A method for monitoring the perturbation and deformation state of a bridge

Abstract

This invention relates to a method for monitoring the deflection and deformation state of bridges. The method includes: acquiring real-time UWB radar data, time-series environmental parameter data, and time-series traffic load data; preprocessing to form a timestamp-aligned multimodal dataset; based on an online-deployed, pre-trained deflection and deformation state prediction model, using the real-time multimodal dataset as input, outputting real-time bridge deflection prediction values ​​and deformation state assessment results; periodically collecting actual bridge deflection annotation data using a high-precision ranging device, and periodically optimizing the deflection and deformation state prediction model using this actual deflection annotation data as a standard. This invention, based on UWB radar monitoring, can non-contactly and with high precision capture minute displacements on the bridge surface. Combined with multi-source environmental and load data, and through deep learning models for deep feature extraction and state assessment, it achieves a leap from simple data collection to intelligent diagnosis and early warning.

Description

Technical Field

[0001] This invention relates to the field of deep learning technology, and more particularly to a method for monitoring the deflection and deformation state of a bridge. Background Technology

[0002] As a crucial component of modern transportation infrastructure, the structural safety and long-term service performance of bridges directly impact public safety and the smooth operation of transportation lifelines. During long-term operation, bridges are subject to the combined effects of environmental erosion, material aging, and traffic loads, easily leading to structural response changes such as deflection and deformation. If these changes are not monitored and warned of in a timely and accurate manner, they may gradually accumulate and ultimately threaten the overall safety and service life of the bridge. Therefore, developing a long-term, real-time, and accurate health monitoring method for bridge deflection and deformation is of paramount importance for ensuring safe bridge operation and preventing structural accidents.

[0003] With the integration of sensing technology and artificial intelligence, bridge health monitoring is evolving towards intelligence and multi-source fusion. Traditional monitoring methods often have limitations in terms of long-term deployment costs, measurement accuracy, or environmental adaptability.

[0004] Current methods for monitoring bridge health have the following limitations: The limited data types monitored make it difficult to comprehensively assess the structural condition. Existing methods primarily focus on acquiring single, direct physical quantities such as deflection or displacement, failing to simultaneously integrate multi-source environmental and operational data, including temperature, humidity, and traffic loads. The lack of comprehensive analysis of the coupling relationship between these influencing factors and structural response restricts their ability to conduct a comprehensive and in-depth assessment of the overall health status of bridges.

[0005] Sensing methods are significantly constrained by environmental and deployment conditions. Visual image-based methods heavily rely on lighting conditions, making them difficult to operate at night, in foggy weather, or when visibility is poor, and are also susceptible to occlusion. Laser measurement methods based on fixed targets, on the other hand, are relatively complex to deploy and maintain, and their long-term stability in extreme weather or strong vibration environments faces challenges.

[0006] There is a lack of deep feature mining and intelligent state diagnosis capabilities. Current technologies mainly focus on the precise measurement of physical quantities, lacking effective automatic mining and analysis methods for the deep patterns hidden behind the collected time-series data. Therefore, intelligent diagnosis and early warning from data to state cannot be achieved, and interpretation still relies on human experience.

[0007] The system lacks long-term adaptive and predictive capabilities. Existing solutions mostly rely on static measurement or monitoring based on fixed rules, making it difficult to adjust model parameters once they are set. They cannot adapt to long-term performance drift in bridge structures caused by material aging and environmental changes, and lack mechanisms to continuously learn from new data to achieve model self-updates, optimize predictive capabilities, and provide early warnings of potential risks.

[0008] Therefore, it is necessary to improve one or more of the problems existing in the above-mentioned related technical solutions.

[0009] It should be noted that this section is intended to provide background or context for the technical solutions of the invention as set forth in the claims. The description herein does not imply acceptance as prior art simply because it is included in this section. Summary of the Invention

[0010] The purpose of this invention is to provide a method for monitoring the deflection and deformation state of bridges, thereby at least partially solving one or more of the problems caused by the limitations and defects of related technologies.

[0011] This invention provides a method for monitoring the deflection and deformation state of a bridge, comprising: Real-time UWB radar data is collected by an array of UWB radar sensors deployed on the bridge, and real-time environmental parameter time-series data and traffic load time-series data are collected by an auxiliary sensor array deployed on the bridge. A timestamp-aligned multimodal dataset is formed by preprocessing the real-time UWB radar data, the environmental parameter time series data, and the traffic load time series data. Based on a pre-trained deflection and deformation state prediction model deployed online, the real-time multimodal dataset is used as input to output the real-time deflection prediction value and deformation state evaluation result of the bridge. The actual deflection data of the bridge is collected periodically by a high-precision ranging device, and the deflection and deformation state prediction model is optimized periodically using the actual deflection data as a standard. The deflection and deformation state prediction model is trained using a historical multimodal dataset formed by historical UWB radar data, environmental parameter time series data, and traffic load time series data, and verified using historical real deflection annotation data and completed annotation health status.

[0012] Optionally, the step of acquiring real-time UWB radar data through a UWB radar sensor array deployed on the bridge includes: The UWB radar sensor arrays are deployed at key structural monitoring points at the mid-span, supports, and quarter points of the bridge.

[0013] Optionally, the step of collecting real-time environmental parameter time-series data and traffic load time-series data through an auxiliary sensor array deployed on the bridge includes: The data collected by the temperature sensor, humidity sensor, and anemometer deployed at the same location as the UWB radar sensor array are used together as the time series data of the environmental parameters. The traffic load time series data is generated by combining data collected from vehicle-mounted weight sensors and traffic flow cameras deployed at the bridgehead.

[0014] Optionally, the step of preprocessing the real-time UWB radar data, the environmental parameter time-series data, and the traffic load time-series data to form a timestamp-aligned multimodal dataset includes: The UWB radar data is preprocessed using a direct path signal separation technique based on cross-correlation analysis. Then, a bridge dynamic response depth feature extraction network is used to extract features from the preprocessed UWB radar data to generate a bridge dynamic response depth feature vector.

[0015] Optionally, the step of preprocessing the real-time UWB radar data, the environmental parameter time-series data, and the traffic load time-series data to form a timestamp-aligned multimodal dataset includes: The environmental parameter time series data and the traffic load time series data are preprocessed based on normalization processing and time dimension splicing processing. Then, the environmental and load feature extraction network of multilayer perceptron is used to extract features from the preprocessed environmental parameter time series data and traffic load time series data to generate a comprehensive environmental and load feature vector.

[0016] Optionally, the step of preprocessing the real-time UWB radar data, the environmental parameter time-series data, and the traffic load time-series data to form a timestamp-aligned multimodal dataset includes: The bridge health status representation module uses an adaptive weighted fusion method to fuse the bridge dynamic response depth feature vector and the comprehensive feature vector of environment and load to generate a bridge health status representation vector.

[0017] Optionally, the training process of the perturbation and deformation state prediction model includes the following steps: Using the bridge health status representation vector as a shared input, two tasks, deflection value regression and deformation status classification, are performed in parallel. The deflection value regression task aims to minimize the loss between the real-time deflection prediction value and the actual deflection labeled data. The deformation state classification task aims to minimize the loss between the deformation state assessment result and the healthy state; wherein, the healthy state is obtained by labeling the real deflection annotation data according to quantization rules. The deflection and deformation state prediction model is obtained by jointly training the optimization objectives of the deflection value regression task and the deformation state classification task.

[0018] Optionally, the quantification rule is set with an absolute deflection safety threshold and a relative change warning threshold. When the deflection estimate is lower than the relative change warning threshold, it is marked as "normal"; when the deflection estimate is greater than or equal to the relative change warning threshold but lower than the absolute deflection safety threshold, it is marked as "slight deformation"; and when the deflection estimate is greater than or equal to the absolute deflection safety threshold, it is marked as "severe deformation".

[0019] Optional, also includes: The bridge is monitored in real time based on the real-time deflection prediction value and deformation state assessment result of the bridge output by the deflection and deformation state prediction model, and an early warning signal is generated according to the dual threshold alarm mechanism.

[0020] The technical solution provided by this invention may include the following beneficial effects: In this invention, the monitoring method based on UWB radar can capture minute displacements on the bridge surface non-contactly and with high precision. Combined with multi-source environmental and load data, and through deep learning models for deep feature extraction and state assessment, it achieves a leap from simple data collection to intelligent diagnosis and early warning. It not only senses the "health pulse" of the bridge in real time but also provides a scientific basis for preventative maintenance and performance degradation assessment of bridge structures, promoting a transformation in bridge management from a passive response to a proactive prediction model. Attached Figure Description

[0021] The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments consistent with the invention and, together with the description, serve to explain the principles of the invention. It is obvious that the drawings described below are merely some embodiments of the invention, and those skilled in the art can obtain other drawings based on these drawings without any inventive effort.

[0022] Figure 1 This is a flowchart illustrating a method for monitoring the deflection and deformation state of a bridge in an exemplary embodiment of the present invention. Figure 2 A more detailed flowchart illustrating a method for monitoring the deflection and deformation state of a bridge in an exemplary embodiment of the present invention is shown. Figure 3This diagram illustrates the structure of a deflection prediction and deformation monitoring model based on multi-task learning and adaptive feature fusion in an exemplary embodiment of the present invention. Figure 4 This diagram illustrates the impact of multi-step UWB signal preprocessing on deflection prediction accuracy in an exemplary embodiment of the present invention. Figure 5 This diagram illustrates the long-term monitoring performance analysis of the multi-task learning model and incremental learning in an exemplary embodiment of the present invention. Detailed Implementation

[0023] Exemplary embodiments will now be described more fully with reference to the accompanying drawings. However, these exemplary embodiments can be implemented in many forms and should not be construed as limited to the examples set forth herein; rather, they are provided so that the invention will be more comprehensive and complete, and will fully convey the concept of the exemplary embodiments to those skilled in the art. The described features, structures, or characteristics may be combined in any suitable manner in one or more embodiments.

[0024] Furthermore, the accompanying drawings are merely illustrative diagrams of embodiments of the present invention and are not necessarily drawn to scale. The same reference numerals in the drawings denote the same or similar parts, and therefore repeated descriptions of them will be omitted. Some block diagrams shown in the drawings are functional entities and do not necessarily correspond to physically or logically independent entities.

[0025] This invention provides a method for monitoring the deflection and deformation state of a bridge, with reference to... Figure 1 As shown, it includes the following steps: Step S101: Collect real-time UWB radar data through the UWB radar sensor array deployed on the bridge, and collect real-time environmental parameter time series data and traffic load time series data through the auxiliary sensor array deployed on the bridge.

[0026] Step S102: A timestamp-aligned multimodal dataset is formed by preprocessing the real-time UWB radar data, the environmental parameter time series data, and the traffic load time series data.

[0027] Step S103: Based on the pre-trained deflection and deformation state prediction model deployed online, the real-time multimodal dataset is used as input to output the real-time deflection prediction value and deformation state evaluation result of the bridge.

[0028] Step S104: Periodically collect the actual deflection data of the bridge using a high-precision ranging device, and periodically optimize the deflection and deformation state prediction model using the actual deflection data as a standard.

[0029] The deflection and deformation state prediction model is trained using a historical multimodal dataset formed by historical ultra-wideband (UWB) radar data, environmental parameter time series data, and traffic load time series data, and validated using historical real deflection annotation data and completed annotation health status.

[0030] It is important to understand that reference Figure 2 As shown, firstly, to meet the needs of bridge response monitoring, a bridge response multimodal dataset is constructed. This is achieved by deploying a high-precision UWB radar sensor array at key structural monitoring locations of the bridge to collect UWB radar distance sequences reflecting minute displacement changes. Simultaneously, an auxiliary sensor array is deployed to collect time-series data of environmental parameters and traffic loads. All collected data are timestamped and aligned, with measurements from a high-precision laser rangefinder used as the actual deflection annotation data. Based on the set deflection safety threshold and deformation mode discrimination criteria, regularized health status annotations are generated. Finally, the cleaned and aligned UWB radar distance sequences, environmental parameter time-series data, traffic load time-series data, and their corresponding annotations are integrated to construct the bridge response multimodal dataset.

[0031] It is also necessary to understand the time-stamp alignment of the collected UWB radar distance sequences, environmental parameter time-series data, and traffic load time-series data. A unified clock source is provided for all sensors using a network time protocol, ensuring that data from different sources have a precise correspondence at the same point in time. Based on this, a high-precision laser rangefinder is introduced as a verification benchmark to measure the actual deflection values ​​of key monitoring points on the bridge at fixed time intervals, obtaining accurate deflection annotation data.

[0032] It is also necessary to understand the integration of the cleaned and aligned UWB radar range sequences, environmental parameter time-series data, and traffic load time-series data with the corresponding actual deflection annotation data and rule-based health status annotation information. Specifically, all data samples within the same time window are packaged to form a bridge health monitoring multimodal data sample containing both raw observation data and rule-based annotation information. Ultimately, a large number of such samples are used to construct a bridge response multimodal dataset for subsequent deep learning model training and validation.

[0033] It is also necessary to understand that the obtained data is input into the UWB radar data processing and feature extraction module based on signal separation and cascaded networks. First, the UWB radar range sequence is preprocessed based on cross-correlation analysis and filtering to obtain an effective range change sequence that excludes multipath effects. Second, a bridge dynamic response depth feature extraction network based on one-dimensional convolution and gated recurrent units is used to extract the encoded local pattern and long-term evolution bridge dynamic response depth feature vectors. At the same time, an environment and load feature extraction network based on multilayer perceptrons is used to process environmental parameter time series data and traffic load time series data to generate an enhanced comprehensive environmental and load feature vector.

[0034] It is also necessary to understand that the extracted features are input into the deflection prediction and deformation monitoring model based on multi-task learning and adaptive feature fusion. First, the bridge state representation module based on adaptive weighted fusion fuses the bridge dynamic response deep feature vector and the comprehensive feature vector of environment and load to generate a bridge health state representation vector. Second, a multi-task learning monitoring model is constructed, which uses this shared representation vector to perform deflection value regression and deformation state classification tasks in parallel, and is collaboratively trained through a joint loss function.

[0035] It is also important to understand that after the model is deployed online, a long-term adaptive optimization strategy based on incremental learning is implemented. New batches of monitoring data are collected regularly, and the model is adaptively optimized by fixing the parameters of the feature extraction network and fusion module, fine-tuning the parameters of the multi-task monitoring model, and combining elastic weights to consolidate the regularization term. This is to adapt to the long-term performance drift of the bridge and avoid catastrophic forgetting.

[0036] It's also important to understand that after online model deployment, a real-time feedback and adaptive optimization closed loop is constructed, encompassing performance evaluation and parameter updates. The core of this strategy is an incremental learning mechanism for bridge health monitoring. During real-time monitoring, the system periodically collects new UWB radar distance sequences, environmental parameter time-series data, traffic load time-series data, and the latest real deflection data labeled through automated measurement, and labels them with health status, forming new monitoring data batches. When the incremental learning mechanism is activated, the parameters of the bridge dynamic response deep feature extraction network and the channel attention-based adaptive weighted fusion module are first fixed, and only the fully connected layer parameters of the regression and classification branches of the multi-task deep learning monitoring model are unfrozen. Subsequently, the model is fine-tuned in small batches using the new monitoring data batches at a smaller learning rate than the initial training. This process, by introducing elastic weights to consolidate the regularization term into the loss function, constrains the variation of important parameters, thereby enabling the model to adapt to long-term performance drift caused by bridge material aging and environmental changes, while effectively avoiding catastrophic forgetting of previously learned knowledge.

[0037] It is also important to understand that in the system integration and intelligent monitoring implementation phase, the optimized complete model is integrated into the bridge health monitoring system software platform and communication is established with the on-site sensor array; the system receives data in real time and automatically executes processing and inference processes, synchronously outputting real-time deflection prediction values ​​and deformation state assessment results, realizing visualization and dual-threshold alarms; the system stores data for a long time and supports analysis, providing support for bridge performance evaluation and maintenance decisions.

[0038] It's also important to understand the hardware and software system integration: the complete deep learning model, trained and optimized through policies, along with all modules, is integrated into a bridge health monitoring system software platform deployed on a cloud server. This platform establishes a stable, low-latency data communication link with the UWB radar sensor array and auxiliary sensor array deployed at the bridge site via an industrial IoT gateway.

[0039] It is also important to understand that existing methods often rely on single types of displacement or image data, making it difficult to comprehensively assess the impact of environment and load on structural response. This invention constructs a spatiotemporally aligned multimodal dataset by simultaneously deploying multiple sensors, including UWB radar, temperature and humidity sensors, wind speed sensors, and traffic load sensors. It then dynamically integrates the heterogeneous features of the structure and environment using an adaptive weighted fusion module. This overcomes the limitations of traditional monitoring methods with their single perspective, enabling the model to comprehensively perceive and correlate the complex coupling relationships between deflection changes and the environment and operational loads, thereby achieving a more comprehensive and reliable assessment of the bridge's health status. This achieves deep fusion of multi-source information and comprehensive condition assessment.

[0040] It's also important to understand that traditional visual methods are susceptible to lighting and weather conditions, while laser measurements require sophisticated deployment and maintenance. This invention innovatively employs UWB direct path signal separation technology based on cross-correlation analysis, effectively suppressing multipath interference caused by the complex metal structure of bridges and ensuring the quality of the original data. Simultaneously, through a cascaded convolutional and gated recurrent unit network specifically designed for time-series signals, it automatically mines deeper features such as local vibration modes and long-term evolution patterns inherent in the bridge response, surpassing the limitations of traditional methods that only provide physical quantity readings. This enhances the robustness of monitoring and the level of intelligent diagnosis in complex environments.

[0041] It is also important to understand that existing technologies typically separate deflection measurement and safety assessment into two independent stages. This invention constructs a multi-task learning model with shared features as input, using a joint loss function to collaboratively optimize the regression and classification branches. This architecture enables the model to pursue numerical accuracy in deflection prediction while also enhancing its ability to discriminate the overall safety status of the structure, improving the practicality of monitoring results and the efficiency of decision support. An innovative multi-task collaborative learning framework is used to simultaneously achieve accurate prediction and state classification.

[0042] It is also important to understand that the condition of bridges undergoes long-term, slow changes due to material aging and environmental variations, and traditional monitoring models, once fixed, struggle to adapt. This invention introduces an incremental learning-based optimization strategy during the online deployment phase. This strategy, by periodically using new data, fixing the core feature extraction, fine-tuning the task network, and combining it with anti-forgetting regularization techniques, enables the model to smoothly adapt to the long-term performance evolution of the bridge structure during continuous service. This effectively solves the problem of performance drift in static models, ensuring the accuracy and reliability of the monitoring system throughout its entire lifecycle. It endows the system with long-term adaptive evolutionary capabilities to cope with structural performance drift.

[0043] The aforementioned method for monitoring bridge deflection and deformation, based on UWB radar, can capture minute displacements on the bridge surface non-contactly and with high precision. Combined with multi-source environmental and load data, and through deep learning models for deep feature extraction and state assessment, it achieves a leap from simple data collection to intelligent diagnosis and early warning. This not only allows for real-time sensing of the bridge's "health pulse" but also provides a scientific basis for preventative maintenance and performance degradation assessment of bridge structures, promoting a shift in bridge management from a passive response to a proactive prediction model.

[0044] The following will describe in more detail each step of the method for monitoring the deflection and deformation state of a bridge in this example embodiment.

[0045] In some embodiments, step S101 includes: The UWB radar sensor arrays are deployed at key structural monitoring points at the mid-span, supports, and quarter points of the bridge.

[0046] It is important to understand that high-precision UWB radar sensor arrays are deployed at key structural monitoring points of the bridge, including mid-span, supports, and quarter points. At least one UWB radar sensor is deployed at each monitoring point. This sensor transmits ultra-wideband pulse signals to preset measuring points on the bridge surface and receives reflected signals. By calculating the time difference between transmission and reception, the UWB radar distance sequence of that measuring point relative to the radar base station is obtained in real time. This sequence reflects minute displacement changes of the bridge at that measuring point, thus indirectly characterizing the bridge's vibration and deflection deformation.

[0047] In some embodiments, step S101 includes: The data collected by the temperature sensor, humidity sensor, and anemometer deployed at the same location as the UWB radar sensor array are used together as the time series data of the environmental parameters.

[0048] The traffic load time series data is generated by combining data collected from vehicle-mounted weight sensors and traffic flow cameras deployed at the bridgehead.

[0049] It is important to understand that an auxiliary sensor array, including temperature sensors, humidity sensors, and anemometers, is deployed simultaneously to collect time-series data of environmental parameters; and vehicle-mounted weight sensors and traffic flow cameras deployed at the bridgeheads collect time-series data of traffic loads to analyze the impact of environmental and operational loads on the bridge structure response.

[0050] In some embodiments, step S102 includes: The UWB radar data is preprocessed using a direct path signal separation technique based on cross-correlation analysis. Then, a bridge dynamic response depth feature extraction network is used to extract features from the preprocessed UWB radar data to generate a bridge dynamic response depth feature vector.

[0051] It is important to understand the UWB radar signal preprocessing based on cross-correlation analysis and filtering: UWB radar range sequences from the obtained bridge response multimodal dataset Preprocessing is performed, where dt represents the raw distance measurement value at time t. First, the sequence for each measurement point is processed. A sliding window approach was used for outlier review and removal, and the data within the window was further processed using the Laida criterion. Subsequently, a filter was applied to smooth the sequence to suppress high-frequency noise. This filter smoothed the data within a local window using polynomial least squares fitting, and its output was the smoothed distance value. To address the issue of UWB signals being susceptible to multipath interference in complex bridge environments, an innovative direct path signal separation technique based on cross-correlation analysis is employed. This technique uses the ideal reflected signal waveform of the target object as a template. The actual signal received by the radar Perform cross-correlation calculation with the template signal: By finding the cross-correlation function The first significant peak and its corresponding delay This allows for the separation of the strongest direct path component from the mixed signal, and based on this, a more accurate effective distance variation sequence that excludes multipath effects can be calculated. ,in This represents the effective distance value of the measurement point obtained at time t after outlier processing, filtering, and direct path separation.

[0052] A deep feature extraction network for bridge dynamic response based on one-dimensional convolution and gated recurrent units: The output is the effective distance change sequence of all M measuring points. The data is organized into an M×T matrix and used as input to the bridge dynamic response deep feature extraction network. This network is a cascaded structure specifically designed for bridge UWB time-series signals, aiming to automatically mine deep spatiotemporal features. Its main body consists of a local pattern-aware one-dimensional convolutional module and a long-term evolution modeling gated recurrent unit module connected in series.

[0053] The specific network structure and forward propagation process are as follows: The input layer receives the aforementioned M×T input matrix.

[0054] The local pattern-aware one-dimensional convolutional module is used to capture the local spatiotemporal correlations and short-term patterns of the bridge response. This module contains two cascaded one-dimensional convolutional sub-blocks. The first sub-block consists of a single one-dimensional convolutional layer using 32 kernels of width 5, performing convolution along the time dimension with a stride of 1, followed by a ReLU activation function and a one-dimensional max-pooling layer of width 2. This sub-block outputs a feature map with dimensions M×F1×T / 2 (F1 being the number of feature maps), which encodes short-term vibration pattern features in the signals from each measurement point, such as high-frequency flutter. The second sub-block also consists of a single one-dimensional convolutional layer using 64 kernels of width 3, followed by a ReLU activation function and a pooling layer of width 2. This sub-block further integrates the local correlations between different measurement points, outputting a mid-level spatiotemporal correlation feature map F with dimensions M×F2×T / 4. mid (where F2=64), where F mid This represents the feature tensor extracted after two layers of convolution and pooling, which contains the preliminary correlation between local patterns and measurement points.

[0055] The long-term evolution modeling gated recurrent unit module is used to capture the long-term temporal dependency of bridge deformation and vibration. The intermediate spatiotemporal correlation feature map F... mid Unfolding along the time dimension, the sequence is input into this module. This module consists of two stacked layers of gated recurrent units (GRUs). The first GRU layer has 64 hidden units and receives F... mid The sequence captures the mid-term evolution patterns over time. The second-layer GRU also has 64 hidden units, receiving the hidden state sequence from the first-layer GRU to further abstract and capture long-term trends and dynamic patterns. Finally, the hidden state of the last time step of the second-layer GRU is taken to output a fixed-length bridge dynamic response deep feature vector H. gru H gru This represents the feature vector output by the GRU module, which encodes the dynamic evolution of bridge deflection deformation throughout the observation period.

[0056] In some embodiments, step S102 includes: The environmental parameter time series data and the traffic load time series data are preprocessed based on normalization processing and time dimension splicing processing. Then, the environmental and load feature extraction network of multilayer perceptron is used to extract features from the preprocessed environmental parameter time series data and traffic load time series data to generate a comprehensive environmental and load feature vector.

[0057] It is important to understand that this is an environment and load feature extraction network based on a multilayer perceptron. Feature extraction is performed on the time-series data of environmental parameters and traffic loads from the obtained bridge response multimodal dataset. First, missing data is filled using linear interpolation, followed by min-max normalization to map each parameter value to the [0,1] interval. The processed multi-channel auxiliary parameter sequences are concatenated along the time dimension to form a two-dimensional matrix, which is then input into a three-layer fully connected feature extraction network. The network structure is as follows: the first layer (input layer) maps the concatenated features to a 128-dimensional space, followed by a ReLU activation function; the second layer (hidden layer) maps the 128-dimensional features to a 64-dimensional space, followed by ReLU activation; the third layer (output layer) maps the 64-dimensional features to a 32-dimensional space, generating a fixed-dimensional integrated environment and load feature vector F. env , where F env This represents the comprehensive feature representation extracted by the MLP network, which integrates static and dynamic information from multiple auxiliary parameters. Simultaneously, the mean, variance, maximum, and minimum statistics of these auxiliary parameters within the corresponding time window are calculated and compared with F... env The vectors are concatenated to form an enhanced integrated feature vector of environment and load. .

[0058] In some embodiments, step S102 includes: The bridge health status representation module uses an adaptive weighted fusion method to fuse the bridge dynamic response depth feature vector and the comprehensive feature vector of environment and load to generate a bridge health status representation vector.

[0059] It is important to understand that reference Figure 3 As shown, the bridge state representation module based on adaptive weighted fusion: The extracted bridge dynamic response depth feature vector and the extracted enhanced integrated feature vector of environment and load The input is an adaptive weighted fusion module based on channel attention to form a unified and robust bridge health status representation vector. The core of this module is a lightweight single-head attention mechanism, and its workflow is as follows: First, the input is... and The initial fusion features are obtained by splicing. , where [;] denotes the concatenation operation. Subsequently, through three learnable linear transformation matrices... , and Calculate the query vector separately Key vector Sum value vector Next, the attention weights are calculated. ,in This refers to the dimension of the key vector. The softmax function normalizes along the last dimension to ensure the sum of the weights is 1. Finally, the attention weights are used to perform a weighted summation of the value vectors to obtain the attention output. This process enables the model to dynamically evaluate and enhance the channel features most relevant to the current monitoring task from UWB and environmental load characteristics. Then, a fully connected layer is used for nonlinear transformation and dimensionality reduction to output the final bridge health status representation vector. ,in This represents the high-level feature vector that comprehensively characterizes the instantaneous health status of the bridge after adaptive channel weighted fusion.

[0060] In some embodiments, step S103 includes: Using the bridge health status representation vector as a shared input, two tasks, deflection value regression and deformation status classification, are performed in parallel.

[0061] The deflection value regression task aims to minimize the loss between the real-time deflection prediction value and the actual deflection annotation data.

[0062] The deformation state classification task aims to minimize the loss between the deformation state assessment result and the healthy state; wherein the healthy state is obtained by labeling the real deflection annotation data according to quantization rules.

[0063] The deflection and deformation state prediction model is obtained by jointly training the optimization objectives of the deflection value regression task and the deformation state classification task.

[0064] It is important to understand that reference Figure 3 As shown, the construction and joint training of the multi-task learning monitoring model are as follows: Construct a multi-task deep learning monitoring model, which uses bridge health status representation vectors. To share input, two tasks, deflection value regression and deformation state classification, are executed in parallel, achieving collaborative optimization through shared representation learning.

[0065] The regression branch is used to accurately predict deflection values: First, a regression-specific fully connected layer (containing 128 neurons, activated using ReLU) is passed through, followed by a fully connected layer containing 64 neurons (also activated using ReLU). Finally, a linear output layer without an activation function is passed through to directly output the real-time deflection prediction values ​​of key bridge measurement points. This branch uses the mean squared error loss function. Optimization was carried out, including The actual deflection data is labeled, and N is the number of training samples.

[0066] The classification branch is used for deformation state early warning: it shares input with the regression branch. It passes through another separate fully connected layer dedicated to classification (containing 64 neurons, using ReLU activation), and is finally connected to a softmax output layer, which outputs the probability distribution of the deformation state. These correspond to the predicted probabilities for the three categories: "normal," "slightly deformed," and "severely deformed," respectively. This branch uses the cross-entropy loss function. Optimization is performed, where C=3 represents the number of categories. This is the one-hot encoded vector corresponding to the health status label based on the rules.

[0067] The two branches are connected by a joint loss function. Co-training is performed, where α and β are hyperparameters used to balance the weights of the regression and classification tasks. This is achieved by minimizing... During backpropagation, the model simultaneously updates all parameters in the adaptively weighted fusion bridge state representation module and the multi-task learning monitoring model construction and joint training, ensuring a shared bridge health state representation vector. It can simultaneously serve accurate deflection estimation and accurate state classification. The training data comes from a historically constructed bridge response multimodal dataset.

[0068] In some embodiments, the quantification rule is set with an absolute deflection safety threshold and a relative change warning threshold. When the deflection estimate is lower than the relative change warning threshold, it is marked as "normal"; when the deflection estimate is greater than or equal to the relative change warning threshold but lower than the absolute deflection safety threshold, it is marked as "slight deformation"; and when the deflection estimate is greater than or equal to the absolute deflection safety threshold, it is marked as "severe deformation".

[0069] It is important to understand that the health status labeling process follows clear quantitative rules: First, referring to the relevant specifications regarding the deflection limits for beam bridges, an absolute deflection safety threshold La and a relative change warning threshold Lr are set. Second, two typical deformation mode discrimination criteria are defined: a uniform settlement mode is characterized by a consistent, slow, unidirectional change in the distance sequence of all measuring points over a long period; a local deformation mode is characterized by a trend divergence or abrupt change in the distance sequence of some adjacent measuring points. The health status of data samples for each time period is labeled according to the following rules: if the estimated deflection values ​​calculated from the UWB radar distance sequence for all measuring points within that time period are all lower than Lr, and trend analysis does not detect any obvious deformation mode, it is labeled as "normal"; if the estimated deflection value of any measuring point exceeds Lr but is lower than La, or trend analysis detects preliminary characteristics of a uniform settlement mode or a local deformation mode, it is labeled as "slight deformation"; if the estimated deflection value of any measuring point exceeds La, or trend analysis confirms the existence of a significant, developing deformation mode, it is labeled as "severe deformation".

[0070] In some embodiments, the following steps are included after step S104: The bridge is monitored in real time based on the real-time deflection prediction value and deformation state assessment result of the bridge output by the deflection and deformation state prediction model, and an early warning signal is generated according to the dual threshold alarm mechanism.

[0071] It's important to understand that real-time intelligent monitoring and early warning involves a complete processing and reasoning process: the system receives sensor data in real time and automatically triggers a UWB radar data processing and feature extraction module based on signal separation and cascaded networks, culminating in a deflection prediction and deformation monitoring model based on multi-task learning and adaptive feature fusion. Specifically, the data is processed by the UWB radar data processing and feature extraction module based on signal separation and cascaded networks, features are extracted, and then input into the deflection prediction and deformation monitoring model based on multi-task learning and adaptive feature fusion. Finally, the system synchronously outputs real-time deflection prediction values ​​and deformation state assessment results. The system then visualizes these results. Simultaneously, a built-in dual-threshold alarm mechanism is implemented: when the real-time deflection prediction value exceeds the design safety threshold, or the deformation state assessment result is "severe deformation," the system automatically triggers multi-level early warning signals to promptly notify bridge maintenance personnel.

[0072] Data Analysis and Decision Support: The system stores all monitoring data, features, and model outputs long-term, supporting multi-dimensional historical data backtracking and health trend analysis by time, location, and event. Through the mining of long-term data, it provides data-driven decision support for assessing bridge performance degradation, developing preventative maintenance plans, and predicting remaining service life, thereby achieving long-term, intelligent, and adaptive health monitoring of bridge deflection and deformation.

[0073] Based on the methods described above for monitoring the deflection and deformation of bridges, the following experimental comparison will provide a more detailed explanation.

[0074] To verify the effectiveness of the bridge health monitoring method based on UWB multimodal data and multi-task learning proposed in this invention, this experiment uses long-term monitoring data from three typical beam bridges (simply supported beam bridge, continuous beam bridge, and T-beam bridge) as the research object, focusing on evaluating the accuracy improvement effect of UWB signal preprocessing and the long-term adaptive monitoring performance of the multi-task learning model. The core experimental indicators include the root mean square error of deflection prediction and the accuracy of deformation state classification.

[0075] 1. Analysis of the impact of multi-step preprocessing of UWB signals on deflection prediction accuracy The purpose of the experiment is to verify the effectiveness of the UWB radar signal preprocessing process (outlier removal, polynomial least squares smoothing filtering, and cross-correlation direct path multipath separation), quantify the contribution of each step to improving the deflection prediction accuracy, and provide reliable data support for subsequent depth feature extraction.

[0076] Twelve key measurement points (four each at mid-span, supports, and quarter points) from three typical beam bridges in the constructed bridge response multimodal dataset were selected, and a total of 1000 continuous time windows of sample data were extracted. Using the unprocessed raw UWB distance sequence as a baseline, three steps were sequentially performed: outlier removal, multinomial least squares smoothing filtering, and cross-correlation direct path multipath separation. The sequences processed in each step were then input into the regression branch of the multi-task learning model. The root mean square error of deflection prediction at each stage was calculated and statistically analyzed, serving as the core evaluation metric.

[0077] refer to Figure 4 As shown in the figure, the experimental results are presented intuitively. It can be clearly observed from the graph that both error lines exhibit a significant monotonically decreasing trend, and their patterns of change are highly consistent: 1. The original UWB distance sequence had the highest deflection prediction error (average error of 0.85 for each measuring point and error of 0.92 for the mid-span measuring point). This is because the original data contains a large amount of random noise, outliers, and multipath interference caused by the complex environment, which cannot accurately characterize the actual deflection changes of the bridge.

[0078] 2. After outlier removal, the error decreased significantly (the average error dropped to 0.62, and the error at the mid-span measurement point dropped to 0.68), indicating that the outlier handling mechanism of the sliding window and the Laida criterion can effectively remove sudden interference in the measurement point data and initially improve the data quality.

[0079] 3. The smoothing filtering step further reduced the error (the average error was reduced to 0.45, and the error at the mid-span measurement point was reduced to 0.49). The local smoothing effect of polynomial least squares fitting can effectively suppress high-frequency noise and restore the true trend of bridge deflection.

[0080] 4. The cross-correlation direct path multipath separation step brought the most significant error reduction (average error reduced to 0.21, mid-span measurement point reduced to 0.23), with a reduction of more than 50%. This shows that the method can effectively separate the direct path component from the mixed signal and eliminate the influence of multipath interference on UWB ranging. It is the core step to improve data accuracy.

[0081] The above results verify the scientific validity and effectiveness of the UWB signal preprocessing process, with each step progressing progressively and complementing each other for enhanced efficiency.

[0082] 2. Long-term monitoring performance analysis of multi-task learning models and incremental learning The purpose of the experiment is to verify the core performance of the multi-task learning model, the effect of the incremental learning strategy on the long-term adaptive optimization of the model, and to evaluate the robustness and practicality of the model in the scenario of slow drift in bridge performance.

[0083] The trained multi-task learning model was deployed to a cloud monitoring platform for three beam bridges for continuous online monitoring over a period of six months. Two control experiments were set up: 1. Control group (no incremental learning model): No parameters are updated after deployment, and long-term online inference is performed directly.

[0084] 2. Experimental group (with incremental learning model): Following the long-term adaptive optimization strategy of the incremental learning model, new data batches are collected monthly. The parameters of the UWB radar data processing and feature extraction module based on signal separation and cascaded network and the bridge state representation module based on adaptive weighted fusion are fixed. Only the parameters of the fully connected layer of the multi-task deep learning monitoring model branch are fine-tuned. Elastic weights are introduced into the loss function to consolidate the regularization term and avoid catastrophic forgetting.

[0085] The root mean square error of deflection prediction and the accuracy of deformation state classification of the two models were statistically analyzed monthly as core evaluation indicators to compare and analyze the long-term performance changes of the two models.

[0086] refer to Figure 5 As shown, the experimental results are obtained through... Figure 5 The performance differences between the two models are clearly illustrated by contrasting trends. 1. Control Group (No Incremental Learning Model): Both performance curves show a clear deterioration trend. The root mean square error of deflection prediction climbed continuously from 0.21 in Month 1 to 0.78 in Month 6, an increase of over 270%; the accuracy of deformation state classification decreased continuously from 0.95 in Month 1 to 0.68 in Month 6, a decrease of nearly 29%. This is because, during long-term use, factors such as material aging, long-term changes in environmental temperature and humidity, and drift in traffic load distribution cause a slow shift in data distribution. The fixed-parameter model cannot adapt to these changes, resulting in catastrophic forgetting. The previously learned knowledge of bridge health status gradually becomes invalid, ultimately leading to a significant decline in performance and failing to meet the engineering requirements for long-term monitoring.

[0087] 2. Experimental Group (with Incremental Learning Model): The two performance curves consistently remained within the stable optimal range, showing no significant deterioration trend. The root mean square error of deflection prediction remained relatively stable at around 0.20, with a maximum of no more than 0.22; the deformation state classification accuracy remained stable between 0.94 and 0.96, with almost no significant decrease. This demonstrates that the incremental learning strategy proposed by the long-term adaptive optimization strategy based on incremental learning can effectively play its role: fixing the core parameters of the feature extraction and fusion modules avoids the degradation of core feature learning capabilities; only fine-tuning the parameters of the fully connected layers in the multi-task branches achieves lightweight updates and reduces computational costs; the elastic weight consolidation regularization term can constrain the variation range of important parameters, effectively avoiding the forgetting of previous knowledge, enabling the model to adapt to the long-term performance drift of the bridge and continuously maintain high-precision deflection prediction and high-accuracy deformation classification.

[0088] The above results verify the core effectiveness of the multi-task learning model (achieving high-precision prediction and high-accuracy classification in the initial stage), and also verify the long-term optimization value of the model's long-term adaptive optimization strategy based on incremental learning. This ensures the robustness and practicality of the model in the long-term monitoring scenario of bridges, and provides technical support for the long-term intelligent health monitoring of bridges.

[0089] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. In addition, those skilled in the art can combine different embodiments or examples described in this specification.

[0090] Other embodiments of the invention will readily occur to those skilled in the art upon consideration of the specification and practice of the invention disclosed herein. This invention is intended to cover any variations, uses, or adaptations of the invention that follow the general principles of the invention and include common knowledge or customary techniques in the art not disclosed herein. The specification and examples are to be considered exemplary only, and the true scope and spirit of the invention are indicated by the appended claims.

Claims

1. A method for monitoring the deflection and deformation state of a bridge, characterized in that, include: Real-time UWB radar data is collected by an array of UWB radar sensors deployed on the bridge, and real-time environmental parameter time-series data and traffic load time-series data are collected by an auxiliary sensor array deployed on the bridge. A timestamp-aligned multimodal dataset is formed by preprocessing the real-time UWB radar data, the environmental parameter time series data, and the traffic load time series data. Based on a pre-trained deflection and deformation state prediction model deployed online, the real-time multimodal dataset is used as input to output the real-time deflection prediction value and deformation state evaluation result of the bridge. The actual deflection data of the bridge is collected periodically by a high-precision ranging device, and the deflection and deformation state prediction model is optimized periodically using the actual deflection data as a standard. The deflection and deformation state prediction model is trained using a historical multimodal dataset formed by historical UWB radar data, environmental parameter time series data, and traffic load time series data, and verified using historical real deflection annotation data and completed annotation health status.

2. The method for monitoring the deflection and deformation state of a bridge according to claim 1, characterized in that, The step of acquiring real-time UWB radar data through a UWB radar sensor array deployed on the bridge includes: The UWB radar sensor arrays are deployed at key structural monitoring points at the mid-span, supports, and quarter points of the bridge.

3. The method for monitoring the deflection and deformation state of a bridge according to claim 1, characterized in that, The steps of collecting real-time environmental parameter time-series data and traffic load time-series data through an auxiliary sensor array deployed on the bridge include: The data collected by the temperature sensor, humidity sensor, and anemometer deployed at the same location as the UWB radar sensor array are used together as the time series data of the environmental parameters. The traffic load time series data is generated by combining data collected from vehicle-mounted weight sensors and traffic flow cameras deployed at the bridgehead.

4. The method for monitoring the deflection and deformation state of a bridge according to claim 1, characterized in that, The step of preprocessing the real-time UWB radar data, the environmental parameter time-series data, and the traffic load time-series data to form a timestamp-aligned multimodal dataset includes: The UWB radar data is preprocessed using a direct path signal separation technique based on cross-correlation analysis. Then, a bridge dynamic response depth feature extraction network is used to extract features from the preprocessed UWB radar data to generate a bridge dynamic response depth feature vector.

5. The method for monitoring the deflection and deformation state of a bridge according to claim 4, characterized in that, The step of preprocessing the real-time UWB radar data, the environmental parameter time-series data, and the traffic load time-series data to form a timestamp-aligned multimodal dataset includes: The environmental parameter time series data and the traffic load time series data are preprocessed based on normalization processing and time dimension splicing processing. Then, the environmental and load feature extraction network of multilayer perceptron is used to extract features from the preprocessed environmental parameter time series data and traffic load time series data to generate a comprehensive environmental and load feature vector.

6. The method for monitoring the deflection and deformation state of a bridge according to claim 5, characterized in that, The step of preprocessing the real-time UWB radar data, the environmental parameter time-series data, and the traffic load time-series data to form a timestamp-aligned multimodal dataset includes: The bridge health status representation module uses an adaptive weighted fusion method to fuse the bridge dynamic response depth feature vector and the comprehensive feature vector of environment and load to generate a bridge health status representation vector.

7. The method for monitoring the deflection and deformation state of a bridge according to claim 6, characterized in that, The training process of the perturbation and deformation state prediction model includes the following steps: Using the bridge health status representation vector as a shared input, two tasks, deflection value regression and deformation status classification, are performed in parallel. The deflection value regression task aims to minimize the loss between the real-time deflection prediction value and the actual deflection labeled data. The deformation state classification task aims to minimize the loss between the deformation state assessment result and the healthy state; wherein, the healthy state is obtained by labeling the real deflection annotation data according to quantization rules. The deflection and deformation state prediction model is obtained by jointly training the optimization objectives of the deflection value regression task and the deformation state classification task.

8. The method for monitoring the deflection and deformation state of a bridge according to claim 7, characterized in that, The quantification rules are set with an absolute deflection safety threshold and a relative change warning threshold. When the deflection estimate is lower than the relative change warning threshold, it is marked as "normal"; when the deflection estimate is greater than or equal to the relative change warning threshold but lower than the absolute deflection safety threshold, it is marked as "slight deformation"; and when the deflection estimate is greater than or equal to the absolute deflection safety threshold, it is marked as "severe deformation".

9. The method for monitoring the deflection and deformation state of a bridge according to any one of claims 1-8, characterized in that, Also includes: The bridge is monitored in real time based on the real-time deflection prediction value and deformation state assessment result of the bridge output by the deflection and deformation state prediction model, and an early warning signal is generated according to the dual threshold alarm mechanism.

Images