Transformer fault detection method

Abstract

According to the transformer fault detection method, vibration signals are collected in real time, transformer operation early-stage changes can be captured in time, data support is provided for fault early-stage early warning, frequency domain signals are converted into binarized images, then features are extracted, image pixel matrix features reflect overall distribution, and the fault detection accuracy is improved. The morphological features of the connected regions describe the shape structure of the fault region, the texture features of the gray-level co-occurrence matrix represent texture information, and the initial weak change of the fault is described in a multi-dimensional manner, so that the initial feature recognition accuracy is improved, and the initial effective monitoring and diagnosis of the fault are realized; meanwhile, real-time collection does not need power failure of a transformer, normal power supply of a power system is guaranteed, and continuous monitoring and real-time online monitoring are achieved; multiple features are integrated into a feature vector space to be input into a machine learning classifier, fault features can be comprehensively reflected from different angles, different types of faults can be accurately recognized and diagnosed, various typical faults are covered, and the coverage range and accuracy of fault diagnosis are effectively improved.

Description

Technical Field

[0001] This invention relates to the field of transformer fault diagnosis technology, and in particular to a method for detecting transformer faults. Background Technology

[0002] As a key piece of equipment in the power system, the operating status of transformers directly affects the stability and security of power supply. During long-term operation, transformers are affected by a variety of factors, such as electrical stress, mechanical vibration and thermal effects, which gradually lead to faults such as insulation aging, winding deformation and core loosening. If these faults are not detected and dealt with in time, they may cause serious power accidents, resulting in huge economic losses and social impact.

[0003] Currently, traditional fault diagnosis methods have many limitations. While traditional electrical parameter measurement methods can reflect transformer faults to some extent, they only provide qualitative references and are insufficient for effective monitoring and diagnosis in the early stages of a fault. Transformer bushing anomaly identification methods based on offline bridge methods require collecting dielectric loss factors while the transformer is running offline, making real-time online monitoring impossible. Data-driven methods are highly dependent on the sample size and completeness of the anomaly data; however, in practical engineering, transformer bushing anomaly data is scarce and fails to cover all typical faults, resulting in poor accuracy for anomaly identification models trained on small sample data. Summary of the Invention

[0004] In view of this, the present invention proposes a transformer fault detection method, which can effectively solve the shortcomings of the existing technology, such as difficulty in effective monitoring and diagnosis in the early stage of faults, inability to achieve real-time online monitoring, and difficulty in covering various typical faults.

[0005] The technical solution of this invention is implemented as follows: A transformer fault detection method, specifically including: Real-time acquisition of vibration signals during transformer operation to obtain raw vibration signals; The acquired raw vibration signals are preprocessed to obtain frequency domain signals; Image generation processing is performed based on the frequency domain signal to obtain a binarized image; Feature extraction is performed on the binarized image to obtain the image pixel matrix features, the morphological features of connected regions, and the gray-level co-occurrence matrix texture features; The image pixel matrix features, the morphological features of connected regions, and the gray-level co-occurrence matrix texture features are integrated into a feature vector space; the feature vector space is configured as the input of a machine learning classifier; the machine learning classifier is configured to identify the fault type of the transformer and output the final fault detection result.

[0006] As a further optional solution to the aforementioned transformer fault detection method, the preprocessing of the acquired raw vibration signal to obtain a frequency domain signal specifically includes: The collected raw vibration signal is filtered to obtain the filtered vibration signal; The filtered vibration signal is converted into a frequency domain signal based on the Fourier transform.

[0007] As a further optional solution to the aforementioned transformer fault detection method, the step of performing image generation processing based on the frequency domain signal to obtain a binarized image specifically includes: Generate a grayscale image based on the frequency domain signal; A threshold is set to binarize the grayscale image, resulting in a binarized image.

[0008] As a further optional solution to the aforementioned transformer fault detection method, the threshold is automatically calculated using the Otsu threshold segmentation algorithm, specifically including: Analyze the grayscale histogram of a grayscale image and calculate the probability of each grayscale level appearing. Iterate through all possible thresholds to divide the image into foreground and background categories; Calculate the inter-class variance between the two classes; Choose the threshold that maximizes the inter-class variance as the optimal segmentation threshold.

[0009] As a further optional solution to the aforementioned transformer fault detection method, the extraction of image pixel matrix features specifically includes: Mathematical morphology preprocessing is performed on the binarized image to obtain a preprocessed binary image. Based on the binary image after mathematical morphology preprocessing, the pixel matrix features of the image are extracted.

[0010] As a further optional solution to the transformer fault detection method, the extraction of the morphological features of the connected region specifically includes: Based on the binarized image, generate a normal-fault hyperbola deviation map; The normal-fault hyperbola deviation map is segmented by thresholding, and pixels with deviation values ​​exceeding a preset threshold are marked as foreground to form a binarized deviation region. Identify closed connected regions in binarized deviation regions based on connected component labeling algorithms; Based on closed connected regions, morphological features such as frequency parameters, morphological parameters, and image area ratio are extracted.

[0011] As a further optional solution to the aforementioned transformer fault detection method, the extraction of the gray-level co-occurrence matrix texture features specifically includes: Convert a binary image to a grayscale image; Based on the converted grayscale image, the joint grayscale distribution probability of pixel pairs is statistically analyzed in four preset directions to generate grayscale co-occurrence matrices in four directions. Based on the gray-level co-occurrence matrix in four directions, texture features such as energy, entropy, contrast, correlation, maximum probability, uniformity, and inverse difference moment are extracted.

[0012] A transformer fault detection system, comprising: The vibration signal acquisition module is used to acquire vibration signals during transformer operation in real time and obtain raw vibration signals. The vibration signal preprocessing module is used to preprocess the acquired raw vibration signal to obtain the frequency domain signal; The image generation module is used to perform image generation processing based on the frequency domain signal to obtain a binarized image. The multimodal feature extraction module is used to extract features from the binarized image to obtain the image pixel matrix features, the morphological features of connected regions, and the gray-level co-occurrence matrix texture features. The fault diagnosis module is used to integrate the image pixel matrix features, the morphological features of connected regions, and the gray-level co-occurrence matrix texture features into a feature vector space; the feature vector space is configured as the input of a machine learning classifier; the machine learning classifier is configured to identify the fault type of the transformer and output the final fault detection result.

[0013] A computing device includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of any of the above-described transformer fault detection methods.

[0014] A computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps of any of the above-described transformer fault detection methods.

[0015] The beneficial effects of this invention are as follows: By acquiring vibration signals in real time, early changes during transformer operation can be captured promptly, providing a data foundation for early fault warning. Simultaneously, after converting the frequency domain signal into a binary image, feature extraction is performed based on the image. The image pixel matrix features can reflect the overall distribution of the image, the morphological features of connected regions can describe the shape and structure of the fault area, and the gray-level co-occurrence matrix texture features can characterize the texture information of the image. These rich features can depict subtle changes in the early stages of a fault from multiple dimensions, improving the accuracy of early fault feature identification, thereby enabling effective monitoring and diagnosis in the early stages of a fault. Furthermore, real-time acquisition of vibration signals allows for the timely detection of early changes in transformer operation. Vibration signals generated during transformer operation do not require transformer shutdown and do not affect the normal power supply of the power system. This allows for continuous monitoring of the transformer's operating status, enabling timely detection of abnormal changes in equipment status and achieving real-time online monitoring. Furthermore, by integrating image pixel matrix features, morphological features of connected regions, and gray-level co-occurrence matrix texture features into a feature vector space and inputting it into a machine learning classifier, these features describe the information of the binary image generated based on the vibration signal from different perspectives. This comprehensively reflects the characteristics of various faults and can accurately identify and diagnose different types of faults, thus covering various typical faults and improving the coverage and accuracy of fault diagnosis. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0017] Figure 1 This is a flowchart illustrating a transformer fault detection method according to the present invention. Figure 2 This is a schematic diagram of the composition of a transformer fault detection system according to the present invention; Figure 3 This is a schematic diagram of the composition of a computing device according to the present invention. Detailed Implementation

[0018] The technical solutions in the embodiments of the present invention will be clearly and completely described below. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. 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.

[0019] refer to Figures 1 to 3A method for detecting transformer faults, specifically including: High-precision vibration sensors are used to set up multiple acquisition points in key parts of the transformer, such as the core and windings, to collect vibration signals during the transformer's operation in real time.

[0020] The acquired raw vibration signal is preprocessed to obtain a frequency domain signal; in some embodiments, the preprocessing of the acquired raw vibration signal to obtain a frequency domain signal specifically includes: The acquired raw vibration signal is filtered to obtain the filtered vibration signal. The filtering process uses a Butterworth bandpass filter with a passband range of 50Hz-2kHz to remove environmental noise. The filtered vibration signal is converted into a frequency domain signal based on the Fourier transform.

[0021] Specifically, high-precision vibration sensors are used, and multiple acquisition points are arranged in key parts such as the transformer core and windings. This arrangement can capture vibration information during transformer operation from all directions and multiple angles. The core and windings are the parts of the transformer that are more active in vibration and sensitive to faults during operation. Multiple acquisition points can obtain richer and more comprehensive vibration data, providing sufficient and accurate basic data for subsequent analysis, and helping to more accurately reflect the actual operating status of the transformer. Real-time acquisition of vibration signals during transformer operation enables maintenance personnel to promptly grasp the dynamic changes of the transformer during operation. Once abnormal vibration occurs, it can be quickly detected, providing the possibility to take timely measures to prevent the fault from worsening, effectively ensuring the safe and stable operation of the transformer. Using a Butterworth bandpass filter to filter the acquired raw vibration signal can remove high-frequency noise and interference components. In actual signal acquisition environments, various noise contaminations are inevitable, such as electromagnetic interference and mechanical vibration interference. Filtering can make the vibration signal purer, highlight useful feature information, improve the signal-to-noise ratio, and provide a more reliable signal source for subsequent frequency domain analysis. The transfer function of the nth-order Butterworth bandpass filter is: ; in, express n The transfer function of a Butterworth bandpass filter. s Denotes the complex variable in the Laplace transform. n This indicates the order of the Butterworth filter. It is a multiplication symbol. Is the Butterworth filter transfer function in s The extreme points of the domain; Convert the normalized low-pass prototype into a frequency map of a band-pass filter: ; in For the bandpass center frequency, and These are the lower and upper cutoff frequencies of the bandpass filter, respectively. The bilinear transformation from analog to digital is as follows: ; in, T Represents the sampling period. This represents the unit delay operator, which means that the signal is delayed by one sampling period in time; ; in, Indicates that the bandpass filter is in Z The transfer function of a domain. and These are the coefficients of the numerator and denominator polynomials, respectively, which can be selected appropriately. and The value of can be used to design bandpass filters that meet specific requirements; The filtered vibration signal is converted into a frequency domain signal based on the Fourier transform. The Fourier transform is an effective tool for converting time domain signals into frequency domain signals. Through this transformation, complex time-domain vibration signals can be decomposed into combinations of different frequency components. The frequency domain signal can more intuitively show the frequency distribution characteristics of the vibration signal. Different types of problems (such as loose windings, core faults, etc.) often show abnormalities in a specific frequency range. Therefore, frequency domain analysis helps to more accurately identify the fault type and degree of transformer.

[0022] Image generation processing based on frequency domain signals yields a binarized image; in some embodiments, this process specifically includes: Based on the frequency domain signal, a grayscale image is generated. The grayscale image is generated by normalizing the frequency domain signal data to the range of 0-255 and arranging it into a pixel image of a specific size according to row and column rules. A threshold is set to binarize the grayscale image, resulting in a binarized image containing only black and white pixel values.

[0023] Specifically, grayscale images are generated based on frequency domain signals. The frequency domain signal data is normalized to the 0-255 range and arranged into pixel images of a specific size according to row and column rules. This processing method transforms abstract frequency domain signals into intuitive visual images, allowing the characteristics of the signal to be presented in the form of images. For example, the intensity of different frequency components can be represented by the grayscale values ​​of different pixels in the grayscale image. Signal features that are difficult to observe intuitively in the frequency domain can be more clearly identified and analyzed in the grayscale image. Normalize to the 0-255 range using the following formula: ; in, x A single value in the frequency domain signal data. min ( X )and max ( X These represent the minimum and maximum values ​​of the frequency domain signal data, respectively. Norm ( X () represents the normalized data; Pixels are arranged according to fixed row and column rules, giving the image a regular structure that facilitates subsequent image processing and analysis. This standardized image generation method helps improve the efficiency and consistency of data processing and reduces analysis errors that may be caused by inconsistent data formats. By setting a threshold to binarize the grayscale image, a binary image containing only black and white pixel values ​​is obtained. Binarization can remove some subtle changes and noise interference in the grayscale image, highlighting the key information in the image. In transformer fault diagnosis, fault features may appear as some specific grayscale variation areas in the grayscale image. Through binarization, these regions with significant features can be highlighted in the form of black and white contrast, making the fault features more obvious and facilitating subsequent feature extraction and recognition. Binarized images contain only two pixel values, which significantly reduces the amount of data and data complexity compared to grayscale images. This not only saves storage space but also improves the efficiency of image processing. When performing feature extraction and machine learning classification, the simple structure of binarized images helps to extract effective features faster and reduces the amount of computation, making the entire fault diagnosis process more efficient. After image generation and binarization, the features in the image become clearer and easier to extract. For example, the morphological features of connected regions and the features of the image pixel matrix are more easily identified and quantified in the binarized image. This rich feature information provides more accurate and representative input for subsequent machine learning classifiers, which helps to improve the accuracy and reliability of fault diagnosis.

[0024] In some embodiments, the threshold is automatically calculated using the Otsu threshold segmentation algorithm, specifically including: Analyze the grayscale histogram of a grayscale image and calculate the probability of each grayscale level appearing. Iterate through all possible thresholds to divide the image into foreground and background categories; Calculate the inter-class variance between the two classes; Choose the threshold that maximizes the inter-class variance as the optimal segmentation threshold.

[0025] Specifically, the Otsu thresholding algorithm can automatically calculate the gray-level histogram of a grayscale image and the probability of each gray level. It then iterates through all possible thresholds to divide the image into foreground and background categories, and finally selects the threshold that maximizes the inter-class variance as the optimal segmentation threshold. This process requires no manual intervention, greatly improving the efficiency and automation of image segmentation. In transformer fault diagnosis, when faced with a large amount of image data, automatic threshold calculation can quickly complete image segmentation, reducing errors and time costs caused by manual operation. Since the algorithm determines the threshold based on the gray-level distribution characteristics of the image itself, it can adapt to grayscale images generated by transformer vibration signals under different conditions. No matter how the gray-level distribution of the image changes, the Otsu algorithm can find a relatively optimal segmentation threshold, ensuring the accuracy and stability of image segmentation. It is suitable for image analysis under various complex transformer operating conditions and fault conditions. By calculating the inter-class variance and selecting the threshold corresponding to the maximum inter-class variance, the Otsu algorithm can accurately distinguish between the foreground (which may represent the fault-related area) and the background in the image. In the binarized image of transformer fault diagnosis, this distinction helps to highlight the fault feature area, making subsequent feature extraction and analysis more focused on key information and improving the pertinence of fault diagnosis. Determining the threshold based on maximizing the inter-class variance can balance the separation degree between the foreground and background to a certain extent, reducing the possibility of misclassifying foreground pixels as background pixels or vice versa. This helps to preserve important details related to the fault in the image, avoids the loss of fault features or misidentification due to missegmentation, and thus improves the accuracy of fault diagnosis.

[0026] Feature extraction is performed on the binarized image to obtain image pixel matrix features, morphological features of connected regions, and gray-level co-occurrence matrix texture features; in some embodiments, the extraction of the image pixel matrix features specifically includes: A binary image is subjected to mathematical morphological preprocessing to obtain a binary image after mathematical morphological preprocessing. The preprocessing includes at least one morphological operation, which is selected from one or a combination of dilation, erosion, opening, closing, top-hat operation or black-hat operation, to eliminate image noise and optimize the contour of the target region. Based on the binary image after mathematical morphology preprocessing, the pixel matrix features are extracted, wherein the pixel matrix features include: The total number of pixels in the target region is used to quantify the area information of the target region; Statistical characteristics of pixel coordinates in the target region, wherein the statistical characteristics include at least one of centroid coordinates, centroid coordinates, and minimum bounding rectangle coordinates, used to characterize the spatial distribution of the target region; The topological features of the target region pixels, wherein the topological features include at least one of the number of holes, the number of connected components, and the Euler number, are used to describe the geometric structure of the target region.

[0027] Specifically, preprocessing a binarized image using at least one morphological operation, such as dilation, erosion, opening, closing, top-hat operation, or black-hat operation, can effectively eliminate noise in the image. In the binarized image generated from the transformer vibration signal, noise may originate from interference during signal acquisition, errors during image generation, and other factors. This noise can interfere with the true features in the image, while morphological operations can remove isolated noise points and smooth the boundaries of the target area, making the image clearer. Morphological operations can optimize the contour of a target region (which may be related to transformer faults). For example, dilation can expand the target region and connect adjacent small regions; erosion can shrink the target region and remove small protrusions. By reasonably selecting and combining these operations, the contour of the target region can be made more regular and clear, more accurately reflecting the manifestation of fault features in the image, which helps to improve the accuracy of feature extraction. Image pixel matrix features are extracted from the binarized image after mathematical morphology preprocessing. Since the preprocessing removes noise and optimizes the contour of the target region, the extracted pixel matrix features can more accurately reflect the real information of the image. The image pixel matrix features contain the distribution and arrangement information of pixels in the image. The preprocessed image makes these features more reflective of the essential features related to transformer faults and reduces the influence of noise and irregular contours on the features. The optimized image makes the pixel matrix features more representative. In transformer fault diagnosis, different types of faults may exhibit different pixel distribution patterns in the image. The preprocessed image can better highlight these patterns, enabling the extracted pixel matrix features to more effectively distinguish different fault types and provide more valuable feature information for subsequent fault diagnosis.

[0028] In some embodiments, the extraction of the morphological features of the connected regions specifically includes: Based on the binarized image, generate a normal-fault hyperbola deviation map; The normal-fault hyperbola deviation map is segmented by thresholding, and pixels with deviation values ​​exceeding a preset threshold are marked as foreground to form a binarized deviation region. Identify closed connected regions in binarized deviation regions based on connected component labeling algorithms; Based on closed connected regions, frequency parameters, morphological parameters, and image area ratio are extracted as morphological features. The frequency parameter is the frequency domain feature of pixel value oscillation within the connected region, including the dominant frequency component, spectral energy concentration, or high-frequency noise ratio, used to quantify the periodicity or randomness of the deviation. The morphological parameter is the geometric feature of the connected region, including the area of ​​the minimum bounding rectangle, aspect ratio, circularity, and the ratio of the convex hull area to the region area, used to describe the shape regularity of the deviation region. The image area ratio is the ratio of the area of ​​the connected region to the area of ​​the entire image, or the ratio of the area of ​​the connected region to the area of ​​a preset reference region, used to characterize the local severity of the deviation.

[0029] Specifically, generating a normal-fault hyperbolic deviation map based on the binarized image can present the differences in image features between the normal and fault states in an intuitive deviation map form. In transformer operation, there are feature differences in the binarized images generated by vibration signals under normal and fault states. The hyperbolic deviation map can highlight these differences, making the fault features more obvious and providing a more targeted data foundation for subsequent analysis. By generating deviation maps, the complex problem of image comparison is transformed into the analysis of deviation maps, which simplifies the feature analysis process. Deviation maps focus on the differences between normal and fault states, reducing the interference of irrelevant information and enabling researchers to focus more on feature changes related to faults. Threshold segmentation is performed on the hyperbolic deviation map, marking pixels with deviation values ​​exceeding a preset threshold as foreground, forming a binarized deviation region. This step accurately marks pixels in the image that differ significantly from the normal state. These pixels are likely related to transformer faults. By setting a reasonable threshold, abnormal regions can be effectively filtered out, improving the accuracy of fault feature localization. Binarization presents the deviation region in black and white, enhancing feature contrast and recognizability. In subsequent connected component identification and feature extraction processes, this clear binarized representation helps to more accurately identify connected components and reduce recognition errors caused by complex grayscale variations. It should be noted that the image pixel matrix similarity between the normal state curve and the fault state curve is obtained using the image pixel matrix similarity calculation formula. The specific calculation formula is as follows: ; in, and Representing the first and second elements in matrices A and B respectively under normal and fault conditions. i line, numberj The column elements, `A` and `B`, represent the average values ​​of their respective matrix elements. By identifying closed connected regions in the binarized deviation region based on the connected component labeling algorithm, the region where a fault may exist can be accurately defined. Closed connected regions have a certain degree of integrity and independence in the image. By identifying these regions, the fault features can be limited to a specific range, avoiding feature dispersion and confusion, and providing a clear target region for subsequent morphological feature extraction. After clearly defining the closed connected regions, the feature extraction process can be carried out more targetedly. Only the closed connected regions need to be analyzed, without the need for comprehensive feature extraction of the entire image, which greatly improves the efficiency of feature extraction and reduces the amount of computation and processing time. Based on the extraction of morphological features such as frequency parameters, shape parameters, and image area ratio of closed connected regions, the characteristics of the fault area can be comprehensively characterized from multiple perspectives. The frequency parameter can reflect the frequency of occurrence of fault features in the image, the shape parameter can describe the shape and structure of the fault area, and the image area ratio can reflect the proportion of the fault area in the overall image. These features complement each other and provide rich information for accurately judging the fault type and degree of transformer, thereby improving the accuracy and reliability of fault diagnosis.

[0030] In some embodiments, the extraction of the gray-level co-occurrence matrix texture features specifically includes: Convert a binary image to a grayscale image; Based on the converted grayscale image, the joint grayscale distribution probability of pixel pairs is statistically analyzed in four directions: 0°, 45°, 90° and 135°, to generate grayscale co-occurrence matrices in four directions; Based on the gray-level co-occurrence matrix in four directions, texture features such as energy, entropy, contrast, correlation, maximum probability, uniformity, and inverse difference moment are extracted.

[0031] Specifically, based on the converted grayscale image, the joint grayscale probability distribution of pixel pairs is statistically analyzed in four directions: 0°, 45°, 90°, and 135°, generating grayscale co-occurrence matrices in four directions. This multi-directional analysis method can comprehensively capture the texture features of the image in different directions. In transformer fault diagnosis, faults may exhibit different texture changes in different directions of the image. Multi-directional analysis can more accurately identify these changes and avoid missing important features due to single-directional analysis. Statistical analysis of the joint grayscale probability distribution of pixel pairs can accurately describe the spatial relationship and grayscale change pattern between pixels in the image. The grayscale co-occurrence matrix quantifies the texture information of the image in this way. Based on the gray-level co-occurrence matrices in four directions, texture features of energy, entropy, contrast, correlation, maximum probability, uniformity, and inverse moment are extracted. These features describe the texture characteristics of the image from different perspectives: energy reflects the uniformity of the image texture; entropy reflects the complexity of the image texture; contrast describes the difference between bright and dark areas in the image; correlation represents the linear correlation between pixels in the image; maximum probability reflects the most common pixel pair gray-level combination in the image; uniformity measures the regularity of the image texture; and inverse moment is related to the local uniformity of the image. Comprehensive and accurate texture feature extraction helps to more accurately identify the type and extent of transformer faults. Different faults may exhibit different combinations of features on image textures. By extracting and analyzing multiple texture features, machine learning classifiers can better learn fault feature patterns, thereby improving the accuracy and reliability of fault diagnosis.

[0032] The extracted image pixel matrix features, morphological features of connected regions, and gray-level co-occurrence matrix texture features are integrated into a feature vector space, which is then input into a machine learning classifier. The machine learning classifier identifies and diagnoses the fault types of the transformer and outputs the final fault diagnosis result.

[0033] Specifically, machine learning classifiers have powerful learning and classification capabilities. By inputting the integrated feature vector space into the classifier, it can learn the feature patterns corresponding to different fault types. Since the integrated features contain multifaceted information, the classifier can more comprehensively identify fault features, thereby improving the accuracy of transformer fault type identification and reducing misdiagnosis and missed diagnosis. Machine learning classifiers can learn and optimize through a large amount of training data. In practical applications, when facing transformer fault diagnosis under different working conditions and environments, the classifier can adaptively adjust according to the learned feature patterns, better adapt to various complex actual situations, and improve the reliability and stability of diagnosis. Fault diagnosis was automated by utilizing a machine learning classifier. Once the feature vector is input into the classifier, the fault diagnosis results can be output quickly without the need for complex manual feature analysis and judgment, greatly improving diagnostic efficiency and enabling timely provision of transformer fault information to maintenance personnel so that appropriate measures can be taken.

[0034] A transformer fault detection system, comprising: The vibration signal acquisition module is used to acquire vibration signals during transformer operation in real time and obtain raw vibration signals. The vibration signal preprocessing module is used to preprocess the acquired raw vibration signal to obtain the frequency domain signal; The image generation module is used to perform image generation processing based on the frequency domain signal to obtain a binarized image. The multimodal feature extraction module is used to extract features from the binarized image to obtain the image pixel matrix features, the morphological features of connected regions, and the gray-level co-occurrence matrix texture features. The fault diagnosis module integrates the extracted image pixel matrix features, the morphological features of connected regions, and the gray-level co-occurrence matrix texture features into a feature vector space, which is then input into a machine learning classifier. The machine learning classifier identifies and diagnoses the fault type of the transformer and outputs the final fault diagnosis result.

[0035] A computing device includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of any of the above-described transformer fault detection methods.

[0036] A computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps of any of the above-described transformer fault detection methods.

[0037] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for detecting transformer faults, characterized in that, The method is based on vibration signals and image binarization, and specifically includes: Real-time acquisition of vibration signals during transformer operation to obtain raw vibration signals; The acquired raw vibration signals are preprocessed to obtain frequency domain signals; Image generation processing is performed based on the frequency domain signal to obtain a binarized image; Feature extraction is performed on the binarized image to obtain the image pixel matrix features, the morphological features of connected regions, and the gray-level co-occurrence matrix texture features; The image pixel matrix features, the morphological features of connected regions, and the gray-level co-occurrence matrix texture features are integrated into a feature vector space; the feature vector space is configured as the input of a machine learning classifier; the machine learning classifier is configured to identify the fault type of the transformer and output the final fault detection result.

2. The transformer fault detection method according to claim 1, characterized in that, The preprocessing of the acquired raw vibration signal to obtain the frequency domain signal specifically includes: The collected raw vibration signal is filtered to obtain the filtered vibration signal; The filtered vibration signal is converted into a frequency domain signal based on the Fourier transform.

3. The transformer fault detection method according to claim 1, characterized in that, The image generation process based on the frequency domain signal to obtain a binarized image specifically includes: Generate a grayscale image based on the frequency domain signal; A threshold is set to binarize the grayscale image, resulting in a binarized image.

4. The transformer fault detection method according to claim 3, characterized in that, The threshold is automatically calculated using the Otsu threshold segmentation algorithm, specifically including: Analyze the grayscale histogram of a grayscale image and calculate the probability of each grayscale level appearing. Iterate through all possible thresholds to divide the image into foreground and background categories; Calculate the inter-class variance between the two classes; Choose the threshold that maximizes the inter-class variance as the optimal segmentation threshold.

5. The transformer fault detection method according to claim 1, characterized in that, The extraction of the image pixel matrix features specifically includes: Mathematical morphology preprocessing is performed on the binarized image to obtain a preprocessed binary image. Based on the binary image after mathematical morphology preprocessing, the pixel matrix features of the image are extracted.

6. The transformer fault detection method according to claim 1, characterized in that, The extraction of the morphological features of the connected regions specifically includes: Based on the binarized image, generate a normal-fault hyperbola deviation map; The normal-fault hyperbola deviation map is segmented by thresholding, and pixels with deviation values ​​exceeding a preset threshold are marked as foreground to form a binarized deviation region. Identify closed connected regions in binarized deviation regions based on connected component labeling algorithms; Based on closed connected regions, morphological features such as frequency parameters, morphological parameters, and image area ratio are extracted.

7. The transformer fault detection method according to claim 1, characterized in that, The extraction of the gray-level co-occurrence matrix texture features specifically includes: Convert a binary image to a grayscale image; Based on the converted grayscale image, the joint grayscale distribution probability of pixel pairs is statistically analyzed in four preset directions to generate grayscale co-occurrence matrices in four directions. Based on the gray-level co-occurrence matrix in four directions, texture features such as energy, entropy, contrast, correlation, maximum probability, uniformity, and inverse difference moment are extracted.

8. A transformer fault detection system, characterized in that, include: The vibration signal acquisition module is used to acquire vibration signals during transformer operation in real time and obtain raw vibration signals. The vibration signal preprocessing module is used to preprocess the acquired raw vibration signal to obtain the frequency domain signal; The image generation module is used to perform image generation processing based on the frequency domain signal to obtain a binarized image. The multimodal feature extraction module is used to extract features from the binarized image to obtain the image pixel matrix features, the morphological features of connected regions, and the gray-level co-occurrence matrix texture features. The fault diagnosis module is used to integrate the image pixel matrix features, the morphological features of connected regions, and the gray-level co-occurrence matrix texture features into a feature vector space. The feature vector space is configured as the input to a machine learning classifier; the machine learning classifier is configured to identify the fault type of the transformer and output the final fault detection result.

9. A computing device, characterized in that, The method includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the transformer fault detection method according to any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, The storage medium stores a computer program, which, when executed by a processor, implements the steps of the transformer fault detection method according to any one of claims 1-7.

Images