Transformer turn-to-turn short circuit detection method based on high-frequency current time-frequency energy distribution characteristics

Abstract

The invention provides a transformer turn-to-turn short circuit detection method based on high-frequency current time-frequency energy distribution characteristics. The transformer turn-to-turn short circuit detection method comprises the following steps: Step 1, sensor installation and signal acquisition; step 2, spectrum analysis and optimal hierarchy selection are carried out; performing spectral analysis on the signal and determining an optimal level of wavelet decomposition, performing wavelet transform discretization and reconstruction, and performing multi-scale decomposition on the signal to obtain a high-frequency component and a low-frequency component under each scale; step 3, carrying out signal de-noising processing; step 4, extracting a characteristic value; extracting three indexes including a pulse amplitude ratio RA, a high-frequency energy focusing degree Ef and a wavelet entropy variation degree CVH; and the turn-to-turn short circuit detection condition of the transformer is judged according to the characteristic value through a three-level diagnosis rule. The number of wavelet analysis dynamic decomposition layers is determined in combination with FFT frequency domain analysis, and collaborative optimization of noise suppression and fault feature enhancement is realized by using a scale adaptive wavelet threshold denoising algorithm. And a time domain-frequency domain-time frequency combined characteristic system is further constructed, a quantitative grading diagnosis standard is established, and the detection capability and the diagnosis reliability of the slight short circuit are remarkably improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of transformer intelligence, and particularly relates to a transformer inter-turn short circuit detection method based on high-frequency current time-frequency energy distribution characteristics. BACKGROUND

[0002] Dry-type transformer is an important part of power system, and its safe and stable operation is of great significance to the reliability and economy of the entire power grid. Transformer faults mainly occur in the core, bushing, tap changer, winding and other parts. Since the winding is the main structure inside the transformer that bears electric, thermal and mechanical stress, winding fault is a common type of fault in various transformer faults. According to statistical data, among the winding faults of dry-type transformers, inter-turn short circuit faults account for 50% to 60%.

[0003] In the early stage of dry-type transformer inter-turn fault, the main characteristics are partial discharge or local overheating. However, with the generation of partial discharge and local overheating, the transformer insulation system may further deteriorate, and the vibration characteristics may also cause the internal structure to loosen and fall off, eventually leading to inter-turn breakdown and inter-turn short circuit fault. The harm of inter-turn short circuit fault is huge, mainly in two aspects: on the one hand, inter-turn short circuit fault will generate a huge short-circuit current in the transformer, which will cause strong electric force, resulting in deformation of the internal structure of the transformer, so that the transformer cannot operate normally, and even cause fire and burn the transformer; on the other hand, inter-turn short circuit fault will change the excitation of the transformer and cause relatively violent vibration of the transformer, affecting the service life and stability of the equipment. Therefore, it is necessary to detect and distinguish the transformer inter-turn fault in the early stage, and prevent the fault from further developing by means of power outage maintenance and other means, to avoid more serious inter-turn short circuit, inter-turn breakdown, phase-to-phase short circuit, main insulation breakdown and other faults of the transformer.

[0004] Some of the prior art proposes three diagnostic techniques for interpreting scan frequency response analysis for power transformer fault inter-turn identification. They include cross-correlation coefficient factor technique, relative factor technique and diagnostic technique for identifying transformer turns to short circuit faults at high and low voltage sides.

[0005] In the method for identifying dry-type transformer turn-to-turn short-circuit fault based on the difference between voltage and current unbalance degrees, ANSYS finite element analysis software is adopted to establish a simulation model based on the field-circuit coupling principle, and the electromagnetic parameter distribution characteristics, power factor, loss factor, active power and other electrical parameters of turn-to-turn short-circuit faults occurring at different positions are studied. The difference between the phase voltage and the phase current unbalance degree is extracted as a characteristic quantity for judging the turn-to-turn short-circuit fault, thereby providing a theoretical basis for online detection of transformer turn-to-turn short-circuit. In the method for distinguishing transformer magnetizing inrush current based on magnetizing impedance change, the magnetizing impedance changes in different ways under the conditions of turn-to-turn short-circuit and magnetizing inrush current, thereby distinguishing the two kinds of large currents. Moreover, this method does not need to consider the transformer parameters and overall system parameters, and uses separate-phase braking to act very quickly and stably. In the method for identifying oil-immersed transformer turn-to-turn short-circuit fault based on fusion analysis of electrical and thermal characteristics, the electromagnetic characteristic parameters and temperature parameters of the winding are comprehensively considered, and a method for identifying oil-immersed transformer turn-to-turn short-circuit based on comprehensive electrical and thermal characteristics is proposed, which can effectively identify slight turn-to-turn short-circuit. The digital twin technology is mainly used to simulate and analyze the short-circuit current and winding temperature under different operating conditions and different turn-to-turn fault conditions, and has high accuracy.

[0006] Frequency response analysis (FRA) as a method for identifying mechanical defects of transformers may be able to identify turn-to-turn fault potentials and provide help in this case. In the frequency response technique to recognize turn-to-turn insulation deterioration in transformer winding, the ability of FRA to discover turn-to-turn insulation deterioration and the rate of turn-to-turn aging is developed. For this purpose, two different test objects are used, and the FRA characteristics under various conditions are recorded. In the modeling and practical application of transformer turn-to-turn short-circuit, a 220kV three-winding transformer with turn-to-turn short-circuit is equivalent to a four-winding transformer, and the short-circuit winding is equivalent to two different windings. The voltages of each winding of the transformer are derived, and it is proved that the voltage of the low-voltage winding decreases more obviously after short-circuit.

[0007] The existing technology for detecting turn-to-turn short-circuit of transformers has the following defects: 1. The traditional protection device relies on the change of power frequency current for detection, but the initial fault current amplitude of turn-to-turn short-circuit is only 1%-5% of the rated current, and the high-frequency component is overwhelmed by the power frequency signal, resulting in serious lack of sensitivity, and often triggering action only when the number of short-circuit turns exceeds 5%, which delays the best handling opportunity.

[0008] 2. Existing high-frequency detection scheme mostly uses fixed-bandwidth filter (such as 10kHz-1MHz) for noise suppression, but the main frequency of actual fault signal dynamically changes with short-circuit position and degree, and fixed-bandwidth design is easy to cause effective signal loss or noise residue.

[0009] 3. The signal denoising link generally uses a unified threshold wavelet algorithm, which excessively suppresses high-level decomposition coefficients, leading to weak fault pulses being misjudged as noise. In addition, existing diagnostic standards mostly rely on a single parameter (such as total harmonic distortion rate), which is easy to produce misjudgment when electromagnetic interference or load fluctuation occurs.

[0010] These defects collectively result in the existing technology being difficult to accurately identify slight inter-turn short circuit at an early stage, and failing to meet the demand of modern power systems for precise perception of equipment state. SUMMARY

[0011] The technical problem to be solved by the present application is to provide a transformer inter-turn short circuit detection method based on high-frequency current time-frequency energy distribution characteristics, to improve the detection capability and diagnostic reliability of slight short circuit.

[0012] To solve the above technical problems, the technical solution adopted by the present application is: The transformer inter-turn short circuit detection method based on high-frequency current time-frequency energy distribution characteristics comprises: Step1, sensor installation and signal acquisition; capture the short-circuit transient current signal of the current signal at a set sampling rate, select the ground-induced current time window with the initial trigger as the midpoint and a set time length, and perform unit conversion and DC bias removal on the collected core ground current signal; Step2, frequency spectrum analysis and optimal level selection; perform frequency spectrum analysis on the signal and determine the optimal level of wavelet decomposition, perform wavelet transform discretization and reconstruction, and perform multi-scale decomposition on the signal to obtain high-frequency components and a low-frequency component at each scale; Step3, signal denoising processing; Step4, feature value extraction; extract the pulse amplitude ratio R A , high-frequency energy focusing degree E f , and wavelet entropy variation degree CV H ; and judge the transformer inter-turn short circuit detection condition through three-level diagnostic rules.

[0013] In the above Step1, the HFCT sensor is installed on the transformer core ground line in a buckle manner to suppress signal reflection interference; the oscilloscope is used to capture the short-circuit transient current signal at a sampling rate of 40M / s The above Step2 specifically comprises the following steps: Step 2.1: Perform spectral analysis on the signal using Fast Fourier Transform (FFT) and filter high-frequency components by power threshold; determine the optimal level of wavelet decomposition based on the high-frequency range to ensure that fault characteristic frequencies are accurately covered in wavelet multi-scale decomposition. Step 2.2: Discretize the wavelet transform; obtain the discrete binary wavelet transform, introduce the scale, i.e. the smoothing function ψ(x), perform discrete binary wavelet transform decomposition calculation and reconstruction; Step 2.3: Select the mother wavelet and decompose the signal at multiple scales, retaining only the detail coefficients, i.e. the high-frequency components, and eliminating the low-frequency approximation coefficients to avoid low-frequency interference; this will yield the high-frequency components and one low-frequency component at each scale.

[0014] The optimal level calculation formula in Step 2.1 above is: ; Among them, OL represents the optimal level. f s Sampling rate ，f max For The maximum frequency of the high-frequency component.

[0015] The specific steps in Step 2.2 above are as follows: Taking the scaling parameter, x, as the position coordinates for signal analysis, we can obtain the discrete binary wavelet transform, and introduce the scaling function ψ(x): ; satisfy ; ; in, To reconstruct the wavelet, define For smoothing operators, For smoothing coefficients or sequences; For input digital signals, These are the detail components of the discrete binary wavelet transform, generated by filtering the smoothed signal from the previous scale using a high-pass filter; therefore, the discrete binary wavelet transform decomposition algorithm is: ; ; The discrete binary wavelet transform reconstruction algorithm is as follows: ; in for The return to common track, , and They are respectively , and The corresponding sequence becomes a band-pass filter.

[0016] The specific steps of Step 2.3 are as follows: Select Daubechies 4, i.e. db4 wavelet, as the mother wavelet to perform multi-scale decomposition on the signal; only retain the detail coefficient, i.e. high-frequency component, and eliminate the low-frequency approximation coefficient to avoid low-frequency interference; obtain the high-frequency component and a low-frequency component at each scale, which can be expressed by the following formula: ; The output signal S, i.e. the transformer core grounding current, is subjected to multi-scale decomposition to obtain a low-frequency component A and a high-frequency component D reflecting the overall change trend of the signal.

[0017] The specific steps of signal denoising processing in Step 3 are as follows: Design a scale-dependent dynamic threshold function. The high-frequency component of the signal obtained by wavelet multi-scale decomposition is essentially a matrix: ; Taking the absolute value of the matrix can obtain: ; Perform denoising processing and fault feature extraction on the signal: ; wherein is the high-frequency component matrix of the signal after wavelet multi-scale decomposition , the screening result of a single element in the matrix is an intermediate variable for signal denoising and fault feature extraction, which performs preliminary screening on the high-frequency component matrix by hard threshold value, and only retains the signals significantly exceeding the noise level; Adopt the median absolute deviation (MAD) method to adaptively calculate the threshold value, set a dynamic soft threshold value for the detail coefficient of each scale signal, and the denoising process is as follows: MAD (Median Absolute Deviation) estimates the noise intensity through the median absolute deviation of the detail coefficient, =1-0.15(j-1) is a decay factor that increases with the decomposition scale; the threshold function adopts an improved semi-soft threshold method to establish a continuous transition between the hard threshold value and the soft threshold value: ; D j (k) is the detail coefficient at scale j after wavelet multi-scale decomposition, reflecting the high-frequency details of the signal at that scale. j (k)|>2Tj: perform soft threshold processing, subtract the threshold value T j , retain strong features while smoothing the transition; Tj <|D j (k)|≤2T j : Establish a continuous transition between hard threshold and soft threshold by 50% attenuation, avoid signal mutation; |D j (k)|≤T j : Directly set to zero, completely suppress noise.

[0018] The feature value extraction in Step4 above; Extract the pulse amplitude ratio R A , high-frequency energy focusing degree E f and wavelet entropy variation CV H Three indicators are calculated as follows: Pulse amplitude ratio R A : ; Where, x is the input ground current signal; High-frequency energy focusing degree E f : ; Where, cD'j is the denoised wavelet coefficient of the j layer, x is the original signal; Wavelet entropy variation CV H is calculated as follows: ; Where, p i Indicates the probability density of wavelet coefficient energy, H j is the entropy of the jth layer wavelet coefficient, std Indicates the standard deviation of each scale entropy.

[0019] The three-level diagnosis rule in Step4 above is as follows: (1) Normal state; When any one of RA<1, Ef<0.05 or CVH<0.05 is satisfied, it is determined that the system is fault-free; (2) Early short circuit, that is, short circuit involves 1~3 turns; When R A increases to 1~2, E f increases to 5%-15%, and CV H fluctuates between 0.05-0.12, a pre-warning signal is triggered; (3) Serious short circuit, that is, the range of short circuit is greater than 3 turns; When R A >2.5, E f >15% and CV H >0.12, it is determined as serious short circuit and a trip is triggered.

[0020] In a preferred scheme, the detection method further comprises: Step 5, LightGBM diagnosis model; using the feature value database extracted by wavelet transform as the input of the model, establishing a current feature value diagnosis model based on the LightGBM machine learning algorithm; the diagnosis index uses the precision, recall and F1-score.

[0021] The LightGBM diagnosis model processes the wide-frequency current waveform generated by the inter-turn short circuit fault through wavelet transform, and obtains the R A , E f , CV H Three characteristic quantities; introduce the machine learning model LightGBM, take the multi-dimensional fault characteristics as the input, and establish a current feature value diagnosis model based on LightGBM: The LightGBM algorithm first initializes the classification state of the feature parameters: ; In the formula, i is the sample number of the data set; j is the feature number of each sample; xij is the feature parameter; k is the category; and F is the predicted state.

[0022] After initialization, the probability of the feature parameters in each category is calculated, and iteration is continuously performed, and the iteration is stopped when the maximum iteration number is reached; the probability calculation formula is: ; In the formula, m is the iteration number, and K is the total number of categories in the classification problem; Then, the algorithm calculates the negative gradient of the feature parameters; in each iteration, the negative gradient is the new target to be fitted, and the calculation formula is: ; In the formula, is the true probability of fitting the feature parameters. After obtaining the value of the negative gradient, the leaf node value after splitting the model needs to be calculated; the node value determines how the samples are divided into different leaf nodes, and has a direct impact on the performance of the model. The calculation formula is: ; In the formula, h is the number of leaf nodes, and Rh,k,m is the sample set on the leaf node. After obtaining the leaf node value after splitting, the algorithm updates and optimizes the model according to the following formula: ; In the formula, is the learning rate set for the model, I is the index of a leaf node sample set, and H is the number of leaf nodes of the decision tree.

[0023] The final model is as follows: ; In the formula, M is the total number of constructed decision trees.

[0024] The transformer inter-turn short circuit detection method based on high-frequency current time-frequency energy distribution characteristics mentioned in the application proposes an intelligent diagnosis method combining wavelet transform and Light Gradient Boosting Machine (LightGBM) machine learning algorithm. The method faces wide-frequency fault current waveform signals, determines the wavelet analysis dynamic decomposition layer number based on wavelet multi-scale analysis combined with FFT frequency domain analysis, extracts pulse response features after scale adaptive wavelet threshold denoising, and then constructs a lightweight machine learning model to classify and judge the fault severity, realizing automatic identification of three fault levels of "normal state", "initial inter-turn short circuit" and "severe inter-turn short circuit". The method extracts transient high-frequency features through wavelet transform, and realizes intelligent classification of fault levels combined with LightGBM, significantly improving the detection ability and diagnosis reliability of slight short circuits. BRIEF DESCRIPTION OF DRAWINGS The application will be further described below in combination with the drawings and examples: Figure 1 is a schematic diagram of the high-frequency current sensor HFCT used in the application; Figure 2 is an example diagram of signal denoising processing and wavelet analysis results in Example 1 of the application; Figure 3 is a schematic diagram of the current signal when there is no short circuit in Example 2 of the application; Figure 4 is a schematic diagram of data wavelet dynamic layering processing and soft threshold denoising when there is 1 turn short circuit in Example 2 of the application; Figure 5 is a schematic diagram of data wavelet dynamic layering processing and soft threshold denoising when there are 5 turns short circuit in Example 2 of the application; Figure 6 is a schematic diagram of data wavelet dynamic layering processing and soft threshold denoising when there are 10 turns short circuit in Example 2 of the application; Figure 7 is a schematic diagram of the Histogram-based decision tree algorithm principle in Example 2 of the application Figure 8 is a flowchart of the application. DETAILED DESCRIPTION

[0025] The technical solutions of the application will be described in detail below in combination with the drawings and examples.

[0026] Example 1: 1. Sensor Installation and Signal Acquisition like Figure 1 As shown, an HFCT sensor with a bandwidth of 10kHz-50MHz was first used, installed 1.5m from the transformer core grounding wire using a snap-fit ​​method to suppress signal reflection interference; a 40M / s sampling rate ensured no aliasing of the 20MHz high-frequency components. A PicoScope 6407 oscilloscope was used to capture the short-circuit transient current signal at a sampling rate of 40M / s. A 1s time window of ground induced current, with the initial trigger as the midpoint, was selected in the signal. Unit conversion and DC bias removal were performed on the acquired core grounding current signal to eliminate low-frequency interference and enhance the contrast of the signal frequency characteristics. This provides clean input for subsequent wavelet analysis.

[0027] 2. Spectrum Analysis and Optimal Hierarchy Selection Fast Fourier Transform (FFT) is used to perform spectral analysis on the signal, and high-frequency components are filtered out using a power threshold. The optimal level of wavelet decomposition is determined based on the high-frequency range to ensure accurate coverage of fault characteristic frequencies in the wavelet multi-scale decomposition. The optimal level is calculated using the following formula: ; Among them, OL represents the optimal level. f s Sampling rate ，f max For The maximum frequency of the high-frequency component.

[0028] Discretize the wavelet and wavelet transform. Take the scale parameter, x, as the position coordinates for signal analysis; this yields the discrete binary wavelet transform, introducing the scale, i.e., the smoothing function ψ(x): ; ; ; To reconstruct the wavelet, define For smoothing operators, For smoothing coefficients (sequence). For input digital signals, These are the detail components of the discrete binary wavelet transform, generated by filtering the smoothed signal from the previous scale using a high-pass filter; therefore, the discrete binary wavelet transform decomposition algorithm is: ; ; The discrete binary wavelet transform reconstruction algorithm is as follows: ; in for complex common rail, , and respectively , and The corresponding sequence becomes a band-pass filter.

[0029] Daubechies 4, namely db4 wavelet, is selected as the mother wavelet, and the signal is subjected to multi-scale decomposition. Only the detail coefficient, namely the high-frequency component, is retained, and the low-frequency approximation coefficient is removed to avoid low-frequency interference. The high-frequency component and a low-frequency component under each scale can be obtained, which can be expressed by the following formula: ; The output signal S, namely the transformer core grounding current in the application, can obtain the low-frequency component A and the high-frequency component D (detail part) reflecting the overall change trend of the signal after multi-scale decomposition. The high-frequency component in the signal to be detected contains the high-frequency component reflecting the signal change (caused by the transformer turn-to-turn short circuit). Therefore, analyzing the high-frequency component of the signal can effectively realize the transformer turn-to-turn short circuit fault diagnosis.

[0030] Aiming at the transient pulse and wide frequency characteristics of the dry-type transformer turn-to-turn short circuit fault current signal, wavelet transform is adopted for multi-layer multi-resolution analysis (MRA) to realize signal time-frequency feature extraction. Wavelet transform can realize localized analysis of transient signals through multi-scale decomposition, and reveal the frequency distribution law of the turn-to-turn short circuit fault characteristics. The Daubechies wavelet (db4) is selected as the mother wavelet function in the application, and the db4 wavelet basis function has tight support, high-order vanishing moments and good smoothness, which can effectively extract the transient characteristics of the turn-to-turn short circuit pulse signal. The classical implementation method of wavelet transform multi-scale decomposition is Mallat algorithm, which is based on the filter bank decomposition principle to recursively decompose the signal. In order to further reduce the calculation complexity and meet the real-time requirement, the lifting wavelet algorithm is adopted to optimize the wavelet decomposition process. The lifting wavelet algorithm realizes multi-scale decomposition of the signal through three steps of signal splitting, prediction and updating.

[0031] 3. Signal denoising processing A scale-dependent dynamic threshold function is designed. The high-frequency component of the signal obtained by wavelet multi-scale decomposition is essentially a matrix: ; The absolute value of this matrix can be obtained: ; Denoising processing and fault feature extraction are performed on the signal: ; wherein is the high-frequency component matrix after wavelet multi-scale decomposition is the screening result of a single element in the matrix, is an intermediate variable for signal denoising and fault feature extraction, which is a hard threshold preliminary screening of the high-frequency component matrix, and only the signals significantly higher than the noise level are retained; The threshold value is adaptively calculated by the median absolute deviation (MAD) method, and a dynamic soft threshold is set for each scale signal detail coefficient. The denoising process is: wherein MAD, i.e. Median Absolute Deviation, estimates the noise intensity by the median absolute deviation of the detail coefficient, =1-0.15(j-1) is a decay factor that increases with the decomposition scale. This design breaks through the limitations of traditional uniform threshold: more weak fault features are retained at high levels (j=5~6) (a5=0.4, a6=0.25), and strong denoising is implemented at low levels (j=1) (a1=1) to suppress broadband noise. The threshold function uses an improved semi-soft threshold method to establish a continuous transition between hard threshold and soft threshold: ; D j (k) is the detail coefficient at scale j after wavelet multi-scale decomposition, reflecting the high-frequency details of the signal at that scale. j (k)|>2Tj: soft threshold processing is performed, and the threshold T j is subtracted to retain strong features while smoothing the transition; T j <|D j (k)|≤2T j : decay by 50% to establish a continuous transition between hard threshold and soft threshold, avoiding signal discontinuity; |D j (k)|≤T j : directly set to zero, completely suppressing noise.

[0032] The fixed threshold denoising method sets a fixed threshold T for each layer of detail coefficients of the signal and hard-truncates all coefficients. This method has good noise suppression effect on noise signals, but it can cause signal discontinuity and distortion. The soft threshold denoising method adjusts the detail coefficients by smoothing: not only suppresses noise, but also retains part of the small amplitude signal components, which can effectively reduce signal distortion and is suitable for extracting weak fault features, such as Figure 2 , which shows an example of signal denoising and wavelet analysis results.

[0033] 4. Feature value extraction The feature extraction module calculates the pulse amplitude ratio (R A ), high-frequency energy focusing degree (Ef ) and wavelet entropy variation (CV H ) of the three indicators.

[0034] ; where x is the input ground current signal. Under normal conditions, R A <1.5, 1~3 turns short-circuit to 1.5-2.0. ; where, cD'j is the denoised wavelet coefficient of the j layer, x is the original signal. The normal time-frequency domain E f <5%.

[0035] ; where, p i denotes the probability density of the wavelet coefficient energy, H j is the entropy of the jth layer wavelet coefficient, std denotes the standard deviation of each scale entropy. The normal time CVH<0.05 short-circuit induced time-frequency energy distribution disorder.

[0036] Inter-turn short-circuit fault will induce high-frequency pulse signal in the core ground current, the pulse amplitude increases significantly. RA can directly reflect the strength of abnormal pulse in the signal, the greater the amplitude of the burst signal, the greater the value of RA. High-frequency energy focusing degree is used to quantify the proportion of high-frequency component in the total energy of the signal. In the early stage of inter-turn short-circuit fault, the signal mainly contains noise component, the entropy value is larger and the change is smaller. With the development of the fault, the regularity of high-frequency pulse signal is enhanced, and the entropy variation gradually increases.

[0037] Based on the 50 sets of experimental data (including 0-10 turns short-circuit samples) of the 100KVA dry-type transformer scaled model experimental platform, the three-level diagnosis rules are established, and the judgment accuracy reaches 95%: (1) Normal state When any one of RA<1, Ef<0.05 or CVH<0.05 is satisfied, it is determined that the system is fault-free.

[0038] (2) Early short-circuit (1~3 turns) When R A rise to 1~2, E f rise to 5%-15%, CV H fluctuate at 0.05-0.12, the early warning signal is triggered.

[0039] (3) Severe short-circuit (>3 turns) When RA >2.5、E f >15% and CV H >0.12 is determined as a serious short circuit and triggers tripping.

[0040] The detection method also includes: Step 5, LightGBM diagnosis model; using the feature value database extracted by wavelet transform as the input of the model, establishing a current feature value diagnosis model based on the LightGBM machine learning algorithm; the diagnosis index uses the precision Precision, recall Recall and F1 value F1-score.

[0041] The above embodiment 2: An inter-turn short circuit test is performed on a 100KVA dry-type transformer scaled model experimental platform, the current core is grounded, and the current signals under the conditions of no short circuit, 1-turn short circuit, 5-turn short circuit and 10-turn short circuit are tested respectively. Figures 3-6 As shown in the wavelet decomposition and threshold processing method of the present application, the denoised waveform diagram and decomposition diagram can be obtained.

[0042] The present application introduces a machine learning model LightGBM, takes the fault features extracted by wavelet transform as the input data set, establishes a current feature value diagnosis model based on LightGBM, and improves the training efficiency and generalization ability. In the inter-turn short circuit diagnosis, LightGBM can accurately identify the nonlinear relationship between the features and the fault level, and realize high-precision multi-classification judgment.

[0043] The LightGBM diagnosis model processes the wide-frequency current waveform generated by the inter-turn short circuit fault through wavelet transform, obtains the R A 、E f 、CV H Three characteristic quantities; introduce a machine learning model LightGBM, take multi-dimensional fault features as input, establish a current feature value diagnosis model based on LightGBM: LightGBM is a gradient boosting model based on decision tree learning; a number of weak classifiers are combined to become a strong learning machine, thereby constructing the final classification model.

[0044] The LightGBM algorithm first initializes the classification state of the feature parameters: ; In the formula, i is the sample number of the data set, j is the feature number of each sample, xij is the feature parameter, k is the category, and F is the predicted state.

[0045] After initialization, the probability of the feature parameter on each category is calculated, and iteration is continuously carried out, and the iteration is stopped when the maximum iteration number is reached; the probability calculation formula is: ; In the formula, m is the iteration number, and K is the total number of categories in the classification problem; Then, the algorithm calculates the negative gradient of the feature parameter; in each iteration, the negative gradient is the new target to be fitted, and the calculation formula is: ; In the formula, is the true probability of fitting the feature parameter. After obtaining the value of the negative gradient, the leaf node value after model splitting needs to be calculated; the node value determines how the sample is divided into different leaf nodes, and has a direct impact on the performance of the model. The calculation formula is: ; In the formula, h is the number of leaf nodes, and Rh,k,m is the sample set on the leaf node; After obtaining the leaf node value after splitting, the algorithm updates and optimizes the model according to the following formula: ; In the formula, is the learning rate set for the model, I is the index of a leaf node sample set, and H is the number of leaf nodes of the decision tree.

[0046] The final model obtained is as follows: ; In the formula, M is the total number of decision trees constructed.

[0047] The training data set is based on 50 groups of experimental data (including 0-10 turns of short circuit samples) of the 100KVA dry-type transformer scaled model experimental platform. The fault data set is input into the LightGBM model for training, and the trained model is used to diagnose different degrees of fault samples to test the classification performance indicators of the model. The categories of the test data set are divided into normal state, early short circuit (1-3 turns of short circuit are involved), and serious short circuit (the range of short circuit is greater than 3 turns). The diagnosis index uses precision (Precision), recall (Recall), and F1 value (F1-score). The higher the three indicators are, the better the diagnosis effect is, and the combination of the three can more truly reflect the diagnosis effect. The main parameters of the LightGBM model are shown in Table 1.

[0048]

[0049] The diagnostic results are compared with the fault categories of the original current data. The accuracy rate of the LightGBM fault diagnosis model based on wavelet transform for diagnosing normal operation data reaches 92.30%, the recall rate reaches 90.67%, and the F1 value reaches 91.48%. For early fault data diagnosis, the accuracy rate reaches 88.23%, the recall rate reaches 89.12%, and the F1 value reaches 91.76%. For severe fault data diagnosis, the accuracy rate reaches 96.78%, the recall rate reaches 93.55%, and the F1 value reaches 94.65%. By learning the data processed by wavelet transform, the model can induce more general fault diagnosis rules. When facing new fault samples, it can also make more accurate judgments based on the learned general rules, thereby improving the accuracy rate in the light fault diagnosis scenario.

[0050] .

Claims

1. A transformer inter-turn short-circuit detection method based on the time-frequency energy distribution characteristics of high-frequency current, characterized in that, include: Step 1: Sensor installation and signal acquisition; capture short-circuit transient current signals with a set sampling rate, select the ground induced current time window with the initial trigger as the midpoint, and perform unit conversion and DC bias removal on the acquired iron core ground current signal; Step 2, Spectrum Analysis and Optimal Level Selection: Perform spectrum analysis on the signal and determine the optimal level of wavelet decomposition. After wavelet transform discretization and reconstruction, perform multi-scale decomposition on the signal to obtain high-frequency components and a low-frequency component at each scale. Step 3: Signal denoising processing; Step 4: Feature value extraction; extract the pulse amplitude ratio R. A High-frequency energy focusing degree E f and wavelet entropy variation CV H Three indicators; and the transformer's inter-turn short circuit detection status is judged by the characteristic values ​​through a three-level diagnostic rule.

2. The transformer inter-turn short-circuit detection method based on the time-frequency energy distribution characteristics of high-frequency current according to claim 1, characterized in that, In Step 1, a high-frequency current sensor (HFCT) is used and installed at the transformer core grounding wire in a snap-fit ​​manner to suppress signal reflection interference; an oscilloscope is used to capture the short-circuit transient current signal at a sampling rate of 40M / s.

3. The transformer inter-turn short-circuit detection method based on the time-frequency energy distribution characteristics of high-frequency current according to claim 2, characterized in that, Step 2 specifically includes the following steps: Step 2.1: Perform spectral analysis on the signal using Fast Fourier Transform (FFT) and filter high-frequency components by power threshold; determine the optimal level of wavelet decomposition based on the high-frequency range to ensure that fault characteristic frequencies are accurately covered in wavelet multi-scale decomposition. Step 2.2: Discretize the wavelet transform; obtain the discrete binary wavelet transform, introduce the scale, i.e. the smoothing function ψ(x), perform discrete binary wavelet transform decomposition calculation and reconstruction; Step 2.3: Select the mother wavelet and decompose the signal at multiple scales, retaining only the detail coefficients, i.e. the high-frequency components, and eliminating the low-frequency approximation coefficients to avoid low-frequency interference; this will yield the high-frequency components and one low-frequency component at each scale.

4. The transformer inter-turn short-circuit detection method based on the time-frequency energy distribution characteristics of high-frequency current according to claim 3, characterized in that, The optimal level calculation formula in Step 2.1 is as follows: ； Among them, OL represents the optimal level. f s Sampling rate ，f max for The maximum frequency of the high-frequency component.

5. The transformer inter-turn short-circuit detection method based on the time-frequency energy distribution characteristics of high-frequency current according to claim 4, characterized in that, The specific steps in Step 2.2 are as follows: Taking the scaling parameter, x, as the position coordinates for signal analysis, we can obtain the discrete binary wavelet transform, and introduce the scaling function ψ(x): ； satisfy ; ； in, To reconstruct the wavelet, define For smoothing operators, For smoothing coefficients or sequences; For input digital signals, These are the detail components of the discrete binary wavelet transform, generated by filtering the smoothed signal from the previous scale using a high-pass filter; therefore, the discrete binary wavelet transform decomposition algorithm is: ； ； The discrete binary wavelet transform reconstruction algorithm is as follows: ； in for The return to common track, , and They are respectively , and The corresponding sequence is called a bandpass filter.

6. The transformer inter-turn short-circuit detection method based on the time-frequency energy distribution characteristics of high-frequency current according to claim 5, characterized in that, The specific steps in Step 2.3 are as follows: The Daubechies 4 (db4) wavelet is selected as the mother wavelet for multi-scale decomposition of the signal. Only detail coefficients, i.e., high-frequency components, are retained, while low-frequency approximation coefficients are removed to avoid low-frequency interference. The high-frequency components and one low-frequency component at each scale are obtained, which can be expressed by the following formula: ； The output signal S, which is the transformer core grounding current, is decomposed into low-frequency component A and high-frequency component D, which reflect the overall trend of the signal change.

7. The transformer inter-turn short-circuit detection method based on the time-frequency energy distribution characteristics of high-frequency current according to claim 6, characterized in that, The specific steps of signal denoising processing in Step 3 include: The scale-dependent dynamic threshold function is designed based on the high-frequency components of the signal obtained from wavelet multi-scale decomposition, which is essentially a matrix: ； Taking the absolute value of a matrix yields: ； Noise reduction and fault feature extraction are performed on the signal: ； in It is the high-frequency component matrix after wavelet multi-scale decomposition. The filtering results for a single element. It is an intermediate variable for signal denoising and fault feature extraction. It performs a hard threshold screening on the high-frequency component matrix to retain only signals that significantly exceed the noise level. The median absolute deviation (MAD) method is used to adaptively calculate the threshold, and a dynamic soft threshold is set for the signal detail coefficients at each scale. The denoising process is as follows: MAD estimates noise intensity using the median absolute deviation of the detail coefficients. =1-0.15(j-1) is the decay factor that increases with the decomposition scale; the threshold function adopts an improved semi-soft thresholding method to establish a continuous transition between hard and soft thresholds: ； D j (k) represents the detail coefficients at scale j after wavelet multi-scale decomposition, reflecting the high-frequency details of the signal at that scale; D j (k)|>2Tj: Perform soft thresholding and subtract the threshold T. j It retains strong features while providing a smooth transition; T j <|D j (k)|≤2T j : Attenuate by 50% to establish a continuous transition between the hard and soft thresholds, avoiding abrupt signal changes; |D j (k)|≤T j : Set directly to zero to completely suppress noise.

8. The transformer inter-turn short-circuit detection method based on the time-frequency energy distribution characteristics of high-frequency current according to claim 7, characterized in that, In Step 4, feature value extraction is performed; the pulse amplitude ratio R is extracted. A High-frequency energy focusing degree E f and wavelet entropy variation CV H The three indicators are calculated as follows: Pulse amplitude ratio R A : ； Where x is the input grounding current signal; High-frequency energy focusing E f : ； in, cD′j For the denoised first j Layer wavelet coefficients, x The original signal; Wavelet entropy variation CV H The calculation is as follows: ； in, p i This represents the probability density of the wavelet coefficient energy. H j Let be the entropy of the wavelet coefficients of the j-th layer. std It represents the standard deviation of entropy at each scale.

9. The transformer inter-turn short-circuit detection method based on the time-frequency energy distribution characteristics of high-frequency current according to claim 8, characterized in that, The three-level diagnostic rules in Step 4 are as follows: (1) Normal state; when any one of RA<1, Ef<0.05 or CVH<0.05 is met, the system is determined to be fault-free; (2) Early short circuit, i.e., the short circuit involves 1 to 3 turns; when R is satisfied at the same time A Upgrade to 1~2, E f Increase to 5%-15%, CV H A warning signal is triggered when the value fluctuates between 0.05 and 0.

12. (3) Severe short circuit, i.e., the short circuit involves more than 3 turns; when R A >2.5, E f >15% and CV H A reading >0.12 indicates a severe short circuit and triggers a trip.

10. The transformer inter-turn short-circuit detection method based on the time-frequency energy distribution characteristics of high-frequency current according to claim 1, characterized in that, The detection method further includes: Step 5: LightGBM Diagnostic Model; Using the feature value database extracted by wavelet transform as the input of the model, a current feature value diagnostic model based on the LightGBM machine learning algorithm is established; the diagnostic indicators used are precision, recall, and F1 score.

11. The transformer inter-turn short-circuit detection method based on the time-frequency energy distribution characteristics of high-frequency current according to claim 10, characterized in that, The LightGBM diagnostic model described above processes the broadband current waveform generated by inter-turn short-circuit faults using wavelet transform to obtain the RC current data. A E f CV H Three feature quantities; a lightweight gradient descent tree (LightGBM) machine learning model is introduced, using multi-dimensional fault features as input, to establish a current feature value diagnostic model based on LightGBM: The LightGBM algorithm first initializes the classification state of the feature parameters: ； In the formula, i is the number of samples in the dataset; j is the number of features for each sample; xij is the feature parameter; k is the class; and F is the prediction state. After initialization, the probability of the feature parameters in each category is calculated and iterated continuously until the maximum number of iterations is reached, at which point the iteration stops. The probability calculation formula is as follows: ； In the formula, m is the number of iterations, and K is the total number of categories in the classification problem; Next, the algorithm calculates the negative gradient of the feature parameters; in each iteration, the negative gradient is the new target to be fitted, and its calculation formula is: ； In the formula, To fit the true probabilities of the feature parameters; after obtaining the negative gradient value, it is necessary to calculate the leaf node values ​​after the model splits; the node values ​​determine how samples are divided into different leaf nodes, which directly affects the model's performance; the calculation formula is: ； In the formula, h is the number of leaf nodes, and Rh,k,m is the sample set on the leaf nodes; After obtaining the values ​​of the split leaf nodes, the algorithm updates and optimizes the model according to the following formula: ； In the formula, The learning rate set for the model, where I is the index of a sample set for a certain leaf node, and H is the number of leaf nodes in the decision tree; The final model form is shown below: ； In the formula, M represents the total number of decision trees constructed.

Images