Transformer iron core defect detection method, device and equipment

Abstract

The invention discloses a transformer iron core defect detection method, device and equipment, and the method comprises the steps: collecting multi-modal data in the operation process of a transformer, the multi-modal data comprising a vibration signal, a temperature field, a gas concentration and a harmonic current; performing feature fusion processing on the multi-modal data according to the modal weight and a preset cascade fusion network to obtain a fusion feature vector; constructing a spatial impedance distribution equation according to the fusion feature vector, and performing electromagnetic-thermal coupling effect iterative analysis to obtain an impedance distribution matrix; and based on a constraint reinforcement learning strategy and a preset iron core diagram model, performing defect detection analysis according to the fusion feature vector and the impedance distribution matrix to obtain a defect evaluation grade. The technical problems that in the prior art, multi-source heterogeneous data fusion is difficult to achieve, the coupling influence between cross-physical fields is not considered, and consequently the defect detection result lacks pertinence, accuracy and reliability can be solved.

Description

Technical Field

[0001] This application relates to the field of transformer fault monitoring, and in particular to a method, apparatus and equipment for detecting defects in transformer cores. Background Technology

[0002] With the expansion of ultra-high voltage power grids and the integration of new energy sources, unplanned outages caused by transformer core defects are becoming increasingly prominent. Traditional detection technologies rely on single-mode threshold alarms, such as oil chromatography or infrared thermometry, which have significant technical drawbacks, including high false alarm rates (>15%) and coarse positioning (centimeter level).

[0003] In recent years, the development of digital twin and multiphysics simulation technologies has provided new approaches to core condition assessment. However, some problems still exist: first, multi-source heterogeneous data are difficult to integrate effectively; second, the coupling effects between physical fields are not considered. These problems result in core defect detection results lacking specificity, accuracy, and reliability, thus failing to meet the application requirements of real-world scenarios. Summary of the Invention

[0004] This application provides a method, apparatus, and equipment for detecting defects in transformer cores, which addresses the technical problem that existing technologies struggle to achieve multi-source heterogeneous data fusion and fail to consider the coupling effects between physical fields, resulting in a lack of specificity, accuracy, and reliability in defect detection results.

[0005] In view of this, the first aspect of this application provides a method for detecting defects in transformer cores, comprising: Multimodal data is collected during the operation of the transformer, including vibration signals, temperature field, gas concentration, and harmonic current. The multimodal data is subjected to feature fusion processing based on modal weights and a preset cascaded fusion network to obtain a fused feature vector; Based on the fused eigenvectors, a spatial impedance distribution equation is constructed, and an iterative analysis of the electromagnetic-thermal coupling effect is performed to obtain the impedance distribution matrix. Based on the constrained reinforcement learning strategy and the preset core diagram model, defect detection and analysis are performed according to the fused feature vector and the impedance distribution matrix to obtain the defect assessment level.

[0006] Preferably, the acquisition of multimodal data during transformer operation includes: The vibration spectrum of the transformer during operation is acquired by FFT according to a preset window length to obtain the vibration signal; The temperature field is determined by acquiring the spatial gradient of the infrared thermal image of the transformer during operation using a thermal imager. The concentration of dissolved gases in transformer oil was determined using a pre-set line chromatograph, and the gas concentration was obtained. Harmonic currents are obtained by measuring the current harmonics during the operation of the transformer.

[0007] Preferably, the step of performing feature fusion processing on the multimodal data based on modal weights and a preset cascaded fusion network to obtain a fused feature vector includes: The kernel density is estimated based on the multimodal data to obtain the mode weights; Based on the modal weights and the multimodal data, cascaded feature fusion calculations are performed using a preset cascaded fusion network to obtain a fused feature vector.

[0008] Preferably, the step of constructing a spatial impedance distribution equation based on the fused feature vector and performing iterative analysis of electromagnetic-thermal coupling effects to obtain the impedance distribution matrix includes: Based on the temperature field and harmonic current in the fused feature vector, Joule heating effect analysis and harmonic excitation analysis are performed, and a spatial impedance distribution equation is generated. Based on the spatial impedance distribution equation, an iterative analysis of the electromagnetic-thermal coupling effect based on a preset time step is performed to obtain the impedance distribution matrix.

[0009] Preferably, the constraint-based reinforcement learning strategy and the preset core diagram model perform defect detection analysis based on the fused feature vector and the impedance distribution matrix to obtain a defect assessment level, including: A three-dimensional mesh topology diagram of the transformer core is constructed to obtain a preset core diagram model; The fused feature vector, the impedance distribution matrix, and the vibration propagation function are input into the preset core diagram model for convolution calculation to obtain the anomaly probability map. Based on the constrained reinforcement learning strategy, defect detection analysis is performed according to the anomaly probability map to obtain the defect assessment level.

[0010] Preferably, the step of inputting the fused feature vector, the impedance distribution matrix, and the vibration propagation function into the preset core diagram model for convolution calculation to obtain the anomaly probability map includes: Construct an adjacency matrix based on the vibration propagation function; Calculate the temperature mask based on the temperature field gradient in the fused feature vector; The graph convolution features of the preset core diagram model are updated and calculated based on the adjacency matrix, the temperature mask, and the impedance distribution matrix to obtain an anomaly probability map.

[0011] Preferably, the defect assessment level is obtained by performing defect detection analysis based on the anomaly probability map using the constraint reinforcement learning strategy, including: Based on Fourier's law, a dual-constraint reward function is configured according to the gas change rate and heat flux. Based on the constrained reinforcement learning strategy, defect detection analysis based on action-state update operation is performed according to the anomaly probability map, the impedance distribution matrix and the dual-constraint reward function to obtain the defect assessment level.

[0012] Preferably, the constraint-based reinforcement learning strategy and the preset core diagram model perform defect detection analysis based on the fused feature vector and the impedance distribution matrix to obtain the defect assessment level, and then further includes: The defect information is obtained by performing a joint inversion analysis based on the defect assessment level and the anomaly probability in the anomaly probability map using a physical information neural network. The defect information includes the defect location, defect area, and defect depth.

[0013] The second aspect of this application provides a transformer core defect detection device, comprising: The data acquisition unit is used to acquire multimodal data during the operation of the transformer, including vibration signals, temperature field, gas concentration, and harmonic current. The feature fusion unit is used to perform feature fusion processing on the multimodal data according to modal weights and a preset cascaded fusion network to obtain a fused feature vector; The coupling analysis unit is used to construct the spatial impedance distribution equation based on the fused feature vector, and to perform iterative analysis of electromagnetic-thermal coupling effect to obtain the impedance distribution matrix; The defect detection unit is used to perform defect detection analysis based on the fused feature vector and the impedance distribution matrix according to the constrained reinforcement learning strategy and the preset core diagram model, and to obtain the defect assessment level.

[0014] A third aspect of this application provides a transformer core defect detection device, the device including a processor and a memory; The memory is used to store program code and transmit the program code to the processor; the transformer core defect detection method of the first aspect.

[0015] As can be seen from the above technical solutions, the embodiments of this application have the following advantages: This application provides a method for detecting defects in transformer cores, comprising: collecting multimodal data during transformer operation, including vibration signals, temperature fields, gas concentrations, and harmonic currents; performing feature fusion processing on the multimodal data based on mode weights and a preset cascaded fusion network to obtain a fused feature vector; constructing a spatial impedance distribution equation based on the fused feature vector and performing iterative analysis of electromagnetic-thermal coupling effects to obtain an impedance distribution matrix; and performing defect detection analysis based on a constrained reinforcement learning strategy and a preset core diagram model, using the fused feature vector and the impedance distribution matrix to obtain a defect assessment level.

[0016] The transformer core defect detection method provided in this application overcomes the difficulties of multi-source heterogeneous data fusion by using a fusion network to perform feature fusion analysis on various heterogeneous data such as vibration, temperature, gas, and current. Furthermore, based on the fused feature vectors, an impedance-electromagnetic-electrothermal coupling model is built, which not only quantifies the interaction between local overheating and multi-point grounding but also comprehensively considers the coupling effects across physical fields. The defect assessment level obtained based on this model conforms to the characteristics of actual scenarios, exhibiting strong specificity and ensuring the accuracy and reliability of the solution results. Therefore, this application solves the technical problems of existing technologies, which struggle to achieve multi-source heterogeneous data fusion and fail to consider the coupling effects between physical fields, resulting in a lack of specificity, accuracy, and reliability in defect detection results. Attached Figure Description

[0017] Figure 1 A flowchart illustrating a method for detecting defects in a transformer core provided in this application embodiment; Figure 2 This is a schematic diagram of a transformer core defect detection device provided in an embodiment of this application. Detailed Implementation

[0018] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present application, and not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present application.

[0019] For easier understanding, please refer to Figure 1 This application provides an embodiment of a transformer core defect detection method, which includes: Step 101: Collect multi-modal data during transformer operation. The multi-modal data includes vibration signals, temperature field, gas concentration, and harmonic current.

[0020] Further, step 101 includes: The vibration spectrum of the transformer during operation is acquired by FFT according to a preset window length to obtain the vibration signal; The temperature field is determined by acquiring the spatial gradient of the infrared thermal image of the transformer during operation using a thermal imager. The concentration of dissolved gases in transformer oil was determined using a pre-set line chromatograph, and the gas concentration was obtained. Harmonic currents are obtained by measuring the current harmonics during the operation of the transformer.

[0021] In this embodiment, the vibration frequency can be extracted using Fourier Transform (FFT). The range is 0-1kHz, and the preset window used in this process is a Hamming window with a length of 1024 points. The infrared thermal imager can extract infrared thermal images. This allows us to obtain the corresponding spatial gradient. Precision control at Specifically, it can be expressed as:

[0022] in, T For temperature, x The x-axis of the image, y The vertical axis represents the image coordinates, and the gradient above represents the product of the rates of change in the horizontal and vertical coordinates.

[0023] Sampling per second on an online chromatograph The concentration of dissolved gases in transformer oil can then be obtained. The types of gases are not unique and may include, but are not limited to, those mentioned above. , , The current harmonics obtained from synchronous measurement are expressed as follows: The specific harmonic current is expressed as ,in, The number of data collections is [number], and the range of values ​​is [range]. .

[0024] To ensure the accuracy and reliability of subsequent data-based analysis, this embodiment can also preprocess the collected multimodal data to improve data quality, remove anomalies, or fill in missing data. Specific preprocessing methods are not limited here and can be designed according to actual conditions and data types. The preprocessed 4-modal data can be denoted as a multimodal data matrix. ,in, The length of the time series. This embodiment integrates four types of heterogeneous data—vibration, temperature, gas, and current harmonics—and then performs subsequent fusion analysis. It fully considers the comprehensive impact of different levels of operational data on core defects during transformer operation, making it more consistent with actual scenario characteristics and providing more accurate data support for the detection results.

[0025] Step 102: Perform feature fusion processing on the multimodal data according to the modal weights and the preset cascaded fusion network to obtain the fused feature vector.

[0026] Further, step 102 includes: calculating kernel density estimation based on multimodal data to obtain mode weights; Based on modal weights and multimodal data, cascaded feature fusion calculations are performed using a pre-defined cascaded fusion network to obtain a fused feature vector.

[0027] Due to the heterogeneity of multimodal data, the physical dimensions and sampling frequencies of vibration signals, temperature fields, gas concentrations, and harmonic currents differ significantly. For example, vibration signals are at the kHz level, while gas concentrations are at the ppm level. Therefore, to enable unified analysis later, it is necessary to standardize the dimensions of these data. This embodiment can quantify the correlation between various data types and faults using modal weight allocation, that is, by quantifying the correlation between each modal data and the fault label Y through mutual information entropy, avoiding the subjectivity of manually setting weights. In addition, a pre-set cascaded fusion network can be used to extract fusion features from the four modal data, and the fusion features can be used for defect analysis.

[0028] Specifically, the modal weights can be calculated based on the kernel density estimation algorithm, and are expressed as:

[0029] in, For the first i A data matrix with multiple modalities can be , , , ; Y For example, a historical fault label vector. This indicates a localized overheating defect in the iron core. express and Y The mutual information entropy is calculated as follows:

[0030] in, The joint probability distribution can also be calculated using kernel density estimation algorithms. , For each modal data With fault Y Each of their respective probabilities. If a certain mode... With fault Y Strong correlation, then Increasing, synchronization will make Increase the weight value. Through an adaptive weight allocation mechanism, the detection rate of weak defects can be significantly improved, achieving a detection sensitivity of 97.3% in real-world scenarios.

[0031] Based on the obtained modal weights The process of feature fusion of multimodal data with a pre-defined cascaded fusion network is as follows:

[0032] in, For the trainable weight matrix in the cascaded fusion network, For bias terms, For the activation function, where z This represents the net input to the neuron; the activation function can suppress negative features. This represents the fusion layer index, with a value of Where L=4; the result is This is the fused feature vector.

[0033] Step 103: Construct the spatial impedance distribution equation based on the fused eigenvectors, and perform iterative analysis of electromagnetic-thermal coupling effect to obtain the impedance distribution matrix.

[0034] Further, step 103 includes: Joule heating effect analysis and harmonic excitation analysis are performed based on the temperature field and harmonic current in the fused eigenvector, and a spatial impedance distribution equation is generated. Based on the spatial impedance distribution equation, an iterative analysis of the electromagnetic-thermal coupling effect with a preset time step is performed to obtain the impedance distribution matrix.

[0035] This embodiment quantifies the coupling effect of core faults by constructing a spatial impedance distribution equation and performing iterative analysis of electromagnetic-thermal coupling effects. Existing technologies treat core impedance as a static parameter, but in real-world scenarios, the core may overheat locally, causing a 20%-30% decrease in the permeability of the silicon steel sheets, thereby altering the eddy current loss distribution. Therefore, this embodiment uses dynamic coupling analysis to reflect the combined influence of temperature field and harmonic current on impedance in real time, forming a closed-loop verification chain of electromagnetic anomaly-temperature anomaly-vibration propagation.

[0036] The spatial impedance distribution equation constructed based on the temperature field and harmonic current components in the fused eigenvectors is expressed as follows:

[0037] in, Indicates the horizontal coordinate position of the iron core space Place t Complex impedance at time t, unit is It can be obtained through actual measurement and calibration using a broadband impedance analyzer, and is used to reflect the change in eddy current path caused by insulation degradation between silicon steel sheet layers; , represents the temperature coefficient of resistance of silicon steel sheet, and the unit is . The result was obtained by fitting the Epstein square ring through a temperature rise experiment at 20-200℃. This indicates the current in the iron core grounding wire under normal operating conditions. ,like If so, a multi-point grounding alarm will be triggered; is the hysteresis loss coefficient, which characterizes the additional iron loss caused by harmonic current; For the first The third harmonic current component and the third harmonic will lead to zero-sequence magnetic flux, which will aggravate the overheating of the iron core edge. The fundamental frequency of the power frequency is the harmonic frequency. The interaction between the harmonic frequency and the power frequency will cause... The frequency harmonic oscillation; Indicates the spatial position of the iron core Place t Spatial gradient of temperature field at time t, in units of And there are:

[0038] This parameter comes from the spatial derivative of the infrared thermal imager, when... At that time, the risk of localized overheating can be determined. This is based on a preset time step. Solve using the above formula. The impedance distribution matrix can be obtained. .

[0039] Step 104: Based on the constraint reinforcement learning strategy and the preset core diagram model, perform defect detection analysis according to the fused feature vector and impedance distribution matrix to obtain the defect assessment level.

[0040] Further, step 104 includes: A three-dimensional mesh topology diagram of the transformer core is constructed to obtain a preset core diagram model; The fused feature vector, impedance distribution matrix and vibration propagation function are input into the preset iron core diagram model for convolution calculation to obtain the anomaly probability diagram; Based on a constrained reinforcement learning strategy, defect detection and analysis are performed using anomaly probability maps to obtain defect assessment levels.

[0041] Furthermore, the fused feature vector, impedance distribution matrix, and vibration propagation function are input into a pre-defined core diagram model for convolution calculation to obtain an anomaly probability map, including: Construct an adjacency matrix based on the vibration propagation function; Calculate the temperature mask based on the temperature field gradient in the fused feature vector; The graph convolution features of the preset iron core diagram model are updated and calculated based on the adjacency matrix, temperature mask, and impedance distribution matrix to obtain the anomaly probability map.

[0042] Furthermore, based on a constrained reinforcement learning strategy, defect detection analysis is performed using anomaly probability maps to obtain defect assessment levels, including: Based on Fourier's law, a dual-constraint reward function is configured according to the gas change rate and heat flux. Based on a constrained reinforcement learning strategy, defect detection analysis based on action-state update operations is performed using anomaly probability maps, impedance distribution matrices, and dual-constraint reward functions to obtain defect assessment levels.

[0043] Furthermore, based on a constrained reinforcement learning strategy and a pre-defined core diagram model, defect detection analysis is performed using fused feature vectors and impedance distribution matrices to obtain a defect assessment level. This process also includes: By using a physical information neural network to perform joint inversion analysis based on the defect assessment level and the anomaly probability in the anomaly probability map, defect information is obtained, including defect location, defect area, and defect depth.

[0044] It should be noted that constructing a 3D mesh topology map of the transformer core yields a preset core model that can be used for defect spatial location and propagation path analysis. The preset core model in this embodiment includes over 500 nodes. This model can deeply integrate vibration propagation paths and temperature variations, thereby fully analyzing the coupled effects of mechanical dynamics and thermodynamics; achieving pixel-level location of defect areas. In contrast, existing technologies can only perform coarse-grained detection of anomalies. The detection method in this embodiment is more precise, achieving an accuracy of 1mm in real-world scenarios. 2 The microscopic region.

[0045] Specifically, the adjacency matrix in the pre-defined core diagram model describes the vibration transmission relationship between core nodes, and the edge weights are determined by the mechanical impedance; it can be constructed as follows:

[0046] in, Angular frequency, For power frequency, Let be the vibration propagation function, expressed as: , , These represent the equivalent mass, damping, and stiffness parameters of the laminated sheets, respectively. For the Laplace operator, where , , It can be obtained through a hammer-beating experiment; if the node , If cracks exist, the value will decrease by 30%-50%.

[0047] The graph convolution feature update process can be expressed as follows:

[0048] in, For temperature mask, , These are all trainable network parameters, which are sensitivity parameters for controlling the effect of temperature. For the temperature gradient field between nodes, when hour, The value approaches 1, highlighting the overheated area. The degree matrix is ​​a diagonal matrix, which can be expressed as: It is used to normalize the adjacency matrix and prevent feature scale drift caused by differences in node degree. For the first l Layer node characteristics, initial layer It refers to the splicing characteristics of the impedance distribution matrix and the vibration spectrum. ,in For the number of nodes, For feature dimensions. The graph convolution kernel parameters are initialized using spectral graph theory and optimized through end-to-end backpropagation. The physical constraints are mandatory. It satisfies the discrete form of the wave equation to ensure physical consistency. The activation function in this embodiment is: This can prevent gradient vanishing. This represents the Hadamard product, which is an element-wise multiplication used to mask features. The model output is the anomaly probability map. This represents the defect area.

[0049] Anomaly detection and analysis based on graph structures can simultaneously analyze the transmission of vibration signals in the spatial dimension and the temperature change in the temporal dimension between iron core laminations. It can solve topological problems that traditional CNNs cannot handle and is more adaptable to the needs of real-world scenarios.

[0050] To improve the efficiency of decision-making after actual anomaly detection, this embodiment further realizes defect level assessment based on a constrained reinforcement learning strategy. This process involves specific dynamic decision-making and risk assessment, which can transform the obtained anomaly probability map into a defect severity level, specifically divided into levels I / II / III, realizing the leap from detection to decision-making. Unlike existing technologies that rely on manual threshold judgment, which leads to a delay in fault detection and causes a series of uncontrollable problems.

[0051] This embodiment uses the second law of thermodynamics as a penalty term in the reward function, forcing the agent to adhere to the principle of energy conservation and avoiding misjudgments that violate physical principles, such as temperature reversal. The specific analysis process is a multi-objective collaborative optimization process that balances detection sensitivity and false alarm risk, maintaining a 97% recall rate while suppressing the false alarm rate to below 0.8%.

[0052] Specifically, the dual-constraint reward function in this embodiment can be derived based on Fourier's law:

[0053] in, Thermodynamic penalty coefficient, derived from Fourier's law, can be used to minimize the false alarm rate in 300 sets of historical fault data through KKT condition optimization. The hydrogen concentration is the rate of change over time, expressed in ppm / s, directly reflecting the degradation rate of the insulating material; when A value >0.5 ppm / s indicates a severe partial discharge. This parameter is obtained by sampling every second using an online chromatograph and then reducing noise using a Kalman filter. Let L be the L2 norm of the Laplace operator for the temperature field, in units of 1. Under normal operating conditions Heat diffuses from high temperature to low temperature. If this occurs, a penalty mechanism is triggered. This parameter can be obtained by discretely solving the five-point difference based on the temperature data.

[0054] The action-state update process can be expressed as:

[0055] in, Indicates the state Next action The long-term expected reward, the dimension is That is, 3 defect levels and 3 maintenance actions. The expected reward is initialized using the prior knowledge matrix generated by fault tree analysis. As a discount factor, it controls the weight of future rewards. This will cause the agent to ignore latent defects. It is an adaptive learning rate, dynamically adjusted through the PPO algorithm, and its update rule is:

[0056] in, To train stride length.

[0057] The state at the next moment is represented by the following calculation process:

[0058] in, This indicates a vectorization operation.

[0059] action a The possible values ​​are as follows: Recording only data is a Level I defect. This triggers an alert and adjusts the load, classifying it as a Level II defect. Emergency shutdown for maintenance is classified as a Level III defect. Based on this, the defect assessment level can be determined. This embodiment embeds the laws of thermodynamics and electromagnetic field equations into the reinforcement learning reward function to ensure that the defect classification results conform to physical laws, with a false alarm rate of less than 0.8%.

[0060] The defect assessment level obtained in this embodiment can be used for both real-time alarms and subsequent joint inversion analysis based on physical constraints. Real-time alarms can push the level (I / II / III) to the substation automation system via GOOSE messages, triggering the following actions: Level I generates a log record and marks a yellow warning on the SCADA interface; Level II automatically reduces the load by 10%-15%; Level III triggers the circuit breaker to trip and starts the backup transformer.

[0061] This embodiment's joint inversion analysis process based on physical constraints only applies to defect assessment levels II and III and... Inversion analysis is performed on high-probability areas to reduce unnecessary computation; moreover, for Class III defects, multi-point grounding initial guess values ​​are preferred to accelerate convergence.

[0062] This embodiment quantifies defect detection results into specific defect parameters using a Physical Information Neural Network (PINN), providing clear quantitative indicators such as defect location, area, and depth, enabling more precise repair guidance. These quantitative indicators achieve millimeter-level accuracy, far exceeding the accuracy of existing inversion schemes. Furthermore, the inversion operation allows for multi-physics consistency verification, ensuring that the inversion results simultaneously satisfy electromagnetic, thermodynamic, and mechanical vibration laws, avoiding ill-conditioned solutions obtained through purely data-driven approaches.

[0063] The loss function of the Physical Information Neural Network (PINN) is:

[0064] This loss function can control the network output. Approaching the measured value ; These are the parameters for predicting defects, specifically expressed as: These correspond to the location coordinates, area, and depth of the defect, respectively. The measured reference value is a parameter obtained by simultaneously performing pulsed eddy current detection and ultrasonic scanning. It is automatically calibrated every 6 hours, and the error can be controlled within 0.3 mm. These are physical constraint weights, dynamically adjusted using KKT conditions. hour, Increase by 20%. This represents the number of valid measurement points in the current batch. M The number of mesh nodes in a finite element simulation is usually... Magnitude. H This represents the magnetic field on the surface of the iron core; Represents the curl of a magnetic field, with units of _____. This can be obtained through simulation, and the boundary conditions are:

[0065] in, This represents the tangential component of the magnetic field strength on the surface of the iron core. This is the radius of the grounding wire.

[0066] Current density, unit is The calculation method is as follows:

[0067] in, For electric field strength, D It is the electric displacement vector. The electrical conductivity of silicon steel sheets can be expressed as:

[0068] in, T Indicates temperature.

[0069] The output of the inversion solution is specific defect information, expressed as: These correspond to the defect location, area, and depth, respectively. A Physical Information Neural Network (PINN) is used to combine simulation and measured data, and then the defect information guides precise repair.

[0070] The transformer core defect detection method provided in this application overcomes the difficulties of multi-source heterogeneous data fusion by using a fusion network to perform feature fusion analysis on various heterogeneous data such as vibration, temperature, gas, and current. Furthermore, based on the fused feature vectors, an impedance-electromagnetic-electrothermal coupling model is built, which not only quantifies the interaction between local overheating and multi-point grounding but also comprehensively considers the coupling effects across physical fields. The defect assessment level obtained based on this model conforms to the characteristics of actual scenarios, exhibiting strong specificity and ensuring the accuracy and reliability of the solution results. Therefore, this application embodiment can solve the technical problems of existing technologies that struggle to achieve multi-source heterogeneous data fusion and fail to consider the coupling effects between physical fields, resulting in a lack of specificity, accuracy, and reliability in defect detection results.

[0071] For easier understanding, please refer to Figure 2 This application provides an embodiment of a transformer core defect detection device, comprising: The data acquisition unit 201 is used to acquire multi-modal data during the operation of the transformer. The multi-modal data includes vibration signals, temperature field, gas concentration, and harmonic current. The feature fusion unit 202 is used to perform feature fusion processing on multimodal data according to modal weights and a preset cascaded fusion network to obtain a fused feature vector; The coupling analysis unit 203 is used to construct the spatial impedance distribution equation based on the fused eigenvectors and to perform iterative analysis of electromagnetic-thermal coupling effects to obtain the impedance distribution matrix. The defect detection unit 204 is used to perform defect detection analysis based on the fused feature vector and impedance distribution matrix according to the constrained reinforcement learning strategy and the preset core diagram model, and obtain the defect assessment level.

[0072] This application also provides a transformer core defect detection device, the device including a processor and a memory; The memory is used to store program code and transfer the program code to the processor; The processor is used to execute the transformer core defect detection method in the above method embodiment according to the instructions in the program code.

[0073] In the several embodiments provided in this application, it should be understood that the disclosed apparatus and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or units may be electrical, mechanical, or other forms.

[0074] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0075] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0076] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions for executing all or part of the steps of the methods described in the various embodiments of this application through a computer device (which may be a personal computer, server, or network device, etc.). The aforementioned storage medium includes: USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, optical disks, and other media capable of storing program code.

[0077] The above-described embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application.

Claims

1. A method for detecting defects in a transformer core, characterized in that, include: Multimodal data is collected during the operation of the transformer, including vibration signals, temperature field, gas concentration, and harmonic current. The multimodal data is subjected to feature fusion processing based on modal weights and a preset cascaded fusion network to obtain a fused feature vector; Based on the fused eigenvectors, a spatial impedance distribution equation is constructed, and an iterative analysis of the electromagnetic-thermal coupling effect is performed to obtain the impedance distribution matrix. Based on the constrained reinforcement learning strategy and the preset core diagram model, defect detection and analysis are performed according to the fused feature vector and the impedance distribution matrix to obtain the defect assessment level.

2. The method for detecting defects in transformer cores according to claim 1, characterized in that, The multimodal data collected during the operation of the transformer includes: The vibration spectrum of the transformer during operation is acquired by FFT according to a preset window length to obtain the vibration signal; The temperature field is determined by acquiring the spatial gradient of the infrared thermal image of the transformer during operation using a thermal imager. The concentration of dissolved gases in transformer oil was determined using a pre-set line chromatograph, and the gas concentration was obtained. Harmonic currents are obtained by measuring the current harmonics during the operation of the transformer.

3. The method for detecting transformer core defects according to claim 1, characterized in that, The step of performing feature fusion processing on the multimodal data based on modal weights and a preset cascaded fusion network to obtain a fused feature vector includes: The kernel density is estimated based on the multimodal data to obtain the mode weights; Based on the modal weights and the multimodal data, cascaded feature fusion calculations are performed using a preset cascaded fusion network to obtain a fused feature vector.

4. The method for detecting transformer core defects according to claim 1, characterized in that, The spatial impedance distribution equation is constructed based on the fused eigenvectors, and an iterative analysis of the electromagnetic-thermal coupling effect is performed to obtain the impedance distribution matrix, including: Based on the temperature field and harmonic current in the fused feature vector, Joule heating effect analysis and harmonic excitation analysis are performed, and a spatial impedance distribution equation is generated. Based on the spatial impedance distribution equation, an iterative analysis of the electromagnetic-thermal coupling effect based on a preset time step is performed to obtain the impedance distribution matrix.

5. The method for detecting defects in transformer cores according to claim 1, characterized in that, The constraint-based reinforcement learning strategy and the preset core diagram model perform defect detection analysis based on the fused feature vector and the impedance distribution matrix to obtain the defect assessment level, including: A three-dimensional mesh topology diagram of the transformer core is constructed to obtain a preset core diagram model; The fused feature vector, the impedance distribution matrix, and the vibration propagation function are input into the preset core diagram model for convolution calculation to obtain the anomaly probability map. Based on the constrained reinforcement learning strategy, defect detection analysis is performed according to the anomaly probability map to obtain the defect assessment level.

6. The method for detecting transformer core defects according to claim 5, characterized in that, The step of inputting the fused feature vector, the impedance distribution matrix, and the vibration propagation function into the preset core diagram model for convolution calculation to obtain an anomaly probability map includes: Construct an adjacency matrix based on the vibration propagation function; Calculate the temperature mask based on the temperature field gradient in the fused feature vector; The graph convolution features of the preset core diagram model are updated and calculated based on the adjacency matrix, the temperature mask, and the impedance distribution matrix to obtain an anomaly probability map.

7. The method for detecting defects in transformer cores according to claim 5, characterized in that, The constraint-based reinforcement learning strategy performs defect detection analysis based on the anomaly probability map to obtain a defect assessment level, including: Based on Fourier's law, a dual-constraint reward function is configured according to the gas change rate and heat flux. Based on the constrained reinforcement learning strategy, defect detection analysis based on action-state update operation is performed according to the anomaly probability map, the impedance distribution matrix and the dual-constraint reward function to obtain the defect assessment level.

8. The method for detecting defects in transformer cores according to claim 5, characterized in that, The constraint-based reinforcement learning strategy and the preset core diagram model perform defect detection analysis based on the fused feature vector and the impedance distribution matrix to obtain the defect assessment level, and then further include: The defect information is obtained by performing a joint inversion analysis based on the defect assessment level and the anomaly probability in the anomaly probability map using a physical information neural network. The defect information includes the defect location, defect area, and defect depth.

9. A transformer core defect detection device, characterized in that, include: The data acquisition unit is used to acquire multimodal data during the operation of the transformer, including vibration signals, temperature field, gas concentration, and harmonic current. The feature fusion unit is used to perform feature fusion processing on the multimodal data according to modal weights and a preset cascaded fusion network to obtain a fused feature vector; The coupling analysis unit is used to construct the spatial impedance distribution equation based on the fused feature vector, and to perform iterative analysis of electromagnetic-thermal coupling effect to obtain the impedance distribution matrix; The defect detection unit is used to perform defect detection analysis based on the fused feature vector and the impedance distribution matrix according to the constrained reinforcement learning strategy and the preset core diagram model, and to obtain the defect assessment level.

10. A transformer core defect detection device, characterized in that, The device includes a processor and a memory; The memory is used to store program code and transmit the program code to the processor; The processor is used to execute the transformer core defect detection method according to any one of claims 1-8 according to the instructions in the program code.

Images