Transformer fault diagnosis method and system based on photoacoustic signal reconstruction and sensitive field optimization

Abstract

The present invention discloses a transformer fault diagnosis method and system based on the optimization of sensitive fields reconstructed by photoacoustic signals. Specifically, the method comprises the following steps: first, obtaining a multivariate sequence matrix of signals by monitoring photoacoustic signals of multiple gases in transformer oil samples; second, mapping the multivariate sequence matrix in the form of scattered points in a polar coordinate diagram by the idea of reconstructing sensitive fields; then extracting shallow local features, and clustering the features into multiple mutually exclusive categories, with the center of each category as a visual representation factor to form a visual feature table; finally, realizing the classification of transformer fault diagnosis based on an improved autonomous learning multi-space (ALMMo‑0) classifier. The method of the present invention reconstructs the multivariate photoacoustic signal reflecting the gas concentration into a sensitive field rather than in the form of a concentration threshold or ratio, so as to better reveal the diagnostic features hidden in the signal, and realize deep feature extraction and encoding of the image at the same time, thereby improving the transparency and interpretability of the model, and automatically updating the model parameters by data-driven without the need for manual adjustment of hyperparameters. The method is of great significance for improving the accuracy and practicality of transformer fault diagnosis.

Description

Technical Field

[0001] The present invention belongs to the field of transformer fault diagnosis, and in particular relates to a transformer fault diagnosis method and system based on photoacoustic signal reconstruction and sensitive field optimization. Background Art

[0002] Transformers play a central role in power systems and are crucial for ensuring stable operation. A transformer failure can not only damage related power facilities but also impact the overall safety and stability of the power grid. In severe cases, it can even cause large-scale power outages with widespread impact. Therefore, transformer condition monitoring and evaluation using fault diagnosis algorithms can predict transformer health, thereby improving the reliability, stability, and safety of power systems.

[0003] Currently, dissolved gas analysis (DGA) is the primary method for transformer operating status estimation and fault diagnosis. This technique focuses on detecting gases such as CO, CO, CH₄, C₂H₄, C₂H₂, and CH₂ dissolved in transformer oil. Common detection methods include gas chromatography (GC), photoacoustic spectroscopy (PAS), Raman spectroscopy (RS), Fourier transform infrared spectroscopy (FTIRS), and fiber grating (FBG). Gas chromatography is a commonly used method for transformer fault detection in power systems. However, gas chromatography requires regular calibration and is cumbersome to operate and has a long detection cycle, making it difficult to quickly and timely diagnose transformer faults. The PAS method uses the photoacoustic effect to obtain the gas concentration by collecting and processing photoacoustic signals. It has the characteristics of fast detection speed, high sensitivity and stable performance. The RS method analyzes the gas in transformer oil based on the Raman scattering principle. It has the characteristics of high scalability but low detection sensitivity. The FTIRS method analyzes the gas in transformer oil based on Fourier transform infrared spectroscopy. It does not require carrier gas and is maintenance-free, but is easily affected by interfering gases and the spectral baseline is prone to distortion or drift. FBG detects dissolved gas in oil based on optical fiber sensors. It has the characteristics of small size, good coupling, polarization insensitivity, easy maintenance and wide bandwidth range. However, optical fiber only measures narrow bands. Measuring multi-component gases requires multiple different optical fiber Bragg gratings, which makes engineering applications difficult. Summary of the Invention

[0004] To achieve the above objectives, the present invention proposes a transformer fault diagnosis method and system based on photoacoustic signal reconstruction and sensitive field optimization. The multivariate photoacoustic signal reflecting the gas concentration is reconstructed into a sensitive field rather than in the form of a concentration threshold or ratio, so as to better reveal the diagnostic features hidden in the signal. At the same time, by realizing deep feature extraction and encoding of the image, it helps to improve the transparency and interpretability of the model. How to reduce the magnitude of diagnostic process parameters is of great significance to improving the accuracy and practicality of transformer fault diagnosis.

[0005] The first aspect is a transformer fault diagnosis method based on photoacoustic signal reconstruction and sensitive field optimization. Specifically, it includes:

[0006] Step 1: Obtain a multivariate sequence matrix of signals by monitoring the photoacoustic signals of multiple gases in the transformer oil sample;

[0007] Step 2: Map the multivariate sequence matrix in the form of scattered points in the polar coordinate diagram by reconstructing the sensitive field;

[0008] Step 3: Extract shallow local features and cluster them into multiple mutually exclusive categories. The shortest probability entropy of each category center is used as the visual representation factor to form a visual feature vector.

[0009] Step 4: Classify transformer fault diagnosis based on autonomous learning multi-space (ALMMo-0) classifier.

[0010] Furthermore, in the transformer fault diagnosis method based on photoacoustic signal reconstruction and sensitive field optimization, the main contents of step 1 are:

[0011] A small amount of oil sample is extracted from the transformer. The dissolved gas in the oil sample is converted into a gas sample through the oil-gas separation module. The modulated laser irradiates the gas sample at a specific wavelength (or band), causing slight changes in local temperature and pressure, thereby generating an acoustic signal sequence. The acoustic wave sequences generated by multiple groups of dissolved gases are photoacoustic signal sequences, which in turn form a multivariate sequence matrix V M N , M is the number of gas types, and N is the length of the sequence sample.

[0012] Furthermore, the transformer fault diagnosis method based on photoacoustic signal reconstruction of sensitive field optimization, the idea of reconstructing the sensitive field in step 2, specifically includes the following.

[0013] Define xmax and xmin as V i N , i=1,2,…,the maximum and minimum values of M, for each time point i, the corresponding amplitude is xi, the hysteresis coefficient a and the angle offset θ l , and the amplification factor ζ. The specific reconstruction relationship is defined by the following formula:

[0014] The polar radius r(i) is expressed as:

[0015]

[0016] The angular offset θ(i) is defined as:

[0017]

[0018] The reflection angle φ(i) is defined as:

[0019]

[0020] Argument offset θ l Defined as:

[0021]

[0022] The value range of l is from 0 to L-1, where L represents the total number of planes in the sensitive field. By adjusting the amplification factor ζ and the hysteresis coefficient a, any point in the multivariate signal matrix can be converted into an intuitive representation in polar coordinates to form a reconstructed sensitive field image.

[0023] Furthermore, in the transformer fault diagnosis method based on photoacoustic signal reconstruction sensitive field optimization, the extraction of shallow local features in step 3 is specifically performed by using SURF features and FAST features to extract two shallow features.

[0024] Furthermore, the transformer fault diagnosis method based on photoacoustic signal reconstruction sensitive field optimization also includes the step 3.

[0025] Use Kmeans++ to cluster the two shallow features into Kc clusters. The shortest probability entropy of each cluster is calculated by the following formula and used as the representation factor of the cluster center. Finally, the shallow features are encoded into a feature vector with a dimension of 1*Kc.

[0026]

[0027] Among them, G is the cluster index, g is the feature point index in the Gth cluster, D min is the shortest distance between the feature point and the cluster center within the cluster, and D(g) is the distance between the g-th feature point and the cluster center.

[0028] Furthermore, in the transformer fault diagnosis method based on photoacoustic signal reconstruction sensitive field optimization, the specific content of the autonomous learning multi-space (ALMMo-0) classifier in step 4 is:

[0029] Define {g1,g2,g3,…,g τ} is the feature vector set with a dimension of 1*Kc and is used as the training data set. The number of transformer fault categories is set to q, that is, the number of subspaces, and each subspace corresponds to a set of feature classes.

[0030] First, a single eigenvector is selected to form a preliminary set of subspaces. In the training phase, {g1, g2, g3, …, g τ The individual feature vectors of the set are input into the autonomous learning multi-space (ALMMo-0) classifier for training. τ The similarity between it and each subspace is calculated according to the following formula.

[0031]

[0032]

[0033]

[0034] in is the average value of all eigenvectors in the i-th subspace, is the average scalar product of the i-th subspace, τ is the sample index of the eigenvector set, ‖‖ is the norm calculation, g ini is a single eigenvector of the preliminary set.

[0035] Furthermore, the transformer fault diagnosis method based on photoacoustic signal reconstruction sensitive field optimization and the autonomous learning multi-space (ALMMo-0) classifier also include.

[0036] For the input sample g τ , the classifier will calculate q R(g i τ ) value. τ Distribution and maximum R(g i τ ) and fill it into the corresponding subspace for the next round of training. After the training is completed, the feature vector of the collected samples can be tested in real time. The allocation result can be expressed as:

[0037] Class i =argmax i＝1,2,3,...,q (R i τ )

[0038] The second aspect is the transformer fault diagnosis system based on photoacoustic signal reconstruction and sensitive field optimization, which includes:

[0039] Oil-gas separation module, photoacoustic spectroscopy detection sensor, edge detection device, and transformer fault diagnosis service platform.

[0040] The edge detection device adopts a modular design, and the hardware specifically includes a communication unit, a main control unit, a positioning unit, a power supply unit, and a storage unit. The communication unit supports WiFi and 4G wireless communications, mainly realizing communication with the transformer fault diagnosis service platform, and supports RS232 and RS485 wired communications, mainly realizing data communication between the photoacoustic spectroscopy detection sensor. The main control unit contains the V7-690T FPGA+D2000-CPU, and the positioning unit mainly realizes the position positioning of the edge detection device, supporting GPS and Beidou positioning. The power supply unit provides power to each module and unit of the device, and realizes functions such as charging the backup battery and power detection. The storage unit supports SD card or SSD hard drive expansion, which is convenient for storing historical data on the edge side.

[0041] The software design includes an operating system, a language environment, data preprocessing, edge computing, and edge-to-service collaboration. Using Python, C, and C++, the edge detection device developed a data preprocessing algorithm and an edge computing model for transformer fault diagnosis. This collaborative service for transformer fault diagnosis includes functions such as storing edge computing results, interacting with device data and service platform data, and updating the edge computing model.

[0042] The oil-gas separation module is connected to the transformer tank to separate the oil and gas in the transformer tank. This sample gas, serving as the dissolved gas in the transformer oil, is connected to a photoacoustic spectroscopy sensor via an inlet and outlet valve. The edge detection device controls the photoacoustic spectroscopy sensor to detect the photoacoustic signal of the sample gas. The transformer fault diagnosis model running on the edge detection device enables real-time transformer fault diagnosis at the edge, and the calculated results are uploaded to the transformer fault diagnosis service platform. The platform enables viewing transformer status, managing edge detection devices and sensors, and updating edge computing models. Operations and maintenance personnel can view the transformer's status via mobile terminals. When a transformer fault warning occurs, an alarm message is sent to them, allowing them to quickly take preventive measures.

[0043] The beneficial effects of the present invention are as follows:

[0044] Reconstructing multivariate photoacoustic signals reflecting gas concentrations as sensitive fields rather than concentration thresholds or ratios better reveals diagnostic features hidden in the signals. To address the weak diagnostic capabilities of high-dimensional, shallow features, we proposed an image feature encoding method based on the shortest probability entropy, forming a visual feature table to achieve deep feature extraction and encoding of images. Finally, we used a data-driven multi-space model to handle complex parameter adjustments under different operating conditions, improving the model's transparency and interpretability. Automatically updating model parameters through data-driven analysis eliminates the need for manual hyperparameter adjustment, reducing significant training and optimization costs. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] Figure 1 Flow chart of the method of the present invention

[0046] Figure 2 The invention is sensitive to the reconstruction of the photoacoustic signal

[0047] Figure 3 The local feature point clustering result of the present invention (Kc=5)

[0048] Figure 4 This is a system framework diagram of the present invention.

[0049] Specific implementation cases

[0050] The transformer fault diagnosis method and system based on photoacoustic signal reconstruction and sensitive field optimization described in the present invention will be further explained and illustrated below in conjunction with the drawings and specific embodiments of the specification. However, such explanation and illustration do not constitute an undue limitation to the technical solution of the present invention.

[0051] Transformer fault diagnosis using photoacoustic spectroscopy (PAS) is primarily based on the analysis of dissolved gases in transformer oil. Faults within the transformer, such as local overheating, discharge, and insulation aging, generate gases of specific types and proportions, which then dissolve in the transformer's insulating oil. By monitoring the types and concentrations of these gases, the transformer's health and potential faults can be indirectly assessed.

[0052] Currently, by analyzing the photoacoustic signals of different gases (such as methane CH4, ethane C2H6, ethylene C2H4, acetylene C2H2, carbon monoxide CO, carbon dioxide CO2, etc.), their respective concentrations can be determined. The types and ratios of these gases are closely related to specific types of transformer faults, forming the so-called "fault characteristic gas ratios", such as the three-ratio method or the five-ratio method. However, since the potential characteristics of the fault are often limited by intuitive thresholds and concentration ratios, they are often limited to the actual environment and application scenarios. It should be noted that in multiple tests of local overheating, arc discharge and other faults, it was found that the existing methods will have delayed alarms in actual applications.

[0053] Reconstructing multivariate photoacoustic signals reflecting gas concentrations as sensitive fields rather than concentration thresholds or ratios better reveals diagnostic features hidden in the signals. To address the weak diagnostic capabilities of high-dimensional, shallow features, we proposed an image feature encoding method based on the shortest probability entropy, forming a visual feature table to achieve deep feature extraction and encoding of images. Finally, we used a data-driven multi-space model to handle complex parameter adjustments under different operating conditions, improving the model's transparency and interpretability. Automatically updating model parameters through data-driven analysis eliminates the need for manual hyperparameter adjustment, reducing significant training and optimization costs.

[0054] As attached Figure 1 As shown, the transformer fault diagnosis method based on photoacoustic signal reconstruction and sensitive field optimization according to the present invention may include the following steps:

[0055] Step 1: Obtain a multivariate sequence matrix of signals by monitoring the photoacoustic signals of multiple gases in the transformer oil sample;

[0056] 1.1 Extract a small amount of oil sample from the transformer and separate the dissolved gas in the oil sample through the oil-gas separation module to form a gas sample;

[0057] 1.2 Use modulated laser to irradiate the gas sample at a specific wavelength (or band), stimulating small changes in local temperature and pressure, thereby generating a sequence of acoustic wave signals.

[0058] 1.3 The acoustic wave sequences generated by multiple groups of dissolved gases are photoacoustic signal sequences, which in turn form a multivariate sequence matrix V M N , where M is the number of gas types and N is the length of the sequence sample.

[0059] Step 2: Map the multivariate sequence matrix in the form of scattered points in the polar coordinate diagram by reconstructing the sensitive field;

[0060] 2.1 Define xmax and xmin as V i N , i=1,2,…,the maximum and minimum values of M;

[0061] 2.2 For each time point i, according to its amplitude xi, hysteresis coefficient a, and angle offset θ l , and the amplification factor ζ, through the following specific reconstruction relationship, any point in the multivariate signal matrix is converted into an intuitive representation in polar coordinates to form a reconstructed sensitive field image:

[0062] The polar radius r(i) is expressed as:

[0063]

[0064] The angular offset θ(i) is defined as:

[0065]

[0066] The reflection angle φ(i) is defined as:

[0067]

[0068] Argument offset θ l Defined as:

[0069]

[0070] The value of l ranges from 0 to L-1, where L represents the total number of planes in the sensitive field;

[0071] 2.3 Optimize the representation effect of the sensitive field image by adjusting the amplification factor ζ and the hysteresis coefficient a.

[0072] Step 3: Extract shallow local features and cluster them into multiple mutually exclusive categories. The shortest probability entropy of each category center is used as the visual representation factor to form a visual feature vector.

[0073] Step 4: Classify transformer fault diagnosis based on autonomous learning multi-space (ALMMo-0) classifier.

[0074] In step 3, two shallow features are extracted using SURF features and FAST features. It should be noted that in local feature extraction, feature detection and description are separable. First, the location information of the feature is detected, and then it is characterized using a digital description.

[0075] The steps of SURF to extract local features are:

[0076] First, the integral image (Integral Image) and the Gaussian difference integral image (DoG) of the image are calculated.

[0077] Potential keypoints are detected using DoG images in Gaussian scale space. This involves comparing DoG images of adjacent scales at different scales to find local extreme points, or keypoints. Candidate keypoints are then thresholded and removed to ensure keypoint stability.

[0078] The key points detected initially are precisely positioned and their scales are fine-tuned. The principal curvature of the Hessian matrix is usually used to optimize the position and scale of the key points.

[0079] The steps of FAST to extract local features are:

[0080] Select an initial pixel point as a candidate corner point and set a threshold constant T (usually a fixed value of the brightness difference of the sensitive field image).

[0081] For each candidate pixel, check the brightness difference of a circle around the pixel (usually 16 consecutive pixels, forming a ring with a radius of 3, but excluding the central pixel itself). Specifically, check whether there are N consecutive pixels (usually N = 12 consecutive pixels) in this ring that are at least T units brighter (or darker) than the central pixel.

[0082] If such continuous pixels are found in the clockwise or counterclockwise direction, it is considered a corner candidate. If continuous pixels that meet the conditions can be found in both directions, the direction with more continuous pixels is selected.

[0083] In order to avoid repeated detection and improve the quality of corner points, it is necessary to perform non-maximum suppression around the detected corner points and remove those feature points that are not the most significant.

[0084] After extracting local features, Kmeans++ is used to cluster the two shallow features into Kc clusters. The shortest probability entropy of each cluster is calculated by the following formula and used as the representation factor of the cluster center. Finally, the shallow features are encoded into a feature vector with a dimension of 1*Kc.

[0085]

[0086] Among them, G is the cluster index, g is the feature point index in the Gth cluster, D min is the shortest distance between the feature point and the cluster center within the cluster, and D(g) is the distance between the g-th feature point and the cluster center.

[0087] In step 4, first define {g1,g2,g3,…,g τ} is the feature vector set with a dimension of 1*Kc and is used as the training data set. The number of transformer fault categories is set to q, that is, the number of subspaces, and each subspace corresponds to a set of feature classes.

[0088] Select a single eigenvector to form a preliminary set of subspaces, and in the training phase, {g1,g2,g3,…,g τ The individual feature vectors of the set are input into the autonomous learning multi-space (ALMMo-0) classifier for training. τ The similarity between it and each subspace is calculated according to the following formula.

[0089]

[0090]

[0091]

[0092] in is the average value of all eigenvectors in the i-th subspace, is the average scalar product of the i-th subspace, τ is the sample index of the eigenvector set, ‖‖ is the norm calculation, g ini is a single eigenvector of the preliminary set.

[0093] For the input sample g τ , the classifier will calculate q R(g i τ ) value. τ Distribution and maximum R(g i τ ) and fill it into the corresponding subspace for the next round of training. After the training is completed, the feature vectors of the collected samples can be tested and diagnosed in real time. The allocation result can be expressed as:

[0094] Class i =argmax i＝1,2,3,...,q (R i τ )

[0095] At the same time, the present invention designs a transformer fault diagnosis system based on photoacoustic signal reconstruction sensitive field optimization, including:

[0096] Oil-gas separation module, photoacoustic spectroscopy detection sensor, edge detection device, and transformer fault diagnosis service platform.

[0097] The edge detection device adopts a modular design, and the hardware specifically includes a communication unit, a main control unit, a positioning unit, a power supply unit, and a storage unit. The communication unit supports WiFi and 4G wireless communications, mainly realizing communication with the transformer fault diagnosis service platform, and supports RS232 and RS485 wired communications, mainly realizing data communication between the photoacoustic spectroscopy detection sensor. The main control unit contains the V7-690T FPGA+D2000-CPU, and the positioning unit mainly realizes the position positioning of the edge detection device, supporting GPS and Beidou positioning. The power supply unit provides power to each module and unit of the device, and realizes functions such as charging the backup battery and power detection. The storage unit supports SD card or SSD hard drive expansion, which is convenient for storing historical data on the edge side.

[0098] The software design includes an operating system, a language environment, data preprocessing, edge computing, and edge-to-service collaboration. Using Python, C, and C++, the edge detection device developed a data preprocessing algorithm and an edge computing model for transformer fault diagnosis. This collaborative service for transformer fault diagnosis includes functions such as storing edge computing results, interacting with device data and service platform data, and updating the edge computing model.

[0099] The oil-gas separation module is connected to the transformer tank to separate the oil and gas in the transformer tank. This sample gas, serving as the dissolved gas in the transformer oil, is connected to a photoacoustic spectroscopy sensor via an inlet and outlet valve. The edge detection device controls the photoacoustic spectroscopy sensor to detect the photoacoustic signal of the sample gas. The transformer fault diagnosis model running on the edge detection device enables real-time transformer fault diagnosis at the edge, and the calculated results are uploaded to the transformer fault diagnosis service platform. The platform enables viewing transformer status, managing edge detection devices and sensors, and updating edge computing models. Operations and maintenance personnel can view the transformer's status via mobile terminals. When a transformer fault warning occurs, an alarm message is sent to them, allowing them to quickly take preventive measures.

[0100] In order to better illustrate the application of the transformer fault diagnosis method based on photoacoustic signal reconstruction and sensitive field optimization described in the present invention, a simulation test is carried out for further explanation.

[0101] Table 1 Diagnostic results

[0102] Fault type Training set (feature vector) Test set (feature vector) Accuracy (Kc=10) Normal state 45 8 7 / 8 High-energy discharge 60 7 7 / 7 Low-energy discharge 60 7 7 / 7 Partial discharge 50 7 7 / 7 High temperature overheating 40 7 6 / 7 total 255 36 94.44％

[0103] Table 1 shows that the transformer fault diagnosis method based on photoacoustic signal reconstruction and sensitive field optimization has an accuracy rate of 92.3%. After performing the transformer fault diagnosis, the edge detection device sends the diagnosis results to the transformer status monitoring cloud platform, allowing operators to view the transformer status in real time at the monitoring center.

[0104] Table 2 Comparison of different diagnostic algorithms (Kc)

[0105] method Kc=5 Kc=8 Kc=10 Probabilistic Entropy-KNN 77.2％ 80.7％ 82.6％ Probabilistic Entropy-SVM 84.3％ 85.3％ 88.7％ Probabilistic Entropy-LSTM 86.2％ 81.5％ 91.9％ Probabilistic Entropy-ALMMo-0 89.9％ 92.7％ 94.4％

[0106] It can be seen from Table 2 that when Kc=10, each method achieves the best performance and the method of the present invention has the highest accuracy.

[0107] As attached Figure 2 As shown in the figure, it can be seen that the sensitive field images reconstructed from the photoacoustic signals collected from different faults are quite different and suitable for subsequent feature recognition.

[0108] As attached Figure 3 As shown, it can be seen that a cluster of Kc=5 is obtained by clustering the feature points.

[0109] As attached Figure 4 As shown, the transformer fault diagnosis system based on photoacoustic signal reconstruction and sensitive field optimization includes:

[0110] Oil-gas separation module, photoacoustic spectroscopy detection sensor, edge detection device, and transformer fault diagnosis service platform.

[0111] The edge detection device adopts a modular design, and the hardware specifically includes a communication unit, a main control unit, a positioning unit, a power supply unit, and a storage unit. The communication unit supports WiFi and 4G wireless communications, mainly realizing communication with the transformer fault diagnosis service platform, and supports RS232 and RS485 wired communications, mainly realizing data communication between the photoacoustic spectroscopy detection sensor. The main control unit contains the V7-690T FPGA+D2000-CPU, and the positioning unit mainly realizes the position positioning of the edge detection device, supporting GPS and Beidou positioning. The power supply unit provides power to each module and unit of the device, and realizes functions such as charging the backup battery and power detection. The storage unit supports SD card or SSD hard drive expansion, which is convenient for storing historical data on the edge side.

[0112] The software design includes an operating system, a language environment, data preprocessing, edge computing, and edge-to-service collaboration. Using Python, C, and C++, the edge detection device developed a data preprocessing algorithm and an edge computing model for transformer fault diagnosis. This collaborative service for transformer fault diagnosis includes functions such as storing edge computing results, interacting with device data and service platform data, and updating the edge computing model.

[0113] The oil-gas separation module is connected to the transformer tank to separate the oil and gas in the transformer tank. This sample gas, serving as the dissolved gas in the transformer oil, is connected to a photoacoustic spectroscopy sensor via an inlet and outlet valve. The edge detection device controls the photoacoustic spectroscopy sensor to detect the photoacoustic signal of the sample gas. The transformer fault diagnosis model running on the edge detection device enables real-time transformer fault diagnosis at the edge, and the calculated results are uploaded to the transformer fault diagnosis service platform. The platform enables viewing transformer status, managing edge detection devices and sensors, and updating edge computing models. Operations and maintenance personnel can view the transformer's status via mobile terminals. When a transformer fault warning occurs, an alarm message is sent to them, allowing them to quickly take preventive measures.

Claims

1. A transformer fault diagnosis method based on photoacoustic signal reconstruction and sensitive field optimization, characterized in that: The steps are as follows: Step 1: By monitoring the photoacoustic signals of multiple gases in the transformer oil sample, the multivariate sequence matrix V of the signal is obtained M N , where M is the number of gas types and N is the length of the sequence sample; Step 2: By reconstructing the sensitive field, the multivariate sequence matrix V M N Mapped in the polar coordinate graph in the form of scattered points, specifically including: Define x max and x min They are V i N , i=1,2,…,the maximum and minimum values of M. For each time point i, the corresponding amplitude is x i , hysteresis coefficient a, angular offset θ l , amplification factor ζ, where l ranges from 0 to L-1, and L represents the number of planes in the total sensitive field; The points in the multivariate signal matrix are converted into polar coordinates through a specific reconstruction relationship formula to form a reconstructed sensitive field image. The reconstruction relationship is as follows: The polar radius r(i) is expressed as: The angular offset θ(i) is defined as: The reflection angle φ(i) is defined as: Argument offset θ l Defined as: Step 3: Extract shallow local features, including SURF features and FAST features, and cluster these features into multiple mutually exclusive categories. The shortest probability entropy of each category center is used as a visual representation factor to form a visual feature vector. Specifically, it includes: S3.1 uses SURF features and FAST features to extract two shallow features; S3.2 uses Kmeans++ to cluster the two shallow features into Kc clusters and calculates the shortest probability entropy P(G) of each cluster. The formula is as follows: Among them, G is the cluster index, g is the feature point index in the Gth cluster, D min is the shortest distance between the feature point and the cluster center within the cluster, and D(g) is the distance between the g-th feature point and the cluster center; S3.3 takes the shortest probability entropy P(G) of each cluster as the representation factor of the cluster center and encodes the shallow features into a feature vector with a dimension of 1*Kc; Step 4: Based on the improved autonomous learning multi-space ALMMo-0 classifier, the formed visual feature vector is trained and classified to achieve the classification of transformer fault diagnosis.

2. The transformer fault diagnosis method based on photoacoustic signal reconstruction and sensitive field optimization according to claim 1 is characterized in that: The step 1 specifically includes: The dissolved gas in the oil sample is extracted through the oil-gas separation module to form a gas sample, and the gas sample is irradiated with a modulated laser at a specific wavelength or band to cause slight changes in local temperature and pressure, generating an acoustic signal sequence, which then forms a multivariate sequence matrix V M N .

3. The transformer fault diagnosis method based on photoacoustic signal reconstruction and sensitive field optimization according to claim 1 is characterized in that: In step 4, the autonomous learning multi-space ALMMo-0 classifier calculates the similarity between the input feature vector and each subspace, assigns the input sample to the fault category with the greatest similarity, and performs the next round of training until the training is completed.

4. The transformer fault diagnosis method based on photoacoustic signal reconstruction and sensitive field optimization according to claim 1 is characterized in that: The specific steps of step 4 are: S4.1 Define {g1,g2,g3,...,g τ } is the feature vector set with a dimension of 1*Kc and is used as the training data set. The number of transformer fault categories is set to q, that is, the number of subspaces. Each subspace corresponds to a set of feature classes. First, a single eigenvector is selected to form a preliminary set of subspaces. In the training phase, {g1,g2,g3,...,g τ The individual feature vectors of the set are input into the autonomous learning multi-space ALMMo-0 classifier for training. τ The similarity between it and each subspace is calculated according to the following formula; in is the average value of all eigenvectors in the i-th subspace, is the average scalar product of the i-th subspace, τ is the sample index of the eigenvector set, ‖‖ is the norm calculation, g ini is a single eigenvector of the preliminary set; S4.2 For input sample g τ , the classifier will calculate q R(g i τ ) is the value of g τ Distribution and maximum R(g i τ ) and fill it into the corresponding subspace for the next round of training. After the training is completed, the feature vector of the collected samples can be tested in real time; the allocation result can be expressed as: Class i =argmax i＝1,2,3,...,q (R i τ )。 5. A transformer fault diagnosis system based on photoacoustic signal reconstruction and sensitive field optimization is applied to execute the fault diagnosis method according to any one of claims 1 to 4, characterized in that: include: Oil-gas separation module, used to extract oil samples from transformer tanks and separate dissolved gases; A photoacoustic spectrum detection sensor is used to detect the photoacoustic signal of the sample gas; an edge detection device for controlling the photoacoustic spectroscopy detection sensor, processing data, and executing a fault diagnosis model; The transformer fault diagnosis service platform is used to receive fault diagnosis results uploaded by the edge detection device and provide transformer status viewing, equipment management, and model update functions.

6. The transformer fault diagnosis system based on photoacoustic signal reconstruction and sensitive field optimization according to claim 5 is characterized in that: The hardware system of the edge detection device includes a communication unit, a main control unit, a positioning unit, a power supply unit and a storage unit. The communication unit supports WiFi and 4G wireless communications, mainly realizing communication with the transformer fault diagnosis service platform, and supports RS232 and RS485 wired communications, mainly realizing data communication between the photoacoustic spectroscopy detection sensor; the main control unit contains a V7-690T FPGA+D2000-CPU, the positioning unit mainly realizes the position positioning of the edge detection device, and supports GPS and Beidou positioning; the power supply unit provides power to each module and unit of the device, and realizes functions such as charging the backup battery and power detection; the storage unit supports SD card or SSD hard disk expansion, which is convenient for storing historical data on the edge side; The software design of the edge detection device includes an operating system, a language environment, data preprocessing, edge computing, and edge-service collaborative services, enabling real-time fault diagnosis of transformers at the edge and uploading the results to a transformer fault diagnosis service platform.

7. The transformer fault diagnosis system based on photoacoustic signal reconstruction and sensitive field optimization according to claim 5 is characterized in that: The oil-gas separation module is connected to the transformer tank to separate the oil and gas in the transformer tank, and the oil and gas are used as sample gas for the dissolved gas in the transformer oil. The sample gas is connected to the photoacoustic spectrum detection sensor through the air inlet valve and the air outlet valve. The edge detection device controls the photoacoustic spectrum detection sensor to detect the photoacoustic signal of the sample gas. The transformer fault diagnosis model running on the edge detection device is used to realize real-time fault diagnosis of the transformer at the edge end, and the calculation results are uploaded to the transformer fault diagnosis service platform. The transformer fault diagnosis service platform can realize the viewing of the transformer status, the management of the edge detection device and the sensor, and the update of the edge computing model. Operation and maintenance personnel can check the status of the transformer through mobile terminals. When a transformer fault warning occurs, the alarm information can be sent to the operation and maintenance personnel, so that the operation and maintenance personnel can take preventive measures quickly.

Images