Bolt online monitoring method and system

Abstract

The invention discloses a bolt on-line monitoring method and system, and the method comprises the steps: generating a micro-motion feature matrix of a bolt according to a multi-dimensional vibration signal and a stress wave propagation characteristic of a bolt connection part; based on the micro-motion characteristic matrix, fusing local stress distribution data acquired by a bolt surface strain gauge, and outputting a health state quantitative index of the bolt; outputting a loosening risk level according to the health state quantitative index and historical loosening evolution data; and based on the looseness risk level and in combination with equipment operation state parameters, a self-adaptive decision tree model is adopted to dynamically adjust monitoring frequency and an alarm threshold value, a real-time monitoring strategy and a visual early warning report are generated, and a wireless transmission module is synchronously triggered to upload the real-time monitoring strategy and the visual early warning report to a cloud operation and maintenance platform. According to the embodiment of the invention, high-precision, self-adaptive and traceable bolt health state evaluation and early warning can be realized.

Description

Technical Field

[0001] The present invention belongs to the technical field of monitoring, and particularly relates to an online bolt monitoring method and system. Background Art

[0002] In the fields of mechanical equipment, bridge engineering, aerospace, etc., the health status of bolt connection structures directly affects the safety and reliability of the overall equipment. Traditional bolt loosening detection mainly relies on manual inspection or off-line measurement, which has problems such as low efficiency and poor real-time performance, and is difficult to meet the intelligent monitoring requirements of modern industry. 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, it is difficult for a single sensor modality to comprehensively characterize the bolt loosening state, resulting in insufficient detection accuracy; second, there is a lack of multi-modal feature fusion analysis of vibration signals and stress distributions, and the microscopic mechanical behavior of the bolt connection interface cannot be accurately reflected; third, existing methods mostly use static threshold alarms and cannot dynamically adjust the monitoring strategy according to the equipment operation state, which is prone to false alarms or missed alarms; fourth, the management and traceability mechanism of monitoring data is imperfect, and it is difficult to support large-scale distributed deployment. In addition, traditional methods have weak prediction ability for bolt loosening trends and cannot early warn of potential risks. Summary of the Invention

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

[0004] An embodiment of the present application provides an online bolt monitoring method, and the method includes: According to the multi-dimensional vibration signals and stress wave propagation characteristics of the bolt connection part, the wavelet packet decomposition algorithm is used to extract the bolt loosening characteristic frequency band, and combined with the time-delay inversion calculation of the stress wave propagation path, a micro-motion characteristic matrix of the bolt is generated; Based on the micro-motion characteristic matrix, by fusing the local stress distribution data collected by the strain gauges on the bolt surface, a deep convolutional neural network is used to construct a multi-modal feature vector, and the cross-modal attention mechanism is used to associate the spatial correlation between vibration characteristics and stress distribution, and a health status quantification index of the bolt is output; According to the health status quantification index and historical loosening evolution data, a graph neural network is used to establish a bolt-connector dynamic relationship graph, and the time convolution network is used to predict the loosening trend, and the loosening risk level is output; Based on the loosening risk level, combined with the equipment operation state 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 the wireless transmission module to upload to the cloud operation and maintenance platform.

[0005] Optionally, based on the multi-dimensional vibration signals and stress wave propagation characteristics of the bolt connection part, the wavelet packet decomposition algorithm is used to extract the characteristic frequency band of bolt loosening, and combined with the time-delay inversion calculation of the stress wave propagation path, a fretting characteristic matrix of the bolt is generated, including: According to the vibration signals collected by the piezoelectric sensor array, the empirical mode decomposition algorithm is used to separate the high-frequency noise and low-frequency structural vibration components, and a denoised multi-band vibration signal is generated; Perform 8-layer wavelet packet decomposition on the multi-band vibration signal, calculate the energy entropy values of each sub-frequency band, and select the frequency bands where the energy entropy mutation is within the threshold of ±15% as the loosening characteristic frequency bands; Based on the distributed fiber optic sensor to detect the stress wave propagation path, the cross-correlation algorithm is used to calculate the time-delay difference between adjacent sensors, and the micro-displacement fluctuation trajectory of the bolt connection interface is inversely obtained; Align the energy spectrum of the loosening characteristic frequency band with the micro-displacement fluctuation trajectory according to the time window, and generate a three-dimensional fretting characteristic matrix through tensor splicing. The matrix dimension includes three characteristic axes: frequency, displacement, and time sequence.

[0006] Optionally, based on the fretting characteristic matrix, fuse the local stress distribution data collected by the strain gauges on the bolt surface, use a deep convolutional neural network to construct a multi-modal feature vector, and associate the spatial correlation between the vibration characteristics and the stress distribution through a cross-modal attention mechanism, and output a health state quantification index of the bolt, including: According to the local stress distribution data collected by the strain gauge array, generate a stress gradient heat map on the bolt surface, and map it to a stress field image of 128×128 pixels through bilinear interpolation; Input the stress field image into the ResNet-50 network to extract high-order features, and input the fretting characteristic matrix into the 3D-CNN network to extract spatio-temporal features, and generate a bimodal feature vector; Calculate the correlation weight between the vibration characteristics and the stress characteristics through a gated attention mechanism, and perform dynamic weighted fusion on the bimodal feature vector to generate a joint feature vector; Use a graph convolutional network to establish a spatial relationship graph between the bolt connection structure and adjacent components, and use the joint feature vector as the node feature for graph embedding learning to obtain a graph embedding vector; Map the graph embedding vector to a scalar value of 0-1 through a fully connected layer, and output a health state quantification index representing the degree of bolt loosening.

[0007] Optionally, based on the health state quantification index and historical loosening evolution data, use a graph neural network to establish a dynamic relationship graph of the bolt-connector, and predict the loosening trend through a temporal convolutional network, and output the loosening risk level, including: Taking the bolt and the connecting piece as graph nodes, calculating the mechanical coupling strength between nodes based on health indicators, and generating the adjacency matrix of the initial dynamic relationship graph; Using a temporal convolutional network to perform multi-scale feature extraction on historical loosening evolution data, and generating a temporal encoding vector of the node state at each time step; Aggregating the state information of adjacent nodes through a graph attention mechanism, and iteratively updating the node hidden state in combination with the temporal encoding vector to capture the loosening propagation path; Inputting the updated node state into a long short-term memory network, predicting the loosening evolution trend in the next 24 hours, and outputting a probabilistic risk curve; Based on the slope change and threshold crossing point of the probabilistic risk curve, using the dynamic time warping algorithm to divide the loosening risk level of the bolt into low risk, medium risk or high risk levels.

[0008] Optionally, based on the loosening risk level, combined with the device operation state parameters, using an adaptive decision tree model to dynamically adjust the monitoring frequency and alarm threshold, generating a real-time monitoring strategy and a visual warning report, and synchronously triggering the wireless transmission module to upload to the cloud operation and maintenance platform, including: Normalizing the device operation speed, load current parameters and loosening risk level to generate a multi-dimensional state feature vector; Using an incremental learning method to train an adaptive decision tree model, dynamically splitting nodes according to the multi-dimensional state feature vector, and generating a monitoring frequency adjustment rule and an alarm threshold; Based on the Monte Carlo tree search algorithm to explore the optimal strategy in the decision tree space, and outputting a real-time monitoring strategy table including the monitoring interval time, alarm priority and data sampling accuracy; Converting the real-time monitoring strategy table into a JSON format instruction set, driving the visualization engine to generate a three-dimensional dynamic warning graph, and at the same time encrypting and uploading it to the cloud operation and maintenance platform through the LoRa wireless module.

[0009] Optionally, using a graph convolutional network to establish a spatial relationship graph between the bolt connection structure and adjacent components, and performing graph embedding learning with the joint feature vector as the node feature to obtain a graph embedding vector, including: Performing maximum-minimum normalization processing on the joint feature vector according to the modal dimension to eliminate the dimension difference and obtain a standardized node feature vector; Based on the standardized node feature vector, through a two-layer graph convolutional network, the first layer aggregates the features of directly adjacent nodes, the second layer aggregates the features of second-order neighborhood nodes, and combines the gated attention mechanism to dynamically adjust the aggregation weight to generate a hidden state vector representing the spatial relationship between the bolt connection structure and adjacent components; Perform graph pooling operation on the hidden state vector, adopt the adaptive mean pooling algorithm to fuse node features, and output a graph embedding vector representing the global structural characteristics, with the vector dimension compressed to 256 dimensions.

[0010] Optionally, based on the slope change and threshold crossing point of the probabilistic risk curve, use the dynamic time warping algorithm to divide the loosening risk level of the bolt into low risk, medium risk or high risk levels, including: Apply the Savitzky-Golay filter to the probabilistic risk curve, where the window length is 15 and the polynomial order is 3, to eliminate high-frequency noise and retain trend features; Calculate the instantaneous slope of each point on the curve based on a 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 the mutation segment; Align the current risk curve with the historical high-risk and medium-risk template curves through dynamic time warping, 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 ≥ 2, it is determined as the high-risk level. If the minimum cumulative distance is greater than or equal to the high-risk threshold and less than the medium-risk threshold, or the minimum cumulative distance is less than the high-risk threshold and the number of path bends < 2, then it is determined as the medium-risk level. Otherwise, if the minimum cumulative distance is greater than or equal to the medium-risk threshold, it is judged as the low-risk level.

[0011] Another embodiment of the present application provides a bolt online monitoring system, and the system includes: An extraction module, configured to extract the bolt loosening characteristic frequency band through the wavelet packet decomposition algorithm according to the multi-dimensional vibration signal and stress wave propagation characteristics of the bolt connection part, and generate a bolt micro-motion characteristic matrix in combination with the time-delay inversion calculation of the stress wave propagation path; A construction module, configured to fuse the local stress distribution data collected by the bolt surface strain gauge based on the micro-motion characteristic matrix, construct a multi-modal feature vector using a deep convolutional neural network, and associate the spatial correlation between the vibration characteristics and the stress distribution through a cross-modal attention mechanism, and output a quantitative index of the bolt health state; A prediction module, configured to establish a bolt-connector dynamic relationship graph using a graph neural network according to the health state quantitative index and historical loosening evolution data, predict the loosening trend through a temporal convolutional network, and output the loosening risk level; A generation module, configured to dynamically adjust the monitoring frequency and alarm threshold based on the loosening risk level and combine the device operation state parameters, adopt an adaptive decision tree model to generate a real-time monitoring strategy and a visual warning report, and synchronously trigger the wireless transmission module to upload to the cloud operation and maintenance platform.

[0012] Another embodiment of the present application provides a storage medium, in which a computer program is stored, and the computer program is configured to execute the method described in any one of the above when running.

[0013] Another embodiment of the present application provides an electronic device, including a memory and a processor. A computer program is stored in the memory, and the processor is configured to run the computer program to execute the method described in any one of the above.

[0014] Compared with the prior art, a bolt online monitoring method provided by the present invention generates a fretting feature matrix of a bolt according to multi-dimensional vibration signals and stress wave propagation characteristics at the bolt connection part; based on the fretting feature matrix, fuses local stress distribution data collected by a strain gauge on the bolt surface, and outputs a quantitative index of the bolt's health state; according to the quantitative index of the health state and historical loosening evolution data, outputs a loosening risk level; based on the loosening risk level, combines device operation state parameters, and uses an adaptive decision tree model to dynamically adjust the monitoring frequency and alarm threshold, generates a real-time monitoring strategy and a visual warning report, and synchronously triggers a wireless transmission module to upload to a cloud operation and maintenance platform, so as to realize high-precision, adaptive, and traceable bolt health state assessment and warning. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 It is a hardware structure block diagram of a computer terminal for a bolt online monitoring method provided by an embodiment of the present invention; Figure 2 It is a schematic flowchart of a bolt online monitoring method provided by an embodiment of the present invention; Figure 3 It is a schematic structural diagram of a bolt online monitoring system provided by an embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0016] The embodiments described below by referring to the accompanying drawings are exemplary and are only used to explain the present invention, and cannot be construed as limiting the present invention.

[0017] An embodiment of the present invention first provides a bolt online monitoring method, which can be applied to an electronic device, such as a computer terminal, specifically, a general computer, etc.

[0018] The following takes running on a computer terminal as an example to describe it in detail. Figure 1 It is a hardware structure block diagram of a computer terminal for a bolt online monitoring method provided by an embodiment of the present invention. As Figure 1 shown, the computer device includes a processor, a memory, and a network interface connected through a system bus. Among them, the memory may include a non-volatile storage medium and an internal memory.

[0019] The non-volatile storage medium can store an operating system and a computer program. The computer program includes program instructions which, when executed, can cause the processor to execute any one of the bolt online monitoring methods.

[0020] The processor is used to provide computing and control capabilities to support the operation of the entire computer device.

[0021] 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, it can cause the processor to execute any one of the bolt online monitoring methods.

[0022] The network interface is used for network communication, such as sending assigned tasks, etc. Those skilled in the art can understand that Figure 1 the structure shown in

[0023] is only a block diagram of some structures related to the solution of this application, and does not constitute a limitation on the computer device to which the solution of this application is applied. The specific computer device may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0024] See Figure 2 , the embodiments of the present invention provide a bolt online monitoring method, which may include the following steps: S201, according to the multi-dimensional vibration signals and stress wave propagation characteristics of the bolt connection part, extract the bolt loosening characteristic frequency band through the wavelet packet decomposition algorithm, and combine the time-delay inversion calculation of the stress wave propagation path to generate the micro-motion characteristic matrix of the bolt; Specifically, the vibration signals collected by the piezoelectric sensor array can be used to separate the high-frequency noise and low-frequency structural vibration components through the empirical mode decomposition algorithm to generate the denoised multi-band vibration signals; The piezoelectric sensor array adopts a circular layout, with 8 sensors evenly deployed around the bolt connection part, and the sampling frequency is set at 20 kHz (meeting the Nyquist sampling requirements for mechanical vibration signals). The collected original vibration signals contain high-frequency noise (such as electromagnetic interference, equipment friction noise) and low-frequency structural vibration components (such as periodic micro-movement caused by bolt loosening). The Empirical Mode Decomposition (EMD) algorithm adaptively decomposes the signal into multiple Intrinsic Mode Functions (IMFs).

[0025] Decomposition process: Perform EMD on each sensor signal, and the decomposition level is dynamically adjusted (usually 6 - 10 layers). For example, in a certain decomposition, IMF1 (high-frequency noise, frequency > 5 kHz), IMF2 - IMF4 (mid-frequency structural vibration, frequency 1 - 5 kHz), and IMF5 - IMF8 (low-frequency vibration, frequency < 1 kHz) are obtained.

[0026] Select effective IMFs through the correlation coefficient: If the correlation coefficient between an IMF and the original signal < 0.2 (empirical threshold), it is determined as noise and removed. For example, IMF1 is removed due to a correlation coefficient of 0.15.

[0027] Signal reconstruction: Retain IMF2 - IMF5 (correlation coefficient > 0.3), and the frequency range of the reconstructed signal is 0.5 - 4 kHz, covering the characteristic frequency band of bolt loosening.

[0028] Evaluation of the denoising effect: The signal-to-noise ratio (SNR) is increased from 15 dB of the original signal to 28 dB.

[0029] Output example: After the EMD processing of a certain bolt vibration signal, 4 frequency bands (0.5 - 1 kHz, 1 - 2 kHz, 2 - 3 kHz, 3 - 4 kHz) are reconstructed, and each frequency band corresponds to different loosening modes (for example, the low-frequency band reflects macroscopic displacement, and the high-frequency band reflects microscopic friction).

[0030] Perform 8-layer wavelet packet decomposition on the multi-band vibration signal, calculate the energy entropy values of each sub-band, and select the frequency bands where the energy entropy mutation is within the threshold of ±15% as the loosening characteristic frequency bands; For the wavelet packet decomposition (WPD), the Daubechies 4 (db4) wavelet basis is selected, and the decomposition level is 8 layers. Each frequency band signal is further divided into 2^8 = 256 sub-bands (frequency resolution ≈ 15.625 Hz).

[0031] Decomposition and energy entropy calculation: Perform 8-layer WPD on the 4 reconstructed frequency bands respectively. For example, the 1 - 2 kHz frequency band is decomposed into sub-band 1 (1 - 1.015 kHz) to sub-band 256 (1.984 - 2 kHz).

[0032] Calculate the energy entropy of each sub-band: , where, It is the ratio of the energy of the sub - frequency band to the total energy. Under normal conditions, the energy entropy distribution is stable; when there is looseness, the local entropy value mutates (for example, the entropy value of a certain sub - frequency band suddenly increases from 0.8 to 1.2).

[0033] Feature frequency band screening: Set the dynamic threshold: Take the moving average (window length 100) of the entropy values of historical normal data, allowing a ±15% fluctuation. For example, if the normal entropy value of a certain sub - frequency band is 1.0, then the threshold range is 0.85 - 1.15, and values exceeding this range are marked as abnormal.

[0034] Example of screening results: In the range of 2.5 - 2.6 kHz, it is detected that the entropy values of 3 sub - frequency bands reach 1.3 (exceeding the threshold of 1.15), which is determined as the loosening feature frequency band.

[0035] Based on the distributed fiber optic sensor to detect the stress wave propagation path, the cross - correlation algorithm is used to calculate the time delay difference between adjacent sensors, and the micro - displacement fluctuation trajectory of the bolt connection interface is inversely obtained; The distributed fiber optic sensors are arranged along the axial direction of the bolt, with a spacing of 10 mm, covering the entire length of the bolt (for example, 5 sensors are arranged for a 50 - mm bolt). When the stress wave (caused by vibration or external impact) propagates in the bolt, each sensor detects the arrival time difference.

[0036] Time delay difference calculation: Perform cross - correlation operation on the signals of adjacent sensors (such as sensors 1 and 2): The sliding window length is 10 ms (corresponding to 200 sampling points), and the step size is 1 ms.

[0037] Calculate the peak position of the cross - correlation coefficient to determine the time delay Δt. For example, if the signal of sensor 2 is delayed by 0.2 ms compared to sensor 1, the corresponding stress wave propagation speed v = 10 mm / 0.2 ms = 50 m / s.

[0038] Micro - displacement inversion: Based on the time delay difference sequence, construct a propagation path matrix, and inversely obtain the micro - displacement fluctuation by the least - squares method. For example, in a certain inversion, the displacement of the bolt head in the x - direction is 0.05 mm, in the y - direction is 0.02 mm, and the frequency is 2 Hz (corresponding to the loosening micro - motion frequency).

[0039] Trajectory visualization: Generate a time - displacement curve, and mark the area where the displacement amplitude exceeds 0.03 mm (safety threshold) as the high - risk time period.

[0040] Parameter example: The sensitivity of the fiber optic sensor is 0.1 με (micro - strain), the time delay resolution is 0.01 ms, and the displacement inversion error is <±0.005 mm.

[0041] Align the energy spectrum of the loosening feature frequency band and the micro - displacement fluctuation trajectory according to the time window, and generate a three - dimensional micro - motion feature matrix through tensor splicing. The matrix dimension contains three feature axes: frequency, displacement, and time sequence.

[0042] Time window alignment: Adopt a sliding window mechanism with a window length of 1 second (including 20,000 sampling points) and a step size of 0.1 second.

[0043] Perform linear interpolation alignment on the energy spectrum of the characteristic frequency band (256 dimensions) and the micro-displacement trajectory (x and y dual channels, each 100 dimensions) within each window to ensure time sequence synchronization.

[0044] Tensor splicing: Dimension definition: Frequency axis: 256 sub-band energy values; Displacement axis: 100 points for displacement in x / y directions (total 200 dimensions); Time sequence axis: 10 time windows (1-second window, step size 0.1 second).

[0045] Data format: A three-dimensional matrix of 256×200×10, where each element is a normalized energy or displacement value (0 - 1).

[0046] Example: During the time period of t = 5 - 6 seconds for a certain bolt, the energy of the characteristic frequency band suddenly increases by 30% at 2.5 kHz, and at the same time, there are periodic fluctuations in the displacement in the x direction (amplitude 0.06 mm). The element value at the corresponding position in the matrix is marked as 0.8 (high risk).

[0047] In this step, vibration signals are collected through a piezoelectric sensor array, and the wavelet packet decomposition algorithm is used to separate the characteristic frequency bands strongly related to bolt loosening (such as 5 - 8 kHz high-frequency components). At the same time, the distributed fiber optic sensor is combined to detect the stress wave propagation time delay difference, and the micron-level displacement fluctuation of the bolt connection interface is inversely calculated. By aligning the frequency domain energy spectrum and the displacement trajectory in the time dimension, a three-dimensional characteristic matrix containing frequency, displacement, and time sequence is constructed to comprehensively characterize the multi-physical field characteristics of bolt loosening, break through the limitations of traditional single vibration monitoring, and significantly improve the sensitivity of micro-motion detection by fusing vibration frequency domain characteristics and stress wave time delay information, providing high-precision input data for subsequent health state assessment.

[0048] S202, based on the micro-motion characteristic matrix, fuse the local stress distribution data collected by the strain gauges on the bolt surface, use a deep convolutional neural network to construct a multi-modal feature vector, and associate the spatial correlation between vibration characteristics and stress distribution through a cross-modal attention mechanism, and output a quantitative index of the bolt's health state; Specifically, according to the local stress distribution data collected by the strain gauge array, a thermal map of the stress gradient on the bolt surface can be generated, and it is mapped into a stress field image of 128×128 pixels through a bilinear interpolation algorithm; The strain gauge array is pasted on the bolt surface in a "cross" layout (such as a 4×4 array). The sampling frequency of each strain gauge (such as model HBM LY41) is set to 1kHz to collect micro-strain data (unit με) when the bolt is loaded in real time. The raw data is pre-processed by an anti-aliasing filter (cut-off frequency 500Hz), and the strain gradient at each point is calculated: Gradient calculation: Take each strain gauge as the center and calculate the strain difference between it and the adjacent 4 gauges (up, down, left, and right). For example, if the strain of the center gauge is 200με and the strain on the right is 220με, the horizontal gradient is 20με / mm (assuming the spacing is 1mm); 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; Bilinear interpolation:

[0049] Divide the 16×16 grid into 15×15 quadrilateral elements; Interpolate the unknown points (such as coordinates (x,y)) within each cell: Calculate the weights of four adjacent known points (inversely proportional to the distance); The strain gradient at the interpolation point is obtained by weighted summation; Repeated interpolation to 128×128 pixels to fill in the missing areas. For example, the original point spacing of 1mm corresponds to a resolution of 0.125mm after interpolation, which improves image details.

[0050] The final generated stress field image is saved in PNG format through the OpenCV library. The image size is 128×128, each pixel corresponds to the actual physical size of 0.125mm×0.125mm, and the grayscale value reflects the degree of local stress concentration (such as grayscale 200 indicates a high stress area).

[0051] ‌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; ResNet-50 network performs transfer learning on stress field images: Input adjustment: adapt the pre-trained ImageNet weights to a single-channel grayscale image, and modify the first-layer convolution kernel to a 1-channel input; Feature extraction: intercept the first 4 residual blocks (conv1~conv4_x) of ResNet-50 and output a 1024-dimensional feature vector. For example, after convolution and pooling of an input 128×128 image, the final feature map size is 4×4×1024; Global average pooling: compress the 4×4×1024 tensor into a 1024-dimensional vector as the stress modal feature.

[0052] The 3D-CNN network processes the micro-motion feature matrix (dimension: frequency × displacement × time series = 64 × 64 × 100): Network structure: Convolutional layer 1: 3D convolutional kernel 5×5×5, output channels 64, stride 2×2×2; Convolutional layer 2: 3D convolutional kernel 3×3×3, output channels 128, stride 1×1×1; Pooling layer: 3D max pooling, kernel size 2×2×2; Feature extraction: After convolution and pooling of the input 64×64×100, an 8×8×25×128 tensor is output; Flattening and fully connected: Flatten 8×8×25×128 into a 204,800-dimensional vector, and compress it to 1,024 dimensions through the fully connected layer as the vibration mode feature.

[0053] The dual-modal feature vector is composed of the stress mode (1,024 dimensions) and the vibration mode (1,024 dimensions) concatenated together, with a total dimension of 2,048.

[0054] Calculate the correlation weights between the vibration feature and the stress feature through the gated attention mechanism, and perform dynamic weighted fusion on the dual-modal feature vector to generate a joint feature vector; The structure of the gated attention mechanism is as follows: Input processing: Stress feature vector S (1,024 dimensions); Vibration feature vector V (1,024 dimensions); Attention weight calculation: Concatenate S and V to obtain a 2,048-dimensional joint input; Generate an intermediate representation through a fully connected layer (output dimension 512) and ReLU activation; Then, through a fully connected layer (output dimension 2) and Softmax, generate a weight vector [α, 1-α], where α ∈ [0,1]; 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%.

[0055] Training optimization: Loss function: Cross-entropy loss combined with weight sparsity regularization (L1 regularization term coefficient 0.01); Backpropagation: Use the Adam optimizer (learning rate 0.001, batch size 32).

[0056] The finally generated joint feature vector maintains a dimension of 1,024, fusing the key information of the dual modality.

[0057] 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 the node feature for graph embedding learning to obtain the graph embedding vector. Specifically, 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 the node feature for graph embedding learning to obtain the graph embedding vector may include: The joint feature vector is processed by maximum-minimum normalization according to the modal dimension to eliminate the dimension difference and obtain the normalized node feature vector. The joint feature vector is composed of the fusion of vibration features and stress features. The original data collected by different sensors have significant differences in numerical ranges due to physical dimension differences (such as the vibration acceleration unit is g and the stress unit is MPa). For example, the vibration characteristic frequency band energy value of a certain bolt is 0.8 (dimensionless), while the surface stress gradient value is 120 MPa. If both are directly input into the graph convolutional network, the model will be biased towards large numerical features due to the magnitude difference.

[0058] Maximum-minimum normalization (Min-Max Normalization) maps each modal feature to the [0, 1] interval through linear transformation, and the formula logic is: For each modal dimension, calculate the maximum value (Max) and minimum value (Min) of all nodes in this dimension, and each eigenvalue x is scaled by (x - Min) / (Max - Min). For example: Vibration energy value range [0.2, 1.6] → After normalization, 0.2 → 0, 1.6 → 1; Stress gradient range [50, 200] MPa → 50 → 0, 200 → 1.

[0059] In practical applications, a sliding window is used to update the extreme values to adapt to dynamic working conditions. For example, recalculate Max and Min every 10 minutes to avoid normalization distortion caused by sudden changes in equipment load. Each dimension of the normalized feature vector is comparable, ensuring the fairness of weight assignment in subsequent graph convolution.

[0060] Based on the normalized node feature vector, through a two-layer graph convolutional network, the first layer aggregates the features of directly adjacent nodes, and the second layer aggregates the features of second-order neighborhood nodes, and combines the gated attention mechanism to dynamically adjust the aggregation weight to generate a hidden state vector representing the spatial relationship between the bolt connection structure and adjacent components. Graph structure definition: Nodes: Bolts and their directly connected components (such as flanges, brackets), and each node represents a physical entity; Edges: Define the adjacency matrix according to the mechanical assembly relationship. The edge weight between directly adjacent nodes is 1 (such as bolt - flange), and the edge weight of the second-order neighborhood (such as bolt - flange - bracket) is 0.5.

[0061] First Graph Convolutional Layer (GCN Layer 1): Input: Normalized node feature vectors (dimension 128); Adjacency matrix A: Only contains directly adjacent nodes (edge weight 1); Aggregation operation: For each node, weighted sum of its own features and neighbor features. The weight matrix W1 is learned through training, and the activation function uses ReLU. For example, the bolt node aggregates its own features (weight 0.6) and flange features (weight 0.4) to generate an intermediate feature vector (dimension 256).

[0062] Gated Attention Mechanism: Calculation of attention scores: For nodes i and j, calculate the correlation score through a learnable parameter matrix, such as score(i,j) = W_a·[h_i || h_j], where || represents concatenation; Gating coefficient: Map the score to (0,1) through the sigmoid function as the aggregation weight. For example, the attention coefficient between the bolt and the flange is 0.8, and with the bracket is 0.3; Dynamic aggregation: The node feature is updated as a combination of its own feature and weighted neighbor features, such as h_i' = h_i +Σ(gate_ij * h_j).

[0063] Second Graph Convolutional Layer (GCN Layer 2): Input: Implicit features output from the first layer (dimension 256); Expansion of the adjacency matrix: Contains second-order neighborhood nodes (edge weight 0.5); Cross-layer aggregation: Combine the gated attention mechanism to dynamically adjust the contribution of distant nodes. 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).

[0064] Output: Implicit state vector (dimension 512), encoding the node's own characteristics and its structural relationships within the multi-hop neighborhood.

[0065] Perform a graph pooling operation on the implicit state vector, adopt the adaptive mean pooling algorithm to fuse node features, and output a graph embedding vector representing the global structural characteristics, with the vector dimension compressed to 256 dimensions.

[0066] Adaptive Mean Pooling: Pooling objective: Convert a graph structure of variable size (such as the number of different bolt connection components) into global features of a fixed dimension; Adaptive mechanism: Dynamically adjust the pooling granularity according to the node importance. The importance is measured by the node feature norm. For example, the larger the feature L2 norm, the more significant the contribution of the node to the global state.

[0067] Pooling steps: Node sorting: Sort the nodes in descending order of feature norm. For example, bolt node (norm 12.5) > flange node (10.2) > bracket node (8.7); Hierarchical pooling: First layer: Retain the original features for the Top 50% nodes (such as bolts and flanges); Second layer: Perform average pooling on the remaining nodes (such as brackets) to generate summary features; Feature concatenation: Concatenate the retained features and the summary features to obtain the pooled features (dimension 1024).

[0068] Dimensionality reduction processing: Fully connected layer: Map the 1024-dimensional features to 256 dimensions, and use the Tanh activation function; Regularization: Use Dropout (ratio 0.3) to prevent overfitting.

[0069] Example output: The graph embedding vector of a certain bolt connection is [0.7, -0.2,..., 0.5] (256-dimensional), where the positive value dimensions indicate high structural compactness, and the negative value dimensions imply potential loosening risks.

[0070] Map the graph embedding vector to a scalar value between 0 and 1 through the fully connected layer, and output a quantitative health status indicator representing the loosening degree of the bolt.

[0071] Fully connected layer structure: Input: 256-dimensional graph embedding vector; Hidden layer: 128-dimensional, ReLU activation; Output layer: 1-dimensional, Sigmoid activation; Training and inference: Label data: 0 represents normal, 1 represents complete loosening, and the labels are obtained through manual calibration or high-precision torque sensors; Loss function: Binary cross-entropy (BCE Loss); Threshold division: Output value ≥ 0.7 is determined as "high loosening risk", 0.3 - 0.7 is "need to monitor", < 0.3 is "normal".

[0072] Example: The graph embedding vector of a certain bolt outputs 0.85 through the fully connected layer, which is determined as a high risk and triggers a real-time warning.

[0073] The micro-motion feature matrix is input into a 3D-CNN to extract spatio-temporal features. At the same time, the stress field image generated by the strain gauge is processed through a ResNet-50 network, and a gated attention mechanism is used to dynamically weight and fuse the two types of features. Further, a graph convolutional network is combined to model the mechanical coupling relationship between the bolt and adjacent components, and finally, a continuously varying health index from 0 to 1 (0 represents fully tightened, 1 represents complete failure) is output, realizing the cross-modal deep association of vibration and stress data, overcoming the coarseness of traditional threshold alarms, and the quantization index can accurately reflect the loosening development stage (e.g., 0.3 - 0.5 indicates a decrease in pre-tightening force, 0.5 - 0.7 indicates visible loosening).

[0074] S203, according to the health state quantization index and historical loosening evolution data, use a graph neural network to establish a dynamic relationship map of the bolt-connector, and predict the loosening trend through a temporal convolutional network, and output the loosening risk level; Specifically, the bolt and the connector can be used as graph nodes, and the mechanical coupling strength between the nodes is calculated based on the health index to generate the adjacency matrix of the initial dynamic relationship map; The construction of the dynamic relationship map of the bolt and the connector is the core of analyzing the loosening propagation path. Each bolt and its directly connected components (such as flange plates, brackets) are abstracted as graph nodes, and the node attributes include the health state quantization index (a 0-1 scalar value, 0 represents fully tightened, 1 represents complete loosening), the material stiffness coefficient (such as the bolt material is 8.8-grade steel, and the stiffness coefficient is 210 GPa), and geometric parameters (such as the bolt diameter M20 and the effective thread length 50 mm).

[0075] The stress wave propagation path analysis method is adopted for calculating the mechanical coupling strength: Stress wave path modeling: Based on the stress wave propagation trajectory detected by the distributed fiber optic sensor, the bolt connection interface is divided into grid cells of 5 mm × 5 mm, and the stress transfer efficiency of each cell is calculated. For example, if a certain cell transfers 85% of the stress wave energy during vibration, its transfer efficiency is 0.85.

[0076] Coupling strength quantization: The mechanical coupling strength is defined as the geometric mean of the stress transfer efficiencies between adjacent nodes. For example, there are 3 stress wave paths between bolt A and flange plate B, and the transfer efficiencies are 0.8, 0.7, and 0.9 respectively, then the coupling strength is (0.8 × 0.7 × 0.9)^(1 / 3) = 0.79.

[0077] Adjacency matrix generation: If the coupling strength ≥ 0.7 (empirical threshold), the corresponding position of the adjacency matrix is set to 1, otherwise it is 0. For a multi-bolt system, the dimension of the adjacency matrix is N × N (N is the total number of nodes). For example, a system containing 10 bolts and 5 connectors generates a 15 × 15 matrix.

[0078] The edge weights of the dynamic relationship graph are adjusted dynamically according to real-time health indicators. For example, when the health indicator of bolt A rises from 0.2 to 0.6 (indicating increased loosening), its coupling strength weight with flange B decreases from 0.79 to 0.65, triggering real-time updates of the edge weights in the adjacency matrix.

[0079] A temporal convolutional network is used to extract multi-scale features from historical loosening evolution data, generating a temporal encoding vector of the node state at each time step. The temporal convolutional network (TCN) is used to capture the long-term dependencies of bolt loosening trends. The input data is the temporal data of health indicators for the past 30 days (sampling interval of 1 hour, a total of 720 time points), and the input dimension of each node is 1 (health indicator value).

[0080] Network structure design: Dilated convolutional layer: 3 layers of dilated convolution with dilation coefficients of 1, 2, and 4 respectively, convolution kernel size of 3×1, and the number of channels increasing from 32 to 64 layer by layer. For example, the first layer of convolution kernel covers a 3-hour window, the second layer is extended to 6 hours through a dilation coefficient of 2, and the third layer is extended to 12 hours.

[0081] Residual connection: The output of each layer is added to the input to prevent gradient disappearance. For example, for the input health indicator sequence [0.1, 0.15, 0.2...], multi-scale fluctuation features are extracted after convolution (such as short-term fluctuations from 0.1→0.12 and medium-term trends from 0.1→0.2).

[0082] Activation function: LeakyReLU (negative slope of 0.1) is used instead of ReLU to avoid neuron death.

[0083] Generation of temporal encoding vector: The output of the TCN is a 64-dimensional feature vector at each time step. For example, the health indicator of 0.3 at a certain time step corresponds to the encoding vector [0.12, -0.05, 0.3,..., 0.18], where each dimension represents loosening features at different time scales (such as the first dimension reflecting hourly fluctuations and the third dimension reflecting daily trends).

[0084] Multi-scale feature fusion: The outputs with different dilation coefficients are combined through cross-layer feature splicing. For example, the convolution results with dilation coefficients of 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 mutation events (such as accelerated loosening caused by impact loads).

[0085] The state information of adjacent nodes is aggregated through the graph attention mechanism, and the node hidden state is iteratively updated in combination with the temporal encoding vector to capture the loosening propagation path. The Graph Attention Network (GAT) is used to dynamically fuse the influence of adjacent nodes. The update of the hidden state of each node is divided into three stages: Calculation of attention weights: 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 a learnable parameter vector, W is a weight matrix, and h_i and h_j are the temporal encoding vectors of nodes i and j.

[0086] Multi-head attention (4 heads) is adopted, and the attention weights are calculated independently for each head and then concatenated. For example, the attention weights of node i to j are 0.3, 0.2, 0.4, and 0.1 in the 4 heads respectively, and the final weight is a 4-dimensional vector after concatenation.

[0087] Aggregation of neighbor information: For each head k, aggregate the neighbor features: ; After concatenating the multi-head outputs, a fully connected layer is used for dimensionality reduction: 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.

[0088] State update: Combine its own state with the aggregation result: h_i_new = σ(h_i' + h_i), where σ is the Sigmoid function.

[0089] Update iteratively 3 times to gradually capture the multi-hop propagation effect. For example, the first iteration captures the influence of direct neighbors, and the third iteration captures the indirect coupling of nodes three hops away.

[0090] Visualization of the loosening propagation path: By calculating the cumulative value of the attention weights between nodes, a weight heatmap is drawn. For example, in the path of bolt A → flange B → bracket C, the attention weights are 0.6 and 0.4 respectively, indicating that the loosening mainly propagates along A → B.

[0091] Input the updated node state into a long short-term memory network to predict the loosening evolution trend in the next 24 hours and output a probabilistic risk curve; In one implementation, a long short-term memory network (LSTM) is used for time series prediction. The input is a sequence of hidden states (64 dimensions), and the output is the loosening probability for the next 24 time steps (1 point per hour).

[0092] Network configuration: Hidden layer: 128 LSTM units, dropout rate 0.2 to prevent overfitting; Input window length: 24 hours (the same as the prediction length); Output layer: fully connected layer + Sigmoid, mapping to probability values between 0 and 1.

[0093] Training strategy: Loss function: weighted cross-entropy, assigning 3 times the weight to high-risk samples (health indicators > 0.7 in history); Optimizer: Adam, learning rate 0.001, batch size 32; Early stopping mechanism: If the validation set loss does not decrease for 5 consecutive rounds, terminate the training.

[0094] Probabilistic risk curve generation: The model outputs 24 probability values, e.g., [0.1, 0.15, 0.2,..., 0.45], representing the probability that the health indicator exceeds the threshold of 0.5 for each hour in the future. Smooth the curve through cubic spline interpolation and mark the 95% confidence interval.

[0095] Risk level classification: Slope mutation detection: Apply the Savitzky-Golay filter (window 15, order 3) to the smoothed curve, and calculate the point-by-point slope. If the absolute value of the slope of 3 consecutive points > 0.5, it is determined as a mutation segment.

[0096] Dynamic Time Warping (DTW): Align the current curve with the historical high-risk template (such as an exponentially rising curve) and calculate the minimum cumulative distance.

[0097] Threshold crossing: If the probability value breaks through 0.7 and the slope > 0, immediately trigger a high-risk alarm.

[0098] Based on the slope change and threshold crossing point of the probabilistic risk curve, use the dynamic time warping algorithm to classify the bolt loosening risk level into low risk, medium risk, or high risk levels.

[0099] In another implementation, more specifically, based on the slope change and threshold crossing point of the probabilistic risk curve, use the dynamic time warping algorithm to classify the bolt loosening risk level into low risk, medium risk, or high risk levels, including:

[0100] Apply the Savitzky-Golay filter to the probabilistic risk curve, where the window length is 15 and the polynomial order is 3, to eliminate high-frequency noise and retain trend features; Savitzky-Golay filter (abbreviated as SG filter) is a smoothing algorithm based on local polynomial fitting. Its core idea is to suppress noise through polynomial regression within a sliding window while retaining the trend characteristics of the signal. In this step, for the bolt loosening risk curve (such as a probability curve within 24 hours, with a numerical range of 0 - 1 representing the loosening probability), smoothing is performed using a window length of 15 (i.e., each fitting covers 15 consecutive sampling points) and a cubic polynomial (order 3).

[0101] Operation process: Window sliding: Starting from the first sampling point of the risk curve, each time take 15 consecutive points as the fitting window. For example, if the sampling interval is 1 minute, the window covers a 15 - minute data segment.

[0102] Polynomial fitting: Perform cubic polynomial fitting (in the form of y = a0 + a1x + a2x² + a3x³) on the 15 points within the window, and calculate the coefficients a0 - a3 by the least squares method. For example, for a window with large data fluctuations, the high - frequency jitters will be smoothed out after fitting.

[0103] Central point replacement: Only retain the fitting value of the central point (the 8th point) of the window as the smoothed output to avoid edge effects. For example, for the original data point sequence [0.2, 0.3, 0.5,..., 0.8], the central point 0.5 may be corrected to 0.48 after fitting.

[0104] Window stepping: Slide the window 1 point to the right and repeat the above process until all data is covered.

[0105] Technical effects: Noise suppression: For example, for a certain section of the original curve with 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 linearly increasing from 0.3 to 0.6), SG filtering can accurately retain this trend and avoid the phase delay caused by the traditional moving average method.

[0106] 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 the mutation segment; The instantaneous slope reflects the change rate of the risk curve and is used to detect the rapid deterioration or alleviation of the loosening risk. The slope is calculated using the sliding window method: Window definition: With the current point as the center, take 2 points before and after each to form a 5 - point window (i.e., the window length is 5). For example, when calculating the slope at the t - th minute, use the data from t - 2 to t + 2 minutes.

[0107] Linear fitting: Perform linear regression on 5 points within the window, and the slope is the instantaneous slope. For example, for the window data [0.4, 0.5, 0.6, 0.7, 0.8], the slope is 0.1 / minute, indicating that the risk increases by 0.1 per minute.

[0108] Determination of mutation segment: Slope threshold: Absolute value > 0.5 (for example, a slope of +0.6 indicates a rapid increase in risk); Duration: The slopes of 3 consecutive sampling points all exceed the threshold. For example, if the slopes of a certain curve segment are 0.6, 0.7, and 0.55 from 10:00 to 10:02 (3 minutes), it is marked as a mutation segment.

[0109] Application example: Rapid increase in high risk: The risk value of a certain bolt soars from 0.3 to 0.8 within 10 minutes, with a corresponding slope of 0.05 / minute. However, due to local fluctuations within the window, an instantaneous slope of 0.6 may be detected (such as from 10:05 to 10:07), and it is marked as a mutation segment; False judgment prevention: A short-term noise (such as a single-point slope of 0.6) will not be misjudged as a mutation because it does not meet the condition of 3 consecutive points.

[0110] Align the current risk curve with the historical high-risk and medium-risk template curves through dynamic time warping, calculate the minimum cumulative distance, and record the number of path bends as a similarity measure; Dynamic time warping (DTW) is an algorithm for elastically aligning two time series to measure their shape similarity. In this step, align the current risk curve with the pre-stored historical templates (high-risk and medium-risk typical curves) using DTW: Template construction: High-risk template: Extract 10 curves from historical data whose risk values rapidly rise above 0.9 and take the average to generate the template; Medium-risk template: Extract 20 curves whose risk values fluctuate between 0.5 and 0.7 to generate the template.

[0111] DTW alignment: Cumulative distance matrix: Construct a point-to-point distance matrix between the current curve and the template curve, and use the Euclidean distance as the distance metric; Path search: Search for the optimal path from the upper left corner to the lower right corner of the matrix so that the sum of the distances of the points on the path (cumulative distance) is the smallest. The path is allowed to move diagonally, horizontally, or vertically, but the slope is restricted to ≤ 2 (to avoid excessive distortion).

[0112] Feature extraction: Minimum cumulative distance: The total distance of the optimal path, and the smaller the value, the more similar it is to the template; Number of path bends: The number of times the path direction changes (e.g., from horizontal to diagonal), reflecting the complexity of alignment.

[0113] If the minimum cumulative distance is less than the high-risk threshold and the number of path bends ≥ 2, it is determined as a high-risk level. If the minimum cumulative distance is greater than or equal to the high-risk threshold and less than the medium-risk threshold, or the minimum cumulative distance is less than the high-risk threshold and the number of path bends < 2, then it is determined as a medium-risk level. Otherwise, if the minimum cumulative distance is greater than or equal to the medium-risk threshold, it is judged as a low-risk level.

[0114] The risk level classification depends on predefined thresholds and path characteristics: Threshold setting: High-risk threshold: Through historical data analysis, the cumulative distance threshold is set to 150 (i.e., a relatively high similarity to the high-risk template); Medium-risk threshold: Set to 250, higher than the high-risk threshold but lower than the low-risk determination line.

[0115] Determination rules: High-risk: Cumulative distance < 150 and number of path bends ≥ 2 (a complex alignment path reflects a mutation pattern); Medium-risk: 150 ≤ cumulative distance < 250 or cumulative distance < 150 but number of bends < 2; Low-risk: Cumulative distance ≥ 250.

[0116] Dynamic adjustment: According to the equipment type and environmental factors (such as vibration intensity), the threshold can float by ±10%. For example, the threshold for heavy machinery is increased to 160 to tolerate a higher baseline risk.

[0117] Application examples: Case 1: The current curve of a wind power bolt has a cumulative distance of 130 from the high-risk template and 3 bends → High-risk; Case 2: A bridge bolt has a cumulative distance of 140 and 1 bend → Medium-risk (due to the simple path, it may be a slow deterioration); Case 3: A machine tool bolt has a cumulative distance of 280 → Low-risk.

[0118] Based on the graph attention mechanism, a dynamic relationship graph of bolts - connectors is constructed. Through a temporal convolutional network, multi-scale temporal features of historical data are extracted (such as hourly fluctuations and daily trends). Combining with LSTM, the probability curve of loosening evolution in the next 24 hours is predicted. Finally, based on the curve slope and threshold intersection point, the low / medium / high-risk levels are divided, breaking through the limitation of single-point monitoring, predicting the loosening propagation path from the system level (such as the cascading failure of adjacent bolts), and the risk level classification makes the maintenance decision more forward-looking (such as high-risk requires immediate shutdown).

[0119] S204, based on the loosening risk level, combined with the equipment operation state parameters, uses an adaptive decision tree model to dynamically adjust the monitoring frequency and alarm threshold, generates a real-time monitoring strategy and a visual warning report, and synchronously triggers the wireless transmission module to upload to the cloud operation and maintenance platform.

[0120] Specifically, the operating speed and load current parameters of the device can be normalized with the loosening risk level to generate a multi-dimensional state feature vector. Device operating parameters (such as speed and load current) and the loosening risk level usually have different dimensions and numerical ranges. The purpose of normalization is to eliminate the dimensional differences so that each parameter participates in the model calculation on the same scale. The specific process is as follows: Data collection and preprocessing: Speed (unit: RPM): Collected by a Hall sensor, with a range of 0 - 3000 RPM, and a typical value such as 1500 RPM; Load current (unit: A): Measured by a current transformer, with a range of 0 - 100 A, and a typical value such as 45 A; Loosening risk level: The low, medium, and high risk levels (coded as 0, 1, 2) output from the previous steps.

[0121] Normalization method:

[0122] Min - Max normalization: Linearly map each parameter to the interval [0, 1]. For example, the normalization formula for speed is: If the actual speed is 1500 RPM, the normalized value is 0.5; Categorical variable encoding: The risk level uses one - hot encoding. Low risk is [1, 0, 0], medium risk [0, 1, 0], and high risk [0, 0, 1].

[0123] Feature vector construction: Concatenate the normalized parameters into a multi - dimensional vector. For example, when the device speed is 1500 RPM (normalized 0.5), load current 45 A (normalized 0.45), and the risk level is medium risk (coded [0, 1, 0]), the generated feature vector is: [0.5, 0.45, 0, 1, 0].

[0124] Technical example: In a scenario of monitoring bolts of a wind turbine, the speed is 1800 RPM (normalized 0.6), load current 60 A (normalized 0.6), and high risk level (coded [0, 0, 1]). The generated feature vector is [0.6, 0.6, 0, 0, 1]. This vector will be used as the input of the adaptive decision tree model.

[0125] ‌Use an incremental learning method to train the adaptive decision tree model, dynamically split nodes according to the multi - dimensional state feature vector, and generate monitoring frequency adjustment rules and alarm thresholds; 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: Initial model construction: Dataset: Use historical data (such as 1000 sets of samples) to train the initial decision tree, and use the C4.5 algorithm to select splitting features. For example, the first split may select "rotation speed" as the root node, and the threshold is set to 0.5 (corresponding to 1500 RPM); Node splitting rule: Based on maximizing the information gain ratio. For example, when the rotation speed > 0.5, enter the right subtree (high-risk working condition), otherwise enter the left subtree (normal working condition).

[0126] Incremental learning mechanism: Sliding window update: Keep the most recent 200 sets of data. When new data arrives, eliminate the earliest data to keep the window size constant; Dynamic splitting strategy: Feature importance evaluation: Dynamically calculate the importance of each feature through the Gini index. For example, if the Gini index of the load current rises from 0.3 to 0.6, it is preferred to use it as a splitting node; Node reorganization: When the classification error rate of a certain node continuously exceeds 15%, trigger local reconstruction. For example, replace the original "rotation speed" splitting node with "load current".

[0127] Rule and threshold generation: Monitoring frequency adjustment: Dynamically set the sampling interval according to the risk level. For example: Low risk: Collect once every 10 minutes; Medium risk: Collect once every 5 minutes; High risk: Collect in real time (once per second); Alarm threshold setting: Combine with the equipment working condition. For example: When the rotation speed > 0.7 (2100 RPM) and the risk level is high risk, trigger a first-level alarm; When the load current > 0.8 (80 A) and the risk level is medium risk, trigger a second-level alarm.

[0128] Technical example: In the monitoring of the joint bolts of an industrial robot, the initial decision tree is based on the rotation speed and load current partitioning strategy. When a new load pattern is introduced (such as frequent start-stop resulting in increased current fluctuations), incremental learning automatically adds "load current change rate" as a new splitting feature and adjusts the monitoring frequency to once every 2 minutes.

[0129] 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 the monitoring interval time, alarm priority, and data sampling accuracy; Monte Carlo tree search (MCTS) finds the optimal strategy in the decision tree space through the simulation-evaluation-backtracking mechanism. The specific steps are as follows: Decision tree space modeling: Node Definition: Each node represents a combination of monitoring strategies. For example, {Monitoring Interval: 5 min, Alarm Priority: 2, Sampling Precision: High}; Edge Definition: The edge represents the adjustment action of policy parameters, such as "increasing the monitoring frequency" or "improving the sampling precision".

[0130] Four-stage Operation of MCTS: Selection: Starting from the root node, select child nodes based on the UCB1 formula (Upper Confidence Bound). The UCB1 formula is: . Among them, 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; Expansion: When encountering an unexplored node, expand new child nodes. For example, add a new policy {Monitoring Interval: 3 min, Alarm Priority: 1}; Simulation: Randomly simulate the execution results of the policy. For example, predict the number of alarms and energy consumption in the next 1 hour under this policy; Backpropagation: Update the path nodes in reverse with the simulation results (such as revenue = 0.8).

[0131] Policy Evaluation and Screening: Revenue Function Design: Considering both monitoring effectiveness (such as alarm accuracy) and resource consumption (such as energy consumption), the formula is: Revenue = 0.7 × Accuracy + 0.3 × (1 - Energy Consumption Ratio).

[0132] Pareto Front Extraction: Screen non-dominated solutions (that is, solutions that cannot improve one objective without harming other objectives). For example: Policy A: Accuracy 90%, Energy Consumption 30%; Policy B: Accuracy 85%, Energy Consumption 20%.

[0133] Technical Example: In the wind power bolt monitoring scenario, after 1000 simulations, MCTS outputs the optimal policy table, as shown in Table 1 for example:

[0134] Table 1

[0135] Convert the real-time monitoring policy table into a JSON format instruction set to drive the visualization engine to generate a three-dimensional dynamic warning map, and at the same time encrypt and upload it to the cloud operation and maintenance platform through the LoRa wireless module.

[0136] JSON Instruction Set Generation: Data Structure Design: json { "device_id": "Bolt_001", "monitoring_strategy": { "interval": 300, / / Unit: seconds "alert_priority": 1, "sampling_precision": "High" }, "timestamp": "2023-10-05T14:30:00Z" }。

[0137] Dynamic Filling: Automatically fill fields according to the content of the policy table. For example, when the policy "3-minute monitoring interval" is selected, the interval field is set to 180 seconds.

[0138] Three-dimensional Dynamic Early Warning Atlas Generation: Visualization Engine: Use the Three.js library to build a WebGL three-dimensional scene. The key components include: Bolt Model: Import the CAD model and color it according to the health indicators (green = normal, red = high risk); Thermodynamic Stress Distribution Map: Map the strain gauge data to frequency bands: 868 MHz (Europe) / 915 MHz (North America); Transmission Power: 14 dBm; Encryption Protocol: AES-256, and the key is rotated every 24 hours.

[0139] Data Encapsulation: Convert the JSON instruction into a binary format, and add a frame header (device ID), a checksum (CRC32), and a frame tail. For example, the transmission frame structure of the instruction Bolt_001 is: [0xAA][Device ID 4B][Data Length 2B][Encrypted Data N B][CRC32 4B][0x55].

[0140] Technical Example: After the bolt monitoring system of a chemical reactor in a chemical plant detects a high risk, it generates a JSON instruction and triggers the following actions: The three-dimensional bolt model flashes red on the local HMI interface, and the superimposed stress thermodynamic map shows local overlimit; The LoRa module encrypts and uploads data to the cloud, and the cloud platform automatically generates a work order and pushes it to the mobile terminal of the maintenance personnel; The maintenance personnel can view the accurate positioning of the bolt through AR glasses and quickly perform the tightening operation.

[0141] Based on the correlation between parameters such as the real-time rotational speed and load of the device 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 atlas is generated. Low-power remote monitoring is achieved through LoRa wireless transmission, intelligent allocation of monitoring resources (reducing the monitoring overhead of low-risk bolts) is realized, the risk spatial distribution is intuitively displayed in the visualization report, and cloud collaboration supports multi-device cluster management.

[0142] It can be seen that according to the multi-dimensional vibration signals and stress wave propagation characteristics of the bolt connection part, a fretting feature matrix of the bolt is generated; based on the fretting feature matrix, local stress distribution data collected by the surface strain gauge of the bolt is fused, and a quantitative index of the bolt's health state is output; according to the quantitative index of the health state and historical loosening evolution data, the loosening risk level is output; based on the loosening risk level, combined with the device operation state 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 visualization early warning report, and synchronously trigger the wireless transmission module to upload to the cloud operation and maintenance platform, so as to realize high-precision, adaptive, and traceable bolt health state evaluation and early warning.

[0143] Another embodiment of the present invention provides a bolt online monitoring system. Refer to Figure 3 , the system may include: An extraction module 301, configured to extract the bolt loosening characteristic frequency band according to the multi-dimensional vibration signals and stress wave propagation characteristics of the bolt connection part, and generate a fretting feature matrix of the bolt by combining the time-delay inversion calculation of the stress wave propagation path; A construction module 302, configured to fuse the local stress distribution data collected by the surface strain gauge of the bolt based on the fretting feature matrix, construct a multi-modal feature vector by using a deep convolutional neural network, and associate the spatial correlation between the vibration feature and the stress distribution through a cross-modal attention mechanism, and output a quantitative index of the bolt's health state; A prediction module 303, configured to establish a dynamic relationship map of the bolt-connector by using a graph neural network according to the quantitative index of the health state and historical loosening evolution data, predict the loosening trend through a temporal convolutional network, and output the loosening risk level; A generation module 304, configured to dynamically adjust the monitoring frequency and alarm threshold based on the loosening risk level, combined with the device operation state parameters, generate a real-time monitoring strategy and a visualization early warning report, and synchronously trigger the wireless transmission module to upload to the cloud operation and maintenance platform.

[0144] The embodiment of the present invention also provides a storage medium, in which a computer program is stored, and the computer program is set to execute the steps in any one of the above method embodiments when running.

[0145] Specifically, in this embodiment, the above storage medium can be set to store a computer program for executing the following steps: S201, according to the multi-dimensional vibration signals and stress wave propagation characteristics of the bolt connection part, extract the bolt loosening characteristic frequency band through the wavelet packet decomposition algorithm, and combine the time delay inversion calculation of the stress wave propagation path to generate the fretting characteristic matrix of the bolt; S202, based on the fretting characteristic matrix, fuse the local stress distribution data collected by the strain gauges on the bolt surface, use a deep convolutional neural network to construct a multi-modal feature vector, and associate the spatial correlation between the vibration characteristics and the stress distribution through a cross-modal attention mechanism, and output the quantitative health state index of the bolt; S203, according to the quantitative health state index and historical loosening evolution data, use a graph neural network to establish a dynamic relationship map of the bolt-connector, predict the loosening trend through a temporal convolutional network, and output the loosening risk level; S204, based on the loosening risk level, combine the device operation state parameters, use an adaptive decision tree model to dynamically adjust the monitoring frequency and alarm threshold, generate a real-time monitoring strategy and a visual warning report, and synchronously trigger the wireless transmission module to upload to the cloud operation and maintenance platform.

[0146] An embodiment of the present invention also provides an electronic device, including a memory and a processor, wherein a computer program is stored in the memory, and the processor is set to run the computer program to execute the steps in any one of the above method embodiments.

[0147] Specifically, the above electronic device may further include a transmission device and an input / output device, wherein the transmission device is connected to the above processor, and the input / output device is connected to the above processor.

[0148] Specifically, in this embodiment, the above processor can be set to execute the following steps through a computer program: S201, according to the multi-dimensional vibration signals and stress wave propagation characteristics of the bolt connection part, extract the bolt loosening characteristic frequency band through the wavelet packet decomposition algorithm, and combine the time delay inversion calculation of the stress wave propagation path to generate the fretting characteristic matrix of the bolt; S202, based on the fretting characteristic matrix, fuse the local stress distribution data collected by the strain gauges on the bolt surface, use a deep convolutional neural network to construct a multi-modal feature vector, and associate the spatial correlation between the vibration characteristics and the stress distribution through a cross-modal attention mechanism, and output the quantitative health state index of the bolt; S203, according to the quantitative health state index and historical loosening evolution data, use a graph neural network to establish a dynamic relationship map of the bolt-connector, predict the loosening trend through a temporal convolutional network, and output the loosening risk level; S204. Based on the loosening risk level and combined with the device operation state 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 warning report, and simultaneously trigger the wireless transmission module to upload to the cloud operation and maintenance platform.

[0149] The structure, features and effects of the present invention have been described in detail based on the embodiments shown in the drawings. The above is only the preferred embodiment of the present invention, but the present invention is not limited to the implementation scope shown in the drawings. All changes made according to the concept of the present invention, or equivalent embodiments modified into equivalent changes, should still be within the protection scope of the present invention when they do not exceed the spirit covered by the description and the drawings.

Claims

1. An on-line monitoring method for bolts, characterized in that, The method includes: According to the multi-dimensional vibration signals and stress wave propagation characteristics of the bolt connection part, extract the bolt loosening characteristic frequency band through the wavelet packet decomposition algorithm, and combine the time-delay inversion calculation of the stress wave propagation path to generate the fretting characteristic matrix of the bolt; Based on the fretting characteristic matrix, fuse the local stress distribution data collected by the strain gauges on the bolt surface, use a deep convolutional neural network to construct a multi-modal feature vector, and associate the spatial correlation between the vibration characteristics and the stress distribution through a cross-modal attention mechanism, and output the quantitative health status index of the bolt; According to the quantitative health status index and historical loosening evolution data, use a graph neural network to establish a bolt-connector dynamic relationship graph, predict the loosening trend through a temporal convolutional network, and output the loosening risk level; Based on the loosening risk level, combined with the equipment operation status parameters, use an adaptive decision tree model to dynamically adjust the monitoring frequency and alarm threshold, generate a real-time monitoring strategy and a visual warning report, and synchronously trigger the wireless transmission module to upload to the cloud operation and maintenance platform.

2. The method according to claim 1, wherein The step of according to the multi-dimensional vibration signals and stress wave propagation characteristics of the bolt connection part, extracting the bolt loosening characteristic frequency band through the wavelet packet decomposition algorithm, and combining the time-delay inversion calculation of the stress wave propagation path to generate the fretting characteristic matrix of the bolt includes: According to the vibration signals collected by the piezoelectric sensor array, separate the high-frequency noise and low-frequency structural vibration components through the empirical mode decomposition algorithm to generate the denoised multi-band vibration signals; Perform 8-layer wavelet packet decomposition on the multi-band vibration signals, calculate the energy entropy values of each sub-band, and select the frequency bands where the energy entropy mutation is within the threshold of ±15% as the loosening characteristic frequency bands; Based on the distributed fiber optic sensor to detect the stress wave propagation path, use the cross-correlation algorithm to calculate the time-delay difference between adjacent sensors, and invert the micro-displacement fluctuation trajectory of the bolt connection interface; Align the energy spectrum of the loosening characteristic frequency band and the micro-displacement fluctuation trajectory according to the time window, and generate a three-dimensional fretting characteristic matrix through tensor splicing. The matrix dimension includes three characteristic axes: frequency, displacement, and time series.

3. The method according to claim 2, wherein The step of based on the fretting characteristic matrix, fusing the local stress distribution data collected by the strain gauges on the bolt surface, using a deep convolutional neural network to construct a multi-modal feature vector, and associating the spatial correlation between the vibration characteristics and the stress distribution through a cross-modal attention mechanism, and outputting the quantitative health status index of the bolt includes: According to the local stress distribution data collected by the strain gauge array, generate a stress gradient heat map on the bolt surface, and map it to a stress field image of 128×128 pixels through the bilinear interpolation algorithm; Input the stress field image into the ResNet-50 network to extract high-order features, and input the fretting characteristic matrix into the 3D-CNN network to extract spatio-temporal features to generate a bimodal feature vector; Calculate the correlation weight between the vibration characteristics and the stress characteristics through the gated attention mechanism, and perform dynamic weighted fusion on the bimodal feature vector to generate a joint feature vector; Use a graph convolutional network to establish a spatial relationship graph between the bolt connection structure and adjacent components, and use the joint feature vector as the node feature for graph embedding learning to obtain the graph embedding vector; Map the graph embedding vector to a scalar value between 0 and 1 through a fully connected layer, and output a health state quantization index representing the degree of bolt loosening.

4. The method according to claim 3, characterized in that, Based on the health state quantization index and historical loosening evolution data, use a graph neural network to establish a bolt-connector dynamic relationship graph, and predict the loosening trend through a temporal convolutional network, and output the loosening risk level, including: Take the bolt and the connector as graph nodes, calculate the mechanical coupling strength between nodes based on the health index, and generate the adjacency matrix of the initial dynamic relationship graph; Use a temporal convolutional network to perform multi-scale feature extraction on historical loosening evolution data, and generate a temporal encoding vector of the node state at each time step; Aggregate the state information of adjacent nodes through a graph attention mechanism, and iteratively update the node hidden state in combination with the temporal encoding vector to capture the loosening propagation path; Input the updated node state into a long short-term memory network, predict the loosening evolution trend in the next 24 hours, and output a probabilistic risk curve; Based on the slope change and threshold crossing point of the probabilistic risk curve, use the dynamic time warping algorithm to divide the loosening risk level of the bolt into low risk, medium risk or high risk levels.

5. The method according to claim 4, characterized in that, Based on the loosening risk level, combined with the device operation state parameters, use an adaptive decision tree model to dynamically adjust the monitoring frequency and alarm threshold, generate a real-time monitoring strategy and a visual warning report, and synchronously trigger the wireless transmission module to upload to the cloud operation and maintenance platform, including: Normalize the device operating speed and load current parameters with the loosening risk level to generate a multi-dimensional state feature vector; Use an incremental learning method to train an adaptive decision tree model, dynamically split nodes according to the multi-dimensional state feature vector, and generate a monitoring frequency adjustment rule and an alarm threshold; Based on the Monte Carlo tree search algorithm, explore the optimal strategy in the decision tree space, and output a real-time monitoring strategy table including the monitoring interval time, 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 graph, and at the same time encrypt and upload it to the cloud operation and maintenance platform through the LoRa wireless module.

6. The method according to claim 3, wherein Use a graph convolutional network to establish a spatial relationship graph between the bolt connection structure and adjacent components, use the joint feature vector as the node feature for graph embedding learning, and obtain the graph embedding vector, including: Perform maximum-minimum normalization on the joint feature vector according to the modal dimension to eliminate the dimension difference and obtain a normalized node feature vector; Based on the normalized node feature vector, through a two-layer graph convolutional network, the first layer aggregates the features of directly adjacent nodes, the second layer aggregates the features of second-order neighborhood nodes, and dynamically adjusts the aggregation weight in combination with the gated attention mechanism to generate a hidden state vector representing the spatial relationship between the bolt connection structure and adjacent components; Perform a graph pooling operation on the hidden state vector, use the adaptive mean pooling algorithm to fuse the node features, and output a graph embedding vector representing the global structure characteristics, and the vector dimension is compressed to 256 dimensions.

7. The method according to claim 4, characterized in that Based on the slope change and threshold crossing point of the probabilistic risk curve, use the dynamic time warping algorithm to divide the loosening risk level of the bolt into low risk, medium risk or high risk levels, including: Apply the Savitzky-Golay filter to the probabilistic risk curve, where the window length is 15 and the polynomial order is 3, to eliminate high-frequency noise and retain trend characteristics; Calculate the instantaneous slope of each point on the curve based on a 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 the mutation segment; Align the current risk curve with the historical high-risk and medium-risk template curves through dynamic time warping, 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 ≥ 2, it is determined to be a high-risk level. If the minimum cumulative distance is greater than or equal to the high-risk threshold and less than the medium-risk threshold, or the minimum cumulative distance is less than the high-risk threshold and the number of path bends < 2, then it is determined to be a medium-risk level. Otherwise, if the minimum cumulative distance is greater than or equal to the medium-risk threshold, it is judged to be a low-risk level.

8. An on-line bolt monitoring system, characterized in that, The system includes: An extraction module, which is used to extract the bolt loosening characteristic frequency band through the wavelet packet decomposition algorithm according to the multi-dimensional vibration signal and stress wave propagation characteristics of the bolt connection part, and generate the fretting characteristic matrix of the bolt by combining the time-delay inversion calculation of the stress wave propagation path; A construction module, which is used to fuse the local stress distribution data collected by the surface strain gauge of the bolt based on the fretting characteristic matrix, construct a multi-modal feature vector using a deep convolutional neural network, and associate the spatial correlation between the vibration characteristics and the stress distribution through a cross-modal attention mechanism, and output the health status quantization index of the bolt; A prediction module, which is used to establish a bolt-connector dynamic relationship graph using a graph neural network according to the health status quantization index and historical loosening evolution data, predict the loosening trend through a temporal convolutional network, and output the loosening risk level; A generation module, which is used to dynamically adjust the monitoring frequency and alarm threshold based on the loosening risk level and combined with the device operation state parameters, generate a real-time monitoring strategy and a visual warning report, and synchronously trigger the wireless transmission module to upload to the cloud operation and maintenance platform.

9. A storage medium, characterized in that, A computer program is stored in the storage medium, wherein the computer program is set to execute the method according to any one of claims 1-7 when running.

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 set to run the computer program to execute the method according to any one of claims 1-7.

Images