Multi-mode fault monitoring system and method for dry-type air-core reactor

Abstract

The invention discloses a multi-mode fault monitoring system and method for a dry-type air-core reactor, and relates to the technical field of power equipment monitoring, and the method comprises the steps: constructing a reactor three-dimensional model through laser scanning, and building a coordinate system; ultrahigh-frequency electromagnetic waves, high-frequency pulse current, ultrasonic waves and infrared thermal imaging signals are synchronously collected and stored in association with spatial positions; performing preprocessing and feature extraction on each modal signal, and generating a comprehensive feature vector through data-level fusion; inputting the comprehensive feature vector into a pre-trained CNN-LSTM hybrid diagnosis model, and outputting a fault type and a probability; fault source space positioning is carried out according to the signal arrival time difference and the three-dimensional model; and a health index is calculated by integrating the fault type, the probability and the positioning result, and graded early warning is generated according to an index threshold.

Description

Technical Field

[0001] This invention relates to the field of power equipment monitoring technology, specifically to a multi-mode fault monitoring system and method for dry-type air-core reactors. Background Technology

[0002] Dry-type air-core reactors, as key reactive power compensation devices in power systems, are widely used due to their simple structure, good linearity, and convenient maintenance. However, because their windings are directly exposed to the air, the reliability of their insulation system depends on the performance of the solid insulation material. During long-term operation, the reactor is subjected to the combined effects of electrical, thermal, mechanical, and environmental stresses, leading to a gradual deterioration of its insulation performance.

[0003] Early signs of insulation degradation include partial discharge. Inside reactors, particularly at weak points in winding turns and interlayer insulation, internal air bubbles, resin cracks, or sharp edges of conductors, the electric field can become severely distorted and concentrated. These areas of concentrated electric field can trigger partial discharge. Partial discharge is both a precursor to insulation deterioration and an accelerator of its continued degradation. Although weak partial discharges may not immediately cause equipment failure in the short term, and the equipment can continue to operate, the various physical effects they generate, such as charged particles, ultraviolet light, sound waves, and electromagnetic radiation, will continuously erode the insulating material, causing chemical bond breakage and carbonization of organic polymer materials, and forming conductive channels. This process has a cumulative effect and, in the long run, will significantly accelerate insulation aging, potentially leading to catastrophic failures such as inter-turn short circuits or main insulation breakdown, posing a fatal threat to the safe and stable operation of the power system.

[0004] Currently, monitoring methods for dry-type reactors mostly focus on macroscopic parameters such as temperature and current. While these methods can reflect some serious faults, they struggle to effectively capture and identify early, weak partial discharge signals, failing to provide early warning of insulation degradation. Traditional single-detection methods, such as those relying solely on pulse current or ultrasonic methods, are susceptible to complex electromagnetic interference and mechanical vibrations in the field, resulting in low detection sensitivity, inaccurate positioning, and high false alarm rates. Therefore, there is an urgent need for an intelligent monitoring solution that can deeply integrate multiple feature information, possess strong anti-interference capabilities, and accurately identify and warn of early insulation faults. Summary of the Invention

[0005] The purpose of this invention is to provide a multi-mode fault monitoring system and method for dry-type air-core reactors to solve the problems raised in the prior art.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a multi-mode fault monitoring method for dry-type air-core reactors, the multi-mode fault monitoring method comprising: A three-dimensional structural model of the dry-type air-core reactor is obtained, and a coordinate system is established in the three-dimensional structural model. The ultra-high frequency electromagnetic wave signal, high frequency pulse current signal, ultrasonic signal and infrared thermal imaging signal of the dry-type air-core reactor are collected simultaneously, and the collected signals are associated with the spatial position in the three-dimensional structural model and stored. The collected ultra-high frequency electromagnetic wave signals, high frequency pulse current signals, ultrasonic signals, and infrared thermal imaging signals were preprocessed and feature extracted to obtain feature vectors for each signal mode. The extracted multi-modal feature vectors were then fused at the data level to generate a comprehensive feature vector. The comprehensive feature vector is input into a pre-trained multimodal fault diagnosis model, which outputs fault type identification results and fault occurrence probability; based on the arrival time difference of high-frequency signals and the propagation time of ultrasonic signals, combined with the three-dimensional structural model, the spatial location calculation of the fault source is performed. Based on the fault type identification results, fault occurrence probability, and fault source location signal, the comprehensive health index of the dry-type air-core reactor is calculated; when the comprehensive health index is lower than the threshold or a specific type of fault is identified, a graded early warning message is generated and sent to the monitoring terminal.

[0007] Furthermore, obtaining a three-dimensional structural model of the dry-type air-core reactor, and establishing a coordinate system in the three-dimensional structural model includes: The dry-type air-core reactor is fully scanned using laser scanning technology to obtain its geometric parameters, and a three-dimensional structural model of the dry-type air-core reactor is constructed using modeling software. In the above steps, the geometric parameters include, but are not limited to, the number of winding layers, the inter-turn spacing, and the position of the support structure. In the three-dimensional structural model, a right-hand rectangular coordinate system is established with the intersection of the bottom central axis of the dry-type air-core reactor and the contact surface of the base as the origin. The X-axis of the right-hand rectangular coordinate system extends horizontally outward along the radial direction of the dry-type air-core reactor, the Y-axis is distributed horizontally along the circumference of the dry-type air-core reactor, and the Z-axis is vertically upward along the central axis of the dry-type air-core reactor. Several feature points with known actual locations on the dry-type air-core reactor are selected, and their actual spatial coordinates are measured. These actual spatial coordinates are then compared with the coordinates of corresponding points in the three-dimensional structural model, and the coordinate deviation is calculated. ;in, Indicates coordinate deviation, where A represents the actual spatial coordinates of the feature point, and a represents the coordinates of the corresponding point of the feature point in the three-dimensional structural model; if Then adjust the scaling of the 3D model or the position of the origin of the coordinate system until all feature points are adjusted. ; In the above steps, "feature points," such as the top edge point of the winding, the middle support point, and the bottom terminal point, are measured using a high-precision total station to determine their actual spatial coordinates. The preset deviation threshold should be set to a value that meets the accuracy requirements for partial discharge positioning. The process involves simultaneously acquiring ultra-high frequency electromagnetic wave signals, high frequency pulse current signals, ultrasonic signals, and infrared thermal imaging signals from the dry-type air-core reactor, and associating and storing the acquired signals with their spatial positions in the three-dimensional structural model, including: Several ultra-high frequency (UHF) sensors are deployed around the dry-type air-core reactor. The center position of each UHF sensor is marked with coordinates [coordinates missing] in the 3D model. And record the shortest distance between the ultra-high frequency sensor and the reactor winding; In the above steps, the UHF sensor can be installed in the positive X-axis direction, negative X-axis direction, positive Y-axis direction, negative Y-axis direction, positive Z-axis direction, and negative Z-axis direction, respectively. The minimum distance between the UHF sensor and the reactor winding should be 5-10cm to avoid electromagnetic interference. The coil of the current sensor is fitted onto the incoming cable of the dry-type air-core reactor, and the center position of the coil is marked in the three-dimensional structural model as follows. ; The above steps are to ensure that the coil and cable are coaxial, thereby reducing signal attenuation; Several ultrasonic sensors are deployed along the Z-axis on the outside of the reactor winding. Each ultrasonic sensor is marked with coordinates [coordinates missing] in the three-dimensional structural model. ; In the above steps, the number of ultrasonic sensors and the spacing between their deployments are adjusted according to the height of the dry-type air-core reactor, and the ultrasonic sensor probes must be in close contact with the outer surface of the reactor windings, using a coupling agent (such as a silicon-based coupling agent) to reduce sound wave reflection loss. Install the infrared thermal imager directly in front of the dry air reactor, align the central axis of the lens with the center of the dry air reactor, and mark the spatial domain corresponding to the field of view of the infrared thermal imager in the three-dimensional structural model. In the above steps, the spatial domain must ensure that it covers the entire winding; A multi-channel data acquisition card is used to connect all sensors, and a GPS synchronization module is used to ensure that the sampling timestamps of each sensor are consistent. During the data acquisition process, each sensor associates the acquired signal with the corresponding sampling time and sensor coordinates, and stores it in the database. The databases used in the above steps include, but are not limited to, MySQL and PostgreSQL.

[0008] Furthermore, the acquired ultra-high frequency electromagnetic wave signals, high frequency pulse current signals, ultrasonic signals, and infrared thermal imaging signals were preprocessed and feature extracted to obtain the feature vectors of each signal mode, including: For ultra-high frequency electromagnetic wave signals, a bandpass filter is used to filter low-frequency interference, and wavelet threshold denoising is used to remove high-frequency noise. The waveform of the denoised ultra-high frequency electromagnetic wave signal is as follows: ; Where t is the sampling time; For high-frequency pulse current signals, a notch filter is used to remove power frequency interference, and adaptive filtering is used to cancel leakage current interference. The noise-reduced high-frequency pulse current signal is as follows: ; In the above steps, the "notch filter" can suppress 50Hz and its harmonics; the "adaptive filter" uses the background current signal when the reactor has no partial discharge as the reference input; For ultrasonic signals, a high-pass filter is used to filter low-frequency vibration noise, and a short-time Fourier transform is used to remove random sound waves generated by non-partial discharges. The noise-reduced ultrasonic signal is as follows: ; In the above steps, the "high-pass filter" can preserve frequency bands above 20kHz; For infrared thermal imaging signals, the inter-frame difference method is used to remove the influence of ambient temperature fluctuations, and median filtering is used to remove isolated noise points caused by lens dust. The corrected winding surface temperature distribution matrix is ​​as follows: ; In the above steps, the "inter-frame difference method" is, for example, to calculate the average temperature of three consecutive thermal images as the background temperature of the current frame, and then subtract the background temperature from the current frame temperature. Feature extraction was performed on each of the preprocessed modal signals, including the feature vector of the ultra-high frequency electromagnetic wave signal. High-frequency pulse current signal feature vector Ultrasonic signal feature vector Infrared thermal imaging signal feature vector Each feature vector includes 5 dimensions. Among them, the characteristic vector of ultra-high frequency electromagnetic wave signal ; in, For the pulse peak value, by traversing Take all sampling points in the middle. The maximum value is obtained, reflecting the electromagnetic radiation intensity of partial discharge; The pulse rise time is calculated. The time difference corresponding to the amplitude point is obtained to reflect the discharge development speed; The pulse repetition frequency represents the frequency per unit time. Exceeding the preset threshold The number of pulses; The centroid of the spectrum is calculated using the following formula: ,in, for The Fourier transform result; The pulse width is obtained by calculating the difference between the pulse start time and the pulse end time. Exceed The duration; Among them, the high-frequency pulse current signal feature vector ; in, For the pulse peak value, by traversing Take all sampling points in the middle. The maximum value in the formula is obtained, reflecting the current intensity of partial discharge; q is the charge transfer amount, calculated using the following formula: ;in, , These are the start and end times of the pulse, respectively. Exceeding the preset threshold ; The pulse repetition frequency represents the frequency per unit time. Exceed The number of pulses; The current rise rate is calculated by taking the maximum value of the ratio of the current difference between adjacent sampling points to the time difference. The pulse energy is calculated using the following formula: Where R is the equivalent resistance of the Rogowski coil, a known parameter provided by the coil manufacturer, typically 50 ohms. ; Among them, the ultrasonic signal feature vector ; in, The peak sound pressure level is determined by traversing... Take all sampling points in the middle. The maximum value is obtained, reflecting the acoustic wave intensity of partial discharge; For the frequency of sound waves, by... Power spectrum analysis was performed, and the frequency point with the highest power was taken as the sound wave frequency. The duration of the sound wave is obtained by calculating the time difference between the start and end of the pulse. Exceeding the preset threshold ; The integral of sound pressure is calculated using the following formula: ,in, , These are the start and end times of the pulse; The sound wave attenuation coefficient is calculated using the following formula: ,in, The sound wave attenuation coefficient represents the degree of attenuation of sound waves as they propagate through air. The initial sound pressure level of the partial discharge source (obtained through calibration experiments, such as measurement at a simulated discharge source with a known discharge quantity). The relationship with d is obtained by fitting. d is the distance between the sensor and the winding; Among them, the infrared thermal imaging signal feature vector ; in, For the highest temperature, take the temperature distribution matrix. The maximum value in the value represents the highest temperature, reflecting the degree of local overheating in the winding. The temperature standard deviation reflects the degree of local overheating in the winding, and the calculation formula is: Where N is the total number of pixels in the temperature matrix. Let i be the temperature of the i-th pixel. The average temperature of all pixels; The hotspot area is determined by statistical analysis of the temperature matrix. The number of pixels is combined with the actual area corresponding to each pixel (calculated by the scaling ratio of the 3D construction model, e.g., 1 pixel corresponds to 0.1mm × 0.1mm) to obtain the threshold. The settings are based on the temperature distribution pattern during normal operation of the reactor; The temperature gradient is calculated using the following formula: ,in, This represents the actual distance between adjacent pixels. The coordinates of the hotspot center are obtained by calculating the average of the coordinates of all pixels in the hotspot area; A fault simulation platform for dry-type air-core reactors was built in the laboratory. Several sets of multimodal signals were collected for each fault type. The accuracy of each modal feature vector for fault type identification was calculated, and weights were calculated through normalization. The calculated weights were then used to... , , , The weighted average of each corresponding feature dimension in the vector is used to obtain the comprehensive feature vector V. In the above steps, the formula for calculating the weight is: ,in, Let be the recognition accuracy for the k-th modality; each feature vector has 5 dimensions. Except for the coordinate features, the coordinate features are directly retained in the calculation; the comprehensive feature vector V integrates the advantages of multimodal signals, avoids the limitations of a single signal, and improves the accuracy of subsequent fault diagnosis.

[0009] Furthermore, the comprehensive feature vector is input into a pre-trained multimodal fault diagnosis model, and the output fault type identification result and fault occurrence probability include: Several sets of data were collected in the laboratory. Each set of data contained a comprehensive feature vector V and a corresponding fault type label. The collected sets of data were divided into training set and validation set according to a pre-set ratio. In the above steps, the fault type labels are manually marked, and the fault types include: normal, bubble discharge, inter-turn discharge, and inter-layer discharge. A hybrid model of convolutional neural network and long short-term memory network is adopted. The training set is input into the hybrid model to train the model. The structure of the trained model is adjusted through the validation set to obtain a multimodal fault diagnosis model. The comprehensive feature vector V, which is collected and fused in real time, is input into the multimodal fault diagnosis model to output the fault type and the probability of fault occurrence. In the above steps, Convolutional Neural Networks (CNNs) extract spatial correlations of features, and Long Short-Term Memory Networks (LSTMs) capture time-series features; "Adjusting the trained model structure using a validation set," for example, adjusting the number of convolutional layers in the CNN and the number of hidden layer nodes in the LSTM, to ensure that the model's recognition accuracy on the validation set is greater than or equal to 95%, and the overfit (training set accuracy - validation set accuracy) is less than or equal to 3%; Based on the arrival time difference of the high-frequency signal and the propagation time of the ultrasonic signal, combined with the aforementioned three-dimensional structural model, the spatial location calculation of the fault source includes: Obtaining the propagation speed of ultra-high frequency electromagnetic waves Ultrasonic propagation speed Signal arrival time difference Signal arrival time difference ; in, Approximately the speed of light in air ; The speed of propagation in air is related to temperature, and the calculation formula is: ,in, Ambient temperature (obtained from background temperature captured by an infrared thermal imager, unit: ...) ); UHF sensor With reference sensor The time difference between receiving the same partial discharge signal ,in, for The time of signal reception for The time when the signal was received; For ultrasonic sensors With reference sensor The time difference between receiving the same partial discharge signal ,in, for The time of signal reception for The time of signal reception; the reference sensor is deployed along the Z-axis at preset distance intervals when the ultrasonic sensor is deployed; Let the coordinates of the fault source be... Based on the relationship between distance, velocity, and time, equations for UHF signals and ultrasonic signals are established; the least squares method is used to fit the coordinates of the fault source. If the calculated fault source coordinates deviate from the actual coordinates of the simulated fault source in the laboratory by less than or equal to a preset threshold, the location result is output; otherwise, the number of reference sensors is increased and the calculation is repeated. In the above steps, the "least squares method" is used, for example, through the lsqcurvefit function in MATLAB, to minimize the sum of squared errors between the calculated distance and "speed × time"; the "preset threshold" should meet the requirements for inter-turn fault location in dry-type air-core reactors.

[0010] Furthermore, based on the fault type identification results, fault occurrence probability, and fault source location signal, the comprehensive health index of the dry-type air-core reactor is calculated as follows: ; in, For fault type weights, Score the fault type. Let P be the probability weight of the failure, where P is the probability of failure occurring. The fault source location weight, Score the location of the fault source. As for the weight of temperature anomalies, Score for temperature anomalies; In the above steps, The severity of the fault is set according to its severity, for example: normal (0), bubble discharge (0.2), inter-turn discharge (0.4), inter-layer discharge (0.6) (fault hazard level classification is based on power industry standards); Fault type scoring, for example: normal (0), bubble discharge (20), inter-turn discharge (50), inter-layer discharge (80) (the higher the score, the more severe the fault); The degree of influence of the failure probability on the health status is determined through experimental verification; the failure occurrence probability P is obtained from the output of the multimodal failure diagnosis model. The distance is set based on the distance between the fault source and the core components of the reactor (such as winding turns and layers), for example: distance < 5mm (0.4), 5-10mm (0.2), > 10mm (0.1) (the distance is calculated from the difference between the fault source coordinates and the winding coordinates in the three-dimensional structural model); fault source location score For example: distance < 5mm (30), 5-10mm (15), > 10mm (5) (the closer the distance, the higher the score); This indicates the degree of impact of abnormal temperature on health status; the value is preset. Based on hotspot temperature Calculations, for example: , , , ( from (obtained from) When the comprehensive health index falls below a threshold or a specific type of fault is identified, a graded early warning message is generated and sent to the monitoring terminal, including: Based on the range of values ​​for the comprehensive health index HI, an early warning level is set; The warning information includes: warning level, comprehensive health index HI, fault type, fault occurrence probability P, and fault source coordinates. Hotspot temperature .

[0011] Furthermore, to better implement the above method, a multi-mode fault monitoring system for dry-type air-core reactors is also provided. This multi-mode fault monitoring system includes: a modeling and calibration module, a signal acquisition module, a signal preprocessing module, a feature extraction module, a feature fusion module, a fault diagnosis model module, a fault location module, a health index calculation module, and an early warning generation module. The modeling and calibration module is used to obtain the three-dimensional structural model of the dry air-core reactor, establish a coordinate system in the three-dimensional structural model, and ensure that the coordinates of the three-dimensional structural model are consistent with the actual spatial coordinates through feature point calibration, so as to provide a spatial reference for signal correlation and fault location. The signal acquisition module is used to simultaneously acquire ultra-high frequency electromagnetic wave signals, high frequency pulse current signals, ultrasonic signals and infrared thermal imaging signals of the dry air reactor, and to associate and store the acquired signals with the spatial position in the three-dimensional structural model. The signal preprocessing module is used to reduce noise and remove interference from the acquired multimodal signals, including filtering, noise reduction and temperature correction, in order to improve signal quality. The feature extraction module is used to extract feature vectors from the preprocessed modal signals, including feature parameters of electromagnetic waves, current, ultrasound and thermal imaging signals, to form a multimodal feature set; The feature fusion module is used to perform data-level fusion of the extracted multimodal feature vectors to generate a comprehensive feature vector, thereby integrating multi-source information to improve the accuracy of fault identification. The fault diagnosis model module is used to input the comprehensive feature vector into the pre-trained multimodal fault diagnosis model and output the fault type identification result and the probability of fault occurrence. The fault location module is used to calculate the spatial coordinates of the fault source based on the arrival time difference of the high-frequency signal and the propagation time of the ultrasonic signal, combined with a three-dimensional structural model. The health index calculation module is used to calculate the comprehensive health index of the dry-type air-core reactor based on the fault type identification results, fault occurrence probability and fault source location signal. The early warning generation module is used to generate graded early warning information and send it to the monitoring terminal when the comprehensive health index is below the threshold or when a specific type of fault is identified.

[0012] Furthermore, the modeling and calibration module includes: a 3D model construction unit, a coordinate system establishment unit, a coordinate calibration unit, and a data association and storage unit; The 3D model building unit is used to perform a full-size scan of the dry air-core reactor using laser scanning technology, obtain geometric parameters, and build a 3D structural model using modeling software. The coordinate system establishment unit is used to establish a right-handed rectangular coordinate system in the three-dimensional structural model, with the intersection of the bottom central axis of the dry-type air reactor and the contact surface of the base as the origin. The coordinate calibration unit is used to select several feature points with known actual positions on the dry-type air-core reactor, measure the actual spatial coordinates, compare them with the coordinates of the three-dimensional structural model to calculate the deviation, and adjust the model until the deviation meets the accuracy requirements. The data association storage unit is used to associate and store sensor deployment locations and signal acquisition data with spatial coordinates in the 3D model.

[0013] Furthermore, the signal preprocessing module includes: an ultra-high frequency signal preprocessing unit, a high frequency pulse current signal preprocessing unit, an ultrasonic signal preprocessing unit, and an infrared thermal imaging signal preprocessing unit; The ultra-high frequency signal preprocessing unit is used to filter low-frequency interference with a bandpass filter and remove high-frequency noise by wavelet threshold noise reduction. A high-frequency pulse current signal preprocessing unit is used to remove power frequency interference using a notch filter and to cancel leakage current interference through adaptive filtering. The ultrasonic signal preprocessing unit is used to filter low-frequency vibration noise with a high-pass filter and remove random sound waves generated by non-partial discharge through short-time Fourier transform. The infrared thermal imaging signal preprocessing unit is used to remove the influence of ambient temperature fluctuations by using the inter-frame difference method and to remove isolated noise points caused by lens dust by using median filtering.

[0014] Furthermore, the fault diagnosis model module includes: a training data management unit, a hybrid model training unit, a fault type identification unit, and a fault probability calculation unit; The training data management unit is used to collect several sets of data containing comprehensive feature vectors and fault type labels in the laboratory and divide them into training sets and validation sets. The hybrid model training unit is used to train the training set with a hybrid model of convolutional neural network and long short-term memory network, and to adjust the model structure with the validation set; The fault type identification unit is used to input the real-time comprehensive feature vector into the trained model and output the fault type; The failure probability calculation unit is used to calculate the probability of failure occurrence based on the model output.

[0015] Furthermore, the health index calculation module includes: fault type scoring unit, fault probability scoring unit, fault location scoring unit, and temperature anomaly scoring unit. The fault type scoring unit is used to calculate the contribution of fault type to the health index based on fault type weight and score. The fault probability scoring unit is used to calculate the contribution of probability to the health index based on the fault probability weight and the probability of fault occurrence. The fault location scoring unit is used to calculate the contribution of the fault location to the health index based on the fault source location weight and score, and the distance between the fault source and the winding. The temperature anomaly scoring unit is used to calculate the contribution of temperature to the health index based on the temperature anomaly weight and score, using hotspot temperatures.

[0016] Compared with the prior art, the beneficial effects of the present invention are: 1. By fusing multi-modal signals from ultra-high frequency electromagnetic waves, high frequency pulse currents, ultrasonic waves, and infrared thermal imaging, the limitations of single detection methods, such as susceptibility to interference and low sensitivity, are overcome, significantly improving the detection capability of early partial discharge and insulation defects.

[0017] 2. The CNN-LSTM hybrid model is used for intelligent diagnosis of multimodal features, which has both spatial and temporal feature extraction capabilities, resulting in high fault identification accuracy and low false alarm rate.

[0018] 3. By combining a three-dimensional structural model with the signal arrival time difference for localization, the precise spatial location of the fault source was achieved, providing clear guidance for maintenance.

[0019] 4. By quantifying equipment status through a comprehensive health index and combining it with a tiered early warning mechanism, full-process monitoring from early warning to serious faults is achieved, effectively improving the safety and reliability of power system operation. Attached Figure Description

[0020] Figure 1 This is a schematic diagram of the method flow of the multi-mode fault monitoring system and method for dry-type air-core reactors according to the present invention; Figure 2 This is a schematic diagram of the system structure of the multi-mode fault monitoring system and method for dry-type air-core reactors according to the present invention. Detailed Implementation

[0021] 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.

[0022] Example 1: As Figure 1 As shown, the present invention provides a technical solution for a multi-mode fault monitoring method for dry-type air-core reactors, the multi-mode fault monitoring method comprising: A three-dimensional structural model of the dry-type air-core reactor is obtained, and a coordinate system is established in the three-dimensional structural model. The ultra-high frequency electromagnetic wave signal, high frequency pulse current signal, ultrasonic signal and infrared thermal imaging signal of the dry-type air-core reactor are collected simultaneously, and the collected signals are associated with the spatial position in the three-dimensional structural model and stored. The collected ultra-high frequency electromagnetic wave signals, high frequency pulse current signals, ultrasonic signals, and infrared thermal imaging signals were preprocessed and feature extracted to obtain feature vectors for each signal mode. The extracted multi-modal feature vectors were then fused at the data level to generate a comprehensive feature vector. The comprehensive feature vector is input into a pre-trained multimodal fault diagnosis model, which outputs fault type identification results and fault occurrence probability; based on the arrival time difference of high-frequency signals and the propagation time of ultrasonic signals, combined with the three-dimensional structural model, the spatial location calculation of the fault source is performed. Based on the fault type identification results, fault occurrence probability, and fault source location signal, the comprehensive health index of the dry-type air-core reactor is calculated; when the comprehensive health index is lower than the threshold or a specific type of fault is identified, a graded early warning message is generated and sent to the monitoring terminal. The process of obtaining a three-dimensional structural model of the dry-type air-core reactor and establishing a coordinate system within that model includes: The dry-type air-core reactor is fully scanned using laser scanning technology to obtain its geometric parameters, and a three-dimensional structural model of the dry-type air-core reactor is constructed using modeling software. In the three-dimensional structural model, a right-hand rectangular coordinate system is established with the intersection of the bottom central axis of the dry-type air-core reactor and the contact surface of the base as the origin. The X-axis of the right-hand rectangular coordinate system extends horizontally outward along the radial direction of the dry-type air-core reactor, the Y-axis is distributed horizontally along the circumference of the dry-type air-core reactor, and the Z-axis is vertically upward along the central axis of the dry-type air-core reactor. Several feature points with known actual locations on the dry-type air-core reactor are selected, and their actual spatial coordinates are measured. These actual spatial coordinates are then compared with the coordinates of corresponding points in the three-dimensional structural model, and the coordinate deviation is calculated. ;in, Indicates coordinate deviation, where A represents the actual spatial coordinates of the feature point, and a represents the coordinates of the corresponding point of the feature point in the three-dimensional structural model; if Then adjust the scaling of the 3D model or the position of the origin of the coordinate system until all feature points are adjusted. ; The process includes simultaneously acquiring ultra-high frequency electromagnetic wave signals, high frequency pulse current signals, ultrasonic signals, and infrared thermal imaging signals from the dry-type air-core reactor, and associating and storing the acquired signals with their spatial positions in the three-dimensional structural model. Several ultra-high frequency (UHF) sensors are deployed around the dry-type air-core reactor. The center position of each UHF sensor is marked with coordinates [coordinates missing] in the 3D model. And record the shortest distance between the ultra-high frequency sensor and the reactor winding; The coil of the current sensor is fitted onto the incoming cable of the dry-type air-core reactor, and the center position of the coil is marked in the three-dimensional structural model as follows. ; Several ultrasonic sensors are deployed along the Z-axis on the outside of the reactor winding. Each ultrasonic sensor is marked with coordinates [coordinates missing] in the three-dimensional structural model. ; Install the infrared thermal imager directly in front of the dry air reactor, align the central axis of the lens with the center of the dry air reactor, and mark the spatial domain corresponding to the field of view of the infrared thermal imager in the three-dimensional structural model. A multi-channel data acquisition card is used to connect all sensors, and a GPS synchronization module is used to ensure that the sampling timestamps of each sensor are consistent. During the data acquisition process, each sensor associates the acquired signal with the corresponding sampling time and sensor coordinates, and stores it in the database. Specifically, the collected ultra-high frequency electromagnetic wave signals, high frequency pulse current signals, ultrasonic signals, and infrared thermal imaging signals were preprocessed and feature extracted to obtain the feature vectors of each signal mode, including: For ultra-high frequency electromagnetic wave signals, a bandpass filter is used to filter low-frequency interference, and wavelet threshold denoising is used to remove high-frequency noise. The waveform of the denoised ultra-high frequency electromagnetic wave signal is as follows: ; For high-frequency pulse current signals, a notch filter is used to remove power frequency interference, and adaptive filtering is used to cancel leakage current interference. The noise-reduced high-frequency pulse current signal is as follows: ; For ultrasonic signals, a high-pass filter is used to filter low-frequency vibration noise, and a short-time Fourier transform is used to remove random sound waves generated by non-partial discharges. The noise-reduced ultrasonic signal is as follows: ; For infrared thermal imaging signals, the inter-frame difference method is used to remove the influence of ambient temperature fluctuations, and median filtering is used to remove isolated noise points caused by lens dust. The corrected winding surface temperature distribution matrix is ​​as follows: ; Feature extraction was performed on each of the preprocessed modal signals, including the feature vector of the ultra-high frequency electromagnetic wave signal. High-frequency pulse current signal feature vector Ultrasonic signal feature vector Infrared thermal imaging signal feature vector Each feature vector includes 5 dimensions. A fault simulation platform for dry-type air-core reactors was built in the laboratory. Several sets of multimodal signals were collected for each fault type. The accuracy of each modal feature vector for fault type identification was calculated, and weights were calculated through normalization. The calculated weights were then used to... , , , The weighted average of each corresponding feature dimension in the vector is used to obtain the comprehensive feature vector V. The integrated feature vector is input into a pre-trained multimodal fault diagnosis model, and the output fault type identification result and fault occurrence probability include: Several sets of data were collected in the laboratory. Each set of data contained a comprehensive feature vector V and a corresponding fault type label. The collected sets of data were divided into training set and validation set according to a pre-set ratio. A hybrid model of convolutional neural network and long short-term memory network is adopted. The training set is input into the hybrid model to train the model. The structure of the trained model is adjusted through the validation set to obtain a multimodal fault diagnosis model. The comprehensive feature vector V, which is collected and fused in real time, is input into the multimodal fault diagnosis model to output the fault type and the probability of fault occurrence. Based on the arrival time difference of the high-frequency signal and the propagation time of the ultrasonic signal, combined with the aforementioned three-dimensional structural model, the spatial location calculation of the fault source includes: Obtaining the propagation speed of ultra-high frequency electromagnetic waves Ultrasonic propagation speed Signal arrival time difference Signal arrival time difference ; Let the coordinates of the fault source be... Based on the relationship between distance, velocity, and time, equations for UHF signals and ultrasonic signals are established; the least squares method is used to fit the coordinates of the fault source. If the calculated fault source coordinates deviate from the actual coordinates of the simulated fault source in the laboratory by less than or equal to a preset threshold, the location result is output; otherwise, the number of sensors is increased and the calculation is repeated. The comprehensive health index of the dry-type air-core reactor is calculated based on the fault type identification results, fault occurrence probability, and fault source location signal, as follows: ; in, For fault type weights, Score the fault type. Let P be the probability weight of the failure, where P is the probability of failure occurring. The fault source location weight, Score the location of the fault source. As for the weight of temperature anomalies, Score for temperature anomalies; Specifically, when the comprehensive health index falls below a threshold or a specific type of fault is identified, generating tiered early warning information and sending it to the monitoring terminal includes: Based on the range of values ​​for the comprehensive health index HI, an early warning level is set; The warning information includes: warning level, comprehensive health index HI, fault type, fault occurrence probability P, and fault source coordinates. Hotspot temperature ; Example 2: Figure 2 As shown, in order to better implement the above method, a multi-mode fault monitoring system for dry-type air-core reactors is also provided. The multi-mode fault monitoring system includes: a modeling and calibration module, a signal acquisition module, a signal preprocessing module, a feature extraction module, a feature fusion module, a fault diagnosis model module, a fault location module, a health index calculation module, and an early warning generation module. The modeling and calibration module is used to obtain the three-dimensional structural model of the dry air-core reactor, establish a coordinate system in the three-dimensional structural model, and ensure that the coordinates of the three-dimensional structural model are consistent with the actual spatial coordinates through feature point calibration, so as to provide a spatial reference for signal correlation and fault location. The signal acquisition module is used to simultaneously acquire ultra-high frequency electromagnetic wave signals, high frequency pulse current signals, ultrasonic signals and infrared thermal imaging signals of the dry air reactor, and to associate and store the acquired signals with the spatial position in the three-dimensional structural model. The signal preprocessing module is used to reduce noise and remove interference from the acquired multimodal signals, including filtering, noise reduction and temperature correction, in order to improve signal quality. The feature extraction module is used to extract feature vectors from the preprocessed modal signals, including feature parameters of electromagnetic waves, current, ultrasound and thermal imaging signals, to form a multimodal feature set; The feature fusion module is used to perform data-level fusion of the extracted multimodal feature vectors to generate a comprehensive feature vector, thereby integrating multi-source information to improve the accuracy of fault identification. The fault diagnosis model module is used to input the comprehensive feature vector into the pre-trained multimodal fault diagnosis model and output the fault type identification result and the probability of fault occurrence. The fault location module is used to calculate the spatial coordinates of the fault source based on the arrival time difference of the high-frequency signal and the propagation time of the ultrasonic signal, combined with a three-dimensional structural model. The health index calculation module is used to calculate the comprehensive health index of the dry-type air-core reactor based on the fault type identification results, fault occurrence probability and fault source location signal. The early warning generation module is used to generate graded early warning information and send it to the monitoring terminal when the comprehensive health index is below the threshold or when a specific type of fault is identified. The modeling and calibration module includes: a 3D model building unit, a coordinate system establishment unit, a coordinate calibration unit, and a data association and storage unit. The 3D model building unit is used to perform a full-size scan of the dry air-core reactor using laser scanning technology, obtain geometric parameters, and build a 3D structural model using modeling software. The coordinate system establishment unit is used to establish a right-handed rectangular coordinate system in the three-dimensional structural model, with the intersection of the bottom central axis of the dry-type air reactor and the contact surface of the base as the origin. The coordinate calibration unit is used to select several feature points with known actual positions on the dry-type air-core reactor, measure the actual spatial coordinates, compare them with the coordinates of the three-dimensional structural model to calculate the deviation, and adjust the model until the deviation meets the accuracy requirements. The data association storage unit is used to associate and store the sensor deployment location and signal acquisition data with the spatial coordinates in the 3D model; The signal preprocessing module includes: a UHF signal preprocessing unit, a high-frequency pulse current signal preprocessing unit, an ultrasonic signal preprocessing unit, and an infrared thermal imaging signal preprocessing unit. The ultra-high frequency signal preprocessing unit is used to filter low-frequency interference with a bandpass filter and remove high-frequency noise by wavelet threshold noise reduction. A high-frequency pulse current signal preprocessing unit is used to remove power frequency interference using a notch filter and to cancel leakage current interference through adaptive filtering. The ultrasonic signal preprocessing unit is used to filter low-frequency vibration noise with a high-pass filter and remove random sound waves generated by non-partial discharge through short-time Fourier transform. The infrared thermal imaging signal preprocessing unit is used to remove the influence of ambient temperature fluctuations by using the inter-frame difference method and to remove isolated noise points caused by lens dust by using median filtering. The fault diagnosis model module includes: a training data management unit, a hybrid model training unit, a fault type identification unit, and a fault probability calculation unit. The training data management unit is used to collect several sets of data containing comprehensive feature vectors and fault type labels in the laboratory and divide them into training sets and validation sets. The hybrid model training unit is used to train the training set with a hybrid model of convolutional neural network and long short-term memory network, and to adjust the model structure with the validation set; The fault type identification unit is used to input the real-time comprehensive feature vector into the trained model and output the fault type; The failure probability calculation unit is used to calculate the probability of failure occurrence based on the model output. The health index calculation module includes: fault type scoring unit, fault probability scoring unit, fault location scoring unit, and temperature anomaly scoring unit. The fault type scoring unit is used to calculate the contribution of fault type to the health index based on fault type weight and score. The fault probability scoring unit is used to calculate the contribution of probability to the health index based on the fault probability weight and the probability of fault occurrence. The fault location scoring unit is used to calculate the contribution of the fault location to the health index based on the fault source location weight and score, and the distance between the fault source and the winding. The temperature anomaly scoring unit is used to calculate the contribution of temperature to the health index based on the temperature anomaly weight and score, using hotspot temperatures. In an embodiment of the present invention, a multimodal monitoring system was deployed on a dry-type air-core reactor in a substation. The system acquires the three-dimensional structure of the reactor using a laser scanner and establishes a spatial coordinate system with the center of the base as the origin. Four ultra-high frequency sensors, one high-frequency current sensor, six ultrasonic sensors, and one infrared thermal imager are arranged around the reactor, and signals are synchronously acquired via GPS. The system acquires multimodal signals in real time, and after filtering, denoising, and feature extraction, they are fused into a comprehensive feature vector, which is then input into a pre-trained CNN-LSTM model for fault diagnosis. In a monitoring session, the system identified an inter-turn discharge fault with a probability of 85% and located the fault point at a height of 2.3 meters on the Z-axis. The health index was 0.62, which was lower than the threshold of 0.75, triggering a level-two early warning. The information was pushed to the operation and maintenance platform in real time to guide on-site maintenance.

[0023] Finally, it should be noted that the above descriptions are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A multi-mode fault monitoring method for dry-type air-core reactors, characterized in that: The multimodal fault monitoring method includes: A three-dimensional structural model of the dry-type air-core reactor is obtained, and a coordinate system is established in the three-dimensional structural model. The ultra-high frequency electromagnetic wave signal, high frequency pulse current signal, ultrasonic signal and infrared thermal imaging signal of the dry-type air-core reactor are collected simultaneously, and the collected signals are associated with the spatial position in the three-dimensional structural model and stored. The collected ultra-high frequency electromagnetic wave signals, high frequency pulse current signals, ultrasonic signals, and infrared thermal imaging signals were preprocessed and feature extracted to obtain feature vectors for each signal mode. The extracted multi-modal feature vectors were then fused at the data level to generate a comprehensive feature vector. The comprehensive feature vector is input into a pre-trained multimodal fault diagnosis model, which outputs fault type identification results and fault occurrence probability; based on the arrival time difference of high-frequency signals and the propagation time of ultrasonic signals, combined with the three-dimensional structural model, the spatial location calculation of the fault source is performed. Based on the fault type identification results, fault occurrence probability, and fault source location signal, the comprehensive health index of the dry-type air-core reactor is calculated; when the comprehensive health index is lower than the threshold or a specific type of fault is identified, a graded early warning message is generated and sent to the monitoring terminal.

2. The multi-mode fault monitoring method for dry-type air-core reactors according to claim 1, characterized in that: Obtaining a three-dimensional structural model of a dry-type air-core reactor and establishing a coordinate system in the three-dimensional structural model includes: The dry-type air-core reactor is fully scanned using laser scanning technology to obtain its geometric parameters, and a three-dimensional structural model of the dry-type air-core reactor is constructed using modeling software. In the three-dimensional structural model, a right-hand rectangular coordinate system is established with the intersection of the bottom central axis of the dry-type air-core reactor and the contact surface of the base as the origin. The X-axis of the right-hand rectangular coordinate system extends horizontally outward along the radial direction of the dry-type air-core reactor, the Y-axis is distributed horizontally along the circumference of the dry-type air-core reactor, and the Z-axis is vertically upward along the central axis of the dry-type air-core reactor. Several feature points with known actual locations on the dry-type air-core reactor are selected, and their actual spatial coordinates are measured. These actual spatial coordinates are then compared with the coordinates of corresponding points in the three-dimensional structural model, and the coordinate deviation is calculated. ;in, Indicates coordinate deviation, where A represents the actual spatial coordinates of the feature point, and a represents the coordinates of the corresponding point of the feature point in the three-dimensional structural model; if Then adjust the scaling of the 3D model or the position of the origin of the coordinate system until all feature points are adjusted. ; The process involves simultaneously acquiring ultra-high frequency electromagnetic wave signals, high frequency pulse current signals, ultrasonic signals, and infrared thermal imaging signals from the dry-type air-core reactor, and associating and storing the acquired signals with their spatial positions in the three-dimensional structural model, including: Several ultra-high frequency (UHF) sensors are deployed around the dry-type air-core reactor. The center position of each UHF sensor is marked with coordinates [coordinates missing] in the 3D model. And record the shortest distance between the ultra-high frequency sensor and the reactor winding; The coil of the current sensor is fitted onto the incoming cable of the dry-type air-core reactor, and the center position of the coil is marked in the three-dimensional structural model as follows. ; Several ultrasonic sensors are deployed along the Z-axis on the outside of the reactor winding. Each ultrasonic sensor is marked with coordinates [coordinates missing] in the three-dimensional structural model. ; Install the infrared thermal imager directly in front of the dry air reactor, align the central axis of the lens with the center of the dry air reactor, and mark the spatial domain corresponding to the field of view of the infrared thermal imager in the three-dimensional structural model. A multi-channel data acquisition card is used to connect all sensors, and a GPS synchronization module is used to ensure that the sampling timestamps of each sensor are consistent. During the data acquisition process, each sensor associates the acquired signal frame with the corresponding sampling time and sensor coordinates, and stores it in the database.

3. The multi-mode fault monitoring method for dry-type air-core reactors according to claim 1, characterized in that: The acquired ultra-high frequency electromagnetic wave signals, high frequency pulse current signals, ultrasonic signals, and infrared thermal imaging signals were preprocessed and feature extracted to obtain the feature vectors of each signal mode, including: For ultra-high frequency electromagnetic wave signals, a bandpass filter is used to filter low-frequency interference, and wavelet threshold noise reduction is used to remove high-frequency noise. The waveform of the denoised ultra-high frequency electromagnetic wave signal is as follows: ; For high-frequency pulse current signals, a notch filter is used to remove power frequency interference, and adaptive filtering is used to cancel leakage current interference. The noise-reduced high-frequency pulse current signal is as follows: ; For ultrasonic signals, a high-pass filter is used to filter low-frequency vibration noise, and a short-time Fourier transform is used to remove random sound waves generated by non-partial discharges. The noise-reduced ultrasonic signal is as follows: ; For infrared thermal imaging signals, the inter-frame difference method is used to remove the influence of ambient temperature fluctuations, and median filtering is used to remove isolated noise points caused by lens dust. The corrected winding surface temperature distribution matrix is ​​as follows: ; Feature extraction was performed on each of the preprocessed modal signals, including the feature vector of the ultra-high frequency electromagnetic wave signal. High-frequency pulse current signal feature vector Ultrasonic signal feature vector Infrared thermal imaging signal feature vector Each feature vector includes 5 dimensions. A fault simulation platform for dry-type air-core reactors was built in the laboratory. Several sets of multimodal signals were collected for each fault type. The accuracy of each modal feature vector for fault type identification was calculated, and weights were calculated through normalization. The calculated weights were then used to... , , , We take a weighted average of each corresponding feature dimension to obtain the comprehensive feature vector V.

4. The multi-mode fault monitoring method for dry-type air-core reactors according to claim 1, characterized in that: The comprehensive feature vector is input into a pre-trained multimodal fault diagnosis model, and the output fault type identification result and fault occurrence probability include: Several sets of data were collected in the laboratory. Each set of data contained a comprehensive feature vector V and a corresponding fault type label. The collected sets of data were divided into training set and validation set according to a pre-set ratio. A hybrid model of convolutional neural network and long short-term memory network is adopted. The training set is input into the hybrid model to train the model. The structure of the trained model is adjusted through the validation set to obtain a multimodal fault diagnosis model. The comprehensive feature vector V, which is collected and fused in real time, is input into the multimodal fault diagnosis model to output the fault type and the probability of fault occurrence. Based on the arrival time difference of the high-frequency signal and the propagation time of the ultrasonic signal, combined with the aforementioned three-dimensional structural model, the spatial location calculation of the fault source includes: Obtaining the propagation speed of ultra-high frequency electromagnetic waves Ultrasonic propagation speed Signal arrival time difference Signal arrival time difference ; Let the coordinates of the fault source be... Based on the relationship between distance, velocity, and time, equations for UHF signals and ultrasonic signals are established; the least squares method is used to fit the coordinates of the fault source. If the calculated fault source coordinates deviate from the actual coordinates of the simulated fault source in the laboratory by less than or equal to a preset threshold, the positioning result is output; otherwise, the number of sensors is increased and the calculation is repeated.

5. The multi-mode fault monitoring method for dry-type air-core reactors according to claim 1, characterized in that: Based on the fault type identification results, fault occurrence probability, and fault source location signal, the comprehensive health index of the dry-type air-core reactor is calculated as follows: ； in, For fault type weights, Score the fault type. Let P be the probability weight of the failure, where P is the probability of failure occurring. The fault source location weight, Score the location of the fault source. As for the weight of temperature anomalies, Score for temperature anomalies; When the comprehensive health index falls below a threshold or a specific type of fault is identified, a graded early warning message is generated and sent to the monitoring terminal, including: Based on the range of values ​​for the comprehensive health index HI, an early warning level is set; The warning information includes: warning level, comprehensive health index HI, fault type, fault occurrence probability P, and fault source coordinates. Hotspot temperature .

6. A multi-mode fault monitoring system for dry-type air-core reactors, used to execute the multi-mode fault monitoring method for dry-type air-core reactors according to any one of claims 1-5, characterized in that: The multimodal fault monitoring system includes: a modeling and calibration module, a signal acquisition module, a signal preprocessing module, a feature extraction module, a feature fusion module, a fault diagnosis model module, a fault location module, a health index calculation module, and an early warning generation module; The modeling and calibration module is used to obtain the three-dimensional structural model of the dry air-core reactor, establish a coordinate system in the three-dimensional structural model, and ensure that the coordinates of the three-dimensional structural model are consistent with the actual spatial coordinates through feature point calibration, so as to provide a spatial reference for signal correlation and fault location. The signal acquisition module is used to simultaneously acquire ultra-high frequency electromagnetic wave signals, high frequency pulse current signals, ultrasonic signals and infrared thermal imaging signals of the dry air reactor, and associate and store the acquired signals with the spatial position in the three-dimensional structural model; The signal preprocessing module is used to perform noise reduction and interference removal on the acquired multimodal signals, including filtering, noise reduction and temperature correction processing, in order to improve signal quality; The feature extraction module is used to extract feature vectors from the preprocessed modal signals, including feature parameters of electromagnetic waves, current, ultrasound and thermal imaging signals, to form a multimodal feature set; The feature fusion module is used to perform data-level fusion of the extracted multimodal feature vectors to generate a comprehensive feature vector, thereby integrating multi-source information to improve the accuracy of fault identification. The fault diagnosis model module is used to input the comprehensive feature vector into the pre-trained multimodal fault diagnosis model and output the fault type identification result and the probability of fault occurrence. The fault location module is used to calculate the spatial coordinates of the fault source based on the arrival time difference of the high-frequency signal and the propagation time of the ultrasonic signal, combined with a three-dimensional structural model. The health index calculation module is used to calculate the comprehensive health index of the dry-type air-core reactor based on the fault type identification results, fault occurrence probability and fault source location signal. The warning generation module is used to generate graded warning information and send it to the monitoring terminal when the comprehensive health index is below the threshold or a specific type of fault is identified.

7. The multi-mode fault monitoring system for dry-type air-core reactors according to claim 6, characterized in that: The modeling and calibration module includes: a 3D model construction unit, a coordinate system establishment unit, a coordinate calibration unit, and a data association and storage unit; The three-dimensional model building unit is used to perform a full-size scan of the dry air-core reactor using laser scanning technology, obtain geometric parameters, and build a three-dimensional structural model using modeling software. The coordinate system establishment unit is used to establish a right-handed rectangular coordinate system in the three-dimensional structural model, with the intersection of the bottom central axis of the dry-type air reactor and the contact surface of the base as the origin. The coordinate calibration unit is used to select several feature points with known actual positions on the dry-type air-core reactor, measure the actual spatial coordinates, compare them with the coordinates of the three-dimensional structural model to calculate the deviation, and adjust the model until the deviation meets the accuracy requirements. The data association storage unit is used to associate and store the sensor deployment location and signal acquisition data with the spatial coordinates in the three-dimensional model.

8. The multi-mode fault monitoring system for dry-type air-core reactors according to claim 6, characterized in that: The signal preprocessing module includes: an ultra-high frequency signal preprocessing unit, a high frequency pulse current signal preprocessing unit, an ultrasonic signal preprocessing unit, and an infrared thermal imaging signal preprocessing unit. The ultra-high frequency signal preprocessing unit is used to filter low-frequency interference with a bandpass filter and remove high-frequency noise with wavelet threshold noise reduction. The high-frequency pulse current signal preprocessing unit is used to remove power frequency interference using a notch filter and to cancel leakage current interference using adaptive filtering. The ultrasonic signal preprocessing unit is used to filter low-frequency vibration noise with a high-pass filter and remove random sound waves generated by non-partial discharge through short-time Fourier transform. The infrared thermal imaging signal preprocessing unit is used to remove the influence of ambient temperature fluctuations by using the inter-frame difference method and to remove isolated noise points caused by lens dust by using median filtering.

9. The multi-mode fault monitoring system for dry-type air-core reactors according to claim 6, characterized in that: The fault diagnosis model module includes: a training data management unit, a hybrid model training unit, a fault type identification unit, and a fault probability calculation unit; The training data management unit is used to collect several sets of data containing comprehensive feature vectors and fault type labels in the laboratory, and divide them into training sets and validation sets. The hybrid model training unit is used to train the training set using a hybrid model of convolutional neural network and long short-term memory network, and to adjust the model structure using a validation set. The fault type identification unit is used to input the real-time comprehensive feature vector into the trained model and output the fault type. The fault probability calculation unit is used to calculate the probability of fault occurrence based on the model output.

10. The multi-mode fault monitoring system for dry-type air-core reactors according to claim 6, characterized in that: The health index calculation module includes: a fault type scoring unit, a fault probability scoring unit, a fault location scoring unit, and a temperature anomaly scoring unit. The fault type scoring unit is used to calculate the contribution value of the fault type to the health index based on the fault type weight and score. The fault probability scoring unit is used to calculate the contribution value of the probability to the health index based on the fault probability weight and the fault occurrence probability. The fault location scoring unit is used to calculate the contribution value of the location to the health index based on the fault source location weight and score, and the distance between the fault source and the winding. The temperature anomaly scoring unit is used to calculate the contribution of temperature to the health index based on the temperature anomaly weight and score, using hotspot temperatures.

Images