A bolt online monitoring method and system

Abstract

The present invention discloses a method and system for online monitoring of bolts. The method comprises: generating a micro-motion characteristic matrix of the bolts based on the multi-dimensional vibration signals and stress wave propagation characteristics of the bolt connection parts; fusing the local stress distribution data collected by the bolt surface strain gauges based on the micro-motion characteristic matrix to output a quantitative indicator of the bolt's health status; outputting a loosening risk level based on the quantitative indicator of the health status and historical loosening evolution data; and dynamically adjusting the monitoring frequency and alarm threshold using an adaptive decision tree model based on the loosening risk level and in combination with equipment operating status parameters, generating a real-time monitoring strategy and a visual early warning report, and synchronously triggering a wireless transmission module to upload the report to a cloud-based operation and maintenance platform. Utilizing the embodiments of the present invention, high-precision, adaptive, and traceable bolt health status assessment and early warning can be achieved.

Description

Technical Field

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

[0002] In fields such as mechanical equipment, bridge engineering, and aerospace, the health of bolted connection structures directly affects the safety and reliability of the entire equipment. Traditional bolt loosening detection mainly relies on manual inspections or offline measurements, which have problems such as low efficiency and poor real-time performance, making it difficult to meet the needs of modern industrial intelligent monitoring. Existing online detection methods based on vibration signals or strain monitoring, although they can achieve a certain degree of real-time monitoring, generally have the following technical defects: First, a single sensor modality is difficult to fully characterize the loose state of the bolt, resulting in insufficient detection accuracy; second, there is a lack of multi-modal feature fusion analysis of vibration signals and stress distribution, which cannot accurately reflect the micromechanical behavior of the bolt connection interface; third, existing methods mostly use static threshold alarms, which cannot dynamically adjust the monitoring strategy according to the operating status of the equipment, and are prone to false alarms or missed alarms; fourth, the management and traceability mechanism of monitoring data is imperfect, making it difficult to support large-scale distributed deployment. In addition, traditional methods have weak predictive capabilities for bolt loosening trends and cannot provide early warning of potential risks. Summary of the Invention

[0003] The purpose of the present invention is to provide a method and system for online monitoring of bolts to address the deficiencies in the prior art and to achieve high-precision, adaptive, and traceable bolt health status assessment and early warning.

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

[0005] Based on the multi-dimensional vibration signals and stress wave propagation characteristics of the bolt connection, the wavelet packet decomposition algorithm is used to extract the characteristic frequency bands of bolt loosening. Combined with the time delay inversion calculation of the stress wave propagation path, the micro-motion characteristic matrix of the bolt is generated.

[0006] Based on the micro-vibration feature matrix, the local stress distribution data collected by the bolt surface strain gauges are integrated, and a deep convolutional neural network is used to construct a multimodal feature vector. The spatial correlation between the vibration characteristics and the stress distribution is associated with the cross-modal attention mechanism to output a quantitative indicator of the bolt health status.

[0007] Based on the health status quantitative indicators and historical loosening evolution data, a graph neural network is used to establish a dynamic relationship map between bolts and connectors, and a temporal convolutional network is used to predict the loosening trend and output the loosening risk level;

[0008] Based on the looseness risk level and combined with the equipment operating status parameters, an adaptive decision tree model is used to dynamically adjust the monitoring frequency and alarm threshold, generate real-time monitoring strategies and visual early warning reports, and simultaneously trigger the wireless transmission module to upload to the cloud operation and maintenance platform.

[0009] Optionally, the method extracts the characteristic frequency band of bolt loosening by wavelet packet decomposition algorithm based on the multi-dimensional vibration signal and stress wave propagation characteristics of the bolt connection part, and generates the micro-motion characteristic matrix of the bolt by combining the time delay inversion calculation of the stress wave propagation path, including:

[0010] Based on the vibration signals collected by the piezoelectric sensor array, the empirical mode decomposition algorithm is used to separate high-frequency noise and low-frequency structural vibration components to generate a denoised multi-band vibration signal.

[0011] Perform 8-layer wavelet packet decomposition on the multi-band vibration signal, calculate the energy entropy value of each sub-band, and select the frequency band with energy entropy mutation within ±15% of the threshold as the loosening characteristic frequency band;

[0012] ‌Distributed optical fiber sensors are used to detect the stress wave propagation path, and the cross-correlation algorithm is used to calculate the time delay difference between adjacent sensors. The micro-displacement fluctuation trajectory of the bolt connection interface is then inverted.

[0013] The energy spectrum of the loosening characteristic frequency band is aligned with the micro-displacement fluctuation trajectory according to the time window, and a three-dimensional micro-motion characteristic matrix is ​​generated by tensor splicing. The matrix dimensions include three characteristic axes: frequency, displacement and time series.

[0014] Optionally, based on the micro-vibration feature matrix, the local stress distribution data collected by the bolt surface strain gauge is integrated, a deep convolutional neural network is used to construct a multimodal feature vector, and the spatial correlation between the vibration characteristics and the stress distribution is associated with the cross-modal attention mechanism to output a quantitative indicator of the bolt health status, including:

[0015] Based on the local stress distribution data collected by the strain gauge array, a stress gradient thermodynamic map of the bolt surface is generated and mapped into a 128×128 pixel stress field image using a bilinear interpolation algorithm.

[0016] The stress field image is input into the ResNet-50 network to extract high-order features, and the micro-motion feature matrix is ​​input into the 3D-CNN network to extract spatiotemporal features to generate a bimodal feature vector.

[0017] ‌The association weights of vibration features and stress features are calculated through the gated attention mechanism, and the dual-modal feature vectors are dynamically weighted fused to generate a joint feature vector;

[0018] A graph convolutional network is used to establish the spatial relationship graph between the bolt connection structure and adjacent components. The joint feature vector is used as the node feature for graph embedding learning to obtain the graph embedding vector.

[0019] The graph embedding vector is mapped to a scalar value of 0-1 through a fully connected layer, and a quantitative health status indicator representing the degree of bolt loosening is output.

[0020] Optionally, the method of establishing a bolt-connector dynamic relationship graph using a graph neural network based on the health status quantitative indicators and historical loosening evolution data, predicting the loosening trend through a temporal convolutional network, and outputting a loosening risk level includes:

[0021] Bolts and connectors are used as graph nodes, and the mechanical coupling strength between nodes is calculated based on health indicators to generate the adjacency matrix of the initial dynamic relationship graph.

[0022] ‌A temporal convolutional network is used to extract multi-scale features from historical loose evolution data, generating a temporal encoding vector of the node state at each time step;

[0023] ‌ Aggregate the state information of adjacent nodes through the graph attention mechanism, and iteratively update the node implicit state in combination with the temporal encoding vector to capture the loose propagation path;

[0024] Input the updated node status into the long short-term memory network, predict the loose evolution trend in the next 24 hours, and output the probabilistic risk curve;

[0025] Based on the slope change of the probabilistic risk curve and the intersection point of the threshold, the dynamic time warping algorithm is used to classify the loosening risk level of the bolts into low risk, medium risk or high risk levels.

[0026] Optionally, based on the looseness risk level and in combination with equipment operating status parameters, an adaptive decision tree model is used to dynamically adjust the monitoring frequency and alarm threshold, generate a real-time monitoring strategy and a visual early warning report, and synchronously trigger a wireless transmission module to upload the report to a cloud operation and maintenance platform, including:

[0027] Normalize the equipment operating speed, load current parameters and looseness risk level to generate a multi-dimensional state feature vector;

[0028] Adopting incremental learning method to train adaptive decision tree model, dynamically splitting nodes according to multi-dimensional state feature vectors, and generating monitoring frequency adjustment rules and alarm thresholds;

[0029] ‌Explore the optimal strategy in the decision tree space based on the Monte Carlo tree search algorithm, and output a real-time monitoring strategy table including monitoring interval, alarm priority, and data sampling accuracy;

[0030] ‌Convert the real-time monitoring strategy table into a JSON format instruction set, drive the visualization engine to generate a three-dimensional dynamic warning map, and upload it to the cloud operation and maintenance platform through the LoRa wireless module.

[0031] Optionally, the method of using a graph convolutional network to establish a spatial relationship graph between the bolt connection structure and adjacent components, and using the joint feature vector as a node feature for graph embedding learning to obtain a graph embedding vector includes:

[0032] The joint eigenvector is normalized to the maximum and minimum according to the modal dimension to eliminate the dimensional difference and obtain the standardized node eigenvector;

[0033] Based on the standardized node feature vector, a two-layer graph convolutional network is used. The first layer aggregates the features of directly adjacent nodes, and the second layer aggregates the features of second-order neighboring nodes. The gated attention mechanism is used to dynamically adjust the aggregation weights to generate an implicit state vector representing the spatial relationship between the bolted connection structure and adjacent components.

[0034] ‌Graph pooling is performed on the latent state vector, and an adaptive mean pooling algorithm is used to fuse node features to output a graph embedding vector that represents the global structural characteristics. The vector dimension is compressed to 256 dimensions.

[0035] Optionally, the step of using a dynamic time warping algorithm to classify the loosening risk level of the bolt into a low risk, medium risk, or high risk level based on the intersection of the slope change of the probabilistic risk curve and the threshold value includes:

[0036] A Savitzky-Golay filter is applied to the probabilistic risk curve with a window length of 15 and a polynomial order of 3 to remove high-frequency noise and preserve trend characteristics;

[0037] ‌Calculate the instantaneous slope of each point on the curve based on the sliding window, and mark the area where the absolute value of the slope exceeds 0.5 and lasts for more than 3 sampling points as a sudden change segment;

[0038] Perform dynamic time warping to align the current risk curve with the historical high-risk and medium-risk template curves, calculate the minimum cumulative distance, and record the number of path bends as a similarity measure;

[0039] If the minimum cumulative distance is less than the high-risk threshold and the number of path bends is ≥2, the risk level is determined to be high. If the minimum cumulative distance is greater than or equal to the high-risk threshold and less than the medium-risk threshold, or if the minimum cumulative distance is less than the high-risk threshold and the number of path bends is <2, the risk level is determined to be medium. Otherwise, if the minimum cumulative distance is greater than or equal to the medium-risk threshold, the risk level is determined to be low.

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

[0041] The extraction module is used to extract the characteristic frequency bands of bolt loosening based on the multi-dimensional vibration signals and stress wave propagation characteristics of the bolt connection part through the wavelet packet decomposition algorithm. Combined with the time delay inversion calculation of the stress wave propagation path, it generates the micro-motion characteristic matrix of the bolt;

[0042] A construction module is used to integrate the local stress distribution data collected by the bolt surface strain gauge based on the micro-vibration feature matrix, construct a multimodal feature vector using a deep convolutional neural network, associate the spatial correlation between the vibration characteristics and the stress distribution through a cross-modal attention mechanism, and output a quantitative indicator of the bolt health status;

[0043] A prediction module is used to establish a dynamic relationship map between bolts and connectors using a graph neural network based on the health status quantitative indicators and historical loosening evolution data, predict the loosening trend through a temporal convolutional network, and output a loosening risk level;

[0044] A generation module is used to dynamically adjust the monitoring frequency and alarm threshold based on the loosening risk level and combined with the equipment operation status parameters using an adaptive decision tree model, generate a real-time monitoring strategy and a visual early warning report, and simultaneously trigger the wireless transmission module to upload to the cloud operation and maintenance platform.

[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 existing technology, the present invention provides a method for online monitoring of bolts, which generates a micro-motion feature matrix of the bolts based on the multi-dimensional vibration signals and stress wave propagation characteristics of the bolt connection parts; based on the micro-motion feature matrix, the local stress distribution data collected by the strain gauges on the bolt surface are integrated to output a quantitative indicator of the health status of the bolts; based on the quantitative indicator of the health status and historical loosening evolution data, the loosening risk level is output; based on the loosening risk level and in combination with the equipment operating status parameters, an adaptive decision tree model is used to dynamically adjust the monitoring frequency and alarm threshold, generate a real-time monitoring strategy and a visual early warning report, and simultaneously trigger the wireless transmission module to upload to the cloud operation and maintenance platform, thereby realizing high-precision, adaptive, and traceable bolt health status assessment and early warning. BRIEF DESCRIPTION OF THE DRAWINGS

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

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

[0050] Figure 3 A schematic structural diagram of a bolt online monitoring system 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 bolts. The method 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 a bolt online monitoring method 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 bolt online 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 bolt online monitoring methods.

[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 bolts, which may include the following steps:

[0060] S201, based on the multi-dimensional vibration signal and stress wave propagation characteristics of the bolt connection part, the characteristic frequency band of bolt loosening is extracted by wavelet packet decomposition algorithm, and combined with the time delay inversion calculation of the stress wave propagation path, the micro-motion characteristic matrix of the bolt is generated;

[0061] Specifically, the vibration signal collected by the piezoelectric sensor array can be used to separate high-frequency noise and low-frequency structural vibration components through the empirical mode decomposition algorithm to generate a denoised multi-band vibration signal;

[0062] The piezoelectric sensor array adopts a circular layout, with eight sensors evenly distributed around the bolted joint. The sampling frequency is set to 20 kHz (meeting the Nyquist sampling requirement for mechanical vibration signals). The collected raw vibration signal contains high-frequency noise (such as electromagnetic interference and equipment friction noise) and low-frequency structural vibration components (such as periodic micro-vibrations caused by loose bolts). The empirical mode decomposition (EMD) algorithm uses adaptive decomposition to split the signal into multiple intrinsic mode functions (IMFs).

[0063] Decomposition process: EMD is performed on each sensor signal, with the number of decomposition layers dynamically adjusted (typically 6 to 10). For example, a decomposition may yield IMF1 (high-frequency noise, frequencies > 5kHz), IMF2-IMF4 (intermediate-frequency structural vibration, frequencies 1-5kHz), and IMF5-IMF8 (low-frequency vibration, frequencies < 1kHz).

[0064] Filter valid IMFs by correlation coefficient: If the correlation coefficient between an IMF and the original signal is less than 0.2 (empirical threshold), it is considered noise and removed. For example, IMF1 is removed due to its correlation coefficient of 0.15.

[0065] Signal reconstruction: IMF2 to IMF5 (correlation coefficient > 0.3) are retained. The reconstructed signal frequency range is 0.5 to 4 kHz, covering the characteristic frequency band of bolt loosening.

[0066] Denoising effect evaluation: The signal-to-noise ratio (SNR) is improved from 15dB of the original signal to 28dB.

[0067] Output example: After EMD processing, the vibration signal of a bolt is reconstructed into four frequency bands (0.5-1kHz, 1-2kHz, 2-3kHz, and 3-4kHz). Each frequency band corresponds to a different loosening mode (for example, the low frequency band reflects macroscopic displacement, and the high frequency band reflects microscopic friction).

[0068] Perform 8-layer wavelet packet decomposition on the multi-band vibration signal, calculate the energy entropy value of each sub-band, and select the frequency band with energy entropy mutation within ±15% of the threshold as the loosening characteristic frequency band;

[0069] Wavelet packet decomposition (WPD) uses the Daubechies 4 (db4) wavelet basis with 8 decomposition layers, further subdividing each frequency band signal into 2^8=256 sub-bands (frequency resolution ≈ 15.625 Hz).

[0070] Decomposition and energy entropy calculation: Eight layers of WPD are performed on each of the four reconstructed frequency bands. For example, the 1-2 kHz frequency band is decomposed into sub-bands 1 (1-1.015 kHz) to 256 (1.984-2 kHz).

[0071] Calculate the energy entropy of each sub-band: ,in, The ratio of the sub-band energy to the total energy. Under normal conditions, the energy entropy distribution is stable; when it is loose, the local entropy value suddenly changes (for example, the entropy value of a sub-band suddenly increases from 0.8 to 1.2).

[0072] Feature frequency band screening:

[0073] Set a dynamic threshold: Take the sliding average of historical normal data entropy values ​​(window length 100), allowing for ±15% fluctuation. For example, if the normal entropy value for a sub-band is 1.0, the threshold range is 0.85 to 1.15. Any value exceeding this threshold is marked as abnormal.

[0074] Example of screening results: Three sub-bands with entropy values ​​of 1.3 (exceeding the threshold of 1.15) were detected in the 2.5-2.6 kHz range and were identified as loose characteristic frequency bands.

[0075] ‌Distributed optical fiber sensors are used to detect the stress wave propagation path, and the cross-correlation algorithm is used to calculate the time delay difference between adjacent sensors. The micro-displacement fluctuation trajectory of the bolt connection interface is then inverted.

[0076] Distributed fiber optic sensors are arranged along the bolt axis, with a spacing of 10 mm, covering the entire length of the bolt (for example, five sensors are arranged for a length of 50 mm). When stress waves (induced by vibration or external impact) propagate through the bolt, each sensor detects the difference in arrival time.

[0077] Delay difference calculation: Cross-correlation operation is performed on adjacent sensor signals (such as sensor 1 and 2): the sliding window length is 10 ms (corresponding to 200 sampling points) and the step size is 1 ms.

[0078] Calculate the peak position of the cross-correlation coefficient and determine the time delay Δt. For example, if the signal from sensor 2 is 0.2ms later than that from sensor 1, the corresponding stress wave propagation velocity v = 10mm / 0.2ms = 50m / s.

[0079] Micro-displacement inversion: A propagation path matrix is ​​constructed based on the time-delay difference sequence, and micro-displacement fluctuations are inverted using the least squares method. For example, one inversion revealed a bolt head displacement of 0.05 mm in the x-direction and 0.02 mm in the y-direction, with a frequency of 2 Hz (corresponding to the loosening micro-vibration frequency).

[0080] Trajectory visualization: Generates a time-displacement curve and marks areas where the displacement amplitude exceeds 0.03mm (safety threshold) as high-risk periods.

[0081] Parameter example: Fiber optic sensor sensitivity 0.1με (microstrain), time delay resolution 0.01ms, displacement inversion error <±0.005mm.

[0082] The energy spectrum of the loosening characteristic frequency band is aligned with the micro-displacement fluctuation trajectory according to the time window, and a three-dimensional micro-motion characteristic matrix is ​​generated by tensor splicing. The matrix dimensions include three characteristic axes: frequency, displacement and time series.

[0083] Time window alignment:

[0084] A sliding window mechanism is used with a window length of 1 second (including 20,000 sampling points) and a step size of 0.1 second.

[0085] The characteristic frequency band energy spectrum (256 dimensions) and micro-displacement trajectory (x and y dual channels, 100 dimensions each) in each window are linearly interpolated and aligned to ensure timing synchronization.

[0086] Tensor concatenation:

[0087] Dimension definition: Frequency axis: 256 sub-band energy values; Displacement axis: 100 points in the x / y direction (200 dimensions in total); Time axis: 10 time windows (1 second window, 0.1 second step).

[0088] Data format: 256×200×10 three-dimensional matrix, each element is a normalized energy or displacement value (0~1).

[0089] For example, during the period t = 5 to 6 seconds, the energy in the characteristic frequency band of a bolt at 2.5 kHz suddenly increases by 30%, accompanied by periodic fluctuations in the x-direction displacement (amplitude 0.06 mm). The element value at the corresponding position in the matrix is ​​marked as 0.8 (high risk).

[0090] This step collects vibration signals through a piezoelectric sensor array and uses a wavelet packet decomposition algorithm to isolate characteristic frequency bands strongly correlated with bolt loosening (such as high-frequency components between 5 and 8 kHz). Distributed fiber optic sensors are then used to detect differences in stress wave propagation delays, thereby inferring micron-scale displacement fluctuations at the bolt connection interface. By aligning the frequency-domain energy spectrum with the displacement trajectory in the time dimension, a three-dimensional feature matrix comprising frequency, displacement, and time series is constructed. This fully characterizes the multi-physics characteristics of bolt loosening, breaking through the limitations of traditional single-source vibration monitoring. By integrating vibration frequency-domain characteristics with stress wave delay information, the sensitivity of micro-vibration detection is significantly improved, providing high-precision input data for subsequent health status assessments.

[0091] S202: Based on the micro-vibration feature matrix, the local stress distribution data collected by the bolt surface strain gauge is integrated, and a multimodal feature vector is constructed using a deep convolutional neural network. The spatial correlation between the vibration characteristics and the stress distribution is correlated through a cross-modal attention mechanism to output a quantitative indicator of the bolt health status;

[0092] Specifically, a bolt surface stress gradient thermal map can be generated based on the local stress distribution data collected by the strain gauge array, and then mapped into a 128×128 pixel stress field image using a bilinear interpolation algorithm.

[0093] A strain gauge array is attached to the bolt surface in a cross-shaped layout (e.g., a 4×4 array). Each strain gauge (e.g., HBM LY41) is sampled at 1kHz, collecting real-time microstrain data (in με) as the bolt is loaded. The raw data is preprocessed with an anti-aliasing filter (cutoff frequency 500Hz), and the strain gradient at each point is calculated:

[0094] Gradient calculation: With each strain gauge as the center, calculate the strain difference between it and the four adjacent gauges (upper, lower, left, and right). For example, if the strain on the center gauge is 200με and the strain on the right gauge is 220με, then the horizontal gradient is 20με / mm (assuming 1mm spacing).

[0095] Thermal map generation: Normalize the gradient value to the range of 0-255, map it to grayscale value, and generate a stress gradient thermal map with an original resolution of 16×16;

[0096] Bilinear interpolation:

[0097] Divide the 16×16 grid into 15×15 quadrilateral elements;

[0098] Interpolate the unknown points (such as coordinates (x,y)) within each cell:

[0099] Calculate the weights of the four adjacent known points (inversely proportional to the distance);

[0100] The strain gradient at the interpolation point is obtained by weighted summation;

[0101] Repeated interpolation to 128×128 pixels fills in missing areas. For example, an original pixel pitch of 1mm corresponds to a resolution of 0.125mm after interpolation, improving image detail.

[0102] The final generated stress field image is saved in PNG format through the OpenCV library. The image size is 128×128, and each pixel corresponds to the actual physical size of 0.125mm×0.125mm. The grayscale value reflects the degree of local stress concentration (for example, grayscale 200 indicates a high stress area).

[0103] The stress field image is input into the ResNet-50 network to extract high-order features, and the micro-motion feature matrix is ​​input into the 3D-CNN network to extract spatiotemporal features to generate a bimodal feature vector.

[0104] ResNet-50 network performs transfer learning on stress field images:

[0105] Input adjustment: The pre-trained ImageNet weights are adapted to a single-channel grayscale image, and the first-layer convolution kernel is modified to a 1-channel input;

[0106] Feature extraction: intercept the first four residual blocks of ResNet-50 (conv1~conv4_x) and output a 1024-dimensional feature vector. For example, after convolution and pooling of a 128×128 input image, the final feature map size is 4×4×1024;

[0107] Global average pooling: compresses the 4×4×1024 tensor into a 1024-dimensional vector as the stress modal feature.

[0108] 3D-CNN network processing micro-motion feature matrix (dimension: frequency × displacement × time series = 64 × 64 × 100):

[0109] Network structure:

[0110] Convolutional layer 1: 3D convolution kernel 5×5×5, 64 output channels, stride 2×2×2;

[0111] Convolutional layer 2: 3D convolution kernel 3×3×3, output channels 128, stride 1×1×1;

[0112] Pooling layer: 3D maximum pooling, kernel size 2×2×2;

[0113] Feature extraction: After convolution and pooling of the input 64×64×100, the output is a 8×8×25×128 tensor;

[0114] Flattening and full connection: The 8×8×25×128 layer is flattened to 204800 dimensions and then compressed to 1024 dimensions through the fully connected layer as the vibration modal feature.

[0115] The bimodal eigenvector is composed of the stress mode (1024 dimensions) and the vibration mode (1024 dimensions), with a total dimension of 2048.

[0116] ‌The association weights of vibration features and stress features are calculated through the gated attention mechanism, and the dual-modal feature vectors are dynamically weighted fused to generate a joint feature vector;

[0117] The structure of the gated attention mechanism is as follows:

[0118] Input processing:

[0119] Stress eigenvector S (1024 dimensions);

[0120] Vibration eigenvector V (1024 dimensions);

[0121] Attention weight calculation:

[0122] Concatenate S and V to get a 2048-dimensional joint input;

[0123] Generate intermediate representation through fully connected layers (output dimension 512) and ReLU activation;

[0124] Then pass through the fully connected layer (output dimension 2) and Softmax to generate the weight vector [α, 1-α], α∈[0,1];

[0125] Dynamic weighted fusion: Joint feature vector F = α* S + (1-α) * V; Example: If α = 0.7, it means that the stress feature contributes 70% and the vibration feature contributes 30%.

[0126] Training optimization:

[0127] Loss function: cross entropy loss combined with weight sparsity regularization (L1 regularization coefficient 0.01);

[0128] Backpropagation: Adam optimizer is used (learning rate 0.001, batch size 32).

[0129] The dimension of the final generated joint feature vector is kept at 1024, integrating the key information of the two modalities.

[0130] A graph convolutional network is used to establish the spatial relationship graph between the bolt connection structure and adjacent components. The joint feature vector is used as the node feature for graph embedding learning to obtain the graph embedding vector.

[0131] Specifically, a graph convolutional network is used to establish a spatial relationship graph between the bolt connection structure and adjacent components, and the joint feature vector is used as a node feature for graph embedding learning to obtain a graph embedding vector, which can include:

[0132] The joint eigenvector is normalized to the maximum and minimum according to the modal dimension to eliminate the dimensional difference and obtain the standardized node eigenvector;

[0133] The joint feature vector is a fusion of vibration and stress features. The raw data collected by different sensors has significantly different numerical ranges due to differences in physical dimensions (e.g., vibration acceleration in g, stress in MPa). For example, the energy value of a bolt's vibration characteristic frequency band is 0.8 (dimensionless), while the surface stress gradient is 120 MPa. Directly inputting these two into a graph convolutional network will bias the model towards large numerical features due to the magnitude difference.

[0134] Min-Max Normalization maps each modal feature to the [0,1] interval through linear transformation. The formula logic is:

[0135] For each modal dimension, calculate the maximum value (Max) and minimum value (Min) of all nodes in that dimension, and scale each eigenvalue x by (x - Min) / (Max - Min). For example:

[0136] Vibration energy value range [0.2, 1.6] → after normalization 0.2→0, 1.6→1;

[0137] Stress gradient range [50, 200]MPa → 50→0, 200→1.

[0138] In practice, a sliding window is used to update extreme values ​​to adapt to dynamic operating conditions. For example, the Max and Min values ​​are recalculated every 10 minutes to avoid normalization distortion caused by sudden changes in equipment load. The normalized feature vectors are comparable across all dimensions, ensuring fair weight distribution in subsequent graph convolutions.

[0139] Based on the standardized node feature vector, a two-layer graph convolutional network is used. The first layer aggregates the features of directly adjacent nodes, and the second layer aggregates the features of second-order neighboring nodes. The gated attention mechanism is used to dynamically adjust the aggregation weights to generate an implicit state vector representing the spatial relationship between the bolted connection structure and adjacent components.

[0140] Graph structure definition:

[0141] Node: Bolt and its directly connected parts (such as flange, bracket), each node represents a physical entity;

[0142] Edges: Define the adjacency matrix based on the mechanical assembly relationship. The weight of the edge between directly adjacent nodes is 1 (e.g., bolt-flange), and the weight of the edge between second-order neighbors (e.g., bolt-flange-bracket) is 0.5.

[0143] First layer of graph convolution (GCN Layer 1):

[0144] Input: Normalized node feature vector (dimension 128);

[0145] Adjacency matrix A: contains only directly adjacent nodes (edge ​​weight 1);

[0146] Aggregation: For each node, a weighted sum of its own features and neighboring features is performed. The weight matrix W1 is learned through training, and the activation function uses Reluctant Unit (ReLU). For example, a bolt node aggregates its own features (weight 0.6) and flange features (weight 0.4) to generate an intermediate feature vector (dimension 256).

[0147] Gated Attention:

[0148] Attention score calculation: For nodes i and j, the relevance score is calculated using a learnable parameter matrix, such as score(i,j) = W_a [h_i || h_j], where || represents concatenation.

[0149] Gating coefficient: The score is mapped to (0, 1) through the sigmoid function as the aggregation weight. For example, the attention coefficient of the bolt and flange is 0.8, and that of the bracket is 0.3;

[0150] Dynamic aggregation: Node features are updated to a combination of their own features and weighted neighbor features, such as h_i' = h_i +Σ(gate_ij * h_j).

[0151] Second layer of graph convolution (GCN Layer 2):

[0152] Input: Hidden features output by the first layer (dimension 256);

[0153] Adjacency matrix expansion: including second-order neighbor nodes (edge ​​weight 0.5);

[0154] Cross-layer aggregation: Combined with the gated attention mechanism, the contribution of distant nodes is dynamically adjusted. For example, the bolt node indirectly aggregates the bracket features through the flange, and the weight is assigned by the attention mechanism (such as 0.2).

[0155] Output: Hidden state vector (dimension 512), encoding the node's own characteristics and its structural relationship in the multi-hop neighborhood.

[0156] ‌Graph pooling is performed on the latent state vector, and an adaptive mean pooling algorithm is used to fuse node features to output a graph embedding vector that represents the global structural characteristics. The vector dimension is compressed to 256 dimensions.

[0157] Adaptive Mean Pooling:

[0158] Pooling goal: Convert variable-size graph structures (such as the number of different bolt connection parts) into global features of fixed dimensions;

[0159] Adaptive mechanism: Dynamically adjusts the pooling granularity based on node importance. Importance is measured by the node feature norm. For example, a larger feature L2 norm indicates a more significant contribution of the node to the global state.

[0160] Pooling step:

[0161] Node sorting: Arrange nodes in descending order of characteristic norm, for example, bolt node (norm 12.5) > flange node (10.2) > bracket node (8.7);

[0162] Hierarchical Pooling:

[0163] First layer: retain the original features of the top 50% nodes (such as bolts and flanges);

[0164] The second layer: perform mean pooling on the remaining nodes (such as brackets) to generate summary features;

[0165] Feature concatenation: Concatenate the retained features with the summary features to obtain pooled features (dimension 1024).

[0166] Dimensionality reduction:

[0167] Fully connected layer: maps 1024-dimensional features to 256 dimensions, and uses Tanh as the activation function;

[0168] Regularization: Dropout (ratio 0.3) is used to prevent overfitting.

[0169] Example output: The graph embedding vector of a bolt connection is [0.7, -0.2, ..., 0.5] (256 dimensions), where positive dimensions indicate high structural tightness and negative dimensions indicate potential looseness risk.

[0170] The graph embedding vector is mapped to a scalar value of 0-1 through a fully connected layer, and a quantitative health status indicator representing the degree of bolt loosening is output.

[0171] Fully connected layer structure:

[0172] Input: 256-dimensional graph embedding vector;

[0173] Hidden layer: 128 dimensions, ReLU activation;

[0174] Output layer: 1 dimension, Sigmoid activation;

[0175] Training and Inference:

[0176] Label data: 0 means normal, 1 means completely loose, the label is obtained through manual calibration or high-precision torque sensor;

[0177] Loss function: binary cross entropy (BCE Loss);

[0178] Threshold classification: Output value ≥0.7 is judged as "high loosening risk", 0.3-0.7 is "need to monitor", and <0.3 is "normal".

[0179] Example: The graph embedding vector of a bolt outputs 0.85 after the fully connected layer, which is judged as high risk and triggers a real-time warning.

[0180] The micro-vibration feature matrix is ​​fed into a 3D-CNN to extract spatiotemporal features. Simultaneously, the stress field images generated by the strain gauges are processed using a ResNet-50 network. A gated attention mechanism is used to dynamically weight and fuse these two types of features. Furthermore, a graph convolutional network is used to model the mechanical coupling between the bolt and adjacent components. Ultimately, a health indicator with a continuously varying range of 0-1 is output (0 indicates complete tightening, 1 indicates complete failure). This achieves deep cross-modal correlation between vibration and stress data, overcoming the crudeness of traditional threshold alarms. The quantitative indicator accurately reflects the stage of loosening development (e.g., 0.3-0.5 indicates a decrease in preload, 0.5-0.7 indicates visible loosening).

[0181] S203, based on the health status quantitative indicators and historical loosening evolution data, using a graph neural network to establish a dynamic relationship map between bolts and connectors, predicting the loosening trend through a temporal convolutional network, and outputting a loosening risk level;

[0182] Specifically, bolts and connectors can be used as graph nodes, and the mechanical coupling strength between nodes can be calculated based on health indicators to generate the adjacency matrix of the initial dynamic relationship graph;

[0183] Constructing a dynamic graph of the relationship between bolts and connectors is central to analyzing loosening propagation paths. Each bolt and its directly connected components (such as flanges and brackets) are abstracted as graph nodes. Node attributes include a quantitative health indicator (a scalar value from 0 to 1, with 0 indicating full tightening and 1 indicating full loosening), material stiffness coefficients (e.g., bolts made of 8.8-grade steel have a stiffness coefficient of 210 GPa), and geometric parameters (e.g., bolt diameter M20, effective thread length 50 mm).

[0184] The mechanical coupling strength is calculated using the stress wave propagation path analysis method:

[0185] Stress wave path modeling: Based on the stress wave propagation trajectory detected by the distributed fiber optic sensor, the bolted joint interface is divided into 5mm x 5mm grid cells, and the stress transfer efficiency of each cell is calculated. For example, if a cell transfers 85% of the stress wave energy during vibration, its transfer efficiency is 0.85.

[0186] Quantifying coupling strength: Mechanical coupling strength is defined as the geometric mean of the stress transfer efficiencies between adjacent nodes. For example, if there are three stress wave paths between bolt A and flange B, with transfer efficiencies of 0.8, 0.7, and 0.9, respectively, the coupling strength is (0.8 × 0.7 × 0.9)^(1 / 3) = 0.79.

[0187] Adjacency matrix generation: If the coupling strength is ≥ 0.7 (an empirical threshold), the corresponding position in the adjacency matrix is ​​set to 1; otherwise, it is set to 0. For multi-bolt systems, the adjacency matrix dimension is N × N (N is the total number of nodes). For example, a system with 10 bolts and 5 connectors generates a 15 × 15 matrix.

[0188] The edge weights of the dynamic relationship graph are dynamically adjusted based on real-time health indicators. For example, when the health indicator of bolt A increases from 0.2 to 0.6 (increasing looseness), the coupling strength weight between bolt A and flange B decreases from 0.79 to 0.65, triggering a real-time update of the edge weight in the adjacency matrix.

[0189] ‌A temporal convolutional network is used to extract multi-scale features from historical loose evolution data, generating a temporal encoding vector of the node state at each time step;

[0190] A temporal convolutional network (TCN) is used to capture the long-term dependencies of bolt loosening trends. The input data is the health indicator time series data of the past 30 days (sampling interval is 1 hour, totaling 720 time points), and the input dimension of each node is 1 (health indicator value).

[0191] Network structure design:

[0192] Dilated convolution layers: 3 layers of dilated convolutions with dilation coefficients of 1, 2, and 4, respectively. The convolution kernel size is 3×1, and the number of channels increases from 32 to 64 layer by layer. For example, the first layer of convolution kernel covers a 3-hour window, the second layer expands to 6 hours with a dilation coefficient of 2, and the third layer expands to 12 hours.

[0193] Residual connections: Each layer's output is added to the input to prevent gradient vanishing. For example, given a health indicator sequence [0.1, 0.15, 0.2...], convolution extracts multi-scale fluctuation features (e.g., short-term fluctuations from 0.1 to 0.12 and medium-term trends from 0.1 to 0.2).

[0194] Activation function: LeakyReLU (negative slope 0.1) is used instead of ReLU to avoid neuron death.

[0195] Temporal encoding vector generation:

[0196] The TCN output is a 64-dimensional feature vector for each time step. For example, a health indicator of 0.3 at a time step corresponds to the encoding vector [0.12, -0.05, 0.3, ..., 0.18], where each dimension represents the fluctuation characteristics of different time scales (for example, the first dimension reflects hourly fluctuations, and the third dimension reflects daily trends).

[0197] Multi-scale feature fusion:

[0198] The outputs of different expansion coefficients are merged through cross-layer feature splicing. For example, the convolution results of expansion coefficients 1, 2, and 4 are spliced ​​into a 192-dimensional vector, and then compressed to 64 dimensions through a fully connected layer to enhance sensitivity to sudden events (such as loosening acceleration caused by impact loads).

[0199] ‌ Aggregate the state information of adjacent nodes through the graph attention mechanism, and iteratively update the node implicit state in combination with the temporal encoding vector to capture the loose propagation path;

[0200] The Graph Attention Mechanism (GAT) is used to dynamically integrate the influence of adjacent nodes. The implicit state update of each node is divided into three stages:

[0201] Attention weight calculation:

[0202] For node i and its neighbor j, calculate the attention coefficient e_ij = LeakyReLU(a^T [W h_i || W h_j]), where a is the learnable parameter vector, W is the weight matrix, h_i and h_j are the temporal encoding vectors of nodes i and j.

[0203] Multi-head attention (4 heads) is used, with each head independently calculating the attention weights and then concatenating them. For example, the attention weights of node i to node j are 0.3, 0.2, 0.4, and 0.1 in the 4 heads, respectively. The final weight is the concatenated 4-dimensional vector.

[0204] Neighbor information aggregation:

[0205] For each head k, aggregate neighbor features: ;

[0206] After concatenating the multiple head outputs, the dimensionality is reduced through a fully connected layer: h_i' = Concat(h_i^1', h_i^2', h_i^3', h_i^4'). For example, the hidden state of node i is expanded from 64 dimensions to 256 dimensions (4 heads × 64), and then compressed back to 64 dimensions.

[0207] Status Update:

[0208] Combine its own state with the aggregation result: h_i_new = σ(h_i' + h_i), where σ is the Sigmoid function.

[0209] The update is iterated three times to gradually capture the multi-hop propagation effect. For example, the first iteration captures the direct neighbor influence, and the third iteration captures the indirect coupling of nodes three hops away.

[0210] Visualization of loose propagation paths:

[0211] By calculating the cumulative value of attention weights between nodes, a weighted heat map is plotted. For example, in the path from bolt A to flange B to bracket C, the attention weights are 0.6 and 0.4, respectively, indicating that loosening mainly propagates along the path from A to B.

[0212] Input the updated node status into the long short-term memory network, predict the loose evolution trend in the next 24 hours, and output the probabilistic risk curve;

[0213] In one implementation, a long short-term memory (LSTM) network is used for time series prediction, with the input being a latent state sequence (64 dimensions) and the output being the loosening probability of the next 24 time steps (1 point per hour).

[0214] Network Configuration:

[0215] Hidden layer: 128 LSTM units with a dropout rate of 0.2 to prevent overfitting;

[0216] Input window length: 24 hours (consistent with the prediction length);

[0217] Output layer: fully connected layer + Sigmoid, mapped to a probability value of 0-1.

[0218] Training strategy:

[0219] Loss function: weighted cross entropy, giving 3 times the weight to high-risk samples (health index > 0.7 in history);

[0220] Optimizer: Adam, learning rate 0.001, batch size 32;

[0221] Early stopping mechanism: If the validation set loss does not decrease for 5 consecutive rounds, the training is terminated.

[0222] Probabilistic risk curve generation:

[0223] The model outputs 24 probability values, such as [0.1, 0.15, 0.2, ..., 0.45], representing the probability that the health indicator will exceed a threshold of 0.5 every hour in the future. The curve is smoothed using cubic spline interpolation and annotated with 95% confidence intervals.

[0224] Risk level classification:

[0225] Slope mutation detection: Apply a Savitzky-Golay filter (window 15, order 3) to the smoothed curve and calculate the slope point by point. If the absolute value of the slope at three consecutive points is greater than 0.5, it is identified as a sudden change.

[0226] Dynamic Time Warping (DTW): Aligns the current curve with a historical high-risk template (such as an exponential rise curve) and calculates the minimum cumulative distance.

[0227] Threshold crossing: If the probability value exceeds 0.7 and the slope is greater than 0, a high-risk alarm is immediately triggered.

[0228] Based on the slope change of the probabilistic risk curve and the intersection point of the threshold, the dynamic time warping algorithm is used to classify the loosening risk level of the bolts into low risk, medium risk or high risk levels.

[0229] In another more detailed implementation, a dynamic time warping algorithm is used to classify the loosening risk level of bolts into low, medium, or high risk levels based on the slope change of the probabilistic risk curve and the intersection of the threshold, including:

[0230] A Savitzky-Golay filter is applied to the probabilistic risk curve with a window length of 15 and a polynomial order of 3 to remove high-frequency noise and preserve trend characteristics;

[0231] The Savitzky-Golay filter (SG filter) is a smoothing algorithm based on local polynomial fitting. Its core concept is to suppress noise while preserving the signal's trend characteristics through polynomial regression within a sliding window. In this step, a window length of 15 (i.e., each fit covers 15 consecutive sampling points) and a cubic polynomial (order 3) are used for smoothing the bolt loosening risk curve (e.g., a 24-hour probability curve with values ​​ranging from 0 to 1 representing the loosening probability) to smooth the bolt loosening risk curve.

[0232] Operation process:

[0233] Sliding Window: Starting from the first sampling point of the risk curve, 15 consecutive points are taken as the fitting window. For example, if the sampling interval is 1 minute, the window covers a 15-minute data segment.

[0234] Polynomial Fitting: Performs a cubic polynomial fit (in the form of y = a0 + a1x + a2x² + a3x³) to the 15 points within the window, calculating coefficients a0 through a3 using the least squares method. For example, if the data within a window fluctuates significantly, the fitted curve can smooth out high-frequency jitter.

[0235] Center point replacement: Only the fitted value of the window center point (the 8th point) is retained as the smoothed output to avoid edge effects. For example, after fitting the original data point sequence [0.2, 0.3, 0.5, ..., 0.8], the center point 0.5 may be corrected to 0.48.

[0236] Window step: The window slides 1 point to the right, and the above process is repeated until all data is covered.

[0237] Technical effects: Noise suppression: For example, a certain section of the original curve has random fluctuations of 0.1~0.3 due to sensor noise. After SG filtering, the fluctuation amplitude is reduced to ±0.05; Trend retention: If the curve has a slow upward trend (such as a linear increase from 0.3 to 0.6), SG filtering can accurately retain the trend and avoid the phase delay caused by the traditional moving average method.

[0238] ‌Calculate the instantaneous slope of each point on the curve based on the sliding window, and mark the area where the absolute value of the slope exceeds 0.5 and lasts for more than 3 sampling points as a sudden change segment;

[0239] The instantaneous slope reflects the rate of change of the risk curve and is used to detect the rapid deterioration or alleviation of loosening risk. The slope is calculated using the sliding window method:

[0240] Window definition: With the current point as the center, take two points before and after it to form a 5-point window (i.e., window length 5). For example, when calculating the slope at minute t, use data from minutes t-2 to t+2.

[0241] Linear Fit: Perform linear regression on the five points within the window. The slope is the instantaneous slope. For example, the slope of the window data [0.4, 0.5, 0.6, 0.7, 0.8] is 0.1 / minute, indicating that the risk increases by 0.1 per minute.

[0242] Mutation segment determination:

[0243] Slope threshold: absolute value > 0.5 (e.g., a slope of +0.6 indicates a rapid increase in risk);

[0244] Duration: The slopes of three consecutive sampling points all exceed the threshold. For example, if the slopes of a curve are 0.6, 0.7, and 0.55 respectively from 10:00 to 10:02 (3 minutes), it is marked as a sudden change segment.

[0245] Application examples:

[0246] High risk rise: The risk value of a bolt soars from 0.3 to 0.8 within 10 minutes, corresponding to a slope of 0.05 / minute. However, due to local fluctuations within the window, an instantaneous slope of 0.6 may be detected (for example, from 10:05 to 10:07), which is marked as a sudden change segment.

[0247] Misjudgment protection: Short-term noise (such as a single-point slope of 0.6) will not be misjudged as a sudden change because it does not meet the three-point continuity condition.

[0248] Perform dynamic time warping to align the current risk curve with the historical high-risk and medium-risk template curves, calculate the minimum cumulative distance, and record the number of path bends as a similarity measure;

[0249] Dynamic Time Warping (DTW) is an algorithm that elastically aligns two time series to measure their shape similarity. In this step, the current risk curve is aligned with the pre-stored historical templates (high-risk and medium-risk typical curves) through DTW:

[0250] Template construction:

[0251] High-risk template: Extract 10 curves from historical data whose risk values ​​quickly rise to above 0.9, and take the average to generate the template;

[0252] Medium-risk template: Extract 20 curves with risk values ​​fluctuating between 0.5 and 0.7 to generate a template.

[0253] DTW alignment:

[0254] Cumulative distance matrix: Constructs the point-to-point distance matrix between the current curve and the template curve, using Euclidean distance as the distance metric;

[0255] Path search: Find the optimal path from the upper left corner to the lower right corner of the matrix, minimizing the sum of the distances to each point on the path (cumulative distance). The path allows diagonal, horizontal, or vertical movement, but is limited to a slope of ≤ 2 (to avoid excessive distortion).

[0256] Feature extraction:

[0257] Minimum cumulative distance: the total distance of the optimal path. The smaller the value, the more similar it is to the template.

[0258] Path bend count: The number of times the path direction changes (e.g., from horizontal to diagonal), reflecting the complexity of the alignment.

[0259] If the minimum cumulative distance is less than the high-risk threshold and the number of path bends is ≥2, the risk level is determined to be high. If the minimum cumulative distance is greater than or equal to the high-risk threshold and less than the medium-risk threshold, or if the minimum cumulative distance is less than the high-risk threshold and the number of path bends is <2, the risk level is determined to be medium. Otherwise, if the minimum cumulative distance is greater than or equal to the medium-risk threshold, the risk level is determined to be low.

[0260] Risk classification relies on predefined thresholds and path characteristics:

[0261] Threshold setting: High-risk threshold: Through historical data analysis, the cumulative distance threshold is set to 150 (i.e., the similarity with the high-risk template is high); Medium-risk threshold: set to 250, which is higher than the high-risk threshold but lower than the low-risk judgment line.

[0262] Judgment rules: High risk: cumulative distance < 150 and the number of path bends ≥ 2 (complex alignment paths reflect mutation patterns); Medium risk: 150 ≤ cumulative distance < 250 or cumulative distance < 150 but the number of bends < 2; Low risk: cumulative distance ≥ 250.

[0263] Dynamic Adjustment: The threshold can be adjusted by ±10% based on the equipment type and environmental factors (such as vibration intensity). For example, the threshold for heavy machinery is adjusted higher to 160 to tolerate a higher baseline risk.

[0264] Application examples: Case 1: The cumulative distance between the current curve of a wind turbine bolt and the high-risk template is 130, and the number of bends is 3 → high risk; Case 2: The cumulative distance between a bridge bolt and the high-risk template is 140, and the number of bends is 1 → medium risk (due to the simple path, it may be slowly deteriorating); Case 3: The cumulative distance between a machine tool bolt and the high-risk template is 280 → low risk.

[0265] Based on the graph attention mechanism, a dynamic relationship map between bolts and connectors is constructed. The multi-scale time series features of historical data (such as hourly fluctuations and daily trends) are extracted through a temporal convolutional network. Combined with LSTM, the loosening evolution probability curve for the next 24 hours is predicted. Finally, the risk level is divided into low, medium, and high levels based on the slope of the curve and the intersection of the threshold. This breaks through the limitations of single-point monitoring and predicts the loosening propagation path (such as the chain failure of adjacent bolts) from a system level. The risk level classification makes maintenance decisions more forward-looking (for example, high-risk requires immediate shutdown).

[0266] S204, based on the looseness risk level and combined with the equipment operating status parameters, an adaptive decision tree model is used to dynamically adjust the monitoring frequency and alarm threshold, generate a real-time monitoring strategy and a visual early warning report, and simultaneously trigger the wireless transmission module to upload to the cloud operation and maintenance platform.

[0267] Specifically, the equipment operating speed, load current parameters and looseness risk level can be normalized to generate a multi-dimensional state feature vector;

[0268] Equipment operating parameters (such as speed and load current) and loosening risk levels usually have different dimensions and value ranges. The purpose of normalization is to eliminate dimensional differences and make all parameters participate in model calculations at the same scale. The specific process is as follows:

[0269] Data collection and preprocessing:

[0270] Speed ​​(unit: RPM): collected by Hall sensor, range 0~3000 RPM, typical value is 1500 RPM;

[0271] Load current (unit: A): measured by current transformer, range 0~100 A, typical value is 45 A;

[0272] Loosening risk level: low, medium, and high risk levels output by the previous steps (coded as 0, 1, and 2).

[0273] Normalization method:

[0274] Maximum and minimum normalization: linearly map each parameter to the [0,1] interval. For example, the speed normalization formula is: If the actual speed is 1500 RPM, the normalized value is 0.5;

[0275] Categorical variable encoding: Risk level uses one-hot encoding, with low risk being [1,0,0], medium risk being [0,1,0], and high risk being [0,0,1].

[0276] Feature vector construction: Concatenate the normalized parameters into a multidimensional vector. For example, when the device speed is 1500 RPM (normalized to 0.5), the load current is 45 A (normalized to 0.45), and the risk level is medium (coded [0, 1, 0]), the generated feature vector is: [0.5, 0.45, 0, 1, 0].

[0277] Technical Example: In a wind turbine bolt monitoring scenario, with a speed of 1800 RPM (normalized to 0.6), a load current of 60 A (normalized to 0.6), and a high risk level (coded [0, 0, 1]), the generated feature vector is [0.6, 0.6, 0, 0, 1]. This vector serves as input to an adaptive decision tree model.

[0278] Adopting incremental learning method to train adaptive decision tree model, dynamically splitting nodes according to multi-dimensional state feature vectors, and generating monitoring frequency adjustment rules and alarm thresholds;

[0279] The core of the adaptive decision tree model is to dynamically adjust the tree structure to adapt to real-time data changes. Its training process is divided into two stages: initial construction and incremental update:

[0280] Initial model building:

[0281] Dataset: Use historical data (e.g., 1000 samples) to train an initial decision tree, using the C4.5 algorithm to select split features. For example, the first split might select "speed" as the root node, with a threshold of 0.5 (corresponding to 1500 RPM).

[0282] Node splitting rule: Based on maximizing the information gain rate, for example, when the speed is greater than 0.5, it enters the right subtree (high-risk working condition), otherwise it enters the left subtree (normal working condition).

[0283] Incremental learning mechanism:

[0284] Sliding window update: retain the most recent 200 sets of data, eliminate the oldest data when new data arrives, and keep the window size constant;

[0285] Dynamic split strategy:

[0286] Feature importance assessment: Dynamically calculate the importance of each feature using the Gini index. For example, if the Gini index of the load current increases from 0.3 to 0.6, it will be prioritized as the split node;

[0287] Node reorganization: When the classification error rate of a node continues to exceed 15%, local reconstruction is triggered, for example, replacing the original "speed" split node with "load current".

[0288] Rule and threshold generation:

[0289] Monitoring frequency adjustment: Dynamically set the sampling interval based on the risk level. For example:

[0290] Low risk: collected every 10 minutes;

[0291] Medium risk: collected every 5 minutes;

[0292] High risk: real-time collection (once per second);

[0293] Alarm threshold setting: Combined with equipment operating conditions, for example:

[0294] When the speed is >0.7 (2100 RPM) and the risk level is high, a level 1 alarm is triggered;

[0295] When the load current is >0.8 (80 A) and the risk level is medium, a level 2 alarm is triggered.

[0296] Technical example: In monitoring joint bolts on an industrial robot, the initial decision tree partitions the strategy based on speed and load current. When a new load pattern is introduced (e.g., frequent starts and stops causing increased current fluctuations), incremental learning automatically adds the "load current change rate" as a new partitioning feature and adjusts the monitoring frequency to every two minutes.

[0297] ‌Explore the optimal strategy in the decision tree space based on the Monte Carlo tree search algorithm, and output a real-time monitoring strategy table including monitoring interval, alarm priority, and data sampling accuracy;

[0298] Monte Carlo Tree Search (MCTS) uses a simulation-evaluation-backtracking mechanism to find the optimal strategy in the decision tree space. The specific steps are as follows:

[0299] Decision Tree Spatial Modeling:

[0300] Node definition: Each node represents a monitoring strategy combination, for example, {monitoring interval: 5min, alarm priority: 2, sampling accuracy: High};

[0301] Edge definition: An edge represents an adjustment action for a policy parameter, such as "increase monitoring frequency" or "improve sampling accuracy."

[0302] MCTS four-stage operation:

[0303] Selection: Starting from the root node, select child nodes based on the UCB1 formula (Upper Confidence Bound). The UCB1 formula is: .in, is the revenue of the i-th node, is the number of visits to the i-th node, N is the total number of visits to the parent node, and c=2 is the exploration factor;

[0304] Expansion: When encountering an unexplored node, expand the new child node. For example, add a new policy {monitoring interval: 3 minutes, alarm priority: 1};

[0305] Simulation: Randomly simulate the results of strategy execution, such as predicting the number of alarms and energy consumption in the next hour under the strategy;

[0306] Backpropagation: Update the path nodes in reverse order using the simulation results (e.g., profit = 0.8).

[0307] Strategy evaluation and screening:

[0308] Benefit function design: Comprehensive monitoring effectiveness (such as alarm accuracy) and resource consumption (such as energy consumption) are combined. The formula is: benefit = 0.7 × accuracy + 0.3 × (1 − energy consumption ratio).

[0309] Pareto front extraction: Screening non-dominated solutions (i.e., it is impossible to improve one objective without compromising other objectives), for example:

[0310] Strategy A: 90% accuracy, 30% energy consumption;

[0311] Strategy B: accuracy 85%, energy consumption 20%.

[0312] Technical Examples:

[0313] In the wind power bolt monitoring scenario, MCTS outputs the optimal strategy table after 1000 simulations, as shown in Table 1:

[0314] Table 1

[0315]

[0316] ‌Convert the real-time monitoring strategy table into a JSON format instruction set, drive the visualization engine to generate a three-dimensional dynamic warning map, and upload it to the cloud operation and maintenance platform through the LoRa wireless module.

[0317] JSON instruction set generation:

[0318] Data structure design:

[0319] json

[0320] {

[0321] "device_id": "Bolt_001",

[0322] "monitoring_strategy": {

[0323] "interval": 300, / / Unit: seconds

[0324] "alert_priority": 1,

[0325] "sampling_precision": "High"

[0326] },

[0327] "timestamp": "2023-10-05T14:30:00Z"

[0328] }.

[0329] Dynamic Population: Automatically populate fields based on the policy table contents. For example, when the policy "3-minute monitoring interval" is selected, the interval field is set to 180 seconds.

[0330] Generation of three-dimensional dynamic warning maps:

[0331] Visualization engine: Uses the Three.js library to build WebGL 3D scenes. Key components include:

[0332] Bolt model: Import CAD model and color it according to health indicators (green = normal, red = high risk);

[0333] Stress distribution heat map: maps strain gauge data to frequency bands: 868 MHz (Europe) / 915 MHz (North America);

[0334] Transmit power: 14 dBm;

[0335] Encryption protocol: AES-256, key rotation every 24 hours.

[0336] Data encapsulation: Convert the JSON command into binary format, add a frame header (device ID), a checksum (CRC32), and a frame trailer. For example, the transmission frame structure of the command Bolt_001 is: [0xAA][Device ID 4B][Data length 2B][Encrypted data NB][CRC32 4B][0x55].

[0337] Technical example: When a chemical plant's reactor bolt monitoring system detects a high risk, it generates a JSON command and triggers the following actions: the local HMI interface displays a flashing red 3D bolt model, and a superimposed stress thermogram shows a localized overrun. The LoRa module encrypts and uploads data to the cloud, where the cloud platform automatically generates a work order and pushes it to the maintenance personnel's mobile device. The maintenance personnel use AR glasses to verify the precise positioning of the bolts and quickly perform tightening operations.

[0338] Based on the correlation between the real-time speed, load and other parameters of the equipment and the risk level, the monitoring strategy is dynamically optimized through an incremental decision tree, and an early warning report containing a three-dimensional dynamic map is generated. Low-power remote monitoring is achieved through LoRa wireless transmission, and intelligent allocation of monitoring resources is realized (reducing the monitoring overhead of low-risk bolts). Visual reports intuitively display the spatial distribution of risks, and cloud collaboration supports multi-device cluster management.

[0339] It can be seen that according to the multi-dimensional vibration signals and stress wave propagation characteristics of the bolt connection parts, the micro-motion characteristic matrix of the bolt is generated; based on the micro-motion characteristic matrix, the local stress distribution data collected by the strain gauges on the bolt surface are integrated to output the quantitative health status index of the bolt; according to the quantitative health status index and the historical loosening evolution data, the loosening risk level is output; based on the loosening risk level and combined with the equipment operating status parameters, the adaptive decision tree model is used to dynamically adjust the monitoring frequency and alarm threshold, generate a real-time monitoring strategy and a visual early warning report, and simultaneously trigger the wireless transmission module to upload to the cloud operation and maintenance platform, thereby realizing high-precision, adaptive and traceable bolt health status assessment and early warning.

[0340] Another embodiment of the present invention provides a bolt online monitoring system, see Figure 3 , the system may include:

[0341] Extraction module 301 is used to extract the characteristic frequency band of bolt loosening based on the multi-dimensional vibration signal and stress wave propagation characteristics of the bolt connection part through wavelet packet decomposition algorithm, and generate the micro-vibration characteristic matrix of the bolt by combining the time delay inversion calculation of the stress wave propagation path;

[0342] A construction module 302 is configured to integrate the local stress distribution data collected by the bolt surface strain gauge based on the micro-vibration feature matrix, construct a multimodal feature vector using a deep convolutional neural network, associate the spatial correlation between the vibration features and the stress distribution through a cross-modal attention mechanism, and output a quantitative indicator of the bolt health status;

[0343] Prediction module 303, configured to establish a bolt-connector dynamic relationship graph using a graph neural network based on the health status quantitative indicators and historical loosening evolution data, predict the loosening trend using a temporal convolutional network, and output a loosening risk level;

[0344] The generation module 304 is used to dynamically adjust the monitoring frequency and alarm threshold based on the looseness risk level and the equipment operating status parameters using an adaptive decision tree model, generate a real-time monitoring strategy and a visual early warning report, and simultaneously trigger the wireless transmission module to upload to the cloud operation and maintenance platform.

[0345] 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.

[0346] Specifically, in this embodiment, the above-mentioned storage medium may be configured to store a computer program for performing the following steps:

[0347] S201, based on the multi-dimensional vibration signal and stress wave propagation characteristics of the bolt connection part, the characteristic frequency band of bolt loosening is extracted by wavelet packet decomposition algorithm, and combined with the time delay inversion calculation of the stress wave propagation path, the micro-motion characteristic matrix of the bolt is generated;

[0348] S202: Based on the micro-vibration feature matrix, the local stress distribution data collected by the bolt surface strain gauge is integrated, and a multimodal feature vector is constructed using a deep convolutional neural network. The spatial correlation between the vibration characteristics and the stress distribution is correlated through a cross-modal attention mechanism to output a quantitative indicator of the bolt health status;

[0349] S203, based on the health status quantitative indicators and historical loosening evolution data, using a graph neural network to establish a dynamic relationship map between bolts and connectors, predicting the loosening trend through a temporal convolutional network, and outputting a loosening risk level;

[0350] S204, based on the looseness risk level and combined with the equipment operating status parameters, an adaptive decision tree model is used to dynamically adjust the monitoring frequency and alarm threshold, generate a real-time monitoring strategy and a visual early warning report, and simultaneously trigger the wireless transmission module to upload to the cloud operation and maintenance platform.

[0351] 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.

[0352] 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.

[0353] Specifically, in this embodiment, the processor may be configured to execute the following steps through a computer program:

[0354] S201, based on the multi-dimensional vibration signal and stress wave propagation characteristics of the bolt connection part, the characteristic frequency band of bolt loosening is extracted by wavelet packet decomposition algorithm, and combined with the time delay inversion calculation of the stress wave propagation path, the micro-motion characteristic matrix of the bolt is generated;

[0355] S202: Based on the micro-vibration feature matrix, the local stress distribution data collected by the bolt surface strain gauge is integrated, and a multimodal feature vector is constructed using a deep convolutional neural network. The spatial correlation between the vibration characteristics and the stress distribution is correlated through a cross-modal attention mechanism to output a quantitative indicator of the bolt health status;

[0356] S203, based on the health status quantitative indicators and historical loosening evolution data, using a graph neural network to establish a dynamic relationship map between bolts and connectors, predicting the loosening trend through a temporal convolutional network, and outputting a loosening risk level;

[0357] S204, based on the looseness risk level and combined with the equipment operating status parameters, an adaptive decision tree model is used to dynamically adjust the monitoring frequency and alarm threshold, generate a real-time monitoring strategy and a visual early warning report, and simultaneously trigger the wireless transmission module to upload to the cloud operation and maintenance platform.

[0358] 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 bolt online monitoring method, characterized in that: The method comprises: Based on the multi-dimensional vibration signals and stress wave propagation characteristics of the bolt connection, the wavelet packet decomposition algorithm is used to extract the characteristic frequency bands of bolt loosening. Combined with the time delay inversion calculation of the stress wave propagation path, the micro-motion characteristic matrix of the bolt is generated. Based on the micro-vibration feature matrix, the local stress distribution data collected by the bolt surface strain gauges are integrated, and a deep convolutional neural network is used to construct a multimodal feature vector. The spatial correlation between the vibration characteristics and the stress distribution is associated with the cross-modal attention mechanism to output a quantitative indicator of the bolt health status. Based on the health status quantitative indicators and historical loosening evolution data, a graph neural network is used to establish a dynamic relationship map between bolts and connectors, and a temporal convolutional network is used to predict the loosening trend and output the loosening risk level; Based on the looseness risk level and combined with the equipment operating status parameters, an adaptive decision tree model is used to dynamically adjust the monitoring frequency and alarm threshold, generate real-time monitoring strategies and visual early warning reports, and simultaneously trigger the wireless transmission module to upload to the cloud operation and maintenance platform.

2. The method according to claim 1, characterized in that The method extracts the characteristic frequency band of bolt loosening based on the multi-dimensional vibration signal and stress wave propagation characteristics of the bolt connection part through the wavelet packet decomposition algorithm, and generates the micro-motion characteristic matrix of the bolt by combining the time delay inversion calculation of the stress wave propagation path, including: Based on the vibration signals collected by the piezoelectric sensor array, the empirical mode decomposition algorithm is used to separate high-frequency noise and low-frequency structural vibration components to generate a denoised multi-band vibration signal. Perform 8-layer wavelet packet decomposition on the multi-band vibration signal, calculate the energy entropy value of each sub-band, and select the frequency band with energy entropy mutation within ±15% of the threshold as the loosening characteristic frequency band; ‌Distributed optical fiber sensors are used to detect the stress wave propagation path, and the cross-correlation algorithm is used to calculate the time delay difference between adjacent sensors. The micro-displacement fluctuation trajectory of the bolt connection interface is then inverted. The energy spectrum of the loosening characteristic frequency band is aligned with the micro-displacement fluctuation trajectory according to the time window, and a three-dimensional micro-motion characteristic matrix is ​​generated by tensor splicing. The matrix dimensions include three characteristic axes: frequency, displacement and time series.

3. The method according to claim 2, characterized in that Based on the micro-vibration feature matrix, the local stress distribution data collected by the bolt surface strain gauge is integrated, and a deep convolutional neural network is used to construct a multimodal feature vector. The spatial correlation between the vibration characteristics and the stress distribution is associated with the cross-modal attention mechanism to output the quantitative health status indicators of the bolt, including: Based on the local stress distribution data collected by the strain gauge array, a stress gradient thermodynamic map of the bolt surface is generated and mapped into a 128×128 pixel stress field image using a bilinear interpolation algorithm. The stress field image is input into the ResNet-50 network to extract high-order features, and the micro-motion feature matrix is ​​input into the 3D-CNN network to extract spatiotemporal features to generate a bimodal feature vector. ‌The association weights of vibration features and stress features are calculated through the gated attention mechanism, and the dual-modal feature vectors are dynamically weighted fused to generate a joint feature vector; A graph convolutional network is used to establish the spatial relationship graph between the bolt connection structure and adjacent components. The joint feature vector is used as the node feature for graph embedding learning to obtain the graph embedding vector. The graph embedding vector is mapped to a scalar value of 0-1 through a fully connected layer, and a quantitative health status indicator representing the degree of bolt loosening is output.

4. The method according to claim 3, characterized in that The method uses a graph neural network to establish a dynamic relationship graph between bolts and connectors based on the health status quantitative indicators and historical loosening evolution data, predicts the loosening trend through a time convolutional network, and outputs the loosening risk level, including: Bolts and connectors are used as graph nodes, and the mechanical coupling strength between nodes is calculated based on health indicators to generate the adjacency matrix of the initial dynamic relationship graph. ‌A temporal convolutional network is used to extract multi-scale features from historical loose evolution data, generating a temporal encoding vector of the node state at each time step; ‌ Aggregate the state information of adjacent nodes through the graph attention mechanism, and iteratively update the node implicit state in combination with the temporal encoding vector to capture the loose propagation path; Input the updated node status into the long short-term memory network, predict the loose evolution trend in the next 24 hours, and output the probabilistic risk curve; Based on the slope change of the probabilistic risk curve and the intersection point of the threshold, the dynamic time warping algorithm is used to classify the loosening risk level of the bolts into low risk, medium risk or high risk levels.

5. The method according to claim 4, characterized in that Based on the looseness risk level and combined with the equipment operating status parameters, the adaptive decision tree model is used to dynamically adjust the monitoring frequency and alarm threshold, generate a real-time monitoring strategy and a visual early warning report, and synchronously trigger the wireless transmission module to upload to the cloud operation and maintenance platform, including: Normalize the equipment operating speed, load current parameters and looseness risk level to generate a multi-dimensional state feature vector; Adopting incremental learning method to train adaptive decision tree model, dynamically splitting nodes according to multi-dimensional state feature vectors, and generating monitoring frequency adjustment rules and alarm thresholds; ‌Explore the optimal strategy in the decision tree space based on the Monte Carlo tree search algorithm, and output a real-time monitoring strategy table including monitoring interval, alarm priority, and data sampling accuracy; ‌Convert the real-time monitoring strategy table into a JSON format instruction set, drive the visualization engine to generate a three-dimensional dynamic warning map, and upload it to the cloud operation and maintenance platform through the LoRa wireless module.

6. The method according to claim 3, characterized in that The graph convolutional network is used to establish a spatial relationship graph between the bolt connection structure and adjacent components, and the joint feature vector is used as a node feature for graph embedding learning to obtain a graph embedding vector, including: The joint eigenvector is normalized to the maximum and minimum according to the modal dimension to eliminate the dimensional difference and obtain the standardized node eigenvector; Based on the standardized node feature vector, a two-layer graph convolutional network is used. The first layer aggregates the features of directly adjacent nodes, and the second layer aggregates the features of second-order neighboring nodes. The gated attention mechanism is used to dynamically adjust the aggregation weights to generate an implicit state vector representing the spatial relationship between the bolted connection structure and adjacent components. ‌Graph pooling is performed on the latent state vector, and an adaptive mean pooling algorithm is used to fuse node features to output a graph embedding vector that represents the global structural characteristics. The vector dimension is compressed to 256 dimensions.

7. The method according to claim 4, characterized in that The dynamic time warping algorithm is used to classify the loosening risk level of the bolt into low risk, medium risk or high risk levels based on the slope change of the probabilistic risk curve and the intersection point of the threshold, including: A Savitzky-Golay filter is applied to the probabilistic risk curve with a window length of 15 and a polynomial order of 3 to remove high-frequency noise and preserve trend characteristics; ‌Calculate the instantaneous slope of each point on the curve based on the sliding window, and mark the area where the absolute value of the slope exceeds 0.5 and lasts for more than 3 sampling points as a sudden change segment; Perform dynamic time warping to align the current risk curve with the historical high-risk and medium-risk template curves, calculate the minimum cumulative distance, and record the number of path bends as a similarity measure; If the minimum cumulative distance is less than the high-risk threshold and the number of path bends is ≥2, the risk level is determined to be high. If the minimum cumulative distance is greater than or equal to the high-risk threshold and less than the medium-risk threshold, or if the minimum cumulative distance is less than the high-risk threshold and the number of path bends is <2, the risk level is determined to be medium. Otherwise, if the minimum cumulative distance is greater than or equal to the medium-risk threshold, the risk level is determined to be low.

8. A bolt online monitoring system, characterized in that: The system comprises: The extraction module is used to extract the characteristic frequency bands of bolt loosening based on the multi-dimensional vibration signals and stress wave propagation characteristics of the bolt connection part through the wavelet packet decomposition algorithm. Combined with the time delay inversion calculation of the stress wave propagation path, it generates the micro-motion characteristic matrix of the bolt; A construction module is used to integrate the local stress distribution data collected by the bolt surface strain gauge based on the micro-vibration feature matrix, construct a multimodal feature vector using a deep convolutional neural network, associate the spatial correlation between the vibration characteristics and the stress distribution through a cross-modal attention mechanism, and output a quantitative indicator of the bolt health status; A prediction module is used to establish a dynamic relationship map between bolts and connectors using a graph neural network based on the health status quantitative indicators and historical loosening evolution data, predict the loosening trend through a temporal convolutional network, and output a loosening risk level; A generation module is used to dynamically adjust the monitoring frequency and alarm threshold based on the loosening risk level and combined with the equipment operation status parameters using an adaptive decision tree model, generate a real-time monitoring strategy and a visual early warning report, and simultaneously trigger the wireless transmission module to upload to the cloud operation and maintenance platform.

9. 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 7 when run.

10. 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 7.

Images