A Video-Based Quantitative Monitoring Method for Manhole Bolt Loosening in Hydropower Station Pressure Vessels

Abstract

This invention discloses a video-based quantitative monitoring method for loose bolts in manholes of pressure vessels in hydropower stations, comprising the following steps: Step 1, constructing a finite element model of bolt-manhole-vessel, and building a database relating preload, vibration characteristics, and system output entropy; Step 2, constructing and training an AI agent model, deploying the AI ​​agent model to the monitoring terminal, quantitatively identifying bolt preload and loosening location, and outputting the bolt quantitative identification results; Step 3, acquiring bolt vibration videos, calculating system output entropy, and extracting vibration characteristic parameters; Step 4, verifying monitoring data and standardizing the output of bolt loosening quantitative data. This invention adopts a non-contact monitoring mode, which is convenient for operation and maintenance, has low monitoring costs, strong adaptability to operating conditions, and outstanding anti-interference ability. It can achieve accurate quantitative identification and location of bolt loosening, effectively overcoming the technical limitations of existing technologies that can only make qualitative judgments. This invention also discloses a storage medium and computer equipment.

Description

Technical Field

[0001] The present invention belongs to the technical field of pressure vessel monitoring, and relates to a method for quantitatively monitoring the loosening of manhole bolts of a hydraulic power station pressure vessel based on video. The present invention also relates to a storage medium and a computer device including the above method for quantitatively monitoring the loosening of manhole bolts of a hydraulic power station pressure vessel based on video. Background Technique

[0002] As a core pressure-bearing device in the fields of industrial production, energy transmission, etc., the safe operation of pressure vessels is directly related to production stability, personnel safety and property protection, and is indispensable in the field of hydropower stations. Among them, manhole bolts are key fasteners of pressure vessels, playing a core role in connecting the manhole and the equipment body and ensuring the sealing of the equipment. They are in a complex working condition of long-term high pressure, frequent vibration and temperature fluctuation. Once the manhole bolts are loose, fatigued or even fractured, it will directly lead to the failure of the pressure vessel seal, causing medium leakage and equipment damage. In severe cases, major safety accidents such as unit shutdown will occur, not only causing huge economic losses, but also threatening the lives of on-site operators. Especially for the supporting bolt groups of the spiral case access door and the tail water access door of the hydropower station, they are directly related to the sealing safety of the spiral case and the tail water system, and are the key links to prevent medium leakage and ensure the normal operation of the unit. Therefore, the loosening monitoring of manhole bolts is the core link of the safe operation and maintenance of pressure vessels.

[0003] At present, for the loosening monitoring of manhole bolts of hydraulic power station pressure vessels, there are mainly three methods in the industry: The first is manual hammering monitoring, in which maintenance personnel regularly hammer the bolts and judge the loosening situation by listening to the sound and identifying the position. This method completely relies on manual experience, has extremely strong subjectivity, and cannot accurately identify slight loosening; moreover, pressure vessels are mostly in high-risk and high-risk working environments, and manual hammering requires close contact with the equipment, which has potential safety hazards; at the same time, the monitoring efficiency is extremely low, and it is impossible to achieve normalized and batch monitoring.

[0004] The second is contact monitoring. By arranging sensors such as strain and pressure on the bolts, the force data of the bolts is collected in real time to judge the loosening state. This method requires a large number of sensors. The number of bolts supporting the manhole of a single pressure vessel is relatively large, and a large number of sensors are required for comprehensive arrangement, increasing the monitoring cost; moreover, in scenarios such as maintenance and entry, the sensors need to be frequently disassembled and assembled, which not only greatly increases the workload of operation and maintenance and reduces the monitoring efficiency, but also easily causes sensor damage and accuracy failure due to improper disassembly and assembly; at the same time, the sensors are prone to signal instability and data distortion in the high-temperature and strong-vibration environment of hydropower stations, and it is difficult to operate stably for a long time.

[0005] The third method is traditional camera monitoring, which involves deploying high-definition cameras to capture images of the bolts' appearance, enabling online monitoring through manual or basic image recognition. While this method achieves non-contact monitoring, it can only determine if the bolts are visibly loose, and cannot quantitatively assess the degree of looseness. Furthermore, the image recognition accuracy is easily affected by environmental factors such as lighting and vibration, further reducing the reliability of the monitoring.

[0006] In summary, existing methods for monitoring manhole bolts in pressure vessels cannot simultaneously meet the four core requirements of "accurate quantification, real-time early warning, convenient non-contact operation and maintenance, and adaptability to complex working conditions". Summary of the Invention

[0007] The purpose of this invention is to provide a video-based quantitative monitoring method for loosening manhole bolts in hydropower station pressure vessels, which solves the problems of low accuracy, inability to quantify, inconvenient operation and maintenance, weak anti-interference, and difficulty in adapting to the complex working conditions of hydropower stations in existing pressure vessel manhole bolt loosening monitoring technologies.

[0008] A second objective of this invention is to provide a storage medium incorporating the above-described video-based quantitative monitoring method for manhole bolt loosening in hydropower station pressure vessels.

[0009] A third objective of this invention is to provide a computer device incorporating the above-described video-based quantitative monitoring method for manhole bolt loosening in hydropower station pressure vessels.

[0010] The technical solution adopted in this invention is a video-based quantitative monitoring method for loosening manhole bolts in hydropower station pressure vessels, comprising the following steps: Step 1: Construct a finite element model of bolt-manhole-container and build a database relating preload and vibration characteristics to system output entropy; Step 2: Build and train the AI ​​agent model, deploy the AI ​​agent model to the monitoring terminal, quantify and identify the bolt preload and loosening position, and output the bolt quantitative identification results; Step 3: Acquire bolt vibration video, calculate system output entropy and extract vibration characteristic parameters; Step 4: Verify monitoring data and standardize the output of bolt loosening quantitative data.

[0011] The invention is further characterized by: Step 1 is implemented in the following steps: Step 1.1: Construct an integrated three-dimensional finite element model of the bolt, manhole, and container body; Draw solid models of each component, define the connection relationships between bolts and flanges, and between flanges and container body according to the actual connection methods, and construct an integrated three-dimensional finite element model of bolts-manhole-container body; Step 1.2, Mesh generation; The integrated three-dimensional finite element model of bolt-manhole-container body was meshed using a structured meshing method, and the mesh of bolt head, thread and contact surface was refined. Step 1.3, set the parameters; Set material parameters and boundary constraints according to the actual material and condition of the pressure vessel, and define the material density, elastic modulus, Poisson's ratio, yield strength, tensile strength and linear expansion coefficient, as well as the contact constraint type and friction coefficient between bolts and flanges. Step 1.4, Working condition simulation; For typical operating vibration conditions of pressure vessels in hydropower stations, cross vibration simulations were conducted using several manhole bolts as the object, with preloads of 180kN, 200kN, 220kN, and 240kN. Step 1.5, analyze vibration characteristics; Modal analysis was used to extract the first three natural frequencies of bolt vibration, harmonic response analysis was used to extract the vibration amplitude under different excitation loads, and probability density analysis was used to calculate the system output entropy corresponding to the vibration signal. Modal analysis is performed as follows: The calculation of the first three natural frequencies and mode shapes of the bolt is related to the bolt's equivalent mass and equivalent stiffness: (1), In the formula, K eff For equivalent stiffness, M eff For equivalent quality, i Take 1, 2 and 3; The harmonic response analysis method is as follows: Based on the natural frequency, the displacement amplitude of the bolt is calculated under different actual vibration loads. The bolt system is subjected to harmonic loads. F ( t )= F 0e jwt The steady-state response equation under the action is: (2), In the formula, C Let { be the damping matrix, { u} represents the displacement response vector, | u ( w | represents the displacement amplitude to be determined; Probability density analysis is used to calculate the system output entropy, as follows: The kernel density estimation method is used to solve the time-domain signal of bolt vibration displacement. d ( t The probability density function of ) p ( dThe kernel function chosen is the Gaussian kernel function, and the system output entropy is calculated based on the definition of information entropy. S : (3), (4), In the formula, N This represents the number of sampling points for the displacement signal. d i For the first i Displacement values ​​at each sampling point σ For kernel function bandwidth, K For the kernel function, the Gaussian kernel function is selected; Step 1.6: Construct a database linking preload, vibration characteristics, and system output entropy; Construct a standardized correlation database of vibration data under different bolt preload combinations, namely, a correlation database of preload, vibration characteristics, and system output entropy.

[0012] In step 1.4, the manhole bolts are M30 grade, there are 20 of them, and the designed preload is 220kN.

[0013] Step 2 is implemented in the following steps: Step 2.1, build the initial AI agent model; An initial AI agent model was built using a deep learning framework. Based on the ResNet-50 network, an SE-Net attention mechanism layer was added after the 3rd and 4th residual blocks. The activation function was uniformly ReLU, and a Dropout layer was added between the pooling layer and the fully connected layer to suppress overfitting. In the Dropout layer, dropout=0.5. The model input layer corresponds to the vibration characteristics of the bolts and the system output entropy, while the output layer corresponds to the preload value of the bolt group and the coordinates of the loosening position. Step 2.2, process the dataset; The database relating preload force, vibration characteristics, and system output entropy was divided into training, validation, and test sets in a 7:2:1 ratio. ±5% Gaussian random noise was added to each feature data in the database to augment the dataset. Step 2.3: Set the initial AI agent model training parameters; The initial AI agent model is trained using gradient descent, with an adaptive learning rate strategy and regularization mechanism. The loss function is the mean squared error loss function, and a weighted term is introduced to modify the loss function to improve the preload prediction accuracy. The modified formula is as follows: (5), In the formula, F i This represents the true value of the preload force. This is the predicted value of the preload force. X i Loose position X The true value of the coordinates, Loose position X Predicted coordinates Y i Loose position Y The true value of the coordinates, Loose position Y Predicted values ​​of coordinates; Step 2.4: Initial AI agent model training and optimization to obtain the AI ​​agent model; The training set data processed in step 2.2 is input into the initial AI agent model for iterative training. After each round of training, the accuracy of the AI ​​agent model is verified using the validation set data. When the loss function value of the validation set no longer decreases after several rounds, training is stopped and the optimal model parameters are saved. An early stopping mechanism is also introduced. Finally, the AI ​​agent model is obtained. Step 2.5: Verify the accuracy of the AI ​​agent model; The AI ​​agent model was validated using test set data to ensure that the model's preload prediction error and loosening location positioning error met the monitoring requirements. Step 2.6: Deploy the AI ​​agent model, quantify and identify bolt preload and loosening location, and output the bolt quantitative identification results; The AI ​​agent model is deployed to the monitoring terminal, input and output interfaces are set, redundant convolution channels are removed, and the TensorRT inference engine is used for acceleration. The vibration feature parameters extracted in step 3 are received in real time, and the bolt quantization identification results are output. The input interface receives vibration characteristic parameters and system output entropy, while the output interface outputs bolt number, preload value, and loosening position coordinates.

[0014] Step 3 is implemented in the following steps: Step 3.1, Equipment Deployment; Deploy cameras at appropriate locations in the manholes of pressure vessels to capture complete images of all bolts in the manholes; Step 3.2, Algorithm Settings; The algorithm is set in the camera's built-in system, and an optimized sparse optical flow feature matching algorithm is installed. An adaptive threshold adjustment mechanism, corner detection optimization strategy, multi-feature point linkage tracking and Kalman filtering algorithm are introduced. Step 3.3, Video capture; Start the camera to continuously capture and store real-time vibration videos of the bolts. Step 3.4, video processing and feature parameter extraction, to obtain vibration feature parameters and system output entropy; Vibration characteristic parameters include the first three frequencies f1, f2, and f3, and the amplitude A; Step 3.5, Data Output; The vibration characteristic data of each extracted bolt and the system output entropy are output to the AI ​​agent model in real time.

[0015] Step 3.4 is implemented in accordance with the following steps: Step 3.4.1: Video frame preprocessing to obtain video grayscale frames; The acquired vibration video stream is processed frame by frame. Color frames are converted to grayscale frames through grayscale processing, and then Gaussian filtering algorithm is used for noise reduction to filter out noise interference from the field environment and obtain grayscale video frames. The specific formula for grayscale conversion is: (6), In the formula, R ( x,y ), G ( x,y ), B ( x,y ) are respectively in the video frame ( x,y The red, green, and blue pixel values ​​at the coordinates. Gray ( x,y () represents the grayscale value at that coordinate. The specific formula for Gaussian filtering algorithm for noise reduction is as follows: (7), In the formula, The standard deviation is Gaussian. Step 3.4.2, Selection and tracking of bolt feature points; The video grayscale frames are divided into regions. A square region larger than the bolt diameter is selected as the independent analysis region, centered on the head of each bolt. Overlapping areas between bolts are segmented using the centerline, and non-bolt areas are directly removed. Then, a gradient-based corner detection algorithm is used, based on the pixel gradient product of the bolt corners. L The feature points are selected based on their maximum characteristics, and the coordinates of each feature point in the initial frame are recorded. x 0 ,y 0): (8), In the formula, L x For pixels in x Pixel gradient in direction, L y For pixels in y Pixel gradient in direction; Subsequently, the optimized sparse optical flow algorithm, combined with the Kalman filter algorithm, is used to track the coordinates of each feature point in each subsequent frame in real time. x t ,y t ),in t The frame number, t =1,2,3...; The specific steps are shown in formulas (9)-(12): First, a neighborhood window is set for each feature point, and the values ​​of each pixel within the neighborhood window are calculated. x direction, y Gray-scale gradient in direction and time direction I x 、I y and I t : (9), (10) (11), In the formula, G ( x,y,t ) is the first t frame( x,y The grayscale value at the coordinate; Based on the assumption of constant pixel grayscale followed by the sparse optical flow method, the optical flow equation is derived as follows: (12), In the formula, ( u,v ) represents the optical flow vector of the feature point; By solving the optical flow vector of all pixels within the neighborhood window using the least squares method, and obtaining the displacement increment of the corresponding feature point, the feature point can be iteratively updated at the 1st... t Frame coordinates ( x t ,y t )=( x t-1 +u,y t-1 +v ); Simultaneously, the Kalman filter method is used to filter out random fluctuations in feature points. The core formulas include the state equation and the observation equation. The state equation is: (13) In the formula, X k For the first k The state vector of a frame feature point, including coordinates ( xk ,y k ) and speed ( v k ,y k ) A Here is the state transition matrix: A = (14) In the formula, Δ t The time interval between two adjacent frames. w k This is process noise; The observation equation is: (15) In the formula, Z k For the first k Observation coordinates of frame feature points ( x k ,y k ), H The observation matrix; The observation matrix is: H = (16) In the formula, Vk represents the observation noise; the precise coordinates of each feature point in each frame are obtained through the prediction-update iteration of Kalman filtering. Step 3.4.3: Construct the displacement-time curve for each feature point of each bolt. d(t) Calculate the corresponding system output entropy; For each feature point of each bolt, calculate the displacement change d of that feature point based on its precise coordinates in consecutive frames. t The calculation formula is: (17) In time t The horizontal axis represents the displacement. dt Plot the displacement-time curve for each feature point with the vertical axis as the ordinate. d(t) This curve represents the time-domain signal of bolt vibration; then, the corresponding system output entropy is calculated based on the displacement information of the bolt group. Step 3.4.4: Frequency domain transformation and frequency extraction to obtain the amplitude |D( ) corresponding to each frequency point. k | with the first three natural frequencies f1, f2, f3; The displacement-time curve of the time-domain displacement signal is obtained by using FFT. d(t)Converting the signal to the frequency domain and extracting the frequency parameters of the bolt vibration, the core formula is the Discrete Fourier Transform: (18) In the formula, D(k) For the complex amplitude of the frequency domain signal, N The number of sampling points. k For frequency point number, k =0,1,2,...,N-1; After the frequency domain transformation is completed, calculate the amplitude |D( ) corresponding to each frequency point. k Select the frequencies f1, f2, and f3 corresponding to the three frequency points with the largest amplitudes. These are the first three natural frequencies of the bolt vibration, where f1 is the first natural frequency, f2 is the second natural frequency, and f3 is the third natural frequency. Step 3.4.5: Vibration amplitude extraction to obtain the amplitude A of bolt vibration; The vibration amplitude is the maximum displacement of the bolt vibration, which can be obtained through time-domain amplitude and frequency-domain amplitude; the time-domain amplitude is obtained by directly extracting the displacement-time curve. d(t) The maximum peak value in the range; the frequency domain amplitude is the frequency domain amplitude values ​​|D(k1)|, |D(k2)|, and |D(k3)| corresponding to the first three natural frequencies, and the maximum value is taken as the frequency domain amplitude. The amplitude A of bolt vibration is obtained by fusing the time-domain and frequency-domain amplitudes using a weighted average method. Step 3.4.6, Verification and correction of vibration characteristic parameters; The vibration characteristic parameters are verified and outliers are removed. Finite element simulations are performed on the vibration characteristic parameters, and the finite element simulation results are compared with the vibration characteristic parameters to correct the extraction error and ensure that the deviation between the frequency and amplitude extracted on site and the simulation data is ≤5%.

[0016] In step 3.4.2, the side length of the square region is 1.5 times the bolt diameter.

[0017] Step 4 is implemented in the following steps: Step 4.1, Data reception; By monitoring the terminal, the bolt quantization results output by the AI ​​agent model, as well as the vibration characteristic parameters and system output entropy, are received synchronously, and a one-to-one correspondence between the quantization results of each bolt and the vibration characteristic parameters is established. Step 4.2, Data Validation; Adopt 3 The criteria remove outliers from the data received in step 4.1 by calculating the average first-order vibration frequency of each bolt. and standard deviation This will exceed [ -3 , +3 Data within the specified range was identified as outliers and removed. Step 4.3, standardize the output; In accordance with the fastener monitoring standards for hydropower stations, a standardized quantitative monitoring data format for bolt loosening was developed, which includes bolt number, acquisition time, preload value, loosening degree level, loosening location coordinates, vibration characteristic parameters, system output entropy, and data verification results. The data was then compiled into standardized monitoring reports and output to the hydropower station operation and maintenance monitoring center in real time, while also being stored in the database.

[0018] The second technical solution adopted in this invention is a storage medium on which a computer program is stored. When the computer program is executed by a processor, it implements the steps of a video-based quantitative monitoring method for manhole bolt loosening in hydropower station pressure vessels.

[0019] The third technical solution adopted in this invention is a computer device, including a memory, a processor, and a computer program stored in the memory and capable of running on the processor. When the computer program is processed and executed, it implements the steps of a video-based quantitative monitoring method for manhole bolt loosening in hydropower station pressure vessels.

[0020] The beneficial effects of this invention are: (1) The present invention can realize precise quantitative identification and location of bolt loosening, effectively breaking through the technical limitation of existing technologies that can only make qualitative judgments.

[0021] (2) The present invention adopts a non-contact monitoring mode, which is convenient to operate and maintain, has low monitoring cost, and does not interfere with the normal operation of the pressure vessel.

[0022] (3) The present invention has strong adaptability to working conditions and can be adapted to bolt monitoring in various complex vibration modes and operating environments.

[0023] (4) The present invention has outstanding anti-interference ability, high operational stability, timely early warning response, and can effectively prevent safety accidents caused by loose bolts.

[0024] (5) This invention has a wide range of applications and strong scalability, and can provide support for the safe operation and maintenance of similar pressure vessels in industries such as iron and steel metallurgy. Attached Figure Description

[0025] Figure 1 This is a schematic diagram of the technical path of the method of the present invention; Figure 2 This is a schematic diagram of the finite element simulation of bolt vibration according to the present invention; Figure 3 This is a schematic diagram of the bolt group preload prediction process of the present invention. Detailed Implementation

[0026] The following detailed description is provided in conjunction with specific implementation methods.

[0027] like Figure 1 As shown, a video-based quantitative monitoring method for loosening manhole bolts in hydropower station pressure vessels is presented. Step 1 involves simulating the working conditions by conducting finite element simulations of bolt vibration and constructing a database of preload and vibration. This leads to the establishment and training of an AI proxy model in Step 2. Step 3 involves obtaining vibration characteristics of feature points through video monitoring of the bolt group and processing the video grayscale frames. This includes the vibration parameters of each bolt under actual operating conditions and the system output entropy. In Step 4, based on the AI ​​proxy model from Step 2 and the vibration characteristics extracted in Step 3, the AI ​​proxy model calculates the bolt group's early warning force information, including the loosening location and preload. Step 4 ends when the bolt is not loose; otherwise, it calculates the loosening location and preload and ends.

[0028] A video-based quantitative monitoring method for loose manhole bolts in hydropower station pressure vessels includes the following steps: Step 1: Construct a finite element model of bolt-manhole-container and build a database relating preload, vibration characteristics, and system output entropy.

[0029] Based on the actual structural parameters, material properties, and connection methods of the manhole bolts, manhole flanges, and vessel body of the hydropower station pressure vessel, an integrated three-dimensional finite element model of bolts-manhole-vessel body was built. Vibration simulation was carried out under simulated working conditions, and vibration characteristic parameters and system output entropy were extracted simultaneously. A quantitative correlation database of preload, vibration characteristics, and system output entropy was constructed.

[0030] Step 1 is implemented in the following steps: Step 1.1: Construct an integrated three-dimensional finite element model of bolts, manholes, and container body.

[0031] Draw solid models of each component, define the connection relationships between bolts and flanges, and between flanges and container body according to the actual connection method, and construct an integrated three-dimensional finite element model of bolts-manhole-container body.

[0032] Use Solidworks for model drawing and construction.

[0033] Step 1.2, Mesh generation.

[0034] like Figure 2 As shown, the integrated three-dimensional finite element model of bolt-manhole-container body is meshed using a structured meshing method, and the bolt head, thread and contact surface are further refined.

[0035] Mesh generation is performed in ANSYS.

[0036] Step 1.3, set the parameters.

[0037] Set material parameters and boundary constraints according to the actual material and condition of the pressure vessel, and define the material density, elastic modulus, Poisson's ratio, yield strength, tensile strength and linear expansion coefficient, as well as the contact constraint type and friction coefficient between bolts and flanges.

[0038] Step 1.4, Working condition simulation.

[0039] For typical operating vibration conditions of pressure vessels in hydropower stations, cross-vibration simulations were conducted using 20 M30 grade manhole bolts as the object, with preloads of 180kN, 200kN, 220kN, and 240kN.

[0040] The design preload is 220kN.

[0041] Step 1.5, analyze the vibration characteristics.

[0042] Modal analysis was used to extract the first three natural frequencies of bolt vibration, harmonic response analysis was used to extract the vibration amplitude under different excitation loads, and probability density analysis was used to calculate the system output entropy corresponding to the vibration signal. Modal analysis is performed as follows: Calculating the first three natural frequencies and mode shapes of a bolt is an inherent property of the bolt, and is only related to the bolt's equivalent mass and equivalent stiffness. (1), In the formula, K eff For equivalent stiffness, M eff For equivalent quality, i Take 1, 2 and 3.

[0043] The harmonic response analysis method is as follows: Based on the natural frequency, calculate the displacement amplitude of the bolt under different actual vibration loads. The bolt system under harmonic loads... F ( t )= F 0e jwt The steady-state response equation under the action is: (2), In the formula, C Let { be the damping matrix, { u} represents the displacement response vector, | u ( w )| represents the displacement amplitude to be determined.

[0044] Probability density analysis is used to calculate the system output entropy, as follows: The system output entropy can quantitatively reflect the impact of bolt preload changes on the overall system stability. The kernel density estimation method is used to solve the time-domain signal of bolt vibration displacement. d ( t The probability density function of ) p ( d The kernel function chosen is the Gaussian kernel function, and finally, the system output entropy is calculated based on the definition of information entropy. S : (3), (4), In the formula, N This represents the number of sampling points for the displacement signal. d i For the first i Displacement values ​​at each sampling point σ For kernel function bandwidth, K For the kernel function, we choose the Gaussian kernel function.

[0045] Step 1.6: Construct a database linking preload, vibration characteristics, and system output entropy.

[0046] Construct a standardized correlation database of vibration data under different bolt preload combinations, namely, a correlation database of preload, vibration characteristics, and system output entropy.

[0047] Step 2: Build and train the AI ​​agent model, deploy the AI ​​agent model to the monitoring terminal, quantify and identify the bolt preload and loosening position, and output the bolt quantitative identification results.

[0048] A deep learning model of preload and loosening degree was built. Based on the associated database of finite element simulation output, the deep learning model was trained with vibration parameters and system output entropy as input and bolt group preload as output. The mapping relationship between bolt vibration characteristics, system output entropy and loosening position and preload loosening degree was constructed, and an AI agent model was established.

[0049] Step 2 is implemented in the following steps: Step 2.1: Build the initial AI agent model.

[0050] An initial AI proxy model was built using a deep learning framework. Based on the ResNet-50 network, an SE-Net attention mechanism layer was added after the 3rd and 4th residual blocks to improve the ability to capture subtle changes in vibration characteristics. The activation function was uniformly ReLU, and a Dropout layer (dropout=0.5) was added between the pooling layer and the fully connected layer to suppress overfitting. The model input layer corresponds to the vibration characteristics of the bolt and the system output entropy, while the output layer corresponds to the preload value of the bolt group and the coordinates of the loosening position.

[0051] Step 2.2, process the dataset.

[0052] The database of preload and vibration characteristics and system output entropy constructed in step 1 was divided into training set, validation set and test set in a ratio of 7:2:1. ±5% Gaussian random noise was added to each feature data in the database to enhance the dataset and avoid model overfitting.

[0053] Step 2.3: Set the initial AI agent model training parameters.

[0054] The initial AI agent model is trained using gradient descent, employing an adaptive learning rate strategy and regularization mechanism to dynamically adjust the learning rate based on the situation. The mean squared error loss function is used, and a weighted term is introduced to modify the loss function and improve the preload prediction accuracy. The modified formula is as follows: (5), In the formula, F i This represents the true value of the preload force. This is the predicted value of the preload force. X i Loose position X The true value of the coordinates, Loose position X Predicted coordinates Y i Loose position Y The true value of the coordinates, Loose position Y Predicted values ​​of coordinates.

[0055] Step 2.4: Initial AI agent model training and optimization to obtain the AI ​​agent model.

[0056] Input the training set data processed in step 2.2 into the initial AI agent model for iterative training. After each round of training, use the validation set data to verify the accuracy of the AI ​​agent model. When the loss function value of the validation set no longer decreases after several rounds, stop training and save the optimal model parameters. At the same time, an early stopping mechanism is introduced to avoid model overfitting, ensure the model's generalization ability, and finally obtain the AI ​​agent model.

[0057] Step 2.5, verify the accuracy of the AI ​​agent model.

[0058] The AI ​​agent model was validated using test set data to ensure that the model's preload prediction error and loosening location positioning error met the monitoring requirements. Step 2.6: Deploy the AI ​​agent model, quantify and identify bolt preload and loosening location, and output the bolt quantitative identification results.

[0059] The AI ​​agent model is deployed to the monitoring terminal. The model input interface (receiving vibration feature parameters and system output entropy) and output interface (outputting bolt number, preload value, and loosening position coordinates) are set. Redundant convolution channels are removed, and the TensorRT inference engine is used to accelerate the process, ensuring that the model can receive the vibration feature parameters extracted in step 3 in real time and quickly output the bolt quantization recognition results.

[0060] Step 3: Acquire bolt vibration video, calculate system output entropy and extract vibration characteristic parameters.

[0061] A bolt vibration monitoring system based on video analytics was constructed to achieve on-site bolt vibration video acquisition. A markerless video vibration measurement and analysis method for bolt vibration was proposed to calculate the vibration parameters of each bolt under actual operating conditions and the system output entropy.

[0062] Step 3 is implemented in the following steps: Step 3.1, Equipment Deployment.

[0063] Deploy high-definition industrial cameras at appropriate locations in the manholes of pressure vessels, and adjust the camera angles to ensure that a clear, complete image of all bolts in the manhole can be captured, avoiding blind spots.

[0064] Step 3.2, Algorithm settings.

[0065] The camera's built-in system incorporates an optimized sparse optical flow feature matching algorithm, along with an adaptive threshold adjustment mechanism, corner detection optimization strategy, multi-feature point linkage tracking, and Kalman filtering algorithm. This allows the algorithm parameters to be automatically adjusted based on the environment, using corner detection to select bolt heads and edges as stable feature points, avoiding feature point loss or mismatches, and filtering out vibration interference.

[0066] Step 3.3, video capture.

[0067] Start the camera and continuously capture real-time vibration video of the bolt. Monitor the camera's working status in real time during the acquisition process to ensure no disconnection or blurring, and store the video data in a standard format.

[0068] Step 3.4: Video processing and feature parameter extraction to obtain vibration feature parameters and system output entropy.

[0069] The camera is activated to continuously capture real-time vibration videos of the bolt. The video frames are processed, corner points are detected, sparse optical flow is tracked, displacement signals are acquired, and parameters are extracted after frequency domain transformation to obtain vibration characteristic parameters and system output entropy. The vibration characteristic parameters include the first three frequencies f1, f2, and f3, and the amplitude A.

[0070] Step 3.4 includes the following steps: Step 3.4.1: Video frame preprocessing to obtain video grayscale frames.

[0071] The acquired vibration video stream is processed frame by frame. First, color frames are converted to grayscale frames through grayscale conversion to eliminate color interference and improve processing efficiency. The specific formula is as follows: (6), In the formula, R ( x,y ), G ( x,y ), B ( x,y ) are respectively in the video frame ( x,y The red, green, and blue pixel values ​​at the coordinates. Gray ( x,y ) represents the grayscale value at that coordinate.

[0072] Subsequently, a Gaussian filtering algorithm is used for noise reduction to filter out noise interference from the on-site environment (such as dust and light fluctuations). The specific formula is as follows: (7), In the formula, The standard deviation is Gaussian.

[0073] This step yields video grayscale frames, namely a clear, interference-free bolt grayscale frame sequence.

[0074] Step 3.4.2, Selection and tracking of bolt feature points.

[0075] The video grayscale frames are divided into regions. A square region 1.5 times the bolt diameter, centered on the head of each bolt, is selected as an independent analysis region. Overlapping areas between bolts are segmented using the centerline, and non-bolt areas are directly discarded. Then, a gradient-based corner detection algorithm is used, based on the pixel gradient product of the bolt corners. L The feature points are selected based on their maximum characteristics, and the coordinates of each feature point in the initial frame are recorded. x 0 ,y 0): (8), In the formula, L x For pixels in x Pixel gradient in direction, L y For pixels in y Pixel gradient in direction.

[0076] Subsequently, the optimized sparse optical flow algorithm, combined with the Kalman filter algorithm, is used to track the coordinates of each feature point in each subsequent frame in real time. x t,y t )( t The frame number, t =1,2,3...), the specific steps are shown in formulas (9)-(12).

[0077] First, set a neighborhood window of appropriate size for each feature point, and calculate the value of each pixel within the neighborhood window according to the following formula. x direction, y Gray-scale gradient in direction and time (adjacent frames) I x 、I y and I t : (9), (10) (11), In the formula, G ( x,y,t ) is the first t frame( x,y The grayscale value at the coordinate.

[0078] Based on the assumption of constant pixel grayscale followed by the sparse optical flow method, the optical flow equation is derived as follows: (12), In the formula, ( u,v ) represents the optical flow vector of the feature point.

[0079] By solving the optical flow vector of all pixels within the neighborhood window using the least squares method, and obtaining the displacement increment of the corresponding feature point, the feature point can be iteratively updated at the 1st... t Frame coordinates ( x t ,y t )=( x t-1 +u,y t-1 +v ).

[0080] Simultaneously, the Kalman filter method is used to filter out random jitter of feature points, ensuring tracking accuracy. Its core formulas include the state equation and the observation equation. The state equation is: (13) In the formula, X k For the first k The state vector of frame feature points (including coordinates) x k,y k ) and speed ( v k ,y k )), A Here is the state transition matrix: A = (14) In the formula, Δ t The time interval between two adjacent frames. w k This is process noise. w k Adjust according to the on-site vibration intensity.

[0081] The observation equation is: (15) In the formula, Z k For the first k Observation coordinates of frame feature points ( x k ,y k ), H The observation matrix; The observation matrix is: H = (16) In the formula, Vk represents the observation noise. Through Kalman filtering and prediction-update iteration, the precise coordinates of each feature point in each frame are obtained, avoiding feature point shifts caused by vibration interference.

[0082] Step 3.4.3: Construct the displacement-time curve for each feature point of each bolt. d(t) Calculate the corresponding system output entropy.

[0083] For each feature point of each bolt, calculate the displacement change d of that feature point based on its precise coordinates in consecutive frames. t The calculation formula is: (17) In time t The horizontal axis represents the displacement. dt Plot the displacement-time curve for each feature point with the vertical axis as the ordinate. d(t) This curve is the time-domain signal of bolt vibration, which intuitively reflects the vibration change law of bolt over time.

[0084] Then, the corresponding system output entropy is calculated based on the displacement information of the bolt group.

[0085] Step 3.4.4: Frequency domain transformation and frequency extraction to obtain the amplitude |D( ) corresponding to each frequency point. k )| with the first three natural frequencies f1, f2, f3.

[0086] The displacement-time curve of the time-domain displacement signal is obtained by using Fast Fourier Transform (FFT). d(t) The signal is converted to the frequency domain to extract the frequency parameters of the bolt vibration. The core formula is the Discrete Fourier Transform: (18) In the formula, D(k) For the complex amplitude of the frequency domain signal, N The number of sampling points. k Frequency point number ( k =0,1,2,...,N-1).

[0087] After the frequency domain transformation is completed, calculate the amplitude |D( ) corresponding to each frequency point. k The frequencies f1, f2, and f3 corresponding to the three frequency points with the largest amplitude are selected as the first three natural frequencies (core frequency parameters) of bolt vibration. Among them, f1 is the first natural frequency, f2 is the second natural frequency, and f3 is the third natural frequency. f1 is the frequency parameter most closely related to the bolt preload.

[0088] Step 3.4.5: Vibration amplitude extraction, obtain the amplitude A of bolt vibration.

[0089] The vibration amplitude is the maximum displacement of the bolt vibration, which can be obtained through two dimensions (time domain amplitude and frequency domain amplitude) to ensure accuracy.

[0090] Time-domain amplitude: Direct extraction of displacement-time curve d(t) The maximum peak value in; Frequency domain amplitude: Extract the frequency domain amplitudes |D(k1)|, |D(k2)|, and |D(k3)| corresponding to the first three natural frequencies, and take the maximum value as the frequency domain amplitude.

[0091] Finally, the amplitude A of bolt vibration was obtained by fusing the time-domain and frequency-domain amplitudes using a weighted average method.

[0092] Step 3.4.6, Verification and correction of vibration characteristic parameters.

[0093] The vibration characteristic parameters (i.e., the first three extracted frequencies f1, f2, f3 and amplitude A) are verified, outliers are removed, and the parameters are ensured to be accurate. At the same time, the vibration characteristic parameters are simulated by finite element method, and the finite element simulation results are compared with the vibration characteristic parameters to correct the extraction error and ensure that the deviation between the frequency and amplitude extracted on site and the simulation data is ≤5%, so as to provide reliable input for subsequent AI model recognition.

[0094] Step 3.5, Data Output.

[0095] The vibration characteristic data of each extracted bolt (the first three frequencies f1, f2, f3, and amplitude A) and the system output entropy are output to the AI ​​agent model in real time. A stable data transmission method is adopted to ensure that the transmission stability and transmission delay meet the requirements.

[0096] like Figure 3 As shown, step 3 obtains the original data of the bolt group through video acquisition, converts the video frames to grayscale to obtain grayscale frames, divides the grayscale frames into regions, and calculates the vibration feature data (first three frequencies f1, f2, f3, amplitude A) and system output entropy of each feature point of each bolt. The initial AI proxy model is trained and optimized using gradient descent and the dataset to obtain the AI ​​proxy model. The extracted vibration feature data (first three frequencies f1, f2, f3, amplitude A) and system output entropy of each bolt are output to the AI ​​proxy model in real time, quantifying the bolt group preload prediction and loosening location.

[0097] Step 4: Verify monitoring data and standardize the output of bolt loosening quantitative data.

[0098] Based on the AI ​​proxy model in step 2, and using the bolt group vibration characteristic parameters and system output entropy extracted in step 3 as inputs, the loosening position and preload are calculated through the AI ​​proxy model, thereby achieving end-to-end quantitative identification of bolt group preload prediction and loosening position location.

[0099] Step 4 is implemented in the following steps: Step 4.1, Data reception.

[0100] By monitoring the terminal, the bolt quantization results output by the AI ​​agent model are received synchronously, along with the vibration characteristic parameters and system output entropy calculated in step 3, establishing a one-to-one correspondence between the quantization results and vibration characteristic parameters of each bolt.

[0101] Step 4.2, data verification.

[0102] Adopt 3 The criteria remove outliers from the data received in step 4.1 by calculating the average first-order vibration frequency of each bolt. and standard deviation This will exceed [ -3 , +3 Data within the specified range is identified as outliers and removed.

[0103] Step 4.3, standardize the output.

[0104] In accordance with the fastener monitoring standards for hydropower stations, a standardized quantitative monitoring data format for bolt loosening was developed, which includes bolt number, acquisition time, preload value, loosening degree level, loosening location coordinates, vibration characteristic parameters, system output entropy, and data verification results. The data was then compiled into standardized monitoring reports and output to the hydropower station operation and maintenance monitoring center in real time, while also being stored in the database.

[0105] A storage medium containing a computer program, which, when executed by a processor, implements the steps of a video-based quantitative monitoring method for manhole bolt loosening in hydropower station pressure vessels.

[0106] Computer equipment, including memory, processor, and computer program stored in memory and executable on the processor, wherein the computer program, when processed and executed, implements the steps of a video-based quantitative monitoring method for manhole bolt loosening in hydropower station pressure vessels.

[0107] Example 1 A video-based quantitative monitoring method for loose manhole bolts in hydropower station pressure vessels includes the following steps: Step 1: Construct a finite element model of bolt-manhole-container and build a database relating preload and vibration characteristics to system output entropy; Step 2: Build and train the AI ​​agent model, deploy the AI ​​agent model to the monitoring terminal, quantify and identify the bolt preload and loosening position, and output the bolt quantitative identification results; Step 3: Acquire bolt vibration video, calculate system output entropy and extract vibration characteristic parameters; Step 4: Verify monitoring data and standardize the output of bolt loosening quantitative data.

[0108] Example 2 A video-based quantitative monitoring method for loose manhole bolts in hydropower station pressure vessels includes the following steps: Step 1: Construct a finite element model of bolt-manhole-container and build a database relating preload and vibration characteristics to system output entropy; Step 2: Build and train the AI ​​agent model, deploy the AI ​​agent model to the monitoring terminal, quantify and identify the bolt preload and loosening position, and output the bolt quantitative identification results; Step 3: Acquire bolt vibration video, calculate system output entropy and extract vibration characteristic parameters; Step 4: Verify monitoring data and standardize the output of bolt loosening quantitative data.

[0109] Step 1 is implemented in the following steps: Step 1.1: Construct an integrated three-dimensional finite element model of the bolt, manhole, and container body; Draw solid models of each component, define the connection relationships between bolts and flanges, and between flanges and container body according to the actual connection methods, and construct an integrated three-dimensional finite element model of bolts-manhole-container body; Step 1.2, Mesh generation; The integrated three-dimensional finite element model of bolt-manhole-container body was meshed using a structured meshing method, and the mesh of bolt head, thread and contact surface was refined. Step 1.3, set the parameters; Set material parameters and boundary constraints according to the actual material and condition of the pressure vessel, and define the material density, elastic modulus, Poisson's ratio, yield strength, tensile strength and linear expansion coefficient, as well as the contact constraint type and friction coefficient between bolts and flanges. Step 1.4, Working condition simulation; For typical operating vibration conditions of pressure vessels in hydropower stations, cross vibration simulations were conducted using several manhole bolts as the object, with preloads of 180kN, 200kN, 220kN, and 240kN. Step 1.5, analyze vibration characteristics; Modal analysis was used to extract the first three natural frequencies of bolt vibration, harmonic response analysis was used to extract the vibration amplitude under different excitation loads, and probability density analysis was used to calculate the system output entropy corresponding to the vibration signal. Modal analysis is performed as follows: The calculation of the first three natural frequencies and mode shapes of the bolt is related to the bolt's equivalent mass and equivalent stiffness: (1), In the formula, K eff For equivalent stiffness, M eff For equivalent quality, i Take 1, 2 and 3; The harmonic response analysis method is as follows: Based on the natural frequency, the displacement amplitude of the bolt is calculated under different actual vibration loads. The bolt system is subjected to harmonic loads. F ( t )= F 0e jwt The steady-state response equation under the action is: (2), In the formula, C Let { be the damping matrix, { u} represents the displacement response vector, | u ( w | represents the displacement amplitude to be determined; Probability density analysis is used to calculate the system output entropy, as follows: The kernel density estimation method is used to solve the time-domain signal of bolt vibration displacement. d ( t The probability density function of ) p ( d The kernel function chosen is the Gaussian kernel function, and the system output entropy is calculated based on the definition of information entropy. S : (3), (4), In the formula, N This represents the number of sampling points for the displacement signal. d i For the first i Displacement values ​​at each sampling point σ For kernel function bandwidth, K For the kernel function, the Gaussian kernel function is selected; Step 1.6: Construct a database linking preload, vibration characteristics, and system output entropy; Construct a standardized correlation database of vibration data under different bolt preload combinations, namely, a correlation database of preload, vibration characteristics, and system output entropy.

[0110] Example 3 A video-based quantitative monitoring method for loose manhole bolts in hydropower station pressure vessels includes the following steps: Step 1: Construct a finite element model of bolt-manhole-container and build a database relating preload and vibration characteristics to system output entropy; Step 2: Build and train the AI ​​agent model, deploy the AI ​​agent model to the monitoring terminal, quantify and identify the bolt preload and loosening position, and output the bolt quantitative identification results; Step 3: Acquire bolt vibration video, calculate system output entropy and extract vibration characteristic parameters; Step 4: Verify monitoring data and standardize the output of bolt loosening quantitative data.

[0111] Step 1 is implemented in the following steps: Step 1.1: Construct an integrated three-dimensional finite element model of the bolt, manhole, and container body; Draw solid models of each component, define the connection relationships between bolts and flanges, and between flanges and container body according to the actual connection methods, and construct an integrated three-dimensional finite element model of bolts-manhole-container body; Step 1.2, Mesh generation; The integrated three-dimensional finite element model of bolt-manhole-container body was meshed using a structured meshing method, and the mesh of bolt head, thread and contact surface was refined. Step 1.3, set the parameters; Set material parameters and boundary constraints according to the actual material and condition of the pressure vessel, and define the material density, elastic modulus, Poisson's ratio, yield strength, tensile strength and linear expansion coefficient, as well as the contact constraint type and friction coefficient between bolts and flanges. Step 1.4, Working condition simulation; For typical operating vibration conditions of pressure vessels in hydropower stations, cross vibration simulations were conducted using several manhole bolts as the object, with preloads of 180kN, 200kN, 220kN, and 240kN. Step 1.5, analyze vibration characteristics; Modal analysis was used to extract the first three natural frequencies of bolt vibration, harmonic response analysis was used to extract the vibration amplitude under different excitation loads, and probability density analysis was used to calculate the system output entropy corresponding to the vibration signal. Modal analysis is performed as follows: The calculation of the first three natural frequencies and mode shapes of the bolt is related to the bolt's equivalent mass and equivalent stiffness: (1), In the formula, K eff For equivalent stiffness, M eff For equivalent quality, i Take 1, 2 and 3; The harmonic response analysis method is as follows: Based on the natural frequency, the displacement amplitude of the bolt is calculated under different actual vibration loads. The bolt system is subjected to harmonic loads. F ( t )= F 0e jwt The steady-state response equation under the action is: (2), In the formula, C Let { be the damping matrix, { u} represents the displacement response vector, | u ( w | represents the displacement amplitude to be determined; Probability density analysis is used to calculate the system output entropy, as follows: The kernel density estimation method is used to solve the time-domain signal of bolt vibration displacement. d ( t The probability density function of ) p ( d The kernel function chosen is the Gaussian kernel function, and the system output entropy is calculated based on the definition of information entropy. S : (3), (4), In the formula, N This represents the number of sampling points for the displacement signal. d i For the first i Displacement values ​​at each sampling point σ For kernel function bandwidth, K For the kernel function, the Gaussian kernel function is selected; Step 1.6: Construct a database linking preload, vibration characteristics, and system output entropy; Construct a standardized correlation database of vibration data under different bolt preload combinations, namely, a correlation database of preload, vibration characteristics, and system output entropy.

[0112] In step 1.4, the manhole bolts are M30 grade, there are 20 of them, and the designed preload is 220kN.

[0113] Example 4 A video-based quantitative monitoring method for loose manhole bolts in hydropower station pressure vessels includes the following steps: Step 1: Construct a finite element model of bolt-manhole-container and build a database relating preload and vibration characteristics to system output entropy; Step 2: Build and train the AI ​​agent model, deploy the AI ​​agent model to the monitoring terminal, quantify and identify the bolt preload and loosening position, and output the bolt quantitative identification results; Step 3: Acquire bolt vibration video, calculate system output entropy and extract vibration characteristic parameters; Step 4: Verify monitoring data and standardize the output of bolt loosening quantitative data.

[0114] Step 4 is implemented in the following steps: Step 4.1, Data reception; By monitoring the terminal, the bolt quantization results output by the AI ​​agent model, as well as the vibration characteristic parameters and system output entropy, are received synchronously, and a one-to-one correspondence between the quantization results of each bolt and the vibration characteristic parameters is established. Step 4.2, Data Validation; Adopt 3 The criteria remove outliers from the data received in step 4.1 by calculating the average first-order vibration frequency of each bolt. and standard deviation This will exceed [ -3 , +3 Data within the specified range was identified as outliers and removed. Step 4.3, standardize the output; In accordance with the fastener monitoring standards for hydropower stations, a standardized quantitative monitoring data format for bolt loosening was developed, which includes bolt number, acquisition time, preload value, loosening degree level, loosening location coordinates, vibration characteristic parameters, system output entropy, and data verification results. The data was then compiled into standardized monitoring reports and output to the hydropower station operation and maintenance monitoring center in real time, while also being stored in the database.

[0115] Example 5 A storage medium containing a computer program, which, when executed by a processor, implements the steps of a video-based quantitative monitoring method for manhole bolt loosening in hydropower station pressure vessels.

[0116] Example 6 Computer equipment, including memory, processor, and computer program stored in memory and executable on the processor, wherein the computer program, when processed and executed, implements the steps of a video-based quantitative monitoring method for manhole bolt loosening in hydropower station pressure vessels.

Claims

1. A video-based quantitative monitoring method for loosening manhole bolts in hydropower station pressure vessels, characterized in that... Includes the following steps: Step 1: Construct a finite element model of bolt-manhole-container and build a database relating preload and vibration characteristics to system output entropy; Step 2: Build and train the AI ​​agent model, deploy the AI ​​agent model to the monitoring terminal, quantify and identify the bolt preload and loosening position, and output the bolt quantitative identification results; Step 3: Acquire bolt vibration video, calculate system output entropy and extract vibration characteristic parameters; Step 4: Verify monitoring data and standardize the output of bolt loosening quantitative data.

2. The video-based quantitative monitoring method for loosening manhole bolts in hydropower station pressure vessels, as described in claim 1, is characterized in that... Step 1 is implemented in the following steps: Step 1.1: Construct an integrated three-dimensional finite element model of the bolt, manhole, and container body; Draw solid models of each component, define the connection relationships between bolts and flanges, and between flanges and container body according to the actual connection methods, and construct an integrated three-dimensional finite element model of bolts-manhole-container body; Step 1.2, Mesh generation; The integrated three-dimensional finite element model of bolt-manhole-container body was meshed using a structured meshing method, and the mesh of bolt head, thread and contact surface was refined. Step 1.3, set the parameters; Set material parameters and boundary constraints according to the actual material and condition of the pressure vessel, and define the material density, elastic modulus, Poisson's ratio, yield strength, tensile strength and linear expansion coefficient, as well as the contact constraint type and friction coefficient between bolts and flanges. Step 1.4, Working condition simulation; For typical operating vibration conditions of pressure vessels in hydropower stations, cross vibration simulations were conducted using several manhole bolts as the object, with preloads of 180kN, 200kN, 220kN, and 240kN. Step 1.5, analyze vibration characteristics; Modal analysis was used to extract the first three natural frequencies of bolt vibration, harmonic response analysis was used to extract the vibration amplitude under different excitation loads, and probability density analysis was used to calculate the system output entropy corresponding to the vibration signal. Modal analysis is performed as follows: The calculation of the first three natural frequencies and mode shapes of the bolt is related to the bolt's equivalent mass and equivalent stiffness: （1）， In the formula, K eff For equivalent stiffness, M eff For equivalent quality, i Take 1, 2 and 3; The harmonic response analysis method is as follows: Based on the natural frequency, the displacement amplitude of the bolt is calculated under different actual vibration loads. The bolt system is subjected to harmonic loads. F ( t )= F 0e jwt The steady-state response equation under the action is: （2）， In the formula, C Let { be the damping matrix, { u } represents the displacement response vector, | u ( w | represents the displacement amplitude to be determined; Probability density analysis is used to calculate the system output entropy, as follows: The kernel density estimation method is used to solve the time-domain signal of bolt vibration displacement. d ( t The probability density function of ) p ( d The kernel function chosen is the Gaussian kernel function, and the system output entropy is calculated based on the definition of information entropy. S : （3）， （4）， In the formula, N This represents the number of sampling points for the displacement signal. d i For the first i Displacement values ​​at each sampling point σ For kernel function bandwidth, K For the kernel function, the Gaussian kernel function is selected; Step 1.6: Construct a database linking preload, vibration characteristics, and system output entropy; Construct a standardized correlation database of vibration data under different bolt preload combinations, namely, a correlation database of preload, vibration characteristics, and system output entropy.

3. The video-based quantitative monitoring method for loosening manhole bolts in hydropower station pressure vessels according to claim 2, characterized in that, In step 1.4, the manhole bolts are M30 grade, there are 20 of them, and the designed preload is 220kN.

4. The video-based quantitative monitoring method for loosening manhole bolts in hydropower station pressure vessels, as described in claim 1, is characterized in that... Step 2 is implemented in the following steps: Step 2.1, build the initial AI agent model; An initial AI agent model was built using a deep learning framework. Based on the ResNet-50 network, an SE-Net attention mechanism layer was added after the 3rd and 4th residual blocks. The activation function was uniformly ReLU, and a Dropout layer was added between the pooling layer and the fully connected layer to suppress overfitting. In the Dropout layer, dropout=0.

5. The model input layer corresponds to the vibration characteristics of the bolts and the system output entropy, while the output layer corresponds to the preload value of the bolt group and the coordinates of the loosening position. Step 2.2, process the dataset; The database relating preload force, vibration characteristics, and system output entropy was divided into training, validation, and test sets in a 7:2:1 ratio. ±5% Gaussian random noise was added to each feature data in the database to augment the dataset. Step 2.3: Set the initial AI agent model training parameters; The initial AI agent model is trained using gradient descent, with an adaptive learning rate strategy and regularization mechanism. The loss function is the mean squared error loss function, and a weighted term is introduced to modify the loss function to improve the preload prediction accuracy. The modified formula is as follows: （5）， In the formula, F i This represents the true value of the preload force. This is the predicted value of the preload force. X i Loose position X The true value of the coordinates, Loose position X Predicted coordinates Y i Loose position Y The true value of the coordinates, Loose position Y Predicted values ​​of coordinates; Step 2.4: Initial AI agent model training and optimization to obtain the AI ​​agent model; The training set data processed in step 2.2 is input into the initial AI agent model for iterative training. After each round of training, the accuracy of the AI ​​agent model is verified using the validation set data. When the loss function value of the validation set no longer decreases after several rounds, training is stopped and the optimal model parameters are saved. An early stopping mechanism is also introduced. Finally, the AI ​​agent model is obtained. Step 2.5: Verify the accuracy of the AI ​​agent model; The AI ​​agent model was validated using test set data to ensure that the model's preload prediction error and loosening location positioning error met the monitoring requirements. Step 2.6: Deploy the AI ​​agent model, quantify and identify bolt preload and loosening location, and output the bolt quantitative identification results; The AI ​​agent model is deployed to the monitoring terminal, input and output interfaces are set, redundant convolution channels are removed, and the TensorRT inference engine is used for acceleration. The vibration feature parameters extracted in step 3 are received in real time, and the bolt quantization identification results are output. The input interface receives vibration characteristic parameters and system output entropy, while the output interface outputs bolt number, preload value, and loosening position coordinates.

5. The video-based quantitative monitoring method for loosening manhole bolts in hydropower station pressure vessels according to claim 1, characterized in that, Step 3 is implemented in the following steps: Step 3.1, Equipment Deployment; Deploy cameras at appropriate locations in the manholes of pressure vessels to capture complete images of all bolts in the manholes; Step 3.2, Algorithm Settings; The algorithm is set in the camera's built-in system, and an optimized sparse optical flow feature matching algorithm is installed. An adaptive threshold adjustment mechanism, corner detection optimization strategy, multi-feature point linkage tracking and Kalman filtering algorithm are introduced. Step 3.3, Video capture; Start the camera to continuously capture and store real-time vibration videos of the bolts. Step 3.4, video processing and feature parameter extraction, to obtain vibration feature parameters and system output entropy; The vibration characteristic parameters include the first three frequencies f1, f2, f3 and amplitude A; Step 3.5, Data Output; The vibration characteristic data of each extracted bolt and the system output entropy are output to the AI ​​agent model in real time.

6. The video-based quantitative monitoring method for loosening manhole bolts in hydropower station pressure vessels according to claim 5, characterized in that, Step 3.4 is implemented in the following steps: Step 3.4.1: Video frame preprocessing to obtain video grayscale frames; The acquired vibration video stream is processed frame by frame. Color frames are converted to grayscale frames through grayscale processing, and then Gaussian filtering algorithm is used for noise reduction to filter out noise interference from the field environment and obtain grayscale video frames. The specific formula for grayscale conversion is: （6）， In the formula, R ( x,y ), G ( x,y ), B ( x,y ) are respectively in the video frame ( x,y The red, green, and blue pixel values ​​at the coordinates. Gray ( x,y () represents the grayscale value at that coordinate. The specific formula for Gaussian filtering algorithm for noise reduction is as follows: （7）， In the formula, The standard deviation is Gaussian. Step 3.4.2, Selection and tracking of bolt feature points; The video grayscale frames are divided into regions. A square region larger than the bolt diameter is selected as the independent analysis region, centered on the head of each bolt. Overlapping areas between bolts are segmented using the centerline, and non-bolt areas are directly removed. Then, a gradient-based corner detection algorithm is used, based on the pixel gradient product of the bolt corners. L The feature points are selected based on their maximum characteristics, and the coordinates of each feature point in the initial frame are recorded. x 0 ,y 0): （8）， In the formula, L x For pixels in x Pixel gradient in direction, L y For pixels in y Pixel gradient in direction; Subsequently, the optimized sparse optical flow algorithm, combined with the Kalman filter algorithm, is used to track the coordinates of each feature point in each subsequent frame in real time. x t ,y t ),in t The frame number, t =1,2,3...; The specific steps are shown in formulas (9)-(12): First, a neighborhood window is set for each feature point, and the values ​​of each pixel within the neighborhood window are calculated. x direction, y Gray-scale gradient in direction and time direction I x 、I y and I t : （9）， （10）， （11）， In the formula, G ( x,y,t ) is the first t frame( x,y The grayscale value at the coordinate; Based on the assumption of constant pixel grayscale followed by the sparse optical flow method, the optical flow equation is derived as follows: （12）， In the formula, ( u,v ) represents the optical flow vector of the feature point; By solving the optical flow vector of all pixels within the neighborhood window using the least squares method, and obtaining the displacement increment of the corresponding feature point, the feature point can be iteratively updated at the 1st... t Frame coordinates ( x t ,y t )=( x t-1 +u,y t-1 +v ); Simultaneously, the Kalman filter method is used to filter out random fluctuations in feature points. The core formulas include the state equation and the observation equation. The state equation is: （13）， In the formula, X k For the first k The state vector of a frame feature point, including coordinates ( x k ,y k ) and speed ( v k ,y k ) A Here is the state transition matrix: A = （14）， In the formula, Δ t The time interval between two adjacent frames. w k This is process noise; The observation equation is: （15）， In the formula, Z k For the first k Observation coordinates of frame feature points ( x k ,y k ), H The observation matrix; The observation matrix is: H = （16）， In the formula, Vk represents the observation noise; the precise coordinates of each feature point in each frame are obtained through the prediction-update iteration of Kalman filtering. Step 3.4.3: Construct the displacement-time curve for each feature point of each bolt. d(t) Calculate the corresponding system output entropy; For each feature point of each bolt, calculate the displacement change d of that feature point based on its precise coordinates in consecutive frames. t The calculation formula is: （17）， In time t The horizontal axis represents the displacement. dt Plot the displacement-time curve for each feature point with the vertical axis as the ordinate. d(t) This curve represents the time-domain signal of bolt vibration; then, the corresponding system output entropy is calculated based on the displacement information of the bolt group. Step 3.4.4: Frequency domain transformation and frequency extraction to obtain the amplitude |D( ) corresponding to each frequency point. k | with the first three natural frequencies f1, f2, f3; The displacement-time curve of the time-domain displacement signal is obtained by using FFT. d(t) Converting the signal to the frequency domain and extracting the frequency parameters of the bolt vibration, the core formula is the Discrete Fourier Transform: （18）， In the formula, D(k) For the complex amplitude of the frequency domain signal, N The number of sampling points. k For frequency point number, k =0,1,2,...,N-1; After the frequency domain transformation is completed, calculate the amplitude |D( ) corresponding to each frequency point. k Select the frequencies f1, f2, and f3 corresponding to the three frequency points with the largest amplitudes. These are the first three natural frequencies of the bolt vibration, where f1 is the first natural frequency, f2 is the second natural frequency, and f3 is the third natural frequency. Step 3.4.5: Vibration amplitude extraction to obtain the amplitude A of bolt vibration; The vibration amplitude is the maximum displacement of the bolt vibration, which can be obtained through time-domain amplitude and frequency-domain amplitude; the time-domain amplitude is obtained by directly extracting the displacement-time curve. d(t) The maximum peak value in the range; the frequency domain amplitude is the frequency domain amplitude values ​​|D(k1)|, |D(k2)|, and |D(k3)| corresponding to the first three natural frequencies, and the maximum value is taken as the frequency domain amplitude. The amplitude A of bolt vibration is obtained by fusing the time-domain and frequency-domain amplitudes using a weighted average method. Step 3.4.6, Verification and correction of vibration characteristic parameters; The vibration characteristic parameters are verified and outliers are removed. Finite element simulations are performed on the vibration characteristic parameters, and the finite element simulation results are compared with the vibration characteristic parameters to correct the extraction error and ensure that the deviation between the frequency and amplitude extracted on site and the simulation data is ≤5%.

7. The video-based quantitative monitoring method for loosening manhole bolts in hydropower station pressure vessels according to claim 6, characterized in that, In step 3.4.2, the side length of the square region is 1.5 times the bolt diameter.

8. The video-based quantitative monitoring method for loosening manhole bolts in hydropower station pressure vessels according to claim 1, characterized in that, Step 4 is implemented in the following steps: Step 4.1, Data reception; By monitoring the terminal, the bolt quantization results output by the AI ​​agent model, as well as the vibration characteristic parameters and system output entropy, are received synchronously, and a one-to-one correspondence between the quantization results of each bolt and the vibration characteristic parameters is established. Step 4.2, Data Validation; Adopt 3 The criteria remove outliers from the data received in step 4.1 by calculating the average first-order vibration frequency of each bolt. and standard deviation This will exceed [ -3 , +3 Data within the specified range was identified as outliers and removed. Step 4.3, standardize the output; In accordance with the fastener monitoring standards for hydropower stations, a standardized quantitative monitoring data format for bolt loosening was developed, which includes bolt number, acquisition time, preload value, loosening degree level, loosening location coordinates, vibration characteristic parameters, system output entropy, and data verification results. The data was then compiled into standardized monitoring reports and output to the hydropower station operation and maintenance monitoring center in real time, while also being stored in the database.

9. A storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the video-based quantitative monitoring method for loosening manhole bolts in hydropower station pressure vessels as described in any one of claims 1-8.

10. A computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the computer program is processed and executed, it implements the steps of the video-based quantitative monitoring method for loosening manhole bolts in hydropower station pressure vessels as described in any one of claims 1-8.

Images