Transformer fastening bolt mechanical fault diagnosis method and system

Abstract

The invention discloses a transformer fastening bolt mechanical fault diagnosis method and system, and the main method comprises the steps: obtaining a thread image texture feature and phase information of an ROI region, calculating a time domain feature and a frequency domain feature, fusing the time domain feature, the frequency domain feature and the thread image texture feature, and carrying out the calculation of the time domain feature and the frequency domain feature. Stacking the fusion feature vectors obtained in all time steps according to a time sequence, and constructing a two-dimensional visual vibration feature matrix; and taking the two-dimensional visual vibration characteristic matrix as the input of a deep learning diagnosis model, and outputting a fault judgment result. According to the invention, a video non-contact measurement mode is adopted to realize mechanical fault diagnosis of the fastening bolt of the transformer, and hidden faults are effectively mined to provide reliable data support for health monitoring of the transformer.

Description

Technical Field

[0001] This invention relates to the field of power equipment condition monitoring and fault diagnosis technology, and in particular to a method for diagnosing mechanical faults in transformer fastening bolts based on visual vibration signal characteristics and deep learning networks. Background Technology

[0002] As a core piece of equipment in the power system, the stability of the transformer's operating status directly affects the safe and reliable operation of the entire power grid. Fastening bolts, as critical connecting components in the transformer structure, can suffer from mechanical failures such as loosening or breakage, leading to changes in the transformer's vibration characteristics. Long-term operation of these bolts may cause serious faults such as winding deformation and core loosening, and even equipment shutdown.

[0003] Traditional methods for diagnosing transformer bolt faults often rely on contact sensors (such as accelerometers), which suffer from problems such as inconvenient installation, complex wiring, and susceptibility to electromagnetic interference. In contrast, vision-based non-contact measurement technologies offer advantages such as flexible operation, no need to intrude into the equipment, and the ability to perform remote monitoring, and are gradually becoming a research hotspot in equipment condition monitoring.

[0004] In recent years, deep learning technology has demonstrated powerful feature extraction and pattern recognition capabilities in the field of fault diagnosis. One-dimensional convolutional neural networks (1D-CNNs) have been used to analyze transformer vibration signals, achieving accurate prediction of voltage fluctuations and inter-turn short-circuit faults. Deep recurrent neural networks (RNNs), including gated recurrent units (GRUs) and long short-term memory networks (LSTMs), have enabled early warning of transformer underexcitation, overexcitation, and inter-turn faults through vibration time series analysis. These studies show that deep learning networks can effectively uncover hidden fault features in vibration signals, providing a new technical approach for equipment fault diagnosis.

[0005] However, existing research on transformer fault diagnosis based on deep learning mainly focuses on electrical faults (such as inter-turn short circuits and voltage anomalies), with limited research on mechanical structural faults such as fastening bolts, and a lack of effective diagnostic methods that combine visual vibration signal features. Therefore, there is an urgent need to construct a method for diagnosing mechanical faults in transformer fastening bolts based on visual vibration signal features and deep learning networks, in order to achieve accurate identification and early warning of bolt faults. Summary of the Invention

[0006] The technical problem to be solved by this invention is how to solve the problems of limited contact measurement, difficulty in feature extraction, and low diagnostic accuracy in the existing fault diagnosis of transformer fastening bolts.

[0007] The present invention solves the above-mentioned technical problems through the following technical means: Methods for diagnosing mechanical faults in transformer fastening bolts include: S1. Acquire video images of the vibration of the transformer fastening bolts; S2. Perform region of interest selection and filtering on the frames of the video image to obtain a thread image; S3. For the thread image, obtain the thread image texture features through filtering and convolution operations; further calculate the temporal and frequency domain features based on the thread image texture features. S4. At each time step within the continuous time window, the temporal features, frequency features, and image texture features of the region of interest are fused to obtain a fused feature vector. Then, the fused feature vectors obtained from all time steps are stacked in chronological order to construct a two-dimensional visual vibration feature matrix. S5. Use the two-dimensional visual vibration feature matrix as input to the deep learning diagnostic model and output the fault judgment result.

[0008] Furthermore, the specific steps for acquiring the video image sequence are as follows: a high-speed industrial camera is used to capture video of the area where the fastening bolts are located on the surface of the transformer.

[0009] Further, the specific steps for region of interest selection and preprocessing are as follows: First, the region of interest is defined for the bolt and its surrounding area in the video image sequence. Then, an adaptive threshold segmentation algorithm is used to extract the bolt contour to achieve automatic localization and tracking of the region of interest. Next, algorithms such as Gaussian filtering are used to remove noise from the image. Finally, histogram equalization is used to enhance the contrast of the image to obtain the thread image.

[0010] Furthermore, the specific steps for extracting vibration feature quantities are as follows: The thread image of the region of interest is convolved with a set of Gabor filters with different directions and frequencies to obtain the thread image texture features. For each thread image texture feature after convolution with Gabor filters, its phase is calculated to obtain the phase information of the region of interest. An optical flow constraint equation is established based on phase-optical flow; partial derivatives are calculated by differential approximation of phase information in time and space, and the optical flow constraint equation is solved to obtain the optical flow velocity components in the horizontal and vertical directions; The temporal characteristics are obtained by integrating the obtained optical flow velocity components; Then, perform a Fourier transform on the time-domain features to obtain the frequency-domain features.

[0011] Furthermore, the specific steps of visual feature fusion are as follows: concatenating temporal features, frequency domain features, and image texture features of the region of interest; and then combining the concatenated base feature vector. Input to a feedforward neural network, output and Attention score vectors of the same dimension Then, the attention score vector is processed using the Softmax function. The process is performed to generate the final weight vector. Finally, the original basic feature vectors Its corresponding weight vector Element-wise multiplication is performed to obtain the final weighted fused feature vector. .

[0012] The present invention also provides a mechanical fault diagnosis system for transformer fastening bolts, comprising: Image sequence acquisition module: Acquires video images of transformer fastening bolt vibration; Region of Interest (ROI) Selection and Preprocessing Module: Performs ROI selection and filtering on the frames of the video image to obtain a spiral image; Vibration feature extraction module: For thread images, texture features of the thread image are obtained through filtering and convolution operations; time-domain features and frequency-domain features are further calculated based on the texture features of the thread image. Visual feature fusion module: At each time step within a continuous time window, temporal features, frequency features, and image texture features of the region of interest are fused to obtain a fused feature vector. Then, the fused feature vectors obtained from all time steps are stacked in chronological order to construct a two-dimensional visual vibration feature matrix. Diagnostic module: It takes the two-dimensional visual vibration feature matrix as input to the deep learning diagnostic model and outputs the fault judgment result.

[0013] Furthermore, the specific steps for acquiring the video image sequence are as follows: a high-speed industrial camera is used to capture video of the area where the fastening bolts are located on the surface of the transformer.

[0014] Further, the specific steps for region of interest selection and preprocessing are as follows: First, the region of interest is defined for the bolt and its surrounding area in the video image sequence. Then, an adaptive threshold segmentation algorithm is used to extract the bolt contour to achieve automatic localization and tracking of the region of interest. Next, algorithms such as Gaussian filtering are used to remove noise from the image. Finally, histogram equalization is used to enhance the contrast of the image to obtain the thread image.

[0015] Furthermore, the specific steps for extracting vibration feature quantities are as follows: The thread image of the region of interest is convolved with a set of Gabor filters with different directions and frequencies to obtain the thread image texture features. For each thread image texture feature after convolution with Gabor filters, its phase is calculated to obtain the phase information of the region of interest. An optical flow constraint equation is established based on phase-optical flow; partial derivatives are calculated by differential approximation of phase information in time and space, and the optical flow constraint equation is solved to obtain the optical flow velocity components in the horizontal and vertical directions; The temporal characteristics are obtained by integrating the obtained optical flow velocity components; Then, perform a Fourier transform on the time-domain features to obtain the frequency-domain features.

[0016] Furthermore, the specific steps of visual feature fusion are as follows: concatenating temporal features, frequency domain features, and image texture features of the region of interest; and then combining the concatenated base feature vector. Input to a feedforward neural network, output and Attention score vectors of the same dimension Then, the attention score vector is processed using the Softmax function. The process is performed to generate the final weight vector. Finally, the original basic feature vectors Its corresponding weight vector Element-wise multiplication is performed to obtain the final weighted fused feature vector. .

[0017] The advantages of this invention are: 1. Non-contact measurement: It uses visual imaging technology to acquire bolt vibration signals, avoiding the installation limitations and electromagnetic interference problems of contact sensors, and is suitable for various complex field environments.

[0018] 2. High-precision feature extraction: It integrates time-domain, frequency-domain and visual texture features to comprehensively depict vibration changes caused by bolt faults; combined with the automatic feature learning capability of deep learning networks, it effectively mines hidden fault information and solves the limitations of traditional manual feature design.

[0019] 3. High engineering practicality: It can realize remote and real-time monitoring of transformer fastening bolts, provide data support for condition-based maintenance of power equipment, reduce operation and maintenance costs, and improve the reliability of power grid operation. Attached Figure Description

[0020] Figure 1 This is a flowchart of a method for diagnosing mechanical faults in transformer fastening bolts according to one embodiment of the present invention; Figure 2 This is a schematic diagram of a transformer mechanical fault diagnosis test scenario in one embodiment of the present invention; Figure 3 This is a schematic diagram of a transformer mechanical fault diagnosis and testing process in one embodiment of the present invention. Detailed Implementation

[0021] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0022] Combination Figures 1 to 3The method for diagnosing mechanical faults in transformer fastening bolts of the present invention includes the following steps: Video image sequence acquisition: A high-speed industrial camera is used to continuously capture video images of the area around the fastening bolts on the transformer surface, obtaining a video image sequence. The camera frame rate is adjusted according to the actual situation; in this embodiment, a frame rate of 1000fps is used, and the video is saved in AVI format. Generally, the acquired video data can be in common video formats such as AVI, MOV, and MP4. Based on the sampling theorem, to accurately capture the minute vibration signals of the bolts, the camera sampling frequency is... Must meet ,in To determine the highest frequency of bolt vibration, the camera sampling frequency was set to be no less than 1kHz based on previous experimental measurements.

[0023] Region of Interest (ROI) Selection and Preprocessing of Video Images: To effectively reduce background interference, regions of interest (ROIs) are precisely defined for bolts and their surrounding areas in the image sequence. During ROI selection, the image is first processed using Gaussian filtering, utilizing the Gaussian function... The generated convolution kernel is convolved with the image, where Two-dimensional coordinates representing image pixels, The standard deviation of the Gaussian function should be reasonably selected based on the image noise level (generally between 1 and 3), and adjusted accordingly. A value is used to balance noise suppression and edge preservation, improving image quality. Then, an adaptive thresholding algorithm is used to extract the bolt contour. This algorithm dynamically calculates the threshold based on local image characteristics. Let the image grayscale value be I(x,y), and the adaptive threshold T(x,y) be calculated using local neighborhood pixel statistics. For example, if a mean filtering method is used... ,in This is an empirical coefficient (generally ranging from 0.8 to 1.2, which can be fine-tuned based on the contrast between the bolt and the background). For The average grayscale value of pixels within the local neighborhood centered on the bolt is used. By continuously iterating and optimizing the threshold calculation process, automatic positioning and precise tracking of the ROI are achieved, ensuring that the ROI remains tightly surrounding the bolt area during bolt vibration. This step yields a clear thread image of the ROI.

[0024] Vibration feature extraction: (1) Select a set of Gabor filters with different directions and frequencies, and perform convolution operations with each Gabor filter on the thread image of the region of interest to obtain the bolt texture features, i.e., the complex result. For the complex result after convolution of each Gabor filter... ,use Calculate its phase This allows us to obtain the phase information of the feature points and their neighborhood, where... and These are the results of complex numbers. The imaginary and real parts.

[0025] (2) Based on the principle of phase-optical flow (POF) technology, establish the optical flow constraint equation. ,in and These are the horizontal and vertical optical flow components, respectively. , and These are the partial derivatives of the phase with respect to time and space, respectively. The phase difference between adjacent frames... Divide by time interval To approximate the calculation; by using the phase information of the current frame in and direction Difference approximation calculation and Assuming horizontal optical flow Solving the optical flow constraint equations yields the vertical optical flow components. Then, integrating the optical flow component yields the vibrational displacement. Similarly, the horizontal displacement is obtained.

[0026] (3) Perform detailed time-domain characteristic calculations on each vibration curve segment, and accurately calculate the peak value. Root mean square Time-domain statistics such as kurtosis K comprehensively reflect the intensity and stability of bolt vibration. The root mean square calculation formula is as follows: ,in For the first The displacement values ​​at each sampling point are N, where N is the number of sampling points. During calculation, the displacement value at each sampling point is squared, then the average is calculated, and finally the square root is taken to ensure that the calculation result accurately reflects the vibration intensity. Kurtosis is defined as... In the formula The mean value of the vibration curve is calculated by first finding the mean value, then calculating the fourth and second powers of the difference between each sampling point and the mean value, and obtaining the kurtosis value through multiple calculations, which reflects the impact characteristics of the vibration signal.

[0027] (4) Perform a Fourier transform on the vibration curve to obtain accurate spectral characteristics. The Discrete Fourier Transform (DFT) is used, and its formula is as follows: Where x(n) is the discrete-time sequence (i.e., the nth sample value of the vibration curve, where n is the index of the sampling point, n=0,1, N 1) X(k) represents the corresponding frequency domain coefficient (k is the frequency index), and j is the imaginary unit (j² = ...). 1) N is the total number of sampling points. During the calculation, the Fast Fourier Transform (FFT) algorithm is used to improve computational efficiency. By continuously grouping the sequence and recursively calculating, the computational complexity is reduced from O(N²) to O(…). (O represents the time complexity symbol). The focus is on analyzing the spectral components near twice the fundamental frequency (100Hz). By setting the frequency bandwidth (90-110Hz), a detailed analysis of the spectral amplitude within this band is conducted. This band is closely related to the vibration of the transformer core and structure. By accurately analyzing the amplitude changes in this band, the impact of bolt faults on the overall structural vibration can be accurately determined.

[0028] Visual feature fusion: The three types of features extracted in the previous steps are structured and organized to form three independent feature vectors: (1) Time-domain dynamic feature vector This vector comprehensively characterizes the statistical properties of bolt vibration displacement signals in the time dimension. It consists of multiple statistical quantities such as peak value, root mean square (RMS), kurtosis, and waveform factor. Among them, the RMS value reflects the energy intensity of the vibration, while the kurtosis value is particularly sensitive to impact vibration signals caused by faults such as bolt loosening.

[0029] (2) Frequency domain dynamic eigenvectors This vector reveals the energy distribution pattern of bolt vibration along the frequency dimension. It primarily includes key characteristics such as the spectral amplitude and centroid of the octave band, which are closely related to the transformer's vibration. These characteristics effectively reflect the impact of changes in bolt tightening status on the vibration transmission characteristics of the entire structure.

[0030] (3) Static visual texture feature vector This vector is used to quantify the static surface information of the bolt and its surrounding area. Texture features such as contrast, correlation, energy, and homogeneity of the image are extracted using gray-level co-occurrence matrix (GLCM) technology. These features can capture visual cues such as minor corrosion, changes in surface oil, or abnormal light reflection that may result from long-term bolt loosening.

[0031] The most direct way to fuse them is to concatenate the three feature vectors mentioned above to form a basic, wider-dimensional fused feature vector. This operation is simple and efficient, preserving all original feature information. However, it assumes that all features contribute equally to the final diagnosis, which may not be optimal in practical applications. To address the limitations of equal feature weighting in the initial fusion, this invention introduces an attention-weighted mechanism to deeply fuse the basic feature vectors and dynamically assign weights to different features. This aims to allow the model to automatically learn and focus on "fault-sensitive features" that contribute most to fault state judgment during training, such as kurtosis and fundamental frequency band amplitude, while suppressing interference from noise or redundant features.

[0032] The specific implementation of this mechanism is as follows: The concatenated basic feature vectors are... Input a small feedforward neural network that learns a nonlinear mapping, whose output is... Attention score vectors with the same dimensions Then, the attention score vector is processed using the Softmax function. The process is performed to generate the final weight vector. The Softmax function guarantees that the sum of the weights of all features is 1, and each weight value is between (0,1), representing the importance of that feature. Finally, the original basic feature vector is... Its corresponding weight vector Element-wise multiplication is performed to obtain the final weighted fused feature vector. Through this mechanism, the model no longer passively receives all features, but actively and selectively enhances key information, which greatly improves the quality of feature representation and the efficiency of subsequent diagnostic models.

[0033] The above fusion process is performed on a single time step, i.e., a very short time segment of the video. To capture the dynamic process of fault development, we perform the above feature extraction and fusion process at every time step within a continuous time window (a 1-second video segment, divided into 50 time steps in 0.02-second increments). Finally, we calculate the weighted fused feature vector obtained from all time steps. Stacked in chronological order, they form a two-dimensional visual vibration feature matrix. The matrix has the dimension of [number of time steps × dimension of fused features], and this matrix is ​​the final input to the subsequent "1D-CNN+LSTM" hybrid diagnostic model. 1D-CNN excels at extracting combinations of local key patterns along the feature dimension, while LSTM can capture the temporal dependencies of vibration signals along the time step dimension. The combination of the two maximizes the utilization of the carefully designed fusion features of this invention, achieving high-precision diagnosis of transformer fastening bolt fault conditions.

[0034] Building a deep learning diagnostic model: (1) Model Architecture Design: Combining the temporal continuity of the vibration signal of transformer fastening bolts with the spatial correlation of visual texture features, a hybrid architecture of "1D-CNN+LSTM" is adopted. This architecture can simultaneously realize the extraction of local key patterns of multi-domain fusion features (temporal domain, frequency domain, texture) by 1D-CNN, solving the problem of missing hidden fault information in traditional manual feature design; LSTM models the temporal dependence of vibration signals, adapting to the dynamic development process of bolt faults from "normal - slight loosening - severe loosening - fracture". The network consists of an input layer, four feature extraction modules (including convolutional layers, batch normalization layers, activation functions, and pooling layers), one LSTM layer, one fully connected layer, and an output layer. The input layer receives data in the following dimensions: The visual vibration feature matrix, where The length of a single sample time series. To fuse feature dimensions (temporal domain + frequency domain + texture), a min-max normalization formula is used before input. The features are mapped to the [0,1] interval to eliminate the interference of dimensional differences on model training. In the formula, x is the original feature value, and xmin and xmax are the minimum and maximum values ​​of that feature in the training set, respectively.

[0035] Convolutional layer parameter settings: A filter number gradient of "32→64→128→256" is used to progressively increase the feature abstraction level. The kernel size is adaptively selected based on the feature type: For temporal features (such as peak values ​​and root mean square), a small kernel with k=3 is used to enhance the capture of local impact features (such as vibration abrupt changes caused by bolt loosening); for frequency domain features (such as amplitude in the 100Hz band), a large kernel with k=7 is used to integrate energy distribution information over a wide frequency range; for texture features (such as contrast and energy of the gray-level co-occurrence matrix), a medium kernel with k=5 is used to balance spatial details and neighborhood correlation. The convolutional layer calculation process is as follows: In the formula, * represents a one-dimensional convolution operation. The convolution kernel weight matrix is... The bias vector is ReLU, and ReLU is the activation function (ReLU(x)=max(0,x)), which can effectively alleviate the gradient vanishing problem and improve the training convergence speed.

[0036] Each convolutional layer is followed by a batch normalization (BN) layer, calculated using the formula... Standardizing the feature distribution accelerates network convergence and improves generalization ability. In the formula... , These are the mean and variance of a small batch of samples (batchsize=32 in this example). To prevent tiny constants with a denominator of zero, , These are learnable scaling and offset parameters.

[0037] (3) LSTM layer design: A unidirectional LSTM layer is adopted (the development of bolt faults is irreversible in time, and the future state does not affect the historical characteristics), and 128 hidden units are set to specifically model the long-term dynamic correlation of vibration signals (such as the cumulative effect of bolt loosening over time).

[0038] LSTM units selectively memorize and update temporal information through input gates, forget gates, output gates, and cell states. The core calculation formula is as follows: 1. Input gate: Controls the input weights of the features at the current time step.

[0039] 2. Forgetting Gate: Determines the percentage of historical cell states retained.

[0040] 3. Cell state update: Storing long-term temporal information:

[0041]

[0042] 4. Output Gate: Generates the current hidden state:

[0043]

[0044] In the formula, The sigmoid activation function ( ), For element-wise multiplication, Xt is the input of the LSTM layer at time t (from the output of the feature extraction module), ht-1 and Ct-1 are the hidden state and cell state at time t-1, respectively, Wxi, Whi, etc. are weight matrices, and bi, bf, etc. are bias vectors.

[0045] Hidden state (dimension) output by LSTM layer After global average pooling, the input is fed into a fully connected layer (256 neurons), where features are further integrated using the ReLU activation function; the final output layer uses the softmax function. The model output is converted into a probability distribution of four fault states (normal, slightly loose, severely loose, and broken). In the formula, zj is the unnormalized output of the fully connected layer for the j-th fault, and pj is the predicted probability of the j-th fault, satisfying... .

[0046] Model training and fault diagnosis: Dataset Construction Experimental subject: Transformer; 6 sets of fastening bolts were selected for the top cover plate. Fault simulation: Different fault states are achieved by controlling the bolt preload with a torque wrench (normal: 40N). m; Slightly loose: 25N m; Severely loose: 10N m; breakage: removal of bolts); Data acquisition: High-speed industrial cameras (1280×720 resolution, 1000fps) were used to capture vibration videos of the bolt area, and the transformer load current (0.5~1.2 times the rated current) and ambient temperature (25~40℃) were recorded simultaneously to ensure that the sample covers the actual operating conditions; Sample size: 1200 vibration videos were collected for each type of fault state (each video is 1 second long, corresponding to 50 segmented samples of 0.02 seconds each), for a total of 4800 videos, generating 240,000 feature samples.

[0047] The dataset is divided into a training set (160,000 samples), a validation set (40,000 samples), and a test set (40,000 samples) in a 4:1:1 ratio, with the following annotation rules: Normal state: label [1,0,0,0]; Slightly loose: label [0,1,0,0]; Severely loose: label [0,0,1,0]; broken: label [0,0,0,1].

[0048] To address the overfitting issue with small sample sizes, the following operations were performed on the feature samples in the training set: Temporal features: ±5% Gaussian noise was added (to simulate environmental vibration interference); Frequency domain features: the amplitude of the 100Hz band was randomly scaled by ±8% (to simulate the impact of load fluctuations on the spectrum); Texture features: the gray-level co-occurrence matrix was randomly scaled by 0.9 to 1.1 times (to simulate the impact of illumination changes on texture).

[0049] Model training Optimizer: Employs the RMSProp optimizer and sets the decay factor. Momentum parameters Through formula Adaptive adjustment of the learning rate. Where gt is the gradient at time t. It is a moving average of the squared gradient.

[0050] Loss function: Cross-entropy loss function is used. This method is adapted for multi-class classification tasks and quantifies the difference between predicted probabilities and true labels. In the formula, M is the number of samples in the batch, yij is the true label (0 or 1) of the j-th class for the i-th sample, and pij is the predicted probability of the j-th class for the i-th sample.

[0051] Training epochs: Maximum 150 epochs, using an early stopping strategy, training stops when the validation set loss does not decrease for 10 consecutive epochs to avoid overfitting; Model saving: Calculate the macro-average F1 score of the validation set after each training round. , The model with the highest F1 score is saved as the optimal model. In the formula, Pj represents the accuracy of the j-th type of fault (…). Rj is the recall rate of the j-th type of fault ( ), TPj, FPj, and FNj are the number of true positive, false positive, and false negative samples of the j-th class, respectively.

[0052] Fault Diagnosis Preprocessing of samples to be detected: For newly acquired bolt vibration videos, the following process is followed: “ROI selection → phase optical flow displacement calculation → multi-domain feature extraction → feature fusion → min-max normalization” to generate a feature matrix with dimensions [50, 15] (50 frames 0.05s, 15-dimensional fused features). Model inference: Input the feature matrix into the optimal model, and output the probability distribution of four types of faults. ; Decision rule: Select the one with the highest probability. If pmax ≥ 0.9 (confidence threshold, determined by ROC curve analysis of the validation set), it is determined to be the corresponding fault state; if pmax < 0.9, it is marked as "suspected fault" and further verification is required.

[0053] Performance validation metrics: Classification accuracy: the percentage of samples correctly classified overall; Macro-average precision, macro-average recall, and macro-average F1 score: evaluate the model's ability to identify various types of faults in a balanced manner; Confusion matrix: visualizes the misclassification of various faults, with a focus on analyzing the ability to distinguish between "slight looseness" and "normal".

[0054] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method of diagnosing mechanical failure of a transformer fastening bolt, characterized by, The method comprises the following steps: S1. Obtain a video image of transformer fastening bolt vibration; S2. Perform region of interest selection and filtering processing on the frames of the video image to obtain a thread image; S3. For the thread image, obtain thread image texture features through filtering convolution operation; further calculate time domain features and frequency domain features according to the thread image texture features; S4. Perform fusion of the time domain features, the frequency domain features and the image texture features of the region of interest at each time step in a continuous time window to obtain a fusion feature vector, then stack the fusion feature vectors obtained at all time steps in time sequence to construct a two-dimensional visual vibration feature matrix; S5. Take the two-dimensional visual vibration feature matrix as input of a deep learning diagnosis model, and output a fault judgment result.

2. The transformer fastening bolt mechanical failure diagnosis method according to claim 1, characterized by, The step S1 specifically comprises using a high-speed industrial camera to perform video acquisition on the fastening bolt region on the surface of the transformer.

3. The transformer fastening bolt mechanical failure diagnosis method according to claim 1, characterized by, The step S2 specifically comprises: defining a region of interest for the bolt and its surrounding region in the video image sequence, extracting the bolt contour by using an adaptive threshold segmentation algorithm to realize automatic positioning and tracking of the region of interest; then, removing noise in the image by using a Gaussian filtering algorithm; and enhancing the contrast of the image by using a histogram equalization method to obtain a thread image.

4. The transformer fastening bolt mechanical failure diagnosis method according to any one of claims 1 to 3, characterized by, The step S3 specifically comprises: performing convolution operation on the thread image of the region of interest and a group of Gabor filters with different directions and frequencies to obtain thread image texture features, calculating the phase of the thread image texture features after convolution of each Gabor filter to obtain phase information of the region of interest, establishing an optical flow constraint equation according to the phase-optical flow, calculating the partial derivative by differentiating the phase information in time and space, and solving the optical flow constraint equation to obtain horizontal and vertical optical flow velocity components; integrating the obtained optical flow velocity components to obtain time domain features; and performing Fourier transform on the time domain features to obtain frequency domain feature quantities.

5. The transformer fastening bolt mechanical failure diagnosis method according to any one of claims 1 to 3, characterized by, The step S4 is specifically, splicing the time domain features, the frequency domain features and the image texture features of the region of interest, splicing the basis feature vectors after splicing Inputting the feedforward neural network, outputting the attention score vector with the same dimension as the basis feature vector ; and processing the attention score vector by a Softmax function to generate a final weight vector , and finally performing element-by-element multiplication between the original basis feature vector and the corresponding weight vector to obtain a final weighted fusion feature vector .

6. A transformer fastening bolt mechanical failure diagnosis system characterized by, The method comprises the following steps: An image sequence acquisition module is configured to acquire a video image of transformer fastening bolt vibration; A region of interest selection and preprocessing module is configured to perform region of interest selection and filtering processing on the frames of the video image to obtain a thread image; A vibration feature quantity extraction module is configured to, for the thread image, obtain thread image texture features through filtering convolution operation; and further calculate time domain features and frequency domain features according to the thread image texture features; A visual feature fusion module is configured to perform fusion of the time domain features, the frequency domain features and the image texture features of the region of interest at each time step in a continuous time window to obtain a fusion feature vector, and then stack the fusion feature vectors obtained at all time steps in time sequence to construct a two-dimensional visual vibration feature matrix; A diagnosis module is configured to take the two-dimensional visual vibration feature matrix as input of a deep learning diagnosis model, and output a fault judgment result.

7. The transformer fastening bolt mechanical failure diagnostic system of claim 6, wherein, The specific steps of the video image sequence acquisition comprise using a high-speed industrial camera to perform video acquisition on the fastening bolt region on the surface of the transformer.

8. The transformer fastening bolt mechanical failure diagnostic system of claim 6, wherein, The specific steps of the region of interest selection and preprocessing are as follows: a region of interest is defined for a bolt and its surrounding area in a video image sequence, a self-adaptive threshold segmentation algorithm is used to extract the bolt contour, and automatic positioning and tracking of the region of interest are realized; then, a Gaussian filter and other algorithms are used to remove noise in the image; a histogram equalization method is used to enhance the contrast of the image, and a thread image is obtained.

9. The transformer fastening bolt mechanical failure diagnostic system according to any one of claims 6 to 8, characterized in that, The specific steps of the vibration feature extraction are as follows: a convolution operation is performed on the thread image of the region of interest and a group of Gabor filters with different directions and frequencies to obtain thread image texture features; for each thread image texture feature after the convolution of a Gabor filter, the phase thereof is calculated to obtain phase information of the region of interest; a light flow constraint equation is established according to the phase-light flow; the phase information is differentiated in time and space to approximate the partial derivative, and the light flow constraint equation is solved to obtain horizontal and vertical light flow velocity components; the obtained light flow velocity components are integrated to obtain time domain features; and Fourier transform is performed on the time domain features to obtain frequency domain feature quantities.

10. The transformer fastening bolt mechanical failure diagnostic system according to any one of claims 6 to 8, characterized in that, The specific steps of the visual feature fusion are as follows: the time domain feature, the frequency domain feature and the image texture feature of the region of interest are spliced, the spliced basic feature vector is input into a convolutional neural network, and a feature vector of the same dimension as the basic feature vector is output The input is a feedforward neural network, and the output is an attention score vector of the same dimension as the basic feature vector ; the attention score vector is processed through a Softmax function to generate a final weight vector ; finally, the original basic feature vector is multiplied element by element with the corresponding weight vector to obtain a final weighted fusion feature vector .​​​​

Images