate fault mode, which includes fault type semantic tags and fault spatial location coordinates. All preliminary diagnostic descriptors are packaged and combined to output a transformer comprehensive diagnostic descriptor that integrates fault type semantic tags and fault spatial location coordinates.

4. The transformer fault feature fusion diagnosis method based on multimodal time series data reconstruction according to claim 3, characterized in that, The process involves inputting the coordinates of the fault initiation energy release point in the transformer's global fault feature topology chain into the spatial positioning reference generation layer of the fault mode semantic parsing network. A three-dimensional spatial coordinate system within the transformer is established using the coordinates of the fault initiation energy release point as the spatial origin. Simultaneously, the coordinates of each distortion point in the fault path physical field distortion point sequence are converted into polar coordinates relative to the spatial origin, resulting in a set of polar coordinates for fault path points containing azimuth, pitch, and radial distance parameters. The coordinates of the fault initiation energy release point are read from the transformer global fault feature topology chain. The coordinates of the fault initiation energy release point are a vector containing three-dimensional values, which represent the positions in the preset transformer global three-dimensional space coordinate system. Using the coordinates of the fault initiation energy release point as the origin of the spatial coordinate system, a new three-dimensional spatial coordinate system inside the transformer is established. The three coordinate axes of the new coordinate system are parallel to the coordinate axes of the transformer's global three-dimensional spatial coordinate system. The fault path physical field distortion point sequence is read from the transformer global fault feature topology chain. The fault path physical field distortion point sequence contains multiple physical field distortion points, and each physical field distortion point contains its distortion point coordinates in the transformer global three-dimensional space coordinate system. For each distorted point coordinate in the sequence of distorted points in the physical field of the fault path, calculate the difference vector obtained by subtracting the coordinate vector of the fault initiation energy release point from the distorted point coordinate vector. This difference vector is used as the position vector of the distorted point relative to the origin of the spatial coordinate system. The calculated position vector is decomposed into components along the X-axis, Y-axis, and Z-axis of the new coordinate system. The azimuth parameter of the distorted point relative to the origin of the spatial coordinate system is calculated based on the projection of the position vector onto the XY plane. The azimuth parameter is defined as the angle traversed by rotating counterclockwise from the positive X-axis of the new coordinate system to the position of the projection vector. The pitch angle parameter of the distorted point relative to the origin of the spatial coordinate system is calculated based on the angle between the position vector itself and the positive direction of the Z-axis. The pitch angle parameter is defined as the complementary angle between the position vector and the positive direction of the Z-axis. Calculate the magnitude of the position vector to obtain the radial distance parameter of the distorted point relative to the origin of the spatial coordinate system; The azimuth, pitch, and radial distance parameters corresponding to the distortion point are combined into a three-dimensional parameter set, which represents the coordinates of the distortion point in the newly established polar coordinate system with the coordinates of the fault initiation energy release point as the origin. For each distortion point in the sequence of distortion points in the physical field of the fault path, the steps of calculating the difference vector obtained by subtracting the coordinate vector of the fault initiation energy release point from the coordinate vector of the distortion point are repeated until the steps of combining the azimuth parameter, pitch parameter, and radial distance parameter corresponding to the distortion point into a three-dimensional parameter set are performed to obtain the three-dimensional parameter set corresponding to all distortion points. Arrange the three-dimensional parameter sets corresponding to all distortion points according to the order of the distortion points in the sequence to form a set of polar coordinates of the fault path points, including azimuth, pitch, and radial distance parameters.

5. The transformer fault feature fusion diagnosis method based on multimodal time series data reconstruction according to claim 3, characterized in that, The process involves inputting the set of polar coordinates of the fault path points into the lesion region delineation layer of the fault pattern semantic parsing network for spatial clustering of the polar coordinate points. Based on the density of the azimuth and pitch angle parameters of each polar coordinate point, spatial sectors corresponding to the main direction of fault energy release are identified. Simultaneously, based on the distribution range of the radial distance parameter of each polar coordinate point, the radial extension length of the fault-affected area is identified. The spatial sectors and radial extension length are then combined and encoded to generate fault spatial positioning coordinates with the coordinates of the fault initiation energy release point as the origin. This includes: Extract the azimuth and pitch parameters corresponding to each fault path point from the set of polar coordinates of the fault path points, and map the azimuth and pitch parameters of each fault path point onto a unit sphere centered on the coordinates of the fault initiation energy release point to obtain the coordinates of the direction point on the unit sphere corresponding to each fault path point. Density-based spatial clustering is performed on the coordinates of all directional points on a unit sphere. Clustering scanning radius parameters and minimum neighborhood point number parameters are set to identify multiple dense regions where the distribution density of directional point coordinates exceeds a preset density threshold. Each dense region corresponds to a set of fault path points with similar spatial orientations. The arithmetic mean of the azimuth parameters corresponding to the coordinates of all directions within each dense region is calculated as the center azimuth parameter of that dense region. The arithmetic mean of the elevation parameters corresponding to the coordinates of all directions within each dense region is calculated as the center elevation parameter of that dense region. The center azimuth parameter and the center elevation parameter of each dense region are combined as the center direction of the space sector corresponding to that dense region. Starting from the coordinates of the fault initiation energy release point, and taking the direction of the center of the spatial sector corresponding to each dense region as the directional ray, a cone-shaped spatial region is constructed with each directional ray as the axis and a preset cone angle as the half cone angle. Each cone-shaped spatial region is marked as the spatial sector corresponding to the dense region. From the set of polar coordinates of the fault path points, all fault path points whose spatial locations belong to each spatial sector are selected. The radial distance parameter of the fault path point corresponding to each spatial sector is extracted. The maximum value search processing is performed on the radial distance parameter corresponding to each spatial sector to obtain the radial extension length of the fault influence area corresponding to each spatial sector. The spatial sector center direction of each spatial sector is combined with the radial extension length of the corresponding fault-affected area to generate multiple fault spatial positioning coordinates with the fault initiation energy release point coordinates as the origin, the spatial sector center direction as the orientation description, and the corresponding fault-affected area radial extension length as the distance description.

6. The transformer fault feature fusion diagnosis method based on multimodal time series data reconstruction according to claim 1, characterized in that, The step of extracting corresponding local evolution trajectory segments from the transformer's multiphysics evolution trajectory based on the fault spatial location coordinates in the transformer's comprehensive diagnostic descriptor, and generating a maintenance decision trigger command containing the fault remaining lifespan consumption rate based on the local evolution trajectory segments, includes: The process involves parsing each preliminary diagnostic descriptor in the transformer comprehensive diagnostic descriptor, extracting the fault spatial location coordinates and associated fault type semantic tags corresponding to each preliminary diagnostic descriptor, and converting the fault spatial location coordinates into the corresponding target spatial region range in the transformer's internal three-dimensional spatial coordinate system. Using the range of each target spatial region as a spatial filtering condition, search for all physical field distortion event moments in the multi-physics field evolution trajectory of the transformer where the spatial position falls within the corresponding target spatial region range. Arrange the searched physical field distortion event moments in chronological order to obtain the local evolution trajectory fragment corresponding to each preliminary diagnostic descriptor. For each local evolution trajectory segment, a time series analysis of the physical field distortion intensity is performed. For each type of physical field contained in the local evolution trajectory segment, the intensity growth rate of the distortion intensity over time is calculated. The intensity growth rates of each type of physical field are normalized and fused to obtain a comprehensive evolution activity index. Local evolution trajectory segments whose comprehensive evolution activity index exceeds a preset threshold are marked as active evolution trajectory segments. For each active evolution trajectory segment, the distortion intensity values ​​of various physical field types contained therein are extracted at the first and last distortion events. The intensity values ​​of each physical field are dimensionless and weighted and fused to generate a comprehensive current fault severity benchmark value and a comprehensive fault initial severity benchmark value. The difference between the comprehensive current fault severity benchmark value and the comprehensive fault initial severity benchmark value for each active evolution trajectory segment is calculated as the cumulative increment of fault severity. Divide the cumulative increment of fault severity corresponding to each active evolution trajectory segment by the total time span of that active evolution trajectory segment to obtain the average growth rate of fault severity corresponding to each active evolution trajectory segment. Multiply the reciprocal of the average growth rate of fault severity by the difference between the preset, dimensionless comprehensive fault severity threshold and the comprehensive current fault severity benchmark value to obtain the predicted value of the remaining lifetime consumption rate of fault corresponding to each active evolution trajectory segment. The corresponding maintenance decision urgency level is determined based on the numerical range to which each dimensionless remaining life of failure rate of failure prediction belongs. The fault type semantic label and fault spatial location coordinates associated with each maintenance decision urgency level and the corresponding active evolution trajectory segment are processed by instruction encoding to generate a maintenance decision trigger instruction containing multiple dimensionless fault remaining life consumption rate prediction values, multiple urgency levels, multiple fault type semantic labels, and multiple fault spatial location coordinates, and the maintenance decision trigger instruction is sent to the transformer online monitoring master station system.

7. The transformer fault feature fusion diagnosis method based on multimodal time series data reconstruction according to claim 1, characterized in that, The pre-constructed transformer operation status reconstruction network is obtained through the following network training method: A set of historical transformer fault case samples is obtained. The set of historical transformer fault case samples includes multiple sets of historical transformer fault case samples with known real fault energy injection time and sample fault type labels. The historical transformer fault case samples include time-series waveform samples of dissolved gas in oil, time-series waveform samples of tank wall vibration, and time-series waveform samples of bushing thermal map. Perform gas concentration curve inflection point detection processing on each group of dissolved gas time series waveform samples in the oil to obtain the set of historical gas concentration inflection point times and the associated set of historical characteristic gas component proportion vectors corresponding to each group of dissolved gas time series waveform samples in the oil. Perform vibration waveform envelope peak point capture processing on each group of the box wall vibration time sequence waveform samples to obtain the set of historical vibration amplitude peak times and the associated set of historical vibration main frequency offset direction identifiers corresponding to each group of box wall vibration time sequence waveform samples; Perform temperature gradient field calculation processing on each set of the casing heat map time series waveform samples to obtain the set of historical temperature field distortion moments and the associated set of historical temperature anomaly region spatial coordinates corresponding to each set of casing heat map time series waveform samples. Initialize the network parameters of the transformer operation status reconstruction network, which includes a time correlation mapping subnetwork and a physical field coupling verification subnetwork; The set of historical gas concentration inflection point times, the set of historical vibration amplitude peak times, and the set of historical temperature field distortion times are input into the time-related mapping subnetwork. The time-related mapping subnetwork outputs the predicted candidate fault energy injection time, the predicted accompanying vibration response time, and the predicted accompanying thermal field response time corresponding to the transformer's historical fault case sample. The historical characteristic gas component ratio 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 set of historical temperature anomaly region spatial coordinates associated with the predicted accompanying thermal field response time are input into the physical field coupling verification subnetwork. The physical field coupling verification subnetwork outputs the first verification Boolean value, the second verification Boolean value, and the third verification Boolean value. Based on the logical AND operation results of the predicted first verification Boolean value, the predicted second verification Boolean value, and the predicted third verification Boolean value, the predicted fault energy injection point confirmation signal output by the transformer operation status reconstruction network is obtained. Calculate the first time alignment loss between the predicted candidate fault energy injection time and the known actual fault energy injection time of the transformer's historical fault case sample; Calculate the first Boolean value matching loss between the predicted first verification Boolean value and the theoretical first verification Boolean value corresponding to the fault type label of the sample; Calculate the second Boolean value matching loss between the predicted second verification Boolean value and the theoretical second verification Boolean value corresponding to the fault type label of the sample; Calculate the third Boolean value matching loss between the predicted third verification Boolean value and the theoretical third verification Boolean value corresponding to the fault type label of the sample; The first time alignment loss, the first Boolean value matching loss, the second Boolean value matching loss and the third Boolean value matching loss are weighted and summed to obtain the total training loss value corresponding to the historical fault case sample of the transformer. The gradient backpropagation algorithm is used to calculate the gradient of the total training loss with respect to all network parameters of the transformer operation status reconstruction network; The network parameters of the transformer operation status reconstruction network are updated based on the calculated gradient. The process involves iterating through all samples in the transformer historical fault case sample set, repeatedly performing the step of detecting the inflection point of the gas concentration curve for each group of dissolved gas time-series waveform samples in the oil, and updating the network parameters of the transformer operation status reconstruction network based on the calculated gradient, until the total training loss value converges on the independent validation set, thus obtaining the pre-constructed transformer operation status reconstruction network after training.

8. The transformer fault feature fusion diagnosis method based on multimodal time series data reconstruction according to claim 2, characterized in that, The method further includes: a step of dynamic credibility assessment and knowledge base enhancement of diagnostic conclusions, performed after the maintenance decision trigger command is sent to the transformer online monitoring master station system. This step includes: Within the preset observation time window after the maintenance decision trigger command is sent, the timing waveforms of the dissolved gas in the transformer oil, the vibration of the transformer tank wall, and the thermal image of the bushing are continuously collected. The following steps are performed on the subsequent dissolved gas in oil time-series waveform, subsequent tank wall vibration time-series waveform, and subsequent bushing thermal image time-series waveform: inputting the dissolved gas in oil time-series waveform, tank wall vibration time-series waveform, and bushing thermal image time-series waveform into the pre-constructed transformer operation status reconstruction network for cross-modal physical field spatiotemporal alignment processing to generate the subsequent transformer multi-physics field evolution trajectory. The subsequent transformer multiphysics field evolution trajectory is processed by performing the topology source tracing processing of the fault feature clues on the transformer multiphysics field evolution trajectory, thereby generating a subsequent transformer global fault feature topology chain; The subsequent transformer global fault feature topology chain is compared and analyzed with the original transformer global fault feature topology chain on which the maintenance decision trigger command is based, and the topology similarity and node attribute offset are calculated. Based on the topological similarity and node attribute offset, determine whether the fault development path predicted in the maintenance decision triggering command is consistent with the subsequent actual observed fault evolution trend, and generate a diagnostic conclusion validity verification label. Extract the original transformer global fault feature topology chain and the original transformer comprehensive diagnostic descriptor corresponding to the current diagnostic case. Associate and store the original transformer global fault feature topology chain, the original transformer comprehensive diagnostic descriptor and the diagnostic conclusion validity verification label to form a diagnostic case record with verification label. If the diagnostic conclusion validity verification label indicates that the diagnostic conclusion is valid, then the original transformer global fault feature topology chain in the diagnostic case record with the verification label is converted into a topology embedding vector and added as a new positive sample to the fault mode standard topology embedding vector library of the fault mode semantic parsing network. If the diagnostic conclusion validity verification label indicates that the diagnostic conclusion is questionable, then the diagnostic process backtracking analysis is triggered. The diagnostic process backtracking analysis includes verifying the time alignment accuracy of physical field distortion events in the original transformer multi-physics field evolution trajectory output by the transformer operation status reconstruction network, and reviewing the matching threshold setting of the fault mode semantic parsing network. A diagnostic uncertainty analysis report is generated based on the results of the diagnostic process backtracking analysis. The diagnostic uncertainty analysis report is used to guide the adjustment of the preset gas response threshold, preset vibration response threshold, or preset thermal field response threshold. The validity verification tag of the diagnostic conclusion, the knowledge base update status, and the diagnostic uncertainty analysis report are integrated to generate a credibility assessment and knowledge enhancement report for this diagnosis, and the credibility assessment and knowledge enhancement report is sent to the transformer online monitoring master station system.

9. A transformer fault feature fusion diagnostic system based on multimodal time-series data reconstruction, characterized in that, include: processor; A machine-readable storage medium for storing machine-executable instructions of the processor; The processor is configured to execute the transformer fault feature fusion diagnosis method based on multimodal time-series data reconstruction as described in any one of claims 1 to 8 by executing the machine-executable instructions.

Images