An oil-immersed power transformer non-destructive testing optimization method and system

Abstract

The application relates to the technical field of transformer detection, and discloses an oil-immersed power transformer nondestructive detection optimization method and system, which establishes a three-dimensional dielectric response coordinate system to generate a defect preliminary positioning map; through spiral sweep frequency excitation detection, aging dominant type and dampness dominant type defects are recognized; based on the defect types, a temperature gradient excitation scheme is established to generate a defect degree quantitative evaluation index; an acoustic wave modulation excitation scheme is designed to generate a high-resolution defect characteristic spectrum; a multi-modal intelligent sensor network is constructed, a defect evolution prediction is realized in combination with a defect development trend prediction model, and a grading early warning signal is generated; the application realizes whole-process accurate detection of transformer insulation defects from positioning, classification, quantitative evaluation to evolution prediction, and provides a reliable basis for operation and maintenance.

Description

Technical Field

[0001] This invention relates to the field of transformer testing technology, and more specifically, to an optimized method and system for non-destructive testing of oil-immersed power transformers. Background Technology

[0002] As a core piece of equipment in the power system, the health of the insulation system of oil-immersed power transformers directly determines the stability and safety of the power grid operation. With increasing service life, transformer insulation is prone to defects due to aging, moisture, and other factors. Failure to accurately detect and assess these defects can lead to serious faults such as insulation breakdown. Therefore, developing efficient non-destructive testing technologies for insulation defects is of great significance for ensuring the reliable operation of the power system.

[0003] In the prior art, Chinese patent application CN116540040A discloses a method, apparatus, and storage medium for evaluating the insulation status of an oil-immersed power transformer under high-voltage excitation. The method obtains the dielectric response of the oil-paper insulation through a single triangular wave excitation, obtains multiple harmonics through spectral analysis, obtains the excitation and response signals under different harmonics through Fourier decomposition, calculates the loss factor based on the phase difference, extracts the aging determination coefficient, and fits its quantitative relationship with the aging degree of the oil-paper insulation. This method achieves rapid detection under high-voltage excitation and eliminates the influence of nonlinear insulation changes on the evaluation results. Chinese patent application CN113935370A discloses a method for identifying mechanical faults in a three-phase oil-immersed distribution transformer. It collects the initial vibration signal on the surface of the transformer tank, decomposes it using the CEEMDAN algorithm to obtain the modal components (IMF), reconstructs the signal and obtains the SDP image in polar coordinates through point symmetry transformation, and then obtains the typical fault binary matrix through a clustering algorithm and trains the RBF neural network to achieve the identification of mechanical faults. This method overcomes the problems of modal aliasing and low efficiency in traditional signal decomposition and avoids the influence of low-frequency interference from the transformer itself and high-frequency interference from the outside on the diagnostic results.

[0004] However, existing technologies still have significant limitations when transformer insulation systems suffer from both aging and moisture defects simultaneously. While Chinese patent application CN116540040A can assess the degree of aging through high-voltage excitation, it does not consider the impact of moisture on dielectric response. Aging increases the dielectric loss tangent at high frequencies, while moisture increases the dielectric constant at low frequencies. In the degenerate time constant band of 0.1Hz-10Hz, these two effects easily cancel each other out. Chinese patent application CN113935370A focuses on mechanical fault identification but cannot address the problem of cross-interference of dielectric parameters. In this case, a single dielectric parameter (such as loss factor or dielectric constant) cannot distinguish the dominant defect type. This may lead to severely aged transformers that should have their insulation replaced being misjudged as only requiring dehydration, resulting in continuous deterioration of insulation performance and subsequent breakdown faults. This can cause transformer outages or even large-scale power grid blackouts, seriously threatening power system safety. Summary of the Invention

[0005] This invention is applicable to the full lifecycle insulation condition detection of oil-immersed power transformers, including quality sampling inspection before new equipment leaves the factory, regular operation and maintenance monitoring of operating equipment, and accurate diagnosis of insulation defects after a fault. It is particularly suitable for assessing the insulation condition of transformers with long service lives that may simultaneously exhibit aging and moisture defects. To overcome the aforementioned shortcomings of existing technologies, this invention provides an optimized method and system for non-destructive testing of oil-immersed power transformers. By constructing a complete technical system encompassing three-dimensional dielectric response localization, multi-physics field coupling detection, and intelligent evolution prediction, it achieves accurate identification, quantitative assessment, and early warning of both aging and moisture defects. This effectively solves the problem of dielectric spectrum feature masking, provides a reliable basis for transformer insulation operation and maintenance, and significantly improves detection accuracy and power grid operation safety.

[0006] To achieve the above objectives, the present invention provides the following technical solution:

[0007] An optimized method for non-destructive testing of oil-immersed power transformers includes:

[0008] A three-dimensional dielectric response coordinate system is established and initial electric field distribution data is obtained. Based on the initial electric field distribution data, a preliminary defect location map is generated.

[0009] Based on the preliminary defect location map, spiral sweep frequency excitation detection is performed to identify the type of insulation defects inside the oil-immersed power transformer and determine the defect type; the defect type includes aging-dominated defects and moisture-dominated defects;

[0010] A temperature gradient excitation scheme is established based on the defect type, defect response is decoupled, and a quantitative evaluation index of defect degree is generated.

[0011] Based on the quantitative assessment index of defect degree, an acoustic wave modulation excitation scheme is designed, and the abnormal sound pressure attenuation region is identified by acoustic-electric coupling detection and a high-resolution defect feature map is generated.

[0012] A multimodal intelligent sensor network is constructed based on high-resolution defect feature maps to form a defect evolution feature fusion vector. The defect evolution is predicted based on the defect evolution feature fusion vector and a pre-constructed defect development trend prediction model. Based on the defect evolution prediction results, a graded early warning signal is generated.

[0013] Furthermore, the method for generating a preliminary defect location map based on the initial electric field distribution data includes:

[0014] In a three-dimensional dielectric response coordinate system, a regular hexagonal characteristic frequency detection region is defined and six characteristic frequency detection points are determined.

[0015] Based on the six identified characteristic frequency detection points, the initial dielectric response data and electric field gradient data of each monitoring point are synchronously collected by a MEMS electric field sensor network deployed on the inner wall of the oil tank, generating an initial dielectric response dataset. Based on the initial dielectric response dataset, a preliminary defect location map is generated, which marks different defect areas and defect location information.

[0016] Furthermore, the method for generating a preliminary defect location map based on the initial dielectric response dataset includes:

[0017] Based on the initial dielectric response dataset, calculate the vector angle and vector length from the dielectric response data point at each MEMS electric field sensor node location to the origin of the three-dimensional dielectric response coordinate system;

[0018] The electric field gradient magnitude between adjacent MEMS electric field sensor nodes is calculated based on the electric field gradient data, and effective detection areas with electric field gradient magnitude values ​​greater than the preset electric field gradient threshold are selected.

[0019] For the selected effective detection area, the existence judgment conditions of defects are established by combining vector angle, vector length and electric field gradient magnitude, and the defect area is determined.

[0020] The identified defect area is spatially marked in a three-dimensional dielectric response coordinate system to generate a preliminary defect location map containing defect location information.

[0021] Furthermore, the six vertices of the regular hexagonal characteristic frequency detection region correspond to six characteristic frequency points, namely, the first low-frequency band frequency point f1, the second low-frequency band frequency point f2, the first mid-frequency band frequency point f3, the second mid-frequency band frequency point f4, the first high-frequency band frequency point f5, and the second high-frequency band frequency point f6. The frequency range of f1-f2 is defined as the low-frequency band, the frequency range of f3-f4 is defined as the mid-frequency band, and the frequency range of f5-f6 is defined as the high-frequency band, wherein f1 < f2 < f3 < f4 < f5 < f6.

[0022] Furthermore, the method for identifying the type of insulation defects inside an oil-immersed power transformer includes:

[0023] A MEMS accelerometer array is deployed, and dielectric response data and vibration response data are synchronously acquired during the frequency sweep process through a MEMS electric field sensor network and the MEMS accelerometer array; the vibration envelope change rate is obtained based on the vibration response data.

[0024] Spectral analysis of the vibration response data yields the vibration resonance frequency data;

[0025] The spiral sweep frequency sequence is coupled with the vibration resonance frequency data for analysis, and the vibration frequency ratio, which reflects the degree of matching between the sweep frequency and the resonance frequency, is calculated.

[0026] Dielectric response time-domain features are extracted from dielectric response data during frequency sweeping. By combining dielectric response time-domain features, vibration envelope change rate, and vibration frequency ratio, the insulation defects inside oil-immersed power transformers are identified.

[0027] Furthermore, the method for identifying the type of insulation defects inside an oil-immersed power transformer by integrating the time-domain characteristics of the dielectric response, the rate of change of the vibration envelope, and the ratio of vibration frequencies includes:

[0028] Extract the rise time and fall time of each sweep cycle from the dielectric response data during the frequency sweep process, and calculate the ratio of rise time to fall time.

[0029] Based on the ratio of rise time to fall time, the rate of change of vibration envelope, and the ratio of vibration frequency, the defect type is preliminarily determined, and a preliminary defect determination result is obtained.

[0030] Statistical analysis is performed on the preliminary defect judgment results of multiple consecutive frequency sweep cycles. The final defect type is determined based on the frequency of occurrence of the defect type, and a defect classification label is generated.

[0031] Furthermore, the method for decoupling defect response and generating quantitative assessment indicators of defect severity includes:

[0032] Based on the determined defect type, a differential temperature gradient is applied to the upper and lower ends of the transformer tank to obtain three-dimensional temperature field dynamic distribution data, and the dielectric response data of each characteristic frequency point is monitored during the establishment of the temperature gradient to calculate the dielectric spectrum change rate in the low-frequency and high-frequency bands.

[0033] Based on the dielectric spectrum change rate and three-dimensional temperature field dynamic distribution data in the low-frequency and high-frequency bands, a quantitative assessment model for the degree of defect with temperature compensation is established, generating a quantitative assessment index for the degree of defect that includes aging degree index and moisture content index.

[0034] Furthermore, the method for identifying abnormal sound pressure attenuation regions and generating high-resolution defect feature maps through acoustic-electric coupling detection includes:

[0035] A MEMS acoustic pressure sensor array was deployed around the defect area inside the fuel tank. After the acoustic wave modulation excitation was activated, the acoustic pressure amplitude data and acoustic pressure phase data of each acoustic pressure sensor node were collected in real time to construct a three-dimensional sound field distribution map and identify the abnormal acoustic pressure attenuation area. For the abnormal acoustic pressure attenuation area, based on the three-dimensional sound field distribution map, dielectric spectrum measurements were performed under two states: with acoustic wave modulation and without acoustic wave modulation, to obtain dielectric response data under the two states.

[0036] Based on the dielectric response data under the two states, the dielectric spectrum difference between the two states is calculated and coupled with the acoustic pressure phase data for analysis. Based on the coupling analysis results of the dielectric spectrum difference and the acoustic pressure phase data, the acoustic excitation parameters are optimized and multi-frequency scanning is performed to generate a high-resolution defect feature map.

[0037] Furthermore, the frequency of the acoustically modulated sound wave is consistent with the vibration resonance frequency in the vibration resonance frequency data.

[0038] An optimization system for non-destructive testing of oil-immersed power transformers, used to implement the aforementioned optimization method for non-destructive testing of oil-immersed power transformers, the system comprising:

[0039] Defect location module: used to establish a three-dimensional dielectric response coordinate system and obtain initial electric field distribution data, and generate a preliminary defect location map based on the initial electric field distribution data;

[0040] Defect classification module: used to perform spiral sweep frequency excitation detection based on the preliminary defect location map, identify the defect type of insulation defects inside the oil-immersed power transformer, and determine the defect type; the defect type includes aging-dominated defects and moisture-dominated defects;

[0041] Defect quantitative assessment module: Establishes a temperature gradient excitation scheme based on defect type, decouples defect response, and generates quantitative assessment indicators of defect severity;

[0042] Defect Feature Map Construction Module: Based on the quantitative assessment index of defect degree, an acoustic wave modulation excitation scheme is designed, and the abnormal sound pressure attenuation region is identified through acoustic-electric coupling detection and a high-resolution defect feature map is generated;

[0043] The graded early warning module constructs a multimodal intelligent sensor network based on high-resolution defect feature maps, forming a defect evolution feature fusion vector. Based on the defect evolution feature fusion vector and a pre-constructed defect development trend prediction model, it predicts the defect evolution and generates graded early warning signals based on the defect evolution prediction results.

[0044] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0045] This invention achieves systematic non-destructive testing of insulation defects in oil-immersed power transformers by constructing a complete technical system encompassing three-dimensional dielectric response analysis, spiral frequency sweep excitation detection, temperature gradient excitation decoupling, acoustic-electric coupling feature map construction, and multimodal evolution prediction. First, by combining a three-dimensional dielectric response coordinate system with a sensor network, the method accurately locates potential defect areas, solving the problem of ambiguous defect locations in traditional detection. Then, through multi-parameter coupling analysis under spiral frequency sweep excitation, it covers the full range of features across low, mid, and high frequency bands, effectively distinguishing between aging-dominated and moisture-dominated defects, avoiding misjudgments caused by the mutual cancellation effect of single parameters in overlapping frequency bands. For example, to address the coupling interference between increased dielectric loss due to aging in the mid-frequency band and increased dielectric constant due to moisture in the low-frequency band, multi-band, multi-parameter collaborative analysis can accurately identify the dominant defect type, preventing severely aged transformers from being misjudged as requiring only dehydration treatment. The temperature gradient excitation scheme based on defect type further achieves effective decoupling of defect response. By quantitatively assessing the combined effects of aging and moisture absorption through the differentiated changes in dielectric parameters under different temperature gradients, it overcomes the qualitative limitations of traditional methods in assessing defect severity. Combined with high-resolution feature maps generated by acoustic modulation excitation, it can capture the unique responses of minute defects under acoustic-electric coupling, improving the identification accuracy of dual defects. Finally, through the fusion of a multimodal intelligent sensor network and a predictive model, it achieves the tracking and early warning of defect evolution trajectories, fundamentally solving the problem of feature masking in the dielectric spectrum caused by dual defects of aging and moisture absorption. This provides comprehensive technical support for the operation and maintenance of transformer insulation systems, from current status assessment to future trend prediction, significantly improving the accuracy, reliability, and foresight of detection, and ensuring the safe and stable operation of transformers. Attached Figure Description

[0046] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0047] Figure 1 This is a flowchart of an optimized non-destructive testing method for oil-immersed power transformers according to the present invention;

[0048] Figure 2 A flowchart of a method for generating a preliminary defect location map provided in an embodiment of the present invention;

[0049] Figure 3 A flowchart illustrating the method for generating quantitative assessment indicators of defect severity provided in this embodiment of the invention;

[0050] Figure 4 This is a schematic diagram of the defect development trend prediction model provided in an embodiment of the present invention;

[0051] Figure 5 This is a functional block diagram of an optimized non-destructive testing system for oil-immersed power transformers provided in an embodiment of the present invention. Detailed Implementation

[0052] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0053] Example 1:

[0054] Please see Figure 1 As shown, this embodiment provides an optimized method for non-destructive testing of oil-immersed power transformers, including:

[0055] Step S10: Establish a three-dimensional dielectric response coordinate system and obtain initial electric field distribution data; generate a preliminary defect location map based on the initial electric field distribution data.

[0056] Please see Figure 2 As shown, step S10 further includes:

[0057] Step S11: Establish a three-dimensional dielectric response coordinate system with the geometric center of the transformer tank as the origin, and deploy a MEMS electric field sensor network on the inner wall of the tank to obtain the initial electric field distribution data.

[0058] Step S12: Based on the collected initial electric field distribution data, delineate a regular hexagonal characteristic frequency detection region in the three-dimensional dielectric response coordinate system and determine six characteristic frequency detection points.

[0059] Step S13: Based on the six determined characteristic frequency detection points, the initial dielectric response data and electric field gradient data of each monitoring point are synchronously collected through the deployed MEMS electric field sensor network to generate the initial dielectric response dataset.

[0060] Step S14: Based on the initial dielectric response dataset, generate a preliminary defect location map, which marks different defect areas and defect location information.

[0061] Further, step S14 includes:

[0062] Step S141: Based on the initial dielectric response dataset, calculate the vector angle and vector length from the dielectric response data point at each MEMS electric field sensor node location to the origin of the three-dimensional dielectric response coordinate system; the dielectric response data point refers to the spatial coordinate point in the three-dimensional dielectric response coordinate system composed of the logarithm of the frequency at a certain measurement time, the dielectric constant, and the tangent of the dielectric loss angle.

[0063] Step S142: Calculate the electric field gradient magnitude between adjacent MEMS electric field sensor nodes based on the electric field gradient data, and filter out effective detection areas where the electric field gradient magnitude is greater than the preset electric field gradient threshold.

[0064] Step S143: For the selected effective detection area, establish the defect existence judgment conditions by comprehensively considering the vector angle, vector length and electric field gradient magnitude, and determine the defect area;

[0065] Step S144: Spatial marking of the determined defect area in the three-dimensional dielectric response coordinate system to generate a preliminary defect location map containing defect location information.

[0066] Specifically, a three-dimensional dielectric response coordinate system is established with the geometric center of the transformer tank as the origin. The X-axis represents the logarithmic frequency, the Y-axis represents the dielectric constant, and the Z-axis represents the dielectric loss tangent. This three-dimensional coordinate system integrates frequency domain parameters and electrical parameters within the same spatial framework, giving spatial correlation to the previously dispersed dielectric response data. Simultaneously, a three-dimensional physical coordinate system is established with the geometric center of the transformer tank as the origin. A MEMS electric field sensor network is deployed on the inner wall of the tank, with each sensor corresponding to a unique spatial coordinate within the tank. The distribution density of the sensor nodes is determined based on the tank's geometric dimensions. For example, in a 1000kVA transformer tank, one sensor is placed every 30 cm circumferentially along the inner wall and one sensor every 50 cm axially, forming a uniformly covered monitoring network. This monitoring network synchronously collects initial data on the electric field strength, dielectric constant, and dielectric loss tangent as a function of frequency at different locations—that is, the initial electric field distribution data. A one-to-one mapping is established between the three-dimensional dielectric response coordinate system and the three-dimensional physical coordinate system through the sensor node IDs. The physical coordinates of each MEMS electric field sensor and the dielectric response data it collects form an associated array, realizing real-time mapping between physical space and dielectric parameter space.

[0067] The initial electric field distribution data is collected by each sensor node. This data not only contains the correlation information between dielectric parameters and frequency, but also includes the spatial location identifier of the acquisition node. This allows the data to be back-mapped to the specific physical location inside the oil tank when analyzing the data in the three-dimensional dielectric response coordinate system. Based on the collected initial electric field distribution data, a regular hexagonal characteristic frequency detection region is delineated in the three-dimensional dielectric response coordinate system. The diameter of the circumscribed circle of the regular hexagonal characteristic frequency detection region is determined according to the influence range of common transformer defects, for example, set to 50 cm. The six vertices correspond to six characteristic frequency points, which include two low-frequency band frequency points f1 and f2 (e.g., the first low-frequency band frequency point f1 is 0.001 Hz and the second low-frequency band frequency point f2 is 0.01 Hz), two mid-frequency band frequency points f3 and f4 (e.g., the first mid-frequency band frequency point f3 is 0.1 Hz and the second mid-frequency band frequency point f4 is 1 Hz), and two high-frequency band frequency points f5 and f6 (e.g., the first high-frequency band frequency point f5 is 10 Hz and the second high-frequency band frequency point f6 is 100 Hz), where f1 < f2 < f3 < f4 < f5 < f6. The frequency range of f1-f2 is defined as the low-frequency band, the frequency range of f3-f4 is defined as the mid-frequency band, and the frequency range of f5-f6 is defined as the high-frequency band. As shown in Table 1, the low-frequency band is used to capture the dielectric constant variation characteristics of moisture-dominated defects in the low-frequency band, reflecting the influence of moisture-dominated defects; the mid-frequency band is used to reflect the coupling effect of aging and moisture; and the high-frequency band is used to extract the variation law of dielectric loss tangent value of aging-dominated defects in the high-frequency band, reflecting the characteristics of aging-dominated defects.

[0068] Table 1: Correspondence Table of Characteristic Frequency Points and Corresponding Frequency Bands, and Defect Types

[0069]

[0070]

[0071] The six frequency points are synchronously monitored using a MEMS electric field sensor network to acquire initial dielectric response data and electric field gradient data for each monitoring point, generating an initial dielectric response dataset. Based on this dataset, the vector angle and length from the dielectric response data point at each electric field sensor node to the origin of the coordinate system are calculated. The vector angle is determined by the angle between the spatial vector formed by the X, Y, and Z axes and each coordinate axis, and the vector length is the magnitude of this spatial vector, calculated as the square root of the sum of the squares of the parameters. The frequency logarithm is obtained by transforming the natural logarithm of the measured frequency, while the dielectric constant and dielectric loss tangent are obtained directly from the sensors. The electric field gradient magnitude between adjacent sensor nodes is calculated based on the electric field gradient data. This magnitude is determined by the ratio of the difference in electric field strength between adjacent nodes to the node spacing. A preset electric field gradient threshold is obtained through statistical analysis of the electric field gradient distribution of 100 normally operating transformers, for example, set to 0.5 kV / m·cm. Regions with electric field gradient magnitudes greater than the threshold are selected as effective detection areas, which are considered to contain potential defects. For the effective detection area, the existence judgment conditions of defects are established by combining the vector angle, vector length and electric field gradient magnitude. For example, when the vector angle is in a specific range (such as the angle with the X-axis is 30°-60°), the vector length exceeds 1.5 times the normal range, and the electric field gradient magnitude is continuously greater than the electric field gradient threshold within a set time period, it is determined that there is a defect area. This defect area is marked in the three-dimensional dielectric response coordinate system to form a preliminary defect location map containing the defect location coordinates and preliminary features.

[0072] The construction of a three-dimensional dielectric response coordinate system enables spatial mapping of dielectric parameters and frequency domain characteristics, solving the problems of parameter dispersion and lack of spatial correlation in traditional dielectric spectrum measurements. This allows for comparative analysis of defect responses at different locations within the same framework. The deployment of a MEMS electric field sensor network eliminates spatial blind spots. Through high-density node coverage, it ensures that electric field changes in all areas inside the transformer can be captured, preventing local defects from being masked by the overall signal. For example, in areas difficult to cover with traditional single-point measurements, such as corners of the tank, the collaborative monitoring of surrounding sensors can identify dielectric constant anomalies caused by localized moisture. The design of the regular hexagonal detection area utilizes its geometric symmetry to evenly distribute six characteristic frequency points in the frequency domain, avoiding frequency coverage bias. Simultaneously, the separation of low-frequency and high-frequency points allows for the separate capture of characteristics of moisture- and aging-dominated defects, reducing the limitations of single-frequency band measurements. The calculation of vector angles and lengths provides quantitative spatial characteristics of defects. Combined with the filtering of electric field gradient magnitudes, this improves the accuracy of defect identification. For example, changes in vector angles can distinguish the dielectric response direction of different types of defects, and the magnitude of the vector length reflects the severity of the defect. The generation of the preliminary defect location map enables spatial visualization of the defect, providing a clear target area for subsequent inspection and avoiding the problem of blindly opening the box in traditional inspection. At the same time, the defect location information contained in the location map allows the subsequent spiral frequency sweep excitation to be specifically focused on the target area, improving inspection efficiency.

[0073] Step S20: Based on the preliminary defect location map, perform spiral sweep frequency excitation detection to identify the type of insulation defects inside the oil-immersed power transformer and determine the defect type.

[0074] Further, step S20 includes:

[0075] Step S21: Based on the different defect regions marked in the preliminary defect location map and the defect location information, design a targeted spiral sweep frequency excitation signal and deploy a MEMS accelerometer array; the spiral sweep frequency excitation signal includes at least a spiral sweep frequency sequence;

[0076] Step S22: Based on the designed targeted spiral frequency sweep excitation signal, perform frequency sweep excitation, and synchronously collect dielectric response data and vibration response data during the frequency sweep process through MEMS electric field sensor network and MEMS accelerometer array; obtain the vibration envelope change rate based on the vibration response data;

[0077] Specifically, when designing targeted spiral frequency sweep excitation signals based on the different defect regions and location information marked in the preliminary defect location map, the starting and ending points of the spiral frequency sweep are defined using the defect location coordinates as a reference. The design of the spiral frequency sweep sequence is based on the characteristic frequency range that the defect may involve, covering the low-frequency band (f1-f2), the mid-frequency band (f3-f4), and the high-frequency band (f5-f6). Among them, the low-frequency band corresponds to the polarization characteristics of moisture-dominated defects, and the mid-to-high-frequency band corresponds to the relaxation characteristics of aging-dominated defects. The frequency change trajectory of the spiral frequency sweep adopts the Archimedean spiral mathematical model, which is expressed as f(t) = f0 + k × θ(t), where f(t) is the excitation frequency at time t, f0 is the starting frequency, k is the frequency change coefficient, and θ(t) is the polar angle at time t. θ(t) increases linearly with time, causing the frequency to rise or fall in a spiral manner over time. The value of the frequency variation coefficient k is determined based on the estimated size of the defect area. For larger defect areas (e.g., diameter exceeding 50cm), k is set to 0.1Hz / rad to ensure coverage of a wider frequency range within the same time period. For smaller defect areas (e.g., diameter less than 20cm), k is set to 0.05Hz / rad to improve the sweep accuracy of specific frequency bands. The deployment of the MEMS accelerometer array must correspond to the defect areas in the preliminary defect location map to ensure that vibration signals from each defect area can be collected.

[0078] When performing frequency sweep excitation based on the designed targeted spiral sweep frequency excitation signal, the spiral sweep frequency signal needs to be converted into a voltage signal by a signal generator, amplified by a power amplifier, and then applied to the high-voltage winding of the transformer. The collected dielectric response data includes the real-time changes of dielectric constant and dielectric loss tangent at different frequencies. The vibration response data includes vibration displacement, velocity, and acceleration time-domain signals. The vibration envelope change rate is obtained by performing a Hilbert transform on the vibration acceleration waveform. Specifically, the Hilbert transform is performed on the acceleration time-domain signal to obtain an analytical signal. The amplitude of the analytical signal is extracted as the envelope, and then the vibration envelope change rate is obtained by calculating the ratio of the difference in envelope amplitude between adjacent sampling points to the time difference.

[0079] Step S23: Based on the dielectric response data and vibration response data during the frequency sweep process, identify the type of insulation defects inside the oil-immersed power transformer.

[0080] Further, step S23 includes:

[0081] Step S231: Perform spectral analysis on the vibration response data to obtain the vibration resonance frequency data;

[0082] Step S232: Couple the spiral sweep frequency sequence with the vibration resonance frequency data for analysis, and calculate the vibration frequency ratio, which reflects the degree of matching between the sweep frequency and the resonance frequency.

[0083] The acceleration time-domain signal is converted to the frequency domain using a Fast Fourier Transform (FFT). Peak points in the spectrum with amplitudes exceeding three times the average noise level are identified to determine the vibration resonance frequency data. When coupling the helical sweep frequency sequence with the vibration resonance frequency data for analysis, frequency points of the same order of magnitude as the vibration resonance frequency are first extracted from the helical sweep frequency sequence. For example, when the vibration resonance frequency is 500Hz, frequency points in the 450-550Hz range of the sweep sequence are selected as the analysis objects. The vibration frequency ratio is calculated as the ratio of a frequency value in the helical sweep frequency sequence to the corresponding vibration resonance frequency value. For each resonance frequency, its ratio to the five adjacent frequency points in the sweep sequence is calculated, and the average value is taken as the vibration frequency ratio corresponding to that resonance frequency. The vibration frequency ratio reflects the degree of matching between the sweep frequency and the resonance frequency. When the ratio is close to 1, it indicates that the sweep frequency and the resonance frequency are coupled, and the dielectric response signal at this time contains richer defect feature information. The vibration frequency ratio ranges of aging-dominated and moisture-dominated defect regions are different. This difference stems from the fact that the decrease in cellulose polymerization caused by aging reduces the mechanical stiffness of the oil-paper interface and lowers the resonance frequency, while the aggregation of water molecules caused by moisture increases the interface damping and raises the resonance frequency. Therefore, the vibration frequency ratio can be used as a characteristic parameter to distinguish defect types.

[0084] Step S233: Extract the dielectric response time-domain features from the dielectric response data during the frequency sweep process. Combine the dielectric response time-domain features, vibration envelope change rate, and vibration frequency ratio to identify the insulation defects inside the oil-immersed power transformer, determine the defect type, and generate a defect classification label.

[0085] Further, step S233 includes:

[0086] Step S2331: Extract the rise time and fall time of each sweep cycle from the dielectric response data during the frequency sweep process, and calculate the ratio of rise time to fall time.

[0087] Step S2332: Based on the ratio of rise time to fall time, the rate of change of vibration envelope, and the ratio of vibration frequency, a preliminary determination of the defect type is made, and a preliminary defect determination result is obtained.

[0088] Step S2333: Perform statistical analysis on the preliminary defect judgment results of multiple consecutive frequency sweep cycles, determine the final defect type based on the frequency of defect type occurrence, and generate defect classification labels.

[0089] Specifically, the rise time and fall time of each sweep cycle are extracted from the dielectric response data during the frequency sweep process. The rise time refers to the time required for the dielectric response signal to rise from 10% to 90% of the steady-state value, and the fall time refers to the time required for the signal to fall from 90% to 10% of the steady-state value. The ratio R of the rise time to the fall time is obtained by the ratio of the absolute difference between the two to the average value. The ratio R can quantify the asymmetry of the dielectric response. The aging-dominated defect has a stronger dielectric relaxation asymmetry due to the breakage of cellulose molecular chains, so the R value of the aging-dominated defect is larger than that of the moisture-dominated defect. The moisture-dominated defect has a smaller R value due to the symmetry characteristics of water molecule polarization.

[0090] A three-dimensional scatter plot clustering method is used to make a preliminary judgment on the defect type. The comparison value R, the rate of change of the vibration envelope, and the ratio of vibration frequency are normalized by min-max and mapped to the 0-1 interval. A three-dimensional coordinate system of medial vibration characteristics is constructed with the ratio R as the X' axis, the rate of change of the vibration envelope as the Y' axis, and the ratio of vibration frequency as the Z' axis. The detection data of each defect region forms a data point in the three-dimensional coordinate system of medial vibration characteristics. The K-means clustering algorithm was used to train a large amount of historical detection data to obtain the cluster centers and cluster boundaries of aging-dominated and moisture-dominated defects. The cluster centers of aging-dominated defects are usually characterized by a large R value, a low rate of change of vibration envelope (due to slower mechanical vibration decay caused by aging), and a small vibration frequency ratio (due to the decrease in resonant frequency caused by aging, the ratio of sweep frequency to resonant frequency is small). The cluster centers of moisture-dominated defects are characterized by a small R value, a high rate of change of vibration envelope (due to the increased damping caused by moisture, vibration decay is accelerated), and a large vibration frequency ratio (due to the increase in resonant frequency caused by moisture, the ratio of sweep frequency to resonant frequency is large). For example, in a historical dataset, the cluster center coordinates of aging-dominant defects are (0.6, 0.2, 0.8), and the cluster center coordinates of moisture-dominant defects are (0.2, 0.6, 1.2). The cluster boundary is determined by calculating the Euclidean distance between the two cluster centers and taking the midpoint. When the Euclidean distance between the newly detected data point (0.5, 0.3, 0.9) and the aging-dominant cluster center is less than the distance with the moisture-dominant cluster center, it is initially determined to be an aging-dominant defect; otherwise, it is initially determined to be a moisture-dominant defect.

[0091] When statistically analyzing the preliminary defect assessment results of multiple consecutive frequency sweep cycles, the number of statistical cycles is set to N (N is determined based on the stability of the transformer's operating state, usually 5-10 cycles; a larger value is used for transformers with stable operation, and a smaller value is used for transformers with recent load fluctuations). The number of times aging-dominant and moisture-dominant defects appear in the preliminary assessment results of each cycle is counted. When the proportion of aging-dominant defects exceeds 50%, the defect is ultimately classified as aging-dominant and a corresponding tag is generated; otherwise, it is classified as moisture-dominant and a tag is generated. For example, in a statistical cycle of N=6, if 4 cycles are initially classified as aging-dominant and 2 cycles as moisture-dominant, accounting for 67% (more than 50%), then the defect is ultimately determined to be aging-dominant.

[0092] Step S20 addresses the vibration-electric coupling resonance mismatch problem. Specifically, the nonlinear coupling between the micro / nano-level mechanical vibration and dielectric response at the oil-paper interface leads to resonance mismatch when the sweep rate and the natural vibration frequency are mismatched, resulting in the loss of key defect characteristic information. A spiral sweep excitation signal based on the Archimedes spiral model is designed for the defect region. The frequency variation coefficient k is dynamically adjusted to ensure the sweep frequency traverses the possible natural vibration frequency range of the defect, avoiding the matching lag of traditional linear sweeps. Vibration response data is collected using a MEMS accelerometer array, and peak points with amplitudes exceeding three times the average noise are extracted using Fourier transform to determine the defect vibration resonance frequency (reflecting inherent vibration characteristics). The ratio of the sweep frequency to the resonance frequency is calculated; a ratio close to 1 indicates coupling, achieving efficient conversion of excitation energy into defect mechanical vibration. Statistical analysis of the results over 5-10 consecutive sweep cycles smooths instantaneous interference and ensures coupling stability. These methods, through dynamic sweep tracking, resonance frequency localization, coupling quantization, and multi-cycle stability assurance, avoid resonance mismatch caused by the mismatch between the sweep rate and the natural vibration frequency, ensuring complete capture of defect characteristic information.

[0093] Step S20, by extracting the time-domain asymmetric feature of the dielectric response, namely the ratio R, compensates for the shortcomings of traditional frequency-domain analysis in distinguishing defect types. The ratio of the rise time to the fall time can quantify the differences in dielectric relaxation dynamics caused by different defects. This is because the cellulose structure damage caused by aging leads to a stronger nonlinearity in the rise and fall processes of the dielectric response, while the water molecule polarization caused by moisture is closer to a symmetrical relaxation process. The combination of this feature with vibration parameters enables multi-physics cross-verification of electrical and mechanical signals, avoiding misjudgments caused by fluctuations in a single physical field signal. The three-dimensional scatter plot clustering method utilizes the spatial distribution characteristics of multiple parameters to solve the problem of mutual cancellation of single parameters in overlapping frequency bands. This allows for the distinction of defect types based on differences in parameter combinations even in the time constant degeneracy band of f3-f5 (e.g., 0.1Hz-10Hz). Statistical analysis over multiple consecutive cycles reduces the impact of instantaneous interference on the judgment results. For example, short-term load changes in transformers may cause abnormal dielectric response in a certain cycle. Multi-cycle statistics can smooth out such interference and improve the stability of the judgment. The generated defect classification labels provide a targeted basis for the temperature gradient excitation scheme in step S30, enabling subsequent temperature compensation and defect severity assessment to focus on the dominant defect type, avoiding resource waste and accuracy degradation caused by indiscriminate excitation. Without step S20, relying solely on a single parameter or measurement result for defect type determination would lead to an inability to distinguish between aging and moisture-dominant defects in overlapping frequency bands. This would render the temperature gradient excitation scheme in step S30 ineffective, failing to achieve effective decoupling of defect responses, ultimately resulting in inaccurate defect severity assessment and impacting subsequent maintenance strategy formulation. Furthermore, quantitative analysis and statistical verification ensure that the defect type determination process is repeatable and traceable, meeting the objectivity and reliability requirements of non-destructive testing methods. Combined with spatial location information acquired by the MEMS sensor network, it can also distinguish defect types at different locations, solving the problem of misclassification caused by the superposition of defect signals at different locations in traditional methods. This lays the foundation for accurate defect localization and classification.

[0094] Step S30: Establish a temperature gradient excitation scheme based on defect classification, decouple the defect response, and generate a quantitative evaluation index of the defect degree.

[0095] Further, step S30 includes:

[0096] Step S31: Based on the determined defect type, apply a differentiated temperature gradient to the upper and lower ends of the transformer tank to establish a controllable temperature excitation field.

[0097] The application of the differential temperature gradient needs to adjust the temperature range and gradient slope according to the type of defect. For aging-dominated defects, due to the poor thermal stability of aging products, such as small molecule compounds generated by cellulose degradation, the temperature range is set as the interval formed by the high-temperature end T1 and the low-temperature end T2. Among them, T1 is in the sensitive interval 5-10 °C below the thermal aging critical temperature of the oil-paper insulation material, and the thermal aging critical temperature is determined by thermogravimetry-differential scanning calorimetry (TG-DSC); T2 is close to the ambient temperature. By moderately heating T1, the dielectric relaxation time constant of the aging product changes significantly, enhancing its dielectric response characteristics in the high-frequency band; the gradient slope of the aging-dominated defect is set as k1, representing the temperature change rate, and the value of k1 is determined through the thermal shock resistance experiment of the insulation material. In the experiment, the dielectric parameter stability of the aged insulation under different heating rates is simulated, and the maximum rate that makes the parameter fluctuation less than 5% is selected to avoid the accelerated decomposition of the aging product caused by a sudden temperature rise. For moisture-dominated defects, because the diffusion coefficient of water molecules in the temperature range [T3, T4] is more sensitive to temperature changes (T3 < T1, T4 < T2), the high-temperature end T3 is lower than T1 of the aging-dominated defect, and the low-temperature end T4 is slightly lower than the ambient temperature, forming a temperature gradient that is more conducive to the directional migration of water molecules at the oil-paper interface; the gradient slope of the moisture-dominated defect is set as k2 (k2 > k1), and the dielectric constant fluctuation brought by moisture migration is strengthened through a faster temperature change rate. The value of k2 is based on the moisture diffusion kinetics experiment, and the rate at which the ratio of the moisture migration distance to time reaches the maximum and there is no insulation damage is selected. The determination experiments of the temperature difference ΔT1 = T1 - T2 set for the aging-dominated defect and the temperature difference ΔT2 = T3 - T4 set for the moisture-dominated defect are different: ΔT1 is obtained by measuring the change in the tangent value of the dielectric loss angle in the high-frequency band of the aged sample at different temperature differences, and the temperature difference with the largest change is selected; ΔT2 is obtained by measuring the change rate of the dielectric constant in the low-frequency band of the moisture-containing sample at different temperature differences, and the temperature difference with the largest change rate is selected. Both need to meet the condition that the thermal weight loss rate of the insulation material at this temperature difference is lower than 1%, and the thermal weight loss rate is determined through a thermogravimetric analysis experiment to ensure no obvious material damage. The application of the temperature gradient is realized through the heating tape wound around the outer wall of the oil tank (covering the high-temperature end area) and the cooling coil at the bottom (covering the low-temperature end area). The power of the heating tape and the cooling coil is adjusted by a PID controller. The proportional coefficient, integral time, and differential time of the controller are tuned according to the temperature response characteristics of the two types of defects respectively: the controller parameters for the aging-dominated defect are aimed at a temperature overshoot less than 5%, and the parameters for the moisture-dominated defect are aimed at the shortest temperature response time, ensuring that the temperature change rate is stable at the preset k1 or k2.

[0098] Step S32, based on the established controllable temperature excitation field, deploy a three-dimensional network of temperature sensors inside the oil tank to obtain real-time dynamic distribution data of the three-dimensional temperature field;

[0099] The deployment of the three-dimensional temperature sensor network corresponds one-to-one with the spatial coordinates of the MEMS electric field sensor network in step S10. Each sensor node has a built-in clock module that synchronously collects temperature values ​​and timestamps. The optical signal is converted into an electrical signal by a distributed fiber optic demodulator and transmitted to the data processing unit in real time to generate three-dimensional temperature field dynamic distribution data. This data uses spatial coordinates (x, y, z) and time t as indices to store the temperature value T(x, y, z, t) at the corresponding location, providing a temperature reference spatially associated with the dielectric response data for subsequent temperature compensation.

[0100] Step S33: Based on the three-dimensional temperature field dynamic distribution data, monitor the dielectric response data of each characteristic frequency point during the temperature gradient establishment process, and calculate the dielectric spectrum change rate of the low-frequency and high-frequency bands.

[0101] The calculation of the dielectric spectrum change rate is performed on the six characteristic frequency points determined in step S12, grouped into low-frequency bands (f1, f2) and high-frequency bands (f5, f6). The low-frequency band corresponds to the dipole polarization response of moisture-dominated defects, and the high-frequency band corresponds to the electronic relaxation response of aging-dominated defects. The calculation method is as follows: During the establishment of the temperature gradient, the dielectric constant ε and the dielectric loss tangent tanδ are recorded at each characteristic frequency point at time intervals Δt, the value of which is determined according to the temperature field change rate, ensuring that the spatial distribution difference of the temperature field at each interval Δt does not exceed 5%. Using the dielectric constants ε2 and ε1 recorded in two adjacent records, and the dielectric loss tangents tanδ2 and tanδ1 recorded in two adjacent records, the dielectric constant change rate (ε2-ε1) / Δt and the dielectric loss change rate (tanδ2-tanδ1) / Δt are calculated. The two together constitute the dielectric spectrum change rate. This frequency band calculation method utilizes the characteristics that the dielectric parameters of aging-dominated defects are more sensitive to temperature in the high-frequency band, while the dielectric parameters of moisture-dominated defects are more sensitive to temperature in the low-frequency band, providing a quantitative basis for the subsequent decoupling of the responses of the two types of defects.

[0102] Step S34: Based on the dielectric spectrum change rate and three-dimensional temperature field dynamic distribution data of the low-frequency and high-frequency bands, establish a temperature-compensated quantitative assessment model for the degree of defects and generate quantitative assessment indicators for the degree of defects.

[0103] Please see Figure 3 As shown, step S34 further includes:

[0104] Step S341: Based on the three-dimensional temperature field dynamic distribution data, calculate the dielectric response offset of each characteristic frequency point after the temperature gradient stabilizes and construct the offset spatial distribution map.

[0105] Step S342: Based on the dielectric spectrum change rates of the low-frequency and high-frequency bands, calculate the aging response rate index and the moisture response rate index respectively to achieve decoupling of the response rates of the two types of defects.

[0106] Step S343: Based on dielectric response offset, aging response rate index and moisture response rate index, establish quantitative assessment models for aging degree and moisture content respectively.

[0107] Step S344: Based on the three-dimensional temperature field dynamic distribution data and offset spatial distribution map, calculate the dynamic temperature compensation coefficient and humidity correction factor, and optimize the aging degree quantitative assessment model and moisture content quantitative assessment model for temperature compensation.

[0108] Step S345: Combine the quantitative assessment results of aging degree after temperature compensation and the quantitative assessment results of moisture content to generate a quantitative assessment index of defect degree that includes aging degree index and moisture content index.

[0109] Specifically, step S341, based on the dynamic distribution data of the three-dimensional temperature field, determines the stabilization time t0 after the temperature gradient stabilizes through temperature field convergence analysis. Stability is defined as the temperature change at each point within 10 consecutive Δt intervals being less than 0.5℃. The dielectric response offset at each characteristic frequency point is then calculated, i.e., the difference between the current dielectric parameters (εt, tanδt) and the baseline values ​​(ε0, tanδ0) before the temperature gradient was applied: Δε = εt - ε0, Δtanδ = tanδt - tanδ0. Δε is the dielectric constant offset, and Δtanδ is the dielectric loss offset. Δε and Δtanδ together constitute the dielectric response offset (Δε, Δtanδ). The spatial distribution map of the offset is generated by mapping the Δε and Δtanδ of each sensor node to a three-dimensional physical coordinate system. An interpolation algorithm is used to supplement the offset at locations where no sensors are deployed, forming a continuous spatial distribution feature map that visually demonstrates the spatial heterogeneity of the dielectric response affected by temperature. In step S342, the dielectric spectrum change rates of the low-frequency and high-frequency bands are normalized. The aging response rate index is obtained by weighted summation of the dielectric spectrum change rates of the high-frequency bands (f5, f6). The weights w1 and w2 (w1+w2=1) of f5 and f6 are determined by fitting experimental data from multiple aging samples. During the fitting process, the correlation between the high-frequency response rate of samples with known aging degree and the actual aging degree is the target, and w1 and w2 are adjusted to maximize the correlation coefficient. The moisture response rate index is obtained by weighted summation of the dielectric spectrum change rates of the low-frequency bands (f1, f2). The weights w3 and w4 (w3+w4=1) of f1 and f2 are determined by fitting experimental data from moisture samples. The optimization is carried out with the correlation between the low-frequency response rate of samples with known moisture degree and the actual moisture content as the target. By assigning weights, the response characteristics of different frequency bands to specific defects are strengthened, thereby decoupling the response rates of the two types of defects.

[0110] In step S343, the aging response rate index, high-frequency dielectric response offset, temperature difference ΔT1 set for aging-dominant defects, temperature difference ΔT2 set for moisture-dominant defects, low-frequency dielectric response offset, and moisture response rate index are first processed to be dimensionless. The quantitative assessment model of aging degree uses the aging response rate index as the core variable and combines the high-frequency dielectric response offset and temperature gradient value to construct a multiple linear regression model. The expression is: Aging degree index = a × aging response rate index + b × high-frequency dielectric response offset + c × ΔT1, where a, b, and c are regression coefficients. The moisture content is determined by least-squares fitting of historical data from multiple transformers with known aging degrees (based on cellulose polymerization degree). During the fitting process, the mean square error between the model's predicted value and the actual aging degree is calculated, and the coefficients are iteratively adjusted until the mean square error is less than a preset threshold. The quantitative assessment model expression for moisture content is: Moisture content index = d × Moisture response rate index + e × Low-frequency dielectric response offset + f × ΔT2, where d, e, and f are regression coefficients, obtained by fitting test data of insulation samples with different moisture contents. The fitted samples cover the dielectric response characteristics of oil-paper insulation materials in the 1%-4% moisture content range. In step S344, the dynamic temperature compensation coefficient T* is calculated based on the deviation between the temperature value T at each point in the three-dimensional temperature field and the reference temperature T0 (e.g., 25℃), with the formula: T* = 1 + k' × (T - T0), where k' is the temperature coefficient, determined through temperature characteristic experiments of oil-paper insulation materials. In the experiment, the rate of change of dielectric parameters at different temperatures is measured, and the average value is taken as the value of k', reflecting the linear change trend of dielectric parameters with temperature.

[0111] The humidity correction factor is determined based on the spatial distribution gradient of the low-frequency band offset in the offset spatial distribution map. When the spatial gradient of the low-frequency band dielectric response offset (the ratio of the offset difference between adjacent points to the distance) in a certain area is greater than the threshold θ', the correction factor is set to k'1; otherwise, it is set to k'2. θ', k'1, and k'2 are determined through experiments on the impact of uneven moisture distribution on the evaluation results. θ' is the critical value of moisture distribution uniformity, and k'1 and k'2 are the correction coefficients for uneven and uniform moisture distribution, respectively, to correct the evaluation bias caused by the difference in spatial moisture distribution.

[0112] For the quantitative assessment model of aging degree, the dynamic temperature compensation coefficient corrects the high-frequency dielectric response offset. The original high-frequency dielectric response offset is multiplied by the dynamic temperature compensation coefficient to obtain the temperature-corrected high-frequency dielectric response offset. The correction formula is: Corrected high-frequency dielectric response offset = High-frequency dielectric response offset × Dynamic temperature compensation coefficient. This is because the dielectric properties of aged products exhibit a linear trend affected by temperature. For every 1°C deviation from the reference temperature, the rate of change of the high-frequency dielectric parameters is quantified by the temperature coefficient k'. The temperature compensation coefficient can offset this linear drift. The influence of the humidity correction factor on the aging degree model is achieved through cross-correction. Although aging-dominant defects are mainly affected by temperature, high humidity environments may accelerate the aging process. Therefore, the humidity correction factor is used as a weighting coefficient multiplied by the temperature-compensated model output, i.e.: Final aging degree index = (a × Aging response rate index + b × Corrected high-frequency dielectric response offset + c × ΔT1) × Humidity correction factor. When the humidity gradient is large, the weight of the aging degree index is appropriately reduced to avoid misjudging dielectric anomalies caused by moisture as accelerated aging.

[0113] For the quantitative assessment model of moisture content, the dynamic temperature compensation coefficient corrects the low-frequency dielectric response offset. The correction formula is: Corrected low-frequency dielectric response offset = Low-frequency dielectric response offset × Dynamic temperature compensation coefficient. The polarization intensity of water molecules decreases with increasing temperature, and the temperature compensation coefficient can correct this temperature dependence. The humidity correction factor directly acts on the corrected low-frequency offset. The formula is: Humidity-corrected low-frequency dielectric response offset = Corrected low-frequency dielectric response offset × Humidity correction factor. When the spatial gradient of the low-frequency dielectric response offset in a certain area is greater than the threshold θ', i.e., the moisture distribution is uneven, the correction factor k'1 can amplify the offset and highlight the response characteristics of local high-humidity areas. When the spatial gradient is less than or equal to θ', i.e., the moisture distribution is uniform, the correction factor k'2 (equal to 1) keeps the offset unchanged to avoid over-correction. The final moisture content index = d × Moisture response rate index + e × Humidity-corrected low-frequency dielectric response offset + f × ΔT2, which better reflects the actual spatial distribution of moisture. Step S345 performs a weighted summation of the aging degree index and the moisture content index to generate a quantitative assessment index of the defect degree that includes both the aging degree index and the moisture content index.

[0114] Step S30 addresses the problem that traditional temperature compensation methods cannot handle nonlinear drift in dielectric response and thermal diffusion boundary effects. For nonlinear drift in dielectric response, differentiated temperature gradients are applied based on defect classification labels (e.g., a specific temperature range and slope for aging-dominant defects, and a different set of parameters for moisture-dominant defects) to amplify the dielectric response characteristics of different defects. Dynamic temperature compensation coefficients are calculated using three-dimensional temperature field data to correct the nonlinear changes in dielectric parameters with temperature gradients, replacing traditional linear compensation. For thermal diffusion boundary effects, a three-dimensional network of temperature sensors is deployed to acquire dynamic temperature field distribution, and the spatial distribution map of offsets is used to identify the interface thermal diffusion influence region. Frequency band calculations of aging and moisture response rate indices achieve response decoupling, and a humidity correction factor is used to avoid spurious signals caused by cross-reversal of defect temperature responses. Through differentiated temperature gradient design, the temperature response characteristics of aging-dominant and moisture-dominant defects form distinguishable patterns in both the spatial and frequency domains. In conjunction with the defect classification labels generated in step S20, the application of the temperature gradient is specifically targeted, avoiding signal confusion caused by indiscriminate temperature excitation. For example, for aging-dominated defects, the focus is on high-frequency response analysis. By amplifying the dielectric characteristics of aging products in the high-temperature range, interference from low-frequency moisture signals is reduced, thus improving the signal-to-noise ratio of the evaluation results. Without step S30, quantitative differentiation between aging-dominated and moisture-dominated defects cannot be achieved, resulting in qualitative evaluation results. This would deprive the acoustic modulation excitation scheme in step S40 of parameter adjustment basis. For example, the acoustic intensity cannot be matched according to the defect severity, potentially failing to excite subthreshold defect signals due to insufficient intensity or causing secondary damage to the insulation material due to excessive intensity. Step S30, by generating quantitative evaluation indicators for defect severity, transforms the defect type identification results of step S20 into quantifiable physical parameters, providing a numerical benchmark for optimizing the acoustic excitation parameters in step S40. This creates a continuous logical chain from defect location and type identification to severity evaluation, ensuring the systematic nature of the detection process and the traceability of the results.

[0115] Step S40: Design an acoustic modulation excitation scheme based on the quantitative evaluation index of defect degree, identify the abnormal sound pressure attenuation region through acoustic-electric coupling detection and generate a high-resolution defect feature map.

[0116] Further, step S40 includes:

[0117] Step S41: Based on the quantitative assessment index of the defect degree and the preliminary defect location map, install a controllable sound wave generator on the outer wall of the transformer tank and deploy a MEMS sound pressure sensor array in the defect area inside the tank.

[0118] Step S42: Start the acoustic wave modulation excitation and collect the acoustic pressure amplitude data and acoustic pressure phase data of each sensor node in real time through the MEMS acoustic pressure sensor array to construct a three-dimensional sound field distribution map and identify the abnormal acoustic pressure attenuation area.

[0119] Step S43: For the abnormal sound pressure attenuation region, based on the three-dimensional sound field distribution map, dielectric spectrum measurements are performed under two states: with and without acoustic modulation, to obtain dielectric response data under the two states; based on the dielectric response data under the two states, the dielectric spectrum difference between the two states is calculated and coupled with the sound pressure phase data for analysis; the acoustic frequency with acoustic modulation is consistent with the vibration resonance frequency in the vibration resonance frequency data.

[0120] Step S44: Based on the coupling analysis results of dielectric spectrum difference and acoustic pressure phase data, optimize the acoustic excitation parameters and perform multi-frequency scanning to generate a high-resolution defect feature map.

[0121] Specifically, based on the aging degree index and moisture content index included in the quantitative assessment indicators of defect severity, the adjustment range of the acoustic excitation intensity is determined. A higher aging degree index corresponds to a higher acoustic excitation intensity range to ensure that the weak dielectric signal generated by the fracture of nanoscale cellulose can be effectively excited; a higher moisture content index corresponds to a lower acoustic excitation intensity range to avoid signal distortion caused by severe moisture disturbance. The installation location of the acoustic generator is determined based on the defect coordinates in the preliminary defect location map. After the acoustic generator is started, the acoustic pressure amplitude and phase data collected by each MEMS acoustic pressure sensor are recorded simultaneously. The construction of the three-dimensional sound field distribution map uses a back-projection algorithm to map the acoustic pressure data of each sensor node to a three-dimensional physical coordinate system. Noise points are processed by Gaussian filtering, and then Kriging interpolation is used to supplement the sound field data at locations where no sensors are deployed, forming a continuous sound field distribution feature. The identification of abnormal sound pressure attenuation areas is based on the attenuation rate threshold. By analyzing the sound field data of 100 normal transformers, the attenuation rate of each area is calculated as (average sound pressure amplitude of the surrounding area - sound pressure amplitude of the target area) / average sound pressure amplitude of the surrounding area × 100%. After fitting these attenuation rates to a normal distribution, the upper limit of the 95% confidence interval is taken as the attenuation rate threshold. When the sound pressure amplitude of a certain area is attenuated by more than the attenuation rate threshold compared with the surrounding area, it is determined to be an abnormal sound pressure attenuation area. Abnormal sound pressure attenuation areas usually correspond to areas with concentrated defects, because the material density changes caused by defects will enhance the scattering and absorption of sound waves.

[0122] For regions with abnormal sound pressure attenuation, firstly, without acoustic modulation, dielectric spectrum data covering a frequency band from low-frequency point f1 to high-frequency point f6 (e.g., f1 is 0.001Hz, f6 is 100Hz) is collected as a reference signal. Then, acoustic modulation is activated, maintaining the acoustic frequency consistent with the vibration resonance frequency, and dielectric response data, including the dielectric constant and dielectric loss tangent, is collected again in the same frequency band. Maintaining consistency between the acoustic frequency and the vibration resonance frequency essentially matches the acoustic excitation with the inherent vibration characteristics of the defect, utilizing the resonance effect to amplify the acoustic-electric coupling response of the defect. For example, if the vibration resonance frequency of a defect region is determined to be 500Hz through spectral analysis in step S20, then the acoustic frequency is set to 500Hz during acoustic modulation. In this case, the acoustic energy can be efficiently converted into mechanical vibration in the defect region, causing detectable changes in the dielectric properties of subthreshold defects such as nanoscale cellulose fracture, avoiding energy loss or weak signal problems caused by frequency mismatch.

[0123] The dielectric spectrum difference is calculated as the difference between the dielectric constant Δε' and the dielectric loss tangent Δtanδ' at the same frequency point under two conditions, where Δε' is the difference between the dielectric constant with and without acoustic modulation, and Δtanδ' is the difference between the dielectric loss tangent with and without acoustic modulation. When coupling the dielectric spectral difference with the acoustic pressure phase data, the correlation between Δε' and the acoustic pressure phase data, and the correlation between Δtanδ' and the acoustic pressure phase data are calculated using coherence functions, resulting in two coherence function values: the dielectric constant difference coherence coefficient and the dielectric loss difference coherence coefficient. For each frequency point, the arithmetic mean of the dielectric constant difference coherence coefficient and the dielectric loss difference coherence coefficient is taken as the comprehensive coherence coefficient for that frequency point. When the comprehensive coherence coefficient is greater than a preset correlation threshold (e.g., 0.8, determined through statistical analysis of effective signals and noise in historical defect detection data), the frequency point is determined to be an effective coupling point, indicating that the dielectric change at that frequency point is mainly caused by the acoustic-electric response of the defect, rather than noise or interference. When optimizing the acoustic excitation parameters based on the coupling analysis results, for effective coupling points, the corresponding excitation intensity weight is adjusted to a coefficient greater than 1 (e.g., 1.2, which is experimentally determined based on the linear relationship between defect response intensity and excitation intensity); for ineffective coupling points, the corresponding excitation intensity weight is adjusted to a coefficient less than 1 (e.g., 0.8). The excitation intensity weight is a coefficient that adjusts the distribution of acoustic excitation energy at different frequency points. Its core principle is to differentiate energy allocation based on the correlation between frequency points and the acoustic-electric response of defects: the weight of effective coupling points is set to be greater than 1 to enhance their corresponding dielectric signal and amplify subtle changes in subthreshold defects; the weight of ineffective coupling points is set to be less than 1 to reduce ineffective energy input and avoid noise amplification. Effective coupling points are increased in excitation intensity according to their weights, while ineffective coupling points are decreased in intensity according to their weights, concentrating energy at effective frequency points, improving the signal-to-noise ratio of the defect signal, and providing a clear signal basis for subthreshold defect identification.

[0124] Multi-frequency scanning covers the frequency band from f1 to above f6, with step sizes divided by frequency band: a smaller step size F1 (e.g., 0.001Hz) is used for the f1 to f2 range, a medium step size F2 (e.g., 0.01Hz) is used for the f2 to f4 range, and a larger step size F3 (e.g., 0.1Hz) is used for the f4 to f6 and above ranges, ensuring denser data points in frequency bands where defect characteristic frequencies are concentrated (e.g., f3 to f5). High-resolution defect feature maps are generated by spatially fusing Δε' and Δtanδ' obtained from multi-frequency scanning with the three-dimensional acoustic field distribution. A tomographic imaging algorithm is used to spatially reconstruct the dielectric spectral differences at different frequency points, enabling the map to clearly present the spatial morphology of the defect. Its resolution meets the identification requirements for subthreshold defects (e.g., nanoscale cellulose fractures).

[0125] Step S40 addresses the issues of acoustic-electric modulation blind zone and Doppler frequency shift interference in subthreshold defects. Specifically, the dielectric response change of early-stage micro-defects (such as nanoscale cellulose fractures) is below the noise threshold of the measurement system, creating a detection blind zone. Furthermore, transformer oil convection causes Doppler frequency shift in the acoustic waves, which, combined with the frequency change caused by the defects, masks the true defect signal. By exciting the device with an acoustic wave frequency consistent with the vibration resonance frequency determined in step S20, energy focusing is achieved through the resonance effect. When the acoustic wave frequency matches the inherent vibration frequency of the defect, the mechanical vibration of the defect area is significantly amplified, raising the dielectric property changes of micro-defects such as nanoscale cellulose fractures from the subthreshold level to a detectable level. This solves the problem of weak signals caused by frequency mismatch in traditional excitation. Simultaneously, by deploying a MEMS acoustic pressure sensor array to construct a three-dimensional sound field distribution map, regions of abnormal acoustic pressure attenuation (corresponding to concentrated defect areas) are identified, ensuring that acoustic excitation is precisely applied to the target area, avoiding energy dispersion, and further enhancing the response signal of micro-defects.

[0126] Doppler frequency shift interference is caused by the superposition of acoustic wave frequency shift due to oil convection and defect signals. Interference is eliminated through a comparison of two states and coherence analysis: Dielectric spectrum data are collected under both acoustic wave modulation and non-acoustic wave modulation conditions. The difference in dielectric constant Δε' and the difference in dielectric loss tangent Δtanδ' at the same frequency point are calculated to highlight the acoustic-electric response changes specific to defects. The correlation between Δε', Δtanδ' and acoustic pressure phase data is calculated using a coherence function to obtain a comprehensive coherence coefficient. When this coefficient is greater than a preset threshold, it is determined to be an effective coupling point—the dielectric change at this point mainly originates from the acoustic-electric response of the defect, while interference such as Doppler frequency shift is screened out because of its low correlation with acoustic pressure phase. Effective signals are enhanced by differentially adjusting the excitation intensity weights: effective coupling points are assigned a weight greater than 1 to increase their acoustic excitation intensity and further amplify the defect signal; ineffective coupling points are assigned a weight less than 1 to reduce ineffective energy input and suppress noise. Finally, multi-frequency scanning (adjusting the step size according to the defect characteristic frequency band) combined with tomographic imaging algorithm fuses the dielectric spectrum difference with the three-dimensional sound field distribution space to generate a high-resolution defect feature map, achieving clear characterization of subthreshold defects and effective removal of interference signals.

[0127] Step S50: Construct a multimodal intelligent sensor network based on high-resolution defect feature maps to form a defect evolution feature fusion vector. Predict defect evolution based on the defect evolution feature fusion vector and a pre-constructed defect development trend prediction model. Generate a graded early warning signal based on the defect evolution prediction results.

[0128] Further, step S50 includes:

[0129] Step S51: Based on the high-resolution defect feature map, integrate the MEMS electric field sensor network, MEMS acceleration sensor array, temperature sensor three-dimensional network and MEMS acoustic pressure sensor array to construct a multimodal intelligent sensor network and configure edge computing nodes.

[0130] The spatial coordinates of defects in the high-resolution defect feature map are used as the spatial calibration benchmark for the sensor network. The three-dimensional coordinates of the defect edges in the map are mapped to the deployment boundaries of the sensor nodes. The boundary range is determined by selecting the geometric center of the defect and edge feature points, ensuring that the monitoring range of each sensor node accurately corresponds to the defect area. The integration process achieves spatial alignment of each sensor network through a coordinate transformation algorithm. The transformation matrix is ​​calculated based on the correspondence between the feature markers in the map and the physical coordinates of the sensors. Three non-collinear defect feature points are selected, and the coordinate transformation parameters are obtained by fitting using the least squares method, making the data collected by different sensors comparable in the same spatial coordinate system. The edge computing nodes adopt an embedded system. The time synchronization module built into the node achieves clock calibration with the sensor network through GPS or BeiDou time signals, with the synchronization error controlled within 1 microsecond, ensuring the consistency of multimodal data in the time dimension.

[0131] Step S52: Based on the real-time monitoring data of the multimodal intelligent sensor network, the characteristic parameters of each mode are calculated through edge computing nodes and fused to generate a comprehensive defect index; the characteristic parameters of each mode include the electric field intensity change rate, vibration spectrum entropy, heat flux density and acoustic impedance characteristic parameters;

[0132] The real-time monitoring data of the multimodal intelligent sensor network includes electric field intensity data at each node collected by the MEMS electric field sensor network, vibration acceleration signals collected by the MEMS accelerometer array, spatial temperature distribution data collected by the three-dimensional temperature sensor network, and sound pressure amplitude and phase data collected by the MEMS sound pressure sensor array. The calculation methods for each modal characteristic parameter are as follows: the electric field intensity change rate is the ratio of the difference in electric field intensity between two adjacent sampling times to the corresponding time interval. The electric field intensity change rate reflects the dynamic change rate of the electric field distribution; a larger value indicates a more severe electric field distortion caused by the defect. The vibration spectrum entropy is obtained by performing a Fourier transform on the vibration acceleration signal to obtain the frequency spectrum, and then calculated according to the information entropy formula. That is, the spectrum entropy is equal to the sum of the negative values ​​of the product of the probability and the logarithm of each frequency component. A higher spectrum entropy value indicates a more complex vibration frequency distribution and a more significant multi-physics coupling corresponding to the defect. The heat flux density is calculated based on Fourier's law of heat conduction; that is, the heat flux density is equal to the conductivity of the insulating material. The thermal conductivity is calculated as the product of the thermal coefficient and the temperature gradient. The thermal conductivity was determined through thermophysical experiments on the oil-paper insulation material. At 25℃, the thermal conductivity of the oil-paper composite insulation was taken as 0.15 W / (m·K). The temperature gradient was calculated as the ratio of the temperature difference between adjacent nodes to the node spacing, collected by a three-dimensional network of temperature sensors. The acoustic impedance characteristic parameter is the ratio of the sound pressure amplitude collected by a MEMS sound pressure sensor to the particle vibration velocity. The particle vibration velocity is obtained by integrating the acceleration signal. Changes in acoustic impedance reflect changes in the material density of the defect area. Aging-induced cellulose loosening reduces acoustic impedance, while moisture accumulation caused by dampness increases it. The comprehensive defect index is fused using a weighted summation algorithm. The weights are determined using the Delphi method. Several transformer operation and maintenance experts were invited to score the importance of each characteristic parameter, with a score range of 1-10. The final weight is the normalized result of the scores for each parameter.

[0133] Step S53: In the three-dimensional dielectric response coordinate system, the dielectric response data points at each measurement time are combined with the corresponding comprehensive defect index to form four-dimensional defect feature points. The four-dimensional defect feature points are connected in chronological order to establish a defect evolution trajectory. The trajectory geometric feature parameters of the defect evolution trajectory are calculated. The four-dimensional defect feature points and the trajectory geometric feature parameters are combined to form a defect evolution feature fusion vector.

[0134] The fourth dimension of the four-dimensional defect feature points is the comprehensive defect index, whose value is mapped to the 0-1 interval through min-max normalization, where 0 corresponds to a defect-free state and 1 corresponds to a severe defect state. Dielectric response data points are combined with the corresponding comprehensive defect index to form four-dimensional defect feature points. These four-dimensional defect feature points are then connected in chronological order to establish a defect evolution trajectory. The trajectory's geometric parameters include curvature, tangent slope, and spatial span. Curvature is calculated using the radius of the arc formed by three adjacent four-dimensional points; a smaller radius indicates a more drastic change in defect state. The radius of the circle determined by the three points is the radius of curvature for that interval. The tangent slope is the ratio of the difference between two adjacent points in the comprehensive defect index dimension to the time difference, reflecting the rate of defect development. A positive slope indicates defect intensification, while a negative slope indicates defect mitigation. The spatial span is the maximum Euclidean distance of the trajectory in the three-dimensional dielectric response coordinate system, i.e., the maximum spatial distance between any two points on the trajectory, reflecting the extent of the defect's influence.

[0135] Step S54: Based on the defect evolution feature fusion vector and the pre-constructed defect development trend prediction model, obtain the comprehensive defect index change rate and final state label for the future preset time step, and generate a graded early warning signal based on the comprehensive defect index change rate.

[0136] The defect development trend prediction model adopts a time-series prediction architecture based on a long short-term memory network. The network consists of an input layer with 7 nodes corresponding to the four dimensions and 3 trajectory geometric feature parameters of the four-dimensional feature points in the defect evolution feature fusion vector; two hidden layers with 64 and 32 nodes respectively, using ReLU activation function; and an output layer with 2 nodes corresponding to the comprehensive defect index change rate and the final defect state category at a preset future time step. The training data comes from the full life cycle monitoring records of multiple transformers, covering evolution data of different operating years and different defect types. Each data point contains a set of defect evolution feature fusion vectors consisting of multiple consecutive four-dimensional defect feature points and corresponding trajectory geometric parameters, as well as a final state label, such as "normal operation", "requires maintenance", and "requires insulation replacement". The training process uses the sliding window method to generate samples, that is, the input is a preset number of consecutive four-dimensional points and their corresponding geometric parameters, and the output is the comprehensive defect index change rate and the final state label at a preset future time step. The loss function is a weighted sum of mean squared error over the rate of change and cross-entropy over the state category, with a weight ratio of 1:1. Iterative training is performed using the Adam optimizer until the model's state prediction accuracy on the test set meets preset requirements. The generation of tiered early warning signals is based on the predicted rate of change of the comprehensive defect index. A Level 1 warning corresponds to a rate of change below the first threshold, requiring strengthened regular monitoring; a Level 2 warning corresponds to a rate of change between the first and second thresholds, requiring planned maintenance; and a Level 3 warning corresponds to a rate of change above the second threshold, requiring emergency handling. The warning thresholds are determined through statistical analysis of multiple transformer fault cases to ensure coverage of the vast majority of pre-fault characteristics.

[0137] Step S50 addresses the problem of bifurcation mutations and the failure to predict multi-steady-state transitions in defect evolution. Specifically, transformer insulation defect development exhibits bifurcation points; minute perturbations can lead to abrupt changes in the evolution path; and the interaction of multiple defects can generate chaotic evolution. Traditional linear extrapolation and deterministic models cannot capture such nonlinear behavior. The construction of a multimodal intelligent sensor network provides a cross-physics data foundation for capturing bifurcation mutations. By integrating a MEMS electric field sensor network, a MEMS accelerometer array, a three-dimensional temperature sensor network, and a MEMS acoustic pressure sensor array, synchronous acquisition of multi-physics data (electric field, vibration, temperature, and acoustic pressure) is achieved. The spatial coordinates are calibrated using a high-resolution defect feature map to ensure accurate correspondence between the data and the defect region. The linkage analysis of multi-physics data can identify weak coupling signals before bifurcation points. For example, when aging and moisture-dominated defects are coupled, the abnormal ratio of the rate of change of electric field intensity to the heat flux density can serve as a precursor to bifurcation. This correlation cannot be captured in single-physics monitoring, avoiding the omission of nonlinear features by traditional single-parameter monitoring. The construction of a comprehensive defect index and a four-dimensional evolution trajectory enables the quantitative tracking of multi-steady-state features. The comprehensive defect index, by weighted fusion of electric field intensity change rate, vibration spectral entropy, heat flux density, and acoustic impedance characteristic parameters, condenses multi-dimensional parameters into quantifiable indicators, eliminating the problem of inconsistent thresholds in traditional multi-parameter evaluations and enabling direct comparison of characteristic differences between different steady states. The evolutionary trajectory of four-dimensional defect feature points (logarithmic frequency, dielectric constant, dielectric loss tangent, and comprehensive defect index) formed by a time series quantifies nonlinear changes through geometric parameters such as curvature, tangent slope, and spatial span: curvature abrupt changes correspond to bifurcation point locations, tangent slope changes reflect steady-state transition rates, and spatial span expansion reflects the nonlinear expansion of the defect's influence range. These parameters provide calculable geometric features for identifying multiple steady-state transitions. The temporal prediction architecture of the Long Short-Term Memory (LSTM) network solves the problem of traditional linear models failing to predict bifurcation abrupt changes. The model input is a defect evolution feature fusion vector containing the four-dimensional feature point sequence and trajectory geometric parameters. The training data covers the entire lifecycle evolution data of different operating years and different defect types, including a large number of bifurcation abrupt changes and steady-state transition cases. LSTM's gating mechanism can memorize key features from long time series. When the comprehensive defect index fluctuates within a certain range, the model can identify subtle trends towards a new steady state, whereas traditional linear models would have such signals masked by data fluctuations. By generating samples using a sliding window method, the model learns the evolutionary patterns of continuous time steps. The loss function combines mean squared error and cross-entropy, balancing rate of change prediction and state classification to ensure accurate fitting of nonlinear features before and after the bifurcation point. A tiered early warning mechanism implements a tiered response to bifurcation risk based on the predicted rate of change of the comprehensive defect index.The warning threshold is determined through statistics of multiple failure cases, covering the characteristic interval before bifurcation mutation. When the model predicts that the rate of change enters the corresponding interval, different levels of warnings are triggered, providing targeted strategies for operation and maintenance, and avoiding the response lag caused by the inability to predict bifurcation in traditional warnings.

[0138] Example 2:

[0139] This embodiment, based on Embodiment 1, provides an optimized non-destructive testing system for oil-immersed power transformers, such as... Figure 5 As shown, it includes:

[0140] Defect location module: used to establish a three-dimensional dielectric response coordinate system and obtain initial electric field distribution data, and generate a preliminary defect location map based on the initial electric field distribution data;

[0141] Defect classification module: used to perform spiral sweep frequency excitation detection based on the preliminary defect location map, identify the defect type of insulation defects inside the oil-immersed power transformer, and determine the defect type; the defect type includes aging-dominated defects and moisture-dominated defects;

[0142] Defect quantitative assessment module: Establishes a temperature gradient excitation scheme based on defect type, decouples defect response, and generates quantitative assessment indicators of defect severity;

[0143] Defect Feature Map Construction Module: Based on the quantitative assessment index of defect degree, an acoustic wave modulation excitation scheme is designed, and the abnormal sound pressure attenuation region is identified through acoustic-electric coupling detection and a high-resolution defect feature map is generated;

[0144] The graded early warning module constructs a multimodal intelligent sensor network based on high-resolution defect feature maps, forming a defect evolution feature fusion vector. Based on the defect evolution feature fusion vector and a pre-constructed defect development trend prediction model, it predicts the defect evolution and generates graded early warning signals based on the defect evolution prediction results.

[0145] In the defect feature map construction module, the method for identifying abnormal sound pressure attenuation regions and generating high-resolution defect feature maps through acoustic-electric coupling detection includes:

[0146] Step S41: Based on the quantitative assessment index of the defect degree and the preliminary defect location map, install a controllable sound wave generator on the outer wall of the transformer tank and deploy a MEMS sound pressure sensor array in the defect area inside the tank.

[0147] Step S42: Start the acoustic wave modulation excitation and collect the acoustic pressure amplitude data and acoustic pressure phase data of each sensor node in real time through the MEMS acoustic pressure sensor array to construct a three-dimensional sound field distribution map and identify the abnormal acoustic pressure attenuation area.

[0148] Step S43: For the abnormal sound pressure attenuation region, based on the three-dimensional sound field distribution map, dielectric spectrum measurements are performed under two states: with and without acoustic modulation, to obtain dielectric response data under the two states; based on the dielectric response data under the two states, the dielectric spectrum difference between the two states is calculated and coupled with the sound pressure phase data for analysis; the acoustic frequency with acoustic modulation is consistent with the vibration resonance frequency in the vibration resonance frequency data.

[0149] Step S44: Based on the coupling analysis results of dielectric spectrum difference and acoustic pressure phase data, optimize the acoustic excitation parameters and perform multi-frequency scanning to generate a high-resolution defect feature map.

[0150] In the graded early warning module, the method for constructing a multimodal intelligent sensor network based on high-resolution defect feature maps to form a defect evolution feature fusion vector includes:

[0151] Based on high-resolution defect feature maps, a multimodal intelligent sensor network is constructed by integrating a MEMS electric field sensor network, a MEMS accelerometer array, a three-dimensional temperature sensor network, and a MEMS acoustic pressure sensor array, and edge computing nodes are configured. Based on real-time monitoring data from the multimodal intelligent sensor network, the edge computing nodes calculate the characteristic parameters of each mode and fuse them to generate a comprehensive defect index. The characteristic parameters of each mode include the rate of change of electric field intensity, vibration spectrum entropy, heat flux density, and acoustic impedance characteristic parameters. In the three-dimensional dielectric response coordinate system, the dielectric response data points at each measurement moment are combined with the corresponding comprehensive defect index to form four-dimensional defect feature points. The four-dimensional defect feature points are connected in chronological order to establish a defect evolution trajectory, and the trajectory geometric characteristic parameters of the defect evolution trajectory are calculated. The four-dimensional defect feature points and the trajectory geometric characteristic parameters are combined to form a defect evolution feature fusion vector.

[0152] The methods and systems of this application may be implemented in many ways. For example, they may be implemented by software, hardware, firmware, or any combination of software, hardware, and firmware. The above-described order of steps for the method is for illustrative purposes only, and the steps of the method of this application are not limited to the order specifically described above, unless otherwise specifically stated.

[0153] In addition, the parts of the technical solutions provided in the embodiments of this application that are consistent with the implementation principles of the corresponding technical solutions in the prior art have not been described in detail, so as to avoid excessive elaboration.

[0154] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the invention. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. An optimized method for non-destructive testing of oil-immersed power transformers, characterized in that, The method includes: A three-dimensional dielectric response coordinate system is established and initial electric field distribution data is obtained. Based on the initial electric field distribution data, a preliminary defect location map is generated. Based on the preliminary defect location map, a spiral sweep frequency excitation detection is performed to identify the insulation defects inside the oil-immersed power transformer and determine the defect type; the defect type includes aging-dominated defects and moisture-dominated defects; A temperature gradient excitation scheme is established based on the defect type, defect response is decoupled, and a quantitative evaluation index of defect degree is generated. Based on the quantitative assessment index of defect degree, an acoustic wave modulation excitation scheme is designed, and the abnormal sound pressure attenuation region is identified by acoustic-electric coupling detection and a high-resolution defect feature map is generated. A multimodal intelligent sensor network is constructed based on high-resolution defect feature maps to form a defect evolution feature fusion vector. The defect evolution is predicted based on the defect evolution feature fusion vector and a pre-constructed defect development trend prediction model. Based on the defect evolution prediction results, a graded early warning signal is generated.

2. The optimized method for non-destructive testing of oil-immersed power transformers according to claim 1, characterized in that, The method for generating a preliminary defect location map based on initial electric field distribution data includes: In a three-dimensional dielectric response coordinate system, a regular hexagonal characteristic frequency detection region is defined and six characteristic frequency detection points are determined. Based on the six identified characteristic frequency detection points, the initial dielectric response data and electric field gradient data of each monitoring point are synchronously collected by a MEMS electric field sensor network deployed on the inner wall of the oil tank, generating an initial dielectric response dataset. Based on the initial dielectric response dataset, a preliminary defect location map is generated, which marks different defect areas and defect location information.

3. The optimized method for non-destructive testing of oil-immersed power transformers according to claim 2, characterized in that, The method for generating a preliminary defect location map based on the initial dielectric response dataset includes: Based on the initial dielectric response dataset, calculate the vector angle and vector length from the dielectric response data point at each MEMS electric field sensor node location to the origin of the three-dimensional dielectric response coordinate system; The electric field gradient magnitude between adjacent MEMS electric field sensor nodes is calculated based on the electric field gradient data, and effective detection areas with electric field gradient magnitudes greater than the preset electric field gradient threshold are selected. For the selected effective detection area, the existence judgment conditions of defects are established by combining vector angle, vector length and electric field gradient magnitude, and the defect area is determined. The identified defect area is spatially marked in a three-dimensional dielectric response coordinate system to generate a preliminary defect location map containing defect location information.

4. The optimized method for non-destructive testing of oil-immersed power transformers according to claim 3, characterized in that, The six vertices of the regular hexagonal characteristic frequency detection region correspond to six characteristic frequency points, namely, the first low-frequency band frequency point f1, the second low-frequency band frequency point f2, the first mid-frequency band frequency point f3, the second mid-frequency band frequency point f4, the first high-frequency band frequency point f5, and the second high-frequency band frequency point f6. The frequency range of f1-f2 is defined as the low-frequency band, the frequency range of f3-f4 is defined as the mid-frequency band, and the frequency range of f5-f6 is defined as the high-frequency band, wherein f1 < f2 < f3 < f4 < f5 < f6.

5. The optimized method for non-destructive testing of oil-immersed power transformers according to claim 4, characterized in that, The method for identifying the type of insulation defects inside an oil-immersed power transformer includes: A MEMS accelerometer array is deployed, and dielectric response data and vibration response data are synchronously acquired during the frequency sweep process through a MEMS electric field sensor network and the MEMS accelerometer array; the vibration envelope change rate is obtained based on the vibration response data. Spectral analysis of the vibration response data yields the vibration resonance frequency data; The spiral sweep frequency sequence is coupled with the vibration resonance frequency data for analysis, and the vibration frequency ratio, which reflects the degree of matching between the sweep frequency and the resonance frequency, is calculated. Dielectric response time-domain features are extracted from dielectric response data during frequency sweeping. By combining dielectric response time-domain features, vibration envelope change rate, and vibration frequency ratio, the insulation defects inside oil-immersed power transformers are identified.

6. The optimized method for non-destructive testing of oil-immersed power transformers according to claim 5, characterized in that, The method for identifying insulation defects inside oil-immersed power transformers by integrating the time-domain characteristics of the dielectric response, the rate of change of the vibration envelope, and the ratio of vibration frequencies includes: Extract the rise time and fall time of each sweep cycle from the dielectric response data during the frequency sweep process, and calculate the ratio of rise time to fall time. Based on the ratio of rise time to fall time, the rate of change of vibration envelope, and the ratio of vibration frequency, the defect type is preliminarily determined, and a preliminary defect determination result is obtained. Statistical analysis is performed on the preliminary defect judgment results of multiple consecutive frequency sweep cycles. The final defect type is determined based on the frequency of occurrence of the defect type, and a defect classification label is generated.

7. The optimized method for non-destructive testing of oil-immersed power transformers according to claim 6, characterized in that, The method for decoupling defect response and generating quantitative assessment indicators of defect severity includes: Based on the determined defect type, a differential temperature gradient is applied to the upper and lower ends of the transformer tank to obtain three-dimensional temperature field dynamic distribution data, and the dielectric response data of each characteristic frequency point is monitored during the establishment of the temperature gradient to calculate the dielectric spectrum change rate in the low-frequency and high-frequency bands. Based on the dielectric spectrum change rate and three-dimensional temperature field dynamic distribution data in the low-frequency and high-frequency bands, a quantitative assessment model for the degree of defect with temperature compensation is established, generating a quantitative assessment index for the degree of defect that includes aging degree index and moisture content index.

8. The optimized method for non-destructive testing of oil-immersed power transformers according to claim 7, characterized in that, The method for identifying abnormal sound pressure attenuation regions and generating high-resolution defect feature maps through acoustic-electric coupling detection includes: A MEMS acoustic pressure sensor array was deployed around the defect area inside the fuel tank. After the acoustic wave modulation excitation was activated, the acoustic pressure amplitude data and acoustic pressure phase data of each acoustic pressure sensor node were collected in real time to construct a three-dimensional sound field distribution map and identify the abnormal acoustic pressure attenuation area. For the abnormal acoustic pressure attenuation area, based on the three-dimensional sound field distribution map, dielectric spectrum measurements were performed under two states: with acoustic wave modulation and without acoustic wave modulation, to obtain dielectric response data under the two states. Based on the dielectric response data under the two states, the dielectric spectrum difference between the two states is calculated and coupled with the acoustic pressure phase data for analysis. Based on the coupling analysis results of the dielectric spectrum difference and the acoustic pressure phase data, the acoustic excitation parameters are optimized and multi-frequency scanning is performed to generate a high-resolution defect feature map.

9. The optimized method for non-destructive testing of oil-immersed power transformers according to claim 8, characterized in that, The acoustic frequency modulated by the acoustic wave is consistent with the vibration resonance frequency in the vibration resonance frequency data.

10. An optimization system for non-destructive testing of oil-immersed power transformers, used to implement the optimization method for non-destructive testing of oil-immersed power transformers according to any one of claims 1-9, characterized in that, The system includes: Defect location module: used to establish a three-dimensional dielectric response coordinate system and obtain initial electric field distribution data, and generate a preliminary defect location map based on the initial electric field distribution data; Defect classification module: used to perform spiral sweep frequency excitation detection based on the preliminary defect location map, identify the defect type of insulation defects inside the oil-immersed power transformer, and determine the defect type; the defect type includes aging-dominated defects and moisture-dominated defects; Defect quantitative assessment module: Establishes a temperature gradient excitation scheme based on defect type, decouples defect response, and generates quantitative assessment indicators of defect severity; Defect Feature Map Construction Module: Based on the quantitative assessment index of defect degree, an acoustic wave modulation excitation scheme is designed, and the abnormal sound pressure attenuation region is identified through acoustic-electric coupling detection and a high-resolution defect feature map is generated; The graded early warning module constructs a multimodal intelligent sensor network based on high-resolution defect feature maps, forming a defect evolution feature fusion vector. Based on the defect evolution feature fusion vector and a pre-constructed defect development trend prediction model, it predicts defect evolution and generates graded early warning signals based on the defect evolution prediction results.

Images