A transformer fault detection method based on transformer ground current

Abstract

The present application relates to the technical field of electrical fault detection, and particularly relates to a transformer fault detection method based on transformer grounding current. The method comprises: acquiring real-time working condition parameters of the transformer, and performing benchmark dynamic correction on standard cloud models of each fault state; inputting measured multi-dimensional fault features into a forward cloud generator, and combining the corrected cloud model parameters to calculate the membership degree of the fault type; using hyper-entropy parameters to scale the absolute distance of feature deviation to obtain a dispersion index, and accordingly adaptively calculating a credibility weight and converting it into a basic probability assignment value of an independent evidence source; extracting the transient pulse energy proportion in the grounding current signal, and dynamically reconstructing the evidence source weight in the sudden transient and gradual steady state scene; calculating the conflict coefficient after the synthesis of each evidence source, switching the standard D-S synthesis or weighted average fusion rule in a double-track system, and generating a comprehensive trust evaluation index; and finally determining the transformer fault result by using a double threshold.

Description

Technical Field

[0001] This invention relates to the field of electrical fault detection technology, and in particular to a transformer fault detection method based on transformer grounding current. Background Technology

[0002] As a core hub device in the power system, the stability of the transformer's operation directly affects the safety and stability of the entire power grid. Among the various operation and maintenance parameters of a transformer, the core grounding current is a key physical indicator reflecting latent faults such as multi-point grounding of the core, insulation aging, and internal partial discharge. Most existing transformer fault detection methods acquire grounding current signals in real time using sensors, extracting features in the time domain, spectrum, or long-period trends, and then combining fuzzy theory, neural networks, or the classic Depsster-Schaffer (DS) evidence theory to comprehensively fuse multi-dimensional features from multiple sources, thereby achieving the assessment of the transformer's operating status and fault diagnosis.

[0003] However, the actual operating environment of transformers is extremely complex. Grounding current signals are often severely interfered with by strong transient electromagnetic pulses and random noise from sensors. Furthermore, the dynamic switching of the transformer's own operating conditions can cause the physical distribution benchmark of feature data to drift. Existing technologies typically use static models or fixed judgment thresholds, which are difficult to adaptively adjust according to real-time operating conditions, easily leading to distortion in the quantitative assessment of feature membership. More importantly, when facing extreme operating conditions such as sudden transient interference, single-channel sensor anomalies, or severe deviations in multi-source signals, strong evidentiary conflicts can arise between different feature dimensions such as trend, spectrum, and time domain. Traditional DS synthesis rules are prone to vetoing when processing highly conflicting data, leading to false alarms or missed alarms in the system. Summary of the Invention

[0004] To overcome the above shortcomings, this invention provides a transformer fault detection method based on transformer grounding current, aiming to improve the problems of high false alarm and missed alarm rates caused by feature reference drift due to dynamic switching of operating conditions and high conflict of multi-dimensional features under extreme interference in traditional detection methods.

[0005] This invention provides the following technical solution: a transformer fault detection method based on transformer grounding current, comprising: The core grounding current signal of the transformer under test is acquired in real time, the current signal is preprocessed and multi-dimensional fault features are extracted; the multi-dimensional fault features include at least trend features, spectral features and time domain features. Real-time operating parameters are collected, and the standard cloud model corresponding to each preset fault state is dynamically corrected to eliminate the interference of operating condition fluctuations on the core current. Using the modified standard cloud model, the membership certainty of the multidimensional fault features relative to different fault types is calculated respectively; Combining the distribution discreteness of the multidimensional fault characteristics in the modified standard cloud model, the membership certainty is transformed into basic probability allocation values ​​of independent evidence sources through weighted and normalized mapping of distribution discreteness; the independent evidence sources include trends, spectrum, and time domain. The conflict handling mechanism based on the characteristics of sudden transient and gradual steady state is used to modify the DS synthesis rule, fuse the basic probability allocation values, and calculate the comprehensive trust evaluation index for each fault type. The transformer fault result is determined based on the comprehensive trust evaluation index.

[0006] Preferably, the current signal is preprocessed and multidimensional fault features are extracted, including: The acquired core grounding current signal is denoised and DC bias interference is eliminated to obtain a purified current signal. The steady-state increment and rate of change are calculated based on the time series of the purification current signal to construct the trend characteristics. The purification current signal is subjected to frequency domain transformation to calculate the harmonic distortion rate and the frequency band energy ratio including low-order harmonic frequency band and high-frequency response frequency band, and the spectrum characteristics are constructed. Extract the abrupt peak attributes from the time-domain waveform of the purification current signal, calculate the peak factor and pulse factor, and construct the time-domain features.

[0007] Preferably, the standard cloud model corresponding to each preset fault state is dynamically corrected based on a benchmark, including: Acquire at least one real-time operating condition parameter, including transformer load, bus voltage, and operating temperature; Based on the preset correlation mapping relationship between the real-time operating condition parameters and the standard cloud model parameters, the corresponding expected correction amount and entropy correction amount are calculated in real time. Using the expected correction amount and the entropy correction amount, the initial parameters of the standard cloud model are adaptively compensated to generate the corrected standard cloud model.

[0008] Preferably, calculating the membership certainty of the multidimensional fault features relative to different fault types includes: The multidimensional fault characteristics are input as input cloud droplets to the corresponding positive cloud generator; Based on the expected, entropy, and hyperentropy parameters of the modified standard cloud model, the random Gaussian membership values ​​of the input cloud droplets under the membership degree distribution of each fault type are calculated. The random Gaussian membership value is used as the membership certainty of the multidimensional fault feature relative to different fault types.

[0009] Preferably, the method for obtaining the distribution discreteness includes: Calculate the absolute distance between the measured multidimensional fault characteristics and the expected value of the corrected standard cloud model; The absolute distance is scaled using the hyperentropy parameter of the modified standard cloud model to obtain a dispersion index characterizing the discreteness of the distribution of the multidimensional fault features in the corresponding cloud model.

[0010] Preferably, converting the membership certainty into basic probability allocation values ​​for independent evidence sources includes: The credibility weight of each feature dimension is adaptively calculated based on the dispersion index, wherein the larger the dispersion index, the smaller the corresponding credibility weight. The confidence weights are used to multiply and weight the corresponding membership certainty, and then normalized and mapped within a preset fault identification space to generate the basic probability allocation value with trend, spectrum, and time domain as independent evidence sources.

[0011] Preferably, the comprehensive trust evaluation index for each fault type is calculated, including: Extract the percentage of transient pulse energy from the core grounding current signal; When the proportion of transient pulse energy is higher than a preset first threshold, it is determined to be a sudden transient-dominated scenario, the weight of the trend feature evidence source is reduced, and the weight of the spectral feature and time domain feature evidence sources is increased. When the proportion of transient pulse energy is lower than a preset second threshold, it is determined to be a gradually steady-state dominant scenario, and the weights of each independent evidence source remain unchanged.

[0012] Preferably, the comprehensive trust evaluation index for each fault type is calculated, including: Calculate the conflict coefficient after combining the evidence from each independent source; When the conflict coefficient is greater than the preset conflict threshold, it is determined that there is a serious conflict between the evidence sources. At this time, the weighted average fusion rule is used to replace the standard DS synthesis rule, and the arithmetic mean of the basic probability allocation values ​​of each evidence source is used as the fusion result. When the conflict coefficient is less than or equal to the conflict threshold, the standard DS synthesis rule is used for fusion. The comprehensive trust evaluation index for each fault type is calculated.

[0013] Preferably, determining the transformer fault result based on the comprehensive trust evaluation index includes: When the comprehensive trust evaluation index of a certain fault type is greater than or equal to the first judgment threshold, the fault type is determined to be valid. When the comprehensive trust evaluation index of all fault types is less than the second judgment threshold, the transformer is judged to be in normal condition. When the comprehensive trust evaluation index is between the first and second judgment thresholds, it is judged as an early warning state, and continuous monitoring and trend tracking are triggered.

[0014] The present invention has the following beneficial effects: 1. In this invention, by introducing dynamic cloud model correction for operating conditions and adaptive weighting of dispersion, the interference of transformer operating condition drift and random environmental noise can be eliminated in real time, which greatly improves the accuracy of quantitative assessment of fault characteristics and effectively avoids response distortion caused by conventional fixed benchmarks.

[0015] 2. In this invention, by constructing a scene weight dynamic reconstruction and dual-track conflict fusion switching mechanism, the problem of strong conflict and non-convergence of multi-source evidence under extreme interference is effectively solved, misleading signal interference is comprehensively weakened, and the false alarm rate and false alarm rate of transformer fault diagnosis can be reduced. Attached Figure Description

[0016] Figure 1 This is a flowchart of a transformer fault detection method based on transformer grounding current proposed in this invention; Figure 2 This is a graph showing the performance comparison test results of different fault detection schemes provided in the embodiments of the present invention; Figure 3 The confusion matrix diagram of the solution provided in this embodiment of the invention under different fault types. Detailed Implementation

[0017] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] Example 1 In a first embodiment of the present invention, the present invention provides a transformer fault detection method based on transformer grounding current, such as... Figure 1 As shown, it includes the following steps: S1. Acquire the core grounding current signal of the transformer under test in real time, preprocess the current signal and extract multi-dimensional fault features; the multi-dimensional fault features include at least trend features, spectrum features and time domain features; Preferably, the current signal is preprocessed and multidimensional fault features are extracted, including: The acquired core grounding current signal is denoised and DC bias interference is eliminated to obtain a purified current signal. The steady-state increment and rate of change are calculated based on the time series of the purification current signal to construct the trend characteristics. The purification current signal is subjected to frequency domain transformation to calculate the harmonic distortion rate and the frequency band energy ratio including low-order harmonic frequency band and high-frequency response frequency band, and the spectrum characteristics are constructed. Extract the abrupt peak attributes from the time-domain waveform of the purification current signal, calculate the peak factor and pulse factor, and construct the time-domain features.

[0019] Specifically, in this embodiment, the real-time acquired transformer core grounding current signal first enters the preprocessing module. Because strong electromagnetic uncertainties exist at the transformer site, and DC bias may be generated due to the influence of rail transit or DC transmission systems, this embodiment first performs noise reduction and DC bias interference removal on the original current signal.

[0020] The system employs a low-pass filter with a cutoff frequency of 1000Hz to remove high-frequency random white noise from the signal. Subsequently, the severity of DC bias is quantitatively assessed by calculating the positive and negative half-wave asymmetry of the signal before purification. The formulas for calculating the positive and negative half-wave asymmetry are as follows: ; In the above formula, The asymmetry is between the positive and negative half-waves. This represents the peak value of the positive current signal within the current power frequency cycle. This represents the negative trough value of the current signal within the current power frequency cycle. In this embodiment, the preset bias threshold is set to 0.15. When the calculated asymmetry... When the value exceeds 0.15, the transformer is determined to be under DC bias interference. At this point, the system invokes a filter, using a flat linear structural element with a length of 5ms to perform multi-scale switching filtering on the signal. This operation effectively removes even-order harmonic groups excited by DC bias and the DC trend term, ultimately outputting a purified current signal.

[0021] The system applies a sliding time window to the aforementioned purification current signal, with the window length set to 10 minutes. The effective value of the current power frequency current is obtained by performing root mean square (RMS) calculation on the data within the window. Based on the time series of the purification current signal, the steady-state increment of the current effective value relative to the historical reference window is calculated using the following formula: ; In the above formula, For steady-state increments; This represents the effective value of the current within the current sliding window. A historical reference for the effective current value at a given time point is used. In this embodiment, this historical reference is taken from a historical window of the same operating conditions 24 hours prior. Further, the rate of change of the effective current value is calculated using the following formula: ; In the above formula, The rate of change; This represents the time difference between the current window and the historical baseline window. This operation can characterize the acceleration of steady-state RMS growth, thus distinguishing between slow natural aging and faster-growing fault degradation.

[0022] The system performs a Fast Fourier Transform (FFT) on the purification current signal, with a sampling frequency set to 10kHz and a data point count of 4096 points per transform. After completing the frequency domain transformation, the total harmonic distortion (THD) of the grounding current is first calculated using the following formula: ; In the above formula, Total harmonic distortion (THD); This is the effective value of the fundamental frequency current. for The effective value of the harmonic current is taken as follows, with the upper limit of the harmonic order set at 50 in this embodiment. Simultaneously, to capture specific high-frequency disturbances excited by localized weak discharges, the energy proportion of a specific frequency band is calculated using the following formula: ; In the above formula, This refers to the energy percentage of a specific frequency band. For the signal at frequency The power spectral density value at that location; The sampling frequency; and This corresponds to the frequency band boundary between the low-order harmonic frequency band and the high-frequency response frequency band. In this embodiment, Set to 500Hz. The setting is 2000Hz. This operation reflects the proportion of power spectral density within a specific frequency range to the total power across the entire frequency band, and is used to lock onto localized weak electric arcs or high-frequency electromagnetic oscillations that are not obvious in the time domain.

[0023] The system captures abrupt spikes in the grounding current signal waveform to identify transient processes such as sudden metallic foreign object contact and instantaneous insulation breakdown. This embodiment extracts the time-domain waveform characteristics of the purification current signal within a 20ms power frequency period window. First, the crest factor is calculated using the following formula: ; In the above formula, Peak factor; This represents the absolute maximum value of the purification current signal within the current power frequency cycle. This represents the effective value of the current for that cycle. The pulse factor is then calculated using the following formula: ; In the above formula, The pulse factor; The first in the current power frequency cycle Current amplitude at each sampling point; This represents the total number of sampling points within a single power frequency cycle. This operation can quantify the ratio of the signal peak value to the rectified average value and has extremely high sensitivity to sudden transient pulses in the waveform, enabling precise capture of transient arc impacts caused by suspended discharge of metallic foreign objects or intermittent overlap.

[0024] S2. Collect real-time operating condition parameters and perform benchmark dynamic correction on the standard cloud model corresponding to each preset fault state to eliminate the interference of operating condition fluctuations on the core current. Preferably, the standard cloud model corresponding to each preset fault state is dynamically corrected based on a benchmark, including: Acquire at least one real-time operating condition parameter, including transformer load, bus voltage, and operating temperature; Based on the preset correlation mapping relationship between the real-time operating condition parameters and the standard cloud model parameters, the corresponding expected correction amount and entropy correction amount are calculated in real time. Using the expected correction amount and the entropy correction amount, the initial parameters of the standard cloud model are adaptively compensated to generate the corrected standard cloud model.

[0025] Specifically, in this embodiment, the system periodically acquires the real-time operating condition parameters of the transformer under test, thereby eliminating the interference of environmental and external load changes on the grounding current judgment standard. The real-time operating condition parameters are collected online through a sensor array and power quality monitoring terminal deployed on the transformer side, including transformer load, bus voltage, and operating temperature.

[0026] In specific implementation, the transformer load is characterized by the ratio of the current load power to the rated power, i.e., the load factor, and the transformer top oil temperature is used to characterize the operating temperature. As a feasible implementation, the real-time operating parameters of a 110kV transformer at the current moment are: load factor 0.75, bus voltage 112kV, and top oil temperature 55°C. After obtaining the real-time operating parameters, the system compares them with the baseline operating conditions. The baseline operating parameters in this embodiment are set as follows: baseline load factor 0.50, baseline bus voltage 110kV, and baseline top oil temperature 20°C. Based on the preset correlation mapping relationship between the real-time operating parameters and the standard cloud model parameters, the corresponding expected correction and entropy correction are calculated in real time. The formula for calculating the expected correction is as follows: ; In the above formula, This is the expected correction amount; This represents the current real-time load factor. The baseline load factor; This is the current top oil temperature; The reference top oil temperature; and This is a preset expected correction coefficient. In this embodiment, Set to 0.45. The value is set to 0.02. The formula for calculating the entropy correction is as follows: ; In the above formula, This is the entropy correction factor; This is the current bus voltage; The reference bus voltage; and This is a preset entropy correction coefficient. In this embodiment, Set to 0.15. Set to 0.05.

[0027] After calculating the expected correction and entropy correction, these are used to adaptively compensate the initial parameters of the standard cloud model, generating a corrected standard cloud model. The formula for adaptive compensation is as follows: ; ; In the above formula, This represents the expected value of the revised standard cloud model. The initial expectation of the uncorrected standard cloud model; The corrected standard cloud model entropy; The initial entropy is the uncorrected standard cloud model. This operation involves applying the deterministic trend influence and random uncertainty influence generated by macroscopic external conditions directly to the internal boundary parameters of the standard cloud model through algebraic superposition, thereby achieving dynamic drift of the judgment benchmark.

[0028] S3. Using the modified standard cloud model, calculate the membership certainty of the multidimensional fault features relative to different fault types; Preferably, calculating the membership certainty of the multidimensional fault features relative to different fault types includes: The multidimensional fault characteristics are input as input cloud droplets to the corresponding positive cloud generator; Based on the expected, entropy, and hyperentropy parameters of the modified standard cloud model, the random Gaussian membership values ​​of the input cloud droplets under the membership degree distribution of each fault type are calculated. The random Gaussian membership value is used as the membership certainty of the multidimensional fault feature relative to different fault types.

[0029] In this embodiment, the system utilizes the forward cloud generator algorithm module to calculate the membership certainty of multidimensional fault features relative to different fault types. This step quantitatively evaluates the degree of conformity of the measured fault features in each dimension as input cloud droplets under various predetermined fault state standard cloud models by using the measured fault features in each dimension as input cloud droplets.

[0030] The specific implementation process is described using the trend characteristics of iron core grounding current as an example. A specific fault type is preset in the system as multi-point grounding fault of the iron core. The trend characteristics corresponding to this fault type have corresponding corrected standard cloud model parameters, including the corrected expectation and the corrected entropy. During the calculation process, the forward cloud generator first combines the corrected entropy parameter and the hyperentropy parameter to generate a random entropy value that satisfies a Gaussian distribution. The formula for calculating the random entropy value is as follows: ; In the above formula, The generated random entropy value; The corrected standard cloud model entropy; The hyperentropy parameter is a preset value; in this embodiment, the hyperentropy parameter is preset to 0.05. To generate random numbers that conform to a standard normal distribution.

[0031] Subsequently, the forward cloud generator, based on the aforementioned random entropy value and the corrected expectation, calculates the random Gaussian membership value of the measured extracted fault feature amplitude under the membership degree distribution of the multi-point grounding fault type in the iron core. The formula for calculating the random Gaussian membership value is as follows: ; In the above formula, The calculated random Gaussian membership value; The amplitude of the fault features was extracted from actual measurements; This represents the expected value of the revised standard cloud model. The random entropy value is used to quantify the geometric distance between the measured fault feature point and the expected center of the corresponding fault model, and to nonlinearly map the feature deviation in the physical space to a continuous confidence interval between zero and one through a Gaussian exponential basis function constrained by random entropy, so as to characterize the probability that the feature belongs to the specific fault type.

[0032] Following the exact same calculation process, the system performs parallel mapping calculations on spectral features and time-domain features in the forward cloud generator, and outputs the membership certainty of each feature dimension of trend, spectrum and time domain relative to different preset fault types, thereby completing the transformation from the original physical features to the probabilistic representation of the cloud model.

[0033] S4. Combining the distribution discreteness of the multidimensional fault characteristics in the modified standard cloud model, the membership certainty is transformed into the basic probability allocation value of independent evidence sources through weighted and normalized mapping of the distribution discreteness; the independent evidence sources include trends, spectrum and time domain. Preferably, the method for obtaining the distribution discreteness includes: Calculate the absolute distance between the measured multidimensional fault characteristics and the expected value of the corrected standard cloud model; The absolute distance is scaled using the hyperentropy parameter of the modified standard cloud model to obtain a dispersion index characterizing the discreteness of the distribution of the multidimensional fault features in the corresponding cloud model.

[0034] In this embodiment, the system quantitatively calculates the distribution discreteness of measured multidimensional fault features in the corresponding fault type cloud model to assess the dispersion of the current feature data and provide a data foundation for subsequent adaptive weighting of independent evidence sources. The process of obtaining the distribution discreteness mainly depends on the geometric distance and spatial scale scaling between the feature amplitude and the core parameters of the cloud model.

[0035] First, the system calculates the absolute distance between the measured multidimensional fault characteristics and the expected value of the corrected standard cloud model. The formula for calculating the absolute distance is as follows: ; In the above formula, This is the calculated absolute distance; The amplitude of specific-dimensional fault features extracted from actual measurements; This represents the expected value of the corresponding modified standard cloud model.

[0036] Subsequently, the system uses the hyperentropy parameter of the corrected standard cloud model to scale the absolute distance, obtaining a dispersion index characterizing the dispersion of the multidimensional fault features in the corresponding cloud model. The formula for calculating the dispersion index is as follows: ; In the above formula, The discreteness index is calculated to characterize the discreteness of the multidimensional fault feature distribution. The absolute distance obtained from the preceding steps; The hyperentropy parameters, which are preset to correspond to the fault type, serve as feasible implementation data. Specifically, regarding the trend characteristic dimension, the hyperentropy parameters... Set to a fixed constant of 0.05.

[0037] Through the above processing, the system calculates the corresponding dispersion index for trend features, spectral features, and time-domain features in parallel. The magnitude of this index directly reflects the discreteness and aggregation state of the measured fault features in the fault space. The larger the index value, the more discrete the distribution of the feature in the corresponding cloud model, and the more unstable the data performance. This provides a quantitative basis for subsequently adjusting the credibility weight of evidence sources in reverse based on the dispersion index.

[0038] Preferably, converting the membership certainty into basic probability allocation values ​​for independent evidence sources includes: The credibility weight of each feature dimension is adaptively calculated based on the dispersion index, wherein the larger the dispersion index, the smaller the corresponding credibility weight. The confidence weights are used to multiply and weight the corresponding membership certainty, and then normalized and mapped within a preset fault identification space to generate the basic probability allocation value with trend, spectrum, and time domain as independent evidence sources.

[0039] In this embodiment, the system converts the membership certainty of each feature dimension obtained in the preceding steps into basic probability assignment values ​​of independent evidence sources that meet the requirements of DS evidence theory. This conversion process is accomplished by adaptively adjusting the credibility weights and implementing normalization mapping within the recognition framework to ensure that features with high dispersion and strong uncertainty adaptively reduce their decision contribution during the fusion stage.

[0040] First, the system adaptively calculates the credibility weights for each feature dimension based on the dispersion index. In this embodiment, the formula for calculating the credibility weights is as follows: ; In the above formula, The confidence weights for specific feature dimensions are calculated. This is the dispersion index for this feature dimension; The preset adjustment coefficient serves as a feasible implementation data point. Set to a fixed constant of 0.25.

[0041] Subsequently, the system uses the confidence weights to perform a product weighting on the corresponding membership certainty. The product weighting calculation formula for a specific fault type is as follows: ; In the above formula, For specific feature dimensions of fault type The temporary intermediate value after product weighting; This represents the credibility weight for that feature dimension; This feature dimension is relative to the fault type. The degree of certainty of membership.

[0042] Finally, the system performs normalization mapping within a preset fault identification space to generate the basic probability allocation values ​​with trend, spectrum, and time domain as independent evidence sources. The fault identification space includes several preset deterministic fault types and a set of uncertain terms representing unknown states. The calculation formula for the normalization mapping is as follows: ; ; In the above formula, For the final generated independent source of evidence regarding the fault type The basic probability allocation value; For independent sources of evidence regarding uncertainties The basic probability allocation value; The total number of deterministic fault types within the preset fault identification space.

[0043] Through the above processing, the system maps the product-weighted feature confidence performance to a probability quality distribution that strictly satisfies the condition that the summation equals 1. Simultaneously, the confidence margin lost due to weight decay—that is, the difference between 1 and the confidence weight—is safely transferred and allocated to the set uncertainty term in the identification space. This allocation mechanism prevents drastic dispersion in feature dimensions from being incorrectly attributed to a specific fault, thereby blocking the formation of conflicting evidence at the data source and ensuring a safe boundary for subsequent fusion of independent evidence sources in the trend, spectrum, and time domains.

[0044] Preferably, the comprehensive trust evaluation index for each fault type is calculated, including: Extract the percentage of transient pulse energy from the core grounding current signal; When the proportion of transient pulse energy is higher than a preset first threshold, it is determined to be a sudden transient-dominated scenario, the weight of the trend feature evidence source is reduced, and the weight of the spectral feature and time domain feature evidence sources is increased. When the proportion of transient pulse energy is lower than a preset second threshold, it is determined to be a gradually steady-state dominant scenario, and the weights of each independent evidence source remain unchanged.

[0045] In this embodiment, the system performs weight correction on each independent evidence source before fusion based on the conflict handling mechanism of sudden transient and gradual steady-state characteristics, so as to dynamically adapt to the physical signal characteristics of the transformer at different fault evolution stages and ensure the accuracy of fusion decision.

[0046] First, the system extracts the proportion of transient pulse energy in the core grounding current signal. Specifically, a high-pass filter is used to extract the high-frequency pulse component from the purification current signal, and the energy ratio is calculated within a set time window. .

[0047] Subsequently, the system compares the calculated transient pulse energy percentage with a preset threshold range, thereby triggering different weight adjustment strategies. In this embodiment, the preset first threshold is set to a fixed constant of 0.30, and the preset second threshold is set to a fixed constant of 0.08.

[0048] When the proportion of transient pulse energy exceeds a preset first threshold, the current transformer fault is determined to be in a sudden transient-dominated scenario. In this scenario, the system reduces the weight of the trend feature evidence source and increases the weight of the spectral feature and time-domain feature evidence sources. The dynamic weight adjustment formula for the sudden transient-dominated scenario is as follows: ; ; ; In the above formula, , and These are the adjusted weights for trend, spectrum, and time-domain feature evidence sources, respectively. , and These are the initial weights of each independent source of evidence before adjustment; This is a preset trend weight decay coefficient. As a feasible implementation data point, the trend weight decay coefficient... Set to a fixed constant of 0.15.

[0049] Through the above processing, when the proportion of transient pulse energy exceeds a preset first threshold, the current transformer fault is determined to be in a sudden transient-dominated scenario. In this scenario, short-term jumps in ground current can easily lead to distortion in the identification of trend features based on long-period statistics, while spectral and time-domain features have higher fault sensitivity. Therefore, by introducing a trend weight attenuation coefficient, the weight of trend feature evidence sources is significantly reduced, and the reduced weight is equally compensated to spectral and time-domain feature evidence sources, thereby improving the system's response speed and diagnostic accuracy to sudden faults.

[0050] When the proportion of transient pulse energy is lower than a preset second threshold, the current transformer fault is determined to be in a gradually steady-state dominant scenario. At this time, the core current exhibits a slow trend drift, indicating that the transformer is in a steady-state development period of multi-point grounding degradation or gradual insulation aging, and the trend characteristics have high reference value. Therefore, the system directly uses the initial weights adaptively calculated by the aforementioned dispersion index without external scenario intervention to ensure the smoothness and reliability of the fusion process.

[0051] If the proportion of transient pulse energy is between the second threshold and the first threshold, the system determines that it is currently in a transitional state, and maintains the initial weights of each independent evidence source unchanged until the signal features cross a clear scene boundary.

[0052] S5. Based on the conflict handling mechanism of sudden transient and gradual steady-state characteristics, the DS synthesis rule is modified, the basic probability allocation value is fused, and the comprehensive trust evaluation index of each fault type is calculated. Preferably, the comprehensive trust evaluation index for each fault type is calculated, including: Calculate the conflict coefficient after combining the evidence from each independent source; When the conflict coefficient is greater than the preset conflict threshold, it is determined that there is a serious conflict between the evidence sources. At this time, the weighted average fusion rule is used to replace the standard DS synthesis rule, and the arithmetic mean of the basic probability allocation values ​​of each evidence source is used as the fusion result. When the conflict coefficient is less than or equal to the conflict threshold, the standard DS synthesis rule is used for fusion. The comprehensive trust evaluation index for each fault type is calculated.

[0053] In this embodiment, the system fuses multi-source information, including the weighted trend, spectrum, and basic probability allocation values ​​of independent time-domain evidence sources, to calculate a comprehensive trust evaluation index for each fault type. This step, by introducing conflict identification and a dual-track fusion switching mechanism, effectively solves the problem of evidence conflicts that may arise from multi-sensor data under extreme interference.

[0054] First, the system calculates the conflict coefficient after combining the evidence from each independent source. The formula for calculating the conflict coefficient is as follows: ; In the above formula, The calculated conflict coefficient; , and These represent independent evidence sources in the trend, spectrum, and time domain, respectively, for their corresponding state subsets. The basic probability assignment value. By traversing all incompatible state combinations with empty intersections and accumulating their probability products, it is possible to accurately reflect the non-convergence or directional deviation of each feature dimension when determining the transformer fault type.

[0055] The system will then calculate the conflict coefficient. The value is compared with a preset conflict threshold. In this embodiment, the conflict threshold is set to a feasible implementation data of 0.70. When the conflict coefficient... When the value is greater than 0.70, a serious conflict is determined to exist between the independent sources of evidence. In this case, the system uses a weighted average fusion rule instead of the standard DS synthesis rule, taking the arithmetic mean of the basic probability assignments of each source of evidence as the final fusion result. The formula for calculating the arithmetic mean of the comprehensive trust evaluation index for each fault type is as follows: ; In the above formula, Fault types generated by fusion in severe conflict scenarios A comprehensive trust evaluation index is used. This operation provides a fault-tolerance guarantee. In extreme conditions such as severe single-channel sensor failure or strong pulse interference, the basic probability allocation of a certain feature may produce extremely strong misleading biases, leading to a veto of the classic DS synthesis rule or extreme diagnostic conclusions that are completely contrary to intuition. Using an arithmetic mean to smoothly allocate probability quality can weaken the negative interference of extreme conflicting evidence and ensure that the fusion decision does not undergo directional reversal.

[0056] When the conflict coefficient When the score is less than or equal to 0.70, all evidence sources are considered to be within the normal tolerable conflict range. In this case, the system uses the DS synthesis rule for fusion. The standard synthesis formula for calculating the comprehensive trust evaluation index for each fault type is as follows: ; This operation enables strong positive focusing of multidimensional evidence. Under normal conditions where the judgment results of each feature dimension are relatively consistent and the conflict is small, it multiplies and aggregates all probability flows pointing to the intersection of the target fault in the form of orthogonal summation.

[0057] Through the aforementioned dual-track fusion processing, the system calculates the comprehensive trust evaluation index for each fault type. Finally, by finding the maximum value among these comprehensive trust evaluation indices and combining it with a preset output threshold, the system can accurately pinpoint specific internal transformer core fault types such as overheating or cooling, multiple grounding points, and partial discharge.

[0058] S6. Determine the transformer fault result based on the comprehensive trust evaluation index.

[0059] Preferably, determining the transformer fault result based on the comprehensive trust evaluation index includes: When the comprehensive trust evaluation index of a certain fault type is greater than or equal to the first judgment threshold, the fault type is determined to be valid. When the comprehensive trust evaluation index of all fault types is less than the second judgment threshold, the transformer is judged to be in normal condition. When the comprehensive trust evaluation index is between the first and second judgment thresholds, it is judged as an early warning state, and continuous monitoring and trend tracking are triggered.

[0060] In this embodiment, the decision control unit uses preset dual thresholds to quantify and evaluate the comprehensive trust evaluation index of each fault type output by the multi-source information fusion module, thereby outputting the determination result of the transformer operating status. As a feasible implementation data, the preset first determination threshold is set to 0.65, and the preset second determination threshold is set to 0.20. The specific determination process is as follows: When the comprehensive trust evaluation index for a certain fault type is greater than or equal to the first judgment threshold, the fault type is determined to be valid. This step identifies the specific fault mode by evaluating whether the index of a specific fault type reaches the lower confidence limit for diagnosis; once the condition is met, the system immediately outputs the corresponding fault alarm signal and identifies the specific fault category.

[0061] When the comprehensive confidence evaluation index for all fault types is less than the second judgment threshold, the transformer is determined to be in a normal state. This step assesses the overall steady state of the transformer core grounding current by establishing a safe upper limit for confidence in the absence of abnormal energy disturbances; when all fault indicators are suppressed below the second judgment threshold, the system determines that the transformer core grounding current is within the safe steady-state control range.

[0062] When the maximum value of the comprehensive trust evaluation index for each fault type falls between the first judgment threshold and the second judgment threshold, a warning state is determined. This step aims to accurately identify the confidence interval of the fault evolution latency period or boundary abnormal state; once the warning state is triggered, the system adaptively increases the data collection frequency from once per hour to once every five minutes, and simultaneously starts continuous monitoring and trend tracking.

[0063] Example 2 To verify the effectiveness and superiority of the transformer fault detection method based on transformer grounding current described in this invention under actual electrical scenarios, this embodiment introduces a standard transformer fault electrical dataset for simulation and comparative testing. The electrical dataset contains a total of 3000 sets of transformer core grounding current sample data obtained through long-term field monitoring and simulated fault tests. It covers various typical electrical fault scenarios, including normal operation, multi-point grounding of the core, partial discharge, and overheating caused by metal foreign object connections. Furthermore, the dataset contains varying proportions of reference drift data caused by strong pulse electromagnetic interference and operating condition switching.

[0064] To objectively evaluate the technical effectiveness, this embodiment establishes three parallel comparison schemes for fault identification testing on the same standard electrical dataset: Comparative Scheme 1 (Neural Network Scheme): Employs a conventional deep backpropagation (BP) neural network model. The extracted time-domain, spectral, and trend feature matrices are directly used as input vectors, nonlinearly mapped through a multilayer perceptron, and finally, the fault classification probability is output by a Softmax layer.

[0065] Comparative Scheme 2 (Ordinary DS Synthesis Scheme): Employs traditional DS evidence theory methods. The basic probability assignment values ​​for each feature dimension are obtained through static calculation based on fixed membership functions, and the fusion stage directly uses classic orthogonal and synthesis rules to integrate multi-source information.

[0066] Table 1 Performance Comparison Test Results of Different Fault Detection Schemes Overall diagnostic accuracy 85.2% 81.7% 93.7% False alarm rate of strongly interfering conflict samples 11.8% 14.5% 4.2% Latent fault sample false negative rate 9.5% 11.2% 3.8%

[0067] Solution 3 of the present invention: The complete solution described in the present invention is adopted.

[0068] All schemes were run in parallel under identical test set samples, with fault diagnosis accuracy, false alarm rate under high-conflict extreme conditions, and false alarm rate as the core evaluation indicators. The test results are shown in Table 1.

[0069] Based on the quantitative test results recorded in Table 1, the technical effectiveness of the transformer fault detection method based on transformer grounding current described in this embodiment of the invention is analyzed as follows: The overall diagnostic accuracy of the proposed solution (Solution 3) reached 93.6%, a significant improvement compared to Comparative Solution 1 (neural network, 85.2%) and Comparative Solution 2 (ordinary DS method, 81.7%). This indicates that the proposed solution, through dynamic benchmark correction of the standard cloud model, successfully eliminates the feature benchmark drift caused by external operating conditions such as load switching and ambient temperature changes in actual operation of power transformers. This allows the extracted multidimensional fault features to accurately match the actual physical state. In contrast, Comparative Solutions 1 and 2, due to their use of static benchmarks, have insufficient generalization ability, resulting in significantly limited accuracy.

[0070] Under conflicting sample conditions containing strong pulse electromagnetic interference, the false alarm rate of the proposed solution was successfully suppressed to 4.2%, far lower than that of Comparative Solution 2 (14.5%). In real-world electrical scenarios, high-frequency pulse noise often leads to severe physical discrepancies in time-domain, spectral, and trend characteristics (i.e., high-conflict evidence). Traditional DS synthesis rules (Comparative Solution 2) often result in "one-vote veto" or non-convergence misjudgments when processing such data. The proposed solution utilizes weighted reconstruction based on the proportion of transient pulse energy and automatically switches to a weighted average fusion rule during strong conflicts, effectively smoothing out extreme evidence interference caused by noise and demonstrating the superior anti-interference robustness of the proposed solution.

[0071] For latent faults such as slight insulation degradation and minor gradual overheating of the core, the false alarm rate of the present invention is only 3.8%. This quantitatively demonstrates the superiority of the dual-threshold judgment mechanism described in this invention. When the characteristic index is between the first and second judgment thresholds, the system does not blindly give an absolute conclusion of normal or fault, but triggers an "early warning state," adaptively increasing the data acquisition frequency and initiating continuous monitoring and trend tracking, thereby cutting off the false alarm chain in the early stages of fault evolution.

[0072] Figure 2 The visualization results are shown in Table 1. To further analyze the feasibility of this solution, Figure 3 This paper presents a heatmap of the full-class confusion matrix of the proposed solution under a set of 3000 standard electrical samples. Figure 3The confusion matrix analysis shows that the main diagonal region (true examples TP) of the present invention exhibits an extremely high confidence distribution. The correct detection rates for normal state, multi-point grounding of the iron core, partial discharge, and overheating faults reach 95.2%, 93.1%, 93.4%, and 94.0%, respectively, which directly verifies the accurate identification capability of the present invention for typical physical anomalies. The horizontal row analysis shows that the number of false negatives (FN) that actually occurred but were missed as normal states is extremely low (18, 14, and 10 groups, respectively), confirming the efficient detection control of latent faults by the adaptive weighting of dispersion. The vertical column analysis shows that the cross-misjudgment between various faults is effectively suppressed. For example, the cumulative number of false positives (FP) for partial discharge, which is susceptible to noise interference, is only 49 groups, which fully proves that the present invention successfully filters out the abrupt interference of non-target fault dimensions through dual-track conflict fusion switching and achieves clear and decoupled boundary division between various categories.

[0073] Finally, it should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A transformer fault detection method based on transformer grounding current, characterized in that, include: The core grounding current signal of the transformer under test is acquired in real time, and the current signal is preprocessed and multi-dimensional fault features are extracted. The multidimensional fault characteristics include at least trend characteristics, spectral characteristics, and time-domain characteristics; Real-time operating parameters are collected, and the standard cloud model corresponding to each preset fault state is dynamically corrected to eliminate the interference of operating condition fluctuations on the core current. Using the modified standard cloud model, the membership certainty of the multidimensional fault features relative to different fault types is calculated respectively; Combining the distribution discreteness of the multidimensional fault characteristics in the modified standard cloud model, the membership certainty is transformed into basic probability allocation values ​​of independent evidence sources through weighted and normalized mapping of distribution discreteness; the independent evidence sources include trends, spectrum, and time domain. The conflict handling mechanism based on the characteristics of sudden transient and gradual steady state is used to modify the DS synthesis rule, fuse the basic probability allocation values, and calculate the comprehensive trust evaluation index for each fault type. The transformer fault result is determined based on the aforementioned comprehensive trust evaluation index; This includes performing baseline dynamic corrections on the standard cloud models corresponding to each preset fault state, including: Acquire at least one real-time operating condition parameter, including transformer load, bus voltage, and operating temperature; Based on the preset correlation mapping relationship between the real-time operating condition parameters and the standard cloud model parameters, the corresponding expected correction amount and entropy correction amount are calculated in real time. Using the expected correction amount and the entropy correction amount, the initial parameters of the standard cloud model are adaptively compensated to generate the corrected standard cloud model; Calculating the membership certainty of the multidimensional fault features relative to different fault types includes: The multidimensional fault characteristics are input as input cloud droplets to the corresponding positive cloud generator; Based on the expected, entropy, and hyperentropy parameters of the modified standard cloud model, the random Gaussian membership values ​​of the input cloud droplets under the membership degree distribution of each fault type are calculated. The random Gaussian membership value is used as the membership certainty of the multidimensional fault feature relative to different fault types; The method for obtaining the distribution discreteness includes: Calculate the absolute distance between the measured multidimensional fault characteristics and the expected value of the corrected standard cloud model; The absolute distance is scaled using the hyperentropy parameter of the modified standard cloud model to obtain a dispersion index characterizing the dispersion of the multidimensional fault features in the corresponding cloud model. The membership certainty is converted into basic probability assignment values ​​for independent evidence sources, including: The credibility weight of each feature dimension is adaptively calculated based on the dispersion index, wherein the larger the dispersion index, the smaller the corresponding credibility weight. The confidence weights are used to multiply and weight the corresponding membership certainty, and then normalized and mapped within a preset fault identification space to generate the basic probability allocation value with trend, spectrum, and time domain as independent evidence sources.

2. The transformer fault detection method based on transformer grounding current according to claim 1, characterized in that, The current signal is preprocessed and multidimensional fault features are extracted, including: The acquired core grounding current signal is denoised and DC bias interference is eliminated to obtain a purified current signal. The steady-state increment and rate of change are calculated based on the time series of the purification current signal to construct the trend characteristics. The purified current signal is subjected to frequency domain transformation to calculate the harmonic distortion rate and the energy proportion of a specific frequency band, thereby constructing the spectral characteristics; Extract the abrupt peak attributes from the time-domain waveform of the purification current signal, calculate the peak factor and pulse factor, and construct the time-domain features.

3. The transformer fault detection method based on transformer grounding current according to claim 1, characterized in that, Calculate the comprehensive trust evaluation index for each fault type, including: Extract the percentage of transient pulse energy from the core grounding current signal; When the proportion of transient pulse energy is higher than a preset first threshold, it is determined to be a sudden transient-dominated scenario, the weight of the trend feature evidence source is reduced, and the weight of the spectral feature and time domain feature evidence sources is increased. When the proportion of transient pulse energy is lower than a preset second threshold, it is determined to be a gradually steady-state dominant scenario, and the weights of each independent evidence source remain unchanged.

4. The transformer fault detection method based on transformer grounding current according to claim 3, characterized in that, Calculate the comprehensive trust evaluation index for each fault type, including: Calculate the conflict coefficient after combining the evidence from each independent source; When the conflict coefficient is greater than the preset conflict threshold, it is determined that there is a serious conflict between the evidence sources. At this time, the weighted average fusion rule is used to replace the standard DS synthesis rule, and the arithmetic mean of the basic probability allocation values ​​of each evidence source is used as the fusion result. When the conflict coefficient is less than or equal to the conflict threshold, the standard DS synthesis rule is used for fusion. The comprehensive trust evaluation index for each fault type is calculated.

5. The transformer fault detection method based on transformer grounding current according to claim 1, characterized in that, The determination of transformer fault results based on the comprehensive trust evaluation index includes: When the comprehensive trust evaluation index of a certain fault type is greater than or equal to the first judgment threshold, the fault type is determined to be valid. When the comprehensive trust evaluation index of all fault types is less than the second judgment threshold, the transformer is judged to be in normal condition. When the comprehensive trust evaluation index is between the first and second judgment thresholds, it is judged as an early warning state, and continuous monitoring and trend tracking are triggered.

Images