Multi-mode bridge vibration monitoring system

Abstract

The invention relates to the technical field of bridge monitoring, and provides a multi-mode bridge vibration monitoring system. The system comprises a data acquisition and processing module, a bridge monitoring module, a monitoring precision optimization module and a bridge health assessment and early warning module. Intelligent analysis and health state evaluation of multi-dimensional data are realized through a bridge safety fusion network; according to the system, a k and r two-parameter optimization algorithm is adopted to dynamically adjust hyper-parameters of the bridge safety fusion network, the learning rate, regularization parameters and batch size are optimized to adapt to different bridge environments and operation conditions, and the monitoring precision and long-term stability are improved; through multi-modal data fusion, intelligent optimization and adaptive adjustment, the comprehensiveness, accuracy and adaptability of bridge vibration monitoring are improved, and efficient and scientific technical guarantee is provided for bridge safety management.

Description

Technical Field

[0001] The invention relates to the technical field of bridge monitoring, in particular to a multi-modal bridge vibration monitoring system. Background Art

[0002] With the rapid development of modern transportation infrastructure, bridges, as an important part of road and rail transportation, have a significant impact on public safety and social economy due to their structural safety and service life. However, the traditional bridge vibration monitoring system has the following shortcomings: First, the traditional bridge vibration monitoring system mainly relies on a single type of sensor, which is difficult to fully reflect the health status of the bridge, resulting in the one-sidedness of the monitoring data and affecting the accuracy of damage identification; second, the traditional system fails to make full use of the multi-source information collected by the sensor and lacks an effective data fusion strategy. The observation results of different sensors are difficult to process in a coordinated manner, resulting in data redundancy and information loss, affecting the reliability of the monitoring results; finally, the traditional system is difficult to effectively capture the spatiotemporal correlation characteristics of bridge vibration data; especially when the bridge suffers early damage, traditional methods are often difficult to accurately identify and warn, resulting in delayed maintenance decisions, increasing bridge safety risks and maintenance costs. Summary of the invention

[0003] The present invention proposes a multi-modal bridge vibration monitoring system, which aims to improve the accuracy and real-time performance of bridge vibration monitoring and ensure the structural safety and long-term stable operation of the bridge. The system collects multi-modal vibration data of the bridge in real time by arranging accelerometers, strain gauges, displacement sensors, temperature sensors and anemometers on the bridge structure, and constructs a bridge safety fusion network by using topological perception, multi-layer gated fusion and spectrum feature extraction to capture the health status of the bridge from multiple dimensions. The system first establishes a sensor interaction graph through the spatial dependence, time dependence and modal dependence of the bridge sensor data, and initializes the hyperparameters of the bridge safety fusion network based on the topological structure to adapt it to the vibration characteristics under different bridge environments. In the process of data analysis, the system The system uses the Qiaoan fusion network to process multimodal data, extracts low-frequency and high-frequency features from the spectrum domain, and accurately identifies the long-term structural changes and sudden damage of the bridge; to further optimize the monitoring accuracy, the system introduces the k, r dual-parameter optimization algorithm, dynamically adjusts the hyperparameters of the Qiaoan fusion network, optimizes the learning rate, momentum coefficient, regularization parameter and batch size, and ensures that the monitoring system can still maintain efficient and accurate bridge health status prediction and vibration monitoring capabilities under complex loads and environmental interference; the present invention constructs a set of efficient, intelligent and stable bridge vibration monitoring system through multimodal data fusion, intelligent feature extraction and adaptive optimization strategy, which provides scientific support for bridge structure health management and improves the level of bridge safety assurance.

[0004] The present invention provides a multi-modal bridge vibration monitoring system, which includes a bridge sensor, a data acquisition module, a bridge monitoring module, an optimization monitoring accuracy module and a bridge health assessment and early warning module;

[0005] The data acquisition module arranges bridge sensors on the bridge. The bridge sensors include accelerometers, strain gauges, displacement sensors, temperature sensors and anemometers. The bridge sensors collect acceleration data, displacement data, strain data, temperature data and wind speed data to obtain multi-modal bridge data.

[0006] The bridge monitoring module builds a bridge fusion network through topology perception, multi-layer gated fusion and spectrum feature extraction, and initializes the hyperparameters of the bridge fusion network; it processes multi-modal bridge data through the bridge fusion network to monitor bridge vibration and generate bridge health status prediction results;

[0007] Optimize the monitoring accuracy module, build a k,r dual-parameter optimization algorithm through Bregman divergence-guided adaptive gradient optimization, k,r-exponential transformation and generalized regularization mechanism, and optimize the hyperparameters of the Qiaoan fusion network through the k,r dual-parameter optimization algorithm;

[0008] Bridge health assessment and early warning module: Generates a bridge health status report based on the bridge health status prediction results, including the degree of structural damage, bearing capacity attenuation rate, vibration mode change rate, displacement over-limit risk, fatigue damage accumulation, structural stability index, health status score and early warning level.

[0009] Furthermore, the bridge monitoring module generates a process of predicting the bridge health status, which specifically includes the following steps:

[0010] Step S1: Establish the topological structure of multimodal bridge data, capture the spatial dependency, temporal dependency and modal dependency of bridge sensors, and construct a sensor interaction graph;

[0011] Spatial dependence: Adjacent sensors on the bridge are interconnected; sensors include accelerometers, strain gauges, displacement sensors, temperature sensors, and anemometers;

[0012] Time dependency: multimodal bridge data at the current moment is linked to multimodal bridge data in the past; Methods: The association between the current multimodal bridge data and the past multimodal bridge data is captured by the LSTM time series model; Association: The multimodal bridge data at the current moment and the multimodal bridge data at the past moment are used as the input of the LSTM time series model. The LSTM time series model automatically models the dependency of the multimodal bridge data in the time dimension by learning the relationship between the current moment data and the historical data, thereby understanding how the data at different time points influence and relate to each other.

[0013] Modal dependence: different multimodal bridge data are interrelated; Method: The modal dependency between different multimodal bridge data is captured through the attention mechanism. The attention mechanism can give different importance weights to different modal data, so as to better capture the relationship between modalities. Association: The attention mechanism processes different multimodal bridge data in parallel, processes each modality independently, and then fuses them to learn the associations between them. In this way, the attention mechanism can identify the relationship between acceleration data, displacement data, strain data, temperature data, and wind speed data, and then reveal the mutual influence between modal data.

[0014] Step S2: Calculate the Laplacian matrix using the sensor interaction graph, perform feature decomposition, and obtain low-frequency features and high-frequency features; construct a multi-layer adaptive gated fusion mechanism through a three-layer gating mechanism, dynamic attention weighting, and gradient constraints; Use the multi-layer adaptive gated fusion mechanism to fuse the low-frequency features and high-frequency features to generate spectrum domain fusion features;

[0015] Step S3: restore missing data of the spectral domain fusion features, and use a multi-head attention mechanism to optimize feature fusion to generate complete bridge data;

[0016] Step S4: Combine the modal recovery loss function and the cross entropy loss function to construct a bridge health classifier; input the complete bridge data into the bridge health classifier to generate a bridge health status prediction result.

[0017] Further, the process of optimizing the monitoring accuracy module and optimizing the hyperparameters of the Qiaoan fusion network includes the following steps:

[0018] Step B1: Define the search space of the hyperparameters of the Qiaoan fusion network, which includes the learning rate, momentum coefficient, regularization parameter and batch size; set the k and r double parameters to optimize the hyperparameters of the Qiaoan fusion network; k controls the smoothness of the hyperparameter change, and r controls the sensitivity of the hyperparameter adjustment;

[0019] Step B2: Define the loss function and calculate the divergence gradient of the hyperparameters of the Bridge-An fusion network to the loss function;

[0020] Step B3: Combine the divergence gradient and use the k and r double parameters to update the hyperparameters of the Qiaoan fusion network;

[0021] Step B4: Set the convergence conditions. If the convergence conditions are met, stop the optimization; otherwise, return to step B2 to continue updating the hyperparameters of the BridgeAn fusion network.

[0022] Further, step B3 specifically includes the following steps:

[0023] Step B31: Use k,r-exponential gradient to optimize the learning rate and momentum coefficient;

[0024] Step B32: Optimizing regularization parameters using k,r-generalized regularization;

[0025] Step B33: Update the batch size using k,r-additive hyperparameters;

[0026] Step B34: Use Bregman divergence to provide a global optimization strategy to coordinately optimize the learning rate, momentum coefficient, regularization parameter, and batch size, optimize high-dimensional hyperparameter search, and ensure the best fit of the hyperparameter combination.

[0027] By adopting the above scheme, the beneficial effects achieved by the present invention are as follows:

[0028] The present invention realizes the efficient collection and intelligent fusion of multimodal vibration data of bridges, and improves the comprehensive perception capability of the health status of bridge structures. Through topological perception, multi-layer gated fusion and spectrum feature extraction, the system can make full use of multimodal data such as acceleration, strain, displacement, temperature and wind speed to build a bridge safety fusion network, so that bridge vibration monitoring is no longer limited to a single data source, but is based on comprehensive analysis of multi-dimensional information. This data fusion method effectively solves the problems of information isolation and one-sided monitoring data in traditional systems, improves the system's ability to analyze the dynamic characteristics of bridges, and makes the monitoring results more accurate and reliable.

[0029] The present invention optimizes the adaptive capability of the bridge vibration monitoring system and enhances the monitoring stability in complex environments. According to the structural characteristics and environmental conditions of different bridges, the system adopts the k, r dual-parameter optimization algorithm to dynamically adjust the hyperparameters of the bridge safety fusion network to achieve adaptive optimization of the learning rate, regularization parameter and batch size. This optimization mechanism ensures that the system can still accurately extract vibration modes and promptly detect damage changes of bridges when facing different loads, temperature changes and wind speeds. Compared with traditional systems with fixed hyperparameters, this system can dynamically adjust parameter configurations to adapt to a variety of bridge working conditions, significantly improving the long-term stability and adaptability of monitoring.

[0030] In summary, the present invention constructs an efficient, accurate and stable bridge vibration monitoring system through the combination of multimodal data fusion and intelligent optimization algorithm; compared with the traditional system, the present invention not only improves the comprehensiveness and accuracy of bridge structure health monitoring, but also optimizes the long-term stability and adaptability of the monitoring system, providing more scientific and efficient technical support for bridge operation and maintenance management. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] Figure 1 A schematic diagram of a module of a multi-modal bridge vibration monitoring system proposed by the present invention;

[0032] Figure 2 A schematic diagram of the structure of the bridge sensor arranged on the bridge proposed by the present invention;

[0033] Figure 3 This is a schematic flow chart of step B3 proposed in Example 6. DETAILED DESCRIPTION

[0034] The technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments; based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0035] Embodiment 1, according to Figure 1 , Figure 2 , the present invention provides a multi-modal bridge vibration monitoring system, the system includes a data acquisition module, a bridge sensor, a bridge monitoring module, an optimization monitoring accuracy module and a bridge health assessment and early warning module;

[0036] The data acquisition module arranges bridge sensors on the bridge. The bridge sensors include accelerometers, strain gauges, displacement sensors, temperature sensors and anemometers. The bridge sensors collect acceleration data, displacement data, strain data, temperature data and wind speed data to obtain multi-modal bridge data.

[0037] ;

[0038] The bridge monitoring module builds a bridge fusion network through topology perception, multi-layer gated fusion and spectrum feature extraction, and initializes the hyperparameters of the bridge fusion network; it processes multi-modal bridge data through the bridge fusion network to monitor bridge vibration and generate bridge health status prediction results;

[0039] Optimize the monitoring accuracy module, build a k,r dual-parameter optimization algorithm through Bregman divergence-guided adaptive gradient optimization, k,r-exponential transformation and generalized regularization mechanism, and optimize the hyperparameters of the Qiaoan fusion network through the k,r dual-parameter optimization algorithm;

[0040] Bridge health assessment and early warning module: Generates a bridge health status report based on the bridge health status prediction results, including the degree of structural damage, bearing capacity attenuation rate, vibration mode change rate, displacement over-limit risk, fatigue damage accumulation, structural stability index, health status score and early warning level;

[0041] ;

[0042] Embodiment 2: This embodiment is based on embodiment 1. In this embodiment, the bridge monitoring module generates a bridge health status prediction result, which specifically includes the following steps:

[0043] Step S1: Establish the topological structure of multimodal bridge data, capture the spatial dependency, temporal dependency and modal dependency of bridge sensors, and construct a sensor interaction graph;

[0044] Spatial dependence: Adjacent sensors on the bridge are interconnected; sensors include accelerometers, strain gauges, displacement sensors, temperature sensors, and anemometers;

[0045] Time dependency: multimodal bridge data at the current moment is linked to multimodal bridge data in the past; Methods: The association between the current multimodal bridge data and the past multimodal bridge data is captured by the LSTM time series model; Association: The multimodal bridge data at the current moment and the multimodal bridge data at the past moment are used as the input of the LSTM time series model. The LSTM time series model automatically models the dependency of the multimodal bridge data in the time dimension by learning the relationship between the current moment data and the historical data, thereby understanding how the data at different time points influence and relate to each other.

[0046] Modal dependence: different multimodal bridge data are interrelated; Method: The modal dependency between different multimodal bridge data is captured through the attention mechanism. The attention mechanism can give different importance weights to different modal data, so as to better capture the relationship between modalities. Association: The attention mechanism processes different multimodal bridge data in parallel, processes each modality independently, and then fuses them to learn the associations between them. In this way, the attention mechanism can identify the relationship between acceleration data, displacement data, strain data, temperature data, and wind speed data, and then reveal the mutual influence between modal data.

[0047] Step S2: Use the sensor interaction graph to calculate the Laplacian matrix, perform feature decomposition, and obtain low-frequency features and high-frequency features; construct a multi-layer adaptive gated fusion mechanism through a three-layer gating mechanism, dynamic attention weighting, and gradient constraints; use a multi-layer adaptive gated fusion mechanism to fuse low-frequency features and high-frequency features to generate spectrum domain fusion features. The formula used is as follows:

[0048] ;

[0049] in, Indicates the bridge sensor index, Indicates the index of the adjacent sensor of the bridge sensor. represents the fusion feature in the spectrum domain, Indicates adjacent sensors Bridge Sensors The size of the contribution; Represents a set of relationships, including spatial dependency, temporal dependency, and modal dependency. express An element in Indicates bridge sensor In relationship The neighbor set below; represents the adaptive weight, controlling the influence of each sensor in the bridge sensor, Represents the low-frequency characteristics of the sensor itself, Represents neighbor nodes The low frequency characteristics of Represents neighbor nodes The high frequency characteristics of Indicates low frequency gating, represents high frequency gating, represents modal enhancement gating, represents the feature gradient, represents the feature gradient weight;

[0050] Low-frequency features: represent the global health of the bridge, including temperature effects and long-term structural changes;

[0051] High-frequency features: represent sudden anomalies, including cracks, local damage, and impact vibration;

[0052] Step S3: restore missing data of the spectral domain fusion features, and use a multi-head attention mechanism to optimize feature fusion to generate complete bridge data;

[0053] Step S4: Combine the modal recovery loss function and the cross entropy loss function to construct a bridge health classifier; input the complete bridge data into the bridge health classifier to generate the bridge health status prediction result. The formula used is as follows:

[0054] Loss function formula:

[0055] ;

[0056] in, represents the mode restoration loss, Indicates different modes, represents acceleration data, represents strain data, Represents displacement data, Represents temperature data, Indicates wind speed, represents the total number of samples, represents the sample index, represents the modal dimension, Represents real data, Indicates recovery of data;

[0057] Classifier formula:

[0058] ;

[0059] in, represents the bridge health status prediction result, represents the cross entropy loss, Indicates the bridge health status category, represents the classification weight matrix, represents the bias vector; Represents the activation function.

[0060] Embodiment 3: This embodiment is based on embodiment 1. In this embodiment, the bridge monitoring module generates a bridge health status prediction result, which specifically includes the following steps:

[0061] Step E1: Establish the topological structure of multimodal bridge data, capture the spatial dependency, temporal dependency and modal dependency of bridge sensors, and construct the sensor interaction graph;

[0062] Step E2: Calculate the Laplacian matrix using the sensor interaction graph, perform feature decomposition, and obtain low-frequency features and high-frequency features; construct an adaptive fusion model using the gating mechanism, attention weighting, and gradient constraint methods; use the adaptive fusion model to fuse the low-frequency features and high-frequency features to generate spectral domain fusion features;

[0063] Step E3: restore missing data of the spectral domain fusion features, and use a multi-head attention mechanism to optimize feature fusion to generate complete bridge data;

[0064] Step E4: Combine the modal recovery loss function and the cross entropy loss function to construct a bridge health classifier; input the complete bridge data into the bridge health classifier to generate a bridge health status prediction result.

[0065] Embodiment 4: This embodiment is based on embodiment 2. In this embodiment, the process of optimizing the monitoring accuracy module and optimizing the hyperparameters of the Qiaoan fusion network specifically includes the following steps:

[0066] Step B1: Define the search space of the hyperparameters of the Bridge Fusion Network, which includes the learning rate, momentum coefficient, regularization parameter, and batch size; set the k and r dual parameters to optimize the hyperparameters of the Bridge Fusion Network; k controls the smoothness of the hyperparameter changes, and r controls the sensitivity of the hyperparameter adjustments;

[0067] Step B2: Define the loss function and calculate the divergence gradient of the hyperparameters of the Bridge-An fusion network to the loss function;

[0068] Step B3: Combine the divergence gradient and use k,r dual parameter optimization to update the hyperparameters of the Qiaoan fusion network;

[0069] Step B4: Set the convergence conditions. If the convergence conditions are met, stop the optimization; otherwise, return to step B2 to continue updating the hyperparameters of the BridgeAn fusion network.

[0070] Embodiment 5: This embodiment is based on embodiment 2. In this embodiment, the process of optimizing the monitoring accuracy module and optimizing the hyperparameters of the Qiaoan fusion network specifically includes the following steps:

[0071] Step U1: Define the search space of the hyperparameters of the Qiaoan fusion network, which includes the learning rate, momentum coefficient, regularization parameter and batch size;

[0072] Step U2: Define the loss function and calculate the model loss of the hyperparameters of the Bridge-An fusion network;

[0073] Step U3: Update the hyperparameters of the Bridgean fusion network according to the model loss;

[0074] Step U4: Set the convergence conditions. If the convergence conditions are met, stop the optimization; otherwise, return to step B2 to continue updating the hyperparameters of the Bridge-An fusion network.

[0075] Embodiment 6, according to Figure 3 This embodiment is based on the fourth embodiment. In this embodiment, step B3 specifically includes the following steps:

[0076] Step B31: Adaptive gradient update: Use k,r-exponential gradient to optimize the learning rate and momentum coefficient. The formula used is as follows:

[0077] ;

[0078] ;

[0079] in, represents the iteration index, Indicated in The learning rate after round iteration update, Indicated in The learning rate after round iteration update; Indicated in The momentum parameter after round iterative update, Indicated in Momentum parameter after round iterative update; Represents k, r multiplication, used to update hyperparameters and ; Represents the update coefficient, which is used to control the step size of the learning rate update. represents the update coefficient, which is used to control the update amplitude of the momentum coefficient. Represents the learning rate For the loss function The gradient of Momentum coefficient For the loss function The gradient of represents the exponential function;

[0080] Step B32: Optimize regularization parameters: Use k,r-generalized regularization to optimize regularization parameters. The formula used is as follows:

[0081] ;

[0082] in, Indicated in The regularization parameter after rounds of iterations, Indicates the current The regularization parameter of the round iteration, represents k,r-generalized regularization multiplication, ensuring nonlinear update of regularization parameters; represents the regularization parameter The learning rate, Represents the loss function right The gradient of represents the update factor of the regularization parameter;

[0083] Step B33: Optimize batch size: Update the batch size using k,r-additive hyperparameters. The formula used is as follows:

[0084] ;

[0085] in, Indicated in The batch size after round iteration, Indicates the current The batch size of the round iteration, represents the k,r-additive hyperparameter update operation, ensuring smooth adjustment of batch size; Batch size The update step size is Represents the loss function right The gradient of represents the batch size update factor;

[0086] Step B34: Collaborative optimization: Bregman divergence is used to provide a global optimization strategy to collaboratively optimize the learning rate, momentum coefficient, regularization parameter, and batch size, optimize high-dimensional hyperparameter search, and ensure the best fit of the hyperparameter combination.

[0087] Embodiment 7, this embodiment is based on embodiment 6. In this embodiment, the bridge health assessment and early warning module: generates a bridge health status report according to the bridge health status prediction result, including the structural damage degree, bearing capacity attenuation rate, vibration mode change rate, displacement over-limit risk, fatigue damage accumulation, structural stability index, health status score and early warning level;

[0088] Generates the following bridge health report:

[0089] ;

[0090] The health status score is "B", and the bridge is in a sub-healthy state. The following measures are taken:

[0091] Repair the cracks by injecting glue into the cracks;

[0092] Strengthen fatigue monitoring in the next quarter;

[0093] Maintain the current monitoring frequency and perform more frequent sampling when wind loads are high;

[0094] If the modal frequency drop continues to expand, a static load test is conducted to verify the bearing capacity.

[0095] The present invention and its embodiments are described above, which is not restrictive. What is shown in the accompanying drawings is only one of the embodiments of the present invention, and the actual structure is not limited to this. In short, if ordinary technicians in this field are inspired by it and do not deviate from the purpose of the invention, they can creatively design structural methods and embodiments similar to the technical solution, which should all fall within the scope of protection of the present invention.

Claims

1. A multi-modal bridge vibration monitoring system, characterized in that: The system includes bridge sensors, data acquisition modules, bridge monitoring modules, optimized monitoring accuracy modules and bridge health assessment and early warning modules; A data acquisition module is used to arrange bridge sensors on the bridge. The bridge sensors include accelerometers, strain gauges, displacement sensors, temperature sensors and anemometers; The acceleration data, displacement data, strain data, temperature data and wind speed data are collected by bridge sensors to obtain multi-modal bridge data; The bridge monitoring module builds a bridge fusion network through topology perception, multi-layer gated fusion and spectrum feature extraction, and initializes the hyperparameters of the bridge fusion network; it processes multi-modal bridge data through the bridge fusion network to monitor bridge vibration and generate bridge health status prediction results; Optimize the monitoring accuracy module, build a k,r dual-parameter optimization algorithm through Bregman divergence-guided adaptive gradient optimization, k,r-exponential transformation and generalized regularization mechanism, and optimize the hyperparameters of the Qiaoan fusion network through the k,r dual-parameter optimization algorithm; Bridge health assessment and early warning module: Generates a bridge health status report based on the bridge health status prediction results, including the degree of structural damage, bearing capacity attenuation rate, vibration mode change rate, displacement over-limit risk, fatigue damage accumulation, structural stability index, health status score and early warning level.

2. A multi-modal bridge vibration monitoring system according to claim 1, characterized in that: The bridge monitoring module generates the bridge health status prediction results, which specifically includes the following steps: Step S1: Establish the topological structure of multimodal bridge data, capture the spatial dependency, temporal dependency and modal dependency of bridge sensors, and construct a sensor interaction graph; Step S2: construct a multi-layer adaptive gated fusion mechanism according to the sensor interaction graph; use the multi-layer adaptive gated fusion mechanism to perform feature fusion and generate spectrum domain fusion features; Step S3: restore missing data of the spectral domain fusion features, and use a multi-head attention mechanism to optimize feature fusion to generate complete bridge data; Step S4: Combine the modal recovery loss function and the cross entropy loss function to construct a bridge health classifier; input the complete bridge data into the bridge health classifier to generate a bridge health status prediction result.

3. A multi-modal bridge vibration monitoring system according to claim 2, characterized in that: Step S2 specifically includes: using the sensor interaction graph to calculate the Laplacian matrix, perform feature decomposition, and obtain low-frequency features and high-frequency features; constructing a multi-layer adaptive gated fusion mechanism through a three-layer gating mechanism, dynamic attention weighting, and gradient constraints; and using a multi-layer adaptive gated fusion mechanism to fuse low-frequency features and high-frequency features to generate spectral domain fusion features.

4. The multi-modal bridge vibration monitoring system according to claim 1, characterized in that: The process of optimizing the monitoring accuracy module and optimizing the hyperparameters of the Qiaoan fusion network includes the following steps: Step B1: define the search space of the hyperparameters of the Bridge-An fusion network and set the k and r double parameters; Step B2: Calculate the divergence gradient of the hyperparameters of the Bridge fusion network; Step B3: Combine the divergence gradient and use the k and r double parameters to update the hyperparameters of the Qiaoan fusion network; Step B4: Set the convergence conditions. If the convergence conditions are met, stop the optimization; otherwise, return to step B2 to continue updating the hyperparameters of the BridgeAn fusion network.

5. A multi-modal bridge vibration monitoring system according to claim 4, characterized in that: The search space in step B1 includes learning rate, momentum coefficient, regularization parameter and batch size; k, r dual parameters: k controls the smoothness of the change of the hyperparameters of the Bridge Fusion Network, and r controls the sensitivity of the adjustment of the hyperparameters of the Bridge Fusion Network.

6. A multi-modal bridge vibration monitoring system according to claim 5, characterized in that: Step B3 specifically includes the following steps: Step B31: Use k,r-exponential gradient to optimize the learning rate and momentum coefficient; Step B32: Optimizing regularization parameters using k,r-generalized regularization; Step B33: Update the batch size using k,r-additive hyperparameters; Step B34: Use Bregman divergence for collaborative optimization.

Images