Transformer heating point detection method based on array ultrasound

Abstract

The invention discloses a transformer heating point detection method based on array ultrasound. The method comprises the steps that the installation position and the scanning angle range of an array ultrasonic probe and the transmitting frequency, the pulse width and the transmitting power parameter of an ultrasonic signal are determined; sending an excitation signal to the array ultrasonic probe, controlling the array ultrasonic probe to emit an ultrasonic signal to a to-be-detected area of the transformer according to a set parameter, and collecting an echo signal received by the array ultrasonic probe in real time; the received echo signals are preprocessed, effective echo signals are obtained, feature extraction is carried out on the effective echo signals, a heating point detection model is constructed, and a mapping relation between signal features and heating point temperatures, positions and fault types is established; and inputting the signal features extracted in real time into the heating point detection model, and outputting whether a heating point exists in the to-be-detected region of the transformer, the specific space coordinates of the heating point, the temperature value of the heating point and the type of the heating fault. Reliable technical guarantee is provided for safe and stable operation of the transformer.

Description

Technical Field

[0001] This invention relates to the field of transformer condition monitoring technology, and in particular to a method for detecting transformer hot spots based on array ultrasound. Background Technology

[0002] As a core piece of equipment in the power system, the transformer's operating status directly determines the stability and security of power transmission. During long-term operation, critical components inside the transformer, such as windings, core, and joints, are prone to localized heating due to factors like current loss, hysteresis loss, and increased contact resistance. If the temperature at these heating points continues to rise and is not detected in time, it can lead to accelerated aging of insulation materials, shortening the equipment's lifespan, or even causing insulation breakdown, winding burnout, and widespread power outages. Therefore, accurate and real-time detection of transformer heating points is a crucial step in ensuring the reliable operation of the power system.

[0003] Current mainstream transformer hotspot detection technologies mainly include the following categories, but all of them have obvious technical limitations:

[0004] (1) Infrared temperature measurement technology: It relies on infrared thermal imagers to capture the temperature distribution on the surface of transformers. It can only detect the external and shallow heating phenomena of equipment, and cannot penetrate the structure such as oil tank and iron core to obtain information on internal heating points. Moreover, it is greatly affected by ambient temperature, dust and electromagnetic interference. The detection accuracy is significantly reduced in high temperature, high humidity or strong electromagnetic environment, and it is especially difficult to locate internal hidden heating points such as winding joints and iron core lamination joints.

[0005] (2) Dissolved gas analysis technology in oil: The fault is determined by detecting the content of characteristic gases (such as methane, ethylene, and acetylene) generated by insulation overheating in transformer oil. However, this technology is a post-event detection - it takes several hours to several days for the gas to be generated, dissolved in the oil and reach a detectable concentration, which cannot achieve real-time monitoring. Moreover, it can only determine whether there is a heat fault, but cannot determine the specific location, temperature and fault type of the heat point, which is difficult to meet the needs of rapid location and accurate handling.

[0006] (3) Traditional ultrasonic testing technology: Single-element ultrasonic probes are used for single-point scanning, which requires manual movement of the probe to test each area. The detection efficiency is extremely low and is greatly affected by human operation errors. At the same time, the coverage of single-element probes is limited, which cannot achieve full coverage of key heat-prone areas of transformers. It is easy to miss heat-prone areas such as winding and bushing interfaces and internal connection parts of the tank wall. In addition, traditional ultrasonic testing can only make a preliminary judgment on the presence of abnormalities by echo amplitude. It lacks in-depth signal processing and feature analysis, cannot accurately calculate the temperature of the heat point, locate the spatial coordinates, and is difficult to distinguish different types of heat-prone faults (such as winding overheating and core overheating).

[0007] (4) Vibration detection technology: Internal faults are analyzed by collecting vibration signals of the transformer shell. However, the vibration signals are easily affected by the operating load of the equipment and external mechanical vibrations (such as fans and pumps). The signal-to-noise ratio is low, and it is not sensitive to vibration changes caused by slight heating. Furthermore, it cannot directly correlate the location and temperature information of the heating point, resulting in poor fault location accuracy.

[0008] In summary, existing technologies for detecting transformer hot spots generally suffer from limitations such as limited detection range, poor real-time performance, low positioning accuracy, and difficulty in distinguishing fault types. These limitations make it difficult to meet the demands of modern power systems for comprehensive, high-precision, real-time, and intelligent transformer condition monitoring. Therefore, developing a detection method that can overcome these technical bottlenecks and achieve precise positioning of internal transformer hot spots, real-time temperature monitoring, and fault type identification has become an urgent technical problem to be solved in this field. Summary of the Invention

[0009] To address the technical problems existing in the prior art, this invention proposes a transformer heating point detection method based on array ultrasound. This method effectively solves the core pain points of existing transformer heating point detection technologies, such as incomplete coverage, inaccurate positioning, poor real-time performance, and difficulty in classifying types. It achieves a technological breakthrough from manual inspection to intelligent monitoring, from post-event handling to pre-event early warning, and from fuzzy diagnosis to precise positioning, providing a reliable technical guarantee for the safe and stable operation of transformers.

[0010] To achieve the above objectives, the present invention provides a method for detecting transformer hot spots based on array ultrasound, comprising:

[0011] Based on the transformer's model, structural dimensions, and the distribution characteristics of the area to be tested, the array ultrasonic probe is calibrated and its parameters are set to determine the installation position, scanning angle range, and transmission frequency, pulse width, and transmission power parameters of the ultrasonic signal. The array ultrasonic probe can cover the transformer's key heat-prone areas, including the transformer's winding joints, core lamination joints, tank wall and internal component connections, and bushing and tank interfaces.

[0012] An excitation signal is sent to the array ultrasonic probe to control the array ultrasonic probe to emit ultrasonic signals to the transformer to be tested according to the set parameters, and the echo signals received by the array ultrasonic probe are collected in real time. The echo signals include reflection signals from different medium interfaces inside the transformer and abnormal scattering signals caused by changes in local medium properties due to heating.

[0013] The received echo signal is preprocessed to obtain a valid echo signal. Features are extracted from the valid echo signal, and a hot spot detection model is constructed based on the extracted signal features. A mapping relationship is established between the signal features and the temperature, location, and fault type of the hot spot.

[0014] The real-time extracted signal features are input into the hot spot detection model, which outputs whether there is a hot spot in the transformer detection area, the specific spatial coordinates of the hot spot, the temperature value of the hot spot, and the type of hot spot fault.

[0015] Preferably, the array ultrasonic probe adopts a multi-element linear arrangement structure with 32-128 elements, each element having a center frequency of 1-10MHz and an element spacing of 0.5-2mm. The outer shell of the array ultrasonic probe is made of a high-temperature resistant, oil-resistant material with electromagnetic shielding function, and the surface of the outer shell is provided with an anti-slip and wear-resistant coating. Ultrasonic signal transmission is achieved between the array ultrasonic probe and the transformer shell through a coupling agent, which is a high-temperature resistant silicon-based coupling agent.

[0016] Preferably, a three-dimensional coordinate positioning system is used to calibrate the position and set the parameters of the array ultrasonic probe. The three-dimensional coordinate positioning system includes a laser locator, a displacement sensor, and a coordinate calculation unit. The laser locator determines the three-dimensional coordinates of several reference positioning points on the transformer shell. The displacement sensor measures the distance and angle between the array ultrasonic probe and each reference positioning point. Finally, the coordinate calculation unit calculates the actual three-dimensional coordinates of each element of the array ultrasonic probe based on the measurement data and compares them with the preset theoretical coordinates. If the coordinate deviation exceeds the preset deviation, the position of the array ultrasonic probe is adjusted by an electric adjustment mechanism until the coordinate deviation is less than the preset deviation.

[0017] Preferably, the received echo signal is preprocessed, including signal denoising, signal amplification, baseline correction, and signal normalization.

[0018] The signal denoising method employs a combination of wavelet transform denoising and adaptive filtering. The original echo signal is decomposed into wavelets, and then the wavelet coefficients at each level are thresholded. A soft threshold function is used to denoise the high-frequency wavelet coefficients, and the denoised wavelet coefficients are reconstructed by wavelets to obtain the pre-denoised signal. The pre-denoised signal is then input into an adaptive filter to further eliminate residual noise and interference signals.

[0019] The signal amplification uses a low-noise operational amplifier, and the amplification factor is adaptively adjusted according to the amplitude of the echo signal.

[0020] The baseline correction adopts a polynomial fitting method. By performing polynomial fitting on the baseline part of the echo signal, a baseline fitting curve is obtained. Then, the baseline fitting curve is subtracted from the original echo signal to achieve baseline correction.

[0021] The signal normalization process employs a maximum-minimum normalization method, which maps the preprocessed signal amplitude to the range of 0-1, eliminating the impact of signal amplitude differences under different detection conditions on subsequent feature extraction.

[0022] Preferably, the features of the extracted effective echo signal include: amplitude variation features, phase shift features, signal propagation time delay features, and spectral distribution features of the echo signal. The amplitude variation features are used to reflect the change in acoustic impedance of the transformer's internal medium due to temperature rise. The phase shift features and signal propagation time delay features are used to locate the spatial position of the heat source. The spectral distribution features are used to distinguish different types of heat-related faults.

[0023] Preferably, the heat point detection model is obtained by training a large number of known heat fault types and transformer ultrasonic signal samples under normal operating conditions using a machine learning algorithm;

[0024] The machine learning algorithm employs a fusion algorithm combining Support Vector Machines (SVMs) and Random Forests. The extracted signal features are divided into training and testing sets. The training set samples undergo feature standardization to eliminate the influence of differences in the units of measurement between different features. Then, SVM and Random Forest models are constructed separately. The SVM model uses a radial basis function kernel, and the penalty coefficient and kernel function parameters are determined through cross-validation. The number of decision trees in the Random Forest model is set to 50-200, and the maximum depth of each decision tree is set to 5-15 layers. The Gini coefficient is used as the feature selection criterion when splitting nodes. The training set samples are input into the SVM and Random Forest models respectively for training, resulting in two independent sub-models. Finally, a fusion model is constructed, and the outputs of the two sub-models are merged using a weighted voting method. The weights of the SVM model and the Random Forest model are determined based on the accuracy of the two sub-models on the testing set.

[0025] Preferably, when determining the specific spatial coordinates of the heating point, a multi-element positioning algorithm is adopted. The propagation time delay of the echo signal of the same heating point received by different elements in the array ultrasonic probe is used to establish a set of spatial positioning equations, and the least squares method is used to solve the set of spatial positioning equations.

[0026] The spatial positioning equations are based on a spherical wave propagation model. The spherical equations are established with the three-dimensional coordinates of each array element as the center and the propagation distance of the ultrasonic signal from the heating point to the array element as the radius. The intersection of the spherical equations corresponding to multiple array elements is the spatial coordinate of the heating point.

[0027] Preferably, the method further includes real-time display of the detection results of the hot spots. The detection results include a location distribution image of the hot spots, a temperature change curve, and a fault type identifier. When the temperature value of the hot spot exceeds a preset safety threshold or a serious overheating fault occurs, an audible and visual alarm signal is triggered. At the same time, the detection results and alarm information are stored in the database and sent to the remote monitoring terminal.

[0028] Compared with the prior art, the present invention has the following advantages and technical effects:

[0029] (1) The present invention uses an array ultrasonic probe with multiple linear array elements, combined with a three-dimensional coordinate positioning system to achieve precise calibration of the probe position. The scanning angle and coverage range can be customized according to the transformer model and structural size to ensure full coverage of all key heat-prone areas such as winding joints, core lamination joints, bushing and tank interfaces, and internal connection parts of the tank wall, thus completely eliminating the detection blind spots of traditional single-element probes with single-point scanning and incomplete coverage.

[0030] (2) This invention extracts the spectral distribution characteristics (center frequency, bandwidth, frequency component power ratio), cross-correlation characteristics (signal correlation between array elements), and entropy characteristics (information entropy, wavelet entropy) of the echo signal, and combines them with a machine learning model that integrates support vector machine and random forest. This invention can accurately distinguish different types of overheating faults such as winding overheating, core overheating, and poor joint contact. It solves the problem that traditional technology only knows the fault but not the type, and provides maintenance personnel with accurate diagnostic results of fault type and severity, which facilitates the development of targeted treatment plans. Attached Figure Description

[0031] The accompanying drawings, which form part of this application, are used to provide a further understanding of this application. The illustrative embodiments and descriptions of this application are used to explain this application and do not constitute an undue limitation of this application. In the drawings:

[0032] Figure 1 This is a flowchart of a transformer heating point detection method based on array ultrasound according to an embodiment of the present invention. Detailed Implementation

[0033] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.

[0034] It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and although a logical order is shown in the flowchart, in some cases the steps shown or described may be executed in a different order than that shown here.

[0035] This embodiment proposes a method for detecting transformer heating points based on array ultrasound, such as... Figure 1 ,include:

[0036] Based on the transformer's model, structural dimensions, and the distribution characteristics of the area to be tested, the array ultrasonic probe is calibrated and its parameters are set to determine the installation position, scanning angle range, and transmission frequency, pulse width, and transmission power parameters of the ultrasonic signal. The array ultrasonic probe can cover the transformer's key heat-prone areas, including the transformer's winding joints, core lamination joints, tank wall and internal component connections, and bushing and tank interfaces.

[0037] An excitation signal is sent to the array ultrasonic probe to control the array ultrasonic probe to emit ultrasonic signals to the transformer to be tested according to the set parameters, and the echo signals received by the array ultrasonic probe are collected in real time. The echo signals include reflection signals from different medium interfaces inside the transformer and abnormal scattering signals caused by changes in local medium properties due to heating.

[0038] The received echo signal is preprocessed to obtain a valid echo signal. Features are extracted from the valid echo signal, and a hot spot detection model is constructed based on the extracted signal features. A mapping relationship is established between the signal features and the temperature, location, and fault type of the hot spot.

[0039] The real-time extracted signal features are input into the hot spot detection model, which outputs whether there is a hot spot in the transformer detection area, the specific spatial coordinates of the hot spot, the temperature value of the hot spot, and the type of hot spot fault.

[0040] This embodiment uses a 110kV oil-immersed power transformer as the test object. The transformer has a rated capacity of 50MVA and the tank size is 4.2m×2.1m×2.8m. The key heat-prone areas include the high-voltage winding joints (6 in total), the low-voltage winding joints (6 in total), the core lamination joints (8 in total), the connection between the tank wall and the core support (4 in total), and the bushing and tank interface (3 in total).

[0041] Furthermore, the array ultrasonic probe adopts a multi-element linear arrangement structure with 32-128 elements, each element having a center frequency of 1-10MHz and an element spacing of 0.5-2mm. The outer shell of the array ultrasonic probe is made of a high-temperature resistant, oil-resistant material with electromagnetic shielding function, and the surface of the outer shell is provided with an anti-slip and wear-resistant coating. Ultrasonic signal transmission is achieved between the array ultrasonic probe and the transformer shell through a coupling agent, which is a high-temperature resistant silicon-based coupling agent.

[0042] Specifically, the number of array elements is determined to be 64 (balancing coverage and detection accuracy; 32 elements would be insufficient, and 128 elements would be too expensive), with a center frequency of 3MHz for each element (adapting to the thickness of the transformer's testing area: approximately 80mm for the winding joint area and approximately 120mm for the core lamination area), and an element spacing of 1mm (to avoid signal interference between elements while ensuring scanning resolution). The probe housing is made of polytetrafluoroethylene composite material, which is resistant to high temperatures and transformer oil corrosion, and an added nickel-copper alloy shielding layer provides electromagnetic shielding (shielding effectiveness ≥40dB, capable of resisting strong electromagnetic interference in substations); the housing surface is coated with an alumina wear-resistant coating (50μm thick, hardness HV800, to prevent wear during installation and use). A coupling agent application groove (5mm wide, 2mm deep) is provided at the contact point between the probe and the transformer housing to ensure uniform distribution of the coupling agent.

[0043] Furthermore, a three-dimensional coordinate positioning system is used to calibrate the position and set the parameters of the array ultrasonic probe. The three-dimensional coordinate positioning system includes a laser locator, a displacement sensor, and a coordinate calculation unit. The laser locator determines the three-dimensional coordinates of several reference positioning points on the transformer shell. The displacement sensor measures the distance and angle between the array ultrasonic probe and each reference positioning point. Finally, the coordinate calculation unit calculates the actual three-dimensional coordinates of each element of the array ultrasonic probe based on the measurement data and compares them with the preset theoretical coordinates. If the coordinate deviation exceeds the preset deviation, the position of the array ultrasonic probe is adjusted by an electric adjustment mechanism until the coordinate deviation is less than the preset deviation.

[0044] Specifically, including:

[0045] Reference point determination: Six reference points were selected on the transformer casing, distributed as follows: ① Upper left corner of the front of the tank (near the high-voltage bushing); ② Lower right corner of the front of the tank (near the low-voltage bushing); ③ Upper left corner of the back of the tank; ④ Lower right corner of the back of the tank; ⑤ Middle of the left side of the tank; ⑥ Middle of the right side of the tank. A laser positioning instrument was used to measure the three-dimensional coordinates of each reference point (with the center of the transformer base as the origin, the X-axis along the length of the tank, the Y-axis along the width, and the Z-axis along the height).

[0046] Initial probe installation: Based on the distribution of critical heat-prone areas in the transformer, the array ultrasonic probes are installed in three groups: ① The first group of probes (2 probes) is installed in the center of the front of the tank, corresponding to the high-voltage winding joint and the core lamination joint; ② The second group of probes (2 probes) is installed in the center of the back of the tank, corresponding to the low-voltage winding joint and the connection between the tank wall and the core support; ③ The third group of probes (1 probe) is installed on the top of the tank near the bushing, corresponding to the bushing and the tank interface. Magnetic fixing brackets (magnetic force ≥500N to ensure tight probe contact) are used during installation. The brackets are equipped with an electric adjustment mechanism (accuracy ±0.01mm, including X / Y / Z three-axis adjustment motors).

[0047] Probe Coordinate Measurement and Adjustment: Displacement sensors are used to measure the distance and angle between the center of each probe and six reference positioning points. For example, the distance from the center of the left probe in the first group to P1 is measured to be 1.5m, with angles of 30° (angle with the X-axis) and 15° (angle with the Y-axis). The actual three-dimensional coordinates of each probe element are calculated using a coordinate calculation unit. For example, the actual coordinates of a certain element are (1.8m, 0.3m, 1.5m), while the preset theoretical coordinates are (1.8005m, 0.3m, 1.5m), resulting in a coordinate deviation of 0.0005m (0.5mm), exceeding the 0.1mm deviation threshold. At this point, the electric adjustment mechanism is activated, controlling the X-axis motor to fine-tune the probe position until the deviation between the actual and theoretical coordinates of the element is less than 0.1mm (e.g., a deviation of 0.08mm after adjustment). This process is repeated to complete the coordinate calibration of all probe elements.

[0048] Calibration result verification: After calibration, the three-dimensional coordinates of the probe array elements were measured again using a laser positioning instrument. The deviation data before and after calibration were compared to ensure that the coordinate deviation of all array elements was ≤0.1mm. At the same time, by transmitting test ultrasonic signals to the probe, the position of the reflected wave peak of the transformer internal structure (such as the winding core interface) in the echo signal was observed to see if it was stable (10 consecutive measurements, wave peak position deviation ≤0.05mm) to verify the calibration effect.

[0049] Furthermore, the received echo signal is preprocessed, including signal denoising, signal amplification, baseline correction, and signal normalization.

[0050] The signal denoising process employs a combination of wavelet transform denoising and adaptive filtering. First, the original echo signal is decomposed into wavelet layers of 3-5, using the db4 wavelet as the wavelet basis function. Then, the wavelet coefficients at each layer are thresholded using a soft thresholding function to reduce noise in the high-frequency wavelet coefficients. The threshold is determined based on the standard deviation of the wavelet coefficients at each layer. Next, the processed wavelet coefficients are reconstructed using wavelet transform to obtain the pre-denoised signal. This pre-denoised signal is then input into an adaptive filter. The adaptive filter uses a minimum mean square error algorithm, employing the ultrasonic signal collected under normal transformer operation as the reference signal. By continuously adjusting the filter coefficients, the mean square error between the filter's output signal and the reference signal is minimized, thereby further eliminating residual noise and interference signals.

[0051] The signal amplification uses a low-noise operational amplifier, and the amplification factor can be adaptively adjusted according to the amplitude of the echo signal. The amplification factor range is 10-1000 times, ensuring that the amplitude of the amplified signal is within the optimal input range of the data acquisition module.

[0052] Baseline correction employs a polynomial fitting method. By performing polynomial fitting on the baseline portion of the echo signal, a baseline fitting curve is obtained. The original echo signal is then subtracted from the baseline fitting curve to achieve baseline correction.

[0053] The signal normalization process uses the maximum-minimum normalization method to map the preprocessed signal amplitude to the range of 0-1, so as to eliminate the influence of signal amplitude differences under different detection conditions on subsequent feature extraction.

[0054] Specifically, the ultrasonic signal collected under normal transformer operation was selected as the reference signal. After wavelet denoising, the signal showed no obvious abnormal fluctuations, the position of the reflection peak was stable (e.g., the reflection peak at the core interface appeared at 85μs), and historical data verified that there were no records of overheating faults under this condition, thus ensuring the purity of the reference signal.

[0055] Filter parameter settings: The adaptive filter uses the minimum mean square error (LMS) algorithm. In LabVIEW, the Adaptive Filter Toolkit is called to set the filter order to 32 (too high an order will increase the amount of computation, while too low an order will result in poor filtering effect; 32 orders can balance the two). The step size factor is set to 0.001 (too large a step size will cause the filter to be unstable, while too small a step size will result in slow convergence; 0.001 can ensure convergence within 100 sampling points).

[0056] Filtering: The pre-denoised signal is used as the input signal, and the reference signal is used as the desired signal, both input to an adaptive filter. The filter iteratively adjusts the weighting coefficients to minimize the mean square error between the output signal and the desired signal. For example, if the pre-denoised signal still contains 100Hz power frequency interference (amplitude ±0.05V), after adaptive filtering, the interference amplitude is reduced to below ±0.01V, the signal waveform becomes smoother, and the amplitude deviation of key characteristic peaks is reduced from ±5% to ±2%, further improving signal quality.

[0057] Furthermore, the features of the extracted effective echo signal include: amplitude variation features, phase shift features, signal propagation time delay features, and spectral distribution features. The amplitude variation features are used to reflect the acoustic impedance changes of the transformer's internal medium due to temperature increases. The phase shift features and signal propagation time delay features are used to locate the spatial position of the heating point. The spectral distribution features are used to distinguish different types of heating faults.

[0058] Specifically, when extracting the amplitude change characteristics of the echo signal, the characteristic peak in the echo signal is first determined. The characteristic peak corresponds to the reflected signal of a specific structural interface inside the transformer. By comparing the amplitude of the same characteristic peak at different detection times, the amplitude change rate is calculated. Amplitude change rate = (amplitude at current time - amplitude at initial time) / amplitude at initial time × 100%. When the absolute value of the amplitude change rate exceeds 5%, it is determined that there is a temperature change in the area corresponding to the characteristic peak.

[0059] When extracting phase shift features, Hilbert transform is used to process the echo signal to obtain the analytical signal. The instantaneous phase of the echo signal is calculated through the analytical signal. The instantaneous phase difference of the same characteristic peak at different detection times is compared to obtain the phase shift. The phase shift is linearly related to the temperature change. The corresponding temperature change value can be calculated based on the phase shift using a pre-established phase shift-temperature change calibration curve.

[0060] When extracting the signal propagation time delay feature, the propagation time of the characteristic wave peak from transmission to reception is determined, the difference in propagation time at different detection times is calculated, and the propagation time delay is obtained. Based on the relationship between the propagation speed of the ultrasonic signal in the medium and temperature, combined with the propagation time delay and the known propagation path length, the location coordinates of the hot spot are calculated.

[0061] When extracting spectral distribution features, a fast Fourier transform is used to perform spectral analysis on the echo signal to obtain the power spectral density distribution of the signal. The center frequency, bandwidth, and power ratio of each frequency component of the power spectrum are calculated. When an internal overheating fault occurs in the transformer, the center frequency of the spectrum will shift towards lower or higher frequencies, the bandwidth will increase, and the power ratio of each frequency component will also change. By analyzing these changes in spectral parameters, different types of overheating faults can be distinguished. For example, overheating faults in the windings will cause the center frequency of the spectrum to shift towards lower frequencies, while overheating faults in the core will cause the bandwidth of the spectrum to increase.

[0062] Furthermore, the heat point detection model is obtained by training a large number of known heat fault types and transformer ultrasonic signal samples under normal operating conditions using a machine learning algorithm.

[0063] The machine learning algorithm employs a fusion algorithm combining Support Vector Machines (SVMs) and Random Forests. The extracted signal features are divided into training and testing sets, with the training set comprising 70%-80% of the total samples and the testing set comprising 20%-30%. The training set samples undergo feature standardization to eliminate the influence of differences in the units of measurement between different features. Then, SVM and Random Forest models are constructed separately. The SVM model uses a radial basis function kernel, and the penalty coefficient and kernel function parameters are determined through cross-validation. The Random Forest model has 50-200 decision trees, with a maximum depth of 5-15 layers per tree. The Gini coefficient is used as the feature selection criterion when splitting nodes. The training set samples are input into the SVM and Random Forest models respectively for training, resulting in two independent sub-models. Finally, a fusion model is constructed, and the outputs of the two sub-models are merged using a weighted voting method. The weights of the SVM and Random Forest models are determined based on the accuracy of the two sub-models on the testing set.

[0064] Specifically, the construction of the Support Vector Machine (SVM) sub-model includes:

[0065] Feature standardization: Z-score standardization is performed on the feature vectors of the training set.

[0066] Parameter optimization: Five-fold cross-validation (dividing the training set into 5 parts, 4 for training and 1 for validation, repeated 5 times) was used to determine the penalty coefficient C and kernel function parameter γ (radial basis function RBF) of the SVM. A grid search method was used, setting the search range of C to [0.1, 1, 10, 100] and the search range of γ to [0.001, 0.01, 0.1, 1, 10]. The accuracy of the model on the validation set under different combinations of (C, γ) was calculated. The optimal parameters were finally determined to be C=10 and γ=0.1 (at which point the validation set accuracy was the highest, reaching 93%).

[0067] Model training: In Python, call the sklearn.svm.SVC function, input the optimized parameters and the standardized training set, train the SVM sub-model, and obtain the model's decision function. This function maps feature vectors to a high-dimensional space, finds the optimal classification hyperplane, and classifies "normal / fault" and fault types.

[0068] The construction of the Random Forest (RF) sub-model includes:

[0069] Parameter settings: Set the number of decision trees to 100, the maximum depth to 10 layers, and the node splitting criterion to the Gini coefficient.

[0070] Model Training: The `sklearn.ensemble.RandomForestClassifier` function was called, inputting the training set feature vectors and labels (normal=0, winding overheat=1, core overheat=2, bushing overheat=3, joint defective=4) to train the RF sub-model. This model constructs 100 independent decision trees, each trained based on randomly sampled samples and features, and finally outputs the majority vote of each tree's prediction as the sub-model's prediction result. After training, the RF sub-model achieved an accuracy of 96% on the training set and 94% on the validation set.

[0071] A weighted voting method is used to fuse the prediction results of the two sub-models. For each test sample, if the SVM predicts "winding overheat" and the RF predicts "winding overheat", it is determined to be "winding overheat". If the SVM predicts "normal" and the RF predicts "slight overheat" (0.506), it is determined to be "slight overheat".

[0072] Furthermore, when determining the specific spatial coordinates of the heating point, a multi-element positioning algorithm is adopted. The propagation time delay of the echo signal of the same heating point received by different elements in the array ultrasonic probe is used to establish a set of spatial positioning equations, and the least squares method is used to solve the set of spatial positioning equations.

[0073] The spatial positioning equations are based on a spherical wave propagation model. The spherical equations are established with the three-dimensional coordinates of each array element as the center and the propagation distance of the ultrasonic signal from the heating point to the array element as the radius. The intersection of the spherical equations corresponding to multiple array elements is the spatial coordinate of the heating point.

[0074] Furthermore, the method also includes real-time display of the detection results of the hot spots. The detection results include the location distribution image of the hot spots, the temperature change curve, and the fault type identifier. When the temperature value of the hot spot exceeds the preset safety threshold or a serious overheating fault occurs, an audible and visual alarm signal is triggered. At the same time, the detection results and alarm information are stored in the database and sent to the remote monitoring terminal.

[0075] Specifically, based on a heating point temperature of 125℃ (moderate overheating, 80℃-120℃ is mild, 120℃-160℃ is moderate, and >160℃ is severe), a moderate overheating alarm is triggered:

[0076] Audible and visual alarm control: Sending control commands to the audible and visual alarm will activate the buzzer in a "short tone at 1-second interval" mode (i.e., a 1-second sound followed by a 1-second pause cycle), and the red alarm light will flash in a 1-second cycle (on for 1 second, off for 1 second). The alarm volume will be set to 85dB (ensuring clear hearing within 50m of the substation maintenance room), and the alarm light brightness will be set to maximum (200cd / m² for easy observation from a distance). Simultaneously, an alarm pop-up window will appear in the upper right corner of the display screen, showing "Medium overheat alarm: A 125℃ hot spot was detected in the winding area. The fault type is winding overheating. Please handle it promptly." The pop-up window uses a red background with white text and will remain displayed until the maintenance personnel click the "Confirm" button.

[0077] Alarm priority handling: If multiple overheating points are detected simultaneously (e.g., winding temperature 125℃, core temperature 90℃), alarm priority is determined according to temperature level—severe overheating (>160℃) has the highest priority, triggering a continuous long tone and a constantly lit alarm; moderate overheating (120℃-160℃) has a medium priority, triggering an alarm at 1-second intervals; and slight overheating (80℃-120℃) has a low priority, triggering an alarm at 2-second intervals. When a high-priority alarm is triggered, low-priority alarms are automatically paused, only resuming after the high-priority alarm is cleared, avoiding confusion for maintenance personnel due to multiple alarms overlapping. For example, if the winding temperature subsequently rises to 165℃ (severe overheating), the buzzer switches to a continuous long tone, the alarm light remains constantly lit, and the slight overheating alarm at 90℃ (core temperature) is paused until the winding temperature drops below 160℃.

[0078] Remote monitoring terminals (computers, mobile phones, tablets) are equipped with dedicated monitoring software (supporting Windows, Android, and iOS systems). The software automatically parses JSON data and displays it in a visual manner: ① Computer terminal: Displays a 3D model of the transformer (with marked hot spots), temperature change curves, and alarm logs, supporting data export (Excel / PDF format) and historical data query (filtered by time and fault type); ② Mobile terminal: Pushes alarm information in the form of pop-up windows, displaying a concise hot spot temperature and fault type, and supports clicking to jump to view detailed data (such as coordinates and handling suggestions).

[0079] This embodiment provides a detailed and practical operational solution covering the entire process from parameter calibration, signal processing, feature extraction, model application, alarm display, model updating, special operating condition adjustment, fault diagnosis, and effect verification. By employing array ultrasonic technology combined with machine learning algorithms, it effectively solves the pain points of traditional detection technologies, such as incomplete coverage, inaccurate positioning, poor real-time performance, and difficulty in classifying types. It achieves full coverage detection, precise positioning, real-time monitoring, and intelligent diagnosis of transformer heating points, significantly reducing operation and maintenance costs, improving fault handling efficiency, and providing reliable technical support for the safe and stable operation of transformers. It can be widely applied to transformer condition monitoring in power systems, industrial plants, and other fields.

[0080] The above are merely preferred embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for detecting a hot spot of a transformer based on array ultrasound, characterized in that, The application relates to a transformer overheating point detection method and device. According to the model, structural size and distribution characteristics of the to-be-detected area of the transformer, position calibration and parameter setting are performed on an array ultrasonic probe, the installation position, scanning angle range and ultrasonic signal transmission frequency, pulse width and transmission power parameters of the array ultrasonic probe are determined, the array ultrasonic probe can cover the key overheating risk area of the transformer, and the key overheating risk area includes the winding joint of the transformer, the core lamination joint, the connection part of the oil tank wall and the internal component and the interface between the sleeve and the oil tank; An excitation signal is sent to the array ultrasonic probe, the array ultrasonic probe transmits ultrasonic signals to the to-be-detected area of the transformer according to the set parameters, and the echo signals received by the array ultrasonic probe are collected in real time, wherein the echo signals contain reflection signals of different medium interfaces in the transformer and abnormal scattering signals caused by local medium characteristic changes due to overheating; The received echo signals are preprocessed to obtain effective echo signals, the effective echo signals are subjected to feature extraction, and a heating point detection model is constructed based on the extracted signal features, so as to establish the mapping relationship among the signal features, the temperature, position and fault type of the heating point; The real-time extracted signal features are input into the heating point detection model, and whether the to-be-detected area of the transformer has a heating point, the specific spatial coordinates of the heating point, the temperature value of the heating point and the type of the heating fault are output.

2. The array ultrasonic based transformer hot spot detection method of claim 1, wherein, The array ultrasonic probe adopts a multi-element linear arrangement structure, the number of array elements is 32-128, the center frequency of each array element is 1-10 MHz, the array element spacing is 0.5-2 mm, the shell of the array ultrasonic probe is made of a material resistant to high temperature, oil and electromagnetic shielding, the shell surface is provided with an anti-skid wear-resistant coating, the array ultrasonic probe and the transformer shell realize ultrasonic signal transmission through a coupling agent, and the coupling agent adopts a high-temperature-resistant silicon-based coupling agent.

3. The array ultrasonic based transformer hot spot detection method of claim 2, wherein, The position calibration and parameter setting of the array ultrasonic probe are performed by using a three-dimensional coordinate positioning system, the three-dimensional coordinate positioning system comprises a laser positioner, a displacement sensor and a coordinate calculation unit, the three-dimensional coordinates of a plurality of reference positioning points on the transformer shell are determined by the laser positioner, the distance and angle between the array ultrasonic probe and each reference positioning point are measured by the displacement sensor, and finally the actual three-dimensional coordinates of each array element of the array ultrasonic probe are calculated by the coordinate calculation unit according to the measurement data, and the actual three-dimensional coordinates are compared with the preset theoretical coordinates; if the coordinate deviation exceeds the preset deviation, the position of the array ultrasonic probe is adjusted through an electric adjustment mechanism until the coordinate deviation is less than the preset deviation.

4. The array ultrasonic based transformer hot spot detection method of claim 1, wherein, The received echo signals are preprocessed, including signal denoising, signal amplification, baseline correction and signal normalization processing. The signal denoising adopts a method combining wavelet transform denoising and adaptive filtering, wavelet decomposition is performed on the original echo signal, threshold processing is performed on the wavelet coefficients of each layer, a soft threshold function is used to denoise the high-frequency wavelet coefficients, the wavelet coefficients after denoising are reconstructed to obtain a signal after preliminary denoising; the signal after preliminary denoising is input into an adaptive filter to further eliminate residual noise and interference signals; The signal amplification adopts a low-noise operational amplifier, and the amplification multiple is adaptively adjusted according to the amplitude of the echo signal; The baseline correction adopts a polynomial fitting method, a baseline fitting curve is obtained by performing polynomial fitting on the baseline part of the echo signal, and the original echo signal is subtracted from the baseline fitting curve to realize baseline correction; The signal normalization processing adopts a maximum-minimum value normalization method, and the amplitude of the preprocessed signal is mapped to the range of 0-1 to eliminate the influence of signal amplitude difference under different detection conditions on subsequent feature extraction.

5. The array ultrasonic based transformer hot spot detection method of claim 1, wherein, The extracted features of the effective echo signal include amplitude variation characteristics, phase shift characteristics, signal propagation time delay characteristics and spectral distribution characteristics of the echo signal, wherein the amplitude variation characteristics are used to reflect the change of acoustic impedance caused by the temperature rise of the transformer internal medium, the phase shift characteristics and the signal propagation time delay characteristics are used to locate the spatial position of the heat point, and the spectral distribution characteristics are used to distinguish different types of heat faults.

6. The array ultrasonic based transformer hot spot detection method of claim 1, wherein, The heat point detection model adopts a machine learning algorithm, which is obtained by training a large number of transformer ultrasonic signal samples under known heat fault types and normal operating conditions; The machine learning algorithm adopts a fusion algorithm combining support vector machines and random forests, the extracted signal features are divided into a training set and a test set, the training set samples are subjected to feature standardization processing to eliminate the influence of the dimension difference between different features; then a support vector machine model and a random forest model are constructed respectively, the support vector machine model adopts a radial basis kernel function, and the penalty coefficient and the kernel function parameter are determined by a cross-validation method; the number of decision trees of the random forest model is set to 50-200, the maximum depth of each decision tree is set to 5-15 layers, and the Gini coefficient is used as the feature selection standard when the node is split; the training set samples are input into the support vector machine model and the random forest model for training to obtain two independent sub-models; finally, a fusion model is constructed, and the output results of the two sub-models are fused by weighted voting, and the weights of the support vector machine model and the random forest model are determined according to the accuracy of the two sub-models on the test set.

7. The array ultrasonic based transformer hot spot detection method of claim 1, wherein, When determining the specific spatial coordinates of the heat point, a multi-element positioning algorithm is adopted, the propagation time delay of the echo signals of the same heat point received by different array elements in the array ultrasonic probe is used to establish a spatial positioning equation set, and the least square method is used to solve the spatial positioning equation set; The spatial positioning equation set is based on a spherical wave propagation model, and a spherical equation is established with a three-dimensional coordinate of each array element as a spherical center and a propagation distance of an ultrasonic signal from a heat point to the array element as a radius. An intersection point of the spherical equations corresponding to multiple array elements is a spatial coordinate of the heat point.

8. The array ultrasonic based transformer hot spot detection method of claim 1, wherein, The method further includes displaying a detection result of the heat point in real time, the detection result including a position distribution image of the heat point, a temperature change curve, and a fault type identification, triggering an audible and light alarm signal when a temperature value of the heat point exceeds a preset safety threshold or a serious heat fault type is detected, and storing the detection result and alarm information into a database and sending the detection result and alarm information to a remote monitoring terminal.

Images