Highway bridge settlement monitoring device

Abstract

The invention discloses a highway bridge settlement monitoring device which comprises an intelligent sensing layer which collects settlement, vibration and environmental data through a multi-source sensor, dynamically compensates vibration errors after electromagnetic shielding and anti-interference processing of a damping support, and transmits the data to an edge layer; the edge calculation layer is used for deploying a CNN (Convolutional Neural Network) to carry out feature extraction and compression, calculating and processing an abnormal value in real time by utilizing an Flink flow, triggering preliminary early warning through a preset static threshold value and uploading the early warning to the cloud; the cloud decision-making layer adopts a VGG16-LSTM fusion model, combines a DQN to dynamically adjust a threshold value, uses an RBF to correct temperature and humidity influence, triggers model retraining when a false alarm rate exceeds 5%, synchronously generates an SHAP thermodynamic diagram and gradient weighted class activation mapping, visualizes settlement key factors and outputs final early warning details; and the execution control layer is used for regulating and controlling the telescopic road pile to shunt vehicles according to a cloud instruction, dynamically adjusting an LED guide rod and a reflective warning flag, and feeding back an execution state to the cloud for verification in real time.

Description

Technical Field

[0001] This invention relates to the field of bridge settlement monitoring technology, and in particular to a highway bridge settlement monitoring device. Background Technology

[0002] Bridge settlement monitoring is a core component of ensuring the safe operation of large-scale infrastructure. This is due to the vital role of bridges as transportation lifelines—enduring long-term vehicle loads, geological changes, and environmental erosion, they are prone to uneven settlement. This settlement can range from affecting travel comfort to causing structural cracking or even collapse. Systematic monitoring can promptly detect settlement trends, providing a basis for maintenance decisions and preventing small problems from escalating into major hazards. It is a key means of preventative maintenance and is of great significance for ensuring public safety and extending the service life of bridges.

[0003] Nowadays, the problem of poor adaptability of bridge settlement monitoring to extreme environments is becoming increasingly prominent. For example, a cross-sea bridge in southern my country encountered strong winds, salt spray corrosion and high-frequency lightning interference during the typhoon season. Traditional sensors frequently experienced data drift and transmission interruption due to insufficient electromagnetic shielding, resulting in the failure of settlement early warning and creating serious hidden dangers. Therefore, a highway bridge settlement monitoring device is proposed. Summary of the Invention

[0004] The purpose of this invention is to address the shortcomings of existing technologies by proposing a highway bridge settlement monitoring device.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: A highway bridge settlement monitoring device, comprising: Intelligent sensing layer: Real-time collection of settlement deformation, vibration frequency and environmental parameters through multi-source heterogeneous sensor network, dynamic compensation of vibration error after three-layer electromagnetic shielding and silicone shock-absorbing bracket anti-interference treatment, and finally transmission to edge computing layer; Edge computing layer: Convolutional neural networks (CNNs) are deployed based on edge computing modules to extract and compress data features. After outliers are processed in real time by the Apache Flink streaming computing framework, the data is transmitted to the cloud. Preliminary warning signals are triggered by preset static thresholds and sent to the cloud decision layer. Cloud-based decision layer: Receives preliminary early warning signals and structured data transmitted from the edge layer, employs a VGG16-LSTM multimodal feature fusion model, optimizes it with five-fold cross-validation, and dynamically adjusts the early warning threshold using the DQN reinforcement learning framework; it fuses confidence scores from multiple sensors, uses an RBF neural network to correct for the effects of temperature and humidity, automatically triggers model retraining when the false alarm rate exceeds 5%, and simultaneously generates SHAP value heatmaps and gradient-weighted class activation maps to visualize key settlement factors and generate final early warning details; Execution control layer: After receiving instructions from the cloud, the sub-control terminal adjusts the speed control of the multi-segment telescopic bollards to perform diversion, flow restriction and flow interruption operations. The LED guide poles and reflective warning flags on the top of the bollards dynamically adjust the flashing frequency and color according to the warning level. At the same time, the video acquisition mechanism monitors the vehicle speed in real time and feeds back the execution status to the cloud for verification.

[0006] Furthermore, the dynamic compensation for vibration error after three layers of electromagnetic shielding and silicone shock-absorbing bracket anti-interference treatment includes the following steps: The three-layer electromagnetic shielding adopts a composite structure of "copper foil layer (inner layer) - aluminum foil layer (middle layer) - conductive cloth layer (outer layer)". The copper foil layer absorbs high-frequency electromagnetic noise (such as high-voltage line radiation), the aluminum foil layer shields medium-frequency interference (such as lightning pulse), and the conductive cloth layer suppresses low-frequency electrostatic coupling, forming a gradient electromagnetic protection to ensure that the original sensor data is not interfered with in a strong electromagnetic environment. The silicone vibration damping bracket uses high-damping silicone material and absorbs the vibration energy of the bridge through a three-stage vibration damping structure of "base-bracket-sensor". The bracket design conforms to the dynamic modal frequency of the bridge, avoids resonance with the natural vibration frequency of the bridge, and reduces the direct impact of mechanical vibration on the sensor measurement. Kalman filtering dynamically corrects measurement errors caused by vibration through a two-stage recursive calculation of "prediction-update". Define the system state vector ,in, For settlement, For the settling rate, For vibration acceleration, the state transition equation is constructed based on Newton's laws of motion:

[0007] Where F is the state transition matrix (describing the evolution of the state over time), and B is the control input matrix (external force). ~N(0,Q) represents process noise (reflecting model uncertainty); Based on the optimal state estimate from the previous moment And error covariance matrix Calculate the prior state estimate at the current time:

[0008] Update prior error covariance:

[0009] Combine sensor measurements (e.g., GNSS settlement, accelerometer output), construct the observation equations:

[0010] Where H is the observation matrix (mapping the state to the measurement space). ~N(0,R) represents the measurement noise (reflecting sensor accuracy); Calculate Kalman gain Balance the weights of predicted and measured values:

[0011] Where R is the measurement noise covariance matrix; By combining predicted and measured values, the posterior optimal state estimate is obtained:

[0012] Updated posterior error covariance:

[0013] Where I is the identity matrix.

[0014] Furthermore, the deployment of a convolutional neural network (CNN) based on the edge computing module for data feature extraction and compression includes the following steps: It adopts low-power edge computing modules such as NVIDIA Jetson Nano (typical power consumption ≤15W, operating temperature -40℃~85℃), integrates GPU acceleration units (such as 128-core Maxwell GPU), supports CUDA parallel computing, and meets the real-time requirements of CNN inference.

[0015] Lightweight convolutional neural network architectures such as MobileNetV3-Small are selected. Through model pruning, quantization (such as INT8 precision), and weight compression techniques, the model size is compressed from the original several GB to tens of MB, ensuring efficient operation in the limited memory of edge devices (such as 4GB LPDDR4). Before model deployment, inference acceleration is performed using optimization tools such as TensorRT to improve the number of frames processed per second (FPS≥30). The edge computing module normalizes the input data (e.g., Min-Max scaling to [0,1]), denoises (e.g., wavelet thresholding), and converts the format (e.g., RGB image to grayscale). For time-series data (e.g. vibration acceleration), it uses a sliding window method (window length 512ms, step size 100ms) to divide it into subsequences to adapt to the input dimension of CNN. CNN extracts spatial features from input data through multiple layers of convolutional kernels. For example, in the image processing branch, convolutional layers gradually extract low-level features such as the edges and textures of bridge cracks, and then use pooling layers (such as max pooling) to reduce the dimensionality and retain key feature maps. In time series data processing, 1D convolutional kernels are used to extract the time-frequency domain features of vibration signals. The feature map is compressed into a feature vector by a global average pooling layer, reducing the dimension to tens to hundreds of dimensions. Then, the feature is further compressed by dimensionality reduction algorithms such as principal component analysis (PCA) or t-SNE, while retaining more than 95% of the information entropy. The compressed feature vector is then losslessly compressed by entropy coding (such as Huffman coding), reducing the data volume to 1 / 10 to 1 / 5 of the original data. For multi-source data (such as images + time series), a feature-level fusion strategy is adopted, where feature vectors from different modalities are concatenated and input into a fully connected layer to generate fused features. The fused features need to be enhanced with the ReLU activation function to improve their non-linear expressive power and discriminative power.

[0016] The compressed feature vectors are normalized using Z-Score or Min-Max to ensure that the output value distribution meets the input requirements of the subsequent Apache Flink streaming computing framework (such as the range [-1,1]). The normalization parameters (mean, standard deviation) need to be pre-calculated and fixed in the edge computing module during system deployment.

[0017] Furthermore, after the outliers are processed in real time by the Apache Flink streaming computing framework and transmitted to the cloud, a preliminary warning is triggered based on a preset static threshold, including the following steps: After receiving the raw data stream, Flink first performs Z-Score normalization to eliminate dimensional differences. For example, the raw GNSS settlement value is converted into a standard normal distribution (mean 0, standard deviation 1) after normalization, making sensor data of different magnitudes comparable. Flink uses time windows (such as a 1-second sliding window) to perform real-time statistics on the data stream, calculating indicators such as the mean and standard deviation within the window. Based on preset static thresholds (such as a settlement threshold of ±1.5mm and an acceleration threshold of ±0.5g), it checks the data point by point within the window. If a value exceeds the threshold, it is immediately marked as an outlier and an anomaly event record is generated.

[0018] For detected outliers, Flink performs dynamic filtering to remove obviously erroneous data (such as outliers caused by sensor drift). At the same time, it fills in missing values ​​through linear interpolation or historical averages to ensure the continuity of the data stream and avoid computational interruptions caused by outliers. The preset static thresholds are based on bridge design specifications and historical monitoring data, and are divided into three levels of warning thresholds: yellow, orange, and red. For example, the settlement threshold is ±1.0 mm for the yellow level, ±1.5 mm for the orange level, and ±2.0 mm for the red level. The acceleration threshold is dynamically adjusted according to the bridge's modal frequency to ensure that it matches the vibration characteristics. When the data points processed by Flink (such as the normalized GNSS settlement) exceed the corresponding static threshold, the system immediately triggers a preliminary warning signal. The warning signal contains structured information such as timestamp, sensor type, exceedance value, and warning level to ensure traceability.

[0019] Furthermore, the VGG16-LSTM multimodal feature fusion model, after optimization with five-fold cross-validation, and combined with the DQN reinforcement learning framework to dynamically adjust the warning threshold, includes the following steps: For spatial data such as bridge crack images and settlement thermal imaging, VGG16 extracts local texture features (such as crack edges and deformation areas) through 13 convolutional layers (such as 3×3 convolutional kernels + ReLU activation) and outputs a 512-dimensional feature vector. For example, the geometric features such as the width and length of cracks on the surface of bridge piers are compressed into low-dimensional features through a global average pooling layer, preserving key spatial information. For time-series data such as settlement, vibration acceleration, temperature and humidity, LSTM captures long-term dependencies (such as periodic changes in settlement trends) through gating mechanisms (forget gate, input gate, output gate). After the input data is segmented by a sliding window (window length 30 minutes, step size 5 minutes), LSTM outputs a 256-dimensional time-series feature vector, reflecting the dynamic change pattern of settlement. A feature-level fusion approach is adopted, which concatenates the 512-dimensional spatial feature vector of VGG16 with the 256-dimensional temporal feature vector of LSTM to form a 768-dimensional fused feature vector. After fusion, the dimensionality is reduced to 128 dimensions through a fully connected layer and input into the classification layer to generate the subsidence risk probability. The fusion process retains the physical meaning of the original features (such as spatial features corresponding to image location and temporal features corresponding to time evolution).

[0020] The historical dataset (including normal / abnormal settlement samples) is randomly divided into 5 mutually exclusive subsets. Each subset is divided into training set, validation set and test set in an 8:1:1 ratio. The model is trained on 4 subsets (e.g. Adam optimizer, learning rate 0.001), and the performance is verified on the remaining 1 subset. The process is repeated 5 times to ensure that each subset is tested once. The model performance (such as accuracy, recall, and F1 score) is quantified through five-fold cross-validation. For example, the number of hidden nodes in the LSTM layer (64-256), the dropout rate (0.2-0.5), and the batch normalization parameter are adjusted to select the hyperparameter combination that performs best on the validation set. The final model must meet the following requirements: five-fold average accuracy ≥ 95% and variance between each fold ≤ 2%.

[0021] Using a fused feature vector (128 dimensions) as the core state input, supplemented by environmental parameters (such as temperature, humidity, and traffic flow) to form a state space, for example, the state vector contains key indicators such as the probability of settlement risk, peak vibration acceleration, and rate of temperature change, comprehensively depicting the current health status of the bridge. Actions are defined as adjustment strategies for warning thresholds, such as discrete actions like "increasing the threshold by 1mm", "decreasing the threshold by 0.5mm", or "keeping the threshold unchanged". The action space must cover a reasonable range of threshold adjustment (e.g., ±3mm) to ensure adaptability to different settlement risk levels. The reward function is designed based on a trade-off between early warning accuracy and false alarm rate. For example, a correct early warning (settlement exceeds the threshold and settlement actually occurs) rewards +10 points, a false alarm (settlement exceeds the threshold but does not actually occur) penalizes -5 points, and a missed warning (settlement does not exceed the threshold but actually occurs) penalizes -10 points. By maximizing the cumulative reward, DQN is guided to learn the optimal threshold adjustment strategy.

[0022] DQN uses a deep neural network to approximate the Q function. The input is a state vector and the output is the Q value of each action. During training, interaction experience (state, action, reward, and next state) is stored in an experience replay pool (capacity 10,000). Random sampling breaks the correlation of samples. The target network updates its parameters every 1,000 steps to make the Q value estimation stable.

[0023] The cloud-based decision layer receives the fused feature vectors in real time, and DQN selects the action with the highest Q value (i.e., the optimal threshold adjustment strategy) based on the current state. For example, during peak traffic hours, the system automatically lowers the threshold to improve warning sensitivity; during off-peak hours, it appropriately raises the threshold to reduce false alarms, and the adjusted threshold is fed back to the execution control layer in real time.

[0024] Furthermore, the fusion of multi-sensor data confidence levels, and the correction of the influence of temperature and humidity using an RBF neural network, includes the following steps: Real-time confidence assessment is performed on data from heterogeneous sensors such as GNSS settlement sensors, accelerometers, and temperature and humidity sensors. Assessment dimensions include: Historical calibration accuracy: The root mean square error of each sensor's measured value compared with the standard value is calculated based on the calibration data of the most recent 30 days (e.g., RMSE≤0.8mm for GNSS settlement sensor). Sensors with higher accuracy are assigned a higher confidence weight. Current operating status: The built-in self-test module monitors the health status of the sensor (e.g., a signal-to-noise ratio (SNR) of ≥30dB is normal, and <20dB is abnormal). Abnormal status triggers a reduction in confidence weight (e.g., reduced to 50% of the original weight). Environmental adaptability score: The confidence level is dynamically adjusted based on environmental parameters (such as temperature, humidity, and electromagnetic interference intensity). For example, in a low-temperature environment of -40℃, the confidence level of the temperature and humidity sensor may decrease by 10% due to the material's shrinkage effect. An adaptive weighted average algorithm is used to fuse data from multiple sensors. Let the measurement value of the i-th sensor be... The confidence weight is (satisfying ∑) =1), fusion result =∑ Calculate the weights:

[0025] A three-layer RBF neural network structure is adopted: Input layer: Contains temperature and humidity sensor data (such as temperature T, humidity H) and fused sedimentation data. and time series characteristics (such as the rate of change of temperature and humidity in the most recent hour); Hidden layer: Consists of several radial basis functions (such as Gaussian functions), with a center vector. and width parameter The typical distribution of temperature, humidity and settlement error is automatically extracted from the training data using the K-means clustering algorithm.

[0026] Output layer: Outputs the corrected settlement amount The hidden layer output is combined using linear weights w, i.e.

[0027] Where b is the bias term; Collect historical datasets, including temperature and humidity sensor data and fused settlement data. The data must cover different seasons, weather conditions (such as high temperature, high humidity, and low temperature) and traffic load scenarios to ensure the model's generalization ability. The weights w and bias b of the RBF network were trained using the least squares method, with the goal of minimizing the mean square error (MSE ≤ 0.05 mm) between the corrected settlement and the true value. 2 During training, the number of hidden layer nodes is adjusted (e.g., 50-200) using leave-one-out cross-validation to avoid overfitting.

[0028] The model performance was verified using an independent test set to ensure that the Pearson correlation coefficient between the corrected settlement and the true value was ≥0.95. After verification, the model was deployed to the cloud decision layer to receive input data and output corrected values ​​in real time. When the edge computing layer transmits the fused settlement amount When temperature and humidity data are sent to the cloud, the RBF network calculates correction values ​​in real time. For example, in high temperature and high humidity environments, if the temperature and humidity cause the GNSS settlement sensor measurement value to be 0.3 mm higher, the RBF network will output a correction value. = -0.3mm, eliminating environmental errors; The correction effect is verified by comparing the corrected settlement with independent calibration data (such as monthly manual measurements). If the correction error exceeds the threshold (e.g., |corrected value - true value| > 0.2 mm), the model retraining mechanism is triggered to update the RBF network parameters using the new data.

[0029] Corrected settlement Together with the corresponding multi-sensor confidence weights, they are fed back to the edge computing layer as inputs to algorithms such as Kalman filtering, forming a complete closed loop of "data acquisition - confidence assessment - fusion - correction - feedback".

[0030] Furthermore, the synchronous generation of SHAP value heatmaps and gradient-weighted class activation mappings to visualize key settlement factors and generate final early warning details includes the following steps: Based on a pre-trained random forest model (simulating the fusion effect of VGG16-LSTM), the SHAP algorithm is applied to calculate the contribution of each feature to the prediction result. For example, the SHAP value of the settlement amount feature to the settlement risk score is 1.2, indicating that it has a high contribution; the acceleration feature is 0.8, indicating a medium contribution. A heatmap of SHAP values ​​is generated using the seaborn library. The horizontal axis represents the feature name (such as settlement amount, acceleration X, temperature), and the vertical axis represents the absolute value of the feature contribution. The color intensity indicates the magnitude of the contribution. The heatmap visually shows that settlement amount is the main influencing factor, followed by acceleration, and temperature has a smaller impact.

[0031] Based on the last convolutional layer of the VGG16 model, a high-dimensional feature map is extracted. The gradient weighted sum of the feature map is calculated by global average pooling to generate a class activation map. The class activation map is superimposed on the original image (such as a thermal image of a bridge structure) to generate a gradient weighted class activation map. In the map, the red area represents the area that plays a key role in settlement prediction (such as cracks at the bottom of the bridge pier), and the blue area represents irrelevant areas. For example, the activation value of the crack area at the bottom of the bridge pier is high, indicating that this area is a key factor in settlement. The final warning details include a timestamp, warning level, key factor analysis, links to visualization results, and recommended measures. For example: Timestamp: 2025-11-14 12:00:00 Warning Level: Orange (Medium Risk) Key factors: High contribution from settlement (SHAP=1.2), moderate contribution from acceleration x (SHAP=0.8) Visualization results: SHAP heatmap, Grad-CAM plot Recommended measures: Inspect the bridge pier foundations and assess potential settlement issues. The warning details are transmitted to the operation and maintenance end through the cloud decision layer, and the model self-optimization mechanism is triggered at the same time. If the false alarm rate exceeds 5%, the model is automatically retrained to adjust the VGG16-LSTM model parameters or RBF neural network to correct the influence of temperature and humidity.

[0032] Furthermore, after receiving instructions from the cloud, the sub-control terminal controls the multi-segment retractable bollards to perform diversion, flow restriction, and flow interruption operations, including the following steps: The sub-control terminal is deployed at key sections of the bridge site (such as near bridge piers or ramp entrances), integrated into an industrial control box with an IP67 dustproof and waterproof rating. It contains a built-in microcontroller (such as the STM32F4 series), a speed control drive module (supporting PWM / CAN bus), and a wireless communication module (such as a 4G / 5G module). Its core function is to receive cloud commands, parse and convert them into precise control signals for the multi-segment telescopic bollards, and simultaneously provide real-time feedback of the execution status to the cloud for verification. The bollards adopt a segmented design (such as 3-5 telescopic sleeves), each segment is made of high-strength aluminum alloy, with built-in servo motors and encoders, supporting millimeter-level telescopic accuracy. The top of the bollard integrates an LED guide rod and a reflective warning flag, which use different colors (red / yellow / green) and flashing frequencies (such as 1Hz constant light / 2Hz fast flash) to map the warning level.

[0033] The sub-control terminal receives instructions from the cloud via a wireless communication module, parses the operation type and parameters, and the speed control drive module adjusts the servo motor speed via PWM signals according to the operation type in the instruction to control the extension speed of the road bollards (e.g., adjustable from 0.5m / s to 2m / s). For example, when executing the "block traffic" instruction, the road bollards extend to full height (e.g., 1.2m) at the fastest speed (e.g., 2m / s) to block the lane; when executing the "limit traffic" instruction, they extend to half height (e.g., 0.6m) at a medium speed (e.g., 1m / s) to limit the vehicle speed to 30km / h.

[0034] Each section of the bollard receives real-time feedback on its extension / retraction position from encoders to the sub-control unit, forming a closed-loop control. For example, when the first section extends to the target position, the extension of the next section is automatically triggered, ensuring synchronized overall movement. Simultaneously, the LED guide poles adjust their color and flashing frequency according to the warning level; for example, red indicates current interruption with a 2Hz rapid flash, while yellow indicates current limiting with a 1Hz constant light, enhancing the visual warning effect.

[0035] The present invention has the following beneficial effects: In this invention, a multi-source heterogeneous sensor network in the intelligent sensing layer collects settlement deformation, vibration frequency, and environmental parameters in real time. After three layers of electromagnetic shielding and anti-interference processing with silicone shock-absorbing supports, vibration errors are dynamically compensated and transmitted to the edge computing layer. The edge layer deploys a CNN for data feature extraction and compression. After outliers are processed in real time by the Apache Flink streaming computing framework, a preliminary warning is triggered by a preset static threshold and transmitted to the cloud. The cloud decision layer uses a VGG16-LSTM multimodal feature fusion model, combined with DQN reinforcement learning to dynamically adjust the warning threshold, and uses an RBF neural network to correct the influence of temperature and humidity. When the false alarm rate exceeds 5%, the model is automatically retrained. Simultaneously, a SHAP value heatmap and gradient weighted class activation mapping are generated to visualize key settlement factors and generate the final warning details. The execution control layer adjusts the speed of multi-segment telescopic bollards according to cloud commands to perform diversion, flow restriction, and flow interruption operations. The LED guide poles and reflective warning flags dynamically adjust their flashing frequency and color, and the execution status is fed back to the cloud for verification in real time, effectively solving the problem of poor adaptability of bridge settlement monitoring in extreme environments. Attached Figure Description

[0036] Figure 1 This is a system block diagram of a highway bridge settlement monitoring device proposed in this invention. Detailed Implementation

[0037] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0038] Please see Figure 1 As shown, the present invention is a highway bridge settlement monitoring device, comprising: Intelligent sensing layer: Real-time collection of settlement deformation, vibration frequency and environmental parameters through multi-source heterogeneous sensor network, dynamic compensation of vibration error after three-layer electromagnetic shielding and silicone shock-absorbing bracket anti-interference treatment, and finally transmission to edge computing layer; Edge computing layer: Convolutional neural networks (CNNs) are deployed based on edge computing modules to extract and compress data features. After outliers are processed in real time by the Apache Flink streaming computing framework, the data is transmitted to the cloud. Preliminary warning signals are triggered by preset static thresholds and sent to the cloud decision layer. Cloud-based decision layer: Receives preliminary early warning signals and structured data transmitted from the edge layer, employs a VGG16-LSTM multimodal feature fusion model, optimizes it with five-fold cross-validation, and dynamically adjusts the early warning threshold using the DQN reinforcement learning framework; it fuses confidence scores from multiple sensors, uses an RBF neural network to correct for the effects of temperature and humidity, automatically triggers model retraining when the false alarm rate exceeds 5%, and simultaneously generates SHAP value heatmaps and gradient-weighted class activation maps to visualize key settlement factors and generate final early warning details; Execution control layer: After receiving instructions from the cloud, the sub-control terminal adjusts the speed control of the multi-segment telescopic bollards to perform diversion, flow restriction and flow interruption operations. The LED guide poles and reflective warning flags on the top of the bollards dynamically adjust the flashing frequency and color according to the warning level. At the same time, the video acquisition mechanism monitors the vehicle speed in real time and feeds back the execution status to the cloud for verification.

[0039] In one embodiment, the dynamic compensation for vibration error after three-layer electromagnetic shielding and silicone shock-absorbing bracket anti-interference treatment includes the following steps: The three-layer electromagnetic shielding adopts a composite structure of "copper foil layer (inner layer) - aluminum foil layer (middle layer) - conductive cloth layer (outer layer)". The copper foil layer absorbs high-frequency electromagnetic noise (such as high-voltage line radiation), the aluminum foil layer shields medium-frequency interference (such as lightning pulse), and the conductive cloth layer suppresses low-frequency electrostatic coupling, forming a gradient electromagnetic protection to ensure that the original sensor data is not interfered with in a strong electromagnetic environment. The silicone vibration damping bracket uses high-damping silicone material and absorbs the vibration energy of the bridge through a three-stage vibration damping structure of "base-bracket-sensor". The bracket design conforms to the dynamic modal frequency of the bridge, avoids resonance with the natural vibration frequency of the bridge, and reduces the direct impact of mechanical vibration on the sensor measurement. Kalman filtering dynamically corrects measurement errors caused by vibration through a two-stage recursive calculation of "prediction-update". Define the system state vector ,in, For settlement, For the settling rate, For vibration acceleration, the state transition equation is constructed based on Newton's laws of motion:

[0040] Where F is the state transition matrix (describing the evolution of the state over time), and B is the control input matrix (external force). ~N(0,Q) represents process noise (reflecting model uncertainty); Based on the optimal state estimate from the previous moment And error covariance matrix Calculate the prior state estimate at the current time:

[0041] Update prior error covariance:

[0042] Combine sensor measurements (e.g., GNSS settlement, accelerometer output), construct the observation equations:

[0043] Where H is the observation matrix (mapping the state to the measurement space). ~N(0,R) represents the measurement noise (reflecting sensor accuracy); Calculate Kalman gain Balance the weights of predicted and measured values:

[0044] Where R is the measurement noise covariance matrix; By combining predicted and measured values, the posterior optimal state estimate is obtained:

[0045] Updated posterior error covariance:

[0046] Where I is the identity matrix.

[0047] In one embodiment, the deployment of a convolutional neural network (CNN) based on an edge computing module for data feature extraction and compression includes the following steps: It adopts low-power edge computing modules such as NVIDIA Jetson Nano (typical power consumption ≤15W, operating temperature -40℃~85℃), integrates GPU acceleration units (such as 128-core Maxwell GPU), supports CUDA parallel computing, and meets the real-time requirements of CNN inference.

[0048] Lightweight convolutional neural network architectures such as MobileNetV3-Small are selected. Through model pruning, quantization (such as INT8 precision), and weight compression techniques, the model size is compressed from the original several GB to tens of MB, ensuring efficient operation in the limited memory of edge devices (such as 4GB LPDDR4). Before model deployment, inference acceleration is performed using optimization tools such as TensorRT to improve the number of frames processed per second (FPS≥30). The edge computing module normalizes the input data (e.g., Min-Max scaling to [0,1]), denoises (e.g., wavelet thresholding), and converts the format (e.g., RGB image to grayscale). For time-series data (e.g. vibration acceleration), it uses a sliding window method (window length 512ms, step size 100ms) to divide it into subsequences to adapt to the input dimension of CNN. CNN extracts spatial features from input data through multiple layers of convolutional kernels. For example, in the image processing branch, convolutional layers gradually extract low-level features such as the edges and textures of bridge cracks, and then use pooling layers (such as max pooling) to reduce the dimensionality and retain key feature maps. In time series data processing, 1D convolutional kernels are used to extract the time-frequency domain features of vibration signals. The feature map is compressed into a feature vector by a global average pooling layer, reducing the dimension to tens to hundreds of dimensions. Then, the feature is further compressed by dimensionality reduction algorithms such as principal component analysis (PCA) or t-SNE, while retaining more than 95% of the information entropy. The compressed feature vector is then losslessly compressed by entropy coding (such as Huffman coding), reducing the data volume to 1 / 10 to 1 / 5 of the original data. For multi-source data (such as images + time series), a feature-level fusion strategy is adopted, where feature vectors from different modalities are concatenated and input into a fully connected layer to generate fused features. The fused features need to be enhanced with the ReLU activation function to improve their non-linear expressive power and discriminative power.

[0049] The compressed feature vectors are normalized using Z-Score or Min-Max to ensure that the output value distribution meets the input requirements of the subsequent Apache Flink streaming computing framework (such as the range [-1,1]). The normalization parameters (mean, standard deviation) need to be pre-calculated and fixed in the edge computing module during system deployment.

[0050] In one embodiment, the abnormal values ​​processed in real time by the Apache Flink streaming computing framework are transmitted to the cloud, and a preliminary warning is triggered by a preset static threshold, including the following steps: After receiving the raw data stream, Flink first performs Z-Score normalization to eliminate dimensional differences. For example, the raw GNSS settlement value is converted into a standard normal distribution (mean 0, standard deviation 1) after normalization, making sensor data of different magnitudes comparable. Flink uses time windows (such as a 1-second sliding window) to perform real-time statistics on the data stream, calculating indicators such as the mean and standard deviation within the window. Based on preset static thresholds (such as a settlement threshold of ±1.5mm and an acceleration threshold of ±0.5g), it checks the data point by point within the window. If a value exceeds the threshold, it is immediately marked as an outlier and an anomaly event record is generated.

[0051] For detected outliers, Flink performs dynamic filtering to remove obviously erroneous data (such as outliers caused by sensor drift). At the same time, it fills in missing values ​​through linear interpolation or historical averages to ensure the continuity of the data stream and avoid computational interruptions caused by outliers. The preset static thresholds are based on bridge design specifications and historical monitoring data, and are divided into three levels of warning thresholds: yellow, orange, and red. For example, the settlement threshold is ±1.0 mm for the yellow level, ±1.5 mm for the orange level, and ±2.0 mm for the red level. The acceleration threshold is dynamically adjusted according to the bridge's modal frequency to ensure that it matches the vibration characteristics. When the data points processed by Flink (such as the normalized GNSS settlement) exceed the corresponding static threshold, the system immediately triggers a preliminary warning signal. The warning signal contains structured information such as timestamp, sensor type, exceedance value, and warning level to ensure traceability.

[0052] In one embodiment, the VGG16-LSTM multimodal feature fusion model, optimized with five-fold cross-validation, and dynamically adjusted for alert thresholds using the DQN reinforcement learning framework, includes the following steps: For spatial data such as bridge crack images and settlement thermal imaging, VGG16 extracts local texture features (such as crack edges and deformation areas) through 13 convolutional layers (such as 3×3 convolutional kernels + ReLU activation) and outputs a 512-dimensional feature vector. For example, the geometric features such as the width and length of cracks on the surface of bridge piers are compressed into low-dimensional features through a global average pooling layer, preserving key spatial information. For time-series data such as settlement, vibration acceleration, temperature and humidity, LSTM captures long-term dependencies (such as periodic changes in settlement trends) through gating mechanisms (forget gate, input gate, output gate). After the input data is segmented by a sliding window (window length 30 minutes, step size 5 minutes), LSTM outputs a 256-dimensional time-series feature vector, reflecting the dynamic change pattern of settlement. A feature-level fusion approach is adopted, which concatenates the 512-dimensional spatial feature vector of VGG16 with the 256-dimensional temporal feature vector of LSTM to form a 768-dimensional fused feature vector. After fusion, the dimensionality is reduced to 128 dimensions through a fully connected layer and input into the classification layer to generate the subsidence risk probability. The fusion process retains the physical meaning of the original features (such as spatial features corresponding to image location and temporal features corresponding to time evolution).

[0053] The historical dataset (including normal / abnormal settlement samples) is randomly divided into 5 mutually exclusive subsets. Each subset is divided into training set, validation set and test set in an 8:1:1 ratio. The model is trained on 4 subsets (e.g. Adam optimizer, learning rate 0.001), and the performance is verified on the remaining 1 subset. The process is repeated 5 times to ensure that each subset is tested once. The model performance (such as accuracy, recall, and F1 score) is quantified through five-fold cross-validation. For example, the number of hidden nodes in the LSTM layer (64-256), the dropout rate (0.2-0.5), and the batch normalization parameter are adjusted to select the hyperparameter combination that performs best on the validation set. The final model must meet the following requirements: five-fold average accuracy ≥ 95% and variance between each fold ≤ 2%.

[0054] Using a fused feature vector (128 dimensions) as the core state input, supplemented by environmental parameters (such as temperature, humidity, and traffic flow) to form a state space, for example, the state vector contains key indicators such as the probability of settlement risk, peak vibration acceleration, and rate of temperature change, comprehensively depicting the current health status of the bridge. Actions are defined as adjustment strategies for warning thresholds, such as discrete actions like "increasing the threshold by 1mm", "decreasing the threshold by 0.5mm", or "keeping the threshold unchanged". The action space must cover a reasonable range of threshold adjustment (e.g., ±3mm) to ensure adaptability to different settlement risk levels. The reward function is designed based on a trade-off between early warning accuracy and false alarm rate. For example, a correct early warning (settlement exceeds the threshold and settlement actually occurs) rewards +10 points, a false alarm (settlement exceeds the threshold but does not actually occur) penalizes -5 points, and a missed warning (settlement does not exceed the threshold but actually occurs) penalizes -10 points. By maximizing the cumulative reward, DQN is guided to learn the optimal threshold adjustment strategy.

[0055] DQN uses a deep neural network to approximate the Q function. The input is a state vector and the output is the Q value of each action. During training, interaction experience (state, action, reward, and next state) is stored in an experience replay pool (capacity 10,000). Random sampling breaks the correlation of samples. The target network updates its parameters every 1,000 steps to make the Q value estimation stable.

[0056] The cloud-based decision layer receives the fused feature vectors in real time, and DQN selects the action with the highest Q value (i.e., the optimal threshold adjustment strategy) based on the current state. For example, during peak traffic hours, the system automatically lowers the threshold to improve warning sensitivity; during off-peak hours, it appropriately raises the threshold to reduce false alarms, and the adjusted threshold is fed back to the execution control layer in real time.

[0057] In one embodiment, fusing multi-sensor data confidence and using an RBF neural network to correct for the effects of temperature and humidity includes the following steps: Real-time confidence assessment is performed on data from heterogeneous sensors such as GNSS settlement sensors, accelerometers, and temperature and humidity sensors. Assessment dimensions include: Historical calibration accuracy: The root mean square error of each sensor's measured value compared with the standard value is calculated based on the calibration data of the most recent 30 days (e.g., RMSE≤0.8mm for GNSS settlement sensor). Sensors with higher accuracy are assigned a higher confidence weight. Current operating status: The built-in self-test module monitors the health status of the sensor (e.g., a signal-to-noise ratio (SNR) of ≥30dB is normal, and <20dB is abnormal). Abnormal status triggers a reduction in confidence weight (e.g., reduced to 50% of the original weight). Environmental adaptability score: The confidence level is dynamically adjusted based on environmental parameters (such as temperature, humidity, and electromagnetic interference intensity). For example, in a low-temperature environment of -40℃, the confidence level of the temperature and humidity sensor may decrease by 10% due to the material's shrinkage effect. An adaptive weighted average algorithm is used to fuse data from multiple sensors. Let the measurement value of the i-th sensor be... The confidence weight is (satisfying ∑) =1), fusion result =∑ Calculate the weights:

[0058] A three-layer RBF neural network structure is adopted: Input layer: Contains temperature and humidity sensor data (such as temperature T, humidity H) and fused sedimentation data. and time series characteristics (such as the rate of change of temperature and humidity in the most recent hour); Hidden layer: Consists of several radial basis functions (such as Gaussian functions), with a center vector. and width parameter The typical distribution of temperature, humidity and settlement error is automatically extracted from the training data using the K-means clustering algorithm.

[0059] Output layer: Outputs the corrected settlement amount The hidden layer output is combined using linear weights w, i.e.

[0060] Where b is the bias term; Collect historical datasets, including temperature and humidity sensor data and fused settlement data. The data must cover different seasons, weather conditions (such as high temperature, high humidity, and low temperature) and traffic load scenarios to ensure the model's generalization ability. The weights w and bias b of the RBF network were trained using the least squares method, with the goal of minimizing the mean square error (MSE ≤ 0.05 mm) between the corrected settlement and the true value. 2 During training, the number of hidden layer nodes is adjusted (e.g., 50-200) using leave-one-out cross-validation to avoid overfitting.

[0061] The model performance was verified using an independent test set to ensure that the Pearson correlation coefficient between the corrected settlement and the true value was ≥0.95. After verification, the model was deployed to the cloud decision layer to receive input data and output corrected values ​​in real time. When the edge computing layer transmits the fused settlement amount When temperature and humidity data are sent to the cloud, the RBF network calculates correction values ​​in real time. For example, in high temperature and high humidity environments, if the temperature and humidity cause the GNSS settlement sensor measurement value to be 0.3 mm higher, the RBF network will output a correction value. = -0.3mm, eliminating environmental errors; The correction effect is verified by comparing the corrected settlement with independent calibration data (such as monthly manual measurements). If the correction error exceeds the threshold (e.g., |corrected value - true value| > 0.2 mm), the model retraining mechanism is triggered to update the RBF network parameters using the new data.

[0062] Corrected settlement Together with the corresponding multi-sensor confidence weights, they are fed back to the edge computing layer as inputs to algorithms such as Kalman filtering, forming a complete closed loop of "data acquisition - confidence assessment - fusion - correction - feedback".

[0063] In one embodiment, the synchronous generation of SHAP value heatmaps and gradient-weighted class activation mappings to visualize key settlement factors and generate final early warning details includes the following steps: Based on a pre-trained random forest model (simulating the fusion effect of VGG16-LSTM), the SHAP algorithm is applied to calculate the contribution of each feature to the prediction result. For example, the SHAP value of the settlement amount feature to the settlement risk score is 1.2, indicating that it has a high contribution; the acceleration feature is 0.8, indicating a medium contribution. A heatmap of SHAP values ​​is generated using the seaborn library. The horizontal axis represents the feature name (such as settlement amount, acceleration X, temperature), and the vertical axis represents the absolute value of the feature contribution. The color intensity indicates the magnitude of the contribution. The heatmap visually shows that settlement amount is the main influencing factor, followed by acceleration, and temperature has a smaller impact.

[0064] Based on the last convolutional layer of the VGG16 model, a high-dimensional feature map is extracted. The gradient weighted sum of the feature map is calculated by global average pooling to generate a class activation map. The class activation map is superimposed on the original image (such as a thermal image of a bridge structure) to generate a gradient weighted class activation map. In the map, the red area represents the area that plays a key role in settlement prediction (such as cracks at the bottom of the bridge pier), and the blue area represents irrelevant areas. For example, the activation value of the crack area at the bottom of the bridge pier is high, indicating that this area is a key factor in settlement. The final warning details include a timestamp, warning level, key factor analysis, links to visualization results, and recommended measures. For example: Timestamp: 2025-11-14 12:00:00 Warning Level: Orange (Medium Risk) Key factors: High contribution from settlement (SHAP=1.2), moderate contribution from acceleration x (SHAP=0.8) Visualization results: SHAP heatmap, Grad-CAM plot Recommended measures: Inspect the bridge pier foundations and assess potential settlement issues. The warning details are transmitted to the operation and maintenance end through the cloud decision layer, and the model self-optimization mechanism is triggered at the same time. If the false alarm rate exceeds 5%, the model is automatically retrained to adjust the VGG16-LSTM model parameters or RBF neural network to correct the influence of temperature and humidity.

[0065] In one embodiment, after receiving instructions from the cloud, the sub-control terminal controls the multi-segment retractable bollards to perform diversion, flow restriction, and flow interruption operations, including the following steps: The sub-control terminal is deployed at key sections of the bridge site (such as near bridge piers or ramp entrances), integrated into an industrial control box with an IP67 dustproof and waterproof rating. It contains a built-in microcontroller (such as the STM32F4 series), a speed control drive module (supporting PWM / CAN bus), and a wireless communication module (such as a 4G / 5G module). Its core function is to receive cloud commands, parse and convert them into precise control signals for the multi-segment telescopic bollards, and simultaneously provide real-time feedback of the execution status to the cloud for verification. The bollards adopt a segmented design (such as 3-5 telescopic sleeves), each segment is made of high-strength aluminum alloy, with built-in servo motors and encoders, supporting millimeter-level telescopic accuracy. The top of the bollard integrates an LED guide rod and a reflective warning flag, which use different colors (red / yellow / green) and flashing frequencies (such as 1Hz constant light / 2Hz fast flash) to map the warning level.

[0066] The sub-control terminal receives instructions from the cloud via a wireless communication module, parses the operation type and parameters, and the speed control drive module adjusts the servo motor speed via PWM signals according to the operation type in the instruction to control the extension speed of the road bollards (e.g., adjustable from 0.5m / s to 2m / s). For example, when executing the "block traffic" instruction, the road bollards extend to full height (e.g., 1.2m) at the fastest speed (e.g., 2m / s) to block the lane; when executing the "limit traffic" instruction, they extend to half height (e.g., 0.6m) at a medium speed (e.g., 1m / s) to limit the vehicle speed to 30km / h.

[0067] Each section of the bollard receives real-time feedback on its extension / retraction position from encoders to the sub-control unit, forming a closed-loop control. For example, when the first section extends to the target position, the extension of the next section is automatically triggered, ensuring synchronized overall movement. Simultaneously, the LED guide poles adjust their color and flashing frequency according to the warning level; for example, red indicates current interruption with a 2Hz rapid flash, while yellow indicates current limiting with a 1Hz constant light, enhancing the visual warning effect.

[0068] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and variations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A highway bridge settlement monitoring device, characterized in that, include: Intelligent sensing layer: Real-time collection of settlement deformation, vibration frequency and environmental parameters through multi-source heterogeneous sensor network, dynamic compensation of vibration error after three-layer electromagnetic shielding and silicone shock-absorbing bracket anti-interference treatment, and finally transmission to edge computing layer; Edge computing layer: Convolutional neural networks (CNNs) are deployed based on edge computing modules to extract and compress data features. After outliers are processed in real time by the Apache Flink streaming computing framework, the data is transmitted to the cloud. Preliminary warning signals are triggered by preset static thresholds and sent to the cloud decision layer. Cloud-based decision layer: Receives preliminary early warning signals and structured data transmitted from the edge layer, employs a VGG16-LSTM multimodal feature fusion model, optimizes it with five-fold cross-validation, and dynamically adjusts the early warning threshold using the DQN reinforcement learning framework; it fuses confidence scores from multiple sensors, uses an RBF neural network to correct for the effects of temperature and humidity, automatically triggers model retraining when the false alarm rate exceeds 5%, and simultaneously generates SHAP value heatmaps and gradient-weighted class activation maps to visualize key settlement factors and generate final early warning details; Execution control layer: After receiving instructions from the cloud, the sub-control terminal adjusts the speed control of the multi-segment telescopic bollards to perform diversion, flow restriction and flow interruption operations. The LED guide poles and reflective warning flags on the top of the bollards dynamically adjust the flashing frequency and color according to the warning level. At the same time, the video acquisition mechanism monitors the vehicle speed in real time and feeds back the execution status to the cloud for verification.

2. The highway bridge settlement monitoring device according to claim 1, characterized in that, The dynamic compensation for vibration error after three-layer electromagnetic shielding and silicone shock-absorbing bracket anti-interference treatment includes the following steps: The three-layer electromagnetic shielding employs a composite structure consisting of an inner copper foil layer, a middle aluminum foil layer, and an outer conductive cloth layer. The copper foil layer absorbs high-frequency electromagnetic noise, the aluminum foil layer shields mid-frequency interference, and the conductive cloth layer suppresses low-frequency electrostatic coupling, forming a gradient electromagnetic protection. The silicone vibration damping bracket uses high-damping silicone material and absorbs bridge vibration energy through a three-stage damping structure of base-bracket-sensor. The bracket design conforms to the bridge's dynamic modal frequencies, avoiding resonance with the bridge's natural vibration frequencies. Kalman filtering dynamically corrects measurement errors caused by vibration through a two-stage recursive calculation of prediction and update.

3. The highway bridge settlement monitoring device according to claim 1, characterized in that, The deployment of a convolutional neural network (CNN) based on an edge computing module for data feature extraction and compression includes the following steps: The edge computing module normalizes, denoises, and converts the format of the input data. For time-series data, it uses the sliding window method to segment it into subsequences to adapt to the input dimension of the CNN. The CNN extracts spatial features from the input data through multiple convolutional kernels. The feature map is compressed into a feature vector through a global average pooling layer, and then the features are further compressed while retaining more than 95% of the information entropy. The compressed feature vector is then losslessly compressed through entropy encoding. For multi-source data, a feature-level fusion strategy is adopted, where feature vectors from different modalities are concatenated and then input into a fully connected layer to generate fused features.

4. The highway bridge settlement monitoring device according to claim 1, characterized in that, The abnormal values, after being processed in real time by the Apache Flink streaming computing framework, are transmitted to the cloud, and a preliminary warning is triggered based on a preset static threshold, including the following steps: After receiving the raw data stream, Flink uses a time window to perform real-time statistics on the data stream, calculating indicators such as mean and standard deviation within the window. Based on a preset static threshold, it checks the data point by point within the window. If a value exceeds the threshold, it is immediately marked as an outlier and an anomaly event record is generated. For detected outliers, Flink performs dynamic filtering to remove obviously erroneous data and fills in missing values ​​using linear interpolation or historical averages. The preset static thresholds are based on bridge design specifications and historical monitoring data and are divided into three levels of warning thresholds: yellow, orange, and red. When the data points processed by Flink exceed the corresponding level of static threshold, the system immediately triggers a preliminary warning signal. The warning signal contains structured information such as timestamp, sensor type, out-of-limit value, and warning level.

5. A highway bridge settlement monitoring device according to claim 1, characterized in that, The VGG16-LSTM multimodal feature fusion model, after optimization with five-fold cross-validation, is combined with the DQN reinforcement learning framework to dynamically adjust the warning threshold, including the following steps: For spatial data, VGG16 extracts local texture features through 13 convolutional layers and outputs a 512-dimensional feature vector. For temporal data, LSTM captures long-term dependencies through a gating mechanism. After the input data is segmented by a sliding window, LSTM outputs a 256-dimensional temporal feature vector, reflecting the dynamic change pattern of settlement. A feature-level fusion method is adopted to concatenate the 512-dimensional spatial feature vector of VGG16 with the 256-dimensional temporal feature vector of LSTM to form a 768-dimensional fused feature vector. After fusion, the dimensionality is reduced to 128 dimensions through a fully connected layer and then input into the classification layer to generate the subsidence risk probability. The fusion process retains the physical meaning of the original features. The historical dataset is randomly divided into 5 mutually exclusive subsets. Each subset is divided into a training set, a validation set, and a test set in an 8:1:1 ratio. The model is trained on 4 subsets and its performance is validated on the remaining subset. This process is repeated 5 times to ensure that each subset is tested once. The model performance is quantified by five-fold cross-validation. The fused feature vector is used as the core state input, supplemented by environmental parameters to form the state space. The action is defined as the adjustment strategy of the warning threshold. The action space needs to cover a reasonable range of threshold adjustment to adapt to different settlement risk levels. The reward function is designed based on a trade-off between early warning accuracy and false alarm rate. By maximizing the cumulative reward, it guides DQN to learn the optimal threshold adjustment strategy. DQN uses a deep neural network to approximate the Q function. The input is a state vector and the output is the Q value of each action. During training, interaction experience is stored in an experience replay pool. The cloud decision layer receives the fused feature vector in real time. DQN selects the action with the largest Q value based on the current state.

6. The highway bridge settlement monitoring device according to claim 1, characterized in that, The confidence level of the fused multi-sensor data, using an RBF neural network to correct for the effects of temperature and humidity, includes the following steps: Real-time confidence assessment of heterogeneous sensor data, including the following dimensions: Historical calibration accuracy: The root mean square error of each sensor's measured value compared to the standard value is calculated based on the most recent calibration data. Sensors with higher accuracy are assigned a higher confidence weight. Current operating status: The built-in self-test module monitors the health status of the sensors, and abnormal status triggers a reduction in confidence weight; Environmental adaptability score: Confidence level is dynamically adjusted based on environmental parameters; An adaptive weighted average algorithm is used to fuse multi-sensor data. A three-layer RBF neural network structure consisting of an input layer, a hidden layer, and an output layer is adopted. Historical datasets are collected, including temperature and humidity sensor data, fused settlement and corresponding ground values. The data covers different seasons, weather conditions and traffic load scenarios. The least squares method is used to train the weights and biases of the RBF network to minimize the mean square error between the corrected settlement and the true value. During training, cross-validation is used to adjust the number of hidden layer nodes. The model performance is verified using an independent test set. After verification, the model is deployed to the cloud decision layer to receive input data and output corrected values ​​in real time. When the edge computing layer transmits the fused settlement and temperature and humidity data to the cloud, the RBF network calculates the correction value in real time. The correction effect is verified by comparing the corrected settlement with the independent calibration data. If the correction error exceeds the threshold, the model retraining mechanism is triggered, and the RBF network parameters are updated using the new data. The corrected settlement and the corresponding multi-sensor confidence weights are fed back to the edge computing layer as input to the Kalman filter.

7. A highway bridge settlement monitoring device according to claim 1, characterized in that, The process of simultaneously generating a heatmap of SHAP values ​​and a gradient-weighted class activation mapping to visualize key settlement factors and generate final early warning details includes the following steps: Based on a random forest model that simulates the fusion effect of VGG16-LSTM, the SHAP algorithm is applied to calculate the contribution of each feature to the prediction result, generating a SHAP value heatmap. The horizontal axis represents the feature name, the vertical axis represents the absolute value of the feature contribution, and the color intensity indicates the magnitude of the contribution. The heatmap intuitively shows that the sedimentation is the main influencing factor, followed by acceleration, while temperature has a smaller impact. Based on the last convolutional layer of the VGG16 model, a high-dimensional feature map is extracted. The gradient weighted sum of the feature map is calculated through global average pooling to generate a class activation map. The class activation map is then superimposed on the original image to generate a gradient-weighted class activation map. In the map, the red area represents the region that plays a key role in settlement prediction, and the blue area represents the irrelevant region. The final warning details include a timestamp, warning level, key factor analysis, visualization results link, and suggested measures. The warning details are transmitted to the operation and maintenance end through the cloud decision layer. If the false alarm rate exceeds 5%, model retraining is automatically triggered.

8. A highway bridge settlement monitoring device according to claim 1, characterized in that, After receiving instructions from the cloud, the sub-control terminal controls the multi-segment retractable bollards to perform diversion, flow restriction, and flow interruption operations, including the following steps: The bollards adopt a segmented design, with each segment made of high-strength aluminum alloy and equipped with a built-in servo motor and encoder. The top of the bollard integrates an LED guide rod and a reflective warning flag, which use different colors and flashing frequencies to map the warning level. The sub-control terminal receives cloud commands through a wireless communication module, analyzes the operation type and parameters, and the speed control drive module adjusts the servo motor speed according to the operation type in the command through a PWM signal to control the bollard's extension and retraction speed. Each segment of the bollard provides real-time feedback of its extension and retraction position to the sub-control terminal through the encoder. At the same time, the LED guide rod adjusts its color and flashing frequency according to the warning level.

Images