A transformer operating state monitoring system

Abstract

The application discloses a transformer operation state monitoring system, and belongs to the field of power equipment state monitoring and automation, and the system comprises a multidimensional physical quantity sensing matrix module, a signal synchronization normalization preprocessing module, a self-adaptive environmental noise cancellation module, a physical mechanism driven state characteristic extraction module, an artificial intelligence deep evaluation module and a multi-criteria decision early warning module. The system acquires electric, acoustic, thermal, chemical and mechanical multidimensional physical parameters, removes environmental interference by using noise cancellation technology, and deeply couples a physical mechanism model and an artificial intelligence algorithm to solve micro-evolution indexes of insulation deterioration. The application aims to solve the problems of single monitoring dimension, poor model self-adaptation and limited early warning sensitivity of the transformer, realizes accurate quantitative evaluation and trend prediction of the health state of the transformer, significantly improves the diagnosis accuracy under complex working conditions, and provides core support for realizing predictive maintenance.

Description

Technical Field

[0001] This invention belongs to the field of power equipment condition monitoring and automation, and specifically relates to a transformer operation condition monitoring system. Background Technology

[0002] In the intelligent evolution of power systems, the operational safety of transformers, as core hub equipment of the power grid, is crucial. Traditional operation and maintenance models are transitioning to predictive maintenance based on real-time monitoring and health assessment. This requires long-term tracking of multiple physical indicators of transformers through highly integrated sensing technologies. The stable operation of power equipment not only affects the reliability of regional power supply but also serves as a fundamental support for building a modern smart grid. Therefore, accurate perception and in-depth analysis of the operating status of key electrical equipment has become a key research focus in the power industry.

[0003] Transformer operation status monitoring systems, as core tools for ensuring equipment safety, typically utilize multiple sensors to collect key parameters such as current, voltage, oil temperature, and dissolved gases in the oil. This technology, combining big data analytics and deep learning algorithms, aims to construct a multi-dimensional condition evaluation system, thereby achieving early warning and online monitoring of faults. The core of this monitoring method lies in capturing potential abnormal fluctuations within the transformer through real-time processing of massive amounts of operational data, providing real-time data support and risk assessment for operation and maintenance decisions.

[0004] Existing technologies still face significant challenges in practical applications. Some monitoring schemes overemphasize the macroscopic impact of external environmental factors on operating indicators, resulting in insufficient deep coupling analysis of the evolution of microscopic faults such as insulation material degradation and partial discharge within transformers. Furthermore, the accuracy of model predictions is highly dependent on the accurate establishment of large-scale environmental influence functions, and their adaptive capability is weak when dealing with cross-interference from multiple physical fields. Other monitoring methods based on acoustic signatures suffer from a single monitoring dimension, limiting their early warning sensitivity when handling early overheating or fluctuations in electrical parameters without significant acoustic variations. Moreover, the acoustic acquisition process is highly susceptible to interference from complex electromagnetic environments and industrial noise, severely impacting the accuracy of feature extraction and classification. Due to the lack of deep fusion of multidimensional parameters and the organic integration of physical mechanisms and data-driven approaches, the existing systems' anti-interference capabilities and diagnostic accuracy under complex operating conditions are insufficient to meet the high-reliability operation and maintenance requirements of modern power grids. Summary of the Invention

[0005] The purpose of this invention is to provide a transformer operation status monitoring system to solve the problems of single monitoring dimensions, weak environmental adaptability, insufficient micro-fault analysis, and poor anti-interference ability in the existing technology.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0007] A transformer operating status monitoring system, comprising:

[0008] The multidimensional physical quantity sensing matrix module is used to acquire the full-dimensional physical parameters of the transformer in real time during operation. The multidimensional physical quantity sensing matrix module includes a high-frequency pulse current sensing unit, an ultrasonic partial discharge sensing unit, a fiber optic grating temperature sensing array, an oil dissolved gas detection unit, and a vibration spectrum acquisition unit.

[0009] The signal synchronization normalization preprocessing module is used to perform clock synchronization alignment and range standardization processing on the heterogeneous data collected by the multidimensional physical quantity sensing matrix module, so as to eliminate the sampling delay deviation between different sensors.

[0010] The adaptive environmental noise cancellation module is used to extract and remove background electromagnetic noise and industrial environmental vibration interference based on the spatial correlation between the reference noise source and the monitoring signal, thereby restoring the true transformer operating characteristic signal.

[0011] The physical mechanism-driven state feature extraction module is used to establish a coupled mathematical description model of the electromagnetic field, fluid thermal field and insulation chemical field inside the transformer, and combined with the transformer operation feature signal output by the adaptive environmental noise cancellation module, calculates the microscopic evolution index characterizing the degradation of the insulation material.

[0012] The artificial intelligence deep evaluation module is used to map and analyze the micro-evolution indicators output by the physical mechanism-driven state feature extraction module with historical operating data to generate a quantitative score of transformer health status and a fault evolution trend prediction result.

[0013] The multi-criteria decision-making and early warning module is used to classify and determine the quantitative scores generated by the artificial intelligence deep evaluation module, and generate multi-level early warning instructions based on the preset risk threshold matrix.

[0014] Furthermore, in the multidimensional physical quantity sensing matrix module, the high-frequency pulse current sensing unit captures nanosecond-level partial discharge pulse signals through a high-frequency current transformer connected to the transformer core grounding wire and neutral point grounding wire; the ultrasonic partial discharge sensing unit adopts a multi-point distributed layout on the outer wall of the transformer tank, and uses a piezoelectric ceramic sensor to collect ultrasonic signals with a frequency range between 20 kHz and 200 kHz; the fiber optic grating temperature sensing array is deployed in the hot spot area inside the transformer winding and key nodes of the oil circuit to realize quasi-distributed temperature measurement; the dissolved gas detection unit in the oil analyzes the volume fraction of hydrogen, methane, ethane, ethylene, acetylene, carbon monoxide, and carbon dioxide in the transformer oil at regular intervals and in quantitative terms through headspace sampling or membrane permeation; and the vibration spectrum acquisition unit uses a triaxial accelerometer to acquire the mechanical vibration waveform of the transformer tank surface.

[0015] As one embodiment of the present invention, the signal synchronization normalization preprocessing module is equipped with a high-precision hardware clock synchronization circuit. By receiving timing signals from the Global Positioning System or the BeiDou Navigation Satellite System, it ensures that the synchronization error of all distributed acquisition units is controlled within 1 microsecond. For the acquired analog voltage or current signals, the module uses an analog-to-digital converter with a resolution of 16 bits or higher for high-speed sampling and uses the least squares method to correct the nonlinearity of the sensor. Finally, it maps all values ​​to a normalized range between 0 and 1, providing a consistent data benchmark for subsequent data fusion.

[0016] As one embodiment of the present invention, the operating logic of the adaptive environmental noise cancellation module is as follows: a reference monitoring point is preset outside the transformer body to capture pure environmental electromagnetic background noise and background mechanical noise of the plant area; the module uses an adaptive filter structure to take the signal of the reference monitoring point as the noise estimation benchmark, and continuously adjusts the weighting coefficient of the filter to minimize the difference between the output signal of the transformer body sensor and the reference noise signal in a statistical sense; for impulse interference, the module adopts a threshold discrimination method based on wavelet decomposition to identify and eliminate external corona discharge interference that is similar to the partial discharge waveform but has different phase characteristics, ensuring that the extracted discharge quantity data is true and reliable.

[0017] Furthermore, the core of the physical mechanism-driven state feature extraction module lies in constructing a multi-field coupling model of transformer insulation aging. This module first uses Maxwell's equations to establish a winding leakage magnetic field distribution model and calculates the power density of local overheating caused by eddy current losses. Then, combining computational fluid dynamics principles, it simulates the convective heat dissipation process of transformer oil and calculates the dynamic temperature distribution of paper insulation materials in real time. This module further combines Arrhenius's law to correlate the temperature gradient with the growth rate of dissolved gases in the oil, and calculates the equivalent value of furfural content characterizing cellulose decomposition and the ratio of carbon dioxide to carbon monoxide. By comparing the measured gas concentration with the predicted value of the mechanism model, this module can identify abnormal gas increments caused by internal partial discharge or local overheating.

[0018] Furthermore, the artificial intelligence deep evaluation module adopts a combined architecture of convolutional neural networks and long short-term memory networks. The convolutional neural network part is used to extract the spatial topological features of vibration spectrum and discharge pulse waveform to identify typical fault modes, such as winding loosening, multi-point grounding of iron core, or air gap discharge in insulation layer. The long short-term memory network part is used to process the time evolution sequence of oil temperature, load current, and gas concentration to capture the slow degradation trend of transformer condition. This module uses feature-level fusion technology to input the solution results of the mechanism model as constraints into the deep neural network to correct the evaluation bias caused by insufficient training samples, and finally outputs a comprehensive health index ranging from 0 to 100.

[0019] Furthermore, the multi-criteria decision-making and early warning module is built on the framework of evidence theory, and performs conflict degree measurement and synthesis processing on monitoring information from five dimensions: electricity, sound, heat, chemistry, and mechanics. When the health index is below 80 points and the rate of increase of acetylene concentration in the oil exceeds the first preset change rate threshold, the system triggers a level 1 general early warning, prompting maintenance personnel to shorten the inspection cycle. When the health index is below 60 points and there is a strong correlation between high-frequency pulse current and ultrasonic signal, the system triggers a level 2 severe early warning and automatically starts the internal discharge location algorithm to calculate the coordinates of the discharge point. When the health index is below 40 points and the vibration spectrum shows a significant fundamental frequency shift and a sudden increase in total hydrocarbon concentration, the system triggers a level 3 emergency early warning, generating a power outage maintenance suggestion and a spare parts allocation list.

[0020] Furthermore, the transformer operation status monitoring system also includes an edge computing gateway, which is deployed in the transformer field area to undertake data aggregation, local storage, and primary logic analysis tasks. The edge computing gateway uploads the processed compressed data packets to the cloud management platform via industrial Ethernet or fiber optic network. The cloud management platform collects the operation data of multiple transformers of the same model, performs group health benchmarking analysis, continuously optimizes the empirical parameters in the physical mechanism-driven state feature extraction module, and distributes the updated model weights to each field monitoring terminal.

[0021] In one embodiment of the present invention, after acquiring the signal, the vibration spectrum acquisition unit converts the time series signal into a frequency distribution spectrum through a fast Fourier transform; the artificial intelligence deep evaluation module monitors the energy proportion of the 100 Hz fundamental frequency and its harmonic components in the frequency distribution spectrum; if the fundamental frequency energy decreases while the high-frequency harmonic energy increases significantly, the system determines this feature as a sign of loose core clamps or winding structure deformation, and includes it in the deduction items of the health assessment.

[0022] Furthermore, the dissolved gas detection unit in the oil has an automatic calibration function. The system automatically connects to standard calibration gas every 240 hours to correct the zero point and sensitivity of the sensor, ensuring that the absolute measurement error under long-term operation is within 5%. The unit also integrates an environmental temperature and humidity compensation algorithm to eliminate the sensor's sensitivity to environmental moisture and ensure the stability of monitoring data under extreme climatic conditions such as extreme cold or high humidity.

[0023] Compared with the prior art, the beneficial technical effects of the present invention are as follows:

[0024] This invention completely changes the traditional monitoring methods' single-dimensional and fragmented information by constructing a multi-dimensional physical quantity sensing matrix covering five physical parameters, including electrical, acoustic, thermal, chemical, and mechanical aspects. Through the complementary cooperation of multiple sensors, the system achieves comprehensive monitoring of transformers from macroscopic operating load to microscopic insulation degradation. In particular, the introduction of high-frequency pulse current and ultrasonic sensing for collaborative monitoring greatly enhances the system's ability to capture early partial discharge behavior inside transformers, and the early warning sensitivity is more than twice that of traditional solutions.

[0025] This invention innovatively employs adaptive environmental noise cancellation technology, which effectively suppresses complex electromagnetic interference and industrial noise around the transformer by utilizing reference source cancellation logic. It solves the technical bottlenecks of existing technologies, such as high false alarm rate and submerged feature signals under complex operating conditions. Through high-precision clock synchronization alignment, it ensures the timeliness consistency of sensor data across regions and multiple nodes, providing a solid underlying data foundation for the deep fusion of multi-dimensional features and effectively improving diagnostic accuracy in environments with strong interference.

[0026] This invention overcomes the shortcomings of purely data-driven models in the power equipment field, such as lack of interpretability and reliance on large-scale labeled samples, by deeply coupling physical mechanism models with artificial intelligence algorithms. The physical mechanism model provides the system with clear fault evolution logic and constraints, enabling the system to understand the microscopic mechanism of aging of internal insulation materials in transformers. Meanwhile, the artificial intelligence model uses powerful pattern recognition capabilities to capture nonlinear fault characteristics. This dual-drive mode significantly improves the accuracy of the system's assessment of transformer health status, enabling the earlier detection of latent fault hazards and providing core technical support for modern power grids to achieve a leap from passive emergency repairs to proactive predictive maintenance.

[0027] This invention effectively solves the problem of logical conflicts between heterogeneous monitoring data by establishing a multi-criteria decision-making and early warning mechanism and an evidence theory synthesis algorithm. The system no longer relies on a single alarm threshold, but makes a comprehensive judgment based on the correlation of evidence from multiple dimensions, which greatly reduces the risk of misjudgment caused by single sensor anomalies or environmental fluctuations. The multi-level early warning system combined with health index scoring provides the operation and maintenance department with an intuitive and scientific basis for decision-making, significantly optimizes the allocation efficiency of operation and maintenance resources, and ensures the stability of core power grid hub equipment. Attached Figure Description

[0028] Figure 1 This is a schematic diagram of the overall technical architecture of the transformer operation status monitoring system proposed in this invention;

[0029] Figure 2 This is a schematic diagram of the core principle framework of the coupling between physical mechanism-driven and artificial intelligence-based deep evaluation in this invention. Detailed Implementation

[0030] 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.

[0031] Example 1

[0032] This embodiment discloses a transformer operation status monitoring system, which constructs a closed-loop system across the entire chain, from bottom-level physical sensing to top-level decision-making and early warning, in its overall logical architecture. The system deeply integrates multi-dimensional sensing technology, signal processing algorithms, physical mechanism models, and deep learning networks, aiming to achieve accurate monitoring and trend prediction of the operating status of power transformers. The entire system consists of multiple core functional modules, which achieve seamless information flow and logical interaction through a high-speed data bus and standardized communication protocols.

[0033] The multidimensional physical quantity sensing matrix module, acting as the sensory nerve endings of the entire system, is deployed in key parts of the transformer body and its auxiliary equipment to acquire real-time physical parameters of the transformer in all dimensions during operation. This module is not a simple stacking of single sensors, but rather a collaborative sensing network constructing five dimensions—electrical, acoustic, thermal, chemical, and mechanical—to address the different physical manifestations of transformer fault evolution.

[0034] In the electrical signal dimension, the multi-dimensional physical quantity sensing matrix module includes a high-frequency pulse current sensing unit. This unit captures nanosecond-level partial discharge pulse signals through a high-frequency current transformer connected to the transformer core grounding wire and neutral point grounding wire. The high-frequency current transformer uses a high-permeability nanocrystalline material as its core, and its sampling frequency bandwidth covers the range of 3 MHz to 30 MHz, enabling it to sensitively sense the pulse current excited by weak discharges inside the insulation. When a partial discharge occurs inside the transformer, the generated electromagnetic waves propagate along the windings and are induced to the grounding wire. The high-frequency pulse current sensing unit records the original waveform at a sampling rate of no less than 200 MHz, acquiring key parameters such as the pulse amplitude, phase, discharge frequency, and pulse polarity, providing original evidence for subsequent identification of insulation defect types.

[0035] In the acoustic dimension, the ultrasonic partial discharge sensing unit is arranged in a multi-point distributed layout on the outer wall of the transformer tank. This unit utilizes piezoelectric ceramic sensors to collect ultrasonic signals in the frequency range of 20 kHz to 200 kHz. Due to the propagation characteristics of sound waves in transformer oil and steel plates, the sensors are tightly attached to the tank surface using a high-viscosity acoustic coupling agent. To eliminate positioning errors caused by sound wave diffraction and refraction, the ultrasonic partial discharge sensing unit is configured with at least four sensor nodes, forming a spatial sensor array. Each node integrates a preamplifier circuit and a bandpass filter circuit, effectively suppressing low-frequency mechanical noise and high-frequency radio interference. By recording the time difference of sound waves arriving at different sensors, the system possesses the physical basis for three-dimensional spatial positioning of discharge points inside the transformer.

[0036] In the temperature dimension, fiber Bragg grating temperature sensor arrays are deployed in hot spots and key nodes of the oil circuit inside the transformer windings to achieve quasi-distributed temperature measurement. Compared to traditional resistance thermometers, fiber Bragg grating sensor arrays have inherent electromagnetic insulation properties and can be directly embedded between winding layers in high-voltage, strong electromagnetic field environments. This array modulates the wavelength by etching multiple Bragg gratings with different center wavelengths onto a single optical fiber and utilizing the change in grating period caused by temperature changes at the measurement point. The fiber Bragg grating temperature sensor array achieves a temperature measurement accuracy of 0.1℃ and a sampling spatial resolution of 10 cm, accurately reproducing the three-dimensional thermal field distribution inside the transformer, particularly capturing the dynamic temperature rise of easily overheated areas such as winding ends and lead joints.

[0037] In the temperature dimension, the fiber Bragg grating temperature sensing array is deployed and operated according to the following steps:

[0038] (1) Sensor array deployment: During the transformer manufacturing or maintenance phase, temperature measurement points are pre-determined in hot spots and key nodes of the oil circuit inside the winding. The hot spots include the winding ends, lead joints, near clamps, and the surface of the core column; the key nodes of the oil circuit include the oil inlet / outlet and the oil flow guide plate. Multiple Bragg gratings with different center wavelengths are etched on a single optical fiber. The gratings are arranged at equal or non-equal intervals along the optical fiber axis, achieving a spatial resolution of 10 cm. The optical fiber with the etched grating array is directly buried between the winding layers and at the oil circuit nodes. Utilizing the electromagnetic insulation properties of the optical fiber itself, it can operate stably in high-voltage and strong electromagnetic field environments without additional insulation treatment.

[0039] (2) Wavelength demodulation and temperature conversion: The completed fiber Bragg grating temperature sensing array is connected to the fiber Bragg grating demodulator. The demodulator emits broadband light pulses into the optical fiber, and each Bragg grating reflects only narrowband light matching its center wavelength. When the temperature at the measurement point changes, the grating period changes due to thermal expansion, causing a shift in the center wavelength of the reflected light. The demodulator captures the reflected wavelength of each grating in real time and converts the wavelength shift into a temperature value according to the pre-calibrated wavelength-temperature correspondence. The temperature measurement accuracy of the fiber Bragg grating temperature sensing array reaches 0.1℃.

[0040] (3) Three-dimensional thermal field distribution reconstruction: The system reconstructs the three-dimensional thermal field distribution map inside the transformer based on the spatial location of each grating and its measured real-time temperature value through an interpolation algorithm. Specifically, the coordinates of each temperature measuring point and its temperature data are input into the spatial interpolation model to calculate the temperature distribution in the area where no temperature measuring points are set, forming a continuous thermal field distribution covering the entire winding and oil circuit. The three-dimensional thermal field distribution can realistically restore the dynamic temperature rise process of easily overheated parts such as the winding ends and lead joints.

[0041] (4) Thermal field data output and fusion: The real-time temperature data and three-dimensional thermal field distribution results of each temperature measurement point are clock aligned and range standardized by the signal synchronization normalization preprocessing module, and then output to the physical mechanism driven state feature extraction module to correct the temperature boundary conditions in the computational fluid dynamics model and compare them with the thermal field distribution predicted by the mechanism model to identify abnormal temperature rise.

[0042] At the chemical level, the dissolved gas detection unit in the oil analyzes the volume fractions of hydrogen, methane, ethane, ethylene, acetylene, carbon monoxide, and carbon dioxide in transformer oil at regular intervals and in quantitative quantities using headspace sampling or membrane permeation. This unit integrates a miniature chromatographic column and a photoacoustic spectroscopy sensor. During the monitoring period, a 50 ml oil sample is drawn from the bottom drain valve of the transformer, and the dissolved gases are separated by a degassing device. After the gas enters the detection chamber, a pressure wave is generated by exciting gas molecules with a specific wavelength of infrared laser. The photoacoustic spectroscopy sensor captures this pressure signal and converts it into an electrical signal. The unit features automatic calibration; the system automatically connects to standard calibration gas every 240 hours to correct the sensor's zero point and sensitivity, ensuring that the absolute measurement error is within 5% over long-term operation. Furthermore, the unit integrates an environmental temperature and humidity compensation algorithm to eliminate the sensor's sensitivity to ambient moisture, ensuring the stability of monitoring data under extreme climatic conditions such as extreme cold or high humidity.

[0043] In the mechanical dimension, the vibration spectrum acquisition unit uses a triaxial accelerometer to acquire the mechanical vibration waveforms on the surface of the transformer tank. During normal operation, the transformer generates vibrations dominated by a 100 Hz fundamental frequency due to the electromagnetic expansion effect of the silicon steel sheets and the electromotive force on the windings. After acquiring the signal, the vibration spectrum acquisition unit converts the time-series signal into a frequency distribution spectrum using a fast Fourier transform. This unit can monitor the frequency distribution of vibration amplitude, velocity, and acceleration, paying particular attention to the energy evolution of high-frequency harmonic components. If the core clamps become loose or the windings deform, the vibration characteristics will shift, manifesting as an abnormal surge in energy within a specific frequency band.

[0044] The signal synchronization and normalization preprocessing module receives all heterogeneous data from the aforementioned sensing matrix. This module is internally equipped with a high-precision hardware clock synchronization circuit, ensuring that the synchronization error of all distributed acquisition units is controlled within 1 microsecond by receiving timing signals from the Global Positioning System (GPS) or the BeiDou Navigation Satellite System. For the acquired analog voltage or current signals, this module uses an analog-to-digital converter with a resolution of 16 bits or higher for high-speed sampling. Due to the significant differences in the dimensions and amplitudes of the outputs from different sensors—for example, partial discharge pulses are millivolt-level signals while winding temperatures range from tens to hundreds of degrees Celsius—this module uses the least squares method to correct the sensor nonlinearity, ultimately mapping all values ​​to a normalized range between 0 and 1. This process is achieved by performing a linear transformation on each dimension of data, i.e., subtracting the historical minimum value of that dimension and dividing by its historical range, thereby providing a consistent data benchmark for subsequent data fusion and eliminating weight imbalances caused by differences in measurement ranges.

[0045] The adaptive environmental noise cancellation module is used to reconstruct the true operating characteristics of the transformer. Pre-set reference monitoring points are located outside the transformer to capture pure environmental electromagnetic background noise and background mechanical noise from the plant area. The core logic of this module is based on an adaptive filter structure, using the signal from the reference monitoring points as a noise estimation benchmark. By continuously adjusting the weighting coefficients of the filter, the difference between the output signal of the transformer's sensors and the reference noise signal is minimized statistically. For impulse interference, the module employs a threshold discrimination method based on wavelet decomposition to decompose the signal into frequency bands of different scales. The system identifies and eliminates external corona discharge interference that is similar to the internal partial discharge waveform but has different phase characteristics. For example, external interference is usually distributed near the voltage peak and has a relatively random phase, while internal discharge follows a specific phase distribution pattern. This discrimination mechanism ensures that the extracted discharge data is accurate and reliable.

[0046] The physical mechanism-driven state feature extraction module is the core computational support of the system. This module aims to construct a multi-field coupled model of transformer insulation aging, rather than relying solely on surface features. First, the module establishes a winding leakage magnetic field distribution model using Maxwell's equations. By inputting real-time acquired load current parameters, it calculates the magnetic induction intensity distribution within the windings and structural components, and then calculates the power density of local overheating caused by eddy current losses. Subsequently, the module combines computational fluid dynamics principles to simulate the convective heat dissipation process of transformer oil. During the calculation, the nonlinear relationship between the oil's viscosity and density and temperature is considered, and the dynamic temperature distribution of the paper insulation material is calculated in real time.

[0047] Furthermore, the physical mechanism-driven state feature extraction module, combined with Arrhenius's law, establishes a quantitative relationship between temperature and chemical degradation. Arrhenius's law describes the exponential growth of chemical reaction rates with increasing temperature. This module correlates the calculated temperature gradient with the growth rate of dissolved gases in the oil, calculating the equivalent furfural content characterizing cellulose decomposition and the carbon dioxide to carbon monoxide ratio. By comparing the measured gas concentrations with the predicted values ​​from the mechanism model, this module can identify abnormal gas increments caused by internal partial discharge or local overheating. For example, when the measured acetylene increase rate is much higher than the theoretical increase simulated based on the current load and oil temperature, the system identifies this difference as abnormal gas production caused by an internal latent fault.

[0048] The AI-driven deep evaluation module receives micro-evolutionary indices from the physical mechanism-driven state feature extraction module. This module employs a combined architecture of convolutional neural networks (CNNs) and long short-term memory (LSM) networks. The CNN portion processes high-dimensional topological features, such as converting vibration spectra and discharge pulse waveforms into grayscale images or topological matrices, and extracting typical fault mode features through convolutional kernels. These modes include, but are not limited to, characteristic harmonic clusters generated by winding loosening, low-frequency circulating current features caused by multi-point grounding of the core, and specific phase clustering features generated by air gap discharge within the insulation layer. The LSM network portion is specifically designed to process temporal evolution sequences. Transformer aging is a slow process with a memory effect; historical changes in oil temperature, load current, and gas concentration strongly constrain the current state. This network captures the long-term degradation trend of the transformer state through a logical combination of forget gates, input gates, and output gates.

[0049] The AI ​​deep assessment module uses feature-level fusion technology to input the solution results of the mechanistic model as constraints into the deep neural network. This approach corrects assessment bias caused by insufficient training samples, ensuring that the model can still output judgments consistent with physical logic under unseen extreme conditions. The module ultimately outputs a comprehensive health index ranging from 0 to 100. The calculation of this index follows the logic below:

[0050]

[0051] In the above formula, A comprehensive health index representing the transformer; This represents the total number of physical dimensions involved in the evaluation, which is 5 in this embodiment; Representing the The weighting coefficients for each dimension are dynamically adjusted based on the transformer's operating years, voltage level, and historical fault frequency. Indicates the first Sensor measurement feature values ​​in multiple dimensions, This represents the theoretically expected value corresponding to this dimension, calculated based on a physical mechanism model; function The degree of dispersion between measured and theoretical values, as well as the nonlinear attenuation law of each characteristic component, are defined. Using this formula, the system can converge heterogeneous monitoring signals into a unified quantization score space.

[0052] The specific steps to converge heterogeneous monitoring signals into a unified quantization score space are as follows:

[0053] (1) Establishing Dimensional Correspondence: This module first establishes the mapping relationship between five physical dimensions and monitoring signals. Specifically, the electrical dimension corresponds to the discharge pulse amplitude, discharge frequency, and phase distribution characteristics output by the high-frequency pulse current sensing unit; the acoustic dimension corresponds to the ultrasonic signal energy and arrival time difference output by the ultrasonic partial discharge sensing unit; the thermal dimension corresponds to the temperature values ​​of each temperature measuring point and the three-dimensional thermal field distribution characteristics output by the fiber optic grating temperature sensing array; the chemical dimension corresponds to the concentrations and ratios of seven characteristic gases output by the dissolved gas in oil detection unit; and the mechanical dimension corresponds to the 100 Hz fundamental frequency energy and the proportion of high-frequency harmonic energy output by the vibration spectrum acquisition unit. The original signals of the above five dimensions have different dimensions and vastly different amplitude ranges, making it impossible to directly make a comprehensive evaluation.

[0054] (2) Normalization and Feature Extraction: By normalizing the original monitoring signals of each dimension, the measured values ​​of each physical quantity are uniformly mapped to the standardized range of 0 to 1, eliminating dimensional differences. Subsequently, sensitive feature values ​​that can characterize the insulation state are extracted for each dimension. The characteristic values ​​include, but are not limited to: the equivalent discharge quantity of the partial discharge pulse, the cumulative energy value of the ultrasonic signal, the temperature difference between the winding hot spot temperature and the top oil temperature, the daily growth rate of acetylene and total hydrocarbons, and the energy ratio of the fundamental frequency to the harmonic frequency in the vibration spectrum.

[0055] (3) Synchronous calculation of theoretical expected values ​​by mechanism model: In sync with the above-mentioned measured feature value extraction, the physical mechanism-driven state feature extraction module calculates the theoretical expected value of each dimension under normal aging conditions based on the current load current, ambient temperature and historical operating data, using Maxwell's equations, computational fluid dynamics model and Arrhenius law. This theoretically expected value represents the baseline health of a transformer under conditions free of latent faults.

[0056] (4) Discreteness calculation and nonlinear mapping: For each dimension, the measured feature values ​​are calculated. Compared with theoretical expected value Input to evaluation function The function first calculates the absolute difference or relative deviation rate between the measured and theoretical values, quantifying the degree of dispersion between them. Then, based on the physical failure characteristics of the transformer insulation material, the function applies a nonlinear decay law to this dispersion: when the dispersion is small, for example, the deviation rate of ratio-type parameters such as gas concentration ≤ 1.5 times, and the absolute deviation of absolute value-type parameters such as temperature ≤ 5℃, the function output increases slowly with the increase of dispersion; when the dispersion exceeds a preset threshold, for example, the deviation rate of ratio-type parameters such as gas concentration > 2 times, and the absolute deviation of absolute value-type parameters such as temperature > 15℃, the function output increases exponentially rapidly to reflect the accelerated stage of insulation degradation.

[0057] (5) Weighted fusion and score convergence: Assign dynamic weight coefficients to each dimension This coefficient is adjusted in real time based on the transformer's operating age, voltage level, and historical fault frequency. Transformers with longer operating ages, higher voltage levels, or higher historical fault frequencies have correspondingly higher weighting coefficients for their chemical and thermal dimensions. Subsequently, the module multiplies the function output values ​​of each dimension by their corresponding weighting coefficients and sums the products of the five dimensions to obtain the comprehensive health index. The index is limited to a range of 0 to 100 points, where 100 points represents a perfectly healthy state and 0 points represents a severely faulty state.

[0058] Through the above five progressive processing steps, the heterogeneous monitoring signals with different dimensions and physical meanings are uniformly mapped to a quantization score space of 0 to 100, realizing the convergence and comprehensive evaluation of multidimensional information.

[0059] The multi-criteria decision-making and early warning module is built on the framework of evidence theory, performing conflict degree measurement and synthesis processing on monitoring information from five dimensions: electrical, acoustic, thermal, chemical, and mechanical. Since different sensors may be affected by local interference or their own malfunctions, the multi-criteria decision-making and early warning module uses the Dempster synthesis rule to calculate the support probability of each dimension's evidence for the fault mode. When the health index is below 80 points and the rate of increase in acetylene concentration in the oil exceeds the first preset change rate threshold, the system triggers a Level 1 general early warning. At this time, the system considers the transformer to have initial overheating or slight discharge signs, prompting maintenance personnel to shorten the inspection cycle.

[0060] When the health index falls below 60 points and is accompanied by a strong correlation between high-frequency pulse current and ultrasonic signals, the system triggers a Level 2 severe warning. In this state, the system automatically activates its internal discharge location algorithm. This algorithm combines the time difference of ultrasonic arrival with the spatial distribution of sound velocity in the oil, calculating the coordinates of the discharge point by solving a system of nonlinear equations. The location results are displayed in a 3D cloud map within the transformer's digital twin model. When the health index falls below 40 points, and the vibration spectrum shows a significant fundamental frequency shift and a sudden increase in total hydrocarbon concentration, the system triggers a Level 3 emergency warning. At this time, the system determines that the transformer faces risks of winding strand breakage, severe short circuit, or large-area insulation breakdown, automatically generates a power outage maintenance suggestion and a spare parts allocation list, and sends a risk pre-control command to the dispatch center.

[0061] This system also includes an edge computing gateway. Deployed at the transformer site, this gateway utilizes a high-performance multi-core processor to handle data aggregation, local storage, and basic logic analysis. The edge computing gateway possesses powerful protocol conversion capabilities, supports communication with various sensor units, and performs feature compression on the raw sampled data. Through industrial Ethernet or fiber optic networks, the edge computing gateway uploads the processed compressed data packets to the cloud management platform. The cloud management platform collects operating data from multiple transformers of the same model to perform group health benchmarking analysis. Utilizing cloud computing resources, the platform periodically executes complex large-scale finite element simulations, continuously optimizing empirical parameters in the physical mechanism-driven state feature extraction module, such as the polymerization degree decay rate constant of the insulating paper, and distributing updated model weights to each field monitoring terminal, enabling continuous algorithm evolution.

[0062] In specific operational scenarios, if the AI-powered deep evaluation module analyzes the frequency distribution spectrum output by the vibration spectrum acquisition unit and detects an anomaly in the energy proportion of the 100 Hz fundamental frequency and its harmonics, specifically, if the 100 Hz fundamental frequency energy decreases while the energy of high-frequency harmonics such as 200 Hz and 300 Hz increases significantly, the system identifies this characteristic as a sign of loose core clamps or winding structure deformation. At this point, the system retrieves load current data from the same period for correlation analysis. If the harmonic energy increases proportionally to the square of the load current, it is further confirmed as winding deformation under stress, and this is included in the deduction items of the health assessment, lowering the overall health score.

[0063] To ensure the rigor of the entire decision-making system, the multi-criteria decision-making early warning module introduces an evidence synthesis formula. It is assumed that the fault probability allocation obtained from the electrical signal components is as follows: The fault probability allocation obtained from the chemical components is as follows: The logic for calculating the support after synthesis is as follows:

[0064]

[0065] in, To present two pieces of evidence and After synthesis, the proposition The joint basic probability assignment, Present evidence One focal element, Present evidence One focal element, Indicate the evidence for the proposition The basic allocation rate (BPA) Indicate the evidence for the proposition The basic allocation rate (BPA) It represents a specific failure mode. This represents the conflict factor between two pieces of evidence. This value is obtained by summing the products of the probabilities of the two pieces of evidence on mutually exclusive sets. If A value close to 1 indicates a complete conflict between the two pieces of evidence. In this case, the system will initiate an anomaly sensor self-check procedure to eliminate false alarms. Through this evidence synthesis mechanism, the system can effectively reduce the impact of false alarms from a single dimension, improving the accuracy and robustness of early warnings.

[0066] Example 2

[0067] Based on Example 1, this example further optimizes the architecture of the transformer operation status monitoring system to improve its adaptability under special climatic conditions. For extreme environments in high-altitude and frigid regions, an environmental compensation link is introduced between the multi-dimensional physical quantity sensing matrix module and the signal synchronization normalization preprocessing module.

[0068] Considering the thin air at high altitudes, which significantly increases interference from transformer bushing edge discharge, the adaptive environmental noise cancellation module adds a noise reduction strategy based on variational mode decomposition to the original adaptive filtering. This strategy decomposes the acquired high-frequency pulse current signal into several eigenmode functions with independent center frequencies. By analyzing the energy operators of each mode, the system can identify high-frequency interference with quasi-periodic characteristics generated by external corona discharge and separate it from broadband random pulses generated by internal insulation breakdown. This dual noise reduction mechanism improves the signal-to-noise ratio of partial discharge signals by approximately 15 dB in environments above 3000 meters in altitude.

[0069] To address the impact of frigid environments on sensor sensitivity, the dissolved gas detection unit in oil incorporates a temperature-controlled chamber in its hardware structure. This chamber utilizes a semiconductor heating element to stabilize the gas chamber temperature at 40°C, with fluctuations controlled within 0.5°C. This ensures that the photoacoustic spectral sensor maintains consistent spectral response characteristics even when the ambient temperature fluctuates up to 60°C. Simultaneously, the system reads data from an ambient temperature and humidity meter in real time via an edge computing gateway to perform secondary corrections for wavelength shifts in the fiber Bragg grating temperature sensing array. Since the coefficient of thermal expansion of optical fiber exhibits nonlinear characteristics under extremely low temperatures, the system employs a piecewise cubic spline interpolation algorithm to compensate for temperature measurements, ensuring that the winding temperature measurement error remains less than 0.5°C even at -40°C.

[0070] In the AI ​​deep evaluation module, to address the issue of sample scarcity under extreme environments, this embodiment introduces a generative adversarial network (GAN). This network runs on a cloud management platform, generating a large number of evolutionary sequences simulating extreme environments using existing fault samples. These artificially synthesized samples are used to strengthen the training of the Long Short-Term Memory (LSTM) network, enabling it to identify the state baseline of the transformer under special operating conditions such as the cold start-up phase and the rapid load fluctuation phase. The system uses these baselines as dynamic thresholds, rather than employing fixed alarm limits.

[0071] Furthermore, the multi-criteria decision-making and early warning module incorporates geographical environmental influence factors into its evidence theory framework. When the system detects that the outside temperature is below -20°C, it automatically lowers the early warning weight of the oil temperature dimension and simultaneously increases the weight of the vibration spectrum dimension. This is because at low temperatures, the viscosity of transformer oil increases significantly, resulting in stronger damping of the windings. In this state, changes in vibration characteristics are more sensitive to structural defects than temperature characteristics. Through this dynamic weight adjustment, the system achieves reliable operation under all-weather conditions.

[0072] At the communication security level, an encrypted channel based on domestically developed cryptographic algorithms has been established between the edge computing gateway and the cloud management platform. All uploaded data packets undergo digital signature processing to ensure that monitoring data is not tampered with or hijacked during transmission. The cloud platform not only performs group health analysis but also maintains an expert database containing over 300 typical transformer fault spectra. When the AI ​​deep evaluation module identifies an unknown feature pattern, the system automatically initiates an expert database matching task. If the matching degree is lower than a set value, the feature is marked and pushed to a remote manual review interface for calibration by senior experts. The calibration results are then fed back to the deep learning network, completing a knowledge loop update.

[0073] The monitoring system described in Example 2 significantly improves its survivability and diagnostic accuracy under complex geographical and climatic conditions through dynamic perception of the sensing environment and adaptive algorithm adjustment, making the transformer operation status monitoring system truly have the potential for large-scale industrial applications.

[0074] Example 3

[0075] This embodiment focuses on optimizing computational efficiency and managing data lifecycle in a transformer operation status monitoring system when dealing with large-scale distributed deployments. When the monitoring system is applied to a regional power grid containing hundreds of transformers, the explosive growth in data volume poses extremely high challenges to the edge computing gateway and cloud management platform.

[0076] To address this issue, the signal synchronization normalization preprocessing module introduces an event-driven sampling mechanism. This mechanism no longer records data across all dimensions at a fixed high-frequency sampling rate. During the transformer's stable operation period, the system operates in a low-frequency inspection mode, recording only the effective values ​​and statistical characteristics of each physical quantity. When any sensing unit in the multi-dimensional physical quantity sensing matrix module detects a sudden change, such as the amplitude of a high-frequency pulse current exceeding the second preset threshold, or a pulse-like increase in vibration displacement, the system immediately triggers a high-speed sampling mode. It retrospectively records the original waveform data of the 10 cycles prior to the trigger moment and maintains high-speed sampling until the signal returns to normal. This mechanism reduces the storage pressure of massive amounts of data by more than 90% while preserving the key characteristics of the fault moment.

[0077] The physical mechanism-driven state feature extraction module is further simplified in this embodiment. To run complex coupled models in real-time on the edge computing gateway, a reduced-order model technique is introduced. By performing intrinsic orthogonal decomposition on the original Maxwell's equations and the computational fluid dynamics model, the high-dimensional computational task with millions of degrees of freedom is projected into a low-dimensional orthogonal subspace. This reduced-order model shortens the solution time from hours to milliseconds while retaining over 98% of the computational accuracy. This enables the edge computing gateway to output the dynamic loss rate of the transformer's internal insulation material in real-time, on a minute-by-minute basis.

[0078] In Example 3, the AI ​​deep evaluation module employs a federated learning framework. While protecting the data privacy of each substation, each field edge computing gateway only uploads gradient update values ​​of model parameters to the cloud platform, rather than the original monitoring data. The cloud management platform aggregates and averages the model increments from each node to generate the globally optimal fault identification model before distributing it. This approach not only solves the data silo problem but also leverages the operational experience of transformers across the entire network to enhance the intelligence level of individual terminals. For newly connected transformers, the system can utilize transfer learning technology to inherit mature model weights from existing transformers of the same type and load level, shortening the system's online trial period.

[0079] The multi-criteria decision-making early warning module adds a closed-loop feedback verification function to the early warning issuance process. When the system issues a Level 2 or Level 3 early warning, it automatically monitors the transformer protection action signals and circuit breaker status. If the expected subsequent evolution of physical parameters is not observed within a set time after the early warning is issued, the system automatically lowers the confidence score of that sensor combination. This self-learning mechanism enables the system to continuously filter out false alarms caused by sensor aging or special load interference.

[0080] In terms of data storage strategy, the system implements hierarchical and layered storage. Raw waveform data of faults with high scientific research value is stored in a high-speed solid-state storage array in the cloud; long-term operational indicator data is stored using distributed compression technology. Each stored data item carries a precise nanosecond-level timestamp and spatial geographic coordinate label. By establishing a data index based on a time-series database, maintenance personnel can retrieve snapshots of the operational status of all transformers in the network at any point within the past five years within seconds.

[0081] Example 3 successfully resolved the contradictions between real-time performance, communication bandwidth, storage costs, and algorithm evolution speed in large-scale monitoring systems through optimized computing architecture and intelligent management methods. The implementation of this solution marks a leap in transformer condition monitoring from individual device monitoring to intelligent group operation and maintenance, significantly improving the level of full lifecycle management of power grid assets.

[0082] In summary, the transformer operation status monitoring system disclosed in this invention achieves deep insight into the transformer's operating status through a dual-driven architecture of physical mechanisms and artificial intelligence, combined with a multi-dimensional perception matrix and adaptive noise reduction technology. From nanosecond-level pulse capture to interannual-level degradation prediction, from microscopic molecular degradation analysis to macroscopic health index assessment, the system constructs a comprehensive, high-precision, and highly robust monitoring system. Its evidence synthesis mechanism and multi-level early warning logic provide scientific and intuitive decision support for power grid operation and maintenance, effectively preventing the occurrence of sudden transformer failures and ensuring the continuity and security of power supply. The system architecture of this invention has strong scalability and environmental adaptability, and can be widely applied to the field of power transformer monitoring under various voltage levels and operating conditions.

[0083] Each functional unit and module described in this specification can be implemented by hardware circuits, software logic running on a general-purpose processor, or a combination of hardware and software. The parameter values, threshold settings, and specific algorithm flows disclosed in each embodiment are typical examples provided to illustrate the technical principles of the present invention and should not be construed as limiting the scope of the claims. Equivalent substitutions, functional reorganizations, or parameter optimizations made by those skilled in the art to the components without departing from the core concept of the present invention should all be covered within the scope of protection of the present invention. The present invention is not limited to the technical solutions listed in the above embodiments; its core ideas also have significant reference value and application prospects for the online monitoring of other high-voltage electrical equipment. Through continuous technological iteration and data accumulation, this system will lay a solid technical foundation for building a digital and intelligent modern power system, and promote the evolution of power grid operation and maintenance models towards a more efficient, green, and safe direction.

[0084] In practical deployment, this system can be configured differently based on the importance level of the transformers. For large-capacity transformers in core hub substations, a full-dimensional physical sensing matrix can be deployed, and the highest frequency sampling mode can be activated. For distribution transformers in remote areas, a simplified sensor combination and a local early warning mode based on edge computing can be used. This flexibility ensures that the monitoring system achieves a balance between technological advancement and engineering economy. With the continuous advancement of sensing technology and artificial intelligence algorithms, the monitoring system proposed in this invention will further integrate quantum sensors, 5G / 6G communication technology, and deeper causal inference models, thus demonstrating greater vitality in complex and ever-changing power systems. In future power grid construction, this condition monitoring system, which integrates cutting-edge technologies from multiple disciplines, will undoubtedly become a core tool for ensuring the security of energy infrastructure.

[0085] This paper, through a rigorous logical structure and detailed technical specifications, fully demonstrates the innovation and application value of this invention. From the underlying physical interaction to the high-level information decision-making, every part has undergone meticulous engineering consideration. This deeply coupled system architecture not only improves the sensitivity of monitoring but, more importantly, enhances the reliability and interpretability of monitoring results, solving the long-standing problems of false alarms and missed fault reports in the power operation and maintenance field. Through the implementation of this invention, transformer maintenance will move away from the traditional experience-based model and truly enter the era of intelligent operation and maintenance based on real-time status and scientific prediction. This invention is not only a significant improvement to existing transformer monitoring technology but also an important practical application of the theory of full life cycle management of power equipment. In the context of national energy security and digital transformation, the autonomous controllability, precision, and efficiency of this system undoubtedly have significant social and economic benefits. All technical descriptions aim to clearly and accurately convey the technical solution of this invention, providing a solid textual basis for subsequent patent examination and practical engineering implementation.

[0086] Through the detailed description of the above embodiments, the technical solution and positive effects of the present invention have been fully demonstrated. The collaborative work between the various modules constitutes an intelligent system with the capabilities of perception, thinking, decision-making, and evolution. Whether facing transient discharge faults or long-term slow aging, this system can provide timely and accurate monitoring data and expert-level diagnostic suggestions. This comprehensive technical guarantee is an indispensable foundation for the stable operation of modern high-voltage power grids. In the ever-changing technological landscape, the present invention will maintain its core competitive advantage and continue to lead the development direction of transformer online monitoring technology.

[0087] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A transformer operating status monitoring system, characterized in that, include: The multidimensional physical quantity sensing matrix module is used to acquire the full-dimensional physical parameters of the transformer in real time during operation. The signal synchronization normalization preprocessing module is used to perform clock synchronization alignment and range standardization processing on the heterogeneous data collected by the multidimensional physical quantity sensing matrix module, so as to eliminate the sampling delay deviation between different sensors, and use the least squares method to correct the nonlinearity of the sensors, thereby mapping each value to a normalized range between 0 and 1. The adaptive environmental noise cancellation module is used to extract and remove background electromagnetic noise and industrial environmental vibration interference based on the spatial correlation between the reference noise source and the monitoring signal, so as to restore the true transformer operating characteristic signal. The physical mechanism-driven state feature extraction module is used to establish a coupled mathematical description model of the electromagnetic field, fluid thermal field and insulation chemical field inside the transformer, and combined with the transformer operation feature signal output by the adaptive environmental noise cancellation module, calculates the micro-evolution index characterizing the degradation of insulation materials. The artificial intelligence deep evaluation module is used to map and analyze the micro-evolution indicators output by the physical mechanism-driven state feature extraction module with historical operating data to generate a quantitative score of transformer health status and a fault evolution trend prediction result. The multi-criteria decision-making and early warning module is used to classify and determine the quantitative scores generated by the artificial intelligence deep evaluation module, and generate multi-level early warning instructions based on the preset risk threshold matrix.

2. The transformer operation status monitoring system according to claim 1, characterized in that: The multidimensional physical quantity sensing matrix module consists of a high-frequency pulse current sensing unit, an ultrasonic partial discharge sensing unit, a fiber optic grating temperature sensing array, an oil dissolved gas detection unit, and a vibration spectrum acquisition unit. The high-frequency pulse current sensing unit captures nanosecond-level partial discharge pulse signals through a high-frequency current transformer connected to the transformer core grounding wire and neutral point grounding wire. The ultrasonic partial discharge sensing unit is arranged in a multi-point distributed manner on the outer wall of the transformer tank, and uses piezoelectric ceramic sensors to collect ultrasonic signals with a frequency range of 20 kHz to 200 kHz. A fiber optic grating temperature sensor array is deployed in hot spots and key nodes of the oil circuit inside the transformer winding to achieve quasi-distributed temperature measurement. The dissolved gas detection unit in the oil analyzes the volume fraction of hydrogen, methane, ethane, ethylene, acetylene, carbon monoxide, and carbon dioxide in transformer oil at regular intervals and in quantitative terms by headspace sampling or membrane permeation. The vibration spectrum acquisition unit uses a triaxial accelerometer to acquire the mechanical vibration waveform of the transformer tank surface.

3. The transformer operation status monitoring system according to claim 1, characterized in that: The signal synchronization normalization preprocessing module is equipped with a high-precision hardware clock synchronization circuit. By receiving timing signals from the Global Positioning System or the BeiDou Navigation Satellite System, it ensures that the synchronization error of all distributed acquisition units is controlled within 1 microsecond. For the acquired analog voltage or current signals, the signal synchronization normalization preprocessing module uses an analog-to-digital converter with a resolution of 16 bits or higher for high-speed sampling.

4. The transformer operation status monitoring system according to claim 1, characterized in that: The operating logic of the adaptive environmental noise cancellation module is as follows: a reference monitoring point is preset outside the transformer body to capture pure environmental electromagnetic background noise and background mechanical noise of the plant area; through an adaptive filter structure, the signal of the reference monitoring point is used as the noise estimation benchmark, and by continuously adjusting the weighting coefficient of the filter, the difference between the output signal of the transformer body sensor and the reference noise signal is minimized in a statistical sense; for impulse interference, the adaptive environmental noise cancellation module adopts a threshold discrimination method based on wavelet decomposition to identify and eliminate external corona discharge interference that is similar to the partial discharge waveform but has different phase characteristics.

5. The transformer operation status monitoring system according to claim 1, characterized in that: The operating logic of the physical mechanism-driven state feature extraction module is as follows: First, a winding leakage magnetic field distribution model is established using Maxwell's equations to calculate the power density of local overheating caused by eddy current losses; then, combined with the principles of computational fluid dynamics, the convective heat dissipation process of transformer oil is simulated, and the dynamic temperature distribution of paper insulation material is calculated in real time; the physical mechanism-driven state feature extraction module further combines Arrhenius's law to correlate the temperature gradient with the growth rate of dissolved gases in the oil, and calculates the equivalent value of furfural content characterizing cellulose decomposition and the ratio of carbon dioxide to carbon monoxide; by comparing the measured gas concentration with the predicted value of the mechanism model, abnormal gas increments caused by internal partial discharge or local overheating are identified.

6. The transformer operation status monitoring system according to claim 1, characterized in that: The artificial intelligence deep evaluation module adopts a combined architecture of convolutional neural network and long short-term memory network. The convolutional neural network part is used to extract the spatial topological features of vibration spectrum and discharge pulse waveform to identify fault modes such as winding loosening, multi-point grounding of iron core or air gap discharge in insulation layer. The long short-term memory network part is used to process the time evolution sequence of oil temperature, load current and gas concentration to capture the degradation trend of transformer condition. The AI ​​deep assessment module uses feature-level fusion technology to input the solution results of the mechanism model as a constraint into the deep neural network, correcting the assessment bias caused by insufficient training samples, and finally outputting a comprehensive health index ranging from 0 to 100.

7. The transformer operation status monitoring system according to claim 1, characterized in that: The system also includes an edge computing gateway and a cloud management platform; The edge computing gateway is deployed in the transformer field area for data aggregation, local storage and basic logical analysis, and uploads the processed compressed data packets to the cloud management platform via the network. The cloud management platform is used to collect operating data from multiple transformers of the same model to conduct group health benchmarking analysis, optimize the empirical parameters in the physical mechanism-driven state feature extraction module, and distribute the updated model weights to each field monitoring terminal.

8. The transformer operation status monitoring system according to claim 1, characterized in that: The multi-criteria decision-making and early warning module is built on the framework of evidence theory. It measures and synthesizes the conflict degree of monitoring information from five dimensions: electricity, sound, heat, chemistry, and mechanics. When the health index is below 80 points and the rate of increase of acetylene concentration in the oil exceeds the first preset change rate threshold, a level 1 general early warning is triggered. When the health index is below 60 points and there is a strong correlation between high-frequency pulse current and ultrasonic signal, a level 2 severe early warning is triggered, and the internal discharge positioning algorithm is automatically started to calculate the coordinates of the discharge point. When the health index is below 40 points and the vibration spectrum shows a fundamental frequency shift and a sudden increase in total hydrocarbon concentration, a level 3 emergency early warning is triggered, and a power outage maintenance suggestion and spare parts allocation list are generated.

9. The transformer operation status monitoring system according to claim 2, characterized in that: After acquiring the signal, the vibration spectrum acquisition unit converts the time series signal into a frequency distribution spectrum through a fast Fourier transform. The system monitors the energy proportion of the 100 Hz fundamental frequency and its harmonic components in the frequency distribution spectrum. If the fundamental frequency energy decreases while the high-frequency harmonic energy increases, the system determines this as a sign of loose core clamps or winding structure deformation and includes it in the deduction items of the health assessment. The dissolved gas detection unit in the oil has an automatic calibration function. The system automatically connects to standard calibration gas every 240 hours to correct the zero point and sensitivity of the sensor.

10. The transformer operation status monitoring system according to claim 6, characterized in that: The calculation logic of the comprehensive health index is as follows: The sensor measurement feature values ​​of multiple physical dimensions involved in the evaluation are compared with the theoretical expected values ​​corresponding to each dimension calculated based on the physical mechanism model; an evaluation function is defined to reflect the dispersion between the measured values ​​and the theoretical values, as well as the nonlinear attenuation law of each feature component; the evaluation function values ​​of each dimension are weighted and summed with their corresponding weight coefficients to obtain the comprehensive health index; wherein, the weight coefficients are dynamically adjusted according to the transformer's operating years, voltage level, and historical fault frequency. The comprehensive health index calculation model is as follows: in, The overall health index representing the transformer. This represents the total number of physical dimensions involved in the evaluation. Representing the The weighting coefficients for each dimension are dynamically adjusted based on the transformer's operating years, voltage level, and historical fault frequency. Indicates the first Sensor measurement feature values ​​in multiple dimensions This represents the theoretically expected value corresponding to this dimension, calculated based on a physical mechanism model. The degree of dispersion between measured and theoretical values ​​and the nonlinear decay law of each characteristic component are defined.

Images