Self-adaptive multi-source fusion roadbed intelligent monitoring and dynamic regulation and control system and method

Abstract

The invention specifically discloses a self-adaptive multi-source fusion roadbed intelligent monitoring and dynamic regulation and control system and method, and relates to the technical field of civil engineering monitoring. The system comprises a differential sensing unit, a data transmission and processing unit, a hybrid model prediction unit and a dynamic regulation and control unit, the differential sensing unit acquires multi-dimensional data through the composite anchoring structure, the magnetic suction follow-up module and the optical fiber monitoring network; the data transmission and processing unit adopts a LoRaWAN protocol to transmit data and preprocess the data; the hybrid model prediction unit predicts a settlement trend and a risk level based on an STGCN-MHSA hybrid model and an MOPSO algorithm; and the dynamic regulation and control unit performs corresponding regulation and control according to the prediction result. The adaptive multi-source fusion roadbed intelligent monitoring and dynamic regulation and control system and method provided by the invention can realize high-precision monitoring and prediction, have an active regulation and control capability, and are suitable for roadbed health management in a complex scene.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of civil engineering monitoring, in particular to a self-adaptive multi-source fusion roadbed intelligent monitoring and dynamic regulation system and method. BACKGROUND

[0002] As the core load-bearing structure in road engineering, the roadbed will be disturbed by various natural environments in daily use, and its stability directly determines the safety and service life of the road. Especially in special environments such as highways and high fill roadbeds, due to the changeable environmental load and complex geological conditions, the roadbed is prone to uneven settlement, leading to serious problems such as road surface cracking and even structural instability. Therefore, it is of great significance to monitor and predict the settlement of the roadbed in real time with high precision and timely regulation to prevent engineering disasters and reduce maintenance costs.

[0003] At present, roadbed settlement monitoring technology mainly includes physical sensing devices and data analysis models. Traditional physical monitoring devices rely on pre-embedded cement blocks or rigid structures (such as cone pipes and spiral blades), which have a complicated construction process and are prone to displacement of the reference point due to precipitation or freezing-thawing in soft soil or frozen soil, affecting the monitoring accuracy. At the same time, existing follow-up monitoring structures lack adaptability to multi-dimensional settlement and cannot accurately capture tilting or local deformation. For monitoring data, it is mostly limited to reporting and does not integrate active regulation modules. In the face of sudden settlement or extreme environments (such as freezing-thawing cycles), it cannot delay the settlement process through dynamic adjustment of load distribution or temperature control, resulting in high passive maintenance costs. From the perspective of data analysis models, although the CNN-BiLSTM architecture used in current mainstream prediction models can extract time sequence features, it does not fully utilize the topological relationship of the roadbed sensor network and lacks sufficient spatio-temporal feature fusion capability. In addition, the use of genetic algorithm (GA) for hyperparameter optimization is prone to local optimization and slow convergence, affecting the efficiency of model deployment. SUMMARY

[0004] The purpose of the present application is to provide a self-adaptive multi-source fusion roadbed intelligent monitoring and dynamic regulation system and method, which significantly improves the problems of traditional rigid sensing devices that cannot adapt to multi-dimensional deformation monitoring in complex geological environments, traditional roadbed settlement prediction models that have low efficiency of hyperparameter optimization and insufficient spatio-temporal feature fusion, and traditional roadbed settlement monitoring systems that lack regulation functions and cannot respond to settlement risks in real time.

[0005] To achieve the above purpose, the present application provides a self-adaptive multi-source fusion roadbed intelligent monitoring and dynamic regulation system, which comprises: The differential sensing unit comprises a composite anchoring structure, a magnetic attraction following module and an optical fiber monitoring network. The composite anchoring structure comprises a tapered pipe, the outer wall of the tapered pipe is provided with staggered spiral blades, and the edge of the blade is integrated with a micro pressure sensor. The magnetic attraction following module is composed of a magnetic sleeve and an elastic plate, the sleeve is provided with a Hall sensor, the magnetic attraction force is used to dynamically adsorb a signal receiver at the end of a flexible pipeline, and the elastic plate is provided with a shape memory alloy (SMA). The optical fiber monitoring network is provided with distributed fiber Bragg grating (FBG) sensors in layers along the roadbed to collect strain and temperature data in real time. The data transmission and processing unit uses a LoRaWAN wireless transmission protocol to upload multi-source data to an edge computing node, and a data fusion layer in the edge computing node performs spatio-temporal alignment and outlier filtering on the multi-source data to generate a standardized feature matrix. The hybrid model prediction unit constructs an STGCN-MHSA hybrid model based on an improved spatio-temporal graph convolution network (STGCN) and a multi-head attention mechanism (MHSA), optimizes the hyperparameters of the STGCN-MHSA hybrid model using a multi-objective particle swarm optimization (MOPSO) algorithm, inputs the standardized feature matrix, and predicts the settlement trend and risk level. The dynamic regulation unit comprises a hydraulic support system, a temperature control module and a blockchain storage module. The hydraulic support system is used to adjust the local load distribution of the roadbed, the temperature control module is used to start a geothermal cycle in a permafrost region, and the blockchain storage module is used to upload all operation records to the blockchain for storage.

[0006] Preferably, in the differential sensing unit, the deformation amount of the shape memory alloy (SMA) is The calculation formula is as follows: ; Wherein, α is the thermal expansion coefficient, β is the stress response coefficient, σ is the stress of the soil, is the real-time monitored soil temperature (unit: ℃), which is collected by the FBG optical fiber network or the temperature sensor, is the reference temperature of the SMA (unit: ℃), i.e. the deformation reference temperature (such as the phase transition starting temperature or the designed initial temperature).

[0007] Preferably, in the hybrid model prediction unit, the STGCN-MHSA hybrid model encodes the roadbed topology structure as graph data, extracts local features through spatio-temporal convolution, and uses multi-head attention to weight key nodes.

[0008] Preferably, in the hybrid model prediction unit, the multi-objective particle swarm optimization (MOPSO) algorithm is set to a double fitness function to simultaneously optimize the prediction error (MAE) and the model inference speed (FPS) to generate a Pareto optimal solution set.

[0009] Preferably, in the dynamic regulation module, the temperature control module starts the geothermal cycle in the frozen soil area to maintain the roadbed temperature at ≥0℃.

[0010] The application also provides a self-adaptive multi-source fusion roadbed intelligent monitoring and dynamic regulation method for realizing the self-adaptive multi-source fusion roadbed intelligent monitoring and dynamic regulation system. Step S1, real-time collection of pressure P, displacement δ, strain ε and temperature H through the micro pressure sensor on the conical pipe spiral blade, the Hall sensor of the magnetic attraction servo module and the distributed FBG optical fiber network, and construction of a space-time alignment matrix X E R N×T×C wherein N is the number of sensor nodes, T is the length of the time window, C is the feature dimension; Step S2, DBSCAN clustering algorithm is used to eliminate outliers and perform normalization processing; Step S3, the preprocessed data are input into the STGCN-MHSA hybrid model to output settlement prediction value Y and risk level R; Step S4, according to the prediction result of the STGCN-MHSA hybrid model, the hydraulic support system is adjusted to regulate the local load distribution of the roadbed or the temperature control module is started to inhibit the freeze-thaw settlement.

[0011] Preferably, in step S2, the normalization formula is as follows: ; wherein, is the element value of the original feature tensor, is the element value of the normalized feature tensor, , are the mean and standard deviation of the first c dimensional feature, respectively.

[0012] Preferably, in step S3, the calculation process of the STGCN-MHSA hybrid model is as follows: Spatial graph convolution: Define the sensor topology graph G = (V, E) , the adjacency matrix A is constructed by threshold truncation method, and the formula is as follows: ; wherein, V is the vertex set (Vertices), i.e. the sensor node set, each node represents a physical monitoring point; E is the edge set (Edges), which shows the connection relationship between nodes, representing the spatial correlation between sensors; is the definition of node edges (only distance The nodes of the graph are connected by edges, is a distance threshold, unit m, is a node is the Euclidean distance between the nodes, is a distance threshold; The Chebyshev polynomial approximation is adopted, and the formula is as follows: ; Wherein, is a spatial feature matrix, is order Chebyshev polynomial, is a scaled Laplace matrix, is a trainable parameter, is the order of Chebyshev polynomial (control the range of receptive field); Temporal expansion convolution: The convolution kernel with expansion rate is used to extract the time sequence feature, and the formula is as follows: ; Wherein, is a time feature extraction module, is a time offset, is the size of the time convolution kernel, is the convolution kernel weight, is the spatial correlation feature matrix of the sensor node extracted by the spatial graph convolution module; Multi-head self-attention weighting: Split into heads, and the formula is as follows: ; ; ; ; Wherein, is the output of the th attention head, , , is the query, key, and value matrix, is the transpose of the key matrix, which is used to calculate the query-key similarity, is the key vector dimension (scaling factor is used to stabilize the gradient), , , is the attention calculation formula, , , is the projection matrix; The final result output is calculated, and the formula is as follows: ; Wherein, is the fusion feature, is the output of the th attention head, is a linear projection matrix (output weight), is the output of splicing multiple attention heads; The settlement result prediction is calculated, and the formula is as follows: ; Wherein, is the output of the th attention head, is a tensor flattening operation (input pre-processing before the fully connected layer), , is a fully connected layer parameter, is a Sigmoid activation function.

[0013] Preferably, in step S3, the STGCN-MHSA hybrid model adopts an NSGA-II double fitness function optimization model parameter, and the formula is as follows: ; Wherein, is the mean absolute error, is the reciprocal of the model inference speed, is the true settlement value (mm, from historical monitoring or field measurement), is the model inference speed (frame / s), is the mean absolute error.

[0014] Preferably, in step S4, according to the settlement prediction result, the dynamic regulation and execution of the hydraulic support system is triggered, specifically: When , the hydraulic support system is triggered: ; Wherein is a PID coefficient, is a differential gain coefficient (kPa·s / mm, inhibiting the settlement change rate), is a settlement risk threshold (mm), is the load amount of the hydraulic system adjustment (kPa, used to compensate for uneven settlement).

[0015] Therefore, the present application proposes an adaptive multi-source fusion roadbed intelligent monitoring and dynamic regulation system and method, which has the following beneficial effects: (1) The application realizes omnidirectional and high-precision monitoring of the roadbed by the cooperation of the composite anchoring structure, the magnetic attraction following module and the optical fiber monitoring network. According to actual test, the monitoring precision of the system can reach ±0.3mm, which is improved by 40% compared with the traditional monitoring equipment, and the multi-dimensional settlement capture error is reduced by 50% compared with the traditional equipment, so that the subtle changes of the roadbed can be detected earlier and more accurately, and sufficient time is gained for subsequent regulation and control.

[0016] (2) In the application, the mixed prediction model is based on the improved space-time graph convolution network (STGCN) and the multi-head attention mechanism (MHSA), which fully excavates the topological relationship and space-time characteristics of the roadbed sensor network, greatly improves the prediction accuracy. At the same time, the multi-objective particle swarm optimization algorithm (MOPSO) optimizes the model hyperparameters, which improves the reasoning speed while ensuring the prediction accuracy. The experimental results show that the average absolute error (MAE) of the application is reduced to 0.8mm, and the reasoning speed is improved by 20%, which can make a reliable prediction on the future roadbed settlement trend and provide strong support for formulating countermeasures in advance.

[0017] (3) The application fully considers the complex geological and extreme climate conditions, and can still operate normally under the extreme environment of-40℃. In the permafrost area, the temperature control module starts the geothermal cycle to effectively maintain the roadbed temperature at ≥0℃, and inhibit the freeze-thaw settlement. In the face of complex geology such as high fill roadbed, the differentiated sensing unit flexibly adapts to the soil deformation to ensure stable data acquisition. Compared with the traditional system which frequently fails and data distortion in extreme environment, the adaptive multi-source fusion system of the application shows excellent stability and reliability.

[0018] (4) The application not only has accurate monitoring and prediction ability, but also innovatively introduces a dynamic regulation unit. When the predicted settlement exceeds the threshold, the hydraulic support system quickly adjusts the local load distribution of the roadbed according to the accurate algorithm; when the freeze-thaw risk occurs, the temperature control module responds in time. In the actual application in the permafrost area, through active regulation, the system reduces the maintenance cost by 30%, and the system design life can reach more than 10 years, which greatly reduces the long-term operation and maintenance cost, and significantly improves the economic benefit and social benefit of road engineering.

[0019] The technical solutions of the application will be further described in detail below with the help of the drawings and examples. BRIEF DESCRIPTION OF DRAWINGS

[0020] Figure 1 The flowchart of the adaptive multi-source fusion roadbed intelligent monitoring and dynamic regulation method of the application. DETAILED DESCRIPTION

[0021] In order to make the technical solutions, advantages and objectives of the present application clearer, the technical solutions of the embodiments of the present application will be clearly and completely described below. The described embodiments are part of the embodiments of the present application, not all of the embodiments. Based on the described embodiments of the present application, all other embodiments obtained by those of ordinary skill in the art without creative labor fall within the protection scope of the present application.

[0022] Unless otherwise defined, the technical terms or scientific terms used in the present application shall have the usual meanings understood by those skilled in the art to which the present application belongs.

[0023] Embodiment one

[0024] The present application proposes a self-adaptive multi-source fusion roadbed intelligent monitoring and dynamic regulation system, which is used to improve the deficiencies of the previous roadbed settlement monitoring and regulation, including: The differentiated sensing unit includes a composite anchoring structure, a magnetic attraction following module and an optical fiber monitoring network. The composite anchoring structure contains a tapered pipe, the outer wall of which is provided with staggered spiral blades, and the edges of the blades are integrated with micro pressure sensors. The magnetic attraction following module is composed of a magnetic sleeve and an elastic plate part, and the elastic plate part is embedded with shape memory alloy (SMA). When the settlement exceeds the threshold value, the deformation is triggered, and the contact area with the roadbed is increased. The sleeve is provided with a Hall sensor, which dynamically captures the displacement change of the signal receiver at the end of the flexible pipeline through magnetic attraction force. The optical fiber monitoring network is composed of distributed fiber Bragg grating sensors (FBG) arranged in layers along the roadbed, which is used to collect strain, temperature and humidity data in real time, and facilitate accurate monitoring and regulation of the real-time state of the roadbed.

[0025] The deformation amount of the shape memory alloy SMA The calculation formula is as follows: ; Among them, α the thermal expansion coefficient, β the stress response coefficient, σ the stress of the soil body; the real-time monitored soil temperature (unit: ℃), collected through the FBG optical fiber network or temperature sensor; the reference temperature of the SMA (unit: ℃), that is, the deformation reference temperature (such as the phase change starting temperature or the designed initial temperature).

[0026] The data transmission and processing unit uploads the multi-source data (pressure, displacement, strain) to the edge computing node by using the LoRaWAN protocol, and the theoretical communication distance is greater than or equal to 10km; the data fusion layer in the edge computing node eliminates the time delay of the sensor data through the space-time alignment algorithm, and filters the abnormal values based on DBSCAN clustering, to generate a standardized feature matrix.

[0027] The hybrid model prediction unit constructs an STGCN-MHSA hybrid model based on an improved spatiotemporal graph convolution network STGCN and a multi-head attention mechanism MHSA, and adopts a multi-objective particle swarm optimization MOPSO to optimize the STGCN-MHSA hybrid model hyperparameters, inputs a standardized feature matrix, and predicts the settlement trend and risk level; wherein the spatiotemporal graph convolution network (STGCN) encodes the subgrade sensor network topology into a graph structure, and extracts spatial correlation features through graph convolution; the multi-head self-attention mechanism (MHSA) dynamically weights the key nodes in the time series; and the multi-objective particle swarm optimization (MOPSO) takes the prediction error (MAE) and the reasoning speed (FPS) as the dual fitness functions to generate a Pareto optimal solution set.

[0028] The dynamic regulation unit comprises a hydraulic support system, a temperature control module and a blockchain storage module; wherein the hydraulic support system is used to adjust the local load distribution of the subgrade according to the prediction result, and suppress uneven settlement; the temperature control module is used to start the geothermal cycle in the permafrost area, maintain the subgrade temperature greater than or equal to 0℃, and suppress freeze-thaw settlement; and the blockchain storage module is used to upload all operation records to the blockchain storage, and ensure that the data cannot be tampered with.

[0029] Embodiment two

[0030] As shown in Figure 1 , the application further provides a self-adaptive multi-source fusion subgrade intelligent monitoring and dynamic regulation method, which is used to realize the self-adaptive multi-source fusion subgrade intelligent monitoring and dynamic regulation system proposed in embodiment one, and the steps are as follows: S1, real-time collection of pressure P, displacement δ, strain ε and temperature H is realized through the micro pressure sensor on the conical pipe spiral blade, the Hall sensor of the magnetic attraction following module and the distributed FBG optical fiber network, and a spatiotemporal alignment matrix is constructed X E R N×T×C , wherein N is the number of sensor nodes, T is the time window length, C is the feature dimension; S2, the DBSCAN clustering algorithm is adopted to eliminate outliers, and normalization processing is performed, and the formula is as follows: ; , wherein is the element value of the original feature tensor, is the element value of the normalized feature tensor, , are the mean and standard deviation of the first c dimensional feature, respectively; S3, input the pretreated data into the STGCN-MHSA hybrid model, output the settlement prediction value Y and the risk level R; the calculation process of the STGCN-MHSA hybrid model is as follows: Spatial graph convolution: Definition of sensor topology graph G = (V, E) The adjacency matrix A is constructed by threshold truncation method, and the formula is as follows: ; Wherein, V is the vertex set (Vertices), that is, the sensor node set, and each node represents a physical monitoring point; E is the edge set (Edges), which shows the connection relationship between nodes, representing the spatial correlation between sensors; is the definition of node edge (only the nodes with a distance are connected, is the distance threshold, unit: m), is the Euclidean distance of node , is the distance threshold; The Chebyshev polynomial approximation is adopted, and the formula is as follows: ; Wherein, is the spatial feature matrix, is the order Chebyshev polynomial, is the scaled Laplacian matrix, is the trainable parameter, is the order of Chebyshev polynomial (control the range of receptive field); Temporal expansion convolution: The convolution kernel with expansion rate is used to extract the time sequence feature, and the formula is as follows: ; Wherein, is the time sequence feature matrix obtained by processing through the time feature extraction module (such as time expansion convolution, LSTM, etc.), is the time offset, which represents the sliding position of the convolution kernel on the time sequence; is the size of the time convolution kernel, which controls the window length of the extracted time sequence feature, is the convolution kernel weight, is the sensor node spatial correlation feature matrix extracted by the spatial graph convolution module; Multi-head self-attention weighting: Split into heads, and the formula is as follows: ; ; ; ; where, is the i-th attention head output, , , is the query, key, value matrix, is the transpose of the key matrix for computing query-key similarity, is the key vector dimension (scaling factor for stabilizing gradients), , is the attention computation formula, , , is the projection matrix; the final result output, formula as follows: ; where, is the fused feature, is the h-th attention head output, is the linear projection matrix (output weight), is the concatenation of the outputs of multiple attention heads; the settlement result prediction, formula as follows: ; where, is the i-th attention head output, is the tensor flattening operation (input pre-processing before fully connected layer), , is the fully connected layer parameter, is the Sigmoid activation function.

[0031] The STGCN-MHSA hybrid model adopts NSGA-II double fitness function to optimize the model hyperparameters, formula as follows: ; where, is the mean absolute error, is the inverse of model inference speed, is the true settlement value (mm, from historical monitoring or field measurement), is the model inference speed (frame / second), is the mean absolute error.

[0032] S4, according to the prediction result of the STGCN-MHSA hybrid model, the local load distribution of the roadbed is adjusted through the hydraulic support system or the temperature control module is started to inhibit the freeze-thaw settlement, and the specific operation is as follows: When , the hydraulic support system is triggered: ; Wherein PID coefficient, is the differential gain coefficient (kPa·s / mm, inhibiting the rate of settlement change), is the settlement risk threshold (mm), is the load amount of the hydraulic system (kPa, used to compensate for uneven settlement).

[0033] The following will be further described through specific implementation cases.

[0034] After being deployed in the high-cold permafrost area of the highway, the STGCN-MHSA model predicts the local settlement trend 72 hours in advance (error 0.7mm), triggers the geothermal cycle to maintain the roadbed temperature at 1.2℃, and the hydraulic system adjusts the load distribution, and the uneven settlement amount is reduced by 62%.

[0035] It should be noted that the contents not described in detail in the present application are all prior art and are well known to those skilled in the art.

[0036] Therefore, the present application provides a self-adaptive multi-source fusion roadbed intelligent monitoring and dynamic regulation system and method, which realizes high-precision monitoring of ±0.3mm (40% higher than the traditional one) through the cooperation of multi-source heterogeneous sensor network and advanced algorithm, reduces the prediction MAE to 0.8mm and improves the reasoning speed by 20%, and can stably operate in an extreme environment of-40℃; the active regulation mechanism reduces the maintenance cost in the permafrost area by 30%, the system life is more than 10 years, and the problems of low monitoring precision, poor prediction timeliness and passive regulation of the traditional monitoring are solved, and an efficient solution is provided for the health management of the roadbed in complex scenes.

[0037] Finally, it should be noted that: the above examples are only used to illustrate the technical solutions of the present application but not to limit it, although the present application has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that: the technical solutions of the present application can still be modified or replaced by equivalents, and these modifications or equivalent replacements cannot make the modified technical solutions deviate from the spirit and scope of the technical solutions of the present application.

Claims

1. An adaptive multi-source fusion roadbed intelligent monitoring and dynamic control system, characterized by: It includes a differentiated sensing unit, a data transmission and processing unit, a hybrid model prediction unit, and a dynamic control unit; The differentiated sensing unit includes a composite anchoring structure, a magnetic follower module, and a fiber optic monitoring network. The composite anchoring structure comprises a tapered tube with staggered spiral blades on its outer wall, and micro-pressure sensors integrated into the blade edges. The magnetic follower module consists of a magnetic sleeve and an elastic plate. The sleeve has a built-in Hall effect sensor that dynamically attracts a signal receiver at the end of the flexible pipe through magnetic attraction. The elastic plate has a built-in shape memory alloy (SMA). The optical fiber monitoring network is layered with distributed fiber grating sensors FBG along the roadbed; The data transmission and processing unit uses the LoRaWAN wireless transmission protocol to upload multi-source data to the edge computing node. The data fusion layer in the edge computing node performs spatiotemporal alignment and outlier filtering on the multi-source data. The hybrid model prediction unit builds the STGCN-MHSA hybrid model based on the improved spatiotemporal graph convolutional network (STGCN) and the multi-head attention mechanism (MHSA). The multi-objective particle swarm algorithm (MOPSO) is used to optimize the hyperparameters of the STGCN-MHSA hybrid model to predict sedimentation trends and risk levels. The dynamic control unit includes a hydraulic support system, a temperature control module, and a blockchain evidence storage module.

2. The adaptive multi-source fusion roadbed intelligent monitoring and dynamic control system according to claim 1 is characterized in that: Deformation of shape memory alloy SMA in differentiated sensing unit The calculation formula is as follows: ； in, α is the coefficient of thermal expansion, β is the stress response coefficient, σ is the soil stress, For real-time monitoring of soil temperature, for SMA The reference temperature.

3. The adaptive multi-source fusion roadbed intelligent monitoring and dynamic control system according to claim 1 is characterized in that: In the hybrid model prediction unit, the STGCN-MHSA hybrid model encodes the roadbed topology structure into graph data, extracts local features through spatiotemporal convolution, and uses multi-head attention to weight key nodes.

4. The adaptive multi-source fusion roadbed intelligent monitoring and dynamic control system according to claim 1 is characterized in that: In the hybrid model prediction unit, the multi-objective particle swarm algorithm MOPSO is set as a dual fitness function to simultaneously optimize the prediction error MAE and the model inference speed FPS to generate the Pareto optimal solution set.

5. The adaptive multi-source fusion roadbed intelligent monitoring and dynamic control system according to claim 1 is characterized in that: In the dynamic control module, the temperature control module starts the geothermal cycle in the permafrost area to maintain the roadbed temperature not less than 0℃.

6. An adaptive multi-source fusion roadbed intelligent monitoring and dynamic control method, used to implement the control system according to any one of claims 1 to 5, characterized in that: The specific steps are as follows: Step S1: Use the micro pressure sensor on the spiral blade of the cone tube, the Hall sensor of the magnetic follower module, and the distributed FBG optical fiber network to collect pressure P, displacement δ, strain ε, and temperature H in real time and construct a time-space alignment matrix. X∈R N×T×C ,in N is the number of sensor nodes, T is the time window length, C is the feature dimension; Step S2: using the DBSCAN clustering algorithm to remove outliers and perform normalization; Step S3: input the pre-processed data into the STGCN-MHSA hybrid model, and output the settlement prediction value Y and risk level R; Step S4: According to the prediction results of the STGCN-MHSA hybrid model, the local load distribution of the roadbed is adjusted through the hydraulic support system or the temperature control module is activated to suppress freeze-thaw settlement.

7. The adaptive multi-source fusion roadbed intelligent monitoring and dynamic control method according to claim 6 is characterized in that: In step S2, the normalization formula is as follows: ； in, is the element value of the original feature tensor, is the normalized feature tensor element value, 、 Respectively c The mean and standard deviation of the dimensional features.

8. The adaptive multi-source fusion roadbed intelligent monitoring and dynamic control method according to claim 6 is characterized in that: In step S3, the STGCN-MHSA hybrid model calculation process is as follows: Spatial Graph Convolution: Define sensor topology G=(V,E) , the adjacency matrix A is constructed according to the threshold truncation method, and the formula is as follows: ； in, V is a set of vertices, E is the edge set, To define the node edges, For nodes The Euclidean distance of is the distance threshold; Using Chebyshev polynomial approximation, the formula is as follows: ； in, is the spatial feature matrix, for Chebyshev polynomials of order, is the scaled Laplacian matrix, is a trainable parameter, is the order of Chebyshev polynomial; Time-Dilated Convolution: Adopting expansion rate The convolution kernel extracts the time series features, and the formula is as follows: ； in, To extract the time feature through the module, is the time offset, is the size of the temporal convolution kernel, is the convolution kernel weight, is the spatial correlation feature matrix of sensor nodes extracted by the spatial graph convolution module; Multi-head self-attention weighted: Will Split into The formula is as follows: ； ； ； ； in, For the The output of an attention head, 、 、 is the query, key, and value matrix, for The transpose of the key matrix, used to calculate query-key similarity, is the key vector dimension, 、 is the attention calculation function, 、 、 is the projection matrix; Calculate the final result, the formula is as follows: ； in, To fusion features, For the The output of an attention head, is the linear projection matrix, To concatenate the outputs of multiple attention heads; The settlement result prediction formula is as follows: ； in, For the The output of an attention head, is the tensor flattening operation, 、 are the parameters of the fully connected layer, is the Sigmoid activation function.

9. The adaptive multi-source fusion roadbed intelligent monitoring and dynamic control method according to claim 6 is characterized in that: In step S3, the STGCN-MHSA hybrid model uses the NSGA-II dual fitness function to optimize the model hyperparameters. The formula is as follows: ； in, is the mean absolute error, is the inverse of the model inference speed, is the true settlement value, is the model inference speed, is the mean absolute error.

10. The adaptive multi-source fusion roadbed intelligent monitoring and dynamic control method according to claim 6 is characterized in that: In step S4, according to the settlement prediction result, the dynamic control of the hydraulic support system is triggered, specifically: when When the hydraulic support system is triggered: ； in is the PID coefficient, is the differential gain coefficient, is the subsidence risk threshold, Adjust the load capacity for the hydraulic system.

Images