Power transformer arc discharge multi-parameter detection simulation platform and fault diagnosis method

Abstract

The invention discloses a power transformer arc discharge multi-parameter detection simulation platform and a fault diagnosis method, and relates to the technical field of power system equipment state monitoring and fault diagnosis. The platform comprises a transformer body, a replaceable discharge module, a multi-parameter sensing unit and a signal processing and diagnosis module, and can truly reproduce various typical arc discharge faults of a needle plate, an air gap, a creeping surface, turn-to-turn and the like. The sensing unit is integrated with ultrahigh frequency and ultrasonic sensing probes, high-frequency current and voltage sensors, optical fiber temperature / pressure / strain sensors and the like, so that synchronous acquisition of multi-physical field signals is realized. According to the diagnosis method, through wavelet denoising and multi-dimensional feature extraction, a feature vector of multi-state parameter fusion in the process from partial discharge to arcing is constructed, and accurate classification of fault types is realized by using a support vector machine (SVM) model. The diagnosis method has high accuracy and early warning capability, effectively overcomes the limitation of single parameter diagnosis, and provides reliable technical support for transformer fault research and intelligent operation and maintenance.

Description

Technical Field

[0001] This invention relates to the field of power system equipment condition monitoring and fault diagnosis technology, specifically to a multi-parameter detection simulation platform for arc discharge of power transformers and a fault diagnosis method. Background Technology

[0002] Oil-immersed power transformers are core equipment in power systems, enabling voltage transformation and energy transmission. Their operational reliability directly impacts the safety and stability of the power grid. During long-term operation, due to factors such as insulation aging, manufacturing defects, or localized electric field concentration, partial discharge can easily occur inside the transformer. In severe cases, this can develop into high-energy arc discharge, leading to major accidents such as insulation breakdown, winding deformation, or even explosions and fires. Therefore, conducting simulation studies and accurate diagnosis of arc discharge faults is of great significance for achieving early fault warning, preventing equipment damage, and ensuring power grid safety.

[0003] Currently, research on transformer discharge faults mainly relies on laboratory simulation devices. However, existing technologies still have significant shortcomings in terms of the realism of fault simulation, the completeness of multi-parameter detection, and diagnostic accuracy, specifically manifested in the following technical bottlenecks:

[0004] First, the discharge model is too simplistic and cannot realistically reproduce complex fault scenarios.

[0005] Existing simulation platforms mostly employ simplified needle-plate or ball-ball electrode structures, which can only simulate localized, concentrated electric field discharges. They cannot accurately reproduce the evolution of complex defects typical within transformers, such as inter-turn oil gap discharges, surface creepage, and floating potential discharges. Especially for the progressive discharge behavior in oil-paper insulation systems caused by bubbles, impurities, or mechanical stress, existing models lack structural adjustability and operating condition matching, leading to distorted discharge signal characteristics and making it difficult to support high-reliability fault mechanism analysis and diagnostic algorithm verification.

[0006] Second, the sensing methods are isolated, and the acquisition of multi-physics information is asynchronous:

[0007] Traditional detection methods often rely on single sensors for partial discharge monitoring, such as ultra-high frequency or ultrasonic sensors. However, these methods are limited by issues like significant signal attenuation and complex propagation paths in transformer oil, resulting in blind spots and low positioning accuracy. Furthermore, macroscopic physical parameters closely related to discharge energy release, such as electromagnetic pulses, temperature, pressure, and mechanical strain, are often neglected or collected independently, lacking the ability to monitor multiple parameters simultaneously. In addition, installing multiple sensors requires densely packed openings in the oil tank, which not only weakens the tank structure but also introduces the risk of oil leakage, severely restricting the feasibility of multi-source information integration.

[0008] Third, the spatiotemporal inaccuracy of the signal affects the accuracy of fault inversion and positioning:

[0009] In existing systems, different types of sensors are usually deployed independently, such as high-frequency current sensors, voltage sensors, and fiber optic sensors. The lack of a unified time reference and synchronous acquisition mechanism leads to deviations in the time axis of multi-source signals, making it difficult to accurately correlate discharge events with thermal, mechanical, and other response processes. This limits the accurate inversion of the discharge energy evolution path and fault development trend.

[0010] Fourth, the diagnostic methods rely on a single parameter, resulting in a low level of intelligence.

[0011] Current fault identification methods mostly rely on single-parameter threshold judgment or simple pattern recognition, which are susceptible to noise interference, environmental fluctuations, and equipment differences, resulting in high false alarm and false negative rates. Although some studies have attempted to introduce multi-parameter fusion approaches, they lack a systematic feature extraction and data fusion framework, failing to effectively uncover the intrinsic relationships between heterogeneous data and making it difficult to achieve accurate classification and early identification of various fault types.

[0012] In summary, existing technologies still have significant shortcomings in terms of the realism of discharge fault simulation, the integration of multi-parameter detection, the synchronization of signal acquisition, and the accuracy of fault diagnosis. Therefore, there is an urgent need to construct a comprehensive experimental platform and method that can realistically simulate various typical arc discharge modes, integrate multi-dimensional sensing units, achieve synchronous acquisition of multi-source signals, and support intelligent collaborative diagnosis. Summary of the Invention

[0013] To address the aforementioned issues, this invention provides a multi-parameter detection and simulation platform for arc discharge in power transformers and a fault diagnosis method. This solution integrates replaceable discharge modules, multi-parameter sensing integration, safety protection design, and support vector machine intelligent diagnosis, effectively improving the realism of fault simulation, the comprehensiveness of monitoring, and the level of intelligent diagnosis.

[0014] The technical solution adopted in this invention is as follows:

[0015] A multi-parameter detection and simulation platform for arc discharge of power transformers includes a transformer body, a replaceable discharge module, a multi-parameter sensing unit, and a signal processing and diagnostic module.

[0016] The transformer body is the basic carrier of the entire platform. It adopts a true oil-immersed power transformer structure, which includes an iron core, windings, and insulating oil. The transformer body can withstand the rated voltage and short-term overvoltage operation.

[0017] The replaceable discharge module is inserted into a designated position inside the transformer body oil tank through the lifting hole on the top of the transformer body. The replaceable discharge module can be independently pressurized to stimulate partial discharge without power interruption.

[0018] The multi-parameter sensing unit integrates an ultra-high frequency ultrasonic probe, a high frequency pulse current sensor, a capacitive voltage divider sensor, and fiber optic temperature, pressure, and strain sensors to construct a multi-dimensional synchronous monitoring network covering electromagnetic, acoustic, electrical, thermal, and mechanical fields, enabling comprehensive perception and data acquisition of transformer arc discharge faults.

[0019] The signal processing and diagnostic module is connected to the multi-parameter sensing unit, which performs multi-parameter information fusion processing and diagnoses discharge faults.

[0020] Furthermore, the replaceable discharge module can be a needle-plate discharge model, an air gap discharge model, a surface discharge model, or an inter-turn discharge model.

[0021] The needle-plate discharge model uses pointed electrodes and flat plates to simulate point discharge caused by local electric field concentration; the air gap discharge model uses spherical electrodes and flat plates to simulate gap breakdown caused by air bubbles or impurities in oil; the surface discharge model uses cylindrical electrodes and flat plates to simulate surface creepage; and the inter-turn discharge model uses flat copper strips wrapped in crepe paper and epoxy resin pads to simulate oil gap discharge caused by the deterioration of inter-turn insulation in windings.

[0022] Each discharge model is connected to an independent voltage regulating station, and a stepped pressurization method is used to stably trigger partial discharge.

[0023] Furthermore, the multi-parameter sensing unit includes an ultra-high frequency ultrasonic integrated probe, a high frequency pulse current sensor, a capacitive voltage divider sensor, and an optical fiber sensor system;

[0024] The ultra-high frequency ultrasonic integrated probe is installed at the side opening of the transformer tank via a flange; the high frequency pulse current sensor is installed in series on the bushing end screen grounding wire, neutral point grounding wire, and core / clamping grounding wire; the capacitive voltage divider voltage sensor is installed on the bushing end screen grounding wire, which uses the bushing's own capacitance as the high voltage arm, and the end screen end is connected in series with a precision measuring capacitor to form the low voltage arm; the fiber optic sensor system includes fiber optic temperature sensors attached to the top and four inner walls of the transformer tank, fiber optic pressure sensors introduced into the transformer tank through a sealed through-plate, and fiber optic strain sensors magnetically fixed to the center point of the side wall of the transformer tank.

[0025] Furthermore, the multi-parameter sensing unit also includes an impedance sensor and a pressure relief protection device;

[0026] The impedance sensor is installed on the grounding wire of the transformer body's pressurization system; the pressure relief protection device is a pressure relief valve, which is installed at the circular pressure relief valve hole on the top of the transformer body.

[0027] A method for diagnosing arc discharge faults in power transformers, based on the aforementioned multi-parameter detection and simulation platform for arc discharge in power transformers, includes the following steps:

[0028] Step 1, Synchronous acquisition and preprocessing of multi-parameter data: Multi-parameter signals are synchronously acquired by each sensor of the multi-parameter sensing unit, and wavelet transform and digital filtering are used for noise reduction and alignment to complete the preprocessing of the discharge data.

[0029] Step 2, Multi-dimensional feature extraction and feature vector construction: Extract multi-dimensional features from partial discharge, electromagnetic pulse and macroscopic physical response, and construct a normalized comprehensive feature vector;

[0030] Step 3, Support Vector Machine (SVM) model construction and training: Using Support Vector Machine (SVM) combined with RBF kernel function and One-vs-One strategy, the model is trained and the parameters are optimized on typical fault samples;

[0031] Step 4, Real-time Fault Diagnosis and Result Output: Input the real-time features into the trained SVM model for classification and reasoning, and output the fault type, confidence level and warning information.

[0032] Furthermore, step 1 includes the following steps:

[0033] Step 1.1, Synchronous Acquisition Mechanism:

[0034] Using the power frequency voltage signal as a time reference, all sensors in the multi-parameter sensing unit are triggered to start synchronously.

[0035] When the amplitude of UHF or ultrasound AE signal exceeds the preset threshold, multi-channel data recording is initiated.

[0036] Record the occurrence time, phase angle φ, and raw waveform data of each fatal discharge pulse;

[0037] All electromagnetic, acoustic, electrical, thermal, and mechanical detection parameters are stored using a unified timestamp to maintain spatiotemporal consistency.

[0038] Step 1.2, Signal Denoising and Interference Suppression:

[0039] For UHF and ultrasonic signals, wavelet transform is used for noise reduction.

[0040] The wavelet decomposition formula is as follows:

[0041]

[0042] The scaling and translation mother wavelet function is defined as follows:

[0043]

[0044] In the formula, C a,b Ψ represents the wavelet coefficients, indicating the similarity between the signal f(t) and the wavelet basis function at scale a and displacement b, reflecting the signal characteristic intensity at that scale and location; f(t) is the original signal, representing the time-domain signal to be analyzed; a,b Ψ(t) is the scaled and translated wavelet basis function, obtained from the mother wavelet function through scaling and translation transformations, used to match local features in the signal; Ψ(t) is the mother wavelet function, representing a pre-selected basic wavelet function with finite energy, rapid decay, and oscillatory characteristics; a is the scaling parameter, controlling the scaling degree of the wavelet function, a>0; b is the translation parameter, controlling the position movement of the wavelet function on the time axis, used for sliding analysis of local features of the signal at different time periods; The normalization coefficients ensure the consistency of wavelet function energy at different scales, making the transformation results comparable. The conjugate of the wavelet basis functions;

[0045] By setting a threshold, the wavelet coefficients are subjected to soft / hard thresholding, and then wavelet reconstruction is performed to separate the discharge signal from the noise.

[0046] For high-frequency pulse current and voltage signals, a bandpass digital filter is used to filter out power frequency interference and other low-frequency noise;

[0047] Step 1.3, Data Alignment and Normalization:

[0048] Data from all channels is precisely aligned by timestamp, and multi-source signals corresponding to the same discharge event are strictly matched in time; physical quantities of different dimensions, such as temperature, pressure, and current, are normalized to eliminate differences in magnitude.

[0049] Furthermore, step 2 includes the following steps:

[0050] Step 2.1, Partial Discharge Feature Set X PD :

[0051] The following statistical features were extracted from the PRPD maps generated from UHF and ultrasound signals:

[0052] Skewness: This reflects the left-right asymmetry of the spectrum;

[0053] Kuroshi: Describe the sharpness of the spectrum;

[0054] Maximum discharge quantity q max With average discharge quantity q avg ;

[0055] Pulse quantity distribution: Statistical analysis of the positive half-cycle N within one power frequency cycle+ With negative half-cycle N - The number of discharges and their ratio N + / N - ;

[0056] In the formula, q is the discharge amount of a single discharge pulse; μ is the mean of the discharge amount, μ=E[q] is the mathematical expectation or arithmetic mean of the discharge amounts of all discharge pulses, reflecting the overall discharge level; E[] is the mathematical expectation, which is the average operation over the random variable; σ is the standard deviation. This indicates the degree of dispersion of the discharge quantity around the mean, reflecting the fluctuation of the discharge intensity; (q-μ) 3 This represents the cube of the deviation from the mean, used to measure the asymmetry of the distribution. A positive deviation indicates the presence of more high-intensity discharges; a negative deviation indicates that low-intensity discharges dominate. (q-μ) 4 The deviation from the mean to the fourth power amplifies the effect of extreme values ​​and makes it more sensitive to heavy tails or sharp peaks; q max The maximum discharge quantity among all detected discharge pulses within one power frequency cycle, reflecting the instantaneous discharge energy intensity, is used to determine whether a high-energy partial discharge exists; q avg N is the arithmetic mean of the discharge quantities of all discharge pulses. + N represents the total number of discharge pulses occurring during the positive half-cycle of the voltage from 0 to 180° within one power frequency cycle. - The total number of discharge pulses occurring during the negative half-cycle of voltage (180–360°) within one power frequency cycle;

[0057] Step 2.2, Electromagnetic Pulse Feature Set X EM :

[0058] Based on high-frequency pulse current I HF (t) and voltage V HF The following features are extracted from the (t) signal:

[0059] Frequency domain characteristics: dominant frequency f p =argmax f |F[I HF [t]|, peak amplitude of spectrum A peak Signal bandwidth BW;

[0060] Time-domain characteristics: Pulse rise time t r Pulse duration t d Pulse energy E p =∫I HF (t) 2 dt;

[0061] In the formula, f p The dominant frequency refers to the frequency component with the strongest energy in the discharge pulse spectrum, reflecting the oscillation characteristics of the discharge channel; I HF(t) is a high-frequency pulse current signal, acquired by a high-frequency current sensor, reflecting the current change at the moment of discharge; F[] is the Fourier transform of the time-domain signal I. HF (t) is converted to the frequency domain representation F[I HF (t)](f), revealing its frequency composition; A peak BW is the maximum amplitude value in the spectrum; BW is the bandwidth, referring to the amplitude at the main frequency f. p Centered on the peak value, the spectral amplitude drops to its maximum. The frequency range corresponding to time; t r Rise time refers to the time required for a pulse to rise from 10% to 90% of its amplitude; t d Duration refers to the time span from the start to the end of a pulse, often defined as the time interval from when the pulse first exceeds the threshold to when it falls back below the threshold; E p The pulse energy represents the total energy of the discharge pulse in the time domain, and is related to I. 2 R loss is proportional to its thermal effect and destructive potential; I HF (t) 2 Instantaneous power;

[0062] Step 2.3, Macroscopic physical quantity characteristic set X Phys :

[0063] Extracting dynamic change features from slowly varying signals acquired by fiber optic temperature, pressure, and strain sensors:

[0064] Rate of change: Temperature change rate dT / dt; Pressure change rate dP / dt;

[0065] Extreme value characteristics: highest temperature T max Maximum pressure P max Maximum strain mutation Δε max ;

[0066] In the formula, T represents temperature, the temperature value of the transformer oil or local structure obtained in real time by a fiber optic temperature sensor; t represents time, the time variable for signal acquisition; dT / dt represents the temperature change rate, indicating the rate of temperature change per unit time, reflecting the development rate of local overheating; P represents pressure, the internal oil pressure of the transformer measured by a fiber optic pressure sensor; dP / dt represents the pressure change rate, indicating the rate of pressure increase within the tank per unit time; T max The highest temperature, P, represents the maximum value among all temperature sampling points within a monitoring period. max ε represents the highest oil pressure measured within a specified time period; ε is the strain, the mechanical deformation of the transformer tank or local structure measured by a fiber optic strain sensor; Δε is the strain change, representing the strain increment relative to the initial state or background value at a given moment; Δε maxThe maximum strain abrupt change represents the peak abrupt change among all strain changes during a single discharge event.

[0067] Step 2.4, Feature Vector Construction:

[0068] After normalizing the above three types of features, they are concatenated into a unified multidimensional feature vector: x = [X PD ,X EM ,X Phys This serves as the input for subsequent classification models.

[0069] Furthermore, step 3 includes the following steps:

[0070] Step 3.1, Dataset Preparation:

[0071] A sample library containing various typical fault types was constructed, including: floating potential discharge, winding insulation defect discharge, oil bubble discharge, and arc discharge caused by inter-turn short circuit.

[0072] Each sample consists of a feature vector x i and the corresponding actual fault label y i Composed of {1,2,3,4};

[0073] Step 3.2, Multi-classification strategy:

[0074] Multiple binary classification SVM sub-models are constructed using a one-vs-one strategy. One classifier is trained for each pair of fault types, and the final category is determined by a voting mechanism.

[0075] Step 3.3, Model Optimization and Kernel Function Selection:

[0076] The objective function is:

[0077] The constraint is: y i (w·φ(x i )+b)≥1-ζ i ,ζ i ≥0;

[0078] The radial basis function (RBF) is used to handle nonlinear separability, expressed as: K(x i ,x j )=exp(-γ||x i -x j || 2 );

[0079] In the formula, min represents minimization, indicating that the goal of this optimization problem is to find the parameter combination that minimizes the objective function; w is the weight vector, representing the normal vector of the hyperplane, which determines the direction of the classification boundary; M is the total number of support vectors; b is the bias term, which determines the position of the classification hyperplane in space, i.e., the intercept; ζ i y is a slack variable, allowing sample points to slightly violate the classification boundary, used to handle non-linearly separable or noisy data; C is a penalty parameter, controlling the degree of penalty for misclassification; i x is the true label of the i-th sample, which is usually +1 or -1 in binary classification to represent the category to which the sample belongs; i Let φ(x) be the feature vector of the i-th sample. i ) is the feature mapping function, which maps the samples x in the original input space. i Mapping to a high-dimensional feature space makes problems that were originally linearly inseparable become linearly separable; w·φ(x i The dot product operation () represents the dot product of the weight vector and the mapped feature vector, used to calculate the projection of a sample onto the decision hyperplane; y i (w·φ(x i K(x) + b) is the functional margin, which measures whether a sample is correctly classified and its distance from the classification boundary; i ,x j ) is a kernel function used to calculate the kernel function for two samples x. i and x j The inner product in a high-dimensional feature space; ||x i -x j || 2 γ is the squared Euclidean distance, representing the squared distance between two samples in the original feature space, reflecting their similarity; γ is the kernel function parameter, controlling the width or range of influence of the RBF function;

[0080] Step 3.4, Model Training and Validation:

[0081] The SVM model is trained using the training set; accuracy, recall, and F1 score are evaluated on the independent validation set; and the parameters are saved after the model converges for real-time diagnostics.

[0082] Furthermore, step 4 includes the following steps:

[0083] Step 4.1, Real-time Feature Extraction:

[0084] The multi-parameter data collected online is preprocessed and feature extracted according to the methods in steps one and two to generate a real-time feature vector x. new ;

[0085] Step 4.2, Fault Classification and Judgment:

[0086] xnew Input the pre-trained SVM model and calculate the decision function:

[0087]

[0088] The final fault type is determined based on the voting results of each subclassifier;

[0089] In the formula, f(x) new The output of the decision function represents the new sample x. new The classification result takes a value of +1 or -1, and the positive or negative class is determined by the sign function; x new For the sample to be classified, the currently input diagnostic feature vector is generated from real-time acquired multi-parameter data after preprocessing and feature extraction; x i Let α be the feature vector of the i-th sample; i The Lagrange multipliers, the weight coefficients associated with the i-th training sample, are obtained from the SVM training process; y i K(x) represents the true label of the i-th sample, which is usually +1 or -1 in binary classification to indicate the category to which the sample belongs; i ,x new ) represents the kernel function; b represents the bias term; M represents the total number of support vectors;

[0090] Step 4.3, Diagnostic results output:

[0091] The system automatically outputs information such as fault type, occurrence time, diagnostic confidence level, and key characteristic data; and displays PRPD graphs, waveform curves, and diagnostic conclusions on the human-machine interface; when a high-risk fault is identified, an audible and visual alarm is triggered and a diagnostic report is generated.

[0092] The beneficial effects of this invention are:

[0093] This invention provides a multi-parameter detection and simulation platform for arc discharge in power transformers, integrating fault simulation, multi-parameter sensing, safety protection, and intelligent diagnosis, as well as a fault diagnosis method. Compared with existing technologies, it has the following significant advantages:

[0094] 1. Realistic reproduction of typical arc discharge faults: By designing four modules for replaceable needle plate, air gap, surface and oil paper insulation inter-turn discharge, the discharge process of various typical insulation defects inside the transformer can be accurately simulated, which significantly improves the consistency between laboratory simulation and actual operating conditions, and provides a high-fidelity experimental basis for fault mechanism research and algorithm verification.

[0095] 2. A multi-physics field collaborative sensing system was constructed: integrating an ultra-high frequency / ultrasound integrated probe, a high frequency pulse current sensor, a capacitive voltage divider sensor, and eight sensing units including fiber optic temperature, pressure, and strain sensors, it realizes the synchronous acquisition of multi-dimensional signals such as electromagnetic, acoustic, electrical, thermal, and force signals, comprehensively captures the multi-field coupling response characteristics induced by arc discharge, and breaks through the information limitations of single parameter monitoring.

[0096] 3. The platform structure design and safety have been optimized: a functional partitioned opening layout is adopted, which ensures the mechanical strength and sealing performance of the oil tank while enabling rapid module replacement and efficient sensor integration; an automatic pressure relief valve is equipped to effectively suppress the risk of combustion and explosion caused by high-energy electric arcs and significantly improve the safety of high-pressure tests.

[0097] 4. Improved data quality and spatiotemporal consistency: Using power frequency voltage as the synchronization benchmark and combining it with a threshold triggering mechanism, high-precision time alignment of multi-source signals is achieved; wavelet denoising and digital filtering techniques are used to effectively suppress power frequency interference and random noise, ensuring the reliability of the original data and providing high-quality input for subsequent feature extraction and fusion analysis.

[0098] 5. A multi-parameter collaborative intelligent diagnostic method is proposed: a multi-dimensional feature vector containing partial discharge, electromagnetic pulse and macroscopic physical quantities is constructed to fully mine the discriminative information in heterogeneous data; a classification model is established by combining support vector machine with RBF kernel function, which overcomes the strong dependence on single parameter and susceptibility to interference of traditional methods, and significantly improves the accuracy and robustness of fault identification.

[0099] 6. Accurate classification and early warning of fault types: The proposed method can effectively distinguish various typical faults such as floating potential discharge, winding insulation defects, oil bubble discharge, and inter-turn short circuit. It still has good identification ability in the weak discharge stage and has strong early warning potential, which helps maintenance personnel to take timely intervention measures to avoid equipment damage and unplanned power outages.

[0100] 7. Promoted the intelligent upgrade of transformer condition-based maintenance: The platform supports automatic data collection, feature extraction, model reasoning and result output, and can link audible and visual alarms and diagnostic report generation, realizing the transformation from "passive monitoring" to "active diagnosis", providing key technical support for the construction of intelligent substations and power equipment condition assessment systems.

[0101] 8. Possesses excellent scalability and engineering application prospects: The modular architecture supports the addition of new discharge models or sensor access. The platform can be used by scientific research institutions to conduct fault mechanism research, and can also serve as a standard test environment for training and verifying new monitoring algorithms. It has broad promotional value and industrialization prospects.

[0102] In summary, this invention not only solves key technical problems in the prior art such as distorted discharge simulation, asynchronous acquisition of multiple parameters, and low diagnostic accuracy, but also constructs a realistic, safe, and intelligent closed-loop system for transformer fault research and diagnosis, significantly improving the early warning capability and intelligent operation and maintenance level of power transformers, and has important scientific significance and engineering application value. Attached Figure Description

[0103] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0104] Figure 1 This is a schematic diagram of the basic framework of the multi-parameter detection simulation platform for arc discharge in power transformers of the present invention.

[0105] Figure 2 This is a flowchart of the power transformer arc discharge fault diagnosis method of the present invention;

[0106] Figure 3 This is a schematic diagram showing the location of the openings on the side of the transformer body of the present invention;

[0107] Figure 4 This is a schematic diagram showing the location of the opening on the top surface of the transformer body of the present invention;

[0108] 1-Transformer body, 2-Rectangular hoisting hole, 3-Circular pressure relief valve hole, 4-Tempered glass observation window, 5-Sensor mounting hole, 6-Lead wire hole, 7-Metal flange. Detailed Implementation

[0109] The technical solutions of 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.

[0110] To address the significant shortcomings of existing technologies in terms of the realism of arc discharge simulation, the integration of multi-parameter detection, the synchronization of signal acquisition, and the accuracy of fault diagnosis, this embodiment provides a multi-parameter detection simulation platform for arc discharge in power transformers. This platform is a comprehensive experimental device integrating fault simulation, multi-physics sensing, safety protection, and synchronous data acquisition. It aims to realistically reproduce typical arc discharge faults inside oil-immersed power transformers, providing a high-fidelity and repeatable experimental environment for fault mechanism research, monitoring technology verification, and the development of intelligent diagnostic algorithms.

[0111] like Figure 1 As shown, this multi-parameter detection simulation platform for arc discharge of a power transformer includes a transformer body, a replaceable discharge module, a multi-parameter sensing unit, and a signal processing and diagnostic module; its detailed structure is as follows:

[0112] 1. Transformer body:

[0113] The transformer body uses a 35kV / 1000kVA true oil-immersed power transformer as the base body, with external dimensions of 2850mm×1200mm×2800mm. It includes the core, windings, insulating oil and power frequency pressurization system, and has the same electromagnetic characteristics and insulation structure as the actual operating equipment. It can withstand the rated voltage and short-term overvoltage operation.

[0114] like Figure 3 and Figure 4 As shown, the oil tank of transformer body 1 has three types of pre-set functional openings to take into account structural strength, sealing performance and module installation requirements:

[0115] A rectangular hoisting hole 2 with a diameter of φ100mm is opened on the top of the transformer body 1 for quick installation and removal of the discharge module; the transformer body 1 is also provided with a circular pressure relief valve hole 3 with a diameter of φ100mm to install a pressure relief protection device. The pressure relief protection device is an automatic pressure relief valve to prevent the oil tank from overpressure and bursting due to high-energy electric arc, thereby improving the safety of the experiment.

[0116] A φ100mm tempered glass observation window 4 is provided on the side of the transformer body 1 to facilitate visual monitoring of the discharge process; sensor mounting holes 5 are also provided on the top and side of the transformer body 1 for connecting pressure, temperature and other probes.

[0117] A φ50mm lead hole 6 is opened on the side of the transformer body 1, and a metal flange 7 with an O-ring and a flow guide groove is provided on the side of the transformer body 1 for the lead-out and oil sealing of the ultra-high frequency and ultrasonic integrated probe.

[0118] 2. Replaceable discharge module:

[0119] The platform is equipped with four typical discharge models, which can be placed in designated positions inside the oil tank through the top hoisting hole, such as the high-voltage winding gap or insulation support area. When the main transformer is not energized, the discharge model can be energized by applying voltage to the discharge model by an independent voltage regulating table to induce partial discharge.

[0120] The structures of typical discharge models are as follows:

[0121] Needle-plate discharge model: Consists of a pointed electrode and a flat plate electrode, simulating point discharge caused by local electric field concentration.

[0122] Air gap discharge model: A combination of spherical and flat electrodes is used to simulate gap breakdown caused by air bubbles or impurities in oil.

[0123] Surface discharge model: Composed of cylindrical electrodes and flat plate electrodes, it simulates the creepage phenomenon on insulating surfaces.

[0124] Oil-paper insulated inter-turn discharge model: A flat copper strip wrapped with three layers of crepe paper is used as a high-voltage electrode. The distance between the copper strip and the grounding electrode is adjusted to 2mm by epoxy resin pads and fixed with insulating cable ties to realistically simulate the oil gap discharge between winding turns.

[0125] In practice, an independent voltage regulating station was used to apply stepwise voltage to the discharge model with a step size of 1kV. Partial discharge could be stably triggered at about 25kV, and the measured discharge quantity was about 1500pC, which met the requirements of typical defect discharge characteristics.

[0126] 3. Multi-parameter sensing unit:

[0127] To achieve simultaneous sensing of multiple physics fields, the platform integrates eight sensing subsystems:

[0128] Ultra-high frequency / ultrasound integrated probe: installed at the φ50mm flange at the bottom of the oil tank, with a frequency band covering 0.5~3GHz UHF and 10kHz~1MHz AE, used to synchronously capture electromagnetic waves and acoustic signals generated by discharge, and the external shielding flange effectively suppresses external interference;

[0129] High-frequency pulse current sensor: bandwidth > 1MHz, installed in series with the bushing end screen grounding wire, neutral point grounding wire and iron core / clamping grounding wire, respectively used to detect internal winding fault current, neutral point fault component and multi-point grounding circulating current.

[0130] Capacitive voltage divider type voltage sensor: It uses the bushing's own capacitance as the high voltage arm, and a measuring capacitor with an accuracy of 0.1% is connected in series at the end of the screen to form the low voltage arm. The response time is in the nanosecond range, and it is used to obtain high frequency voltage changes.

[0131] A fiber optic sensor system consisting of a fiber optic temperature sensor, a fiber optic pressure sensor, a fiber optic strain sensor, and an impedance sensor:

[0132] Fiber optic temperature sensors are attached to the inner wall of the tank top and four side walls, with a range of -55 to 125°C, to monitor oil temperature distribution in real time. Fiber optic pressure sensors are introduced into the tank via a sealed through-plate to sense dynamic changes in oil pressure. Fiber optic strain sensors are magnetically fixed to the center point of the tank side wall, with a range of ±750 με, to detect mechanical deformation caused by electric arc impact. Impedance sensors are installed on the main grounding wire of the simulated transformer pressurization system to monitor abnormal grounding circuit impedance, aiding in the diagnosis of insulation degradation or poor contact.

[0133] All sensor signals are led out through a unified junction box and transmitted using a hybrid of fiber optic and coaxial cables to minimize electromagnetic interference.

[0134] 4. Signal Processing and Diagnostic Module:

[0135] The signal processing and diagnostic module connects to various sensors and is responsible for data acquisition, preprocessing, feature extraction, and fault classification. The acquisition system uses power frequency voltage as the synchronization reference, with a sampling rate ≥100MS / s and a time resolution down to the nanosecond level. Multi-channel data recording is triggered when ultra-high frequency or ultrasonic signals exceed a preset threshold. All parameters are strictly aligned according to timestamps to ensure accurate correlation of multi-dimensional information for the same discharge event. Data can be remotely monitored, stored, and visualized via a host computer.

[0136] Furthermore, based on the aforementioned multi-parameter detection and simulation platform for arc discharge in power transformers, this embodiment also provides a method for diagnosing arc discharge faults in power transformers. This method integrates multi-dimensional sensor information such as ultra-high frequency, ultrasound, current, voltage, temperature, pressure, and strain, and combines signal preprocessing, multi-feature extraction, and support vector machine (SVM) classification models to achieve accurate identification and early warning of typical arc discharge fault types.

[0137] like Figure 2 As shown, the method for diagnosing arc discharge faults in power transformers includes the following steps:

[0138] Step 1, Synchronous Acquisition and Preprocessing of Multi-parameter Data:

[0139] This step aims to obtain high-quality, spatiotemporally consistent raw data, solving the feature distortion problem caused by signal asynchrony and severe noise interference in existing technologies, and laying the foundation for subsequent accurate feature extraction.

[0140] Step 1.1, Synchronous Acquisition Mechanism:

[0141] Using the power frequency voltage signal as a time reference, all sensors in the multi-parameter sensing unit are triggered to start synchronously.

[0142] When the amplitude of UHF or ultrasound AE signal exceeds the preset threshold, multi-channel data recording is initiated.

[0143] Record the occurrence time, phase angle φ, and raw waveform data of each fatal discharge pulse;

[0144] All electromagnetic, acoustic, electrical, thermal, and mechanical detection parameters are stored with a unified timestamp to maintain spatiotemporal consistency.

[0145] Step 1.2, Signal Denoising and Interference Suppression:

[0146] Differentiated denoising strategies are employed for different types of signals:

[0147] For UHF and ultrasonic signals: wavelet transform is used for noise reduction.

[0148] The wavelet decomposition formula is as follows:

[0149]

[0150] The scaling and translation mother wavelet function is defined as follows:

[0151]

[0152] In the formula, C a,b Ψ represents the wavelet coefficients, indicating the similarity between the signal f(t) and the wavelet basis function at scale a and displacement b, reflecting the signal characteristic intensity at that scale and location; f(t) is the original signal, representing the time-domain signal to be analyzed; a,b Ψ(t) is the scaled and translated wavelet basis function, obtained from the mother wavelet function through scaling and translation transformations, used to match local features in the signal; Ψ(t) is the mother wavelet function, representing a pre-selected basic wavelet function with finite energy, rapid decay, and oscillatory characteristics; a is the scaling parameter, controlling the scaling degree of the wavelet function, a>0; b is the translation parameter, controlling the position movement of the wavelet function on the time axis, used for sliding analysis of local features of the signal at different time periods; The normalization coefficients ensure the consistency of wavelet function energy at different scales, making the transformation results comparable. It is the conjugate of the wavelet basis functions.

[0153] By selecting appropriate wavelet bases such as db4 and decomposition levels, setting soft thresholds to process the wavelet coefficients, and then reconstructing them, weak discharge signals can be effectively separated from background noise.

[0154] For high-frequency pulse current and voltage signals: a 0.1–10MHz bandpass digital filter is used to filter out 50 / 60Hz power frequency interference and other low-frequency interference, while retaining the transient components in the fault current.

[0155] Step 1.3, Data Alignment and Normalization:

[0156] Data from all channels is precisely aligned by timestamp, and multi-source signals corresponding to the same discharge event are strictly matched in time; physical quantities of different dimensions, such as temperature, pressure, and current, are normalized to eliminate differences in magnitude.

[0157] This step overcomes the time synchronization problem caused by independent acquisition by multiple sensors in traditional monitoring systems, and improves the accuracy of multi-source signal correlation; wavelet transform can adaptively match the non-stationary characteristics of discharge signals, and more effectively preserves transient pulse details compared with traditional filtering methods; the multi-channel synchronous acquisition mechanism ensures the accurate correspondence between electromagnetic, acoustic, thermal, and mechanical responses, and provides reliable data support for fault evolution process inversion.

[0158] Step 2, Multi-dimensional Feature Extraction and Feature Vector Construction:

[0159] This step transforms the complex raw signal into numerical features with classification and discrimination capabilities, fully exploring the intrinsic relationship between partial discharge, electromagnetic pulse, and macroscopic physical response, and overcoming the problem of one-sided information from a single feature dimension.

[0160] Specifically, the preprocessed data is divided into three feature sets, normalized, and then concatenated into a unified feature vector:

[0161] Step 2.1, Partial Discharge Feature Set X PD :

[0162] The following statistical features were extracted from the PRPD maps generated from UHF and ultrasound signals:

[0163] Skewness: This reflects the left-right asymmetry of the spectrum;

[0164] Kuroshi: Describe the sharpness of the spectrum;

[0165] Maximum discharge quantity q max With average discharge quantity q avg ;

[0166] Pulse quantity distribution: Statistical analysis of the positive half-cycle N within one power frequency cycle + With negative half-cycle N - The number of discharges and their ratio N + / N - ;

[0167] Among them, skewness and kurtosis can quantify the asymmetry and sharpness of the spectrum, which can be used to distinguish different discharge modes; maximum discharge quantity and average discharge quantity can reflect the discharge intensity; the number of positive and negative half-cycle pulses and their ratio can characterize the discharge phase distribution characteristics, which can help identify types such as floating potential and internal discharge.

[0168] In the formula, q is the discharge amount of a single discharge pulse; μ is the mean of the discharge amount, μ=E[q] is the mathematical expectation or arithmetic mean of the discharge amounts of all discharge pulses, reflecting the overall discharge level; E[] is the mathematical expectation, which is the average operation over the random variable; σ is the standard deviation. This indicates the degree of dispersion of the discharge quantity around the mean, reflecting the fluctuation of the discharge intensity; (q-μ) 3 This represents the cube of the deviation from the mean, used to measure the asymmetry of the distribution. A positive deviation indicates the presence of more high-intensity discharges; a negative deviation indicates that low-intensity discharges dominate. (q-μ) 4 The deviation from the mean to the fourth power amplifies the effect of extreme values ​​and makes it more sensitive to heavy tails or sharp peaks; q max The maximum discharge quantity among all detected discharge pulses within one power frequency cycle, reflecting the instantaneous discharge energy intensity, is used to determine whether a high-energy partial discharge exists; q avg N is the arithmetic mean of the discharge quantities of all discharge pulses. + N represents the total number of discharge pulses occurring during the positive half-cycle of the voltage from 0 to 180° within one power frequency cycle. - This represents the total number of discharge pulses occurring during the negative half-cycle of the voltage from 180° to 360° within one power frequency cycle.

[0169] Step 2.2, Electromagnetic Pulse Feature Set X EM :

[0170] Based on high-frequency pulse current I HF (t) and voltage V HF The following features are extracted from the (t) signal:

[0171] Frequency domain characteristics: dominant frequency f p =argmax f |F[I HF [t]|, peak amplitude of spectrum A peak , signal bandwidth BW.

[0172] Time-domain characteristics: Pulse rise time t r Pulse duration t d Pulse energy E p =∫I HF (t) 2 dt reflects the discharge energy and propagation characteristics.

[0173] In the formula, f p The dominant frequency refers to the frequency component with the strongest energy in the discharge pulse spectrum, reflecting the oscillation characteristics of the discharge channel; I HF (t) is a high-frequency pulse current signal, acquired by a high-frequency current sensor, reflecting the current change at the moment of discharge; F[] is the Fourier transform of the time-domain signal I. HF (t) is converted to the frequency domain representation F[IHF (t)](f), revealing its frequency composition; A peak BW is the maximum amplitude value in the spectrum; BW is the bandwidth, referring to the amplitude at the main frequency f. p Centered on the peak value, the spectral amplitude drops to its maximum. The frequency range corresponding to time; t r Rise time refers to the time required for a pulse to rise from 10% to 90% of its amplitude; t d Duration refers to the time span from the start to the end of a pulse, often defined as the time interval from when the pulse first exceeds the threshold to when it falls back below the threshold; E p The pulse energy represents the total energy of the discharge pulse in the time domain, and is related to I. 2 R loss is proportional to its thermal effect and destructive potential; I HF (t) 2 This refers to instantaneous power.

[0174] Step 2.3, Macroscopic physical quantity characteristic set X Phys :

[0175] Extracting dynamic change features from slowly varying signals acquired by fiber optic temperature, pressure, and strain sensors:

[0176] Rate of change: temperature change rate dT / dt, pressure change rate dP / dt; rapid pressure increase is an important indicator of high-energy electric arcs.

[0177] Extreme value characteristics: highest temperature T max The highest pressure P max Maximum strain mutation Δε max This reflects the degree of energy accumulation during a fault.

[0178] In the formula, T represents temperature, the temperature value of the transformer oil or local structure obtained in real time by a fiber optic temperature sensor; t represents time, the time variable for signal acquisition; dT / dt represents the temperature change rate, indicating the rate of temperature change per unit time, reflecting the development rate of local overheating; P represents pressure, the internal oil pressure of the transformer measured by a fiber optic pressure sensor; dP / dt represents the pressure change rate, indicating the rate of pressure increase within the tank per unit time; T max The highest temperature, P, represents the maximum value among all temperature sampling points within a monitoring period. max ε represents the highest oil pressure measured within a specified time period; ε is the strain, the mechanical deformation of the transformer tank or local structure measured by a fiber optic strain sensor; Δε is the strain change, representing the strain increment relative to the initial state or background value at a given moment; Δε max The maximum strain mutation represents the peak mutation value among all strain changes during a single discharge event.

[0179] Step 2.4, Feature Vector Construction:

[0180] After normalizing the above three types of features, they are concatenated into a unified multidimensional feature vector: x = [X PD ,X EM ,X Phys This serves as the input for subsequent classification models.

[0181] This step achieves a leap from "waveform perception" to "feature understanding," avoiding the dimensionality curse caused by directly using the original waveform; multi-dimensional feature fusion compensates for the defects of single parameters being susceptible to interference and fuzzy criteria, such as using only UHF amplitude; the introduction of slowly changing features such as the rate of change of macroscopic physical quantities enhances the ability to distinguish between high-energy arcs and gradual degradation, and improves diagnostic robustness.

[0182] Step 3, Support Vector Machine (SVM) model construction and training:

[0183] This step establishes a high-precision, highly generalizable intelligent classification model to solve the problem that traditional threshold methods or simple rule-based judgments cannot accurately distinguish similar fault types.

[0184] Step 3.1, Dataset Preparation:

[0185] Construct a sample library containing various typical fault types, including:

[0186] Floating potential discharge;

[0187] Discharge due to winding insulation defects;

[0188] Oil bubble discharge;

[0189] Arc discharge caused by inter-turn short circuit.

[0190] Each sample consists of a feature vector x i and the corresponding actual fault label y i It consists of {1,2,3,4}.

[0191] Step 3.2, Multi-classification strategy:

[0192] A one-vs-one strategy is adopted for joint training. The binary classification SVM sub-models are used to determine the category through a voting mechanism.

[0193] Step 3.3, Model Optimization and Kernel Function Selection:

[0194] The radial basis function (RBF) is used to handle nonlinear separability, expressed as:

[0195] K(x i ,x j )=exp(-γ||x i-x j || 2 )

[0196] It can effectively handle nonlinear separable problems, mapping low-dimensional features to high-dimensional space and improving the flexibility of classification boundaries.

[0197] The objective function is:

[0198] The constraint is: y i (w·φ(x i )+b)≥1-ζ i ,ζ i ≥0.

[0199] In the formula, min represents minimization, indicating that the goal of this optimization problem is to find the parameter combination that minimizes the objective function; w is the weight vector, representing the normal vector of the hyperplane, which determines the direction of the classification boundary; M is the total number of support vectors; b is the bias term, which determines the position of the classification hyperplane in space, i.e., the intercept; ζ i y is a slack variable, allowing sample points to slightly violate the classification boundary, used to handle non-linearly separable or noisy data; C is a penalty parameter, controlling the degree of penalty for misclassification; i x is the true label of the i-th sample, which is usually +1 or -1 in binary classification to represent the category to which the sample belongs; i Let φ(x) be the feature vector of the i-th sample. i ) is the feature mapping function, which maps the samples x in the original input space. i Mapping to a high-dimensional feature space makes problems that were originally linearly inseparable become linearly separable; w·φ(x i The dot product operation () represents the dot product of the weight vector and the mapped feature vector, used to calculate the projection of a sample onto the decision hyperplane; y i (w·φ(x i K(x) + b) is the functional margin, which measures whether a sample is correctly classified and its distance from the classification boundary; i ,x j ) is a kernel function used to calculate the kernel function for two samples x. i and x j The inner product in a high-dimensional feature space; ||x i -x j || 2 γ is the squared Euclidean distance, representing the squared distance between two samples in the original feature space, reflecting their similarity; γ is the kernel function parameter, controlling the width or range of influence of the RBF function.

[0200] Step 3.4, Model Training and Validation:

[0201] The model training process is as follows: the SVM model is learned using the training set; the optimal penalty parameter C and kernel function parameter γ are determined through cross-validation to ensure that the model is neither overfitting nor underfitting.

[0202] The validation and evaluation process involves testing metrics such as accuracy, recall, and F1 score on an independent validation set. Once the model performance is confirmed to be stable, the parameters are saved for real-time diagnosis.

[0203] Compared to traditional methods such as BP neural networks, SVM has stronger small-sample learning capabilities, making it suitable for laboratory scenarios with limited data. The RBF kernel function can capture the nonlinear relationship between features, which is significantly better than linear methods such as linear discriminant analysis (LDA). The "one-to-one" strategy effectively improves the classification accuracy of multiple fault types, with an average recognition accuracy of over 96%, which is much higher than the traditional method of about 70%.

[0204] Step 4, Real-time Fault Diagnosis and Result Output:

[0205] This step achieves a closed-loop response from data acquisition to diagnostic decision-making, promoting the transformation of transformer condition monitoring from "passive alarm" to "proactive diagnosis".

[0206] Step 4.1, Real-time Feature Extraction:

[0207] The multi-parameter data collected online is preprocessed and feature extracted according to the methods in steps one and two to generate a real-time feature vector x. new .

[0208] Step 4.2, Fault Classification and Judgment:

[0209] x new Input the pre-trained SVM model and calculate the decision function:

[0210]

[0211] The final fault type is determined based on the voting results of each subclassifier.

[0212] In the formula, f(x) new The output of the decision function represents the new sample x. new The classification result takes a value of +1 or -1, and the positive or negative class is determined by the sign function; x new For the sample to be classified, the currently input diagnostic feature vector is generated from real-time acquired multi-parameter data after preprocessing and feature extraction; x i Let α be the feature vector of the i-th sample; i The Lagrange multipliers, the weight coefficients associated with the i-th training sample, are obtained from the SVM training process; y iK(x) represents the true label of the i-th sample, which is usually +1 or -1 in binary classification to indicate the category to which the sample belongs; i ,x new ) is the kernel function; b is the bias term; M is the total number of support vectors.

[0213] Step 4.3, Diagnostic results output:

[0214] The system automatically outputs the following information:

[0215] Fault type: such as "inter-turn short circuit";

[0216] The occurrence time is in milliseconds.

[0217] Diagnostic confidence, which is based on classification distance;

[0218] Key characteristic data: such as PD pulse count, pressure rise rate;

[0219] Simultaneously, the PRPD spectrum, waveform curve, and diagnostic conclusion are displayed on the human-machine interface;

[0220] When a high-risk fault is identified, an audible and visual alarm is triggered and a diagnostic report is generated.

[0221] This step automates and intelligently processes the diagnostic process, significantly reducing reliance on expert experience; the output includes confidence levels and key features, enhancing the interpretability of diagnostic results and facilitating decision-making by maintenance personnel; it supports early warning, effectively identifying fault buds even when the discharge level is below 2000pC, outperforming the lag of traditional methods; and the alarm linkage mechanism helps prevent fault escalation and reduces the risk of unplanned power outages.

[0222] In summary, the power transformer arc discharge fault diagnosis method proposed in this invention not only solves the key technical problems of low diagnostic accuracy, susceptibility to interference, and lack of early warning capability in the existing technology, but also realizes a closed loop from "data acquisition" to "intelligent decision-making", which significantly improves the accuracy, reliability and engineering practicality of transformer fault diagnosis and has broad application prospects.

[0223] The above-described embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be included within the protection scope of the present invention.

Claims

1. A multi-parameter detection and simulation platform for arc discharge in power transformers, characterized in that: It includes the transformer body, a replaceable discharge module, a multi-parameter sensing unit, and a signal processing and diagnostic module; The transformer body is the basic carrier of the entire platform. It adopts a true oil-immersed power transformer structure, which includes an iron core, windings, and insulating oil. The transformer body can withstand the rated voltage and short-term overvoltage operation. The replaceable discharge module is inserted into a designated position inside the transformer body oil tank through the lifting hole on the top of the transformer body. The replaceable discharge module can be independently pressurized to stimulate partial discharge without power interruption. The multi-parameter sensing unit integrates an ultra-high frequency ultrasonic probe, a high frequency pulse current sensor, a capacitive voltage divider sensor, and fiber optic temperature, pressure, and strain sensors to construct a multi-dimensional synchronous monitoring network covering electromagnetic, acoustic, electrical, thermal, and mechanical fields, enabling comprehensive perception and data acquisition of transformer arc discharge faults. The signal processing and diagnostic module is connected to the multi-parameter sensing unit, which performs multi-parameter information fusion processing and diagnoses discharge faults.

2. The multi-parameter detection and simulation platform for arc discharge of power transformers according to claim 1, characterized in that: The replaceable discharge module is a needle-plate discharge model, an air gap discharge model, a surface discharge model, or an inter-turn discharge model. The needle-plate discharge model uses pointed electrodes and flat plates to simulate point discharge caused by local electric field concentration; the air gap discharge model uses spherical electrodes and flat plates to simulate gap breakdown caused by air bubbles or impurities in oil; the surface discharge model uses cylindrical electrodes and flat plates to simulate surface creepage; and the inter-turn discharge model uses flat copper strips wrapped in crepe paper and epoxy resin pads to simulate oil gap discharge caused by the deterioration of inter-turn insulation in windings. Each discharge model is connected to an independent voltage regulating station, and a stepped pressurization method is used to stably trigger partial discharge.

3. The multi-parameter detection and simulation platform for arc discharge of power transformers according to claim 1, characterized in that: The multi-parameter sensing unit includes an ultra-high frequency ultrasonic integrated probe, a high frequency pulse current sensor, a capacitive voltage divider sensor, and a fiber optic sensor system. The ultra-high frequency ultrasonic integrated probe is installed at the side opening of the transformer tank via a flange; the high frequency pulse current sensor is installed in series on the bushing end screen grounding wire, neutral point grounding wire, and core / clamping grounding wire; the capacitive voltage divider voltage sensor is installed on the bushing end screen grounding wire, which uses the bushing's own capacitance as the high voltage arm, and the end screen end is connected in series with a precision measuring capacitor to form the low voltage arm; the fiber optic sensor system includes fiber optic temperature sensors attached to the top and four inner walls of the transformer tank, fiber optic pressure sensors introduced into the transformer tank through a sealed through-plate, and fiber optic strain sensors magnetically fixed to the center point of the side wall of the transformer tank.

4. The multi-parameter detection and simulation platform for arc discharge of power transformers according to claim 1, characterized in that: The multi-parameter sensing unit also includes an impedance sensor and a pressure relief protection device; The impedance sensor is installed on the grounding wire of the transformer body's pressurization system; the pressure relief protection device is a pressure relief valve, which is installed at the circular pressure relief valve hole on the top of the transformer body.

5. A method for diagnosing arc discharge faults in power transformers, wherein the method is based on the multi-parameter detection and simulation platform for arc discharge in power transformers as described in any one of claims 1-4, characterized in that: The method for diagnosing arc discharge faults in power transformers includes the following steps: Step 1, Synchronous acquisition and preprocessing of multi-parameter data: Multi-parameter signals are synchronously acquired by each sensor of the multi-parameter sensing unit, and wavelet transform and digital filtering are used for noise reduction and alignment to complete the preprocessing of the discharge data. Step 2, Multi-dimensional feature extraction and feature vector construction: Extract multi-dimensional features from partial discharge, electromagnetic pulse and macroscopic physical response, and construct a normalized comprehensive feature vector; Step 3, Support Vector Machine (SVM) model construction and training: Using Support Vector Machine (SVM) combined with RBF kernel function and One-vs-One strategy, the model is trained and the parameters are optimized on typical fault samples; Step 4, Real-time Fault Diagnosis and Result Output: Input the real-time features into the trained SVM model for classification and reasoning, and output the fault type, confidence level and warning information.

6. The method for diagnosing arc discharge faults in power transformers according to claim 5, characterized in that: Step 1 includes the following steps: Step 1.1, Synchronous Acquisition Mechanism: Using the power frequency voltage signal as a time reference, all sensors in the multi-parameter sensing unit are triggered to start synchronously. When the amplitude of UHF or ultrasound AE signal exceeds the preset threshold, multi-channel data recording is initiated. Record the occurrence time, phase angle φ, and raw waveform data of each fatal discharge pulse; All electromagnetic, acoustic, electrical, thermal, and mechanical detection parameters are stored using a unified timestamp to maintain spatiotemporal consistency. Step 1.2, Signal Denoising and Interference Suppression: For UHF and ultrasonic signals, wavelet transform is used for noise reduction. The wavelet decomposition formula is as follows: The scaling and translation mother wavelet function is defined as follows: In the formula, C a,b Ψ represents the wavelet coefficients, indicating the similarity between the signal f(t) and the wavelet basis function at scale a and displacement b, reflecting the signal characteristic intensity at that scale and location; f(t) is the original signal, representing the time-domain signal to be analyzed; a,b Ψ(t) is the scaled and translated wavelet basis function, obtained from the mother wavelet function through scaling and translation transformations, used to match local features in the signal; Ψ(t) is the mother wavelet function, representing a pre-selected basic wavelet function with finite energy, rapid decay, and oscillatory characteristics; a is the scaling parameter, controlling the scaling degree of the wavelet function, a>0; b is the translation parameter, controlling the position movement of the wavelet function on the time axis, used for sliding analysis of local features of the signal at different time periods; The normalization coefficients ensure the consistency of wavelet function energy at different scales, making the transformation results comparable. The conjugate of the wavelet basis functions; By setting a threshold, the wavelet coefficients are subjected to soft / hard thresholding, and then wavelet reconstruction is performed to separate the discharge signal from the noise. For high-frequency pulse current and voltage signals, a bandpass digital filter is used to filter out power frequency interference and other low-frequency noise; Step 1.3, Data Alignment and Normalization: Data from all channels is precisely aligned by timestamp, and multi-source signals corresponding to the same discharge event are strictly matched in time; physical quantities of different dimensions, such as temperature, pressure, and current, are normalized to eliminate differences in magnitude.

7. The method for diagnosing arc discharge faults in power transformers according to claim 5, characterized in that: Step 2 includes the following steps: Step 2.1, Partial Discharge Feature Set X PD : The following statistical features were extracted from the PRPD maps generated from UHF and ultrasound signals: Skewness: This reflects the left-right asymmetry of the spectrum; Kuroshi: Describe the sharpness of the spectrum; Maximum discharge quantity q max With average discharge quantity q avg ; Pulse quantity distribution: Statistical analysis of the positive half-cycle N within one power frequency cycle + With negative half-cycle N - The number of discharges and their ratio N + / N - ; In the formula, q is the discharge amount of a single discharge pulse; μ is the mean of the discharge amount, μ=E[q] is the mathematical expectation or arithmetic mean of the discharge amounts of all discharge pulses, reflecting the overall discharge level; E[] is the mathematical expectation, which is the average operation over the random variable; σ is the standard deviation. This indicates the degree of dispersion of the discharge quantity around the mean, reflecting the fluctuation of the discharge intensity; (q-μ) 3 This represents the cube of the deviation from the mean, used to measure the asymmetry of the distribution. A positive deviation indicates the presence of more high-intensity discharges; a negative deviation indicates that low-intensity discharges dominate. (q-μ) 4 The deviation from the mean to the fourth power amplifies the effect of extreme values ​​and makes it more sensitive to heavy tails or sharp peaks; q max The maximum discharge quantity among all detected discharge pulses within one power frequency cycle, reflecting the instantaneous discharge energy intensity, is used to determine whether a high-energy partial discharge exists; q avg N is the arithmetic mean of the discharge quantities of all discharge pulses. + N represents the total number of discharge pulses occurring during the positive half-cycle of the voltage from 0 to 180° within one power frequency cycle. - The total number of discharge pulses occurring during the negative half-cycle of voltage (180–360°) within one power frequency cycle; Step 2.2, Electromagnetic Pulse Feature Set X EM : Based on high-frequency pulse current I HF (t) and voltage V HF The following features are extracted from the (t) signal: Frequency domain characteristics: dominant frequency f p =arg max f |F[I HF [t]|, peak amplitude of spectrum A peak Signal bandwidth BW; Time-domain characteristics: Pulse rise time t r Pulse duration t d Pulse energy E p =∫I HF (t) 2 dt; In the formula, f p The dominant frequency refers to the frequency component with the strongest energy in the discharge pulse spectrum, reflecting the oscillation characteristics of the discharge channel; I HF (t) is a high-frequency pulse current signal, acquired by a high-frequency current sensor, reflecting the current change at the moment of discharge; F[] is the Fourier transform of the time-domain signal I. HF (t) is converted to the frequency domain representation F[I HF (t)](f), revealing its frequency composition; A peak BW is the maximum amplitude value in the spectrum; BW is the bandwidth, referring to the amplitude at the main frequency f. p Centered on the peak value, the spectral amplitude drops to its maximum. The frequency range corresponding to time; t r Rise time refers to the time required for a pulse to rise from 10% to 90% of its amplitude; t d Duration refers to the time span from the start to the end of a pulse, often defined as the time interval from when the pulse first exceeds the threshold to when it falls back below the threshold; E p The pulse energy represents the total energy of the discharge pulse in the time domain, and is related to I. 2 R loss is proportional to its thermal effect and destructive potential; I HF (t) 2 Instantaneous power; Step 2.3, Macroscopic physical quantity characteristic set X Phys : Extracting dynamic change features from slowly varying signals acquired by fiber optic temperature, pressure, and strain sensors: Rate of change: Temperature change rate dT / dt; Pressure change rate dP / dt; Extreme value characteristics: highest temperature T max Maximum pressure P max Maximum strain mutation Δε max ; In the formula, T represents temperature, the temperature value of the transformer oil or local structure obtained in real time by a fiber optic temperature sensor; t represents time, the time variable for signal acquisition; dT / dt represents the temperature change rate, indicating the rate of temperature change per unit time, reflecting the development rate of local overheating; P represents pressure, the internal oil pressure of the transformer measured by a fiber optic pressure sensor; dP / dt represents the pressure change rate, indicating the rate of pressure increase within the tank per unit time; T max The highest temperature, P, represents the maximum value among all temperature sampling points within a monitoring period. max ε represents the highest oil pressure measured within a specified time period; ε is the strain, the mechanical deformation of the transformer tank or local structure measured by a fiber optic strain sensor; Δε is the strain change, representing the strain increment relative to the initial state or background value at a given moment; Δε max The maximum strain abrupt change represents the peak abrupt change among all strain changes during a single discharge event. Step 2.4, Feature Vector Construction: After normalizing the above three types of features, they are concatenated into a unified multidimensional feature vector: x = [X PD ,X EM ,X Phys This serves as the input for subsequent classification models.

8. The method for diagnosing arc discharge faults in power transformers according to claim 5, characterized in that: Step 3 includes the following steps: Step 3.1, Dataset Preparation: A sample library containing various typical fault types was constructed, including: floating potential discharge, winding insulation defect discharge, oil bubble discharge, and arc discharge caused by inter-turn short circuit. Each sample consists of a feature vector x i and the corresponding actual fault label y i Composed of {1,2,3,4}; Step 3.2, Multi-classification strategy: Multiple binary classification SVM sub-models are constructed using a one-vs-one strategy. One classifier is trained for each pair of fault types, and the final category is determined by a voting mechanism. Step 3.3, Model Optimization and Kernel Function Selection: The objective function is: The constraint is: y i (w·φ(x i )+b)≥1-ζ i ,ζ i ≥0; The radial basis function (RBF) is used to handle nonlinear separability, expressed as: K(x i ,x j )=exp(-γ||x i -x j || 2 ); In the formula, min represents minimization, indicating that the goal of this optimization problem is to find the parameter combination that minimizes the objective function; w is the weight vector, representing the normal vector of the hyperplane, which determines the direction of the classification boundary; M is the total number of support vectors; b is the bias term, which determines the position of the classification hyperplane in space, i.e., the intercept; ζ i y is a slack variable, allowing sample points to slightly violate the classification boundary, used to handle non-linearly separable or noisy data; C is a penalty parameter, controlling the degree of penalty for misclassification; i x is the true label of the i-th sample, which is usually +1 or -1 in binary classification to represent the category to which the sample belongs; i Let φ(x) be the feature vector of the i-th sample. i ) is the feature mapping function, which maps the samples x in the original input space. i Mapping to a high-dimensional feature space makes problems that were originally linearly inseparable become linearly separable; w·φ(x i The dot product operation () represents the dot product of the weight vector and the mapped feature vector, used to calculate the projection of a sample onto the decision hyperplane; y i (w·φ(x i K(x) + b) is the functional margin, which measures whether a sample is correctly classified and its distance from the classification boundary; i ,x j ) is a kernel function used to calculate the kernel function for two samples x. i and x j The inner product in a high-dimensional feature space; ||x i -x j || 2 γ is the squared Euclidean distance, representing the squared distance between two samples in the original feature space, reflecting their similarity; γ is the kernel function parameter, controlling the width or range of influence of the RBF function; Step 3.4, Model Training and Validation: The SVM model is trained using the training set; accuracy, recall, and F1 score are evaluated on the independent validation set; and the parameters are saved after the model converges for real-time diagnostics.

9. The method for diagnosing arc discharge faults in power transformers according to claim 5, characterized in that: Step 4 includes the following steps: Step 4.1, Real-time Feature Extraction: The multi-parameter data collected online is preprocessed and feature extracted according to the methods in steps one and two to generate a real-time feature vector x. new ; Step 4.2, Fault Classification and Judgment: x new Input the pre-trained SVM model and calculate the decision function: The final fault type is determined based on the voting results of each subclassifier; In the formula, f(x) new The output of the decision function represents the new sample x. new The classification result takes a value of +1 or -1, and the positive or negative class is determined by the sign function; x new For the sample to be classified, the currently input diagnostic feature vector is generated from real-time acquired multi-parameter data after preprocessing and feature extraction; x i Let α be the feature vector of the i-th sample; i The Lagrange multipliers, the weight coefficients associated with the i-th training sample, are obtained from the SVM training process; y i K(x) represents the true label of the i-th sample, which is usually +1 or -1 in binary classification to indicate the category to which the sample belongs; i ,x new ) represents the kernel function; b represents the bias term; M represents the total number of support vectors; Step 4.3, Diagnostic results output: The system automatically outputs information such as fault type, occurrence time, diagnostic confidence level, and key characteristic data; and displays PRPD graphs, waveform curves, and diagnostic conclusions on the human-machine interface; when a high-risk fault is identified, an audible and visual alarm is triggered and a diagnostic report is generated.

Images