A method and system for fault diagnosis of secondary circuit of current transformer

Abstract

This application relates to the field of equipment condition monitoring technology, and discloses a method and system for fault diagnosis of the secondary circuit of a current transformer. The method includes: connecting a detection device in series to the secondary circuit to establish a physical circuit; injecting a step-scanning high-frequency excitation signal; synchronously sampling the input power signal and output voltage signal to obtain time-series discrete data; performing a fast Fourier transform on the data to extract the frequency domain complex components; calculating the amplitude ratio and phase difference at different scanning frequencies to generate a joint amplitude-frequency-phase-frequency distribution curve; extracting fusion features from the curve; dividing the dataset after normalization; constructing a basic classifier based on the LightGBM algorithm; introducing the TPE algorithm for hyperparameter global optimization; training a state identification model and outputting fault diagnosis results; and using the SHAP interpretation model to calculate the contribution value of the fusion features and outputting feature distribution information. This invention improves the accuracy and engineering interpretability of secondary circuit fault classification.

Description

Technical Field

[0001] This invention relates to the field of equipment condition monitoring technology, specifically to a method and system for diagnosing faults in the secondary circuit of a current transformer. Background Technology

[0002] Current transformers are devices used for energy metering and relay protection in power systems. The operating status of their secondary circuits directly affects the correct operation of protection devices. During long-term operation, the terminal block nodes and wiring parts of the secondary circuits are prone to faults such as poor contact, open circuits, or short circuits due to mechanical vibration, insulation aging, or human error. When such faults occur in the secondary circuits, the current signals received by the measurement, control, and protection devices will be distorted, which can easily lead to maloperation or failure to operate of the relay protection devices, thereby affecting the operational safety of the power grid.

[0003] Currently, fault diagnosis of the secondary circuit of current transformers often employs passive monitoring of power frequency signals or offline conventional testing methods. Passive monitoring primarily relies on the steady-state power frequency current in the circuit, exhibiting low sensitivity to early, minute impedance changes and failing to effectively identify early terminal loosening or minor contact issues. Offline testing requires power outages, impacting grid continuity and failing to reflect the dynamic impedance characteristics of the equipment under actual operating conditions. To address the limitations of single monitoring methods, some diagnostic techniques have begun to incorporate high-frequency signal injection combined with machine learning algorithms for condition assessment. However, these data-driven diagnostic models typically operate as black boxes, directly mapping collected time-domain or frequency-domain features to fault categories. While this approach improves the automation of fault identification, the output diagnostic results lack support at the physical mechanism level. Because the data features extracted by the algorithm are not effectively combined and aligned with the impedance evolution patterns of the circuit's underlying resistance, inductance, and capacitance, field maintenance personnel struggle to trace the physical reasons behind the algorithm's specific diagnostic conclusions, resulting in low reliability of the diagnostic model in practical engineering applications. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides a method and system for diagnosing faults in the secondary circuit of a current transformer. This solves the problem that existing methods for diagnosing faults in the secondary circuit of current transformers are not sensitive to early-stage hidden dangers, and that existing data-driven diagnostic models lack underlying physical mechanism support, resulting in low reliability in engineering applications.

[0005] To achieve the above objectives, the present invention provides the following technical solution: The first aspect of the present invention provides a method for diagnosing faults in the secondary circuit of a current transformer, comprising: Connect the detection device in series to the secondary circuit of the current transformer to establish a high-frequency signal physical circuit with a signal receiving end. A step-scanning high-frequency excitation signal is injected into the secondary circuit of the current transformer as the input power signal, and the input power signal and the output voltage signal of the signal receiver are sampled simultaneously to obtain discrete time data. Perform a Fast Fourier Transform on the time-series discrete data to extract the complex components of the input power signal and the complex components of the output voltage signal at the frequency position with the maximum amplitude in the frequency domain; Calculate the amplitude ratio and phase difference between the complex components of the output voltage signal and the complex components of the input power supply signal at different scanning frequencies, and generate a joint amplitude-frequency and phase-frequency distribution curve; The fusion features are extracted from the amplitude-frequency and phase-frequency joint distribution curves, and the fusion features are normalized and then divided into training and testing sets. A basic classifier based on the LightGBM algorithm is constructed, and the TPE algorithm is introduced to globally optimize the hyperparameters of the basic classifier. A base classifier is trained using the training set to obtain a state identification model, and the state identification model is used to process the test set to output fault diagnosis results. Based on the fault diagnosis results and the fusion features contained in the test set, the contribution value of the fusion features is calculated using the SHAP interpretation model, and the feature distribution information is output.

[0006] In the aforementioned fault diagnosis method, the high-frequency signal physical circuit is established by connecting the detection device in series at the terminal block node of the current transformer's secondary circuit, thereby constructing a high-frequency broadband equivalent circuit model. At the physical level, the equivalent circuit model is transformed into a series topology of resistors, inductors, and capacitors, where the matching capacitor is a pre-set fixed matching capacitor, and the sampling resistor at the signal receiving end is equivalent to a constant equivalent inductance and a continuously adjustable resistor. The system traverses three typical operating conditions—normal operation, secondary open circuit, and terminal short circuit—by switching the physical connection state.

[0007] During the data acquisition phase, the system employs a fixed-frequency step value for discrete step scanning to generate discrete test frequency points. For each discrete step frequency point, a uniform signal injection duration is set, removing transitional data from the initial stage and extracting data within the steady-state time window as valid acquisition data for state identification. The analog-to-digital converter performs synchronous discretization sampling on the input power signal and output voltage signal contained in the valid acquisition data for state identification through a multi-channel synchronous hardware triggering mechanism. After acquiring the time-series discrete data, the system applies a window function for smooth truncation and performs a Fast Fourier Transform to map it into a frequency domain complex sequence. The system presets a frequency tolerance interval around the center of the scanning frequency point, iterates through and calculates the amplitude of the frequency domain complex sequence, locks the frequency coordinate position corresponding to the maximum amplitude in the frequency domain distribution of the input power signal and the frequency domain distribution of the output voltage signal, and extracts the corresponding complex components.

[0008] The generation of the joint distribution curve is based on the calculation of the complex components in the frequency domain. The system extracts the magnitude of the complex components to calculate the amplitude ratio, and introduces a minimal positive constant in the denominator as a bias protection to avoid division by zero overflow during the calculation; it also extracts the argument of the complex components to calculate the phase difference. The system aggregates the amplitude ratio and phase difference at each frequency point in sequence, and generates a joint amplitude-frequency-phase-frequency distribution curve by combining the underlying hardware parameters and operating conditions. Subsequently, the system extracts multi-dimensional fusion features covering statistical characteristics, physical mechanism characteristics, and operating condition parameter characteristics. A standardization strategy is applied to the fusion features, mapping the data to a standard distribution interval with a mean of zero and a variance of one. A minimal positive constant is forcibly introduced as a bias protection when performing division operations to calculate the standard deviation. The sample distribution ratio of each category state in the constrained subset is consistent with the original global sample library. The overall fusion feature set is divided into training and test sets through a hierarchical cross-validation strategy.

[0009] In the model building and state identification stages, the LightGBM base classifier approximates the true value by integrating multiple weak evaluators, and the objective function includes a regularization term. The TPE algorithm is used for hyperparameter optimization. The algorithm uses a fixed percentage quantile of the model error in historical iteration records to set an error threshold, divides the parameter set into a dominant group and a subordinate group, constructs a nonparametric probability density estimate, and uses an expectation boosting function to find the parameter sampling point with the maximum expected gain. After the state identification model outputs diagnostic results, the SHAP additive interpretation model deconstructs the predicted value into a linear function of binary variables. Based on the marginal contribution principle, the system calculates the expectation across all feature permutations and combinations, allocates the deviation of the global prediction expectation to each input feature, and calculates the contribution value of each fused feature. The system takes the absolute value of the contribution of each feature dimension and calculates the global average, generating a local feature contribution analysis map. For typical operating conditions, it outputs corresponding local feature attribution and mechanism mapping maps.

[0010] After the model interpretation output, the system establishes a physical simulation model of resistance, inductance, and capacitance, and extracts the amplitude-frequency and phase-frequency attenuation curves of the theoretical state under the same operating conditions. The pure mechanism features and local feature attributions obtained from the physical simulation model are compared with the high-contribution feature intervals shown in the mechanism mapping spectrum through dual alignment between time and operating conditions. Based on the overlap between the simulation theoretical extreme values ​​and the high-contribution intervals of the model, the consistency of the positive and negative push-pull logic of features, and the evolution law of circuit impedance, a multi-dimensional weighted score is completed, realizing a deep alignment verification between data-driven logic and actual circuit characteristics.

[0011] A second aspect of the present invention provides a fault diagnosis system for the secondary circuit of a current transformer, comprising: The high-frequency signal injection module is connected in series to the secondary circuit of the current transformer and is used to apply a step-scanning high-frequency excitation signal within a preset frequency range as an input power signal to the secondary circuit of the current transformer. The signal sampling module is connected to the high-frequency signal injection module and the secondary circuit of the current transformer. It is used to synchronously sample the input power signal and the output voltage signal of the signal receiver to obtain time-series discrete data. The frequency domain feature construction module, connected to the signal sampling module, is used to receive the acquired time-series discrete data, convert the time-domain sampled data to the frequency domain using fast Fourier transform, extract the complex components of the input power signal and the output voltage signal at the frequency position with the maximum amplitude in the frequency domain, calculate the amplitude ratio and phase difference of the complex components of the output voltage signal and the complex components of the input power signal at different scanning frequencies, and generate the amplitude-frequency-phase-frequency joint distribution curve. The feature processing module, connected to the frequency domain feature construction module, is used to extract fusion features from the amplitude-frequency and phase-frequency joint distribution curves. After normalization, the fusion features are divided into training and test sets. The model identification module, connected to the feature processing module, is used to build a basic classifier based on the LightGBM algorithm. The TPE algorithm is introduced to globally optimize the hyperparameters of the basic classifier. The basic classifier is trained using the training set to obtain the state identification model. The state identification model is used to process the test set and output the fault diagnosis results. The decision analysis module, connected to the model identification module, is used to calculate the contribution value of the fusion features based on the fault diagnosis results and the fusion features contained in the test set, and output the feature distribution information using the SHAP interpretation model.

[0012] This invention provides a method and system for diagnosing faults in the secondary circuit of a current transformer. It has the following beneficial effects: 1. This invention injects a step-scanning high-frequency excitation signal into the secondary circuit of a current transformer, and combines synchronous sampling and fast Fourier transform to extract the complex components at the position of the maximum amplitude, generating a joint amplitude-frequency and phase-frequency distribution curve. This technical feature integrates the amplitude ratio and phase difference of the output and input signals, and can objectively reflect the broadband impedance evolution law of the secondary circuit under different operating conditions such as normal, open circuit and short circuit. Compared with the single-dimensional time domain characteristics, it provides a more complete data foundation for fault state identification.

[0013] 2. This invention uses the LightGBM algorithm to construct the basic classifier and introduces the TPE algorithm to perform global optimization of hyperparameters. By extracting fusion features containing statistical, mechanistic, and operating condition parameters, and introducing bias protection during normalization, the scale consistency of the underlying input data is ensured. At the same time, the TPE algorithm divides the parameter set based on the model error in the historical iteration record and uses the expectation boosting function to guide the convergence of the search space, which directly improves the accuracy and computational efficiency of the state identification model for fault classification under different load conditions.

[0014] 3. This invention introduces the SHAP interpretation model to calculate the feature contribution output distribution information and simultaneously establishes a corresponding physical simulation model to extract theoretical extreme values. This mechanism performs a dual alignment comparison between the local feature attribution map output by the SHAP model and the impedance change logic of the actual physical circuit in terms of time and operating conditions, completing the mutual verification between the data-driven algorithm and the underlying physical characteristics. This makes up for the lack of mechanistic support in pure black-box machine learning methods and improves the engineering interpretability of diagnostic results. Attached Figure Description

[0015] Figure 1 This is a flowchart of the method of the present invention; Figure 2 This is a schematic diagram of the system architecture of the present invention; Figure 3 This is a graph showing the amplitude-frequency response of the equivalent circuit of the present invention. Figure 4 This is a comparison chart of the classification accuracy of the present invention; Figure 5 This is a block diagram of the current transformer performance verification system of the present invention; Figure 6 This is an equivalent circuit diagram of the detection device of the present invention; Figure 7 This is a Simulink simulation diagram of the present invention; Figure 8 This is a comparison diagram of the input and output voltage waveforms of the present invention; Figure 9 This is a graph showing the combined amplitude-phase frequency distribution of the secondary circuit of the current transformer of the present invention. Figure 10 This is a training set diagram of the simulation results of the present invention; Figure 11 This is a test set diagram of the simulation results of the present invention; Figure 12 This is a global contribution diagram of the present invention; Figure 13 This is a schematic diagram of the normal operating state of the present invention; Figure 14 This is a schematic diagram of the short-circuit operation state of the present invention. Detailed Implementation

[0016] 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.

[0017] Please see the appendix Figure 1This invention provides a method for diagnosing faults in the secondary circuit of a current transformer, comprising the following steps: S10, connect the detection device in series to the secondary circuit of the current transformer to establish a high-frequency signal physical circuit; S20, inject a step-scanning high-frequency excitation signal into the secondary circuit of the current transformer, and synchronously sample the injected power signal and the output voltage signal of the signal receiving end. S30 performs a fast Fourier transform on the sampled power supply signal and output voltage signal to extract the complex components of the input signal and the complex components of the output signal at the frequency position with the maximum amplitude in the frequency domain. S40, calculate the amplitude ratio and phase difference between the output signal and the input signal at different scanning frequencies, and generate the amplitude-frequency and phase-frequency joint distribution curve; S50 extracts fusion features from the amplitude-frequency and phase-frequency joint distribution curves, and after normalizing the fusion features, divides them into training and test sets. S60, construct a basic classifier based on the LightGBM algorithm, and introduce the TPE algorithm to globally optimize the hyperparameters of the basic classifier; S70 uses the training set to train the basic classifier to obtain the state identification model, uses the state identification model to process the test set, and outputs the fault diagnosis results of the secondary circuit of the current transformer. S80 uses the SHAP interpretation model to calculate the contribution value of the fused features and outputs the feature contribution distribution information.

[0018] Please see the appendix Figure 2 The present invention provides a fault diagnosis system for the secondary circuit of a current transformer, comprising: a high-frequency signal injection module, a signal sampling module, a frequency domain feature construction module, a feature processing module, a model identification module, and a decision analysis module.

[0019] A high-frequency signal injection module is connected in series to the secondary circuit of the current transformer. This module applies a stepped-scanning high-frequency excitation signal within a preset frequency range to the secondary circuit of the current transformer. The secondary circuit of the current transformer exhibits different equivalent impedance characteristics under different operating conditions, and the high-frequency excitation signal generates a corresponding voltage response in the equivalent circuit.

[0020] The signal sampling module is connected to the high-frequency signal injection module and the secondary circuit of the current transformer. The signal sampling module is equipped with an analog-to-digital converter. The signal sampling module is used to synchronously sample the injected power signal and the output voltage signal from the receiving end to obtain discrete timing data.

[0021] The frequency domain feature construction module is connected to the signal sampling module. It receives the sampled data output from the signal sampling module. The frequency domain feature construction module uses a Fast Fourier Transform (FFT) to convert the time-domain sampled data to the frequency domain, extracting the complex components of the input and output signals at specific frequency points. By calculating the amplitude ratio and phase difference between the output and input signals at each scanning frequency, the frequency domain feature construction module generates amplitude and phase feature vectors.

[0022] The feature processing module is connected to the frequency domain feature construction module. The feature processing module extracts fused features from the amplitude and phase feature vectors. It then normalizes the extracted fused features to eliminate dimensional differences and divides the processed feature set into training and testing sets.

[0023] The model identification module is connected to the feature processing module. The model identification module constructs a state classification model based on the LightGBM algorithm. It also introduces the TPE algorithm to iteratively optimize the hyperparameters of the state classification model in the global space. The model identification module uses the training set output by the feature processing module to train the model and outputs classification results for the normal state, open-circuit state, and terminal short-circuit state of the current transformer secondary circuit.

[0024] The decision analysis module is connected to the model identification module. The decision analysis module receives the classification results and input features from the model identification module. Based on the SHAP model, the decision analysis module calculates the contribution value of each feature to the classification result. The decision analysis module outputs the feature contribution distribution information to verify the consistency between the algorithm output and the evolution of the underlying circuit characteristics.

[0025] The specific implementation process of connecting the detection device in series with the secondary circuit of the current transformer to establish a high-frequency signal physical circuit includes the following steps: S101, In this embodiment, the detection device for state detection is connected in series to the secondary circuit of the current transformer under test; please refer to the appendix. Figure 5 , Figure 5 The figure shows a block diagram of the current transformer performance verification system of the present invention. CT (Current Transformer) represents the current transformer. This detection device integrates a high-frequency test signal source (labeled AC in the figure) and a sampling measurement circuit. It is connected to the terminal block node of the secondary circuit via physical cables, forming a high-frequency observation path with the existing primary high-voltage side of the power system and the secondary metering and protection devices. To avoid interference from the high-frequency detection signal to the power frequency operation of the existing secondary protection and metering equipment, this observation path is established based on the high impedance isolation characteristics of the high-frequency signal in the power frequency band, thereby ensuring that the primary load current of the distribution network is not affected.

[0026] S102, Based on the high-frequency observation path established above, a high-frequency broadband equivalent circuit model of the current transformer considering distributed parameters is established; please refer to the appendix. Figure 6 , Figure 6 This is an equivalent circuit diagram of the detection device of the present invention. (In conjunction with...) Figure 6 As can be seen, the input impedance (labeled Zin in the figure) exhibits complex high-frequency characteristics. The system can be equivalent to a resistance by treating the entire physical circuit, including the secondary circuit of the current transformer and the detection device, as a resistor. ,inductance and capacitor A series topology is formed. Among them, the capacitors... For the fixed matching capacitor pre-configured inside the detection device, this capacitor It is used to construct a resonant circuit for excitation of state characteristics with the inherent inductive element of the circuit under test. In this equivalent circuit model, the high-frequency detection signal voltage at the injection terminal is defined as... The output voltage signal acquired by the receiving end is defined as follows: The signal receiver is equipped with a sampling resistor for extracting voltage characteristics. To accurately reflect the actual physical characteristics at high frequencies, the input impedance at the signal injection point is ignored, and the sampling resistor is equivalent to a series resistor and inductor structure. This series structure consists of an equivalent inductor. and adjustable resistor Composition. As a specific lower-level feature implementation of this functional generalization, considering the typical physical magnitudes of the leakage inductance on the secondary side of a real current transformer and the distributed inductance of the leads, an equivalent current is set. The value is kept constant at 0.5mH. Simultaneously, the adjustable resistor... The resistance value is adjustable within a range of 0.5Ω to 100Ω. This setting method can simulate the impact of different load conditions, from extremely light to heavy loads, on the measurement terminal in actual transformer substations, thereby providing underlying hardware support for building a feature sample library covering a variety of typical operating conditions.

[0027] S103. After completing the equivalent circuit model construction, the physical impedance response characteristics of the current transformer secondary circuit under different operating states are further analyzed. In actual power distribution network operation, the current transformer secondary circuit exhibits three typical operating conditions: normal operation, secondary open circuit, and terminal short-circuit. Based on the principle of electromagnetic induction, when the circuit is in the secondary open circuit state, it induces core flux saturation, leading to a sharp decrease in the secondary excitation inductance. When the circuit is in the terminal short-circuit state, it alters the original current shunt path, effectively short-circuiting part of the load impedance and secondary lead inductance in the electrical circuit. These three changes directly affect the equivalent inductance in the equivalent circuit model. A significant numerical difference is observed. Equivalent inductance The differential changes will disrupt the original impedance matching conditions of the series circuit, directly causing the resonant frequency of the established resonant circuit to shift, accompanied by distortion of the amplitude and phase response characteristics throughout the entire scanning frequency band. This correspondence between the underlying physical electrical connection state and the frequency domain response characteristics provides an objective physical basis for subsequent high-frequency active flaw detection and algorithm classification and identification. For the steady-state voltage solution and basic impedance calculation process of the series resonant circuit under a given high-frequency excitation, those skilled in the art can derive it based on Kirchhoff's laws and the complex phasor method. Its basic circuit response analysis is a well-known technique in this field and will not be elaborated upon here.

[0028] The specific implementation process of injecting a step-scanning high-frequency excitation signal into the secondary circuit of a current transformer and performing synchronous sampling includes the following steps: S201, combining the established high-frequency observation path, an active detection excitation is applied to the secondary circuit of the current transformer under test. In this embodiment, the starting frequency of the injected signal is selected as 10kHz, and a discrete step scan is performed with a fixed frequency step value of 500Hz, defining the full-band scan coverage range as 10kHz to 40kHz. The physical basis for selecting this specific scanning frequency band is that low-frequency signals are easily submerged by the inherent power frequency and its low-order harmonic background noise of the distribution network system, while high-frequency bands exceeding 40kHz are prone to causing severe skin effect and nonlinear aliasing distortion in the transformer core. Using the above frequency band and step strategy, the extracted high-frequency broadband impedance response characteristics can be ensured to have high linearity and anti-interference capability while ensuring signal penetration. Based on this sweep frequency range and frequency step value setting, a total of 61 discrete test frequency points are generated during the full-band scan process.

[0029] S202, In executing the aforementioned step-sweep strategy, strict data capture time window control is required to obtain steady-state physical characteristics that accurately reflect the loop impedance. For each set discrete step frequency point, the duration of the high-frequency detection signal voltage injection is uniformly set to 1 second. When the high-frequency signal is first connected to a physical circuit containing energy storage components such as inductors and capacitors, a transient transition process is inevitably triggered. To eliminate the interference of this transient attenuation component on the accuracy of subsequent frequency domain complex number operations, the system is configured to discard the initial transient data and only capture the data from 0.5 seconds to 1 second after signal injection as valid acquisition data for state identification. This steady-state time window, set based on the physical transient attenuation law, effectively avoids subsequent feature calculation deviations caused by the inclusion of transient distorted data.

[0030] S203 further performs synchronous discretization sampling processing on multi-source signals within the intercepted time window. The sampling frequency of the analog-to-digital converter inside the detection device is set to 100kHz, and the injected high-frequency detection signal voltage is sampled separately. and the output voltage signal acquired by the receiving end Sampling is performed. The 100kHz sampling rate is determined based on the Nyquist sampling theorem, and its value is greater than twice the highest scan frequency of 40kHz, thus theoretically ensuring that frequency domain aliasing does not occur during the high-frequency signal discretization process. To ensure the rigor of the subsequent impedance phase calculation logic, the analog-to-digital converter adopts a hardware-level multi-channel synchronous triggering mechanism to ensure strict alignment of the input and output signals in the operating conditions and time dimensions. Based on the aforementioned 100kHz sampling frequency and a 0.5s effective steady-state data truncation time window, the length of the time-series discrete data acquired at a single scan frequency point is precisely controlled to 50,000 points. This data length provides sufficient data redundancy to support high-resolution frequency domain transformation operations while taking into account the computational load of the underlying control chip. The multi-channel synchronous hardware clock allocation and underlying triggering implementation method of the analog-to-digital converter are described below.

[0031] The specific implementation process of frequency domain transformation and complex component extraction for the synchronous sampling data obtained at each step scanning frequency includes the following steps: S301, perform frequency domain mapping transformation on the time-series discrete data acquired within the truncation time window. In this embodiment, considering that directly truncating the time-domain signal would cause spectral leakage, a window function is applied to the synchronously sampled input power signal and output voltage signal before performing the frequency domain transformation. As a specific implementation of this functional operation, a Hanning window or a Hamming window can be used to weight the original sequence to effectively suppress spectral sidelobe interference. After smoothing and truncation, the one-dimensional time-domain discrete data sequence is mapped to the frequency dimension using the Fast Fourier Transform algorithm to generate the corresponding frequency domain complex sequence.

[0032] S302, after acquiring the frequency domain complex sequence, the specific characteristic frequency position is locked based on the spectral information. Based on the aforementioned active frequency injection mechanism, the actual excitation energy is concentrated at the currently applied specific scanning frequency point. To avoid the failure of the simple maximum amplitude determination logic due to sudden high-frequency pulse noise in the industrial field, the peak search is not blindly traversed across the entire Nyquist frequency range. The currently set active injection scanning frequency value is used as the center point, and a preset frequency tolerance range is set. Within this limited tolerance range, the amplitude of the frequency domain response sequence is traversed and calculated, and the specific frequency coordinates where the amplitude reaches its maximum value in the frequency domain distribution of the input power signal and the frequency domain distribution of the output voltage signal are locked respectively. The technical purpose of implementing the above-mentioned peak search within the limited range is to automatically filter the power frequency background noise of the distribution network and harmonic interference in unrelated frequency bands, ensuring that the subsequently extracted frequency domain response parameters strictly correspond to the current actual scanning frequency.

[0033] S303, based on the locked maximum amplitude frequency position, extract the corresponding frequency domain complex components. At the locked frequency coordinate point, extract the frequency domain complex components of the input signal and the output signal respectively. The extracted input signal complex components are denoted as... The extracted complex components of the output signal are denoted as The complex components at these two specific frequency points characterize the steady-state response properties of the physical circuit under high-frequency excitation. The complex data not only reflects the absolute amplitude of the voltage signal through its magnitude, but the algebraic relationship between its real and imaginary parts also contains information about the voltage and current phase delays caused by the equivalent impedance of the current transformer's secondary circuit. Using these extracted frequency-domain complex components, the physical changes in the underlying electrical connection state can be mapped to the data layer, thus providing a fundamental calculation basis for subsequently constructing a broadband amplitude-phase characteristic fingerprint.

[0034] The specific implementation process of generating broadband amplitude-frequency and phase-frequency feature fingerprints by combining the extracted frequency domain complex components includes the following steps: S401, based on the extracted complex components of the input and output signals, calculate the impedance response characteristics of the physical circuit at a specific scanning frequency. Before the calculation step, according to the general principle of steady-state response of linear circuits, the transfer characteristics of the secondary circuit of the current transformer under high-frequency excitation can be directly quantified by the ratio of the input and output complex phasors. In this embodiment, the amplitude ratio of the output signal to the input signal is calculated using the magnitude of the complex components, and the phase difference between the two is calculated using the phase angle of the complex components. To avoid system overflow caused by the denominator of the division operation approaching zero due to the extremely small input signal under extreme impedance attenuation conditions, a preset minimum positive constant is explicitly introduced into the denominator as bias protection when calculating the amplitude ratio in the underlying code. The expressions for calculating the amplitude ratio and phase difference at different scanning frequencies are as follows: ; ; In the formula, This represents the amplitude ratio at the current scanning frequency. The magnitude of the complex components in the frequency domain of the output signal; The magnitude of the complex component of the input signal in the frequency domain; This represents the phase difference at the current scanning frequency. The angle of the complex component in the frequency domain of the output signal; The amplitude ratio represents the angle of the complex component of the input signal in the frequency domain. This amplitude ratio physically reflects the degree of attenuation of the injected high-frequency signal by the equivalent impedance of the secondary circuit at the current frequency, while the phase difference characterizes the inductive delay or capacitive lead characteristics of the circuit at this frequency.

[0035] S402, after completing the response calculation at a single frequency point, the feature data at each discrete step frequency point need to be aggregated sequentially to construct a wideband feature vector that evolves with frequency. Based on the above-described active frequency sweep strategy, which includes multiple step frequency points, the amplitude and phase features ultimately generated by the system are represented as a feature vector that evolves with frequency. This wideband vectorized representation can completely characterize the response drift trajectory of the equivalent impedance throughout the entire high-frequency band. The expressions for the amplitude and phase feature vectors are as follows: ; ; In the formula, It is the amplitude eigenvector; It is the phase eigenvector; For the first Amplitude ratio at a specific scanning frequency; For the first Phase difference at a specific scanning frequency; This represents the total number of scan frequency points. The upper and lower limits of the scan interval and the step frequency are jointly determined, and its expression is as follows: ; In the formula, This is the maximum scan frequency value set. The set starting scan frequency value; The set frequency step value. In the constructed feature vector, each element sequence strictly corresponds to the amplitude response and phase response at a specific frequency. These two vector sequences together constitute a standardized amplitude and phase characteristic fingerprint for fault diagnosis.

[0036] S403, relying solely on characteristic fingerprints under a single static operating condition is insufficient to support the generalization training of subsequent data-driven models. Therefore, it is necessary to establish a library of amplitude-frequency and phase-frequency joint distribution curves covering multiple typical states by changing the underlying hardware configuration parameters and physical operating conditions. In this embodiment, the equivalent inductance is maintained. The sampling resistor is kept constant at 0.5mH and continuously adjusted within the range of 0.5Ω to 100Ω. The resistance value was determined. For each resistance value configuration, the physical connection state of the current transformer's secondary circuit was switched to traverse typical operating conditions including normal operation, secondary open circuit, and terminal short circuit. Using the multi-source sequences collected under these differentiated operating conditions, a combined amplitude-frequency and phase-frequency distribution curve encompassing all circuit states and load conditions was generated, providing complete raw data support for subsequent multi-dimensional feature engineering and model construction.

[0037] Please see Figure 9 , Figure 9The evolution trajectory of amplitude-frequency and phase difference-frequency response characteristics under specific operating conditions is shown.

[0038] Based on the constructed amplitude-frequency and phase-frequency joint distribution curves, the specific implementation process of multi-angle feature engineering and data preprocessing includes the following steps: S501, based on the aforementioned amplitude-frequency and phase-frequency joint distribution curve, multi-dimensional fusion features are extracted from the frequency response sample library. When processing raw high-frequency response data, directly inputting high-dimensional wideband sequences into the classifier not only causes extreme redundancy in computational dimensions, but also makes it difficult for simple discrete sequence points to highlight the core physical differences characterizing fault states. Based on these considerations, in this embodiment, the system performs in-depth information mining on the raw frequency response curve from multiple perspectives, including statistical distribution, physical mechanisms, and operating conditions. The extracted fusion features cover statistical features, physical mechanism features, and operating condition parameter features. Statistical features cover the mean, skewness, and kurtosis of the amplitude and phase sequences. The mean characterizes the overall average response benchmark of the equivalent impedance of the loop across the entire frequency band, while skewness and kurtosis are used to quantify the degree of asymmetry in the overall distribution of the amplitude-frequency curve and the steepness of the resonant peaks. Physical mechanism features cover... The value, resonant frequency, and phase oscillation intensity. The resonant frequency directly corresponds to the extreme coordinates of the impedance response curve, and its numerical drift maps the physical change trajectory of the underlying equivalent inductance parameters; The value assesses the energy loss state of the physical system within the resonant frequency band. The operating parameter characteristics incorporate external load parameters, such as the set sampling resistor value, as prior information into the feature space.

[0039] S502, for the extracted multi-dimensional feature data, normalization processing is performed to achieve a unified overall data scale. In the multi-dimensional feature set, the numerical magnitudes of features in different dimensions vary across several orders of magnitude. If this type of multi-source heterogeneous data is not aligned, features with extremely large numerical scales will dominate weight allocation during model optimization. This embodiment uses a Z-score normalization strategy to process all continuous variables, strictly mapping all dimensions of data to a standard distribution interval with a mean of 0 and a variance of 1. To verify and improve the completeness of the algorithm logic, considering that some physical feature columns are constants under certain static conditions, causing the standard deviation to approach zero, the system forces the introduction of a very small positive number as a bias protection in the denominator when performing the division by standard deviation operation, thereby effectively avoiding the risk of division by zero overflow in the underlying code.

[0040] In S503, a constraint-based partitioning mechanism is introduced to segment the processed feature dataset, constructing an independent dataset for model training and generalization capability verification. After unifying the data scale, the system employs a hierarchical cross-validation strategy to segment the entire feature set. The algorithm enforces that the sample distribution ratio of each category state in the partitioned subsets remain strictly consistent with the original global sample library. While ensuring balanced distribution alignment, the overall fused feature set is divided into a model training set and a model test set in a rigorous 8:2 ratio. This sampling logic, which introduces category label distribution constraints, effectively avoids the training imbalance risk that is easily caused by conventional random partitioning, ensuring that the test set can perform fair generalization tests on the typical physical operating states of the current transformer.

[0041] The specific implementation process of constructing the TPE-LightGBM model based on the preprocessed multidimensional feature set and performing global hyperparameter optimization includes the following steps: S601, Determine the basic classification model architecture and construct the iterative objective function. In this embodiment, there is a highly nonlinear mapping relationship between the fault state of the secondary circuit of the current transformer to be diagnosed and the underlying electrical characteristics. To overcome the limitations of traditional linear judgment rules, the LightGBM algorithm is selected as the basic classifier. As a necessary prerequisite for model construction, its input data is strictly defined as the aforementioned normalized preprocessed multidimensional fusion feature vector, and its output directly maps to three specific physical operating states: normal operation, secondary open circuit, and terminal short circuit. Through continuous iterative updates, the system achieves high-precision nonlinear fitting of the fault state of the current transformer. During the model training phase, the algorithm approximates the true value by integrating multiple weak evaluators. The goal of each iteration is to minimize the objective function, which includes the regularization term. ; In the formula, For the first The objective function of the round iteration; The total number of samples participating in the training; The loss function used to measure prediction error is configured here as the multi-class log loss function; For the first The true state label of each sample; For the front Round of 1 Cumulative predicted value for each sample; For the first Predicted values ​​for new tree rotation; This is a regularization term used to prevent overfitting caused by excessively complex tree structures.

[0042] S602, performs Taylor expansion of the loss function and leaf node weight analysis. To achieve fast solution and optimization of the objective function, LightGBM... loss function Perform a second-order Taylor expansion. During the expansion, the corresponding first-order gradient... and second gradient Defined as: ; In the formula, For the first The first gradient corresponding to each sample; For the first The second-order gradients corresponding to each sample. Based on the above gradient metric statistics, the optimal output weights of the leaf nodes are... The calculation formula is: ; In the formula, For the first The optimal output weights for each leaf node; For the first The set of samples contained in each leaf node region; This is the regularization coefficient. In the underlying operational logic of the algorithm, this regularization coefficient not only penalizes excessively large weight values ​​on a macroscopic level, but also acts as a consistently positive bias constant in the denominator.

[0043] S603 introduces a Tree-structured Parsons Estimator (TPE) to implement the probability density partitioning of hyperparameters. LightGBM contains numerous hyperparameters such as learning rate and maximum tree depth. The complex nonlinear relationships between these parameters make traditional manual trial-and-error optimization inefficient. The TPE algorithm is introduced for global optimization. TPE utilizes historical iteration records and, based on a set error threshold... The tested parameter set is divided into an advantageous group and a disadvantageous group, and nonparametric probability density estimates are constructed for each group: ; In the formula, Let be the probability density distribution of the parameters given the model error. The parameter distribution when performance is excellent; represents the parameter distribution when performance is poor; y represents the model error under this hyperparameter combination. The set error threshold is set as the 15th percentile of the model error in the historical iteration record. This statistically significant adaptive threshold effectively avoids the problem of rigidity in the exploration space caused by a single fixed value.

[0044] S604, based on the expectation boosting function to guide the rapid convergence of the hyperparameter search space. TPE uses expectation boosting as the sampling function to find the parameter sampling points that bring the maximum expected gain to the model performance: ; In the formula, For a given combination of parameters The expected improvement is as follows; the integration operation is a continuous integration from negative infinity to a set error threshold; through derivation, maximizing... Equivalent to finding the ratio The parameter point to be maximized. When performing this ratio optimization calculation in the underlying code, the probability distribution of the disadvantaged group in some less popular parameter regions is taken into account. It may be extremely close to zero, so the system forces the addition of a preset minimal positive number as a bias protection when calculating the actual denominator.

[0045] The specific implementation process of performing prediction, identification, and post-attribution mechanisms on the trained state classification model includes the following steps: S701, to obtain the classification results output by the TPE-LightGBM model, please refer to [link / reference]. Figure 10 and Figure 11 , Figure 10 and Figure 11 The confusion matrix prediction distributions of the model on the training and test sets are shown respectively, intuitively reflecting the model's high-precision classification performance for normal, short-circuit, and open-circuit states. A post-attribution mechanism is also introduced. Although the aforementioned classification model has high prediction accuracy, as an ensemble tree model, its internal structure of hundreds of interwoven decision trees is extremely complex. The mapping process from feature data to the final state label lacks intuitive physical interpretability. This embodiment introduces a game theory-based SHAP additive interpretation model. The technical purpose of this interpretation model is to transform the invisible model decision logic into a physical contribution score that human engineers can understand through rigorous mathematical allocation theory, thereby meeting the reliability assessment requirements of fault identification results in industrial settings.

[0046] S702, based on additive attribution theory, deconstructs the model's predicted values ​​into linear functions of binary variables. As a general-purpose technique, SHAP does not alter the internal weights of the original classification model; instead, it distributes the deviation of the global prediction expectation to each input feature by externally observing minute changes in the feature inputs and outputs. Its additive feature attribute formula is as follows: ; In the formula, This represents the model's actual predicted output value under the current specific sample input. The number of input features; The baseline value is the average predicted value of the model for the sample without any feature information. For the first The SHAP value of each feature can be decomposed into the sum of the basic expectation and the independent contribution of each physical feature.

[0047] S703 quantifies the specific impact of a single feature by exhaustively enumerating the boundaries of feature combinations. This is to accurately evaluate specific features. The physical impact needs to be calculated based on the idea of ​​marginal contribution across all possible feature permutations. Specific features SHAP value The calculation formula is: ; In the formula, The set of all features; For features not included A subset of features; For subset The number of features contained within; For the model in subset The predicted output is as follows; To feature Join a subset The predicted output of the post-model. Before performing the above calculations, the system first confirms the number of extracted multidimensional fusion features. It must always be a positive integer greater than or equal to 1. This constraint naturally guarantees the denominator term in the underlying code logic. The value is always greater than zero, completely avoiding the software overflow risk that factorial division operations might cause. In this computational architecture, the difference term within square brackets represents the feature. The marginal decision change brought about by the introduction of [the term], while the fractional term serves as a weighting coefficient, representing a specific subset. The statistical probability of occurrence in all possible combinations.

[0048] In step S704, after mathematically calculating the SHAP values ​​for all samples across all feature dimensions, the system uses this massive local attribution data matrix as input to the physical property alignment analysis module. This step establishes a conversion interface from pure numerical tensors to human-readable graphs, providing a data source for subsequent mechanistic cross-validation.

[0049] The specific implementation process of performing physical interpretability verification and mechanism alignment based on the SHAP mechanism includes the following steps: S801, extract the local feature contribution analysis map to evaluate the global importance of multidimensional features. In purely data-driven algorithms, ensemble models are prone to overfitting, which involves memorizing background noise rather than learning true physical laws. In this embodiment, the system calculates the absolute value of the contribution of each feature dimension based on the SHAP values ​​of all samples obtained above and calculates the global average, thereby generating a local feature contribution analysis map. Please refer to [link to relevant documentation]. Figure 12 , Figure 12 This provides a visualization of the global physical feature contribution analysis results based on the SHAP mechanism. The technical purpose of this analysis step is to quantify the overall contribution weight of frequency response features in each dimension to state identification, and then verify whether the top-ranking key features are reasonably concentrated at the resonant frequency. The multi-dimensional weighted ranking verification can objectively verify from the data level that the classification model has indeed captured and utilized the key physical mechanism parameters such as amplitude and skewness.

[0050] S802 constructs local feature attribution and mechanism mapping maps, implementing deep alignment between data-driven logic and underlying electrical characteristics. Since simple global weight ranking only indicates the relative importance of features but cannot reveal how changes in specific feature values ​​specifically push and pull the model to change the decision direction, the system outputs corresponding local feature attribution and mechanism mapping maps for typical operating conditions such as normal operation, secondary open circuit, and terminal short circuit. Please refer to [link / reference]. Figure 13 and Figure 14 , Figure 13 This is a schematic diagram of the normal operation state of the present invention (i.e., a local feature attribution and mechanism mapping map under normal operation state). Figure 14 This is a schematic diagram of the short-circuit operation state of the present invention (i.e., a local feature attribution and mechanism mapping map under the terminal short-circuit state). In the underlying visual mapping logic of this map, the horizontal axis is clearly defined as the specific value of SHAP contribution, the vertical axis represents different fusion feature dimensions, and the depth of the scatter color strictly maps the level of the original value of the feature itself.

[0051] S803, in conjunction with the physical simulation model, performs multi-dimensional cross-validation of fault characteristics. To fundamentally close the loop and verify the causal relationships output by the aforementioned SHAP plot, the system simultaneously incorporates a physical simulation platform for theoretical cross-checking. Please refer to the appendix. Figure 7 , Figure 7 This is a Simulink simulation diagram of the present invention. In the Simulink simulation platform, the discrete solution step size is set to 1e-06 s (i.e., 1 microsecond), and a rigorous simulation is established based on the known hardware parameters of the actual current transformer secondary circuit. - - Equivalent circuit physical simulation model. As the basis for implementing the underlying parameters of this step, the aforementioned initial hardware parameters, such as the equivalent inductance and distributed capacitance, can be obtained by reading the factory nameplate parameters of the transformer under test. High-precision offline bridge measurement acquisition. After obtaining accurate equivalent physical parameters, the same high-frequency scanning excitation is injected into the physical model under the same operating conditions to extract the amplitude and phase frequency attenuation curves under theoretical conditions.

[0052] After acquiring multi-source comparison data, S804 performs a feature alignment judgment based on weighted logic. The system compares the purely mechanistic features, such as resonance extrema and phase reversal frequency bands obtained from simulation, with the high-contribution feature intervals shown in the aforementioned SHAP beehive diagram, using both time and operating conditions for rigorous alignment. When determining the validity of the final identification results, the system avoids relying solely on a single extremum for judgment. Instead, it performs multi-dimensional weighted scoring based on the overlap between the simulation theoretical extrema and the model's high-contribution intervals, and the consistency between the positive and negative push-pull logic of the features and the evolution law of circuit impedance. After this rigorous verification step, the system confirms the deep alignment between the data-driven logic and the actual circuit characteristics, thereby outputting fault diagnosis conclusions that meet the high reliability requirements of industrial sites. When abnormal states such as secondary open circuits or terminal short circuits are diagnosed, the system automatically generates and pushes corresponding alarm prompts and maintenance work orders through the operation and maintenance monitoring system, assisting on-site personnel in quickly locating and eliminating faults, thus achieving intelligent closed-loop management of equipment status.

[0053] Specific application examples: Application scenario settings: A 10kV distribution network substation contains several electromagnetic current transformers that have been in operation for many years. To investigate potential safety hazards in the secondary circuits, maintenance personnel carried a portable testing device with an integrated high-frequency signal source and high-speed sampling module to conduct online diagnostics.

[0054] Implementation steps: Physical access and signal injection: Maintenance personnel attach the test probes of the detection device to the secondary terminal block of the current transformer. The device begins operation, automatically emitting a high-frequency step-scan excitation of 10 kHz to 40 kHz (500 Hz step, 61 frequency points in total). This high-frequency signal utilizes the high impedance isolation characteristics of the power frequency band and will not affect the original relay protection and metering services.

[0055] Timing sampling and frequency domain transformation: The signal is injected and maintained for one second at each frequency, and steady-state data for 0.5 seconds is captured afterward. Input power and output voltage signals are simultaneously acquired at a sampling rate of 100 kHz. Please refer to the appendix. Figure 8 , Figure 8This is a comparison diagram of the input and output voltage waveforms of the present invention. In the diagram, the horizontal axis Time (s) represents time, and the vertical axis Amplitude (V) represents voltage amplitude. Figure 8 As shown, within the steady-state time window, the system synchronously captures the input power signal (blue solid line Input Ui in the figure) and the output voltage signal (red solid line Output Uo in the figure), which has phase delay and amplitude attenuation, using these as the underlying computational basis. The device's built-in microprocessor calls the Hanning window and Fast Fourier Transform algorithm to extract the maximum complex component at the current injection frequency.

[0056] Amplitude and phase characteristic calculation and characteristic analysis (in conjunction with appendix) Figure 3 illustrate): After calculation, the system outputs a wideband feature sequence reflecting the underlying state. Please refer to the appendix. Figure 3 The equivalent circuit amplitude-frequency response curve is shown in the graph, which visually reflects the mapping relationship between the above calculation results and physical characteristics. The horizontal axis (scanning frequency, in kilohertz) represents the current frequency value of the high-frequency step excitation signal actively injected by the detection device into the secondary circuit of the current transformer; the vertical axis (ratio of output to input voltage amplitude) represents the ratio of the steady-state output voltage amplitude collected by the receiver to the injected input voltage amplitude at the corresponding frequency, which visually reflects the degree of signal attenuation caused by the overall equivalent impedance of the secondary circuit at the current frequency.

[0057] Combination Figure 3 As can be seen, the black solid line and circle marking (normal operating state) in the figure represent the physical frequency response characteristics of the secondary circuit of the current transformer under normal and intact conditions. The curve shows a significant bulge near 25 kHz, which is the resonance peak formed by the inherent inductance of the circuit and the matching capacitance of the detection device.

[0058] When a terminal short circuit occurs during simulation, the black dashed line and square mark in the figure (terminal short circuit state) represent the frequency response characteristics of the secondary circuit when a short circuit occurs. Because the short circuit causes part of the inductance to be bypassed, the equivalent inductance decreases and the damping increases, causing the overall resonant peak to shift towards a lower frequency direction (about 18 kHz), and the peak value becomes lower and flatter, resulting in a significant decrease in the quality factor.

[0059] If a secondary open circuit occurs, the corresponding black dotted line and triangle mark in the diagram (secondary open circuit state) represent the frequency response characteristics when the secondary circuit is open. At this time, the transformer core is deeply saturated, the secondary leakage inductance drops sharply, causing it to completely lose its resonance characteristics in this detection frequency band, exhibiting a monotonically increasing capacitive impedance response, and the resonant frequency shifts directly to beyond 40 kHz.

[0060] Model diagnosis and attribution analysis: After normalizing the extracted features, including mean, skewness, resonant frequency, and quality factor, they are input into the pre-trained optimized model of this invention in the cloud. The model instantly outputs the current state label as normal operation with a confidence level as high as 99.2%.

[0061] Subsequent physical explanation and verification: The backend system automatically generates an attribution graph based on an additive explanation model using game theory. The graph indicates that the top-ranked core contributing feature for this diagnosis, which is considered normal, is that the resonant frequency falls within the normal range of 24.5 kHz to 25.5 kHz. The second-ranked feature is that the phase reversal kurtosis (i.e., the kurtosis) conforms to the nominal inductor-capacitor resonance characteristics. This explanation aligns the black-box algorithm with the underlying Kirchhoff's laws mechanism, significantly increasing the on-site maintenance personnel's confidence in the diagnostic results.

[0062] Experimental verification and effect comparison: Experimental setup: To verify the authenticity and effectiveness of the invention, the R&D team built a hardware test bench in the laboratory, including a real 10kV current transformer, typical secondary cable distributed capacitance, and adjustable sampling load resistance (0.5 ohms to 100 ohms). Through physical switching, a total of 1,500 broadband characteristic samples of amplitude and phase frequencies were collected and generated (500 each for normal operation, secondary open circuit, and terminal short circuit). These were divided into a training set (1,200 samples) and a test set (300 samples) at an 8:2 ratio.

[0063] The algorithm comparison experiment setup and results are as follows: Support Vector Machine (SVM) model: As a baseline comparison algorithm, this model shows insufficient generalization ability when dealing with high-dimensional broadband frequency response features containing complex distributions. Its overall diagnostic accuracy on the experimental test set is only 84.6%, and it is particularly prone to confusion between terminal short circuits and normal states under light load conditions. Furthermore, the model optimization process is time-consuming.

[0064] Random Forest model: As a classic ensemble tree algorithm, it outperforms Support Vector Machines in handling non-linear features. With the same feature set, the overall diagnostic accuracy on the test set is improved to 90.3%. However, the deep structure of Random Forest results in a relatively bloated model, with significant computational overhead for predicting a single sample.

[0065] Basic Lightweight Gradient Boosting Machine Model (Manually Tuned Hyperparameters): Introducing this algorithm significantly improves training speed by leveraging its histogram algorithm and mutually exclusive feature binding mechanism. Without global hyperparameter optimization (using only default or manually set learning rates and tree depths), the test set accuracy reached 94.1%, but some edge samples still exhibited misclassification.

[0066] The optimized model of this invention, after introducing a tree-structured Parsons estimator for global Bayesian optimization, adaptively locks in the optimal combination (e.g., the optimal learning rate is locked at 0.035, and the number of leaf nodes is locked at 31). Experimental results show that the model's overall diagnostic accuracy on the test set jumps to 98.6%, and the inference time for single sample state identification is shortened to less than three milliseconds. The comprehensive evaluation index, which balances precision and recall, reaches 0.985.

[0067] To intuitively quantify the performance differences among the above algorithms, please refer to the appendix. Figure 4 This is a comparison chart of classification accuracy.

[0068] This figure details the performance of the four algorithms on the independent test set. The horizontal axis (classification model name) represents the four different data-driven algorithm models used in the experimental comparison. To avoid using English letters, they are referred to as traditional machine learning algorithms (Support Vector Machine, Random Forest), basic tree models without parameter optimization (basic lightweight gradient booster), and the optimized model proposed in this invention with global parameter optimization (this invention's optimized model).

[0069] The vertical axis of the graph (diagnostic accuracy, percentage) represents the percentage of samples that each algorithm model successfully identified on an independent test sample set, showing normal operation, terminal short circuit, and secondary open circuit states out of the total number of samples.

[0070] The gray bars in the graph visually represent the diagnostic accuracy of an algorithm model; taller bars indicate better performance. The percentage figures above the bars represent the precise accuracy values ​​for that model. Figure 4 As can be clearly seen, the gray bar on the far right, which belongs to the optimization model of this invention, is the tallest, and the value above it is 98.6%. This objectively quantifies the significant improvement in classification accuracy achieved by the present invention after introducing hyperparameter optimization.

[0071] Summary of Results: Comparative experiments fully demonstrate that this invention, through the physical hardware features excited by active high-frequency injection and combined with the efficient nonlinear fitting and parameter optimization capabilities of the optimization model, effectively overcomes the shortcomings of traditional methods in extracting weak fault features of distribution networks and the low recognition rate, and achieves high-precision and robust secondary circuit state identification.

Claims

1. A method for diagnosing faults in the secondary circuit of a current transformer, characterized in that, Includes the following steps: Connect the detection device in series to the secondary circuit of the current transformer to establish a high-frequency signal physical circuit with a signal receiving end. A step-scanning high-frequency excitation signal is injected into the secondary circuit of the current transformer as an input power signal, and the input power signal and the output voltage signal of the signal receiving end are sampled simultaneously to obtain time-series discrete data. Perform a Fast Fourier Transform on the time-series discrete data to extract the complex components of the input power signal and the complex components of the output voltage signal at the frequency position with the maximum amplitude in the frequency domain. Calculate the amplitude ratio and phase difference between the complex components of the output voltage signal and the complex components of the input power supply signal at different scanning frequencies, and generate a joint amplitude-frequency and phase-frequency distribution curve; The fusion features are extracted from the amplitude-frequency and phase-frequency joint distribution curves, and after normalization, they are divided into training and testing sets. A basic classifier based on the LightGBM algorithm is constructed, and the TPE algorithm is introduced to perform global hyperparameter optimization on the basic classifier. The base classifier is trained using the training set to obtain a state identification model, and the test set is processed using the state identification model to output fault diagnosis results. Based on the fault diagnosis results and the fusion features of the test set, the contribution value of the fusion features is calculated using the SHAP interpretation model, and the feature contribution distribution information is output.

2. The fault diagnosis method for the secondary circuit of a current transformer according to claim 1, characterized in that, The detection device is connected in series to the secondary circuit of the current transformer to establish a high-frequency signal physical circuit with a signal receiving end, including: The detection device is connected in series to the terminal block node of the secondary circuit of the current transformer to establish a high-frequency broadband equivalent circuit model including the physical secondary circuit of the current transformer and the detection device. The high-frequency broadband equivalent circuit model is transformed into a series topology consisting of resistors, inductors, and capacitors. The capacitor is a fixed matching capacitor pre-set inside the detection device; The signal receiving end is provided with a sampling resistor for extracting voltage characteristics. The sampling resistor is equivalent to a series resistor and inductor structure, which consists of an equivalent inductor and an adjustable resistor. The value of the equivalent inductance is constant, and the resistance value of the adjustable resistor is continuously adjusted within a set value range. The physical connection status of the secondary circuit of the current transformer is switched, and the three typical operating conditions are traversed: normal operation, secondary open circuit, and terminal short circuit.

3. The fault diagnosis method for the secondary circuit of a current transformer according to claim 1, characterized in that, A stepped scanning high-frequency excitation signal is injected into the secondary circuit of the current transformer as an input power signal. The input power signal and the output voltage signal of the signal receiver are simultaneously sampled to obtain discrete time-series data, including: Select the starting frequency of the step scanning high-frequency excitation signal, and perform discrete step scanning using a fixed frequency step value to generate discrete test frequency points; For each discrete stepping frequency point in the discrete test frequency points, the duration of the injected high-frequency excitation signal for stepping scanning is uniformly set; Discard the transitional data at the initial injection stage, and extract the data within the steady-state time window of the step-scan high-frequency excitation signal injection as the effective acquisition data for state identification; By using the multi-channel synchronous hardware triggering mechanism of the analog-to-digital converter, the input power signal and the output voltage signal included in the state identification effective acquisition data are synchronously discretized and sampled to obtain the time-series discrete data.

4. The fault diagnosis method for the secondary circuit of a current transformer according to claim 1, characterized in that, Perform a Fast Fourier Transform on the time-series discrete data to extract the complex components of the input power signal and the output voltage signal at the frequency position of maximum amplitude in the frequency domain, including: A window function is applied to the time-series discrete data to perform smooth truncation. The time-series discrete data is mapped to the frequency dimension using the Fast Fourier Transform algorithm to generate the corresponding frequency domain complex sequence; Set a preset frequency tolerance range around the center of the currently applied discrete scan frequency point; Within the preset frequency tolerance range, the amplitude of the frequency domain complex sequence is calculated and the frequency coordinate positions in the frequency domain complex sequence corresponding to the maximum amplitude of the input power signal and the output voltage signal are locked respectively. At the locked frequency coordinate position, the complex components of the input power signal and the complex components of the output voltage signal are extracted respectively.

5. The fault diagnosis method for the secondary circuit of a current transformer according to claim 1, characterized in that, Calculate the amplitude ratio and phase difference between the complex components of the output voltage signal and the complex components of the input power supply signal at different scanning frequencies, and generate a joint amplitude-frequency and phase-frequency distribution curve, including: The amplitude ratio of the output voltage signal and the input power signal is calculated using the magnitudes of the complex components of the output voltage signal and the complex components of the input power signal. A preset minimum constant is introduced as a bias protection in the denominator of the division operation for calculating the amplitude ratio. The phase difference between the output voltage signal and the input power signal is calculated using the phase angle between the complex components of the output voltage signal and the complex components of the input power signal. The amplitude ratio and phase difference at each discrete step frequency point are aggregated sequentially to construct a wideband feature vector that evolves with frequency. The broadband feature vector includes an amplitude feature vector and a phase feature vector; By changing the underlying hardware parameters and physical operating conditions of the detection device, the amplitude-frequency and phase-frequency joint distribution curves containing all loop states and load conditions are generated.

6. The fault diagnosis method for the secondary circuit of a current transformer according to claim 1, characterized in that, The fusion features are extracted from the amplitude-frequency and phase-frequency joint distribution curves. After normalization, the fusion features are divided into training and testing sets, including: The fusion features are extracted from the amplitude-frequency and phase-frequency joint distribution curve; the fusion features encompass statistical features, physical mechanism features, and operating condition parameter features; A standardization strategy involving division by standard deviation is used to normalize all continuous variables in the fusion features, mapping all dimensions of data to a standard distribution interval with a mean of zero and a variance of one. When performing the division operation to calculate the standard deviation, a very small positive constant is forcibly introduced into the denominator as a bias protection. The fused features are segmented using a hierarchical cross-validation strategy, and the sample distribution ratio of each category state in the segmented subset is constrained to be consistent with the original global sample library, thus dividing the subset into the training set and the test set.

7. The fault diagnosis method for the secondary circuit of a current transformer according to claim 1, characterized in that, Construct a base classifier based on the LightGBM algorithm, and introduce the TPE algorithm to perform global hyperparameter optimization on the base classifier, including: The LightGBM algorithm was selected as the base classifier. The input data of the basic classifier is the fusion feature after normalization, and the output result is mapped to three operating states: normal operation, secondary open circuit, and terminal short circuit. The base classifier approximates the true value corresponding to the fused feature by integrating multiple weak evaluators and iteratively updates it by minimizing the objective function containing a regularization term. The TPE algorithm is introduced, which utilizes historical iteration records and divides the parameter set that has been tried into an advantage group and an disadvantage group according to a set error threshold, and constructs nonparametric probability density estimates for each group. The error threshold is set as a fixed percentage quantile of the model error in the historical iteration records; The expected boost function is used as the sampling function to find the parameter sampling point that brings the maximum expected gain to the model performance, thus guiding the convergence of the hyperparameter search space.

8. The fault diagnosis method for the secondary circuit of a current transformer according to claim 1, characterized in that, Based on the fault diagnosis results and the fusion features contained in the test set, the contribution values ​​of the fusion features are calculated using the SHAP interpretation model, and the feature contribution distribution information is output, including: The fusion features contained in the test set and the corresponding fault diagnosis results are imported into the SHAP interpretation model as pre-input items; The SHAP interpretation model deconstructs the predicted values ​​of the imported fault diagnosis results into a linear function of binary variables; Based on the principle of marginal contribution, the expectation is obtained from all permutations and combinations of the fusion features contained in the test set, and the deviation of the global prediction expectation is allocated to each input feature to calculate the contribution value of a single fusion feature. The absolute value of the contribution value of each of the fused feature dimensions is taken and the global average is calculated to generate a local feature contribution analysis map, which is output as the feature contribution distribution information.

9. The fault diagnosis method for the secondary circuit of a current transformer according to claim 8, characterized in that, After outputting the feature contribution distribution information, the method further includes: Establish a physical simulation model of the RLC equivalent circuit; Under the same operating conditions, the same high-frequency scanning excitation is injected into the physical simulation model to extract the amplitude-frequency and phase-frequency attenuation curves under the theoretical state. The pure mechanistic features obtained from the physical simulation model are compared with the high contribution feature intervals shown in the local feature contribution analysis map by double alignment of time and working conditions. Based on the overlap between the simulation theory extrema and the high contribution range of the model, the consistency of the characteristic positive and negative push-pull logic and the evolution law of circuit impedance, a multi-dimensional weighted score is performed to complete the deep alignment verification of data-driven logic and actual circuit characteristics.

10. A fault diagnosis system for the secondary circuit of a current transformer, characterized in that, The method for diagnosing faults in the secondary circuit of a current transformer according to any one of claims 1-9 includes: The high-frequency signal injection module is connected in series to the secondary circuit of the current transformer and is used to apply a step-scanning high-frequency excitation signal within a preset frequency range as an input power signal to the secondary circuit of the current transformer. The signal sampling module is connected to the high-frequency signal injection module and the secondary circuit of the current transformer, respectively, and is used to synchronously sample the input power signal and the output voltage signal of the signal receiving end to obtain time-series discrete data. The frequency domain feature construction module, connected to the signal sampling module, is used to receive the time-series discrete data and perform a fast Fourier transform on it, extract the complex components of the input power signal and the complex components of the output voltage signal at the frequency position with the maximum amplitude in the frequency domain, calculate the amplitude ratio and phase difference between the complex components of the output voltage signal and the complex components of the input power signal at different scanning frequencies, and generate a joint amplitude-frequency-phase-frequency distribution curve. The feature processing module, connected to the frequency domain feature construction module, is used to extract fused features from the amplitude-frequency and phase-frequency joint distribution curve, and divide them into training set and test set after normalization processing. The model identification module, connected to the feature processing module, is used to construct a basic classifier based on the LightGBM algorithm, introduce the TPE algorithm to perform global hyperparameter optimization on the basic classifier, train the basic classifier using the training set to obtain a state identification model, and use the state identification model to process the test set to output fault diagnosis results. The decision analysis module, connected to the model identification module, is used to calculate the contribution value of the fusion feature based on the fault diagnosis result and the fusion feature of the test set using the SHAP interpretation model and output the feature contribution distribution information.

Images