A method and system for online monitoring of bolt loosening

Abstract

The present invention discloses a method and system for online monitoring of bolt loosening, which includes: collecting an original data set containing vibration signals, temperature gradient data, and structural stress distribution of bolt connections through a multimodal sensor array, performing spatiotemporal alignment and frequency domain decomposition processing on the original data set to obtain a multidimensional feature matrix of the bolt nodes; based on the multidimensional feature matrix, using a graph neural network model to calculate the probability of bolt loosening, and outputting a real-time loosening probability value for each bolt; generating a risk level map based on time evolution according to the real-time loosening probability value; mapping the risk level map in real time through a three-dimensional visualization interface, and when it is detected that the bolt loosening risk level exceeds a preset threshold, generating a warning message and uploading it to the operation and maintenance platform to complete a closed-loop monitoring response. Utilizing the embodiments of the present invention, accurate assessment, dynamic prediction, and visual warning of loosening risks can be achieved, thereby improving the intelligent level of structural safety monitoring.

Description

Technical Field

[0001] The present invention belongs to the technical field of online monitoring, and in particular to a method and system for online monitoring of bolt loosening. Background Art

[0002] In large steel structures such as power towers and bridges, the reliability of bolt connections directly affects the safety performance of the overall structure. Traditional bolt loosening detection relies primarily on manual inspections or periodic torque checks, which suffer from low efficiency, poor real-time performance, and difficulty in detecting early-stage loosening risks. Existing loosening detection methods based on vibration signals or strain monitoring typically use a single sensor to collect data, are susceptible to interference from environmental noise, and lack the ability to comprehensively analyze the coupled influences of multiple factors such as temperature changes and structural force transmission paths. In addition, traditional methods make it difficult to achieve dynamic prediction and visual early warning of loosening risks, resulting in delayed operation and maintenance responses. Summary of the Invention

[0003] The purpose of the present invention is to provide a method and system for online monitoring of bolt loosening to address the deficiencies in the prior art, enable accurate assessment, dynamic prediction and visual early warning of loosening risks, and enhance the intelligence level of structural safety monitoring.

[0004] An embodiment of the present application provides a method for online monitoring of bolt loosening, the method comprising:

[0005] The original data set containing vibration signals, temperature gradient data, and structural stress distribution of the bolted joint is collected through a multimodal sensor array, and the original data set is subjected to spatiotemporal alignment and frequency domain decomposition to obtain a multidimensional feature matrix of the bolted joint. The multimodal sensor array includes a piezoelectric vibrating piece, an infrared thermal imaging unit, and a fiber Bragg grating sensor.

[0006] Based on the multi-dimensional feature matrix, a graph neural network model is used to calculate the probability of bolt loosening. By integrating the vibration frequency domain energy entropy, temperature-stress coupling coefficient and structural topology characteristics, key nodes are dynamically weighted through the attention mechanism to output the real-time loosening probability value of each bolt.

[0007] Based on the real-time loosening probability value, combined with the tower's historical operating data and environmental load time series prediction results, the loosening risk level is dynamically corrected using a LSTM network enhanced by physical information to generate a time-evolution-based risk level map, wherein the risk level map includes the risk evolution trend of each bolt node within a preset time period in the future;

[0008] The risk level map is mapped in real time through a three-dimensional visualization interface. When it is detected that the risk level of bolt loosening exceeds a preset threshold, a warning message is generated and uploaded to the operation and maintenance platform to complete the closed-loop monitoring response.

[0009] Optionally, the multimodal sensor array is used to collect an original data set including vibration signals, temperature gradient data, and structural stress distribution of the bolt connection, and the original data set is subjected to spatiotemporal alignment and frequency domain decomposition processing to obtain a multi-dimensional feature matrix of the bolt node, wherein the multimodal sensor array includes a piezoelectric vibrating piece, an infrared thermal imaging unit, and a fiber Bragg grating sensor, including:

[0010] According to the bolt vibration signal collected by the piezoelectric vibration plate, wavelet packet transform is used to perform noise reduction processing to obtain the denoised vibration time domain signal;

[0011] Based on the denoised vibration time domain signal, the energy distribution of the 0-5kHz frequency band is calculated by short-time Fourier transform, and the energy proportion and energy entropy of each frequency band are extracted to obtain the vibration frequency domain eigenvector;

[0012] Based on the temperature distribution image collected by the infrared thermal imaging unit, the bicubic interpolation algorithm is used to improve the spatial resolution, calculate the radial and axial temperature gradients of the bolt connection surface, and obtain the temperature gradient characteristic matrix;

[0013] Based on the strain data measured by the fiber Bragg grating sensor and the temperature gradient characteristic matrix, the stress fluctuation coefficient caused by temperature change is calculated to obtain the temperature-stress coupling characteristic vector;

[0014] According to the timestamps of the vibration frequency domain eigenvector, temperature gradient eigenvector and temperature-stress coupling eigenvector, the dynamic time warping algorithm is used to perform time series alignment, and finally a multi-dimensional characteristic matrix of the bolt node is generated.

[0015] Optionally, based on the multi-dimensional feature matrix, a graph neural network model is used to calculate the bolt loosening probability. By fusing the vibration frequency domain energy entropy, the temperature-stress coupling coefficient, and the structural topology characteristics, key nodes are dynamically weighted through an attention mechanism to output a real-time loosening probability value for each bolt, including:

[0016] According to the structural mechanics model of the tower, a bolt connection topology diagram is constructed, where the nodes of the topology diagram represent the bolt positions and the edges represent the structural force transmission paths;

[0017] According to the multi-dimensional feature matrix, the vibration frequency domain energy entropy, temperature gradient characteristics and stress coupling coefficient of each bolt node are extracted as the initial node features of the graph neural network;

[0018] According to the node characteristics and topological relationships, a multi-head attention mechanism is used to calculate the feature propagation weights between nodes, focusing on nodes with abnormal vibration energy and sudden temperature gradient changes;

[0019] According to the node characteristics after weighted aggregation and combined with the theoretical calculation formula of preload force, the real-time loosening probability value of each bolt is calculated through a fully connected neural network.

[0020] Optionally, based on the real-time loosening probability value, combined with the tower's historical operating data and the environmental load time series prediction results, the loosening risk level is dynamically corrected using the LSTM network enhanced by physical information to generate a risk level map based on time evolution, wherein the risk level map includes the risk evolution trend of each bolt node within a preset time period in the future, including:

[0021] Based on the historical cases of bolt loosening recorded in the operation and maintenance database, the environmental parameters and equipment status characteristics at the time of loosening are extracted to build a historical working condition feature library;

[0022] Based on weather forecast data, a temporal convolutional network is used to predict the wind speed, temperature, and humidity trends for the next 24 hours, and the environmental load prediction curve is obtained.

[0023] Based on the real-time loosening probability value, historical working condition feature library and environmental load prediction curve, the physical information-enhanced LSTM network is input for time series prediction;

[0024] According to the output results of the LSTM network, the risk level change curve of each bolt node in the next 24 hours is generated, and a three-dimensional spatiotemporal risk level map is constructed.

[0025] Optionally, the risk level map is mapped in real time through a three-dimensional visualization interface. When it is detected that the risk level of bolt loosening exceeds a preset threshold, a warning message is generated and uploaded to the operation and maintenance platform to complete a closed-loop monitoring response, including:

[0026] Based on the tower BIM model and bolt location information, a digital twin is created on a 3D visualization platform;

[0027] Based on the risk level map data, the risk status of each bolt node is dynamically rendered on the digital twin, using red, yellow, and green to represent different risk levels;

[0028] Automatically detect high-risk bolt nodes based on preset warning thresholds and generate warning messages containing location coordinates, risk levels, and maintenance recommendations;

[0029] According to the warning level, the message is pushed to the corresponding operation and maintenance terminal, and the disposal feedback information is tracked to complete the closed-loop process of monitoring-warning-disposal.

[0030] Optionally, the strain data measured by the fiber Bragg grating sensor is combined with the temperature gradient characteristic matrix to calculate the stress fluctuation coefficient caused by temperature change to obtain the temperature-stress coupling characteristic vector, including:

[0031] Based on the original wavelength offset data collected by the fiber Bragg grating sensor, the strain value of each measuring point is calculated using the Bragg wavelength demodulation algorithm to obtain the initial strain distribution data;

[0032] According to the initial strain distribution data, the spatial Kalman filter algorithm is used to eliminate the measurement noise and obtain the smoothed strain field matrix;

[0033] According to the temperature gradient characteristic matrix provided by the infrared thermal imaging unit, the temperature change rate of the corresponding measuring point is extracted, and the temperature-strain mapping relationship table is established in combination with the strain field matrix;

[0034] According to the temperature-strain mapping relationship table, the temperature compensation coefficient is fitted by the least squares method to calculate the effective stress fluctuation caused by pure mechanical load, and finally the temperature-stress coupling eigenvector is generated.

[0035] Optionally, the method of calculating the real-time loosening probability value of each bolt through a fully connected neural network based on the weighted aggregated node features and a theoretical calculation formula for preload force includes:

[0036] Based on the weighted node feature vector output by the graph neural network, three key indicators are extracted: vibration energy entropy, temperature-stress coupling coefficient, and topological connection strength.

[0037] According to the theoretical formula of preload, the stress fluctuation data in the node characteristics is converted into the percentage of preload loss to obtain the theoretical looseness benchmark value;

[0038] The theoretical looseness benchmark value and weighted node features are input into a three-layer fully connected neural network. The first layer performs feature fusion, the second layer adds physical constraints, and the third layer outputs probability values.

[0039] The neural network output value is mapped to the 0-1 interval through the Sigmoid activation function, and the probability threshold is calibrated in combination with historical fault data to finally generate the real-time loosening probability value of each bolt.

[0040] Another embodiment of the present application provides an online monitoring system for bolt loosening, the system comprising:

[0041] a processing module for collecting a raw data set including vibration signals, temperature gradient data, and structural stress distribution of the bolted joint using a multimodal sensor array, performing spatiotemporal alignment and frequency domain decomposition on the raw data set to obtain a multidimensional feature matrix of the bolted joint, wherein the multimodal sensor array includes a piezoelectric vibrating piece, an infrared thermal imaging unit, and a fiber Bragg grating sensor;

[0042] An output module is used to calculate the probability of bolt loosening based on the multi-dimensional feature matrix using a graph neural network model. By integrating the vibration frequency domain energy entropy, the temperature-stress coupling coefficient, and the structural topology characteristics, and dynamically weighting the key nodes through an attention mechanism, the module outputs a real-time loosening probability value for each bolt.

[0043] A correction module is used to dynamically correct the loosening risk level based on the real-time loosening probability value, combined with the tower's historical operating data and the environmental load time series prediction results, using a physical information-enhanced LSTM network to generate a risk level map based on time evolution, wherein the risk level map includes the risk evolution trend of each bolt node within a preset time period in the future;

[0044] The generation module is used to map the risk level map in real time through a three-dimensional visualization interface. When it is detected that the risk level of bolt loosening exceeds a preset threshold, an early warning message is generated and uploaded to the operation and maintenance platform to complete the closed-loop monitoring response.

[0045] Yet another embodiment of the present application provides a storage medium, wherein the storage medium stores a computer program, wherein the computer program is configured to execute any of the above methods when run.

[0046] Yet another embodiment of the present application provides an electronic device, comprising a memory and a processor, wherein the memory stores a computer program, and the processor is configured to run the computer program to execute any of the above methods.

[0047] Compared with the prior art, the present invention provides an online monitoring method for bolt loosening. The method collects original data sets including vibration signals, temperature gradient data and structural stress distribution of bolt connections through a multimodal sensor array, performs spatiotemporal alignment and frequency domain decomposition processing on the original data sets to obtain a multi-dimensional feature matrix of the bolt nodes; based on the multi-dimensional feature matrix, a graph neural network model is used to calculate the probability of bolt loosening, and a real-time loosening probability value of each bolt is output; according to the real-time loosening probability value, a risk level map based on time evolution is generated; the risk level map is mapped in real time through a three-dimensional visualization interface. When it is detected that the bolt loosening risk level exceeds a preset threshold, an early warning message is generated and uploaded to the operation and maintenance platform to complete the closed-loop monitoring response, thereby enabling accurate assessment, dynamic prediction and visual early warning of loosening risks, thereby improving the intelligence level of structural safety monitoring. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] Figure 1 A hardware structure block diagram of a computer terminal for an online bolt loosening monitoring method provided by an embodiment of the present invention;

[0049] Figure 2 A schematic flow chart of a method for online monitoring of bolt loosening provided by an embodiment of the present invention;

[0050] Figure 3 A schematic structural diagram of an online monitoring system for bolt loosening provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0051] The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and are not to be construed as limiting the present invention.

[0052] The embodiment of the present invention first provides a method for online monitoring of bolt loosening, which can be applied to electronic equipment, such as a computer terminal, specifically a common computer.

[0053] The following describes it in detail by taking running on a computer terminal as an example. Figure 1 The hardware structure block diagram of a computer terminal for an online monitoring method for bolt loosening provided by an embodiment of the present invention. Figure 1 As shown, the computer device includes a processor, a memory, and a network interface connected via a system bus, wherein the memory may include a non-volatile storage medium and an internal memory.

[0054] The non-volatile storage medium can store an operating system and a computer program. The computer program includes program instructions, and when the program instructions are executed, the processor can execute any one of the online bolt loosening monitoring methods.

[0055] The processor is used to provide computing and control capabilities and support the operation of the entire computer equipment.

[0056] The internal memory provides an environment for the operation of the computer program in the non-volatile storage medium. When the computer program is executed by the processor, the processor can execute any one of the online monitoring methods for bolt loosening.

[0057] The network interface is used for network communication, such as sending assigned tasks, etc. Those skilled in the art will understand that Figure 1 The structure shown in the figure is only a block diagram of a part of the structure related to the solution of the present application, and does not constitute a limitation on the computer device to which the solution of the present application is applied. The specific computer device may include more or fewer components than shown in the figure, or combine certain components, or have a different component arrangement.

[0058] It should be understood that the processor may be a central processing unit (CPU), other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field-programmable gate arrays (FPGA), other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor may be a microprocessor or any conventional processor, etc.

[0059] See also Figure 2 , an embodiment of the present invention provides a method for online monitoring of bolt loosening, which may include the following steps:

[0060] S201, collecting an original data set including vibration signals, temperature gradient data, and structural stress distribution of a bolted connection using a multimodal sensor array, performing spatiotemporal alignment and frequency domain decomposition on the original data set to obtain a multidimensional feature matrix of the bolted connection, wherein the multimodal sensor array includes a piezoelectric vibrating piece, an infrared thermal imaging unit, and a fiber Bragg grating sensor;

[0061] Specifically, the vibration signal of the bolt collected by the piezoelectric vibration plate can be subjected to noise reduction processing by using wavelet packet transform to obtain a denoised vibration time domain signal;

[0062] The piezoelectric vibrator samples the vibration signal of the bolted joint at a sampling rate of 100,000 times per second (100 kHz). The original signal contains high-frequency noise (such as wind noise and electromagnetic interference from equipment). Noise reduction is performed using the Wavelet Packet Transform (WPT). Its core is to extract the effective components of the signal through multi-scale decomposition. The Symlet wavelet basis (Sym8) was chosen because its symmetry reduces phase distortion and is well-suited to the non-stationary nature of mechanical vibration signals.

[0063] Signal decomposition: The original signal is decomposed into 8 subbands (2^3=8) using a three-layer wavelet packet decomposition method. The frequency range is 0-50 kHz. For example, the first layer is decomposed into 0-25 kHz and 25-50 kHz, and the second layer is further subdivided into 0-12.5 kHz, 12.5-25 kHz, and so on. Finally, each subband in the third layer is approximately 6.25 kHz wide.

[0064] Threshold denoising: This algorithm uses an improved Sureshrink threshold algorithm to dynamically adjust the threshold based on the noise level of each sub-band. For example, for high-frequency sub-bands (such as 25-50 kHz), where noise energy accounts for a high proportion, the threshold is set at 1.5 times the standard deviation of the signal energy in that band. For low-frequency sub-bands (such as 0-6.25 kHz), the threshold is set at 0.8 times the standard deviation to preserve more valid signal.

[0065] Signal reconstruction: Only sub-bands with significant energy (above a threshold) are retained for inverse transformation. For example, if the characteristic frequency caused by a loose bolt is concentrated in the 8-12 kHz range, the corresponding sub-band (a local frequency band within the 12.5-25 kHz sub-band of the second layer) is retained, and the remaining high-frequency noise sub-bands are suppressed.

[0066] The signal-to-noise ratio (SNR) of the signal after noise reduction is improved from 15 dB to 35 dB, and the periodic impact components in the time domain waveform (such as micro-slip vibration caused by loose bolts) are clearly visible.

[0067] Based on the denoised vibration time domain signal, the energy distribution of the 0-5kHz frequency band is calculated by short-time Fourier transform, and the energy proportion and energy entropy of each frequency band are extracted to obtain the vibration frequency domain eigenvector;

[0068] The denoised vibration signal is subjected to time-frequency analysis using Short-Time Fourier Transform (STFT), focusing on the 0-5 kHz frequency band (the characteristic frequency range of bolt loosening).

[0069] Parameter configuration:

[0070] Window type: A Hanning window is used with a window length of 256 samples (corresponding to a time window of 2.56 milliseconds) and a 75% overlap ratio (64 samples per slide). This window design balances time resolution and frequency resolution, reducing spectral leakage.

[0071] Frequency range: The signal sampling rate is 100 kHz, and the STFT analysis range covers 0-50 kHz. However, only the 0-5 kHz frequency band data is extracted, which contains the micro-slip vibration characteristics caused by bolt loosening.

[0072] Sub-band division and energy calculation: 0-5 kHz is divided into 5 equal-width sub-bands (0-1 kHz, 1-2 kHz, 2-3kHz, 3-4 kHz, 4-5 kHz), each with a bandwidth of 1 kHz.

[0073] Calculate the energy contribution of each sub-band: For example, if the energy of the 3-4 kHz sub-band accounts for 40% of the total energy of 0-5 kHz within a certain time window, it indicates that there is a significant vibration anomaly in this frequency band.

[0074] Energy entropy calculation: Based on the Shannon entropy principle, this method quantifies the degree of dispersion of energy distribution within each subband. Higher energy entropy values ​​indicate a more dispersed energy distribution, potentially indicating the superposition of multiple frequency bands due to loosening. For example, the entropy value of a properly functioning bolt is typically less than 1.8, but may exceed 2.5 when loose.

[0075] Feature vector generation: Each time window outputs a 6-dimensional vector containing the energy proportions of 5 subbands (such as [0.1, 0.2, 0.3, 0.25, 0.15]) and an energy entropy value (such as 2.7).

[0076] The vector is spliced ​​through time series to form a continuous frequency domain feature stream for subsequent graph neural network analysis of the bolt status.

[0077] Based on the temperature distribution image collected by the infrared thermal imaging unit, the bicubic interpolation algorithm is used to improve the spatial resolution, calculate the radial and axial temperature gradients of the bolt connection surface, and obtain the temperature gradient characteristic matrix;

[0078] The infrared thermal imaging unit captures temperature distribution images of the bolted joint area at a rate of one frame per second, with an original resolution of 640 × 480 pixels. Bicubic interpolation is used to increase the resolution to 1280 × 960 pixels to accurately capture temperature gradient details.

[0079] Interpolation processing:

[0080] Interpolation principle: A bicubic polynomial is fitted to the 16×16 neighborhood (original image) of each low-resolution pixel to calculate the temperature value of the high-resolution pixel. For example, if the temperature distribution in a certain area of ​​the original image is 25°C to 30°C, the interpolation method can detect a small difference of 26.5°C.

[0081] Improved results: The edge clarity of the contact surface between the bolt head and the flange in the interpolated image is improved, the transition of temperature jump areas (such as friction hot spots) is smoother, and the gradient calculation error is reduced.

[0082] Temperature gradient calculation:

[0083] Radial gradient: Calculate the temperature change rate (°C / mm) every 0.5 mm along the radial direction, centered on the bolt axis. For example, the temperature gradient at the edge of a bolt due to friction can reach 1.2°C / mm, significantly higher than the 0.3°C / mm at the center.

[0084] Axial gradient: The temperature change is measured at 1mm intervals along the length of the bolt. For example, the axial gradient near the flange surface is 0.8°C / mm, while the gradient drops to 0.2°C / mm away from the flange, reflecting the attenuation of heat conduction along the bolt.

[0085] Characteristic matrix construction: The bolted connection surface is divided into a 10×10 grid (a total of 100 grids), each of which stores radial and axial gradient values. For example, a grid with gradient values ​​of [1.2, 0.8] indicates a radial gradient of 1.2°C / mm and an axial gradient of 0.8°C / mm.

[0086] Finally, a 20×10 temperature gradient feature matrix (each grid contains 2 gradient values) is generated to comprehensively characterize the thermal distribution anomaly of the bolted connection surface.

[0087] Based on the strain data measured by the fiber Bragg grating sensor and the temperature gradient characteristic matrix, the stress fluctuation coefficient caused by temperature change is calculated to obtain the temperature-stress coupling characteristic vector;

[0088] Specifically, based on the strain data measured by the fiber Bragg grating sensor and combined with the temperature gradient characteristic matrix, the stress fluctuation coefficient caused by the temperature change is calculated to obtain the temperature-stress coupling characteristic vector, which may include:

[0089] Based on the original wavelength offset data collected by the fiber Bragg grating sensor, the strain value of each measuring point is calculated using the Bragg wavelength demodulation algorithm to obtain the initial strain distribution data;

[0090] The core principle of fiber Bragg grating (FBG) sensors is to reflect strain changes in the measured object through Bragg wavelength shift. Each FBG sensor is factory-set with a reference Bragg wavelength (e.g., 1550 nm). When a bolted connection deforms, the periodic variation in the FBG causes a shift in the reflected wavelength. Assuming the sensor is calibrated so that a wavelength shift of 1 picometer (pm) corresponds to 1 microstrain (με), the measured wavelength shift can be directly converted to strain.

[0091] The specific process of the Bragg wavelength demodulation algorithm is as follows:

[0092] Spectral Data Acquisition: A high-precision spectrometer (resolution 0.1 pm) is used to scan the fiber Bragg grating's reflection spectrum and capture the peak wavelength. For example, in a measurement, the reference wavelength is 1549.8 nm, and the current peak wavelength is 1550.3 nm. The total offset is 0.5 nm (500 pm).

[0093] Strain calculation: According to the calibration factor (1 pm = 1 με), a 500 pm offset corresponds to a strain of 500 με.

[0094] Initial temperature compensation: Because fiber Bragg gratings are sensitive to temperature (temperature coefficient approximately 10 pm / °C), changes in ambient temperature must be recorded. For example, if the temperature rises by 2°C, 20 pm (2°C x 10 pm / °C) must be deducted from the total offset, leaving 480 pm, corresponding to an effective strain of 480 με.

[0095] By processing all measurement points one by one (for example, 50 fiber Bragg grating sensors are arranged on a bolted joint), initial strain distribution data is generated. For example, the strain at measurement point A is 300 με, and at measurement point B is -150 με (negative values ​​indicate compressive strain).

[0096] According to the initial strain distribution data, the spatial Kalman filter algorithm is used to eliminate the measurement noise and obtain the smoothed strain field matrix;

[0097] The spatial Kalman filter algorithm suppresses random noise by integrating the spatial correlation of adjacent measurement points. Assuming the measurement points are distributed in a 5×10 grid, the state vector of each measurement point contains the strain value and its first-order derivative (strain change rate). The filtering process is divided into two stages: prediction and update:

[0098] Prediction stage: Predict the current strain based on the strain value and change rate at the previous moment. For example, if the strain at measuring point A was 300με at the previous moment and the change rate was 10με / s, the current strain is predicted to be 310με.

[0099] Considering the influence of spatial neighborhood, if the strain of the adjacent measuring point B is -150με, the predicted value is adjusted by the weight coefficient (such as 0.2), and the final predicted strain = 310με×0.8 + (-150με)×0.2 = 218με.

[0100] Update phase: Take the weighted average of the predicted value and the actual measured value (for example, the actual measured value is 320με). If the Kalman gain is set to 0.7, the updated strain = 218με × 0.3 + 320με × 0.7 = 289.4με.

[0101] Update the covariance matrix to reduce the impact of noise.

[0102] After 10 iterations of filtering, random noise in the raw data (such as ±50με fluctuations) is suppressed to within ±5με, generating a smooth strain field matrix. For example, the strain at measuring point A converges from 320±50με to 295±5με, and at measuring point B from -160±50με to -155±5με.

[0103] According to the temperature gradient characteristic matrix provided by the infrared thermal imaging unit, the temperature change rate of the corresponding measuring point is extracted, and the temperature-strain mapping relationship table is established in combination with the strain field matrix;

[0104] The infrared thermal imaging unit captures temperature distribution images of the bolted joint surface at a rate of 5 frames per second (resolution 0.1°C). Image registration technology aligns the temperature data with the fiber Bragg grating (FBG) measurement points. For example, at measurement point A, corresponding to image coordinates (x=120, y=80), the temperature at that location increases from 30°C to 32°C (rate of change 0.4°C / minute).

[0105] The process of building the temperature-strain mapping relationship table:

[0106] Time alignment: Align temperature data and strain data with millisecond-level timestamps. For example, at a certain time t, the temperature of measurement point A is 31.5°C and the strain is 295με.

[0107] Rate of change calculation: Calculates the temperature change rate within a 10-second window. For example, if the temperature changes from 31°C to 31.5°C in 30 seconds, the rate of change is 1°C / minute.

[0108] Mapping table filling: Store data in five columns: measurement point number, timestamp, temperature value, temperature change rate, and strain value, as shown in Table 1:

[0109] Table 1

[0110]

[0111] By analyzing the mapping table, we can find the correlation between temperature and strain. For example, for every 1°C increase in temperature, the strain at measuring point A increases by 20με (thermal expansion effect), while the strain at measuring point B, due to structural constraints, increases by -10με (compression).

[0112] According to the temperature-strain mapping relationship table, the temperature compensation coefficient is fitted by the least squares method to calculate the effective stress fluctuation caused by pure mechanical load, and finally the temperature-stress coupling eigenvector is generated.

[0113] The least square method is used to separate the temperature effect from the mechanical load effect. Assuming that the strain (ε) is given by the temperature strain (ε t ) and mechanical strain (ε m ) Linear superposition: ε = ε t +ε m =αΔT+ε m , where α is the temperature compensation coefficient and ΔT is the temperature change.

[0114] Fitting process:

[0115] Data sampling: 100 sets of temperature-strain data were selected (ΔT from -5°C to +10°C, with a step size of 0.5°C).

[0116] Linear regression: Use the least squares method to fit the relationship between ε and ΔT and obtain the slope α. For example, the fitting result for measurement point A is α = 20 με / °C, with R² = 0.95 (high correlation).

[0117] Temperature compensation: subtract the temperature effect from the total strain. For example, if the temperature of measuring point A rises by 2°C at a certain moment, the temperature strain ε is calculated. t =20με / ℃×2℃=40με. If the measured total strain ε=300με, the calculated mechanical strain is ε m =300-40=260με.

[0118] Calculation of effective stress fluctuation: According to Hooke's law, stress σ=Ε×ε m , where E is the elastic modulus of the material (200 GPa for steel). For example, ε m =260με=260×10 -6 , then σ=200×10 9 Pa ×260×10 -6 = 52 MPa.

[0119] The resulting temperature-stress coupling eigenvector contains the following dimensions: temperature compensation coefficient α (unit: με / °C); mechanical strain ε m (unit: με); effective stress σ (unit: MPa); temperature change rate (unit: ℃ / min).

[0120] For example, the eigenvector of a measurement point is [20, 260, 52, 1.0], which means that the stress caused by the current mechanical load is 52 MPa and the temperature is rising at a rate of 1°C / minute.

[0121] According to the timestamps of the vibration frequency domain eigenvector, temperature gradient eigenvector and temperature-stress coupling eigenvector, the dynamic time warping algorithm is used to perform time series alignment, and finally a multi-dimensional characteristic matrix of the bolt node is generated.

[0122] Multimodal data has asynchronous timestamps due to different sensor sampling rates (vibration signal 100 kHz, temperature 1 Hz, strain 1 kHz), and requires dynamic time warping (DTW) to align the timing.

[0123] Data segmentation: Using 1-second intervals for temperature data as the baseline segment, the vibration and strain data are downsampled to 1 Hz (taking the mean or peak value within each segment). For example, the vibration frequency domain feature vector is downsampled from 1000 Hz to 1 Hz, and the mean entropy value of 1000 samples is taken for each segment.

[0124] DTW alignment: aligns three 1 Hz sequences, vibration, temperature, and strain, pairwise, and calculates the optimal path. For example, if the starting point of a temperature segment is delayed by 0.2 seconds, synchronization can be achieved by stretching the strain sequence for 0.2 seconds.

[0125] Feature fusion: The aligned features are concatenated by time point: vibration (6 dimensions), temperature (20×10=200 dimensions), and stress coupling (5 dimensions), for a total of 211 dimensions.

[0126] The resulting multidimensional feature matrix is ​​N × 211, where N is the number of time points. For example, one hour of data (3600 seconds) forms a 3600 × 211 matrix for subsequent loose probability calculations.

[0127] By deploying a multimodal sensing network consisting of piezoelectric vibrating plates, infrared thermal imaging units, and fiber Bragg grating sensors, heterogeneous data such as the vibration spectrum, temperature field distribution, and structural stress of bolted joints are synchronously collected. Wavelet packet transform and short-time Fourier transform are used to perform joint time-frequency domain analysis of the raw signals. A dynamic time warping algorithm is used to achieve spatiotemporal synchronization of multi-source data, constructing a multidimensional feature matrix containing vibration energy entropy, temperature gradient coefficients, and stress coupling characteristics. This overcomes the limitations of traditional single-modal detection and comprehensively reflects the state of bolted joints through collaborative sensing of multiple physical quantities. Spatiotemporal alignment processing eliminates acquisition delays and spatial deviations between sensors, and frequency domain decomposition extracts characteristic frequency bands strongly correlated with loosening, providing high-quality input data for subsequent intelligent diagnosis.

[0128] S202, based on the multi-dimensional feature matrix, a graph neural network model is used to calculate the probability of bolt loosening. By integrating the vibration frequency domain energy entropy, the temperature-stress coupling coefficient, and the structural topology characteristics, key nodes are dynamically weighted through an attention mechanism to output a real-time loosening probability value for each bolt;

[0129] Specifically, a bolt connection topology diagram can be constructed based on the tower structure mechanics model, where the nodes of the topology diagram represent the bolt positions and the edges represent the structural force transmission paths.

[0130] The bolt connection topology diagram is the core model for describing the mechanical properties of tower structures. First, based on the tower's 3D CAD design drawings or BIM model, the spatial coordinates and connection relationships of all bolt nodes are extracted. For example, in a high-voltage transmission tower, the main bolts are typically located at key locations such as tower legs and crossarms. Each bolt is represented as an independent node in the topology diagram, with node attributes including its installation location (e.g., tower leg node number A1), preload torque (e.g., 300 N·m), and material properties (e.g., 8.8-grade high-strength steel).

[0131] The force transmission path is constructed using the finite element analysis method:

[0132] Static analysis: Apply standard loads (such as wind load and wire tension) in software such as ANSYS to calculate the stress distribution of each bolt node;

[0133] Path tracing: Identify the main force transmission paths through stress cloud maps, such as tower leg bolts bearing vertical pressure and crossarm bolts bearing bending moment;

[0134] Edge weight assignment: Set edge weights based on the stress transfer ratio between nodes. For example, an edge weight of 0.8 from node A to node B means 80% of the load is transferred through this path.

[0135] For complex structures, topological relationships are calibrated using experimental data. Strain gauges are placed on prototype towers to measure load distribution under various operating conditions, which then adjust the edge weights of the finite element model. For example, if a crossarm bolt actually bears an additional 12% load under wind vibration, the corresponding edge weight is adjusted from 0.75 to 0.87. The resulting topological graph is stored in an adjacency matrix format, with the matrix elements representing the weights of force transmission between nodes.

[0136] According to the multi-dimensional feature matrix, the vibration frequency domain energy entropy, temperature gradient characteristics and stress coupling coefficient of each bolt node are extracted as the initial node features of the graph neural network;

[0137] The multi-dimensional feature matrix contains three types of core features, which need to be extracted and fused separately:

[0138] Energy entropy in the vibration frequency domain:

[0139] Data source: time domain signal collected by piezoelectric vibrating plate (sampling rate 10kHz);

[0140] Processing flow: Decompose the wavelet packet into 8 layers to obtain 32 frequency bands (0-5kHz); calculate the energy contribution of each frequency band (for example, the energy of frequency band 3 accounts for 15%); calculate the energy entropy according to the information entropy formula: the more evenly the energy distribution in each frequency band, the higher the entropy value (about 1.2-1.8 in normal state, and the entropy value drops to 0.6-1.0 when loose because the energy is concentrated in the low frequency band).

[0141] Temperature gradient characteristics:

[0142] Data source: Bolt surface temperature field collected by infrared thermal imager (resolution 0.1°C);

[0143] Processing flow: Bicubic interpolation improves the temperature field resolution to 0.05°C. The temperature gradient (ΔT / Δx, unit: °C / mm) is calculated along the bolt axis, which is approximately 0.2-0.5°C / mm under normal conditions. The radial gradient reflects the contact state between the bolt and the connecting plate, and is abnormal when loose (e.g., reaching 1.2°C / mm in some areas).

[0144] Stress coupling coefficient:

[0145] Data source: strain data measured by fiber Bragg grating sensor (accuracy 1με);

[0146] Processing flow: Kalman filtering eliminates temperature drift (such as compensating strain of 50με at 25°C); calculates mechanical stress σ=E×ε (E is the elastic modulus, 210GPa for steel); combines temperature gradient data to calculate the temperature-stress coupling coefficient α=Δσ / ΔT (normal value is approximately 0.05 MPa / °C; when loose, due to changes in contact thermal resistance, α increases to 0.1-0.15 MPa / °C).

[0147] Feature fusion: The feature vector of each node is three-dimensional (energy entropy, axial gradient, α coefficient). For example, the feature vector of node A is [0.73, 0.48°C / mm, 0.12 MPa / °C], indicating that it is in a potential loose state.

[0148] According to the node characteristics and topological relationships, a multi-head attention mechanism is used to calculate the feature propagation weights between nodes, focusing on nodes with abnormal vibration energy and sudden temperature gradient changes;

[0149] The Multi-Head Attention (MHA) mechanism is used to capture the dynamic relationship between nodes. Take 4-head attention as an example:

[0150] Query-key-value generation: The feature vector of each node is linearly transformed to generate the query, key, and value matrices with dimensions of 64×4. For example, the query vector of node A is [0.32, -0.15, ..., 0.07] (64 dimensions).

[0151] Attention score calculation: Calculate the attention score of node A to node B: If the vibration energy entropy of node B is abnormal (such as 0.5) and the temperature gradient changes suddenly (1.1°C / mm), its Key vector has a high value in the corresponding dimension, resulting in a significant increase in the Score.

[0152] Weight distribution and aggregation:

[0153] Generate attention weights by normalizing the scores with Softmax (e.g., the weight of node A to B is 0.8, and to C is 0.2);

[0154] Weighted aggregate Value vector: New feature of node A = 0.8×Value_B + 0.2×Value_C.

[0155] Abnormal node strengthening:

[0156] Vibration energy entropy threshold: Entropy values ​​< 1.0 are set as abnormal, and the query vector of the corresponding node is given a higher learning rate during training;

[0157] Temperature gradient mask: When the gradient is greater than 0.8°C / mm, the key vector weight is forced to increase by 50%.

[0158] For example, the vibration entropy of a crossarm bolt (node ​​D) is 0.6 and the temperature gradient is 1.0°C / mm. Its average weight in the four attention heads is 0.7, which significantly affects the feature propagation of adjacent nodes.

[0159] According to the node characteristics after weighted aggregation and combined with the theoretical calculation formula of preload force, the real-time loosening probability value of each bolt is calculated through a fully connected neural network.

[0160] Specifically, based on the weighted aggregated node features and combined with the theoretical calculation formula for preload force, the real-time loosening probability value of each bolt is calculated through a fully connected neural network, which may include:

[0161] Based on the weighted node feature vector output by the graph neural network, three key indicators are extracted: vibration energy entropy, temperature-stress coupling coefficient, and topological connection strength.

[0162] From the weighted node feature vector output by the graph neural network, three core indicators need to be extracted to quantify the loosening risk of the bolts.

[0163] Vibrational energy entropy:

[0164] Data source: The time-domain vibration signal collected by the piezoelectric vibrating piece is decomposed into eight sub-bands (such as 0-625Hz, 625-1250Hz, etc.) using wavelet packet transform, and the energy contribution of each frequency band is calculated.

[0165] Entropy calculation: Assuming the energy distribution of a bolt in the sub-band is [0.3, 0.2, 0.1, 0.05, 0.05, 0.1, 0.1, 0.1], the energy entropy calculated using the Shannon entropy formula (without expansion) is 1.85. A higher value indicates a more dispersed energy distribution, which may reflect the complexity of the vibration mode caused by loosening.

[0166] Anomaly detection: Historical data shows that the normal bolt energy entropy range is 1.2~1.6. If the current value exceeds 1.8, it is marked as abnormal.

[0167] Temperature-stress coupling coefficient:

[0168] Data Fusion: After aligning the strain data measured by the fiber Bragg grating sensor with the temperature gradient matrix from infrared thermal imaging, the temperature compensation coefficient is fitted using the least squares method. For example, when the temperature of a bolt rises by 10°C, the strain increases by 0.02%. After deducting the temperature effect, the pure mechanical strain is 0.15%.

[0169] Calculation of coupling coefficient: Define the coefficient = mechanical strain / temperature strain. If the coefficient is > 5 (such as 0.15% / 0.02% = 7.5), it indicates that the mechanical load is dominant and the bolt stress state is stable. If the coefficient is < 2, the temperature influence is significant and requires close monitoring.

[0170] Topological connection strength:

[0171] Structural analysis: Extract the connectivity of bolt nodes based on the tower BIM model (for example, if a bolt connects three beams, the connectivity is 3).

[0172] Dynamic weight: Combined with the graph neural network attention weight (e.g., 0.92), if the attention weight of a high-connectivity bolt (connectivity ≥ 4) is lower than 0.5, its topological importance is judged to have decreased, and loosening may cause a change in the force transmission path.

[0173] For example, the vibration energy entropy of a bolt is 1.9 (abnormal), the temperature-stress coupling coefficient is 6.2 (normal), and the topological connection strength weight is 0.45 (low), which indicates a high risk of loosening.

[0174] According to the theoretical formula of preload force, the stress fluctuation data in the node characteristics is converted into the percentage of preload force loss to obtain the theoretical looseness benchmark value;

[0175] Preload loss is a direct indicator of bolt loosening, and the real-time monitored stress fluctuations need to be converted into theoretical looseness.

[0176] Stress-Preload Mapping:

[0177] Material parameters: Assume that the bolt material is 8.8 grade steel, the elastic modulus is 210GPa, and the initial preload design value is 50kN.

[0178] Application of theoretical formula: According to Hooke's law, stress σ = strain ε × elastic modulus E. For example, if the measured strain decreases by 0.05% (i.e., ε = 0.0005), the stress loss Δσ = 0.0005 × 210 GPa = 105 MPa.

[0179] Calculation of preload loss: Bolt cross-sectional area A = 200mm², preload loss ΔF = Δσ × A = 105MPa × 200mm² = 21kN, loss percentage = 21kN / 50kN = 42%.

[0180] Temperature compensation correction:

[0181] Influence of thermal expansion: When the temperature of the bolt rises by 10°C, the thermal expansion strain ε_thermal = α × ΔT (α = 12e-6 / °C), ε_thermal = 12e-6 × 10 = 0.012%.

[0182] Effective mechanical strain: From the total strain of 0.05%, after deducting 0.012% of thermal strain, the actual mechanical strain is 0.038%. The corresponding preload loss correction is ΔF=0.00038×210GPa×200mm²=15.96kN, and the loss percentage is 31.9%.

[0183] Looseness benchmark value generation:

[0184] Classification standard: Preset threshold: loss <20% is normal (green), 20%~40% is warning (yellow), >40% is high risk (red).

[0185] Dynamic adjustment: Based on historical data, if a certain type of bolt frequently fails when the loss is 35%, the node threshold will be lowered to 30%.

[0186] Example: The theoretical looseness benchmark value of a bolt is 35%, which falls into the yellow warning range and triggers subsequent neural network analysis.

[0187] The theoretical looseness benchmark value and weighted node features are input into a three-layer fully connected neural network. The first layer performs feature fusion, the second layer adds physical constraints, and the third layer outputs probability values.

[0188] The design of a fully connected neural network (FCNN) requires integrating data-driven and physical laws:

[0189] First layer (feature fusion layer):

[0190] Input dimensions: 5 features (vibration energy entropy, temperature-stress coefficient, topological connection strength, theoretical looseness, and historical failure frequency).

[0191] Processing: 128 neurons are used, with the activation function ReLU. For example, if a node input is [1.9, 6.2, 0.45, 35%, 2 times / year], it will output a high-dimensional feature vector after full connection.

[0192] Feature enhancement: Perform nonlinear transformation (such as square term) on vibration energy entropy and theoretical looseness to capture the coupling effect between the two.

[0193] The second layer (physical constraint layer):

[0194] Constraint injection:

[0195] Preload is irreversible: if the theoretical looseness is greater than 100%, the forced output probability is 1.0;

[0196] Energy entropy lower limit: If the energy entropy is less than 0.5 (sensor failure), the probability is set to 0.

[0197] Network structure: 64 neurons, ReLU activation function, weight initialization is constrained (for example, the theoretical looseness weight is not less than 0.3).

[0198] The third layer (probability output layer):

[0199] Binary classification processing: single neuron output, Sigmoid activation function maps the value to 0~1.

[0200] Training data: Use 5000 sets of historical samples (3000 normal and 2000 loose), cross entropy loss function, Adam optimizer (learning rate 0.001).

[0201] Example reasoning: After the input features are processed by the first three layers, the output probability is 0.78, which exceeds the threshold of 0.65 and is judged as high risk.

[0202] Model verification: Confusion matrix test: accuracy 92%, recall 88%; real-time performance: single-node inference time <3ms, meeting online monitoring requirements.

[0203] The neural network output value is mapped to the 0-1 interval through the Sigmoid activation function, and the probability threshold is calibrated in combination with historical fault data to finally generate the real-time loosening probability value of each bolt.

[0204] Probability threshold calibration is key to balancing false positives and false negatives:

[0205] Sigmoid mapping:

[0206] Function: Compress the original output of the neural network (such as 2.3) to 0~1, for example, 2.3→0.91, -1.2→0.23.

[0207] Probability explanation: 0.8 indicates an 80% chance of loosening, which needs to be determined in conjunction with the threshold.

[0208] Threshold dynamic calibration:

[0209] ROC curve analysis: Draw a curve using historical data and select the optimal threshold (such as 0.65) to achieve a true positive rate (TPR) of 85% and a false positive rate (FPR) of <10%.

[0210] Adaptive working conditions: During extreme weather (such as typhoons), the threshold is temporarily lowered to 0.6 to improve warning sensitivity.

[0211] Post-processing of probability values:

[0212] Time smoothing: Take a moving average of the probability values ​​over 10 seconds (e.g., [0.7, 0.72, 0.75, 0.8]→0.74) to avoid false alarms due to instantaneous fluctuations.

[0213] Spatial correlation: If the probability of adjacent bolts is greater than 0.6, a regional warning is triggered (such as the risk of the entire flange loosening).

[0214] Example output: Bolt A: Probability 0.73 → exceeds the threshold of 0.65 → generates a high-risk warning; Bolt B: Probability 0.58 → falls below the threshold → marked as under monitoring; Historical calibration: After adjusting the threshold, the false negative rate dropped from 12% to 6%.

[0215] A topological relationship model of bolted connection systems is constructed based on a graph neural network. Features such as vibration energy entropy and temperature-stress coupling coefficient are used as node attributes. An attention mechanism is used to dynamically capture the propagation paths of abnormal features at key nodes. Integrating prior knowledge of structural mechanics, a weighted aggregation approach is used to calculate the loosening probability of each node, enabling collaborative diagnosis of the entire connection system rather than isolated judgments. Using a graph structure to model the mechanical dependencies between bolts, the attention mechanism automatically focuses on high-risk nodes, preventing traditional methods from missing local anomalies. Topological feature propagation enhances the mutual verification of the states of adjacent bolts, significantly improving the accuracy and robustness of loosening detection.

[0216] S203, based on the real-time loosening probability value, combined with the tower's historical operating data and the environmental load time series prediction results, dynamically correct the loosening risk level using the LSTM network enhanced by physical information, and generate a risk level map based on time evolution, wherein the risk level map includes the risk evolution trend of each bolt node within a preset future time period;

[0217] Specifically, based on the historical cases of bolt loosening recorded in the operation and maintenance database, the environmental parameters and equipment status characteristics at the time of loosening can be extracted to build a historical working condition feature library;

[0218] The construction of the historical operating condition feature library requires the integration of multi-source heterogeneous data, including bolt loosening event records, environmental sensor data, and equipment operation logs. First, the original records of all loosening cases in the past five years are extracted from the operation and maintenance database. Each case contains:

[0219] Environmental parameters: wind speed (0~30m / s), temperature (-20℃~50℃), humidity (10%~100%RH);

[0220] Equipment status: vibration acceleration (0~10g), preload loss rate (0%~100%), bolt torque value (100~500Nm);

[0221] Timestamp: Time series data before and after the loosening occurs (sampling frequency 1 Hz).

[0222] Data preprocessing process:

[0223] Missing value filling: KNN (K nearest neighbor, K=5) algorithm is used to fill missing values ​​based on the mean of samples with similar working conditions;

[0224] Outlier elimination: Based on the 3σ principle, data points exceeding the mean ± 3 times the standard deviation are eliminated;

[0225] Feature standardization: Z-score normalization is performed on parameters of different dimensions such as wind speed and temperature.

[0226] Feature Engineering:

[0227] Time domain characteristics: calculate the root mean square (RMS) and peak-to-peak (P2P) value of the vibration signal;

[0228] Frequency domain features: Extract the energy center of gravity (FC) of the 0-1kHz frequency band through FFT (Fast Fourier Transform);

[0229] Coupling characteristics: Pearson correlation coefficient between temperature gradient and vibration energy was calculated (threshold > 0.6 for strong correlation).

[0230] The final historical operating condition feature library contains more than 10,000 samples, each of which is represented by a 15-dimensional feature vector, for example: [wind speed 6.5m / s, temperature 25°C, humidity 60%, RMS=2.3g, P2P=8.7g, FC=320Hz, correlation coefficient 0.72].

[0231] Based on weather forecast data, a temporal convolutional network is used to predict the wind speed, temperature, and humidity trends for the next 24 hours, and the environmental load prediction curve is obtained.

[0232] The Temporal Convolutional Network (TCN) uses a causal dilation convolutional structure. Its core advantage lies in its ability to capture long-term temporal dependencies. The network configuration is as follows:

[0233] Input layer: meteorological data (wind speed, temperature, humidity) for the past 72 hours, with a time resolution of 1 hour;

[0234] Convolutional layer: 4 layers of dilated convolution, with dilation coefficients of 1, 2, 4, and 8, 64 filters, and a kernel size of 3;

[0235] Output layer: predicted values ​​of three variables (wind speed, temperature, and humidity) for the next 24 hours.

[0236] Training details:

[0237] Loss function: weighted sum of MAE (mean absolute error) and RMSE (root mean square error) (weight 0.7:0.3);

[0238] Optimizer: AdamW (learning rate 0.001, weight decay 0.01);

[0239] Regularization: Dropout rate 0.2 to prevent overfitting.

[0240] Example of prediction process:

[0241] Input data from the past 72 hours: wind speed sequence [5.2, 5.8, ..., 6.1] m / s, temperature sequence [22, 23, ..., 25] °C, humidity sequence [55%, 58%, ..., 60%];

[0242] TCN outputs a 24-hour forecast: peak wind speed of 8.3 m / s (12th hour), temperature fluctuation range of 23°C to 28°C, and humidity rising to 70%.

[0243] Post-processing: Smooth the prediction curve through Kalman filtering to reduce mutation noise.

[0244] The prediction results are stored in the form of a time series matrix with a time resolution of 15 minutes for subsequent risk prediction.

[0245] Based on the real-time loosening probability value, historical working condition feature library and environmental load prediction curve, the physical information-enhanced LSTM network is input for time series prediction;

[0246] Physical Information Enhanced LSTM (PI-LSTM) introduces a physical constraint layer based on traditional LSTM to ensure that the prediction results conform to the laws of mechanics. The network structure is as follows:

[0247] Input layer: real-time loosening probability value (0-1), historical feature similarity (nearest neighbor matching score with K=10), and environmental load prediction value (wind speed, temperature, humidity);

[0248] LSTM layer: 2 stacked layers, 128 units per layer, dropout rate 0.3;

[0249] Physical constraint layer: Embedded preload attenuation model. When predicting the loosening probability P, the theoretical preload loss ΔF=αP (α=0.85 is the material attenuation coefficient) is calculated simultaneously. If ΔF exceeds a threshold (e.g., 30%), the predicted value is forced to be corrected.

[0250] Output layer: the risk level of loosening every hour in the next 24 hours (0 to 5 levels).

[0251] Training strategy:

[0252] Joint loss function:

[0253] Prediction error: cross entropy loss (weight 0.6);

[0254] Physical constraint loss: |ΔF measured - ΔF predicted| (weight 0.4);

[0255] Data enhancement: Generate extreme working condition data (such as wind speed 15m / s + humidity 90%) through GAN to improve model robustness;

[0256] Transfer learning: Historical feature libraries are used in the pre-training phase, and real-time data is added in the fine-tuning phase.

[0257] Example prediction:

[0258] Input: Real-time loosening probability 0.35, similar working condition matching score 0.82, predicted wind speed 7.2m / s;

[0259] LSTM output: The risk level will increase from level 2 to level 4 in the next 6 hours;

[0260] The physical constraint layer detected ΔF=0.35×0.85=29.75%, close to the threshold of 30%, triggering the risk level to be revised up to level 5.

[0261] According to the output results of the LSTM network, the risk level change curve of each bolt node in the next 24 hours is generated, and a three-dimensional spatiotemporal risk level map is constructed.

[0262] The construction of the three-dimensional spatiotemporal risk level map is divided into two stages: data fusion and visualization rendering:

[0263] Data Fusion:

[0264] Spatial mapping: Based on the GPS coordinates of the bolts on the tower (accuracy ±0.1m), the risk level is mapped to a three-dimensional grid (grid size 0.5m×0.5m×0.5m);

[0265] Time interpolation: Perform cubic spline interpolation on the hourly risk level output by LSTM to generate a minute-level continuous curve;

[0266] Heat map generation: Define a color gradient based on the risk level value (0-5): Level 0 (green, RGB(0,255,0)); Level 3 (yellow, RGB(255,255,0)); Level 5 (red, RGB(255,0,0)).

[0267] Visualization rendering:

[0268] Engine selection: Using Unity3D engine, supporting WebGL export;

[0269] Dynamic interaction: Clicking a bolt node will pop up detailed forecast data (such as risk trends for the next 6 hours);

[0270] Multi-view display: supports top view, side view and structural section view, and can overlay and display wind speed vector field.

[0271] Example: The predicted risk level for bolt node #32 of a transmission tower is Level 3 (yellow) at 2:00 PM, rising to Level 5 (red) at 6:00 PM. The 3D model also shows an increase in the surrounding wind speed from 6 m / s to 9 m / s.

[0272] Operation and maintenance personnel can use VR equipment to view the risk diffusion path and lock in high-risk areas in advance.

[0273] By integrating historical operating condition data with environmental load prediction data, the physical evolution laws of bolt loosening (such as the vibration-temperature-stress coupling equation) are embedded in the LSTM network to dynamically correct the current loosening probability. Time-series prediction generates a future risk evolution curve, upgrading static monitoring to a dynamic risk assessment that incorporates a time dimension. The physical information enhancement mechanism ensures that predictions conform to mechanical laws, avoiding the risk of overfitting associated with purely data-driven approaches. This spatiotemporal risk map can predict loosening trends, support preventative maintenance decision-making, and significantly enhance the monitoring system's foresight.

[0274] S204, mapping the risk level map in real time through a three-dimensional visualization interface. When it is detected that the risk level of bolt loosening exceeds a preset threshold, a warning message is generated and uploaded to the operation and maintenance platform to complete the closed-loop monitoring response.

[0275] Specifically, a digital twin can be established on a 3D visualization platform based on the tower BIM model and bolt location information;

[0276] BIM model analysis and lightweighting: Extract original BIM (Building Information Modeling) data from the tower design files in IFC (Industry Foundation Classes) format. Use BIM analysis tools (such as IFC Open Shell) to extract the geometric coordinates (accuracy ±1mm), material properties (e.g., M20 bolt tensile strength 8.8), and connection relationships of the bolt nodes.

[0277] The model was lightweighted: Mesh simplification algorithms (such as Quadric Edge Collapse Decimation) were used to reduce the number of triangles from 100,000 to 10,000, ensuring smooth rendering in the WebGL engine. Key topological information, such as bolt hole spacing (standard 50mm) and flange thickness (20mm), was also retained.

[0278] Digital twin construction and data mapping: Import the lightweight model into Unity3D or Three.js, and assign a unique ID to each bolt node (e.g., Bolt_001). This ID links multi-source sensor data, including the vibration signal acquisition timestamp (e.g., 2023-10-05T14:30:00Z), temperature gradient (e.g., axial gradient of 2.5°C / m), and stress coupling coefficient (0.85).

[0279] Establish a dynamic data channel: Receive risk level data (JSON format) from the backend server in real time through the WebSocket protocol, with a data update frequency of 1Hz.

[0280] Spatial coordinate calibration and fault tolerance mechanism: The ICP (Iterative Closest Point) algorithm is used to align the BIM model with actual point cloud scan data (such as LiDAR acquisition) to eliminate installation errors (such as bolt hole offset ≤ 2mm).

[0281] Set tolerance threshold: If the deviation between sensor data and model position exceeds 5mm, an abnormal alarm is triggered and a manual review process is started.

[0282] Based on the risk level map data, the risk status of each bolt node is dynamically rendered on the digital twin, using red, yellow, and green to represent different risk levels;

[0283] Risk level color coding rules:

[0284] Red (high risk): loosening probability ≥ 0.8 or temperature gradient ≥ 5°C / m;

[0285] Yellow (medium risk): 0.5≤loosening probability<0.8 and temperature gradient<5℃ / m;

[0286] Green (low risk): loosening probability < 0.5 and temperature gradient < 2°C / m.

[0287] Dynamic rendering technology implementation:

[0288] Shader Programming: Custom shaders are written on the GPU to calculate color interpolation in real time based on the risk value. For example, using HSL (hue, saturation, lightness) interpolation: red (H=0°), yellow (H=60°), green (H=120°), a risk value of 0-1 is mapped to an H value range of 0-120°.

[0289] Enhanced particle effects: A particle system (such as a red pulse halo) is superimposed on high-risk bolts, and the particle emission frequency is positively correlated with the risk value (for example, when the risk is 0.9, the frequency is 10Hz).

[0290] Multiple Level of Detail (LOD) optimizations:

[0291] Close viewing angle (<10 meters): Displays the complete bolt model and the suspension risk label (e.g., "Risk: 0.92");

[0292] Far viewing angle (≥10 meters): Simplified to a colored cube mark, the size of which increases with the risk value (for example, the high-risk mark is enlarged to 1.2 times).

[0293] Cross-platform compatibility guarantee:

[0294] For mobile devices (iOS / Android), a compressed texture format (ASTC 4x4) is used to control rendering latency within 30ms.

[0295] The desktop version supports VR headsets (such as Oculus Rift) and reduces dizziness through asynchronous time warp (ATW) technology.

[0296] Automatically detect high-risk bolt nodes based on preset warning thresholds and generate warning messages containing location coordinates, risk levels, and maintenance recommendations;

[0297] Warning trigger logic and threshold setting

[0298] The system dynamically triggers different levels of alarms by monitoring the bolt loosening probability value (range 0-1) and temperature gradient change rate (unit: ℃ / m / min) in real time, combined with preset multi-level warning thresholds:

[0299] Level 1 Alert (Emergency): Triggered when the loosening probability of a bolt is ≥ 0.9 for three consecutive samplings (each sampled every 1 second), or when the temperature gradient suddenly increases by more than 5°C / m within 1 minute. For example, if the loosening probability of a bolt is 0.91, 0.93, and 0.95 at 2:30:00 PM, 2:30:01 PM, and 2:30:02 PM, respectively, the Level 1 alert is immediately activated.

[0300] Level 2 warning (serious): The probability of loosening in a single sample is ≥0.8, or the temperature gradient change rate is ≥2°C / m / min (e.g., it only takes 1 minute to rise from 2°C / m to 4°C / m).

[0301] Level 3 warning (prompt): The probability of loosening is ≥ 0.6 and does not decrease for 10 minutes (such as continuously above 0.6 from 14:30 to 14:40).

[0302] Early warning message content generation

[0303] When the threshold conditions are met, the system automatically generates a structured early warning message containing the following core information:

[0304] Position coordinates: Accurate positioning is achieved through the three-dimensional coordinates of the bolts bound in the BIM model (such as X=125.6m, Y=34.8m, Z=89.2m), with an accuracy of millimeter level.

[0305] Risk level: The value is marked as "high risk (0.92)", "medium risk (0.75)", etc. according to the threshold matching result, and the value is rounded to two decimal places.

[0306] Maintenance recommendations:

[0307] Match historical work order data from the knowledge graph library. For example, if the temperature gradient is greater than 4°C / m, it is recommended to "check the sealing of the flange and replace the aging gasket."

[0308] Combine the bolt specifications (e.g., M20 bolts) with the recommended torque value (e.g., “immediately tighten to 120 Nm ± 5%)” to ensure compliance with mechanical standards.

[0309] Additional data: Associated with the current infrared thermal image (such as Bolt001_thermal_image.png) so that operation and maintenance personnel can visually identify areas with abnormal temperatures.

[0310] According to the warning level, the message is pushed to the corresponding operation and maintenance terminal, and the disposal feedback information is tracked to complete the closed-loop process of monitoring-warning-disposal.

[0311] Tiered push strategy:

[0312] Level 1 warning: triggered simultaneously via SMS, App push (high-priority notification) and sound and light alarms. If not confirmed within 10 seconds, it will be automatically escalated to the operation and maintenance supervisor; Level 2 warning: App push + email notification, with a response timeout threshold of 30 minutes; Level 3 warning: only recorded in the work order system, and an inspection task will be generated the next day.

[0313] Feedback tracking and work order management:

[0314] Work order state machine: pending → assigned → in process → pending → closed;

[0315] Mobile integration: Operations and maintenance personnel use the app to scan the bolt QR code (encoding format QR Code Model 2) to quickly locate the bolt and upload a comparison chart of the vibration spectrum before and after treatment (for example, after treatment, the energy entropy is reduced from 0.9 to 0.3).

[0316] Closed-loop verification mechanism:

[0317] Automatic review: After the warning is closed, the system continuously monitors the data of the bolt for the next hour. If the risk value does not drop to the safe range (<0.5), the order will be automatically reopened.

[0318] Manual sampling inspection: 5% of closed work orders will be randomly selected and required to upload the torque wrench calibration report (PDF format, including the calibration certificate number).

[0319] Data archiving and knowledge iteration: Warning data is stored in a time series database (such as InfluxDB) with a retention period of 10 years. The risk prediction model is updated quarterly through offline reinforcement learning, for example, adding characteristics of bolt loosening patterns in typhoon weather.

[0320] A 3D visualization platform based on digital twin technology maps risk levels to corresponding locations on the tower's BIM model. Color coding is used to intuitively display risk distribution. When detection values ​​exceed thresholds, a graded early warning mechanism is automatically triggered. A work order with precise location and maintenance recommendations is generated and pushed to the operation and maintenance terminal, enabling visual presentation of monitoring results and rapid response, forming a complete closed loop from data collection to action and feedback. 3D visualization lowers the barrier to data comprehension, and the graded early warning mechanism optimizes resource allocation efficiency, ensuring that high-risk nodes receive priority treatment.

[0321] It can be seen that the original data set containing vibration signals, temperature gradient data and structural stress distribution of bolt connections is collected through a multimodal sensor array, and the original data set is subjected to time-space alignment and frequency domain decomposition processing to obtain a multi-dimensional feature matrix of the bolt node; based on the multi-dimensional feature matrix, the graph neural network model is used to calculate the probability of bolt loosening, and the real-time loosening probability value of each bolt is output; according to the real-time loosening probability value, a risk level map based on time evolution is generated; the risk level map is mapped in real time through a three-dimensional visualization interface. When it is detected that the bolt loosening risk level exceeds the preset threshold, an early warning message is generated and uploaded to the operation and maintenance platform to complete the closed-loop monitoring response, thereby realizing accurate assessment, dynamic prediction and visual early warning of loosening risks, and improving the intelligence level of structural safety monitoring.

[0322] Another embodiment of the present invention provides an online monitoring system for bolt loosening, see Figure 3 , the system may include:

[0323] Processing module 301 is configured to collect a raw data set including vibration signals, temperature gradient data, and structural stress distribution of the bolted joint using a multimodal sensor array, and perform spatiotemporal alignment and frequency domain decomposition on the raw data set to obtain a multidimensional feature matrix of the bolted joint. The multimodal sensor array includes a piezoelectric vibrating plate, an infrared thermal imaging unit, and a fiber Bragg grating sensor.

[0324] Output module 302 is used to calculate the bolt loosening probability based on the multi-dimensional feature matrix using a graph neural network model. By integrating the vibration frequency domain energy entropy, the temperature-stress coupling coefficient, and the structural topology characteristics, and dynamically weighting the key nodes through an attention mechanism, the real-time loosening probability value of each bolt is output;

[0325] Correction module 303 is used to dynamically correct the loosening risk level based on the real-time loosening probability value, combined with the tower's historical operating data and the environmental load time series prediction results, using the LSTM network enhanced by physical information to generate a risk level map based on time evolution, wherein the risk level map includes the risk evolution trend of each bolt node within a preset future time period;

[0326] The generation module 304 is used to map the risk level map in real time through a three-dimensional visualization interface. When it is detected that the risk level of bolt loosening exceeds a preset threshold, an early warning message is generated and uploaded to the operation and maintenance platform to complete the closed-loop monitoring response.

[0327] An embodiment of the present invention further provides a storage medium storing a computer program, wherein the computer program is configured to execute the steps of any one of the above method embodiments when running.

[0328] Specifically, in this embodiment, the above-mentioned storage medium may be configured to store a computer program for performing the following steps:

[0329] S201, collecting an original data set including vibration signals, temperature gradient data, and structural stress distribution of a bolted connection using a multimodal sensor array, performing spatiotemporal alignment and frequency domain decomposition on the original data set to obtain a multidimensional feature matrix of the bolted connection, wherein the multimodal sensor array includes a piezoelectric vibrating piece, an infrared thermal imaging unit, and a fiber Bragg grating sensor;

[0330] S202, based on the multi-dimensional feature matrix, a graph neural network model is used to calculate the probability of bolt loosening. By integrating the vibration frequency domain energy entropy, the temperature-stress coupling coefficient, and the structural topology characteristics, key nodes are dynamically weighted through an attention mechanism to output a real-time loosening probability value for each bolt;

[0331] S203, based on the real-time loosening probability value, combined with the tower's historical operating data and the environmental load time series prediction results, dynamically correct the loosening risk level using the LSTM network enhanced by physical information, and generate a risk level map based on time evolution, wherein the risk level map includes the risk evolution trend of each bolt node within a preset future time period;

[0332] S204, mapping the risk level map in real time through a three-dimensional visualization interface. When it is detected that the risk level of bolt loosening exceeds a preset threshold, a warning message is generated and uploaded to the operation and maintenance platform to complete the closed-loop monitoring response.

[0333] An embodiment of the present invention further provides an electronic device, comprising a memory and a processor, wherein the memory stores a computer program, and the processor is configured to run the computer program to perform the steps in any one of the above method embodiments.

[0334] Specifically, the electronic device may further include a transmission device and an input / output device, wherein the transmission device is connected to the processor, and the input / output device is connected to the processor.

[0335] Specifically, in this embodiment, the processor may be configured to execute the following steps through a computer program:

[0336] S201, collecting an original data set including vibration signals, temperature gradient data, and structural stress distribution of a bolted connection using a multimodal sensor array, performing spatiotemporal alignment and frequency domain decomposition on the original data set to obtain a multidimensional feature matrix of the bolted connection, wherein the multimodal sensor array includes a piezoelectric vibrating piece, an infrared thermal imaging unit, and a fiber Bragg grating sensor;

[0337] S202, based on the multi-dimensional feature matrix, a graph neural network model is used to calculate the probability of bolt loosening. By integrating the vibration frequency domain energy entropy, the temperature-stress coupling coefficient, and the structural topology characteristics, key nodes are dynamically weighted through an attention mechanism to output a real-time loosening probability value for each bolt;

[0338] S203, based on the real-time loosening probability value, combined with the tower's historical operating data and the environmental load time series prediction results, dynamically correct the loosening risk level using the LSTM network enhanced by physical information, and generate a risk level map based on time evolution, wherein the risk level map includes the risk evolution trend of each bolt node within a preset future time period;

[0339] S204, mapping the risk level map in real time through a three-dimensional visualization interface. When it is detected that the risk level of bolt loosening exceeds a preset threshold, a warning message is generated and uploaded to the operation and maintenance platform to complete the closed-loop monitoring response.

[0340] The above describes in detail the structure, features and effects of the present invention based on the embodiments shown in the drawings. The above is only a preferred embodiment of the present invention, but the scope of implementation of the present invention is not limited to what is shown in the drawings. Any changes made in accordance with the concept of the present invention, or modifications to equivalent embodiments with equivalent changes, which do not exceed the spirit covered by the description and drawings, should be within the scope of protection of the present invention.

Claims

1. A method for online monitoring of bolt loosening, characterized in that: The method comprises: The original data set containing vibration signals, temperature gradient data, and structural stress distribution of the bolted joint is collected through a multimodal sensor array, and the original data set is subjected to spatiotemporal alignment and frequency domain decomposition to obtain a multidimensional feature matrix of the bolted joint. The multimodal sensor array includes a piezoelectric vibrating piece, an infrared thermal imaging unit, and a fiber Bragg grating sensor. Based on the multi-dimensional feature matrix, a graph neural network model is used to calculate the probability of bolt loosening. By fusing the vibration frequency domain energy entropy, temperature-stress coupling coefficient and structural topological features, key nodes are dynamically weighted through the attention mechanism to output the real-time loosening probability value of each bolt; the method comprises: constructing a bolt connection topological relationship diagram according to the tower structure mechanics model, wherein the nodes of the topological relationship diagram represent the bolt positions and the edges represent the structural force transmission paths; extracting the vibration frequency domain energy entropy, temperature gradient characteristics and stress coupling coefficient of each bolt node according to the multi-dimensional feature matrix as the initial node features of the graph neural network; and calculating the node features and topological relationships using a multi-head attention mechanism. The inter-point feature propagation weights are used to extract three key indicators: vibration energy entropy, temperature-stress coupling coefficient, and topological connection strength, based on the weighted node feature vectors output by the graph neural network. According to the theoretical formula for preload, the stress fluctuation data in the node features is converted into a percentage of preload loss to obtain a theoretical looseness benchmark value. The theoretical looseness benchmark value and the weighted node features are input into a three-layer fully connected neural network. The first layer performs feature fusion, the second layer adds physical constraints, and the third layer outputs a probability value. The neural network output value is mapped to the 0-1 interval using the Sigmoid activation function, and the probability threshold is calibrated in combination with historical fault data to ultimately generate a real-time loosening probability value for each bolt. Based on the real-time loosening probability value, combined with the tower's historical operating data and environmental load time series prediction results, the loosening risk level is dynamically corrected using a LSTM network enhanced by physical information to generate a time-evolution-based risk level map, wherein the risk level map includes the risk evolution trend of each bolt node within a preset time period in the future; The risk level map is mapped in real time through a three-dimensional visualization interface. When it is detected that the risk level of bolt loosening exceeds a preset threshold, a warning message is generated and uploaded to the operation and maintenance platform to complete the closed-loop monitoring response.

2. The method according to claim 1, characterized in that The multimodal sensor array collects an original data set including vibration signals, temperature gradient data, and structural stress distribution of the bolt connection, performs spatiotemporal alignment and frequency domain decomposition on the original data set, and obtains a multi-dimensional feature matrix of the bolt node, wherein the multimodal sensor array includes a piezoelectric vibrating piece, an infrared thermal imaging unit, and a fiber Bragg grating sensor, including: According to the bolt vibration signal collected by the piezoelectric vibration plate, wavelet packet transform is used to perform noise reduction processing to obtain the denoised vibration time domain signal; Based on the denoised vibration time domain signal, the energy distribution of the 0-5kHz frequency band is calculated by short-time Fourier transform, and the energy proportion and energy entropy of each frequency band are extracted to obtain the vibration frequency domain eigenvector; Based on the temperature distribution image collected by the infrared thermal imaging unit, the bicubic interpolation algorithm is used to improve the spatial resolution, calculate the radial and axial temperature gradients of the bolt connection surface, and obtain the temperature gradient characteristic matrix; Based on the strain data measured by the fiber Bragg grating sensor and the temperature gradient characteristic matrix, the stress fluctuation coefficient caused by temperature change is calculated to obtain the temperature-stress coupling characteristic vector; According to the timestamps of the vibration frequency domain eigenvector, temperature gradient eigenvector and temperature-stress coupling eigenvector, the dynamic time warping algorithm is used to perform time series alignment, and finally a multi-dimensional characteristic matrix of the bolt node is generated.

3. The method according to claim 2, characterized in that According to the real-time loosening probability value, combined with the tower's historical operating data and the environmental load time series prediction results, the loosening risk level is dynamically corrected using the LSTM network enhanced by physical information to generate a risk level map based on time evolution, wherein the risk level map includes the risk evolution trend of each bolt node within a preset time period in the future, including: Based on the historical cases of bolt loosening recorded in the operation and maintenance database, the environmental parameters and equipment status characteristics at the time of loosening are extracted to build a historical working condition feature library; Based on weather forecast data, a temporal convolutional network is used to predict the wind speed, temperature, and humidity trends for the next 24 hours, and the environmental load prediction curve is obtained. Based on the real-time loosening probability value, historical working condition feature library and environmental load prediction curve, the physical information-enhanced LSTM network is input for time series prediction; According to the output results of the LSTM network, the risk level change curve of each bolt node in the next 24 hours is generated, and a three-dimensional spatiotemporal risk level map is constructed.

4. The method according to claim 3, characterized in that The risk level map is mapped in real time through a three-dimensional visualization interface. When it is detected that the risk level of bolt loosening exceeds a preset threshold, a warning message is generated and uploaded to the operation and maintenance platform to complete a closed-loop monitoring response, including: Based on the tower BIM model and bolt location information, a digital twin was established on a 3D visualization platform; Based on the risk level map data, the risk status of each bolt node is dynamically rendered on the digital twin, using red, yellow, and green to represent different risk levels; Automatically detect high-risk bolt nodes based on preset warning thresholds and generate warning messages containing location coordinates, risk levels, and maintenance recommendations; According to the warning level, the message is pushed to the corresponding operation and maintenance terminal, and the disposal feedback information is tracked to complete the closed-loop process of monitoring-warning-disposal.

5. The method according to claim 2, characterized in that The strain data measured by the fiber Bragg grating sensor is combined with the temperature gradient characteristic matrix to calculate the stress fluctuation coefficient caused by temperature change and obtain the temperature-stress coupling characteristic vector, including: Based on the original wavelength offset data collected by the fiber Bragg grating sensor, the strain value of each measuring point is calculated using the Bragg wavelength demodulation algorithm to obtain the initial strain distribution data; According to the initial strain distribution data, the spatial Kalman filter algorithm is used to eliminate the measurement noise and obtain the smoothed strain field matrix; According to the temperature gradient characteristic matrix provided by the infrared thermal imaging unit, the temperature change rate of the corresponding measuring point is extracted, and the temperature-strain mapping relationship table is established in combination with the strain field matrix; According to the temperature-strain mapping relationship table, the temperature compensation coefficient is fitted by the least squares method to calculate the effective stress fluctuation caused by pure mechanical load, and finally the temperature-stress coupling eigenvector is generated.

6. A bolt loosening online monitoring system, characterized in that: The system comprises: a processing module for collecting a raw data set including vibration signals, temperature gradient data, and structural stress distribution of the bolted joint using a multimodal sensor array, performing spatiotemporal alignment and frequency domain decomposition on the raw data set to obtain a multidimensional feature matrix of the bolted joint, wherein the multimodal sensor array includes a piezoelectric vibrating piece, an infrared thermal imaging unit, and a fiber Bragg grating sensor; The output module is used to calculate the probability of bolt loosening based on the multi-dimensional feature matrix using a graph neural network model, and output the real-time loosening probability value of each bolt by fusing the vibration frequency domain energy entropy, temperature-stress coupling coefficient and structural topological features, dynamically weighting key nodes through the attention mechanism; including: constructing a bolt connection topological relationship diagram based on the tower structural mechanics model, wherein the nodes of the topological relationship diagram represent the bolt positions and the edges represent the structural force transmission paths; extracting the vibration frequency domain energy entropy, temperature gradient characteristics and stress coupling coefficient of each bolt node based on the multi-dimensional feature matrix as the initial node features of the graph neural network; adopting a multi-head attention mechanism based on the node features and topological relationships The feature propagation weights between nodes are calculated. Based on the weighted node feature vectors output by the graph neural network, three key indicators are extracted: vibration energy entropy, temperature-stress coupling coefficient, and topological connection strength. According to the preload theory formula, the stress fluctuation data in the node features is converted into a preload loss percentage to obtain a theoretical looseness benchmark value. The theoretical looseness benchmark value and the weighted node features are input into a three-layer fully connected neural network. The first layer performs feature fusion, the second layer adds physical constraints, and the third layer outputs a probability value. The neural network output value is mapped to the 0-1 interval using the Sigmoid activation function. The probability threshold is calibrated in combination with historical fault data to ultimately generate a real-time loosening probability value for each bolt. A correction module is used to dynamically correct the loosening risk level based on the real-time loosening probability value, combined with the tower's historical operating data and the environmental load time series prediction results, using a physical information-enhanced LSTM network to generate a risk level map based on time evolution, wherein the risk level map includes the risk evolution trend of each bolt node within a preset time period in the future; The generation module is used to map the risk level map in real time through a three-dimensional visualization interface. When it is detected that the risk level of bolt loosening exceeds a preset threshold, an early warning message is generated and uploaded to the operation and maintenance platform to complete the closed-loop monitoring response.

7. A storage medium, characterized in that: The storage medium stores a computer program, wherein the computer program is configured to execute the method according to any one of claims 1 to 5 when executed.

8. An electronic device comprising a memory and a processor, characterized in that: A computer program is stored in the memory, and the processor is configured to run the computer program to perform the method according to any one of claims 1 to 5.

Images