Method for sensing abnormity of intelligent transformer based on three-dimensional magnetic flux reconstruction

Abstract

The invention provides a three-dimensional magnetic flux reconstruction intelligent transformer anomaly sensing method, which comprises the following steps: arranging a three-dimensional magnetic flux density distribution sensing array in a transformer iron core, collecting magnetic flux density multipoint spatio-temporal data in real time, and generating an original sampling matrix; according to the method, a three-dimensional magnetic flux field distribution diagram is constructed by adopting space fitting and interpolation processing and is used for representing magnetic flux vector direction and local disturbance characteristics; carrying out difference analysis on the magnetic flux field maps in different time periods, extracting magnetic flux disturbance characteristics and generating a physical characteristic parameter set; synchronously collected vibration data and the parameter set are compared and analyzed so as to eliminate external disturbance interference and obtain an abnormal response signal after interference suppression; and based on physical characteristic parameters in the signal, executing a multi-threshold combination judgment logic, identifying an abnormal evolution trend and judging a corresponding fault type, finally automatically generating a load adjustment or cooling control instruction according to an identification result, and executing intelligent self-adaptive adjustment on the transformer through a control interface.

Description

Technical Field

[0001] This invention relates to the field of power equipment condition monitoring and intelligent control technology, and in particular to an intelligent transformer anomaly sensing method based on three-dimensional magnetic flux reconstruction. Background Technology

[0002] In existing technologies, monitoring the operating status of transformers typically relies on equipment such as temperature sensors, current transformers, voltage transformers, and gas analyzers. The health status of the transformer is indirectly determined by detecting parameters such as oil temperature, load current, voltage fluctuations, and dissolved gases in the oil. Some solutions use vibration sensors or partial discharge detection devices to assist in analyzing structural anomalies or electrical breakdowns. Other research attempts to identify fault trends by monitoring magnetic field changes using leakage flux sensors. These monitoring methods are mostly based on single-point information, are easily affected by load fluctuations and environmental interference, and primarily employ a threshold-triggered passive response mode.

[0003] Existing technologies have significant limitations in dealing with early, weak anomaly signals. On the one hand, fault diagnosis based on characterizing physical quantities (such as temperature, gas concentration, and vibration amplitude) often exhibits a response lag, making it difficult to reflect early evolution processes such as internal inter-turn short circuits, local aging, or discharges in a timely manner. On the other hand, existing methods lack real-time observation means of changes in the internal structure of the magnetic flux field, cannot effectively obtain the complete distribution characteristics of the magnetic flux vector in space, and have not formed an effective interference suppression mechanism or multi-factor fusion judgment logic, leading to frequent misjudgments and omissions, affecting the reliability of the monitoring system and the timeliness of the control response.

[0004] In view of the above-mentioned shortcomings in the existing technology, it is necessary to propose a new technical solution to improve the sensitivity, accuracy and intelligent response capability of transformer anomaly detection. Summary of the Invention

[0005] This application provides a three-dimensional magnetic flux reconstruction-based intelligent transformer anomaly sensing method to improve the safety, stability, and autonomous response capability of transformer operation.

[0006] This application provides a method for intelligent transformer anomaly detection based on three-dimensional magnetic flux reconstruction, including: A three-dimensional magnetic flux density distribution sensor array is deployed inside the transformer core to collect multi-point spatiotemporal data of magnetic flux density in real time during transformer operation and generate the original magnetic flux density sampling matrix. The original magnetic flux density sampling matrix is ​​spatially fitted and interpolated to obtain a three-dimensional magnetic flux field distribution map for the current operating cycle. The three-dimensional magnetic flux field distribution map is used to characterize the directionality and local perturbation features of the magnetic flux vector. The differences in the three-dimensional magnetic flux field distribution map corresponding to the continuous operation cycle are analyzed to extract magnetic flux disturbance features, including magnetic flux disturbance amplitude, magnetic flux change slope and local abnormal clustering areas, and a set of physical feature parameters is generated accordingly. Vibration data of the transformer body are collected synchronously, and the vibration data is compared and analyzed with the set of physical characteristic parameters. Interference components caused by external disturbances are eliminated, and abnormal response signals after interference suppression are generated. Based on the physical characteristic parameters in the abnormal response signal, a multi-threshold combination judgment logic is executed to identify the abnormal evolution trend and determine whether the abnormal response signal corresponds to a preset fault type, including internal insulation aging, winding inter-turn short circuit or partial discharge. Based on the identified fault type, the corresponding handling strategy rules are invoked to generate a set of control instructions, including load adjustment commands or local cooling control commands, and sent to the target execution component through the transformer control interface to achieve intelligent adaptive adjustment and control of the transformer.

[0007] The beneficial effects of the technical solution provided in this application include: (1) By deploying a three-dimensional magnetic flux density distribution sensor array inside the transformer core, high-precision real-time reconstruction of the magnetic flux vector spatial distribution is achieved. Compared with the traditional single-point monitoring method, it can comprehensively capture small local abnormal magnetic flux disturbances, significantly improving the spatial resolution and timeliness of fault identification. (2) By adopting a synchronous comparison mechanism between magnetic flux disturbance characteristics and vibration data, non-fault signals caused by external mechanical disturbances or environmental interference are effectively eliminated, improving the accuracy of abnormal perception, reducing the false alarm rate, and enhancing the stability of the system in complex operating environments. (3) Through the correlation analysis of multi-threshold combination judgment logic and physical characteristic parameters, it is possible to predict the abnormal evolution trend in advance and has the ability to identify different types of faults (such as inter-turn short circuit, partial discharge, insulation aging, etc.), meeting the requirements of refined condition assessment. (4) By combining the automatically generated control instruction set with the transformer control interface, a closed-loop response of load regulation and local cooling is realized, enabling the system to have intelligent adaptive adjustment capabilities, which helps to delay fault development, reduce operation and maintenance risks, and improve the operating safety and service life of the transformer. Attached Figure Description

[0008] Figure 1 This is a flowchart of an intelligent transformer anomaly sensing method for three-dimensional magnetic flux reconstruction provided in the first embodiment of this application. Detailed Implementation

[0009] Many specific details are set forth in the following description to provide a full understanding of this application. However, this application can be implemented in many other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of this application; therefore, this application is not limited to the specific embodiments disclosed below.

[0010] The first embodiment of this application provides a method for intelligent transformer anomaly detection based on three-dimensional magnetic flux reconstruction. Please refer to... Figure 1 This figure is a schematic diagram of the first embodiment of this application. The following is in conjunction with... Figure 1 The first embodiment of this application provides a detailed description of an intelligent transformer anomaly sensing method based on three-dimensional magnetic flux reconstruction.

[0011] Step S101: A three-dimensional magnetic flux density distribution sensor array is deployed inside the transformer core to collect multi-point spatiotemporal data of magnetic flux density during transformer operation in real time and generate the original magnetic flux density sampling matrix.

[0012] In the process of implementing the intelligent transformer anomaly sensing method for three-dimensional magnetic flux reconstruction of the present invention, step S101 is the foundation and starting point of the whole method. Its core lies in reasonably arranging a three-dimensional magnetic flux density distribution sensing array inside the transformer core, and based on the array, realizing high spatiotemporal resolution data acquisition of magnetic flux density information of the transformer during operation, thereby constructing the original magnetic flux density sampling matrix.

[0013] First, the selected magnetic flux density sensor should possess high sensitivity, low noise, wide bandwidth, and good heat resistance. Preferably, a three-dimensional vector magnetic sensor device based on the Hall effect, fluxgate magnetometry, or miniature integrated magnetoresistive principles should be used. Each sensor node should be able to simultaneously detect magnetic flux density components in at least three spatially orthogonal directions (X, Y, Z) to completely record the directional changes of the magnetic flux vector. The sensor node must be encapsulated in a non-magnetic structural material with high electrical insulation and thermal stability to avoid magnetic flux sensing deviations due to structural interference or electromagnetic shielding.

[0014] Regarding the deployment method, the spatial arrangement of the sensor array should be optimized based on the structural characteristics of the transformer core. For example, for a three-phase stacked core structure, sensor arrays can be arranged along the three basic components of each phase core (core column, yoke, and oblique joint), while layering deployment planes at different heights in the vertical direction. Each deployment plane should at least cover the junction area of ​​the core column and the yoke, thus forming a three-dimensional sampling network. The spatial spacing between each sensor should be as small as possible without affecting the normal insulation structure and thermal field layout of the transformer to improve the magnetic flux reconstruction accuracy. For example, an equidistant deployment strategy of 20~50 mm can be used to form a typical cubic or hexahedral magnetic flux sampling area.

[0015] For signal acquisition, each sensor node should be equipped with a high-speed sampling module, and the sampled signal should be transmitted to the integrated processing unit via shielded twisted-pair cable or optical fiber. All sensor nodes must be time-synchronized with the main acquisition and control system. GPS time synchronization or an industrial Ethernet-based synchronization mechanism (such as IEEE 1588) can be used to ensure that different nodes complete unified time-base sampling within milliseconds or even sub-milliseconds, thereby ensuring the timeliness and accuracy of subsequent data fitting and reconstruction.

[0016] Within each sampling period, each sensor node encapsulates the real-time acquired three-dimensional vector magnetic flux density values ​​into structured data packets in a unified format. Each data packet includes a sensor identifier, spatial coordinates, a timestamp, and three-axis magnetic flux density values ​​(Bx, By, Bz). All data packets are aggregated to form the original magnetic flux density sampling matrix for that period. This matrix, indexed by spatial location and using magnetic flux density vectors as values, constructs a multi-dimensional spatiotemporal data volume describing the three-dimensional magnetic field state inside the iron core.

[0017] It should be noted that the design of the entire step S101 must take into account the electromagnetic compatibility of the sensor, safe insulation distance, embedded installation process of the sensing element, and long-term adaptability to the operating environment of the transformer body (high temperature, high voltage, strong electromagnetic field). After the sensor is deployed, calibration and compensation should be performed, including sensor zero drift correction, temperature drift model fitting, and systematic deviation correction, to ensure the stability and usability of the original sampling data.

[0018] In summary, by establishing a high-density, synchronized, three-dimensional vectorized magnetic flux density distribution sensing array inside the iron core, and combining it with a strict data synchronization and sampling strategy, a basic sampling foundation for subsequent three-dimensional magnetic flux field reconstruction can be constructed. This provides accurate, stable, and complete basic sensing data support for the entire anomaly sensing method, ensuring that the system has the ability to deeply understand the dynamic behavior of the transformer's magnetic field.

[0019] To further illustrate the construction process of the original magnetic flux density sampling matrix, a specific example is provided below: Taking a three-phase transformer with a rated capacity of 1000 kVA as the object, a three-dimensional magnetic flux density distribution sensing array containing 4 layers × 4 rows × 4 columns is deployed in the area where the core column and yoke meet, with a total of 64 three-dimensional vector magnetic flux sensors. Each sensor node corresponds to a unique spatial coordinate index (i, j, k), where i represents the layer height, j represents the row position on the core column cross-section, and k represents the column position. Each node can simultaneously output magnetic flux density values ​​(Bx, By, Bz) in three directions, in mT (millitales). For example, at a certain sampling time... At time , the magnetic flux density values ​​collected by the sensor located at node (2,3,1) are Bx = 2.35 mT, By = -1.12 mT, and Bz = 0.87 mT, with corresponding timestamps of . = 2025-05-25 10:32:05.125. The outputs of all sensor nodes at this moment will be aggregated and recorded in the following structure: The sampling matrix entries are formatted as follows: [(i, j, k), , Bx(i,j,k), By(i,j,k), Bz(i,j,k)].

[0020] Where (i, j, k) represents the spatial position index of the current sensor in the three-dimensional magnetic flux density distribution sensing array, i represents the layer number of the sensor in the vertical direction (along the height direction of the core column); j represents the row number of the sensor in the horizontal direction (usually along the width direction of the core column or yoke); and k represents the column number of the sensor in the horizontal direction (usually along the depth direction of the core column or yoke). This triplet uniquely determines the physical position of the sensor in the core structure and is an important reference coordinate for spatial interpolation and magnetic flux reconstruction.

[0021] This indicates the timestamp of the current sampled data, i.e., the actual time point when this magnetic flux density data was collected. This time information is used for subsequent time series analysis, comparison of continuous period differences, and extraction of abnormal evolution trends, ensuring the temporal consistency and traceability of the data.

[0022] Bx(i,j,k) represents the time of the sensor located at position (i,j,k). The component of the collected magnetic flux density in the X direction (e.g., along the longitudinal axis of the transformer core), measured in millitalas (mT), reflects the magnetic flux intensity at that point in that direction.

[0023] By(i,j,k) represents the time of the sensor located at position (i,j,k). The component of the collected magnetic flux density in the Y direction (e.g., along the width of the yoke), measured in millitalas (mT), reflects the magnetic flux intensity at that point in that direction.

[0024] Bz(i,j,k) represents the time of the sensor located at position (i,j,k). The component of the collected magnetic flux density in the Z direction (e.g., perpendicular to the cross-section of the iron core), measured in millitalas (mT), is the third component vector in the three-dimensional vector properties of the magnetic flux.

[0025] This structured data, uniformly input into the subsequent spatial fitting and magnetic flux field reconstruction processing flow in matrix form, constitutes a complete original magnetic flux density sampling matrix. This not only preserves the directional information of the magnetic flux vectors at each node but also ensures the continuity of the reconstructed magnetic flux field and the ability to sensitively capture local changes through dense spatial point placement, thus providing a solid perceptual foundation for the entire anomaly identification and control chain. From a data structure perspective, this matrix can also be viewed as a four-dimensional tensor of size [I × J × K × 3], where the first three dimensions are spatial dimensions, and the fourth dimension represents the three directional components of the magnetic flux density.

[0026] Furthermore, the method of deploying a three-dimensional magnetic flux density distribution sensor array inside the transformer core to collect multi-point spatiotemporal data of magnetic flux density during transformer operation in real time and generate an original magnetic flux density sampling matrix includes: A three-dimensional magnetic flux density distribution sensing array with an equilateral cubic structure is constructed inside the transformer core. A set of triaxial magnetic flux density sensor nodes is arranged at the center of each cubic unit to obtain the component information of the magnetic flux density at the center point in three spatial directions, forming a basic sampling point set. Based on the spatial gradient value of magnetic flux density at each point in the basic sampling point set, the region where the magnetic flux density gradient exceeds the preset threshold is identified. The equilateral cube structure is locally refined in this region, and a densified sampling point set is formed by adding a triaxial magnetic flux density sensor node. The magnetic flux density three-component values ​​of each sampling point in the encrypted sampling point set are encapsulated in a unified format to construct a data structure with five dimensions, where the five dimensions correspond to sampling time, X spatial coordinate, Y spatial coordinate, Z spatial coordinate and magnetic flux density direction component, respectively, forming an uncompressed multidimensional magnetic flux density data set. Real-time online structural compression processing is performed on the uncompressed multidimensional magnetic flux density dataset. The compression operation is based on the spatial redundancy calculation between sampling points. The compressed magnetic flux density dataset is generated by retaining the complete data of the key regions of the magnetic flux density gradient and compressing the data of the regions with stable changes. The compressed magnetic flux density data set is time-aligned with the transformer body vibration data collected within the same sampling period. The alignment method is based on a set unified sampling period control signal with a time resolution of no more than 1 millisecond, so as to obtain a multi-source data group after time synchronization. Finally, the time-synchronized magnetic flux density data part is used as the original magnetic flux density sampling matrix for spatial fitting and interpolation processing of the original magnetic flux density sampling matrix.

[0027] To achieve high spatiotemporal resolution sensing of magnetic flux density information during transformer operation and provide a precise data foundation for subsequent flux reconstruction, anomaly identification, and intelligent control, a dynamically responsive three-dimensional magnetic flux density acquisition mechanism must be constructed inside the transformer core. This mechanism not only needs to cover the spatial structural features of the transformer core but also must possess high-density sensing capabilities for areas sensitive to magnetic flux changes, while simultaneously meeting stability, time synchronization, and rational data structure organization requirements under long-term operating conditions.

[0028] First, when constructing a three-dimensional magnetic flux density distribution sensing array inside the transformer core, an equilateral cubic spatial structure should be preferred. This structure helps achieve uniformity in spatial sampling and analytical symmetry of the three-axis magnetic flux field data. Specifically, the core is divided into multiple non-overlapping equilateral cubic units based on its dimensional parameters and thermo-electric-magnetic environment characteristics. A set of triaxial magnetic flux density sensor nodes is placed at the geometric center of each cubic unit. These sensor nodes should simultaneously detect the vector components of magnetic flux density in the X, Y, and Z directions. This can be achieved using integrated Hall sensors, micro-magnetic reluctance arrays, or fluxgate elements, requiring linear response characteristics and low-temperature drift stability within typical operating magnetic flux densities (usually in the range of 0.3T to 1.8T). Sensor installation should be integrated with insulation design, preferably embedded in the interlayer insulation sheet structure of the core or attached to the core surface, and connected to the acquisition unit via flexible PCB traces to ensure no interference with the magnetic circuit integrity of the core itself.

[0029] After constructing the basic sampling point set, the system performs spatial gradient analysis on the magnetic flux density vector data acquired in each sampling period. This analysis requires calculation based on the local gradient operator of the vector field, that is, solving for the partial derivatives of the magnetic flux density in each direction in space through finite difference or local tensor analysis methods, thereby obtaining the rate of change of magnetic flux density in the region where each sensing node is located. If the gradient magnitude of a certain region exceeds a preset threshold (e.g., exceeding 150% of the average gradient of the surrounding area), it is considered a region with active magnetic flux disturbance. At this time, the system will refine the original equilateral cubic mesh in this local region. The refinement method is to divide the original unit into eight sub-cubic units in equal proportion, and re-deploy the triaxial magnetic flux density sensor node at the geometric center of each new sub-unit, dynamically forming a "densified sampling point set". Through this mechanism, the system realizes the adaptive sensing capability of magnetic flux field structure, which can form a high-density monitoring network in areas with drastic magnetic flux changes, while maintaining a low-redundancy sparse deployment in areas with stable magnetic flux, thereby optimizing the balance between acquisition resources and data quality.

[0030] After data acquisition, the output data from all sampling points must be entered into the sampling matrix according to a unified structure and encapsulation format. The output of each sensor node includes: a sampling timestamp, spatial coordinates (X, Y, Z), and three directional components of the magnetic flux density (Bx, By, Bz). To support subsequent spatial reconstruction and temporal analysis, this dataset is organized into a five-dimensional tensor structure: time, X-coordinate, Y-coordinate, Z-coordinate, and directional components. The time dimension can be linearly indexed according to the sampling frequency (e.g., 200 frames per second), the three spatial dimensions use fixed coordinate numbers or relative index encoding, and the directional dimension is fixed at three directional channels. This multidimensional data structure enables continuous temporal tracking of the three-dimensional vector magnetic flux field and supports rapid retrieval and local area focusing analysis.

[0031] Because a large number of spatially uniform regions with slow magnetic flux changes exist during sampling, directly storing the five-dimensional tensor would result in a massive and redundant data volume. Therefore, after data encapsulation, the system performs real-time online structural compression on the uncompressed tensor. The compression method is based on spatial redundancy assessment between sampling points, such as using a similarity evaluation index based on the angle between adjacent point vectors to identify sub-regions with redundancy exceeding a set threshold, and then performing vector quantization or low-rank reconstruction on them. The compression operation strictly preserves the complete data of regions with significant magnetic flux gradients, ensuring that magnetic flux distortion and perturbation features are not diluted, while significantly reducing the data bandwidth occupied in stable regions. The output is a compressed magnetic flux density dataset with efficient transmission and storage.

[0032] Next, to achieve the function of co-comparing magnetic flux disturbance characteristics with mechanical disturbances in this invention, the magnetic flux density data and vibration data need to be precisely aligned in time. Vibration data is collected by high-sensitivity triaxial accelerometers deployed on the transformer body casing, winding supports, or cooling structures. Their sampling frequency is usually different from that of the magnetic flux data, therefore alignment must be performed using a unified sampling control signal. This invention proposes using a unified clock source (e.g., a GPS synchronization signal or the IEEE 1588 protocol) to drive the magnetic flux sampling and vibration sampling systems, setting a unified sampling period control signal (e.g., one frame every 5 milliseconds), and embedding timestamps in both the magnetic flux and vibration data. During the alignment process, the system will extract the data frame closest to that moment within each sampling period window, ensuring that the time deviation between the two types of data does not exceed 1 millisecond. Through this multi-source sampling alignment mechanism, the system can accurately correlate magnetic flux disturbances with the corresponding structural vibration responses within each sampling period, constructing a "time-synchronized multi-source data set."

[0033] Finally, the magnetic flux density component is extracted from the time-synchronized multi-source data set and output as the final content of the "raw magnetic flux density sampling matrix". This raw sampling matrix not only features adjustable spatial resolution, a standardized data organization structure, and high sensitivity to magnetic flux change gradients, but also possesses the functional foundation for coupling with subsequent vibration signal comparison, anomaly trend identification, and control response. In subsequent magnetic flux field spatial fitting and interpolation processing, this matrix will serve as a direct input for the numerical reconstruction of the three-dimensional magnetic flux field, thereby enabling the entire system to possess a fine-grained perception capability of the magnetic flux dynamics inside the transformer.

[0034] The above process covers the entire process from sensor array construction, gradient-driven adaptive encryption, structured data encapsulation, redundancy compression processing, time alignment mechanism establishment, to the formation of the original sampling matrix. Each step has clear input / output interfaces and quantifiable control parameters, ensuring effective deployment and implementation on power transformers of different models and structural arrangements. Through this approach, the present invention solves the problems of limited flux monitoring dimensions, response lag, high spatial redundancy, and poor interference coupling in existing technologies, establishing a fundamental data support framework for state perception of intelligent evolutionary transformers.

[0035] Step S102: Perform spatial fitting and interpolation processing on the original magnetic flux density sampling matrix to obtain a three-dimensional magnetic flux field distribution map of the current operating cycle, wherein the three-dimensional magnetic flux field distribution map is used to characterize the directionality and local perturbation characteristics of the magnetic flux vector.

[0036] In the intelligent transformer anomaly sensing method based on three-dimensional magnetic flux reconstruction described in this invention, step S102 is a crucial step connecting the original sensing data with subsequent anomaly feature extraction and analysis. It aims to reconstruct a continuous and physically consistent three-dimensional magnetic flux field distribution map from the original magnetic flux density sampling matrix collected at multiple locations within the iron core, through precise spatial fitting and interpolation. This step not only restores the distribution pattern of the magnetic flux vector within the iron core in three-dimensional space but also preserves subtle perturbation features, providing a complete and reliable magnetic flux field basis for subsequent fault location and trend identification.

[0037] Specifically, in step S101, a multi-dimensional data volume has been obtained, recording the magnetic flux density vector values ​​collected by multiple sensing nodes within a certain sampling period. This data forms a structured sampling matrix based on the coordinate indices of the nodes in three-dimensional space. Due to the limited number of sensors, this sampling matrix is ​​essentially a sparse spatial point cloud, which cannot directly represent the continuous distribution of the magnetic field inside the iron core. Therefore, in step S102, this sparse point cloud data must be spatially reconstructed to complete the magnetic flux vector distribution in the unmeasured areas in order to construct a complete three-dimensional magnetic flux field map.

[0038] When implementing this step, a unified three-dimensional spatial coordinate system corresponding to the transformer core structure should first be established, typically a Cartesian coordinate system, with the transformer's central axis as the axis of symmetry or reference plane. Data points from each sensing node are mapped into this coordinate system, preserving their spatial position and corresponding three-component magnetic flux density values ​​(Bx, By, Bz). Based on this, independent spatial fitting processing is performed on each of the three-axis components. Preferred fitting methods include cubic spline interpolation, radial basis function interpolation, inverse distance weighted (IDW), or a Gaussian process-based spatial regression algorithm to achieve smooth extension of the magnetic flux density in each direction. For each unmeasured spatial point, the interpolation function calculates an estimated value based on the spatial distance and magnetic flux direction gradient of several surrounding known measurement points, thereby generating a spatially continuous magnetic flux vector field.

[0039] To ensure the physical validity of the fitting results, the fitting accuracy should be constrained and controlled. Specifically, this includes: setting interpolation error limits (e.g., less than 5%) to prevent local overfitting; using local weighting methods to enhance boundary preservation in regions of abrupt changes in magnetic flux density; and using higher-order fitting in regions of continuous magnetic flux to enhance the overall smoothness of the field map. Furthermore, during data preprocessing, outliers or invalid data points in the original sampling matrix must be identified and removed to prevent the introduction of reconstruction errors. It is recommended to combine spatial gradient analysis and multivariate statistical methods such as Mahalanobis distance to filter out distorted data.

[0040] After reconstruction, the three-dimensional magnetic flux field distribution map can be stored in the form of a three-dimensional tensor, with the data structure F(x,y,z) = [Bx(x,y,z), By(x,y,z), Bz(x,y,z)]. Each spatial coordinate point (x,y,z) contains a magnetic flux vector, which reflects the directionality and intensity of the magnetic flux at that location within that period. This magnetic flux field not only possesses continuity and smoothness in its structure but also retains important physical characteristics such as local magnetic flux perturbations, distortion concentrations, and gradient abrupt changes in actual operation.

[0041] The final generated three-dimensional magnetic flux field distribution map can be visualized through slice projection, streamline diagrams, vector diagrams, etc., to assist in subsequent steps to identify local anomaly features. Especially when local inter-turn short circuits, core saturation, or partial discharge precursors occur in the transformer, the magnetic flux field map will exhibit specific disturbance modes or local clustering effects. These phenomena can only be clearly captured by high-precision field map reconstruction. The three-dimensional magnetic flux field fitting and interpolation processing completed in step S102 is one of the key technical foundations for the early anomaly detection and intelligent decision-making of the method of this invention. Its accuracy and stability directly determine the reliability of subsequent diagnostic logic and the timeliness of system response.

[0042] Furthermore, the step of performing spatial fitting and interpolation processing on the original magnetic flux density sampling matrix to obtain a three-dimensional magnetic flux field distribution map for the current operating cycle includes: The magnetic flux density multi-point spatiotemporal data contained in the original magnetic flux density sampling matrix are spatially divided according to X spatial coordinates, Y spatial coordinates and Z spatial coordinates to form multiple spatial sub-regions, and the three-component data of magnetic flux density in each spatial sub-region are extracted. For each spatial sub-region, the magnetic flux density three-component data are subjected to joint fitting processing. The joint fitting processing includes performing a weighted superposition operation of cubic spline interpolation and inverse distance weighted interpolation on the magnetic flux density three-component data respectively, so as to simultaneously ensure the continuity of derivatives in the boundary region and the fitting stability in the central region, and generate the initial fitting result of three-dimensional magnetic flux density in the spatial sub-region. Based on the initial fitting results of the three-dimensional magnetic flux density, the divergence value of the three components of magnetic flux density at each spatial coordinate point is calculated, the magnetic flux conservation constraint condition of the three-dimensional magnetic flux field is established, the spatial location points with non-zero divergence are identified as fitting error points, and the statistical results of magnetic flux density residuals containing the coordinates of each error point and its corresponding error value are output. Based on the statistical results of the magnetic flux density residuals, for spatial locations where the residuals exceed a preset threshold, the spatiotemporal data of the neighboring magnetic flux density multi-points in the original magnetic flux density sampling matrix are called in the corresponding spatial sub-region. Local data augmentation is performed, and joint fitting processing of cubic spline interpolation and inverse distance weighted interpolation is performed again. Finally, the three-dimensional magnetic flux field distribution map of the current running cycle is output, which is used to perform difference analysis on the three-dimensional magnetic flux field distribution maps corresponding to continuous running cycles.

[0043] To achieve high-fidelity three-dimensional reconstruction of the internal magnetic flux behavior of a transformer, after obtaining the original magnetic flux density sampling matrix, the irregular spatial distribution formed by the sampling points must be uniformly organized, and a high-precision fitting algorithm must be used to recover the continuous magnetic flux vector field of the current operating cycle. This process not only requires the fitting results to have geometric continuity and smoothness, but also to satisfy the fundamental constraint of magnetic flux conservation in electromagnetic physics, and to be able to perform residual feedback and adaptive adjustment of the fitting accuracy, thereby providing stable and reliable data support for subsequent anomaly extraction and trend analysis.

[0044] Before performing spatial fitting and interpolation, the spatial data in the original magnetic flux density sampling matrix must first be structured. Since sensor deployment is typically non-uniform, the distribution density of the original data points in three-dimensional space is inconsistent. Therefore, it is necessary to construct a unified three-dimensional spatial coordinate system according to the geometry of the transformer core. This coordinate system is usually a Cartesian coordinate system, with the X, Y, and Z axes corresponding to the longitudinal, transverse, and layer height directions of the core, respectively. Based on the coordinate range and sampling point distribution density, the entire core space is divided into several spatial sub-regions, for example, the core volume is divided into several regular hexahedral voxels. Each spatial sub-region contains several magnetic flux density sampling points, forming a local magnetic flux vector point set. Then, the three components of the magnetic flux density data in each spatial sub-region are extracted from the original magnetic flux density sampling matrix, i.e., the magnetic flux density values ​​of each point in the X, Y, and Z directions. These data will serve as input for local fitting processing.

[0045] After spatial partitioning and data allocation, three-dimensional magnetic flux density fitting is performed for each sub-region. Since the magnetic flux distribution inside a transformer often includes both large-scale stable changes and localized violent disturbances, this invention does not employ a single interpolation method, but rather a weighted joint algorithm combining cubic spline interpolation and inverse distance weighted interpolation. Specifically, in each sub-region, the magnetic flux density components in the X, Y, and Z directions are fitted separately. Cubic spline interpolation has good mathematical continuity and differentiability, making it suitable for function construction at boundary changes, ensuring a smooth fitted surface while avoiding oscillations. Inverse distance weighted interpolation, on the other hand, is suitable for regions with uneven data point density distribution and concentrated local changes. It can allocate fitting weights according to the distance between sampling points, providing a stronger approximation capability for the central region.

[0046] To achieve joint fitting of the two interpolation methods, a weighting function is set for each directional component. A local gradient-driven weighting strategy is typically employed, such as increasing the spline interpolation weight in regions with high gradients and increasing the proportion of inverse distance weighting in regions with low gradients. The results of the two fitting methods are then weighted and superimposed to form the initial 3D magnetic flux density fitting result for this sub-region. This initial fitting result possesses spatial continuity globally and preserves key magnetic flux perturbation characteristics locally, serving as an important foundation for subsequent physical consistency verification.

[0047] After obtaining the fitted value of the three-dimensional magnetic flux density at each spatial point, the system will enter the accuracy verification stage based on physical constraints. According to the requirements of Maxwell's equations for the static magnetic field, the magnetic flux vector field in the source-free region should satisfy the condition of zero divergence, i.e. Therefore, in this invention, partial derivative operations are performed on the fitted magnetic flux density components Bx, By, and Bz at each spatial point to calculate their spatial divergence value. The partial derivatives in each direction can be calculated using the central difference method or a higher-order finite difference method, and the divergence at that point is obtained by weighted summation. ; The divergence of a vector field B represents the degree to which the magnetic flux density "diversifies" at a given point. In electromagnetism, it describes whether magnetic field lines "originate" or "converge" at a particular point.

[0048] If the absolute value of the divergence at a certain spatial point is greater than the preset tolerance (e.g.) If the magnetic flux density is T / m, then the point is considered a fitting error point that violates the law of flux conservation. The system will record the spatial coordinates of all error points and their corresponding divergence residual values, and output a residual statistical result matrix containing the position, direction, and error amount. This matrix not only identifies local fitting deviations but also provides a basis for subsequent data correction.

[0049] Subsequently, the system performs local adaptive data augmentation based on the aforementioned magnetic flux density residual statistics. Within regions where the error exceeds a threshold, the system searches for neighboring sampling points in the original magnetic flux density sampling matrix to expand the data support range and enhance the local density of the fitted dataset. Based on the augmented dataset, the system re-fits within this sub-region. At this point, the joint interpolation method is still used, and the weighting parameters are appropriately adjusted to improve fitting rigidity and suppress residual diffusion. This process can be iterative; if the augmented fitting result still fails to meet the divergence constraint, the data augmentation radius can be further expanded until the fitting accuracy meets the requirements or the maximum number of iterations is reached. Through this mechanism, this invention introduces a "physical consistency" constraint into the fitting result, overcoming the shortcomings of previous methods that only pursued mathematical fitting accuracy while neglecting electromagnetic constraints.

[0050] After all sub-regions have been fitted and verified against conservation constraints, the system integrates all local fitting results into a unified three-dimensional magnetic flux field data structure. This data structure is organized in the form of a regular voxel grid, with each grid cell recording its center coordinates and three-dimensional magnetic flux vector components, forming a complete three-dimensional magnetic flux field distribution map for the current operating cycle. This distribution map can be further used for tasks such as three-dimensional visualization, magnetic flux streamline tracing, and local disturbance aggregation analysis, and serves as input for subsequent cycle difference analysis and fault diagnosis.

[0051] Furthermore, to adapt to the dynamic changes in the magnetic flux distribution pattern of the transformer under different operating conditions, this invention also allows for the introduction of an adaptive adjustment mechanism for spatial fitting parameters during multi-cycle operation. For example, when a sustained increase in the amplitude of magnetic flux disturbance in a certain region is detected, the spatial sub-region division accuracy, sampling point density, or fitting weight coefficient of that region can be automatically adjusted to improve the local response capability of the fitting. In this way, the entire spatial fitting process has dynamic evolution capability, adapting to the transition process from normal to abnormal transformer operating status, and improving the sensitivity of anomaly detection.

[0052] The spatial fitting and interpolation method described in this invention not only integrates multiple interpolation techniques and error feedback mechanisms in its technical approach, but also considers feasibility and computational efficiency in engineering applications. All calculations can be implemented using a parallel computing structure based on tensor operations, making it suitable for embedded industrial processors, GPU arrays, or edge computing nodes. By deploying the aforementioned spatial fitting process at the software level, combined with the raw magnetic flux density sampling matrix data output by the hardware sensor array, the entire three-dimensional magnetic flux reconstruction module can form a complete and implementable closed-loop system.

[0053] In summary, from data structure preprocessing, fitting method selection, application of physical conservation constraints, and local error feedback control, to the final three-dimensional magnetic flux field output, this part of the technical solution of the present invention provides a high-resolution, high-accuracy, and physically consistent foundation for spatiotemporal modeling of transformer magnetic flux behavior. It enables visualized tracking and quantitative identification of the early evolution of abnormal magnetic flux behavior, laying a solid data foundation for subsequent intelligent anomaly diagnosis and adaptive control. The advantage of this method compared to existing technologies lies not only in the improved mathematical accuracy but also in its establishment of a consistent link between magnetic flux data and physical laws, which is a crucial core element in realizing intelligent sensing of transformers.

[0054] Step S103: Perform difference analysis on the three-dimensional magnetic flux field distribution map corresponding to the continuous operation cycle, extract magnetic flux disturbance features including magnetic flux disturbance amplitude, magnetic flux change slope and local abnormal clustering areas, and generate a set of physical feature parameters accordingly.

[0055] In the intelligent transformer anomaly sensing method for three-dimensional magnetic flux reconstruction described in this invention, the purpose of step S103 is to perform difference analysis based on the three-dimensional magnetic flux field distribution map obtained from continuous operation cycles, extract key disturbance features that reflect abnormal trends, and construct a set of physical feature parameters accordingly to provide support for subsequent anomaly identification and control decisions.

[0056] The first step in this process is to obtain the magnetic flux field spectra for two consecutive operating cycles, denoted as the _____. Period and the Period. Within each period, the magnetic flux vector distribution function is reconstructed in the three-dimensional space of the transformer core through spatial interpolation, and is expressed as: ; in: Indicates time period At that moment, located at the spatial coordinate point Magnetic flux vector; These represent the components of magnetic flux density in the three spatial directions, with units of millitalas (mT). It represents three-dimensional spatial coordinates, covering the entire iron core area.

[0057] Subsequently, based on the magnetic flux field distribution of the two periods, a magnetic flux change vector field is constructed: ; in, This represents the change in the magnetic flux vector at that spatial point between two periods; These represent the differences in magnetic flux density along the X, Y, and Z directions, respectively.

[0058] To quantitatively assess the intensity of the magnetic flux disturbance, the amplitude of the magnetic flux disturbance is calculated at each spatial point: ; in, It represents the magnitude change of the magnetic flux vector at that point between two periods, and is an important indicator for judging whether the magnetic flux has been significantly disturbed; all square terms and square root operations are performed at the same spatial point.

[0059] Next, by calculating the rate of change of the disturbance amplitude in multiple consecutive periods, the slope of the magnetic flux disturbance (i.e., the rate of disturbance evolution) can be obtained: ; in, This represents the rate of change of the magnetic flux disturbance amplitude at that point with time, in units of... ; Two cycles respectively and Medium disturbance amplitude; Two sampling periods and The time interval between them is in seconds.

[0060] In addition, it is necessary to identify potential localized anomalous clusters within regions of significant magnetic flux disturbance. To this end, an anomalous threshold for the disturbance amplitude can be set. To satisfy all: ; Spatial points are used as anomaly candidates and divided into one or more local anomaly clusters based on 3D connectivity or density clustering algorithms (such as DBSCAN). Each cluster will extract the following features: Center point coordinates ; Maximum disturbance amplitude ; Average rate of change of disturbance ; Aggregate volume (i.e., the number of grid cells contained in the anomaly region) Unit volume).

[0061] Ultimately, all the above features will be organized into a set of physical feature parameters, forming the following structural diagram: Node location: ; Disturbance amplitude: ; Perturbation rate of change: ; Cluster ID; Clustering characteristics: cluster center coordinates, maximum perturbation, average perturbation rate, volume, etc. These parameter sets will serve as the input basis for comparison with vibration data in subsequent step S104, and as the core criterion in fault trend identification and classification. Therefore, step S103 not only completes the numerical calculation of magnetic flux disturbance, but also realizes anomaly focusing processing at the spatial level, giving the present invention high anomaly sensitivity and spatial resolution.

[0062] Furthermore, the differential analysis of the three-dimensional magnetic flux field distribution map corresponding to the continuous operating cycle is performed to extract magnetic flux disturbance features, including magnetic flux disturbance amplitude, magnetic flux change slope, and local abnormal accumulation regions, and a set of physical feature parameters is generated accordingly, including: The magnetic flux density components of each corresponding spatial coordinate point in the three-dimensional magnetic flux field distribution map obtained in each continuous operation cycle are calculated point by point to obtain the magnetic flux disturbance vector of the spatial coordinate point. Based on the magnetic flux disturbance vector, a three-dimensional magnetic flux disturbance vector field containing all spatial coordinate points is constructed. Based on the three-dimensional magnetic flux disturbance vector field, the magnetic flux disturbance amplitude at each spatial coordinate point is calculated, and a time-domain sliding window is constructed based on the magnetic flux disturbance amplitude sequence of the spatial coordinate point in adjacent periods. The slope of the change of the disturbance amplitude within the window is extracted using a weighted least squares fitting method as the magnetic flux change slope value of the spatial coordinate point. The intermediate feature result set containing the magnetic flux disturbance amplitude and the magnetic flux change slope is output. A local tensor field is constructed for the three-dimensional magnetic flux perturbation vector field. Based on the gradient of the change in the perturbation vector direction of each spatial coordinate point in its neighborhood, a directional mutation factor is calculated. This directional mutation factor is used to characterize the discontinuity of the perturbation direction in three-dimensional space, serving as the directional criterion for subsequent spatial clustering analysis, and a spatial feature map with the directional mutation factor is output. In the spatial feature map with directional mutation factors, spatial coordinate points whose mutation factor values ​​exceed a preset clustering threshold are selected, and density clustering algorithm is executed to form local abnormal clustering regions. In each abnormal clustering region, the cluster mean and dispersion of the three features included—magnetic flux disturbance amplitude, magnetic flux change slope, and directional mutation factor—are statistically analyzed, and a cluster density index for the abnormal clustering region is constructed as the output content of the spatial structural anomaly features. Finally, the magnetic flux disturbance amplitude, magnetic flux change slope, directional mutation factor, and the cluster density index of the abnormal clustering region to which each spatial coordinate point belongs are uniformly integrated to generate a set of physical feature parameters.

[0063] To achieve early detection and classification of magnetic flux anomalies within transformers, it is necessary to further extract key disturbance features after completing high-precision reconstruction of the three-dimensional magnetic flux field. The core of this process is to perform differential analysis on the three-dimensional magnetic flux field distribution maps obtained over continuous operating cycles to reveal subtle trends in the spatial distribution of magnetic flux. These changes are then structured and quantified to form a set of physical characteristic parameters that can be used for subsequent vibration interference elimination and fault trend identification.

[0064] After reconstructing the three-dimensional magnetic flux field within a certain period, the system saves the magnetic flux density vector values ​​of all spatial coordinate points in that period. The magnetic flux density of each spatial point consists of three directional components, representing the intensity of the magnetic flux in the X, Y, and Z directions, respectively. Subsequently, after the end of the next consecutive period, the corresponding three-dimensional magnetic flux field distribution map is obtained. To extract the magnetic flux change of each spatial coordinate point between two periods, the system performs a one-to-one vector difference calculation on all spatial points in the two periods, that is, the difference is calculated on the magnetic flux density vector of the same coordinate point in the two consecutive periods to obtain the magnetic flux perturbation vector of that point in one sampling period. Based on the perturbation vectors of all spatial points, a three-dimensional magnetic flux perturbation vector field covering the entire internal region of the iron core can be constructed. This perturbation vector field directly reflects the magnetic flux evolution process and represents the direction and intensity of the magnetic field change at each spatial point in this period.

[0065] Based on the constructed magnetic flux perturbation vector field, the system further calculates the perturbation amplitude for each spatial coordinate point. The perturbation amplitude refers to the total change in magnetic flux density at that spatial point within the current period, i.e., the magnitude of the vector difference at that point. This quantifies the absolute strength of the perturbation. Next, to identify whether the perturbation exhibits a sustained growth trend, the system establishes a time window, typically a sequence of perturbation amplitudes encompassing the current period and several preceding periods. This time series is called a time-domain sliding window. For the data sequence within this window, the system uses weighted least squares to perform trend fitting to extract the slope of the perturbation amplitude change over time. This fitting method assigns higher weights to data closer to the current period, thereby improving the sensitivity to the current evolution trend. The slope obtained from the fitting is the magnetic flux change slope at that spatial coordinate point, a crucial indicator reflecting the speed of anomaly development.

[0066] At this point, the system has established two basic disturbance characteristics for each spatial coordinate point: the amplitude of the magnetic flux disturbance and the slope of the magnetic flux change. These two characteristics have been unified into an intermediate feature result set, providing a basis for subsequent structural analysis. However, in actual transformer operation, magnetic flux anomalies are often not only reflected in the numerical value of the disturbance intensity, but more importantly in their spatial distribution morphology and directional structure. Therefore, it is necessary to further extract spatial characteristics describing the directional changes of the magnetic flux disturbance.

[0067] To this end, the system constructs a local tensor field based on the existing three-dimensional magnetic flux perturbation vector field to characterize the gradient change of the perturbation vector direction in the three-dimensional neighborhood. Specifically, at each spatial coordinate point, the system analyzes the degree of directional change of the perturbation vector in its neighboring spatial points. If the perturbation vector directions of a spatial point and its surrounding points are consistent, it indicates that the perturbation is propagating continuously; however, if the direction changes significantly, it may be an anomaly source caused by partial discharge, short circuit, or structural damage. To quantify this degree of change, the system introduces the calculation of a directional mutation factor. This factor can be established by evaluating the standard deviation of the vector angle between a point and its surrounding points; that is, the greater the inconsistency in direction, the larger the mutation factor. After the directional mutation factors of all spatial points are calculated, a complete spatial characteristic map of the directional mutation factors is formed, providing directional constraints for clustering analysis.

[0068] Suppose we are currently processing a spatial coordinate point A in a three-dimensional magnetic flux perturbation vector field. The coordinates of this point are (5,5,5), indicating that it is located at the center of a standard sampling grid inside the iron core. To analyze the degree of directional abrupt change at this point, the system first constructs a neighborhood window around it, selecting 26 neighboring points (i.e., all points in a 3×3×3 cube except the center point). These points include neighboring sampling points in the up, down, left, right, front, back, and diagonal directions.

[0069] The system extracts the direction of the magnetic flux disturbance vector at point A, which is the spatial direction of the magnetic flux change, for example, pointing "slightly to the right and upward". Then, the system compares the direction of point A with the magnetic flux disturbance vector direction of each surrounding point. The comparison method is to normalize the vectors of the two points and then measure the angle between them in space. If the magnetic flux disturbance directions of the two points are exactly the same, their angle is close to zero degrees; if they are completely opposite, the angle is close to 180 degrees; and if the directions are not exactly the same but not completely opposite, the angle is somewhere in between.

[0070] For example, among these 26 neighboring points, suppose the perturbation direction of 20 points is basically consistent with that of point A, with only minor deviations, such as all pointing "right front" or "right front lower"; while the perturbation direction of another 6 points deviates significantly from that of point A, some even almost perpendicular or opposite, such as pointing "left rear" or "lower". This indicates that the perturbation direction in this area is relatively consistent overall, but there are drastic changes in direction in some local areas.

[0071] The system aggregates the values ​​of all these included angles and analyzes their dispersion. In other words, it doesn't just look at whether there's a deviation at one or two points, but rather statistically analyzes the "distribution" of all included angles. If these included angles are all small, then the differences between them are also small, and the directional abrupt change factor for that point will be assigned a low value; conversely, if these included angles vary in size, especially if there are several points with angles far from the overall average direction, then the system will identify this directional "inconsistency" and assign a higher directional abrupt change factor to point A.

[0072] To further enhance accuracy, the system also considers the weight of "outlier directions" on the overall impact. For example, when a neighboring point is detected to have a direction almost opposite to that of point A, the system will assign a higher weight to this contrast, making the point's influence on the final mutation factor more significant. This avoids the dilution of local outlier directions caused by simple averaging.

[0073] Finally, the directional mutation factor at point A is quantified into a numerical value, which may be a floating-point number from 0 to 1. The higher the value, the more inconsistent the direction of the disturbance in its neighborhood, and the more likely it is to be the source of some kind of physical anomaly, such as a change in magnetic flux direction caused by partial discharge, loose winding, or transient impact.

[0074] The system repeats the above analysis process for all spatial points and plots the mutation factors of each point into a "directional mutation factor map," similar to a heat map, where darker colors represent more severe mutations. This map can not only be used for visualization but also serve as input for subsequent spatial clustering or anomaly partitioning to identify which regions are forming structural electromagnetic anomalies.

[0075] Next, based on the directional mutation factor map, the system selects all spatial points whose mutation factor values ​​exceed a set clustering threshold as candidate points and performs spatial cluster identification. This process typically employs a density-based clustering algorithm to search for clusters of points that are close to each other and have consistent directional mutations in three-dimensional space, forming local anomalous clusters. These regions may be the source points of electromagnetic anomaly evolution. After clustering, the system performs statistical analysis on the characteristic indicators within each anomalous cluster. Specifically, it calculates the average value of the magnetic flux disturbance amplitude, the average value of the slope of change, and the average value of the directional mutation factor within each cluster, while also calculating the dispersion of these features to assess the concentration and stability of the anomaly. Based on this, the system constructs a cluster density index for each anomalous cluster, which quantifies whether the region possesses "source-type" anomalous characteristics, i.e., whether it constitutes a highly clustered anomalous point cluster in space.

[0076] After the above analysis is completed, the system will integrate the four key indicators on a spatial coordinate point basis. For each spatial point, the final characteristic parameters include: the magnetic flux disturbance amplitude, the magnetic flux change slope, the directional abrupt change factor, and whether the point belongs to a certain abnormal clustering region and the clustering density index of that region. These data form a set of structured physical characteristic parameters, which can serve as a direct input data source for subsequent interference identification, fault type determination, and intelligent response control. This set has key technical advantages such as point location, interpretable indicators, and traceable evolution, making it highly practical in intelligent transformer operation diagnosis systems.

[0077] Furthermore, various parameters in this process, such as the time-domain sliding window length, directional mutation threshold, and cluster density threshold, can be trained and optimized based on the transformer model, operating environment, and historical data, thereby adapting to transformer products with different voltage levels, insulation structures, and cooling methods. All algorithms can run on conventional embedded platforms or edge computing nodes, and possess excellent real-time performance and computational controllability.

[0078] In summary, by extracting perturbation vectors point by point from the periodic differences of the three-dimensional magnetic flux field and constructing anomaly structure features by combining four dimensions—temporal trend, spatial gradient, abrupt change in direction, and density clustering—this method achieves accurate localization and qualitative description of magnetic flux anomalies. It provides a high-quality, physically consistent, and engineering-adaptable feature foundation for subsequent vibration interference elimination and fault classification, which is significantly better than existing traditional detection methods that rely solely on threshold judgment based on single-point changes in current and voltage.

[0079] Step S104: Synchronously collect vibration data of the transformer body, compare and analyze the vibration data with the set of physical characteristic parameters, remove interference components caused by external disturbances, and generate abnormal response signals after interference suppression.

[0080] In implementing this invention, step S104 plays a crucial role in connecting perception and judgment. Its core function is to synchronously compare the physical characteristic parameters extracted through magnetic flux reconstruction analysis with the vibration data collected from the transformer itself. This eliminates spurious changes caused by external disturbances, thereby extracting the truly effective response signal related to internal electromagnetic anomalies. This step ensures the accuracy of subsequent anomaly detection and fault classification, and is a key link in improving the system's anti-interference capability and judgment reliability.

[0081] This step first requires deploying highly sensitive vibration sensors. These sensors can be MEMS accelerometers, piezoelectric sensors, or laser interferometer non-contact vibrometers, and should be installed at critical nodes in the transformer's structure, such as winding supports, core fixing components, the outer wall of the tank, and cooling pipe supports. These locations can sensitively respond to changes in the transformer's vibration characteristics during operation, especially structural disturbances caused by local loosening, winding deformation, discharge impacts, or electromagnetic excitation. The sensors should be synchronized with the magnetic flux density acquisition system in time. It is recommended to use a triggering mechanism consistent with the magnetic flux sampling or a unified clock system to ensure that within each sampling period, the magnetic flux data and vibration data can correspond one-to-one and form a time series with the same frequency.

[0082] Within each cycle, the vibration sensor acquires the raw vibration signal waveform, which is a time-domain acceleration curve, typically measured in m / s². For ease of comparison, the vibration data should first be preprocessed, including baseline drift elimination, bandpass filtering (e.g., 10–1000 Hz) to remove low-frequency temperature drift and high-frequency electromagnetic noise interference, and its spectral distribution obtained using Fourier transform. Next, characteristic indicators matching the set of magnetic flux disturbance parameters should be extracted, such as: Vibration energy amplitude (integral intensity within a specific frequency band); dominant frequency component and amplitude; spectral centroid; The RMS value, peak value, and impact factor in the time domain. These indicators constitute a set of vibration characteristic parameters.

[0083] Subsequently, the set of vibration characteristic parameters is compared one-to-one with the set of magnetic flux physical characteristic parameters. The comparison method can employ a strategy of time alignment plus spatial proximity. That is, for a node identified as abnormal by magnetic flux disturbance within a certain spatial region, the system checks whether the vibration sensor at the corresponding location in that region also shows significant changes within the same period. If the vibration intensity exceeds a set threshold, or is consistent with the trend of magnetic flux disturbance changes (e.g., both showing a sudden increase), then the abnormal response is considered to be possibly affected by external mechanical excitation or load disturbance; conversely, if the magnetic flux anomaly in that region is obvious but the vibration response is stable, then the anomaly is more likely to originate from an electromagnetic anomaly inside the core rather than a structural disturbance.

[0084] Based on the above comparison results, by designing interference removal criteria, abnormal signals dominated by mechanical disturbances can be excluded from the magnetic flux disturbance data. For example, a set of joint judgment conditions can be set: If a spatial point satisfies the condition that the magnetic flux disturbance amplitude is higher than the first threshold However, the vibration intensity is below the second threshold. If the spectrum does not contain obvious mechanical shock components, then the anomaly at that point is determined to be an electromagnetic source. If there is a high positive correlation between magnetic flux disturbance and vibration change (e.g., Pearson correlation coefficient greater than 0.9), then this point is considered a mechanically induced disturbance and is not included in subsequent abnormal response analysis.

[0085] Through the above elimination mechanism, a set of abnormal response signals that have been removed from mechanical disturbance pseudo-signals is finally generated. This set is spatially consistent with the magnetic flux disturbance spectrum, but only retains the real response data that are highly correlated with nonlinear instability of magnetic flux inside the iron core, current distortion, and eddy currents induced by partial discharge.

[0086] These abnormal response signals will serve as input for abnormal trend judgment and fault type identification in subsequent step S105. It is important to emphasize that this step not only purifies the data but is also a crucial step in enhancing the overall system's resistance to environmental interference.

[0087] To more clearly illustrate the implementation process of step S104, the following specific example demonstrates the complete application path of the interference removal logic, and provides the structural form and specific content of the abnormal response signal after interference suppression.

[0088] Taking a 500 kVA three-phase distribution transformer in operation as an example, a 4×4×4 three-dimensional magnetic flux density distribution sensor array is installed inside its core. Simultaneously, a total of 12 triaxial vibration sensors are installed at the core yoke, winding support, and the center of the transformer housing. Assume the system is analyzing the 100th sampling period (numbered as follows). During this period, the flux reconstruction module identified a significant increase in the amplitude of the flux disturbance at the spatial coordinate point (2,3,1), reaching ΔB = 2.1 mT, with a slope of S = 0.43 mT / s, which is far higher than the first threshold. It was initially identified as a high-risk anomaly.

[0089] The system then automatically retrieved vibration sensor data corresponding to the spatial location of that point. It was found that after bandpass filtering, the vibration waveform had concentrated energy in the frequency band of 10–500 Hz, but the total energy integral was 0.015 m² / s³, which was lower than the vibration energy threshold. Meanwhile, its dominant frequency is 150 Hz, and it does not exhibit typical mechanical shock characteristics, such as no obvious instantaneous jumps or harmonic broadening. Based on these characteristics, the system considers that the vibration response at this point is insufficient to explain the drastic changes in magnetic flux. Therefore, external mechanical interference factors are excluded, and this node is determined to be the effective point of the internal electromagnetic anomaly response.

[0090] In contrast, another flux anomaly node (1,2,4) has ΔB = 1.8 mT and S = 0.35 mT / s. However, the vibration sensor at the corresponding location recorded an impact energy as high as 0.21 m² / s³ within the same period, with the dominant frequency concentrated at 100 Hz and exhibiting sudden spikes. The vibration waveform changes were highly correlated with the flux disturbance, with a Pearson coefficient of 0.93. Based on this, the system concluded that the flux disturbance at this node was likely affected by a load step, mechanical loosening, or resonance of the cooling structure, and therefore did not include it in the final anomaly response signal set.

[0091] After completing the item-by-item comparison and screening of all nodes, the system generates a set of abnormal response signals after interference suppression for all retained magnetic flux anomaly nodes, as shown in the following structure: [ { "Position": [2, 3, 1], Sampling period: 100, "ΔB": 2.1, "S": 0.43, "Vibration energy": 0.015, Clock speed: 150 Judgment: "Internal electromagnetic anomaly" }, { "Position": [3, 1, 2], Sampling period: 100, "ΔB": 1.7, "S": 0.31, "Vibrational Energy": 0.012, "Core frequency": 120, Judgment: "Internal electromagnetic anomaly" } ] The above response data is provided to the system in JSON structure or tensor list format. It still maintains the coordinate consistency with the magnetic flux disturbance spectrum in space. However, after comparison with vibration data, false anomalies induced by mechanical disturbances have been eliminated. This ensures that the dataset focuses on reflecting the real local aging, inter-turn short circuits or early discharge anomalies that may exist inside the transformer core, providing an accurate and reliable input basis for subsequent diagnosis and control.

[0092] Furthermore, the step of comparing and analyzing the vibration data with the set of physical characteristic parameters, eliminating interference components caused by external disturbances, and generating an abnormal response signal after interference suppression includes: Extract the magnetic flux disturbance amplitude and magnetic flux change slope corresponding to each spatial coordinate point in the current sampling period from the set of physical feature parameters, and simultaneously extract the vibration data of the transformer body corresponding to the position of the spatial coordinate point. The vibration data includes the time series of triaxial vibration acceleration signals, which constitute the input data group for comparative analysis. Based on the relationship between the magnetic flux disturbance amplitude and the acceleration time series in three directions in the vibration data, the correlation index is calculated respectively. The correlation index includes the Pearson correlation coefficient, which is used to determine the dynamic correlation strength between the magnetic flux disturbance and the vibration response. The output is a set of spatial point comparison results containing the correlation score. In the set of spatial point comparison results, spatial coordinate points with correlation scores lower than the set interference judgment threshold are selected. The system inputs the magnetic flux disturbance amplitude, magnetic flux change slope and vibration data corresponding to these spatial coordinate points into the anomaly source discrimination calculation process. The calculation process integrates the dominant frequency distribution characteristics, energy density index and magnetic flux change slope of the vibration data to generate anomaly source discrimination factor for judging the attributes of the anomaly source, and outputs a set of anomaly response points with anomaly source confidence labels. Based on the set of abnormal response points, spatial coordinate points that overlap with the high vibration response region in terms of spatial structure and whose anomaly source confidence label is driven by a mechanical source are removed. At the same time, spatial coordinate points that are in the stable region of vibration data and have large magnetic flux disturbance amplitude, continuously increasing magnetic flux change slope, and anomaly source discrimination factor greater than the threshold are retained. Finally, the magnetic flux disturbance amplitude, magnetic flux change slope, anomaly source discrimination factor and time index corresponding to these spatial coordinate points are integrated to generate an abnormal response signal after interference suppression. This signal is used to execute multi-threshold combination judgment logic based on the physical feature parameters in the abnormal response signal.

[0093] To achieve accurate identification of electromagnetic anomalies within transformers, this invention, building upon three-dimensional magnetic flux reconstruction and anomaly feature extraction, further introduces an interference elimination mechanism based on multi-source data comparison. This mechanism aims to eliminate pseudo-anomaly signals originating from external mechanical excitation or system vibration by cross-analyzing the transformer's structural response signals, thereby retaining the effective magnetic flux disturbance information that truly represents the evolution trend of internal electromagnetic anomalies. The entire interference suppression process closely revolves around the physical characteristic parameters of the magnetic flux disturbance and the synchronously acquired vibration data. By constructing a multi-component correlation model and anomaly source discrimination factors, high-precision, high-confidence anomaly response signal purification is achieved.

[0094] First, before performing interference analysis, the system has obtained the three-dimensional magnetic flux disturbance characteristics inside the transformer core within the current sampling period through preprocessing. These characteristics are indexed in spatial coordinates and include the magnetic flux disturbance amplitude and flux change slope at each coordinate point. The disturbance amplitude reflects the overall degree of change in magnetic flux density within the current period, while the change slope reflects the accelerating trend of this change over time. These two indicators together constitute the core input for anomaly detection.

[0095] To establish a comparative relationship between magnetic flux disturbance characteristics and mechanical response, the system will simultaneously collect triaxial vibration data deployed at different structural locations within the transformer body. The vibration data sources include high-precision accelerometers installed in the winding support frame, core clamping plate, and outer casing connecting beam. Each accelerometer can collect acceleration sequences at equal time intervals in the X, Y, and Z directions, forming a time series of vibration data. The spatial positions of these sensors are pre-converted to match the spatial coordinate system of the magnetic flux disturbance through coordinate mapping, thus enabling precise spatial pairing of the vibration data and the magnetic flux disturbance data.

[0096] After aligning the data in time and space, the system extracts the current magnetic flux disturbance amplitude for each spatial coordinate point and selects the acceleration time series in three directions collected by the vibration sensor closest to that point. Subsequently, the statistical correlation between the magnetic flux disturbance amplitude and the acceleration signals in the three directions is calculated to quantify whether the magnetic flux change is driven by structural vibration.

[0097] The correlation analysis here uses the Pearson correlation coefficient calculation method. This method standardizes the correlation coefficient to a certain range by statistically analyzing the linear correlation between two variables, for example, from -1 to 1, where 1 represents a perfect positive correlation, -1 represents a perfect negative correlation, and 0 represents no correlation. The system calculates the correlation coefficients between the magnetic flux disturbance amplitude and the vibration acceleration in the X, Y, and Z directions, and takes the largest absolute value as the comprehensive correlation score for that point. This score reflects whether the change in magnetic flux at that point synchronously responds to the structural vibration fluctuations. For example, if the magnetic flux disturbance at a certain point shows strong consistency with the vibration change in the Y direction, the correlation score will approach 1.

[0098] After calculating the correlation scores for all spatial points, the system compiles the results into a comparative analysis set, with each spatial point corresponding to a correlation score. The system then performs a threshold filtering operation on this set, removing points whose correlation scores exceed a preset threshold. These points are considered to have magnetic flux disturbances potentially dominated by mechanical excitation and lacking independent electromagnetic anomaly characteristics. The retained points are considered to have low correlation, and their magnetic flux disturbances are preliminarily believed to be not directly caused by mechanical vibration.

[0099] However, to avoid misjudgments caused by accidental factors such as transient resonance, the system further introduces an anomaly source discrimination mechanism. This mechanism performs a more in-depth signal structure analysis on the low-correlation spatial coordinate points selected above. The system re-extracts the triaxial vibration signal corresponding to each point in its neighborhood and performs spectral analysis to calculate the dominant frequency of the vibration signal in the neighborhood of that point, i.e., the frequency component where the maximum energy concentration occurs. At the same time, the system also calculates the total energy density value of the vibration signal to reflect the overall vibration intensity of the region.

[0100] Next, the system combines these vibration characteristic indicators with the slope of the magnetic flux disturbance change to construct an anomaly source discrimination factor. This factor can be numerically represented by a set of weighted rules. For example, if the dominant frequency falls within the transformer's operating frequency resonance range, and the slope of the magnetic flux disturbance change is stable, then this point is more likely a pseudo-anomaly caused by structural resonance. Conversely, if the vibration signal energy is low, the dominant frequency is discrete, and the slope of the magnetic flux disturbance change increases rapidly, then it can be preliminarily determined to be an electromagnetically driven anomaly. This discrimination factor can be output as a confidence label, marking the spatial point as a "highly reliable anomaly" or "suspicious interference."

[0101] Ultimately, the system obtains a set of anomalous response points with anomaly source confidence levels. This set will be used to construct the final anomalous response signal, eliminating all spatial points that overlap with high-energy-density vibration regions and whose anomaly source is determined to be dominated by mechanical sources. Even if these points have low correlation in the initial screening, they will be excluded because their vibration characteristics match the characteristics of mechanical anomalies. On the other hand, if the vibration data of a certain spatial region is stable for a long time, with discrete dominant frequencies and low energy density, but the magnetic flux disturbance amplitude is consistently high, the slope of change is continuously increasing, and the directional mutation factor is significantly increased, then the system retains it, considering it to be caused by changes in internal electromagnetic behavior.

[0102] After completing the above retention and elimination operations, the system integrates all retained spatial coordinates and their corresponding magnetic flux disturbance amplitude, magnetic flux change slope, anomaly source discrimination factor values, vibration response statistical summaries, and time index information to form the final anomaly response signal after interference suppression. This response signal has the characteristics of spatial positioning, temporal traceability, and complete physical attributes, and can be directly input into the subsequent multi-threshold combination judgment logic module to participate in the evolution identification and classification of anomaly trends.

[0103] The entire interference suppression process not only ensures a one-to-one mapping between magnetic flux and vibration information in the data structure, but also embodies a progressive logic of "weak correlation filtering first, then structural coupling identification, and finally signal fusion output" in the algorithm path, significantly different from existing interference removal methods that are simply based on amplitude or mean comparison. Its technical effects include: improving the anti-interference capability of anomaly detection, reducing the false alarm rate, and enhancing the early identification rate of internal electromagnetic failures. It is a crucial supporting module for ensuring the stable operation of the intelligent transformer fault early warning system. This method has clear interface standards in engineering implementation, and all calculation processes can be completed on a real-time industrial control processor, making it applicable to various types of transformers, including distribution and transmission level transformers.

[0104] Step S105: Based on the physical characteristic parameters in the abnormal response signal, execute the multi-threshold combination judgment logic, identify the abnormal evolution trend, and determine whether the abnormal response signal corresponds to the preset fault type, including internal insulation aging, winding inter-turn short circuit or partial discharge.

[0105] In the method of this invention, the core of step S105 lies in performing structured discrimination on the abnormal response signal after interference suppression processing, combining multiple physical characteristic parameters to execute comprehensive threshold judgment logic, thereby identifying whether the current abnormality is evolving into a potential fault, and further determining its specific fault type. This step is a key link in realizing the transition from abnormality perception to fault classification. Its judgment process must be based on clear and repeatable logic to ensure stability and reliability in actual applications during transformer operation.

[0106] First, the physical characteristic parameters carried in the abnormal response signal need to be deconstructed point by point in space. Each spatial point typically includes multiple indicators such as the amplitude of magnetic flux disturbance, the slope of magnetic flux change, the residual characteristic value after vibration response compensation, the spatial clustering weight, and the duration of the disturbance. These indicators reflect the dynamic behavior of the magnetic flux field in the region over recent periods and have significant predictive value for fault evolution. The system sets corresponding judgment thresholds for each type of feature and makes a comprehensive judgment based on combined conditions. For example, a magnetic flux disturbance amplitude exceeding 1.8 mT, a change rate exceeding 0.3 mT / s, and a duration greater than 3 periods can be judged as a high-risk anomaly with an evolutionary tendency.

[0107] Furthermore, the judgment logic in this step does not employ a linear triggering method based on a single variable. Instead, it constructs a nonlinear recognition model through a multi-threshold joint judgment strategy. Logically, the system can be described using Boolean combinations. For example, if condition A and condition B are satisfied, or condition C is satisfied but condition D is not, then the output is "Abnormal evolution in progress." This judgment logic can not only be constructed using empirical rules but can also be optimized through backtracking analysis of actual operational data to improve judgment sensitivity and fault tolerance.

[0108] After identifying anomalies with a clear evolutionary trend, the system will proceed to the fault type determination stage. This stage no longer focuses on whether an anomaly exists, but rather on determining the possible fault type based on multi-dimensional factors such as the spatial distribution, frequency domain characteristics, and local gradient morphology of the anomaly's characteristic patterns. The system has a built-in preset feature template set corresponding to various typical transformer internal faults, including inter-turn short circuits, insulation aging, partial discharge, and core loosening. For example, if anomaly response signals are concentrated in the upper region of the winding, accompanied by a continuously increasing magnetic flux disturbance slope but stable vibration, and high-amplitude disturbances frequently occur in this region in the previous few cycles, the system can initially identify it as an inter-turn short circuit; if the anomaly spreads slowly, the disturbance distribution is wide but not concentrated, and there is no obvious thermal shock, it may be caused by insulation aging; if the anomaly is sudden, accompanied by high-frequency disturbance pulses within a very short time, it suggests the possible presence of partial discharge.

[0109] The identification results are ultimately recorded in a structured format, along with a confidence level label, as input for the control module. For example, the system can output the following diagnostic information: "Judgment type: winding inter-turn short circuit, area center: (x=2, y=3, z=1), confidence level = 0.91, suggested response type: emergency load transfer and local cooling." Such judgments not only facilitate subsequent automatic control responses but can also be used for manual review by maintenance personnel, system learning optimization, and maintenance strategy decision-making.

[0110] Therefore, step S105 is not only a logical filter for abnormal signals, but also an attribution judgment for the complex state evolution process. Its multi-threshold joint identification method and template matching classification strategy ensure timely and accurate identification of multiple types of internal faults during transformer operation, and constitute the diagnostic core of the entire intelligent anomaly perception system.

[0111] To illustrate the practical application of step S105 more specifically, the following case study uses a 220 kV main transformer under operation as an example. Assume that in the 180th sampling period, the system identifies a flux disturbance point located at spatial coordinates (3, 2, 1). The abnormal response signal after interference suppression contains the following physical characteristics: the flux disturbance amplitude is 2.3 mT, the slope of change is 0.42 mT / s, the abnormal duration reaches 5 consecutive sampling periods, and this point is located in the longitudinal middle region between winding layers. Simultaneously, there is no significant vibration response, and spatial clustering indicators show that this point, along with 6 surrounding disturbance points, forms a cluster area.

[0112] After retrieving the corresponding fault identification rules, the system first assesses the evolution trend of the point. Since the magnetic flux disturbance amplitude has exceeded the preset threshold... The slope is higher than The duration exceeds the minimum evolution time threshold The period is longer and the spatial clustering is greater than the set value. The system satisfies the Boolean logic "if ΔB ≥ 0 And S ≥ And T ≥ And C ≥ If the condition is met, it is determined to be an "evolutionary anomaly". Therefore, the point is marked as "evolving anomaly" by the system and enters the fault classification process.

[0113] In fault type determination, the system further analyzes the characteristic distribution of the abnormal cluster area. All abnormal points are detected concentrated in the middle of the winding near the same vertical channel, exhibiting high consistency in magnetic flux disturbance; vibration signals do not show resonance or structural response in this area; the disturbance slope shows a significant increasing trend from the first five cycles, and the spatial disturbance gradient is distributed radially without any outward diffusion. According to the system's built-in fault mode template, this mode matches the "early development type of inter-turn short circuit" with a 92% match rate, exceeding the system's warning threshold of 85%. Therefore, the system outputs the fault determination result as "inter-turn short circuit," along with the following diagnostic data: Judgment type: Inter-turn short circuit in winding; Anomaly center coordinates: (x=3, y=2, z=1); Flux disturbance amplitude: 2.3mT; Change slope: 0.42 mT / s; Spatial clustering: 6 points; Anomaly duration: 5 cycles; Judgment confidence level: 0.92; Suggested response: Immediately reduce operating load by 20%, start the cooling fan group in the middle of the winding, and mark this area into the monitoring priority list.

[0114] As can be seen from the above process, step S105 not only achieves multi-parameter joint anomaly trend identification, but also provides accurate fault classification and response suggestions based on actual evolution characteristics, providing a basis for subsequent intelligent control of the system. In this example, all parameters, judgment logic, and template matching algorithm have clear definitions and implementation paths, and can be adapted and deployed according to transformer model and historical operating data.

[0115] Step S106: Based on the identified fault type, call the corresponding handling strategy rules to generate a set of control instructions, including load adjustment commands or local cooling control commands, and send them to the target execution component through the transformer control interface to achieve intelligent adaptive adjustment and control of the transformer.

[0116] In the final step of the method described in this invention, the system completes the identification of abnormal evolution trends and accurate determination of fault types, subsequently entering the crucial control response stage. The objective of step S106 is to match the pre-defined response rules based on the fault type identified in the previous step, thereby generating corresponding control commands and applying them to the transformer operating system to achieve intelligent adaptive adjustment of its state. The successful implementation of this step marks the completion of the entire perception-identification-response closed loop, truly realizing the transformation from passive monitoring to active control.

[0117] The process begins by receiving diagnostic results from the fault identification module, including the fault type (e.g., internal insulation aging, inter-turn short circuit, partial discharge, etc.), coordinates of the abnormal area, fault evolution level, and system recommended confidence level. The control strategy module then matches the corresponding set of handling rules based on the fault type. These rules are pre-defined based on transformer operating experience, simulation analysis, and equipment manufacturer recommendations, and are stored in a strategy library. Each rule clearly specifies the optimal operating method for a particular type of fault, including adjustment range, response speed, execution priority, and safety protection conditions.

[0118] Taking a winding inter-turn short circuit as an example, when the system determines that the fault level is "medium risk and accelerating," the control strategy library will return a set of handling rules, instructing the system to reduce the load power by 15% to 20% to alleviate the pressure on the short-circuit branch current; at the same time, it will activate the auxiliary fan unit located near the cooling duct in the abnormal area and set the target oil temperature to 5°C lower than the normal value; if the fault signal persists for two cycles without relief, it will issue a warning message and suggest manual inspection. After the above response scheme is called by the strategy library, it will be automatically parsed into a structured control instruction set, which includes the control type (such as load adjustment, cooling activation), target parameters (such as power limit, oil temperature setpoint), execution object (such as load connection, cooling fan number), and effective time.

[0119] Subsequently, the instruction set is sent to the transformer control and execution system through a standardized control interface. This interface can be a communication interface based on the IEC 61850 standard or a custom communication protocol integrated with the local SCADA system. Control commands enter the transformer's low-voltage side control module, cooling system controller, or excitation regulation unit through the interface, and the execution status is fed back in real time by each actuator. The system will continuously monitor the feedback changes in magnetic flux and vibration after the response to determine whether the regulation effect meets the target. If necessary, it will automatically adjust the control parameters or switch to the backup control path to ensure the stability and closed-loop reliability of the entire regulation process.

[0120] Throughout the control process, the system automatically generates control logs, recording information such as command issuance time, execution target, feedback status, and effect evaluation. These logs are then submitted to the backend data platform as a basis for subsequent scheduling optimization and policy updates. If a manual intervention mechanism is deployed, authorized nodes can be set up, requiring certain critical commands to be confirmed by operations and maintenance personnel before taking effect, provided that security protocols are met, thus balancing automation efficiency with operational safety.

[0121] In summary, step S106 does not simply transmit the identification results to the control system, but rather possesses a high degree of adaptability and response matching capability in its control logic. Based on different anomaly types and evolution trends, it achieves full-chain linkage from the strategy layer to the execution layer, ensuring that the transformer can make rapid, accurate, and controlled adjustments when facing potential fault threats, thereby improving the safety, stability, and reliability of the system operation.

[0122] The second embodiment of the application provides an electronic device, the electronic device comprising: processor; The memory is used to store a program, which, when read and executed by the processor, executes the intelligent transformer anomaly sensing method for three-dimensional magnetic flux reconstruction provided in the first embodiment of this application.

[0123] The third embodiment of this application provides a computer-readable storage medium storing a computer program thereon. When the program is executed by a processor, it executes a three-dimensional magnetic flux reconstruction intelligent transformer anomaly sensing method provided in the first embodiment of this application.

[0124] Although this application discloses preferred embodiments as described above, it is not intended to limit this application. Any person skilled in the art can make possible changes and modifications without departing from the spirit and scope of this application. Therefore, the scope of protection of this application should be determined by the scope defined in the claims of this application.

Claims

1. A method for intelligent transformer anomaly detection based on three-dimensional magnetic flux reconstruction, characterized in that, include: A three-dimensional magnetic flux density distribution sensor array is deployed inside the transformer core to collect multi-point spatiotemporal data of magnetic flux density in real time during transformer operation and generate the original magnetic flux density sampling matrix. The original magnetic flux density sampling matrix is ​​spatially fitted and interpolated to obtain a three-dimensional magnetic flux field distribution map for the current operating cycle. The three-dimensional magnetic flux field distribution map is used to characterize the directionality and local perturbation features of the magnetic flux vector. The differences in the three-dimensional magnetic flux field distribution map corresponding to the continuous operation cycle are analyzed to extract magnetic flux disturbance features, including magnetic flux disturbance amplitude, magnetic flux change slope and local abnormal clustering areas, and a set of physical feature parameters is generated accordingly. Vibration data of the transformer body are collected synchronously, and the vibration data is compared and analyzed with the set of physical characteristic parameters. Interference components caused by external disturbances are eliminated, and abnormal response signals after interference suppression are generated. Based on the physical characteristic parameters in the abnormal response signal, a multi-threshold combination judgment logic is executed to identify the abnormal evolution trend and determine whether the abnormal response signal corresponds to a preset fault type, including internal insulation aging, winding inter-turn short circuit or partial discharge. Based on the identified fault type, the corresponding handling strategy rules are invoked to generate a set of control instructions, including load adjustment commands or local cooling control commands, and sent to the target execution component through the transformer control interface to achieve intelligent adaptive adjustment and control of the transformer.

2. The intelligent transformer anomaly sensing method for three-dimensional magnetic flux reconstruction according to claim 1, characterized in that, The method of deploying a three-dimensional magnetic flux density distribution sensor array inside the transformer core for real-time acquisition of multi-point spatiotemporal data of magnetic flux density during transformer operation, and generating an original magnetic flux density sampling matrix, includes: A three-dimensional magnetic flux density distribution sensing array with an equilateral cubic structure is constructed inside the transformer core. A set of triaxial magnetic flux density sensor nodes is arranged at the center of each cubic unit to obtain the component information of the magnetic flux density at the center point in three spatial directions, forming a basic sampling point set. Based on the spatial gradient value of magnetic flux density at each point in the basic sampling point set, the region where the magnetic flux density gradient exceeds the preset threshold is identified. The equilateral cube structure is locally refined in this region, and a densified sampling point set is formed by adding a triaxial magnetic flux density sensor node. The magnetic flux density three-component values ​​of each sampling point in the encrypted sampling point set are encapsulated in a unified format to construct a data structure with five dimensions, where the five dimensions correspond to sampling time, X spatial coordinate, Y spatial coordinate, Z spatial coordinate and magnetic flux density direction component, respectively, forming an uncompressed multidimensional magnetic flux density data set. Real-time online structural compression processing is performed on the uncompressed multidimensional magnetic flux density dataset. The compression operation is based on the spatial redundancy calculation between sampling points. The compressed magnetic flux density dataset is generated by retaining the complete data of the key regions of the magnetic flux density gradient and compressing the data of the regions with stable changes. The compressed magnetic flux density data set is time-aligned with the transformer body vibration data collected within the same sampling period. The alignment method is based on a set unified sampling period control signal with a time resolution of no more than 1 millisecond, so as to obtain a multi-source data group after time synchronization. Finally, the time-synchronized magnetic flux density data part is used as the original magnetic flux density sampling matrix for spatial fitting and interpolation processing of the original magnetic flux density sampling matrix.

3. The intelligent transformer anomaly sensing method for three-dimensional magnetic flux reconstruction according to claim 1, characterized in that, The step of performing spatial fitting and interpolation processing on the original magnetic flux density sampling matrix to obtain the three-dimensional magnetic flux field distribution map for the current operating cycle includes: The magnetic flux density multi-point spatiotemporal data contained in the original magnetic flux density sampling matrix are spatially divided according to X spatial coordinates, Y spatial coordinates and Z spatial coordinates to form multiple spatial sub-regions, and the three-component data of magnetic flux density in each spatial sub-region are extracted. For each spatial sub-region, the magnetic flux density three-component data are subjected to joint fitting processing. The joint fitting processing includes performing a weighted superposition operation of cubic spline interpolation and inverse distance weighted interpolation on the magnetic flux density three-component data respectively, so as to simultaneously ensure the continuity of derivatives in the boundary region and the fitting stability in the central region, and generate the initial fitting result of three-dimensional magnetic flux density in the spatial sub-region. Based on the initial fitting results of the three-dimensional magnetic flux density, the divergence value of the three components of magnetic flux density at each spatial coordinate point is calculated, the magnetic flux conservation constraint condition of the three-dimensional magnetic flux field is established, the spatial location points with non-zero divergence are identified as fitting error points, and the statistical results of magnetic flux density residuals containing the coordinates of each error point and its corresponding error value are output. Based on the statistical results of the magnetic flux density residuals, for spatial locations where the residuals exceed a preset threshold, the spatiotemporal data of the neighboring magnetic flux density multi-points in the original magnetic flux density sampling matrix are called in the corresponding spatial sub-region. Local data augmentation is performed, and joint fitting processing of cubic spline interpolation and inverse distance weighted interpolation is performed again. Finally, the three-dimensional magnetic flux field distribution map of the current running cycle is output, which is used to perform difference analysis on the three-dimensional magnetic flux field distribution maps corresponding to continuous running cycles.

4. The intelligent transformer anomaly sensing method for three-dimensional magnetic flux reconstruction according to claim 1, characterized in that, The process involves performing a difference analysis on the three-dimensional magnetic flux field distribution map corresponding to continuous operating cycles, extracting magnetic flux disturbance features including magnetic flux disturbance amplitude, magnetic flux change slope, and local abnormal clustering regions, and generating a set of physical feature parameters based on these features, including: The magnetic flux density components of each corresponding spatial coordinate point in the three-dimensional magnetic flux field distribution map obtained in each continuous operation cycle are calculated point by point to obtain the magnetic flux disturbance vector of the spatial coordinate point. Based on the magnetic flux disturbance vector, a three-dimensional magnetic flux disturbance vector field containing all spatial coordinate points is constructed. Based on the three-dimensional magnetic flux disturbance vector field, the magnetic flux disturbance amplitude at each spatial coordinate point is calculated, and a time-domain sliding window is constructed based on the magnetic flux disturbance amplitude sequence of the spatial coordinate point in adjacent periods. The slope of the change of the disturbance amplitude within the window is extracted using a weighted least squares fitting method as the magnetic flux change slope value of the spatial coordinate point. The intermediate feature result set containing the magnetic flux disturbance amplitude and the magnetic flux change slope is output. A local tensor field is constructed for the three-dimensional magnetic flux perturbation vector field. Based on the gradient of the change in the perturbation vector direction of each spatial coordinate point in its neighborhood, a directional mutation factor is calculated. This directional mutation factor is used to characterize the discontinuity of the perturbation direction in three-dimensional space, serving as the directional criterion for subsequent spatial clustering analysis, and a spatial feature map with the directional mutation factor is output. In the spatial feature map with directional mutation factors, spatial coordinate points whose mutation factor values ​​exceed a preset clustering threshold are selected, and density clustering algorithm is executed to form local abnormal clustering regions. In each abnormal clustering region, the cluster mean and dispersion of the three features included—magnetic flux disturbance amplitude, magnetic flux change slope, and directional mutation factor—are statistically analyzed, and a cluster density index for the abnormal clustering region is constructed as the output content of the spatial structural anomaly features. Finally, the magnetic flux disturbance amplitude, magnetic flux change slope, directional mutation factor, and the cluster density index of the abnormal clustering region to which each spatial coordinate point belongs are uniformly integrated to generate a set of physical feature parameters.

5. The intelligent transformer anomaly sensing method for three-dimensional magnetic flux reconstruction according to claim 1, characterized in that, The step of comparing and analyzing the vibration data with the set of physical characteristic parameters, removing interference components caused by external disturbances, and generating an abnormal response signal after interference suppression includes: Extract the magnetic flux disturbance amplitude and magnetic flux change slope corresponding to each spatial coordinate point in the current sampling period from the set of physical feature parameters, and simultaneously extract the vibration data of the transformer body corresponding to the position of the spatial coordinate point. The vibration data includes the time series of triaxial vibration acceleration signals, which constitute the input data group for comparative analysis. Based on the relationship between the magnetic flux disturbance amplitude and the acceleration time series in three directions in the vibration data, the correlation index is calculated respectively. The correlation index includes the Pearson correlation coefficient, which is used to determine the dynamic correlation strength between the magnetic flux disturbance and the vibration response. The output is a set of spatial point comparison results containing the correlation score. In the set of spatial point comparison results, spatial coordinate points with correlation scores lower than the set interference judgment threshold are selected. The system inputs the magnetic flux disturbance amplitude, magnetic flux change slope and vibration data corresponding to these spatial coordinate points into the anomaly source discrimination calculation process. The calculation process integrates the dominant frequency distribution characteristics, energy density index and magnetic flux change slope of the vibration data to generate anomaly source discrimination factor for judging the attributes of the anomaly source, and outputs a set of anomaly response points with anomaly source confidence labels. Based on the set of abnormal response points, spatial coordinate points that overlap with the high vibration response region in terms of spatial structure and whose anomaly source confidence label is driven by a mechanical source are removed. At the same time, spatial coordinate points that are in the stable region of vibration data and have large magnetic flux disturbance amplitude, continuously increasing magnetic flux change slope, and anomaly source discrimination factor greater than the threshold are retained. Finally, the magnetic flux disturbance amplitude, magnetic flux change slope, anomaly source discrimination factor and time index corresponding to these spatial coordinate points are integrated to generate an abnormal response signal after interference suppression. This signal is used to execute multi-threshold combination judgment logic based on the physical feature parameters in the abnormal response signal.

Images