Electric reactor impedance characteristic analysis system and method

Abstract

The invention discloses an electric reactor impedance characteristic analysis system and method, relates to the technical field of electric reactors, and aims to solve the problems of signal distortion, complex data processing, poor small sample generalization and single fault characteristic of a traditional method. The method comprises the following steps: S1, multi-dimensional sensing acquisition and anti-interference processing: acquiring a multi-dimensional operation state signal and ensuring the data quality through an anti-interference design; s2, processing original impedance data, and extracting multi-dimensional impedance features by adopting adaptive frequency domain decomposition (based on a variational principle) and transient feature capture (based on time-frequency redistribution) algorithms of a simplified mathematical expression; s3, small sample adaptation and model optimization: samples are enhanced and expanded through data of a simplified formula, small sample generalization is realized in combination with transfer learning, and a lightweight dynamic compensation model is constructed; and S4, performing fault diagnosis, and fusing multi-dimensional features. According to the method, mathematical expression is simplified, precision is guaranteed, the protection range is expanded, different reactor scenes are adapted, and the on-line monitoring requirement of a power system is met.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of electric reactors, in particular to an electric reactor impedance characteristic analysis system and method. BACKGROUND

[0002] The present application relates to the technical field of electric reactors, in particular to the impedance characteristic measurement, operation state monitoring and fault diagnosis technology of electric reactors in power systems, which is suitable for online analysis scenarios of dry / oil-immersed electric reactors of different voltage levels (35kV~220kV), and is especially suitable for impedance characteristic analysis requirements under small sample new equipment and complex working conditions.

[0003] Electric reactors are core equipment of power systems, and their impedance characteristics directly affect the stability of power grids. Traditional electric reactor impedance characteristic analysis methods have the following problems: source signal distortion: only relying on voltage / current single signal acquisition, without designing anti-interference mechanism for strong electromagnetic environment of power systems, weak common-mode noise suppression of sampling circuit, large signal transmission interference, low signal-to-noise ratio of original data, and affecting subsequent analysis accuracy; Only relying on impedance characteristics to construct discrimination criteria, without fusing temperature, vibration, insulation and other related features, easy to miss or misjudge faults, especially poor adaptability to multi-fault coupling scenarios.

[0004] Therefore, there is an urgent need for an electric reactor impedance characteristic analysis system and method. SUMMARY

[0005] The present application aims to provide an electric reactor impedance characteristic analysis system and method to solve the problem of low signal-to-noise ratio of original data and poor subsequent analysis accuracy in the prior art.

[0006] The technical solution of the present application to solve the above technical problems is as follows: The electric reactor impedance characteristic analysis method comprises the following steps: S1: Multi-dimensional sensing data acquisition and anti-interference processing S11 Multi-dimensional data acquisition: synchronously acquiring data through four types of sensing elements: electric signal acquisition elements: voltage and current acquisition elements suitable for electric reactor scenarios, acquiring voltage amplitude, current amplitude and phase at the inlet and outlet ends; temperature sensing elements: acquisition elements suitable for the heating characteristics of electric reactors, acquiring the temperature of key parts of the winding and core; vibration sensing elements: acquisition elements suitable for mechanical vibration, acquiring the vibration amplitude and frequency spectrum of the core and winding support structure; insulation state sensing elements: acquisition elements suitable for insulation monitoring, acquiring insulation-related feature parameters (such as partial discharge quantity).

[0007] S12 Anti-interference processing: sampling circuit: adopt common mode noise suppression conditioning circuit, cooperate with high precision signal conversion element, sampling rate adaptive adjustment with working condition (steady state / transient state); signal transmission: adopt low interference transmission link (such as optical fiber), guarantee multi-channel synchronous transmission, the link has anti-static interference ability; power module: adopt isolation power supply, add electromagnetic compatibility filter components at the entrance, meet the industrial electromagnetic compatibility standard.

[0008] S2: raw impedance data processing S21 raw impedance calculation: based on the collected voltage U, current I and phase difference , calculate the raw impedance frequency characteristic data .

[0009] S22 adaptive frequency domain decomposition (based on variational principle): construct variational model: minimize the sum of modal component bandwidth, constrain modal superposition equal to Z, simplify mathematical expression as: ; wherein, u is the modal component, is the center frequency, is the Dirac function, is the convolution operation. Alternating iteration solution: introduce Lagrange multiplier and penalty factor , the iteration formula is simplified as: , Modal selection: based on Match the fundamental wave (such as 4555Hz) and harmonic (such as 50Hz integer multiple) frequency band, eliminate noise modal, and reconstruct pure fundamental wave , harmonic Impedance component.

[0010] S23 Transient feature capture (based on time-frequency redistribution): wavelet transform: transform the transient period Z, the simplified expression is: wherein W is the wavelet coefficient, a is the scale parameter, b is the translation parameter, is the wavelet conjugate function. Scale-frequency mapping: (f is the frequency, is the wavelet center angular frequency). Time-frequency redistribution: redistribute W to the corresponding frequency-time point, the simplified expression is: wherein T is the time-frequency coefficient after redistribution, is the scale increment. Transient extraction: identify the transient time, impedance mutation amplitude ( is the transient impedance, is the normal impedance) and duration from T.

[0011] S3: small sample adaptation and model optimization S31 Data Enhancement: Time Domain Enhancement: Stretching and shifting the original impedance timing data, simplified expression is: , in, The elongation factor is 1. This represents the time offset. Frequency domain enhancement: Gaussian white noise is added to the modal component u, and the simplified expression is: in, The standard deviation of noise. Standard Gaussian white noise (mean 0, variance 1). Feature domain enhancement: enhancement of multi-dimensional feature vectors. (T is temperature, For vibration, Interpolation for insulation characteristics, simplified expression is: in, Interpolation coefficients ( ).

[0012] S32 Transfer Learning: Source Domain Pre-training: Using sufficient source domain reactor data (such as 35kV dry-type reactors), train a "time-series feature extraction network + multi-label classification network". The loss function is simplified to: Where N is the number of source domain samples, and y is the true label. To predict probabilities. Target domain fine-tuning: For small sample target domain reactors (such as 110kV oil-immersed reactors), freeze the bottom layer network of the pre-trained model and train only the top layer classification network. The total loss is simplified to: in, For classifying losses, For knowledge distillation loss, For distillation weight.

[0013] S33 Lightweight Dynamic Compensation: Model Input: Fundamental Wave ,harmonic Impedance components and temperature deviation ( For actual measured temperature, (Normal temperature), input features The gate control unit calculates the compensation coefficient k using the simplified formula: , , in: , Here, c represents the input gate and forget gate, and h represents the hidden state. , Here is the weight matrix, and b is the bias. For the Sigmoid function, This is element-wise multiplication. The impedance after compensation is: ( The error is <0.3% after compensation.

[0014] S4: Fault Diagnosis and Decision Output S41 Feature Fusion: Constructing Cognitive Enhancement Feature Vectors ,in Impedance characteristics (fundamental frequency deviation, harmonic content). Temperature characteristics (hot spot deviation, rate of temperature rise). Vibration characteristics (amplitude exceeding the standard rate). Insulation characteristics (average discharge quantity). S42 distance calculation: Euclidean distance (simplified expression): in, These are normal feature vectors. , For vector components. Mahalanobis distance (simplified expression): Where S is the normal sample covariance matrix, It is the inverse matrix.

[0015] S43 Fault Diagnosis and Decision-Making: Diagnosis Rule: If , If all values ​​are less than the corresponding threshold, the system is considered normal; otherwise, it is considered faulty, and the fault type is matched according to the "minimum distance principle" (e.g., inter-turn short circuit, core grounding). Location: Combine the fault spatial distribution function (predefined fault location probability) to generate a fault location heatmap. Maintenance recommendations: Based on the health index HI (0100), the system predicts the fault trend and outputs the maintenance priority (e.g., HI<30: 24-hour maintenance; 30≤HI<60: weekly maintenance).

[0016] The reactor impedance characteristic analysis system is used to implement the above method and includes the following modules: Multi-dimensional acquisition module: containing electrical signal, temperature, vibration and insulation status acquisition elements, combined with anti-interference circuit and low interference transmission unit to realize high signal-to-noise ratio acquisition of multi-dimensional signals; Data processing module: Configured with a variational mode decomposition unit (simplifying iterative formulas) and a time-frequency redistribution-based transient capture unit (simplifying time-frequency coefficient calculations) to complete impedance data decomposition and transient feature extraction; Model optimization module: Configured with a data augmentation unit (simplifying augmentation formulas), a transfer learning unit (simplifying loss functions), and a lightweight temporal compensation unit (simplifying gating calculations) to achieve small sample adaptation and dynamic impedance data compensation; Fault diagnosis module: Configured with a multi-dimensional feature fusion unit, a distance calculation unit (simplified Euclidean / Mahavir distance formula), and a decision output unit to realize fault identification, location, and maintenance suggestion generation.

[0017] The present invention has the following beneficial effects: The technical effects of this invention can be achieved step by step along the logical chain of "source data quality - data processing accuracy - scenario adaptability - fault diagnosis reliability": First, through multi-dimensional sensing acquisition and full-link anti-interference design, it not only makes up for the lack of fault characteristics in traditional single electrical signal acquisition, but also reduces the interference of the strong electromagnetic environment of the power system on the original data, providing high-quality and high-reliability basic data for subsequent analysis. Secondly, the core data processing algorithms (frequency domain decomposition, transient capture, and dynamic compensation) simplify mathematical expressions to improve engineering adaptability while retaining the core capabilities of accurately separating fundamental and harmonic waves, timely capturing transient faults, and correcting impedance measurement errors. This solves the dual contradiction of traditional algorithms being "cumbersome and difficult to implement" and "insufficient in accuracy and difficult to use". Furthermore, by combining multi-dimensional data augmentation with transfer learning, the model overfitting problem in small-sample new scenarios (such as new reactors and equipment of different voltage levels) is effectively alleviated, the scope of technical solution adaptation is greatly broadened, and the technical adaptation cycle of new equipment is shortened. Finally, by integrating multi-dimensional features of "electricity, heat, mechanics, and insulation" and combining simplified Euclidean and Mahalanobis distances to construct fault discrimination criteria, we can avoid misjudgment and missed judgment of faults caused by single impedance features, and achieve integrated output of fault type identification, location positioning, and operation and maintenance suggestions. Ultimately, this forms a comprehensive technical advantage of "reliable data, efficient processing, wide applicability, and accurate diagnosis", which fully meets the engineering needs of power systems for online monitoring and fault diagnosis of reactors. Attached Figure Description

[0018] Figure 1 This is a block diagram of the system of the present invention.

[0019] Figure 2 The diagram illustrates the specific steps of the method of the present invention.

[0020] Figure 3 This is a flowchart for fault diagnosis.

[0021] Figure 4 The time-frequency diagram of the wavelet transform for time-frequency redistribution. Detailed Implementation

[0022] The principles and features of the present invention are described below with reference to the accompanying drawings. The examples given are only for explaining the present invention and are not intended to limit the scope of the present invention.

[0023] refer to Figures 1-4As shown, a method for analyzing the impedance characteristics of a reactor includes the following steps: S1: Multi-dimensional sensor data acquisition and anti-interference processing, acquiring multi-dimensional operating status signals during reactor operation, and reducing source signal distortion through anti-interference design; S2: Raw impedance data processing, calculating raw impedance frequency characteristic data from the acquired electrical signals, and processing the raw impedance frequency characteristic data using an adaptive frequency domain decomposition algorithm and a transient feature capture algorithm to extract multi-dimensional impedance features; S3: Small sample adaptation and model optimization, expanding the sample size through data augmentation, and combining transfer learning to achieve model generalization in small sample scenarios, constructing a lightweight feature extraction and dynamic compensation model; S4: Fault diagnosis and decision output, integrating multi-dimensional features to construct fault discrimination criteria, realizing reactor fault type identification, location positioning, and trend prediction, and generating operation and maintenance decision suggestions. The overall framework of the method covers the entire process of "data acquisition-processing-optimization-diagnosis," breaking through the limitations of traditional methods with single features and low generalization. It can adapt to analysis scenarios of different types and voltage levels of reactors, providing basic protection for subsequent refined technical solutions and maximizing coverage of potential application scenarios.

[0024] Step S1, "Multi-dimensional sensor data acquisition and anti-interference processing," specifically includes: S11: Multi-dimensional sensor data acquisition, which obtains data through the following sensing elements: Electrical signal acquisition elements: voltage and current acquisition elements adapted to reactor scenarios, acquiring the voltage amplitude, current amplitude, and phase at the reactor's input or output terminals; Temperature sensing elements: temperature acquisition elements adapted to the reactor's heating characteristics, acquiring the temperature of key parts of the reactor windings and core; Vibration sensing elements: vibration acquisition elements adapted to the reactor's mechanical vibration, acquiring the vibration amplitude and phase of the reactor core and winding support structure. Spectrum; Insulation status sensing element: an acquisition element adapted for reactor insulation monitoring, collecting relevant characteristic parameters of reactor insulation; S12: Anti-interference processing, including: Sampling circuit anti-interference: adopting a conditioning circuit with common-mode noise suppression capability, combined with high-precision signal conversion elements, the sampling rate is adaptively adjusted according to the reactor operating conditions; Signal transmission anti-interference: adopting a low-interference transmission link to ensure synchronous transmission of multi-channel data, and the link has anti-static interference capability; Power supply anti-interference: adopting an isolated power supply module, with an electromagnetic compatibility filter component added to the power input, conforming to industrial electromagnetic compatibility standards. Ensuring data quality from the source, multi-dimensional sensing elements enrich fault correlation characteristics, avoiding missed detections of single electrical signals; the full-link anti-interference design reduces interference from the strong electromagnetic environment of the power system, improves the data signal-to-noise ratio, and provides highly reliable raw data for subsequent processing and diagnosis.

[0025] In step S2, the "adaptive frequency domain decomposition algorithm" is a mode decomposition algorithm based on the variational principle. The specific processing steps include: S211: Constructing a variational model with the objective of minimizing the sum of the bandwidths of each mode component. The constraint condition is that the superposition of each mode component equals the original impedance data Z. The simplified mathematical expression is: Where u is the modal component, For the center frequency, For the Dirac function, For convolution operations, S212: The above model is solved using an alternating iterative method, introducing Lagrange multipliers. (The time partial derivative is Z, and Z represents the original impedance data.) With penalty factor The iterative formula simplifies to: , in: For the new iterative modal components, For the old iterative mode components, S213: Based on the center frequency of the new iteration; By matching the fundamental and harmonic frequency bands, eliminating noise modes, and reconstructing pure fundamental and harmonic impedance components, the mathematical expression of variational mode decomposition is simplified while retaining the core capabilities of "accurate mode separation and anti-aliasing." This effectively separates the fundamental frequency from multiple harmonics, improves the purity of impedance components, and lays a high-precision data foundation for subsequent compensation and feature extraction.

[0026] The "transient feature capture algorithm" in step S2 is a wavelet transform algorithm based on time-frequency redistribution. The specific processing steps include: S221: Perform wavelet transform on the original impedance data Z, and the mathematical expression is simplified to: Where W represents the wavelet coefficients, a is the scaling parameter, and b is the translation parameter. S222: Wavelet conjugate function; S222: Scale-frequency mapping: (f is the frequency, (W is the wavelet center angular frequency), S223: Time-frequency redistribution, redistributing the wavelet coefficients W to the corresponding frequency-time points, simplified expression is: Where T represents the time-frequency coefficients after redistribution. S224: Extracts the transient occurrence time, impedance change amplitude, and duration from the time-frequency coefficient T. While simplifying the mathematical expression of wavelet transform and time-frequency redistribution, it still overcomes the limitation of traditional time-frequency analysis that "resolution cannot be simultaneously achieved," accurately capturing transient fault characteristics such as lightning strikes and switching operations, providing a basis for early fault warning.

[0027] Step S3, "data augmentation," specifically includes: S311: Time-domain augmentation: stretching and shifting the original impedance time-series data Z, simplified as follows: , in: , To enhance the data, The elongation factor is 1. For time offset; S312: Frequency domain enhancement: Add random noise to the modal component u, the simplified expression is: in: To enhance the postmodal components, The standard deviation of noise. Standard Gaussian white noise; S313: Feature domain enhancement: interpolation of multi-dimensional feature vector X (including impedance, temperature, vibration, and insulation features), simplified expression is: in: To enhance the feature vector, , The original feature vector, These are the interpolation coefficients. Through a simplified multi-dimensional data augmentation formula, the amount of training data in small-sample scenarios is effectively expanded, impedance changes under complex working conditions are simulated, model overfitting is alleviated, and a sufficient sample base is provided for transfer learning.

[0028] Step S3, "transfer learning," specifically includes: S321: Source domain pre-training: Using sufficient source domain reactor data, train a "temporal feature extraction network + multi-label classification network," with the loss function simplified as follows: in: The loss function is the pre-training loss, where N is the number of samples in the source domain and y is the true fault label. For predicting probabilities; S322: Target domain fine-tuning: For small sample target domain reactors, freeze the bottom layer network of the pre-trained model and train only the top layer classification network. The total loss is simplified to: in: To fine-tune the total loss, For classifying losses, For knowledge distillation loss, Distillation weights are used. The expression of the transfer learning loss function is simplified, and source domain knowledge is reused through "pre-training-fine-tuning". There is no need to train the model from scratch in new scenarios with small samples (such as new reactors), which shortens the adaptation cycle and reduces the cost of engineering applications.

[0029] In step S3, the "dynamic compensation model" is a lightweight timing model based on a gating mechanism, specifically including: S331: The model input is the fundamental impedance component. Harmonic impedance components and temperature deviation The input feature vector is S332: The compensation coefficient k is calculated through the gating unit, and the simplified formula is: , , in , Here, c represents the input gate and forget gate, and h represents the hidden state. , Here is the weight matrix, and b is the bias. For the Sigmoid function, Element-wise multiplication; S333: Impedance data after compensation: Where Z represents the original impedance data. To compensate for the data, the gating timing model formula is simplified while retaining the core capability of "dynamically capturing timing dependencies." It corrects impedance deviations caused by sampling errors and environmental interference, improving measurement accuracy. The lightweight design adapts to edge device deployment, meeting the real-time requirements of online monitoring.

[0030] In step S4, the "fault discrimination criterion" is constructed based on Euclidean distance and Mahalanobis distance, specifically including: S411: fusing multi-dimensional features to obtain an enhanced feature vector. (Including impedance, temperature, vibration, and insulation characteristics); S412: Calculate the Euclidean distance (simplified expression): in: For Euclidean distance, These are normal feature vectors. , For vector components; S413: Calculate the Mahalanobis distance (simplified expression): in: Let S be the Mahalanobis distance, and S be the covariance matrix of the normal samples. It is the inverse matrix; S414: Fault detection: If , If all distances are less than the corresponding threshold, the system is considered normal; otherwise, it is considered faulty, and the fault type is matched according to the minimum distance. The distance calculation expression is simplified, and the system uses "Euclidean distance (intuitive difference) + Mahalanobis distance (to eliminate interference from feature correlation)" for discrimination, avoiding misjudgment based on a single feature, improving the accuracy of identifying complex faults (such as multiple coupled faults), and providing accurate diagnostic basis for operation and maintenance.

[0031] A reactor impedance characteristic analysis system includes: a multi-dimensional acquisition module containing acquisition elements for electrical signals, temperature, vibration, and insulation status, which, together with an anti-interference circuit and a low-interference transmission unit, enables high signal-to-noise ratio multi-dimensional data acquisition; Data processing module: Configured with a mode decomposition unit based on variational principle and a transient capture unit based on time-frequency redistribution to complete impedance data decomposition and transient feature extraction; Model optimization module: Configures data augmentation, transfer learning, and lightweight time series compensation units to achieve small sample adaptation and dynamic compensation of impedance data; Fault Diagnosis Module: This module is equipped with multi-dimensional feature fusion, distance calculation, and decision output units to achieve fault identification, location, and maintenance suggestion generation. It engineers and implements analytical methods, with a modular design adaptable to different scenarios, providing stable hardware and software support for reactor impedance analysis. This promotes the industrial application of the technical solution and meets the needs of online monitoring projects in power systems.

[0032] Taking "Impedance Characteristic Analysis of 110kV Oil-Immersed Reactor" as an example, the implementation process is explained in detail: Hardware deployment Electrical signal acquisition: 0.2-grade voltage acquisition element (JDZ10-110) and current acquisition element (LZZBJ9-110) are selected and installed at the reactor input terminal; Temperature acquisition: Fiber optic temperature acquisition element (FBG-T200, accuracy ±0.5℃) is selected and attached to the middle of the winding (3 points) and the bottom of the iron core (2 points). Vibration acquisition: A piezoelectric acceleration acquisition element (CA-YD-103, range ±5g) was selected and installed on the iron core bracket (2 points) and the winding end plate (2 points). Insulation acquisition: An ultra-high frequency partial discharge acquisition element (HFCT-01, 300-1500MHz) is selected and adsorbed on the outer wall of the oil tank (evenly distributed at 3 points). Anti-interference configuration: Fiber optic transmission link (HFBR-1521Z, delay <10ms) and isolated power supply (RECOMR-78E5.0-0.5) are adopted, and an EMC filter component is added to the power input.

[0033] Data Acquisition and Processing Sampling rate control: 500Hz for steady-state operation (load rate <80%), and 10kHz for transient operation (load rate ≥80% or switching operation); Frequency domain decomposition (variational principle): number of modes Punishment factor Iterate 50 times (convergence threshold) The fundamental mode (50Hz) and the 5th harmonic mode were obtained by decomposition. After removing the noise mode, the fundamental / harmonic components were reconstructed. Transient capture (time-frequency redistribution): using complex Morlet wavelets (center frequency) ), with 18 levels of decomposition scale and frequency resolution. In the lightning strike simulation test, the internal impedance change of 2ms (125Ω→82Ω) was successfully captured.

[0034] Model training and optimization Source domain data: 1000 sets of data from 35kV dry-type reactors (including 5 types of faults), pre-trained "GRU+CNN" model, loss function The convergence was achieved to 0.08, with an accuracy of 96.2%. Target domain data: 50 sets of data for 110kV oil-immersed reactors (10 sets for normal operation and 40 sets for fault operation), fine-tuning the loss function. The accuracy rate improved from 58% to 89.5%; Dynamic compensation: The lightweight gating model is deployed on an STM32H7 chip with 8 input feature dimensions, 64 hidden layer nodes, and a compensation impedance error of 0.25% and a latency of 18ms.

[0035] Fault diagnosis verification Single fault test (inter-turn short circuit): ,calculate , The short circuit between turns was matched, and the location was found in the middle of the winding (probability 0.82), which is consistent with the actual situation. Multi-fault test (inter-turn short circuit + partial discharge): , It can identify dual-fault coupling, locate the middle of the winding + the left side of the oil tank (probability 0.75 + 0.68), with an accuracy of 88%; Long-term monitoring: After 30 days of continuous operation, the Health Index (HI) dropped from 92 to 75, and the prediction for one month later... It provides a "repair within 1 week" suggestion to avoid malfunctions in advance.

[0036] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. 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 method for analyzing the impedance characteristics of a reactor, characterized in that, Includes the following steps: S1: Multi-dimensional sensor data acquisition and anti-interference processing: Acquire multi-dimensional operating status signals during reactor operation, and reduce source signal distortion through anti-interference design; S2: Raw impedance data processing: The raw impedance frequency characteristic data is calculated from the collected electrical signal. The raw impedance frequency characteristic data is processed by an adaptive frequency domain decomposition algorithm and a transient feature capture algorithm to extract multi-dimensional impedance features. S3: Small sample adaptation and model optimization. Expand the sample size through data augmentation, combine transfer learning to achieve model generalization in small sample scenarios, and build a lightweight feature extraction and dynamic compensation model. S4: Fault diagnosis and decision output, integrates multi-dimensional features to construct fault discrimination criteria, realizes reactor fault type identification, location and trend prediction, and generates operation and maintenance decision suggestions.

2. The method for analyzing the impedance characteristics of a reactor according to claim 1, characterized in that, Step S1, "Multi-dimensional sensor data acquisition and anti-interference processing," specifically includes: S11: Multi-dimensional sensing data acquisition, acquiring data through the following sensing elements: Electrical signal acquisition components: Voltage and current acquisition components adapted for reactor scenarios, which acquire the voltage amplitude, current amplitude, and phase at the reactor's input or output terminals; Temperature sensing element: A temperature acquisition element adapted to the heating characteristics of the reactor, which collects the temperature of key parts of the reactor winding and core; Vibration sensing element: A vibration acquisition element adapted to the mechanical vibration of the reactor, which collects the vibration amplitude and spectrum of the reactor core and winding support structure; Insulation status sensing element: an acquisition element adapted for reactor insulation monitoring, which collects the insulation characteristic parameters of the reactor; S12: Anti-interference processing, including: Sampling circuit anti-interference: adopts a conditioning circuit with common-mode noise suppression capability, combined with high-precision signal conversion components, and the sampling rate is adaptively adjusted according to the reactor operating conditions; Signal transmission anti-interference: adopts a low-interference transmission link to ensure synchronous transmission of multi-channel data, and the link has anti-electrostatic interference capability.

3. The method for analyzing the impedance characteristics of a reactor according to claim 1, characterized in that, The "adaptive frequency domain decomposition algorithm" in step S2 is a mode decomposition algorithm based on the variational principle. The specific processing steps include: S211: Construct a variational model with the objective of minimizing the sum of the bandwidths of all modal components. The constraint is that the superposition of all modal components equals the original impedance data Z. The mathematical expression is: ; S212: The above model is solved using the alternating iterative method, introducing Lagrange multipliers. With penalty factor The iterative formula simplifies to: , in: For the new iterative modal components, For the old iterative mode components, The new iteration center frequency; S213: Based on center frequency Match the fundamental and harmonic frequency bands, eliminate noise modes, and reconstruct pure fundamental and harmonic impedance components.

4. The method for analyzing the impedance characteristics of a reactor according to claim 1, characterized in that, The "transient feature capture algorithm" in step S2 is a wavelet transform algorithm based on time-frequency redistribution. The specific processing steps include: S221: Perform wavelet transform on the original impedance data Z. The mathematical expression simplifies to: Where: W represents the wavelet coefficients, a is the scaling parameter, and b is the translation parameter. It is the wavelet conjugate function; S222: Scale-frequency mapping: Where f is the frequency, The wavelet center angular frequency; S223: Time-frequency redistribution, redistributing wavelet coefficients W to corresponding frequency-time points, simplified expression is: where: T represents the time-frequency coefficients after redistribution. For scale increments; S224: Extract the transient occurrence time, impedance change amplitude, and duration from the time-frequency coefficient T.

5. The method for analyzing the impedance characteristics of a reactor according to claim 4, characterized in that, Step S3, "data augmentation," specifically includes: S311: Time-domain enhancement: Stretching and shifting the original impedance time-series data Z, simplified expression is: , in: , To enhance the data, The elongation factor is 1. This is the time offset; S312: Frequency Domain Enhancement: Random noise is added to the modal component u, and the simplified expression is: in: To enhance the postmodal components, The standard deviation of noise. It is Gaussian white noise; S313: Feature Domain Enhancement: Interpolation of multi-dimensional feature vector X, simplified expression is: in: To enhance the feature vector, , The original feature vector, These are the interpolation coefficients.

6. The method for analyzing the impedance characteristics of a reactor according to claim 5, characterized in that, Step S3, "transfer learning," specifically includes: S321: Source Domain Pre-training: Using sufficient source domain reactor data, train a "temporal feature extraction network + multi-label classification network". The loss function is simplified to: in: The loss function is the pre-training loss, where N is the number of samples in the source domain and y is the true fault label. To predict probabilities; S322: Target Domain Fine-tuning: For small-sample target domain reactors, freeze the bottom layer network of the pre-trained model and train only the top-level classification network. The total loss is simplified to: in: To fine-tune the total loss, For classifying losses, For knowledge distillation loss, For distillation weight.

7. The method for analyzing the impedance characteristics of a reactor according to claim 5, characterized in that, In step S3, the "dynamic compensation model" is a lightweight timing model based on a gating mechanism, specifically including: S331: Model input is the fundamental impedance component. Harmonic impedance components and temperature deviation The input feature vector is ; S332: The compensation coefficient k is calculated using the gating unit, and the simplified formula is: , , in: , Here, c represents the input gate and forget gate, and h represents the hidden state. , Here is the weight matrix, and b is the bias. For the Sigmoid function, Element-wise multiplication; S333: Impedance data after compensation: Where Z represents the original impedance data. This is the data after compensation.

8. The method for analyzing the impedance characteristics of a reactor according to claim 5, characterized in that, The "fault discrimination criterion" in step S4 is constructed based on Euclidean distance and Mahalanobis distance, and specifically includes: S411: Fusing multi-dimensional features to obtain an enhanced feature vector ; S412: Calculate Euclidean distance (simplified expression): in: For Euclidean distance, These are normal feature vectors. , For vector components; S413: Calculate Mahalanobis distance (simplified expression): in: Let S be the Mahalanobis distance, and S be the covariance matrix of the normal samples. It is the inverse matrix; S414: Fault diagnosis: If both are less than the corresponding threshold, the system is considered normal; otherwise, the system is considered faulty, and the fault type is matched according to the minimum distance.

9. A reactor impedance characteristic analysis system, used to implement the method described in any one of claims 1 to 8, characterized in that, include: Multi-dimensional acquisition module: Includes electrical signal, temperature, vibration, and insulation status acquisition elements, combined with anti-interference circuit and low-interference transmission unit to achieve high signal-to-noise ratio multi-dimensional data acquisition; Data processing module: Configured with a mode decomposition unit based on variational principle and a transient capture unit based on time-frequency redistribution to complete impedance data decomposition and transient feature extraction; Model optimization module: Configures data augmentation, transfer learning, and lightweight time series compensation units to achieve small sample adaptation and dynamic compensation of impedance data; Fault diagnosis module: Configured with multi-dimensional feature fusion, distance calculation, and decision output units to realize fault identification, location, and maintenance suggestion generation.

Images