A transformer component-level equivalent test and intelligent data cleaning method and system

Abstract

The present application belongs to the technical field of power equipment insulation test, and relates to a transformer component level equivalent test and intelligent data cleaning method and system. The equivalent test and intelligent data cleaning method adopts an improved isolated forest model based on multi-scale feature fusion and adaptive threshold for data anomaly point detection, uses a deep learning model based on conditional generative adversarial network to complete missing data filling of high voltage test waveforms, and adopts an adaptive denoising algorithm based on dual-tree complex wavelet transform and an improved threshold function for feature extraction and denoising. The equivalent test and intelligent data cleaning system comprises a central control and scheduling module, an equivalent simulation module, a high voltage test module, a multi-parameter acquisition module and a data cleaning and intelligent analysis module, and the data cleaning and intelligent analysis module comprises a data anomaly point detection unit, a data missing filling unit and a feature extraction and denoising unit. The present application realizes automatic anomaly cleaning, missing filling and feature denoising of test data.

Description

Technical Field

[0001] This invention belongs to the field of power equipment insulation testing technology, specifically relating to a transformer component-level equivalent test and intelligent data cleaning method and system for weak insulation links. Background Technology

[0002] As a core hub device in the power system, the insulation performance of transformers directly determines the safety and stability of the power grid operation. The internal insulation structure of transformers is complex, especially in areas such as winding turns, sections, and bushing risers, where electric fields are concentrated and structures are weak, making them prone to insulation failure.

[0003] Existing transformer insulation testing technologies are mainly divided into two categories: one is the withstand voltage test for the entire equipment (such as the overall withstand voltage test specified in GB50150-2016). This type of test can only determine whether the overall insulation of the equipment is qualified, but cannot locate the weak points in the local insulation. Moreover, excessive test voltage can easily cause insulation damage to non-faulty parts. The other type is the performance test for a single insulating material (such as insulating paper or insulating oil). This type of test ignores the actual structural assembly relationship and electric field distribution characteristics of the key parts of the transformer, resulting in significant deviations between the test data and the actual operating conditions, making it difficult to accurately reflect the actual load-bearing capacity of the transformer insulation. Summary of the Invention

[0004] The purpose of this invention is to address the shortcomings of existing technologies by providing a transformer component-level equivalent testing method and system for weak insulation links. This method and system, through "structure-electric field" equivalent modeling and innovative modular design, can accurately simulate and test the real-world conditions of various key weak insulation links within the transformer. Furthermore, it incorporates artificial intelligence algorithms for intelligent analysis of the test data, thereby achieving precise and efficient evaluation of insulation performance.

[0005] To achieve the above objectives, the technical solution adopted by the present invention is as follows: A method for equivalent testing and intelligent data cleaning at the transformer component level includes the following steps: An experimental model was built to perform component-level equivalent simulation of the transformer; Provide various types of high-voltage excitations for the test model to conduct equivalent tests; Under high-voltage excitation, multiple physical quantity signals reflecting the insulation state are synchronously acquired as raw experimental data. The raw experimental data undergoes automatic anomaly cleaning, missing data filling, and feature noise reduction, including: S1. An improved isolated forest model based on multi-scale feature fusion and adaptive thresholding is used for outlier detection. The specific steps are as follows: S11. Multi-scale feature extraction: Construct a three-dimensional feature vector for waveform data, including macroscopic waveform parameter features, microscopic oscillation features, and statistical distribution features; S12. Input the three-dimensional feature vector into the isolated forest model and calculate the anomaly score for each data point; S13. An adaptive threshold determination method based on K-nearest neighbor local density is adopted. For each data point, the difference between its abnormal score and the abnormal scores of its K nearest neighbor data points is calculated. If the difference exceeds a preset multiple of the standard deviation of the scores in the neighborhood, it is determined to be an outlier. S2. Using a deep learning model based on conditional generative adversarial networks, missing data imputation of high-voltage test waveforms is completed. This includes: the generator adopts an encoder-decoder structure combining Bi-LSTM and a self-attention mechanism. Bi-LSTM captures the causal temporal dependence of the impulse voltage waveform, and the self-attention mechanism captures the global parameter constraint relationship between the wavefront and the wave tail; three mask simulation modes—continuous missing data, random point missing data, and segmented missing data—are designed for training based on the causes of missing data in the high-voltage test waveforms; a weighted combination of adversarial loss and L1 reconstruction loss is used as the total loss function. S3. An adaptive denoising algorithm based on dual-tree complex wavelet transform and improved threshold function is used for feature extraction and denoising.

[0006] Preferably, the specific steps for feature extraction and noise reduction in S3 are as follows: S31. Perform dual-tree complex wavelet transform on the original waveform, using two parallel decomposition paths: real part tree and imaginary part tree. The number of decomposition levels is automatically determined based on the waveform sampling frequency and the effective signal bandwidth. The real part tree and imaginary part tree have complementary singularity responses to the signal. S32. During the decomposition process, wavelet basis functions are selected differently according to the characteristics of the signal frequency band: Daubechies series wavelet basis is used for low-frequency steady-state signal components, and Symlet series wavelet basis is used for high-frequency transient signal components. S33. At each decomposition layer, wavelet coefficient thresholds are calculated for the real and imaginary parts respectively. The threshold selection adopts an adaptive method based on SURE. S34. Wavelet coefficients below the threshold are contracted by using an improved soft-hard threshold trade-off function. A continuous transition region is introduced near the threshold, and the width of the continuous transition region is adaptively correlated with the number of decomposition layers: the transition region of the high-frequency layer is narrow to preserve the steep leading edge characteristics and accurate amplitude of the discharge pulse; the transition region of the low-frequency layer is wide to ensure the overall smoothness of the waveform and to keep the coefficients above the threshold unchanged. S35. Perform inverse wavelet transform on the processed real and imaginary coefficients respectively, and average the two reconstruction results to obtain the final denoised waveform. S36. Normalize the denoised test data along with the oil passage number and maximum voltage to eliminate the differences in data dimensions and magnitudes under different test conditions.

[0007] Preferably, the macroscopic waveform parameter features include voltage peak value, wavefront time T1, and wave tail time T2; the microscopic oscillation features include the root mean square amplitude of local waveform oscillations and zero-crossing rate; and the statistical distribution features include the kurtosis and skewness of waveform sampling points.

[0008] Preferably, continuous loss is used to simulate continuous data loss caused by loose fiber optic cable interface or long-term interruption of transmission link; random point loss is used to simulate sporadic data loss caused by instantaneous interference of strong electromagnetic pulse on the acquisition channel; segmented loss is used to simulate multiple discontinuous loss caused by multi-channel switching or intermittent fluctuation of sensor performance.

[0009] Preferably, the test model is built using replaceable simulation units, which include: a winding turn / segment and end insulation simulation unit for simulating oil gap insulation structure; a coil potential gradient simulation unit for simulating the inner coil or ground potential environment of the transformer; and a bushing riser long oil gap simulation unit for simulating the equalizing ball long oil gap insulation structure.

[0010] Preferably, the method for acquiring multi-physical quantity signals reflecting the insulation state includes: measuring the transient electric field intensity in the oil-paper composite insulation using a broadband electric field sensor; recording the excitation waveform, discharge current, and breakdown time of the test circuit using voltage and current sensors; measuring the potential gradient using a probe array; and collecting and storing the test data using a multi-channel digital recorder.

[0011] Preferably, a series resonant test device with an impulse voltage generator is used to provide various types of high-voltage excitation for the test model.

[0012] A transformer component-level equivalent testing and intelligent data cleaning system is provided to implement the aforementioned transformer component-level equivalent testing and intelligent data cleaning method. The system includes an equivalent simulation module, a high-voltage testing module, a multi-parameter acquisition module, and a data cleaning and intelligent analysis module connected to a central control and scheduling module. The data cleaning and intelligent analysis module includes a data anomaly detection unit, a data missing data filling unit, and a feature extraction and denoising unit. The data anomaly detection unit employs an improved isolated forest algorithm based on multi-scale feature fusion and adaptive thresholding. The data missing data filling unit utilizes a deep learning model based on conditional generative adversarial networks to fill in missing data in the high-voltage test waveform. The feature extraction and denoising unit employs an adaptive denoising algorithm based on dual-tree complex wavelet transform and an improved threshold function for feature extraction and denoising.

[0013] Preferably, the winding turn / segment and end insulation simulation unit includes an oil gap insulation structure including an electrostatic ring and a corner ring; the coil potential gradient simulation unit has a metal cylinder set on the inner diameter side of the winding model and grounded, and a measurement probe interface and an optical fiber through disk are pre-set at a specified coil coil disc; the bushing riser long oil gap simulation unit includes a replaceable equalizing ball insulation structure, which is divided into bare electrode, insulated electrode, shaped electrode and paper tube electrode, and is selected and combined according to different voltage levels.

[0014] Preferably, the impulse voltage generator includes a generator body, a weakly damped voltage divider, a cutoff device, and a load capacitor.

[0015] Compared with the prior art, the beneficial effects of the present invention are as follows: Unlike overall withstand voltage tests, which cannot locate local weak points, and single-material tests, which ignore the actual electric field distribution, this invention constructs a component-level test model through "structure-electric field" equivalent modeling, which restores the actual working conditions of key areas of the transformer. At the same time, it introduces an improved isolated forest algorithm, conditional generative adversarial network, and improved wavelet transform adaptive threshold denoising algorithm to realize automatic anomaly cleaning, missing data filling, and feature denoising of test data, thereby improving data quality and analysis efficiency. Attached Figure Description

[0016] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort. Figure 1 This is an architecture diagram of the equivalent testing and intelligent data cleaning system according to an embodiment of the present invention; Figure 2 This is a field schematic diagram of the impact generator according to an embodiment of the present invention; Figure 3 This is a schematic diagram of the power frequency series resonance test device according to an embodiment of the present invention. Figure 4 This is a schematic diagram of an electric field probe according to an embodiment of the present invention, wherein part 4a is the actual electric field probe and part 4b is a schematic diagram of the installation of the electric field probe. Detailed Implementation

[0017] The technical solution of the present invention will be clearly and completely described below with reference to the embodiments. Obviously, the embodiments described 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.

[0018] like Figure 1 As shown, a transformer component-level equivalent testing and intelligent data cleaning system includes: an equivalent simulation module, a high-voltage testing module, a multi-parameter acquisition module, a data cleaning and intelligent analysis module, and a central control and scheduling module. The central control and scheduling module is connected to each module to achieve data interaction. All modules work collaboratively to achieve a closed-loop test of the entire component-level insulation process, from scenario construction, excitation application, data acquisition to data analysis.

[0019] The equivalent simulation module is used to reproduce the "structure-electric field" characteristics of key weak points in a transformer at a 1:1 scale. This module adopts a reconfigurable design, consisting of multiple quickly replaceable simulation units to adapt to the testing needs of transformers with different voltage levels and structural types. Each simulation unit is designed with finite element simulation for electric field optimization, ensuring that the test model is equivalent to the key parts of a real transformer. The simulation unit includes: The winding turn / segment and end insulation simulation unit includes coils wound to actual dimensions and with the same manufacturing process, along with their complete end insulation structure. This unit includes oil gap insulation structures such as electrostatic rings and corner rings, and the coils employ the same winding structure as the actual product.

[0020] The coil potential gradient simulation unit is used to simulate the potential environment of the inner coil or ground of a transformer. This unit has a grounded metal cylinder installed on the inner diameter side of the winding model. Measurement probe interfaces and fiber optic connection trays are pre-installed at designated coil sections. The measurement probe interfaces are metal connection points embedded in the coil sections. The fiber optic connection tray is mounted on the tank wall and has a built-in connector for connecting internal and external optical fibers, providing conditions for full-link measurement of the coil potential distribution under impulse voltage. Before the test, the winding model is vacuum dried, oil-filled, and left to stand, simulating the process flow of a real transformer.

[0021] The bushing riser long oil gap simulation unit is used to simulate the long oil gap insulation structure of the equalizing sphere. This unit includes replaceable equalizing sphere insulation structures, which come in various forms such as bare electrodes, electrodes with insulation, electrodes with molded parts, and paper-tube electrodes, allowing for selection and combination according to different voltage levels. This unit connects to a high-voltage bushing and an adjustable ground electrode system controlled by a high-precision mechanical structure. The adjustable ground electrode adopts a large-size flat electrode structure, and the electrode distance is precisely adjusted by controlling the guide rod transmission through the mechanical structure. The unit as a whole adopts a sealed cavity design, capable of withstanding negative and positive pressure conditions. Key connection parts utilize rubber seals and sealing grooves to achieve negative pressure adsorption design, ensuring reliable sealing during the test process.

[0022] The high-voltage test module is used to provide various types of high-voltage excitation for the test model. This module consists of an impulse voltage generator and a power frequency series resonant test device.

[0023] like Figure 2 As shown, the impulse voltage generator is an existing product used to generate standard lightning impulse and switching impulse voltage waveforms. The test sample is connected in parallel to the output terminal of the impulse voltage generator. The impulse voltage generator mainly consists of four parts: the generator body, a weakly damped voltage divider, a cutoff device, and a load capacitor.

[0024] like Figure 3 As shown, the power frequency series resonant test device is an existing product used to generate a high power frequency voltage, with the test object connected in series in the resonant circuit.

[0025] The impulse voltage generator and the power frequency series resonant test device can be switched and connected according to the test command to programmatically implement high voltage excitation of various waveforms and amplitudes, such as standard lightning impulse, switching impulse, and power frequency withstand voltage.

[0026] The multi-parameter acquisition module is used to synchronously acquire multiple physical quantity signals reflecting the insulation state as raw experimental data under high-voltage excitation. This module consists of a front-end sensor array and a high-speed synchronous acquisition unit.

[0027] The front-end sensor array includes: a broadband electric field sensor, a high-speed response voltage and current sensor, and a potential gradient measurement probe array.

[0028] Broadband electric field sensors employ fully insulated electric field probes based on the photoelectric effect, such as... Figure 4As shown, the electric field probe is an existing product with a package size on the order of millimeters. Because the electric field probe has a fully insulated structure and is not easily magnetized, it will not affect the surrounding magnetic field distribution. The electric field probe is fixed to key locations such as the inner and outer diameter sides of the electrostatic ring, the inner layer of the corner ring, and the surface of the first disc of the winding, using insulating adhesive or crepe paper wrapping. It is used to directly measure the transient electric field intensity in the oil-paper composite insulation without distortion. The electric field probe is connected to a flange connector on the side wall of the test chamber via an internal optical fiber, and then to an external optical fiber via the flange connector to achieve reliable signal extraction.

[0029] High-speed response voltage and current sensors are used to accurately record the excitation waveform, discharge current, and breakdown time of the test circuit.

[0030] The potential gradient measurement probe array is directly electrically connected to a designated coil coil, ensuring reliable contact between the metal probe and the flat copper wire. Specifically, different measurement points are selected on the coil coil and electrically connected to the metal probe via cables, forming a probe array.

[0031] The high-speed synchronous acquisition unit employs a high-resolution multi-channel digital recorder, specifically the TR-AS 200-12-4 digital impulse voltage testing system manufactured by Strauss GmbH, Germany. It features a TFT color display, a sampling rate of 200MHz, a 12-bit resolution, and four channels. The system comprises a digital recorder, measurement channels, a laser printer, and software. All channels sample synchronously to ensure complete reproduction of the transient process. The measurement system has functions such as automatic recording, automatic analysis, and report output, and the calculation methods comply with relevant international standards.

[0032] The data cleaning and intelligent analysis module is used to process and analyze the collected raw test data. High-voltage test data has the following characteristics: (1) The impulse voltage generator generates extremely strong spatial electromagnetic radiation at the moment of discharge, and interference signals are inevitably coupled in the sensor circuit, resulting in occasional abnormal points in the collected data; (2) The electric field sensor may experience performance fluctuations under high temperature vacuum drying process conditions, and the fiber optic through-cart transfer link is relatively long, which poses a risk of occasional signal interruption, resulting in missing segments in the data record; (3) The amplitude of partial discharge and pre-breakdown signals is weak and is often submerged in background noise, and the background noise level varies with the test voltage level and test type. In view of the above data characteristics, the following artificial intelligence algorithm subunits are embedded in this module: 1. Data Anomaly Detection Unit: An improved isolated forest algorithm based on multi-scale feature fusion and adaptive thresholding is adopted. Considering the large amount of discharge test data and the fact that electromagnetic coupling and instantaneous sensor jitter often lead to data distortion, missing data, or other anomalies, the specific improvements of this algorithm compared with the standard isolated forest algorithm are as follows: (a) At the feature construction level, a multi-dimensional feature vector containing macroscopic waveform parameters, microscopic oscillation features, and statistical distribution features is extracted from the original waveform sampling points, replacing the direct use of the original sampling points as input. This makes the anomaly criteria physically interpretable, and the feature dimension is significantly reduced, improving the algorithm's processing efficiency for high-sampling-rate waveform data; (b) At the threshold determination level, an adaptive threshold based on K-nearest neighbor local density is used instead of a globally fixed threshold, so that the anomaly determination criteria are automatically adjusted according to the differences in data distribution under different voltage levels and test types, avoiding the problem of missed detection under low voltage conditions and misjudgment under high voltage conditions when the fixed threshold is used. The specific steps of the algorithm are as follows: S11. Multi-scale feature extraction. Based on the physical meaning of the high-pressure test data, a three-dimensional feature vector is constructed for each recorded waveform data: Macroscopic waveform parameters include peak voltage, wavefront time T1, and wave tail time T2, which are the core indicators required by the impulse voltage test standard and directly reflect whether the impulse voltage waveform meets the standard waveform requirements. Microscopic oscillation characteristics include the root mean square amplitude of local waveform oscillations and the zero-crossing rate. These two characteristics can effectively distinguish between small oscillations superimposed on normal waveforms and abnormal distortions. Statistical distribution characteristics include the kurtosis and skewness of waveform sampling points, which are sensitive to changes in waveform distribution morphology and thus reflect abnormal discharges or sensor failures.

[0033] S12. Input the above three-dimensional feature vectors into the isolated forest model and calculate the anomaly score for each data point.

[0034] S13. Adaptive Threshold Determination. An adaptive threshold determination method based on K-nearest neighbor local density is adopted. The rationale for this choice is that the discharge characteristics differ significantly under different voltage levels (e.g., 220kV and 500kV) and different test types (lightning impulse and switching impulse), making it difficult to adapt to all operating conditions using a fixed threshold. This method calculates the difference between the anomaly score of each data point and the anomaly scores of its K nearest neighbor data points. If this difference exceeds a preset multiple of the standard deviation of the scores within the neighborhood, it is determined to be an outlier. The threshold is adaptively adjusted according to the local density of the data, avoiding misjudgments caused by empirically setting a global threshold.

[0035] 2. Data Missing Information Imputation Unit: This unit utilizes a deep learning model based on Conditional Generative Adversarial Networks (cGANs) to impute missing data in high-pressure test waveforms. The data studied in this project requires high consistency between the overall waveform shape, wavefront steepness, and wave tail attenuation trend and the actual waveform; that is, the imputed data must be indistinguishable from the actual data at the "distribution level." Therefore, cGANs were chosen for this step because ordinary GAN generators generate data from random noise, making it impossible to specify the imputation location and potentially altering existing normal data segments. cGANs use a missing location mask as a conditional input, allowing the generator to only impute the missing regions while leaving existing data segments unchanged, precisely solving the practical problem of "partially missing, partially complete" high-pressure test data.

[0036] Compared with the standard generative adversarial network (GAN) imputation method, the improvements of this unit are as follows: (a) At the network structure level, the generator adopts an encoder-decoder structure that combines Bi-LSTM and self-attention mechanism. Bi-LSTM captures the causal temporal dependence of the impulse voltage waveform, and the self-attention mechanism captures the global parameter constraint relationship between the wavefront and the wave tail, solving the problem that the standard convolutional generator is difficult to model local temporal and global dependencies at the same time; (b) At the training strategy level, three mask simulation modes are designed for the causes of missing data in high-voltage test data - continuous missing, random point missing and segmented missing, covering three actual fault scenarios: fiber optic link interruption, electromagnetic pulse interference and intermittent sensor fluctuations, so that the model's generalization ability to real missing modes is better than the standard method that only uses random mask training; (c) At the loss function level, a weighted combination of adversarial loss and L1 reconstruction loss is adopted to ensure both the distribution authenticity of the imputed waveform and the point-by-point accuracy of the imputed value, overcoming the numerical bias problem that may be caused by simply using adversarial loss.

[0037] Compared to traditional interpolation methods (such as linear interpolation and spline interpolation), which only utilize local information from known points at both ends of the missing segment and cannot reconstruct the global features of the waveform, performance degrades significantly when the missing segment is long or located in a region of abrupt waveform change. Generative Adversarial Networks (GANs), through adversarial training between the generator and discriminator, can learn the overall distribution patterns of complete waveform data, resulting in incomplete data that is statistically closer to the real data.

[0038] The selection criteria for the model structure are as follows: The generator encoder uses a Bi-LSTM: LSTM excels at capturing timing dependencies. The impulse voltage waveform is a strictly causal timing process (voltage rises from zero to peak and then decays back to zero), with clear sequential dependencies between sampling points. Bi-directional LSTM can simultaneously capture forward and backward information, fully modeling the complete data before and after the missing segment, thereby generating filler data consistent with the context waveform characteristics. The generator introduces a self-attention mechanism: Although the wavefront and wavetail of the impulse voltage waveform are at different time stages, there are overall constraints between the waveform parameters (such as the ratio of wavefront time to wavetail time, the relationship between peak time and half-peak time, etc.). The self-attention mechanism can capture this global dependency across distances, so that the filling data meets the overall constraints of the waveform. The discriminator uses a convolutional neural network: CNN is good at extracting local morphological features of waveforms and can effectively judge whether the transition between the filled segment and the adjacent segment is smooth and natural, and whether the filled segment has unreasonable local oscillations or morphological abnormalities.

[0039] Regarding the training strategy, three missing masks were randomly constructed on the complete experimental data for training: Continuous loss: Simulates continuous data loss caused by loose fiber optic cable interface or long-term interruption of transmission link; Random point loss: Simulates sporadic data loss caused by instantaneous interference of strong electromagnetic pulse on the acquisition channel; Segmented loss: Simulates multiple discontinuous loss caused by multi-channel switching or intermittent fluctuations in sensor performance.

[0040] The total loss function is a weighted combination of adversarial loss and L1 reconstruction loss. The reason for introducing L1 loss is that adversarial loss ensures the "realism" of the filled waveform (indistinguishable at the distribution level), while L1 loss ensures the "accuracy" of the filled waveform (the point-by-point values ​​are close to the true values). The combination of the two makes the filling result conform to the global distribution law of the waveform while maintaining numerical accuracy.

[0041] 3. Feature Extraction and Denoising Unit: An adaptive denoising algorithm based on dual-tree complex wavelet transform and an improved threshold function is adopted. This algorithm was chosen because the partial discharge and pre-breakdown signals of the impulse voltage waveform have pulse characteristics, with weak amplitude and short duration. The denoising algorithm must simultaneously meet three requirements: (a) effectively suppressing background noise; (b) preserving the steep leading edge characteristics and accurate amplitude of the discharge pulse; and (c) ensuring the pulse position does not shift after denoising to guarantee the accuracy of breakdown time measurement. Traditional wavelet transform has two main drawbacks: first, it lacks translation invariance; a small shift in the input signal will cause the wavelet coefficient energy to redistribute across different scales, resulting in a drift in the pulse position after denoising; second, it has limited directional selectivity, making it difficult to accurately capture the singularity position of the discharge pulse. Dual-tree complex wavelet transform, through a parallel decomposition structure of real and imaginary trees, utilizes the approximate Hilbert transform pair relationship between the two trees to achieve approximate translation invariance and better directional selectivity, effectively overcoming the above-mentioned drawbacks. The specific steps are as follows: S31. Dual-tree complex wavelet decomposition. The original waveform undergoes dual-tree complex wavelet transform, using two parallel decomposition paths: a real-tree and an imaginary-tree. A half-sampling-period delay exists between the real-tree and the imaginary-tree, ensuring complementary singular responses of the two trees. The number of decomposition levels is automatically determined based on the waveform sampling frequency and the effective signal bandwidth.

[0042] S32. Differentiated Selection of Wavelet Basis Functions. During the decomposition process, wavelet basis functions are selected differently based on the characteristics of the signal frequency band: Daubechies series wavelet bases are used for low-frequency steady-state signal components, as their orthogonality and tight support are suitable for processing smooth steady-state signals and can better preserve the overall trend of power frequency voltage; Symlet series wavelet bases are used for high-frequency transient signal components, as their good symmetry and small phase distortion when processing pulse signals are beneficial for preserving the rising edge characteristics and accurate occurrence time of discharge pulses.

[0043] S33. Hierarchical Adaptive Threshold Calculation. At each decomposition layer, wavelet coefficient thresholds are calculated separately for the real and imaginary parts. Threshold selection employs an adaptive method based on SURE (Stein Unbiased Likelihood Estimation). SURE is chosen because the threshold selection directly affects the denoising effect in wavelet denoising; an excessively large threshold will filter out weak discharge signals, while an excessively small threshold will leave too much noise residue. SURE is a blind threshold estimation method that does not rely on the actual noise-free signal; it estimates the optimal threshold based on the noisy data itself, mathematically satisfies unbiasedness, and allows for the calculation of the optimal threshold for each decomposition layer. This unit introduces a penalty factor related to the number of decomposition layers based on the standard SURE: the threshold is appropriately increased in higher layers (high-frequency detail layers) because high-frequency layers have a high proportion of noise energy; the threshold is appropriately decreased in lower layers (low-frequency approximation layers) to retain the key features of low-frequency signals. The threshold that minimizes the penalized SURE estimate is selected as the optimal threshold for that layer.

[0044] S34. Improved Threshold Function Shrinkage Processing. Wavelet coefficients below the threshold are shrinked. This unit uses an improved soft-hard threshold compromise function, introducing a continuous transition region. The reason for choosing the compromise function is that: the hard threshold function (directly zeroing below the threshold) is discontinuous at the threshold point, which may introduce additional pseudo Gibbs oscillations after denoising, easily generating false pulses when evaluating weak signals such as partial discharge; the soft threshold function (shrunk by the same amount below the threshold) has good continuity, but will cause overall signal amplitude shrinkage, reducing the accuracy of discharge pulse amplitude estimation. The compromise function in this unit introduces a continuous transition region near the threshold, and the width of the transition region is adaptively correlated with the number of decomposition layers: the transition region of the high-frequency layer is narrow, close to the effect of hard thresholding, preserving the steep leading edge characteristics and accurate amplitude of the discharge pulse; the transition region of the low-frequency layer is wide, close to the effect of soft thresholding, ensuring the overall smoothness of the waveform. Coefficients above the threshold remain unchanged.

[0045] S35. Dual-tree inverse transform reconstruction. Inverse wavelet transforms are performed on the processed real and imaginary coefficients respectively, and the average of the two reconstruction results is taken to obtain the final denoised waveform. Due to the translation invariance of the dual-tree complex wavelet, the position of the denoised partial discharge pulse will not shift, ensuring the accuracy of subsequent breakdown time measurements.

[0046] S36. Normalization and Comparative Analysis. The denoised experimental data, along with parameters such as oil passage number and maximum voltage, are normalized to eliminate differences in data dimensions and magnitudes under different experimental conditions. The results are then compared and analyzed with the wave process simulation results to verify the rationality of the experimental data.

[0047] Compared with the standard wavelet thresholding denoising method, the improvements of this unit are as follows: (a) a dual-tree complex wavelet transform is used instead of the traditional real wavelet transform. The approximate translation invariance of the relationship between the real and imaginary trees is achieved by using the approximate Hilbert transform of the real and imaginary trees, which solves the problem of pulse position drift after traditional wavelet denoising and ensures the accuracy of breakdown time measurement; (b) at the threshold selection level, a penalty factor related to the number of decomposition layers is introduced on the basis of the standard SURE method to suppress high-level noise while retaining the signal characteristics of low-level layers, avoiding the problem of the standard SURE having a small threshold in high-noise environments; (c) at the threshold function level, a soft-hard threshold trade-off function that is adaptively related to the width of the transition region and the number of decomposition layers is used instead of a single hard threshold or soft threshold. The steep leading edge characteristics and accurate amplitude of the discharge pulse are retained in the high-frequency layer (narrow transition region), and the overall smoothness of the waveform is ensured in the low-frequency layer (wide transition region).

[0048] This unit can effectively separate weak characteristic signals such as partial discharge or pre-breakdown from strong background noise, significantly improving the signal-to-noise ratio. Simultaneously, due to the translation invariance of dual-tree complex wavelets, the position of the denoised partial discharge pulse will not shift, ensuring the accuracy of subsequent breakdown time measurements. This unit can also normalize the denoised experimental data along with parameters such as oil passage number and maximum voltage, and compare and analyze the results with wave process simulations.

[0049] The central control and scheduling module, connected to the aforementioned modules, is used to automate the test process, manage data, and implement safety interlocks. This module, centered on a programmable logic controller (PLC) and an industrial control computer, communicates via fiber optic transmission. It sets and adjusts the excitation waveform of the high-voltage test module, controls the precise movements of mechanical actuators such as rotation, lifting, and oil gap adjustment of the equivalent simulation module, and simultaneously triggers the multi-parameter acquisition module for data acquisition. This module also provides a visual operating interface, displaying the test status and voltage / current waveforms in real time, enabling fully automated testing from voltage application, data acquisition, data analysis to report generation.

[0050] Through the coordinated operation of the above modules, this invention can achieve equivalent simulation and accurate testing of various weak insulation links such as transformer winding end insulation, coil potential gradient distribution and riser oil gap, providing a system-level test verification method for the optimized design of transformer insulation structure.

Claims

1. A method for equivalent testing and intelligent data cleaning at the transformer component level, characterized in that, Includes the following steps: An experimental model was built to perform component-level equivalent simulation of the transformer; Provide various types of high-voltage excitations for the test model to conduct equivalent tests; Under high-voltage excitation, multiple physical quantity signals reflecting the insulation state are synchronously acquired as raw experimental data. The raw experimental data undergoes automatic anomaly cleaning, missing data filling, and feature noise reduction, including: S1. An improved isolated forest model based on multi-scale feature fusion and adaptive thresholding is used for outlier detection. The specific steps are as follows: S11. Multi-scale feature extraction: Construct a three-dimensional feature vector for waveform data, including macroscopic waveform parameter features, microscopic oscillation features, and statistical distribution features; S12. Input the three-dimensional feature vector into the isolated forest model and calculate the anomaly score for each data point; S13. An adaptive threshold determination method based on K-nearest neighbor local density is adopted. For each data point, the difference between its abnormal score and the abnormal scores of its K nearest neighbor data points is calculated. If the difference exceeds a preset multiple of the standard deviation of the scores in the neighborhood, it is determined to be an outlier. S2. Using a deep learning model based on conditional generative adversarial networks, missing data imputation of high-voltage test waveforms is performed. This includes: the generator adopts an encoder-decoder structure combining Bi-LSTM and a self-attention mechanism. Bi-LSTM captures the causal temporal dependence of the impulse voltage waveform, while the self-attention mechanism captures the global parameter constraint relationship between the wavefront and the wave tail; three mask simulation modes—continuous missing data, random point missing data, and segmented missing data—are designed for training based on the causes of missing data in the high-voltage test waveforms; and a weighted combination of adversarial loss and L1 reconstruction loss is used as the total loss function. S3. An adaptive denoising algorithm based on dual-tree complex wavelet transform and improved threshold function is used for feature extraction and denoising.

2. The method for equivalent testing and intelligent data cleaning at the transformer component level according to claim 1, characterized in that, The specific steps for feature extraction and noise reduction in S3 are as follows: S31. Perform dual-tree complex wavelet transform on the original waveform, using two parallel decomposition paths: real part tree and imaginary part tree. The number of decomposition levels is automatically determined based on the waveform sampling frequency and the effective signal bandwidth. The real part tree and imaginary part tree have complementary singularity responses to the signal. S32. During the decomposition process, wavelet basis functions are selected differently according to the characteristics of the signal frequency band: Daubechies series wavelet basis is used for low-frequency steady-state signal components, and Symlet series wavelet basis is used for high-frequency transient signal components. S33. At each decomposition layer, wavelet coefficient thresholds are calculated for the real and imaginary parts respectively. The threshold selection adopts an adaptive method based on SURE. S34. Wavelet coefficients below the threshold are contracted by using an improved soft-hard threshold trade-off function. A continuous transition region is introduced near the threshold, and the width of the continuous transition region is adaptively correlated with the number of decomposition layers: the transition region of the high-frequency layer is narrow to preserve the steep leading edge characteristics and accurate amplitude of the discharge pulse; the transition region of the low-frequency layer is wide to ensure the overall smoothness of the waveform and to keep the coefficients above the threshold unchanged. S35. Perform inverse wavelet transform on the processed real and imaginary coefficients respectively, and average the two reconstruction results to obtain the final denoised waveform. S36. Normalize the denoised test data along with the oil passage number and maximum voltage to eliminate the differences in data dimensions and magnitudes under different test conditions.

3. The method for equivalent testing and intelligent data cleaning at the transformer component level according to claim 2, characterized in that, Macroscopic waveform parameter characteristics include voltage peak value, wavefront time T1, and wave tail time T2; microscopic oscillation characteristics include the root mean square amplitude of local waveform oscillations and zero-crossing rate; and statistical distribution characteristics include the kurtosis and skewness of waveform sampling points.

4. The method for equivalent testing and intelligent data cleaning at the transformer component level according to claim 2, characterized in that, Continuous loss is used to simulate continuous data loss caused by loose fiber optic cable interface or long-term interruption of transmission link; random point loss is used to simulate sporadic data loss caused by instantaneous interference of strong electromagnetic pulse on the acquisition channel. Segmented missing segments are used to simulate multiple discontinuous missing segments caused by multi-channel switching or intermittent fluctuations in sensor performance.

5. The method for equivalent testing and intelligent data cleaning at the transformer component level according to claim 2, characterized in that, The test model is built using replaceable simulation units, which include: a winding turn / segment and end insulation simulation unit to simulate the oil gap insulation structure; a coil potential gradient simulation unit to simulate the transformer inner coil or ground potential environment; and a bushing riser long oil gap simulation unit to simulate the equalizing ball long oil gap insulation structure.

6. The method for equivalent testing and intelligent data cleaning at the transformer component level according to claim 5, characterized in that, Methods for acquiring multi-physical quantity signals reflecting the insulation state include: measuring the transient electric field intensity in the oil-paper composite insulation using a broadband electric field sensor; recording the excitation waveform, discharge current, and breakdown time of the test circuit using voltage and current sensors; measuring the potential gradient using a probe array; and collecting and storing test data using a multi-channel digital recorder.

7. The method for equivalent testing and intelligent data cleaning at the transformer component level according to claim 6, characterized in that, The test model is provided with various types of high-voltage excitation through an impulse voltage generator and a power frequency series resonant test device.

8. A transformer component-level equivalent testing and intelligent data cleaning system, characterized in that, The method for implementing the transformer component-level equivalent test and intelligent data cleaning method as described in claim 7 includes an equivalent simulation module, a high-voltage test module, a multi-parameter acquisition module, and a data cleaning and intelligent analysis module connected to a central control and scheduling module. The data cleaning and intelligent analysis module includes a data anomaly detection unit, a data missing data filling unit, and a feature extraction and denoising unit. The data anomaly detection unit uses an improved isolated forest algorithm based on multi-scale feature fusion and adaptive thresholding. The data missing data filling unit uses a deep learning model based on conditional generative adversarial networks to fill in missing data in the high-voltage test waveform. The feature extraction and denoising unit uses an adaptive denoising algorithm based on dual-tree complex wavelet transform and an improved threshold function for feature extraction and denoising.

9. The transformer component-level equivalent testing and intelligent data cleaning system according to claim 8, characterized in that, The winding inter-turn / inter-segment and end insulation simulation unit includes an oil gap insulation structure including an electrostatic ring and a corner ring; the coil potential gradient simulation unit has a metal cylinder set on the inner diameter side of the winding model and grounded, and a measurement probe interface and an optical fiber through disk are pre-set at a specified coil coil disc; the bushing riser long oil gap simulation unit includes a replaceable equalizing ball insulation structure, which is divided into bare electrode, insulated electrode, shaped electrode and paper tube electrode, and can be combined and selected according to different voltage levels.

10. The transformer component-level equivalent testing and intelligent data cleaning system according to claim 8, characterized in that, The impulse voltage generator includes a generator body, a weakly damped voltage divider, a cutoff device, and a load capacitor.

Images