Bridge deflection reconstruction method and system based on time multi-scale feature fusion

Abstract

The invention provides a bridge deflection reconstruction method and system based on time multi-scale feature fusion, relates to the technical field of micro expression recognition, and aims to solve the problems that the existing monitoring technology is easily influenced by environmental variables, only depends on long-term stable reference points, is inaccurate in monitoring data and is low in model reconstruction precision. The method comprises the steps of obtaining health monitoring data of a bridge, wherein the health monitoring data at least comprise strain data and vibration data; preprocessing the health monitoring data; inputting the preprocessed strain data and vibration data into a strain feature extraction module and a vibration feature extraction module respectively, and extracting strain features and vibration features respectively; and inputting to a fusion feature extraction module for fusion to obtain a reconstruction deflection. According to the method, the problems that the method is easily influenced by environmental variables and can only depend on long-term stable reference points are solved, the limitation of a traditional monitoring means is avoided, and accurate reconstruction of the bridge deflection is achieved under the condition that extra hardware cost is not increased.

Description

Technical Field

[0001] The present invention belongs to the technical field of traffic engineering, and in particular relates to a bridge deflection reconstruction method and system based on time multi-scale feature fusion. Background Art

[0002] The statements in this section merely provide background information related to the present invention and do not necessarily constitute prior art.

[0003] As a key infrastructure of modern transportation networks, the health of bridges is directly related to traffic safety and efficiency. As bridges around the world gradually age, many bridges face the risk of structural degradation and functional failure, and there is an urgent need to strengthen the monitoring and maintenance of their safety and structural integrity. In recent years, the number of bridges in China has continued to increase, and the growth rate is accelerating. According to the "Statistical Communiqué on the Development of the Transportation Industry" released by the Ministry of Transport in June 2024, by the end of 2023, the number of highway bridges in the country will reach 1.0793 million, an increase of 46,100 from the end of the previous year. The statistical data of the Statistical Communiqué on the Development of the Transportation Industry in recent years show that the number of highway bridges in China is huge and the growth trend is obvious.

[0004] In the field of bridge structural health monitoring (SHM), bridge deflection is an important indicator of bridge structural performance and can reflect internal structural problems of the bridge. Deflection is the vertical deformation of the bridge under load, which directly reflects the stress state and structural stiffness of the bridge. Therefore, accurate monitoring of these parameters is of great significance for timely detection and prevention of structural problems.

[0005] Currently, the methods for obtaining bridge deflection include monitoring methods and reconstruction methods. However, existing detection methods have various problems in long-term measurements. Traditional metrology methods are flexible but difficult to automate in long-term observations. Advanced monitoring methods include machine vision-based monitoring, laser sensor monitoring, continuous wave monitoring, and sensor fusion methods. However, these methods are easily affected by environmental variables such as occlusion, wind-induced shaking, and changes in lighting conditions, making the collected data inaccurate. In addition, since these methods only rely on long-term stable reference points, the model accuracy is low. They are not always feasible in practical applications. Therefore, the application of existing monitoring methods faces many limitations. Summary of the invention

[0006] To overcome the deficiencies of the above-mentioned existing technologies, the present invention provides a method and system for reconstructing bridge deflection based on time multi-scale feature fusion. By combining data-driven with a convolutional neural network model based on time multi-scale feature fusion, feature extraction and fusion are respectively performed on strain data and vibration data, thereby integrating strain, vibration, and environmental data. This can not only avoid the limitations of traditional monitoring methods but also achieve accurate reconstruction of bridge deflection without increasing additional hardware costs. It provides more scientific data support for bridge health assessment, helps extend the service life of bridges, reduces the risk of accidents, and ultimately ensures public traffic safety and reduces losses.

[0007] To achieve the above object, one or more embodiments of the present invention provide the following technical solutions:

[0008] The first aspect of the present invention discloses a method for reconstructing bridge deflection based on time multi-scale feature fusion, including:

[0009] Obtain the health monitoring data of the bridge, where the health monitoring data of the bridge at least includes strain data and vibration data;

[0010] Preprocess the health monitoring data of the bridge to obtain preprocessed strain data and vibration data;

[0011] Input the preprocessed strain data and vibration data into a strain feature extraction module and a vibration feature extraction module respectively to extract strain features and vibration features;

[0012] Input the strain features and vibration features into a fusion feature extraction module for fusion to obtain a reconstructed deflection;

[0013] Among them, the strain feature extraction module and the vibration feature extraction module have the same structure, and both include a short-term feature extraction module, a medium-term feature extraction module, and a long-term feature extraction module. Based on three different time scales of short-term, medium-term, and long-term, three parallel convolutional neural networks are used for feature extraction.

[0014] As a further technical solution, the process of preprocessing the health monitoring data of the bridge is as follows:

[0015] Perform multi-scale decomposition on the vibration data using wavelet transform;

[0016] Perform smoothing and denoising on the strain data through local polynomial fitting;

[0017] Normalize the decomposed vibration data and the smoothed and denoised strain data to obtain preprocessed vibration data and strain data.

[0018] As a further technical solution, the parallel convolutional neural networks all include a convolutional layer and a max pooling layer;

[0019] In the first convolutional neural network, multi-scale features of strain data are extracted through the convolutional layer, and max pooling is performed on the strain data through the max pooling layer to obtain the max pooling output features;

[0020] The multi-scale features of the strain data and the max pooling output features are added together to obtain the first short-term features of the strain data.

[0021] As a further technical solution, the preprocessed strain data is input into the strain feature extraction module to extract strain features. The specific process is as follows:

[0022] The preprocessed strain data is input into the first short-term feature extraction module, and based on the short time scale, the first convolutional neural network is used to extract the short-term features of the strain data;

[0023] The preprocessed strain data is input into the first medium-term feature extraction module, and based on the medium time scale, the second convolutional neural network is used to extract the medium-term features of the strain data;

[0024] The preprocessed strain data is input into the first long-term feature extraction module, and based on the long time scale, the third convolutional neural network is used to extract the long-term features of the strain data;

[0025] The first short-term feature, the first medium-term feature, and the first long-term feature are concatenated to obtain the strain features.

[0026] As a further technical solution, the preprocessed vibration data is input into the vibration feature extraction module to extract vibration features. The specific process is as follows:

[0027] The preprocessed vibration data is input into the second short-term feature extraction module, and based on the short time scale, the fourth convolutional neural network is used to extract the short-term features of the vibration data;

[0028] The preprocessed vibration data is input into the second medium-term feature extraction module, and based on the medium time scale, the fifth convolutional neural network is used to extract the medium-term features of the vibration data;

[0029] The preprocessed vibration data is input into the second long-term feature extraction module, and based on the long time scale, the sixth convolutional neural network is used to extract the long-term features of the vibration data;

[0030] The second short-term feature, the second medium-term feature, and the second long-term feature are concatenated to obtain the vibration features.

[0031] As a further technical solution, the strain features and the vibration features are input into the fusion feature extraction module for fusion to obtain the reconstructed deflection. The fusion feature extraction module includes a convolutional layer and a fully connected layer. The specific process is as follows:

[0032] Process and splice the strain feature and the vibration feature to obtain the spliced feature;

[0033] Input the spliced feature into the convolutional layer and perform a sliding window operation to obtain the spliced feature map;

[0034] Input the spliced feature map into the fully connected layer for processing to obtain the reconstructed deflection.

[0035] As a further technical solution, process and splice the strain feature and the vibration feature to obtain the spliced feature. Specifically, align and perform weighted sum on the strain feature and the vibration feature to obtain the spliced feature.

[0036] The second aspect of the present invention discloses a bridge deflection reconstruction system based on time multi-scale feature fusion, including:

[0037] A data acquisition module for acquiring the health monitoring data of the bridge, where the health monitoring data of the bridge includes at least strain data and vibration data;

[0038] A data preprocessing module for preprocessing the health monitoring data of the bridge to obtain the preprocessed strain data and vibration data;

[0039] A feature extraction module for respectively inputting the preprocessed strain data and vibration data into a strain feature extraction module and a vibration feature extraction module to respectively extract strain features and vibration features. Among them, the strain feature extraction module and the vibration feature extraction module have the same structure, and both include a short-term feature extraction module, a medium-term feature extraction module, and a long-term feature extraction module. Based on three different time scales of short-term, medium-term, and long-term, three parallel convolutional neural networks are used for feature extraction;

[0040] A feature fusion module for inputting the strain feature and the vibration feature into a fusion feature extraction module for fusion to obtain the reconstructed deflection.

[0041] The third aspect of the present invention discloses a computer device, including a memory, a processor, and a computer program stored on the memory and executable on the processor. When the processor executes the program, the steps of the method described in the first aspect are implemented.

[0042] The fourth aspect of the present invention aims to provide a computer-readable storage medium, on which a program is stored. When the program is executed by a processor, the steps in the method described in the first aspect of the present invention are implemented.

[0043] The above one or more technical solutions have the following beneficial effects:

[0044] In this embodiment, strain, vibration and environmental data are integrated to reveal the correlation between environmental factors and bridge displacement, determine the optimal location of the temperature sensor inside the chamber, reveal the correlation between the stress and deformation of the bridge under load, and determine the optimal location of the strain sensor for deflection reconstruction, thus overcoming the influence of the external environment, making the collected data more accurate, and being able to comprehensively monitor the structural response and environmental conditions of the bridge.

[0045] In this embodiment, a multi-scale feature fusion model based on CNN uses a convolutional neural network to extract the strain characteristics of strain data and the vibration characteristics of vibration data at three different time scales: short-term, medium-term and long-term, and fuse them to obtain the reconstructed deflection. This solves the problem that the prior art can only rely on long-term stable reference points, avoids the limitations of traditional monitoring methods, and can accurately reconstruct the bridge deflection without increasing additional hardware costs, significantly reducing the monitoring cost while improving the coverage and real-time performance of monitoring.

[0046] The bridge deflection reconstruction method based on data-driven and time multi-scale in this embodiment can make full use of the multi-scale characteristics of strain and vibration data to achieve more accurate deflection reconstruction. It has its advancedness and practical value, and will provide more scientific data support for bridge health assessment, help extend the service life of bridges, reduce the risk of accidents, and ultimately ensure public transportation safety and reduce losses.

[0047] Advantages of additional aspects of the present invention will be given in part in the following description, and in part will become obvious from the following description, or will be learned through practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] The accompanying drawings in the specification, which constitute a part of the present invention, are used to provide a further understanding of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute improper limitations on the present invention.

[0049] Figure 1 This is a schematic diagram of the structure of a bridge deflection reconstruction model based on time multi-scale feature fusion in the first embodiment of the present invention;

[0050] Figure 2 The comparison between the measured and reconstructed deflections of the continuous beam bridge of the first embodiment;

[0051] Figure 3 This is the measured and reconstructed deflection of the cable-stayed bridge of the first embodiment. DETAILED DESCRIPTION

[0052] It should be noted that the following detailed description is exemplary and is intended to provide further illustration of the present invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by those of ordinary skill in the technical field to which the present invention belongs.

[0053] It should be noted that the terms used herein are only for describing specific embodiments and are not intended to limit the exemplary embodiments according to the present invention.

[0054] Without conflict, the embodiments in the present invention and the features in the embodiments can be combined with each other.

[0055] Embodiment 1

[0056] This embodiment discloses a method for reconstructing bridge deflection based on time multi-scale feature fusion.

[0057] To more clearly illustrate this embodiment, a process for reconstructing bridge deflection based on time multi-scale feature fusion can be specifically described as follows:

[0058] This embodiment provides a method for reconstructing bridge deflection based on time multi-scale feature fusion, including:

[0059] S1. Obtain the health monitoring data of the bridge, where the health monitoring data of the bridge includes at least strain data and vibration data;

[0060] S2. Preprocess the health monitoring data of the bridge to obtain the preprocessed strain data and vibration data;

[0061] S3. Input the preprocessed strain data and vibration data into a strain feature extraction module and a vibration feature extraction module respectively to extract strain features and vibration features;

[0062] Among them, the strain feature extraction module and the vibration feature extraction module have the same structure, and both include a short-term feature extraction module, a medium-term feature extraction module, and a long-term feature extraction module. Based on three different time scales of short-term, medium-term, and long-term, three parallel convolutional neural networks are used for feature extraction;

[0063] S4. Input the strain features and vibration features into a fusion feature extraction module for fusion to obtain the reconstructed deflection.

[0064] As Figure 1 shown, in step S1, obtain the health monitoring data of the bridge, where the health monitoring data of the bridge includes at least strain data and vibration data.

[0065] In this embodiment, the health monitoring data of the bridge is obtained by deploying devices such as fiber Bragg grating sensors, optoelectronic deflection meters, and vibration sensors.

[0066] S1-1. Fiber grating sensors, photoelectric deflectometers, vibration sensors and other equipment were installed.

[0067] For two typical bridges, continuous beam bridges and cable-stayed bridges, fiber Bragg grating sensors, photoelectric deflectometers, vibration sensors and other equipment are deployed. In order to ensure the accuracy and reliability of bridge structure health monitoring, this embodiment uses advanced sensors for data collection. These sensors have the characteristics of high precision, strong stability, and good anti-interference ability, and can comprehensively monitor the structural response and environmental conditions of the bridge.

[0068] S1-2. Obtain health monitoring data of the bridge, including at least strain data and vibration data.

[0069] In this embodiment, the fiber Bragg grating sensor is used to collect structural strain, structural temperature, ambient temperature and humidity, etc., the photoelectric deflectometer is used to collect general deflection, and the vibration sensor is used to collect structural vibration, etc. The specific sensor models corresponding to the monitoring content are shown in Table 1.

[0070] Table 1 Sensor Models

[0071]

[0072]

[0073] In this embodiment, the acquired health monitoring data of the bridge includes strain data, vibration data, ambient temperature and humidity, wind speed data, structural temperature, displacement and deflection data.

[0074] At the same time, the time of data collection for each type is recorded to provide a data basis for the subsequent use of different time scales to extract features.

[0075] like Figure 1 As shown, in step S2, the health monitoring data of the bridge is preprocessed to obtain preprocessed strain data and vibration data.

[0076] S2-1. Preprocess the health monitoring data of the bridge.

[0077] In this embodiment, since the mid-span area is the core location for sensor deployment due to significant deformation under stress, data preprocessing is divided into two parts: low frequency and high frequency:

[0078] (1) For low-frequency data such as ambient temperature, humidity, and displacement, the box plot method and Z-Score were used to detect outliers, missing values were filled by linear interpolation, and invalid wind speed data were eliminated to ensure data quality.

[0079] (2) For the three high-frequency data of strain data, vibration data, and deflection data, wavelet transform and Savgol filtering are used to reduce noise, and the data is cleaned and normalized.

[0080] Among them, wavelet multi-scale decomposition is adopted for deflection and vibration data; local polynomial fitting is used to smooth the noise of strain data; the cleaned data is normalized to provide a reliable basis for subsequent analysis.

[0081] S2-2. Study the characteristics of bridge deflection data. Through Pearson correlation coefficient analysis, reveal the correlation between environmental factors and bridge displacement, and determine the optimal positions of sensors.

[0082] (1) Statistical analysis of monitoring data characteristics.

[0083] The Pearson Correlation Coefficient (PCC) is an important statistical tool for measuring the linear correlation between two continuous variables. The calculation formula of the Pearson correlation coefficient is:

[0084]

[0085] where, x i and y i are the sample values of the two variables respectively, and are the means of the samples, and r is the final correlation coefficient value.

[0086] The covariance of the two variables is calculated through formula (1) to evaluate the linear dependence between the variables. The closer the absolute value of the correlation coefficient is to 1, the stronger the correlation between the two variables. When |r|>0.7, it is considered that there is a strong correlation between the two variables; when 0.3≤|r|≤0.7, it indicates a medium correlation; and when |r|<0.3, the correlation is weak or there is no obvious correlation.

[0087] The Pearson correlation coefficient is widely used in fields such as bridge health monitoring and engineering. For example, in bridge monitoring, the relationship between environmental factors and bridge structural response can be analyzed through the Pearson correlation coefficient to evaluate the impact of environmental changes on the bridge structure.

[0088] (2) Reveal the correlation between environmental factors and bridge displacement.

[0089] 1) According to formula (1), it can be known that the temperature inside the box girder is strongly negatively correlated with the longitudinal displacement (for continuous girder bridges, R≈-0.97, for cable-stayed bridges, R≈-0.93), indicating that temperature changes dominate the longitudinal expansion and contraction deformation; the lateral displacement is weakly affected by temperature (R≈-0.8), and the effects of environmental factors such as wind load are limited.

[0090] 2) The linear correlation between strain and deflection is relatively low (R < 0.32 for continuous beam bridges), but the specific strain measurement points in the side spans of cable-stayed bridges are significantly correlated with deflection (R ≈ 0.81), providing a basis for the feature selection of the reconstruction model.

[0091] (3) Determine the optimal strain sensor positions for deflection reconstruction based on the correlation between environmental factors and bridge displacement.

[0092] By analyzing the changing trends of bridge displacement under different environmental factors, it provides a basis for the health assessment of bridges under different climate conditions and determines the positions of the optimal internal temperature sensors in the chambers for displacement reconstruction. In the analysis of high-frequency monitoring data, the relationship between bridge strain and deflection is mainly analyzed. Through the correlation analysis of strain and deflection data, the correlation between the stress and deformation of the bridge under load is revealed, and the optimal strain sensor positions for deflection reconstruction are determined.

[0093] As Figure 1 shown, in step S3, the preprocessed strain data and vibration data are respectively input into the strain feature extraction module and the vibration feature extraction module to extract strain features and vibration features respectively.

[0094] S3-1. Construct and train a neural network model based on time multi-scale feature fusion.

[0095] In this embodiment, a neural network model based on time multi-scale feature fusion is used for bridge deflection reconstruction. The neural network model based on time multi-scale feature fusion includes a strain feature extraction module, a vibration feature extraction module, and a fusion feature extraction module.

[0096] Among them, the strain feature extraction module and the vibration feature extraction module have the same structure, both including a short-term feature extraction module, a medium-term feature extraction module, and a long-term feature extraction module. Based on three different time scales of short-term, medium-term, and long-term, three parallel convolutional neural networks are used for feature extraction.

[0097] The strain feature extraction module includes a first short-term feature extraction module, a first medium-term feature extraction module, and a first long-term feature extraction module. The three parallel convolutional neural networks are: the first convolutional neural network, the second convolutional neural network, and the third convolutional neural network.

[0098] The vibration feature extraction module includes a second short-term feature extraction module, a second medium-term feature extraction module, and a second long-term feature extraction module. The three parallel convolutional neural networks are: the fourth convolutional neural network, the fifth convolutional neural network, and the sixth convolutional neural network.

[0099] In this embodiment, the parallel convolutional neural networks all include a convolutional layer and a max pooling layer.

[0100] The basic components of a convolutional neural network include several key layers: the convolutional layer, activation function, pooling layer, fully connected layer, and output layer. The convolutional layer is the core of the CNN, which extracts features from the input data through convolutional operations. The convolutional layer uses multiple learnable convolutional kernels (filters) to perform sliding window operations on the input data to generate feature maps. Each convolutional kernel is responsible for capturing different features, such as edges and textures.

[0101] After the convolutional layer, an activation function, such as ReLU (Rectified Linear Unit), is usually applied to introduce non-linearity, enabling the network to learn complex patterns.

[0102] The specific parameters of all convolutional layers and fully connected layers are shown in Table 2, where Conv represents the convolutional layer.

[0103] Table 2 Model Parameter Table

[0104]

[0105] The time multi-scale feature extraction strategy refers to extracting features through different time scales when processing time series data, so as to comprehensively capture the dynamic changes in the data. In this strategy, the feature extraction module is designed with three different time scales: short-term, medium-term, and long-term. Short-term features focus on instantaneous changes, while long-term features capture more stable trends. This multi-scale feature extraction can not only enhance the model's understanding ability of time series data but also effectively improve the prediction accuracy. By fusing information from different time scales, the model can better adapt to complex signal features, and then in specific applications, such as tasks like bridge deflection reconstruction, provide more accurate and reliable results. This strategy of combining CNN with multi-scale feature extraction can make full use of the time series characteristics of the signal and provide strong support for the monitoring and analysis of complex systems.

[0106] Among them, in this embodiment, for the setting of the three time scales of short-term, medium-term, and long-term, the short-term is set in days, the medium-term is set in months, and the long-term is set in years. It can be set according to the actual situation. For example, if set in days, the short-term is 1 day, the medium-term is 1 month, and the long-term is 1 day.

[0107] In this embodiment, the fused feature extraction module realizes the comprehensive utilization of information by uniformly processing and splicing the extracted strain and vibration features.

[0108] First, align the feature dimensions extracted by the two modules to ensure the consistency of the input features.

[0109] Subsequently, in combination with the weighted sum or attention mechanism, the contributions of each feature are evaluated to highlight the features that are more important for deflection reconstruction. In this process, the model can not only capture the subtle changes of the bridge structure under different working conditions but also effectively eliminate the interference of noise on feature extraction, improving the accuracy of deflection reconstruction.

[0110] The method of multi-scale feature fusion enables the model to comprehensively consider various features in the reconstruction of bridge deflections, ultimately achieving more accurate deflection reconstruction, providing a scientific basis for structural health monitoring, and thus providing strong support for the safety assessment and maintenance decision-making of bridges.

[0111] In this embodiment, a neural network model based on time multi-scale feature fusion is constructed through a strain feature extraction module, a vibration feature extraction module, and a fusion feature extraction module. Using an existing bridge dataset and an existing training method, the neural network model based on time multi-scale feature fusion is trained to obtain a trained neural network model based on time multi-scale feature fusion.

[0112] Then, the neural network model based on time multi-scale feature fusion can be used for bridge deflection reconstruction.

[0113] S3-2. Input the preprocessed strain data into the strain feature extraction module to extract strain features.

[0114] In this embodiment, the process of inputting the preprocessed strain data into the strain feature extraction module to extract strain features is as follows:

[0115] (1) Input the preprocessed strain data into the first short-term feature extraction module, and based on the short-term time scale, use the first convolutional neural network to extract the short-term features of the strain data.

[0116] In the first convolutional neural network, multi-scale features of the strain data are extracted through the convolutional layer, and max-pooling is performed on the strain data through the max-pooling layer to obtain the max-pooling output features.

[0117] In this embodiment, the convolutional layer uses a 1*3 convolutional kernel to extract multi-scale features of the strain data to capture the dynamic changes of the strain signal, and in combination with batch normalization and the ReLU activation function, the nonlinearity of the neural network model is enhanced.

[0118] The strain data is downsampled through the max-pooling layer to obtain the max-pooling output features.

[0119] The multi-scale features of the strain data and the max-pooling output features are added together to obtain the first short-term features of the strain data.

[0120] In this embodiment, a residual connection is adopted for the addition operation, adding the multi-scale features of the strain data and the maximum pooling output features to obtain the first short-term feature of the strain data.

[0121] (2) Input the preprocessed strain data into the first medium-term feature extraction module, and based on the medium time scale, use a second convolutional neural network to extract the medium-term features of the strain data.

[0122] In the second convolutional neural network, the multi-scale features of the strain data are extracted through the convolutional layer, and the maximum pooling is performed on the strain data through the maximum pooling layer to obtain the maximum pooling output features;

[0123] In this embodiment, the convolutional layer uses a 1*5 convolutional kernel to extract the multi-scale features of the strain data to capture the dynamic changes of the strain signal, and combines batch normalization and the ReLU activation function to enhance the nonlinearity of the neural network model.

[0124] The maximum pooling layer performs downsampling on the strain data to obtain the maximum pooling output features.

[0125] The multi-scale features of the strain data and the maximum pooling output features are added to obtain the first medium-term feature of the strain data.

[0126] In this embodiment, a residual connection is adopted for the addition operation, adding the multi-scale features of the strain data and the maximum pooling output features to obtain the first medium-term feature of the strain data.

[0127] (3) Input the preprocessed strain data into the first long-term feature extraction module, and based on the long time scale, use a third convolutional neural network to extract the long-term features of the strain data.

[0128] In the third convolutional neural network, the multi-scale features of the strain data are extracted through the convolutional layer, and the maximum pooling is performed on the strain data through the maximum pooling layer to obtain the maximum pooling output features.

[0129] In this embodiment, the convolutional layer uses a 1*7 convolutional kernel to extract the multi-scale features of the strain data to capture the dynamic changes of the strain signal, and combines batch normalization and the ReLU activation function to enhance the nonlinearity of the neural network model.

[0130] The maximum pooling layer performs downsampling on the strain data to obtain the maximum pooling output features.

[0131] The multi-scale features of the strain data and the maximum pooling output features are added to obtain the first long-term feature of the strain data.

[0132] In this embodiment, a residual connection is adopted for the addition operation, and the multi-scale features of the strain data and the max-pooling output features are added together to obtain the first long-term feature of the strain data.

[0133] (4) Concatenate the first short-term feature, the first mid-term feature, and the first long-term feature to obtain the strain feature.

[0134] Align the first short-term feature, the first mid-term feature, and the first long-term feature to ensure the consistency of the input features. Subsequently, combined with a weighted sum or an attention mechanism, evaluate the contributions of each feature, so as to highlight the features that are more important for deflection reconstruction, that is, obtain the strain feature.

[0135] After the above steps, that is, during the strain feature extraction process, different convolutional kernels are used, which can capture the dynamic changes of the strain signal at different time scales. At the same time, combined with batch normalization and the ReLU activation function, the nonlinear expression ability of the network is enhanced. In addition, through downsampling by the max-pooling layer, not only the dimension of the feature map is reduced, but also the most significant feature information is effectively retained. At the same time, residual connections are adopted to enhance the information flow, ensuring that the information in the deep network will not be lost due to the increase in the number of layers.

[0136] S3-3. Input the preprocessed vibration data into the vibration feature extraction module to extract vibration features.

[0137] In this embodiment, input the preprocessed vibration data into the vibration feature extraction module to extract vibration features. The specific process is as follows:

[0138] (1) Input the preprocessed vibration data into the second short-term feature extraction module, and based on the short-term time scale, use the fourth convolutional neural network to extract the short-term features of the vibration data.

[0139] In the fourth convolutional neural network, extract the multi-scale features of the vibration data through the convolutional layer, and perform max-pooling on the vibration data through the max-pooling layer to obtain the max-pooling output features.

[0140] In this embodiment, the convolutional layer uses a 1*3 convolutional kernel to extract the multi-scale features of the vibration data to capture the dynamic changes of the vibration signal, and combines batch normalization and the ReLU activation function to enhance the nonlinearity of the neural network model.

[0141] Perform downsampling on the vibration data through the max-pooling layer to obtain the max-pooling output features.

[0142] Perform an addition process on the multi-scale features of the vibration data and the max-pooling output features to obtain the second short-term feature of the vibration data.

[0143] In this embodiment, a residual connection is adopted for the addition operation, adding the multi-scale features of the vibration data and the maximum pooling output features to obtain the second short-term feature of the vibration data.

[0144] (2) Input the preprocessed vibration data into the second medium-term feature extraction module, and based on the medium time scale, use the fifth convolutional neural network to extract the medium-term features of the vibration data.

[0145] In the fifth convolutional neural network, the multi-scale features of the vibration data are extracted through the convolutional layer, and the vibration data is max-pooled through the max-pooling layer to obtain the max-pooling output features;

[0146] In this embodiment, the convolutional layer uses a 1*5 convolutional kernel to extract the multi-scale features of the vibration data to capture the dynamic changes of the vibration signal, and combines batch normalization and the ReLU activation function to enhance the nonlinearity of the neural network model.

[0147] The vibration data is downsampled through the max-pooling layer to obtain the max-pooling output features.

[0148] The multi-scale features of the vibration data and the max-pooling output features are added to obtain the second medium-term feature of the vibration data.

[0149] In this embodiment, a residual connection is adopted for the addition operation, adding the multi-scale features of the vibration data and the max-pooling output features to obtain the second medium-term feature of the vibration data.

[0150] (3) Input the preprocessed vibration data into the second long-term feature extraction module, and based on the long time scale, use the sixth convolutional neural network to extract the long-term features of the vibration data.

[0151] In the sixth convolutional neural network, the multi-scale features of the strain data are extracted through the convolutional layer, and the strain data is max-pooled through the max-pooling layer to obtain the max-pooling output features;

[0152] In this embodiment, the convolutional layer uses a 1*7 convolutional kernel to extract the multi-scale features of the vibration data to capture the dynamic changes of the vibration signal, and combines batch normalization and the ReLU activation function to enhance the nonlinearity of the neural network model.

[0153] The vibration data is downsampled through the max-pooling layer to obtain the max-pooling output features.

[0154] The multi-scale features of the vibration data and the max-pooling output features are added to obtain the second long-term feature of the strain data.

[0155] In this embodiment, a residual connection is adopted for the addition operation, and the multi-scale features of the vibration data and the maximum pooling output features are added together to obtain the second long-term features of the vibration data.

[0156] (4) Concatenate the second short-term features, the second medium-term features, and the second long-term features to obtain the vibration features.

[0157] Align the second short-term features, the second medium-term features, and the second long-term features to ensure the consistency of the input features. Subsequently, combined with the weighted sum or attention mechanism, evaluate the contributions of each feature, so as to highlight the features that are more important for deflection reconstruction. That is, the vibration features are obtained.

[0158] After the above steps, with the same structural settings, focusing on capturing the dynamic features and time-domain information of the vibration signal, this design ensures the full learning of two different signal features, enabling the model to comprehensively understand the dynamic response of the bridge.

[0159] As Figure 1 shown, in step S4, input the strain features and the vibration features into the fusion feature extraction module for fusion to obtain the reconstructed deflection.

[0160] In this embodiment, the strain features and the vibration features are processed and concatenated to obtain the concatenated features. Specifically, the strain features and the vibration features are aligned and weighted-summed to obtain the concatenated features.

[0161] Input the concatenated features into the convolutional layer for a sliding window operation to obtain the concatenated feature map.

[0162] In this embodiment, input the concatenated features into the convolutional layer with a 1*8 convolutional kernel to generate the concatenated feature map.

[0163] Input the concatenated feature map into the fully connected layer for processing to obtain the reconstructed deflection.

[0164] After the above steps, it can not only capture the subtle changes of the bridge structure under different working conditions, but also effectively eliminate the interference of noise on feature extraction, and improve the accuracy of deflection reconstruction. The method of multi-scale feature fusion enables the model to comprehensively consider various features in the reconstruction of the bridge deflection, and finally realizes more accurate deflection reconstruction, providing a scientific basis for structural health monitoring, and thus providing strong support for the safety assessment and maintenance decision-making of the bridge.

[0165] This embodiment also evaluates the neural network model based on time multi-scale feature fusion.

[0166] Multiple evaluation indexes are selected, including the root mean square error (RMSE), the coefficient of determination (R 2) and the mean absolute error (MAE). The root mean square error (RMSE) is the square root of the MSE and provides an error measure in the same unit as the actual data. The coefficient of determination (R2) reflects the ability of the model to explain the variation in the data, with values ranging from 0 to 1, and the closer to 1, the more effective the model. The mean absolute error (MAE) provides the average of the absolute differences between the predicted and actual values, which helps to intuitively understand the prediction accuracy of the model. The calculation formulas for RMSE, MAE, and R2 are as follows:

[0167]

[0168] where m is the total number of samples, y i is the i-th true value, is the i-th predicted value.

[0169] The proposed bridge deflection reconstruction method evaluates R 2 , the root mean square error (RMSE), and the mean absolute error (MAE) to quantify the reconstruction accuracy by studying continuous beam bridges (with data acquisition frequencies of 33 Hz for deflection, 20 Hz for strain, and 200 Hz for vibration) and cable-stayed bridges (with data acquisition frequencies of 33 Hz for deflection, 50 Hz for strain, and 200 Hz for vibration). The R2 values for the continuous beam bridge and the cable-stayed bridge are 0.51 and 0.55 respectively, which shows that the model has a certain explanatory ability. The RMSE values are 0.35 and 4.21 respectively, indicating that the average deviation between the reconstructed values and the actual values is small, while the MAE values are 0.24 and 2.92 respectively, further confirming the accuracy of the reconstruction results. These indicators show that the method in this paper meets the requirements of bridge deflection reconstruction to a certain extent and can provide strong data support for the structural health monitoring of bridges. The comparison of the measurement and reconstruction results of the two bridges in the test set is shown respectively as Figure 2 and Figure 3 shown.

[0170] In order to demonstrate the advanced nature of this patented method, as shown in Tables 3 and 4, this paper also compares with a variety of mainstream algorithms, including long short-term memory network (LSTM), gated recurrent unit (GRU), Transformer, ordinary convolutional neural network (CNN), multilayer perceptron (MLP) and support vector regression (SVR). These algorithms have their own advantages, but they have certain limitations when processing time series data and nonlinear relationships. LSTM and GRU are suitable for capturing long-term dependencies in time series, but they may not achieve the best results when processing complex feature fusion. Although Transformer performs well in modeling sequence data, its computational complexity is high, especially when the data volume is large. Ordinary CNN has advantages in spatial feature extraction, but its processing of time series is relatively simple. As a basic deep learning model, MLP is easy to implement, but its performance is limited when dealing with complex bridge deflection changes. Although SVR can effectively deal with nonlinear problems, its performance is limited under high-dimensional data.

[0171] Table 3 Comparison of deflection reconstruction accuracy of different models in continuous beam bridges

[0172]

[0173] Table 4 Comparison of deflection reconstruction of different models in cable-stayed bridges

[0174]

[0175] By comparing with the above algorithms, the results show that the proposed method has obvious advantages in the accuracy of deflection reconstruction. This is not only reflected in the evaluation indicators, but also in the ability to make full use of the multi-scale characteristics of strain and vibration data to achieve more accurate deflection reconstruction. This comparison emphasizes the application potential of the proposed method in bridge deflection reconstruction and further verifies its advanced nature and practical value.

[0176] In this embodiment, by deploying fiber grating sensors, photoelectric deflectometers, vibration sensors and other equipment, the health monitoring data of the bridge is obtained and preprocessed; the characteristics of the bridge deflection data are studied, and the correlation between environmental factors and bridge displacement is revealed through Pearson correlation coefficient analysis, overcoming the problem that the existing technology is easily affected by the environment; the changes of the bridge structure under different working conditions are captured through multi-scale feature fusion technology. The bridge deflection is reconstructed based on the neural network model of time multi-scale feature fusion.

[0177] Embodiment 2

[0178] This embodiment discloses a bridge deflection reconstruction system based on time multi-scale feature fusion, including:

[0179] A data acquisition module for acquiring the health monitoring data of a bridge, where the health monitoring data of the bridge includes at least strain data and vibration data;

[0180] A data preprocessing module for preprocessing the health monitoring data of the bridge to obtain preprocessed strain data and vibration data;

[0181] A feature extraction module for respectively inputting the preprocessed strain data and vibration data into a strain feature extraction module and a vibration feature extraction module to extract strain features and vibration features respectively; among them, the strain feature extraction module and the vibration feature extraction module have the same structure, and both include a short-term feature extraction module, a medium-term feature extraction module, and a long-term feature extraction module, and feature extraction is performed using three parallel convolutional neural networks based on three different time scales of short-term, medium-term, and long-term;

[0182] A feature fusion module for inputting the strain features and vibration features into a fusion feature extraction module for fusion to obtain a reconstructed deflection.

[0183] Based on providing a bridge deflection reconstruction system based on time multi-scale feature fusion to implement the method steps in Embodiment 1.

[0184] Embodiment 3

[0185] The purpose of this embodiment is to provide a computer device, including a memory, a processor, and a computer program stored on the memory and executable on the processor, and when the processor executes the program, the steps of the above method are implemented.

[0186] Embodiment 4

[0187] The purpose of this embodiment is to provide a computer-readable storage medium, on which a program is stored, and when the program is executed by a processor, the steps in the method as described above are implemented.

[0188] Those skilled in the art should understand that the above-mentioned modules or steps of the present invention can be implemented by a general computer device. Optionally, they can be implemented by program codes executable by a computing device, so that they can be stored in a storage device and executed by the computing device, or they can be respectively made into individual integrated circuit modules, or multiple modules or steps among them can be made into a single integrated circuit module to implement. The present invention is not limited to any specific combination of hardware and software.

[0189] Although the specific implementation manners of the present invention are described above in conjunction with the accompanying drawings, it is not a limitation to the protection scope of the present invention. Those skilled in the art should understand that based on the technical solutions of the present invention, various modifications or deformations that can be made by those skilled in the art without creative labor are still within the protection scope of the present invention.

Claims

1. A bridge deflection reconstruction method based on time multi-scale feature fusion, characterized in that Including: Obtain the health monitoring data of the bridge, where the health monitoring data of the bridge includes at least strain data and vibration data; Preprocess the health monitoring data of the bridge to obtain the preprocessed strain data and vibration data; Input the preprocessed strain data and vibration data into a strain feature extraction module and a vibration feature extraction module respectively to extract strain features and vibration features; Input the strain features and vibration features into a fusion feature extraction module for fusion to obtain a reconstructed deflection; Among them, the strain feature extraction module and the vibration feature extraction module have the same structure, both including a short-term feature extraction module, a medium-term feature extraction module, and a long-term feature extraction module. Based on three different time scales of short-term, medium-term, and long-term, three parallel convolutional neural networks are used for feature extraction.

2. The bridge deflection reconstruction method based on time multi-scale feature fusion according to claim 1, wherein, The specific process of preprocessing the health monitoring data of the bridge is as follows: Perform multi-scale decomposition on the vibration data using wavelet transform; Perform smoothing denoising on the strain data through local polynomial fitting; Normalize the decomposed vibration data and the smoothed and denoised strain data to obtain the preprocessed vibration data and strain data.

3. The bridge deflection reconstruction method based on time multi-scale feature fusion according to claim 1, characterized in that The parallel convolutional neural networks all include a convolutional layer and a max pooling layer; In the first convolutional neural network, extract the multi-scale features of the strain data through the convolutional layer, and perform max pooling on the strain data through the max pooling layer to obtain the max pooling output features; Add the multi-scale features of the strain data and the max pooling output features to obtain the short-term features of the strain data.

4. A method for reconstructing bridge deflection based on time multi-scale feature fusion according to claim 1, characterized in that Input the preprocessed strain data into the strain feature extraction module to extract strain features. The specific process is as follows: Input the preprocessed strain data into the first short-term feature extraction module, and based on the short-term time scale, use the first convolutional neural network to extract the short-term features of the strain data; Input the preprocessed strain data into the first medium-term feature extraction module, and based on the medium-term time scale, use the second convolutional neural network to extract the medium-term features of the strain data; Input the preprocessed strain data into the first long-term feature extraction module, and based on the long-term time scale, use the third convolutional neural network to extract the long-term features of the strain data; Concatenate the first short-term feature, the first medium-term feature, and the first long-term feature to obtain the strain features.

5. The bridge deflection reconstruction method based on time multi-scale feature fusion according to claim 1, wherein, Input the preprocessed vibration data into the vibration feature extraction module to extract vibration features. The specific process is as follows: Input the preprocessed vibration data into the second short-term feature extraction module, and based on the short-term time scale, use the fourth convolutional neural network to extract the short-term features of the vibration data; Input the preprocessed vibration data into the second medium-term feature extraction module, and based on the medium-term time scale, use the fifth convolutional neural network to extract the medium-term features of the vibration data; Input the preprocessed vibration data into the second long-term feature extraction module, and based on the long-term time scale, use the sixth convolutional neural network to extract the long-term features of the vibration data; Concatenate the second short-term feature, the second medium-term feature, and the second long-term feature to obtain the vibration features.

6. The bridge deflection reconstruction method based on time multi-scale feature fusion according to claim 1, characterized in that The strain feature and the vibration feature are input into a fusion feature extraction module for fusion to obtain a reconstructed deflection. The fusion feature extraction module includes a convolutional layer and a fully connected layer. The specific process is as follows: The strain feature and the vibration feature are processed and concatenated to obtain a concatenated feature; The concatenated feature is input into the convolutional layer for a sliding window operation to obtain a concatenated feature map; The concatenated feature map is input into the fully connected layer for processing to obtain a reconstructed deflection.

7. The bridge deflection reconstruction method based on time multi-scale feature fusion according to claim 6, characterized in that, The strain feature and the vibration feature are processed and concatenated to obtain a concatenated feature. Specifically, the strain feature and the vibration feature are aligned and weighted-summed to obtain a concatenated feature.

8. A bridge deflection reconstruction system based on time multi-scale feature fusion, characterized in that, It includes: A data acquisition module for acquiring the health monitoring data of the bridge. The health monitoring data of the bridge includes at least strain data and vibration data; A data preprocessing module for preprocessing the health monitoring data of the bridge to obtain preprocessed strain data and vibration data; A feature extraction module for respectively inputting the preprocessed strain data and vibration data into a strain feature extraction module and a vibration feature extraction module to respectively extract a strain feature and a vibration feature. Among them, the strain feature extraction module and the vibration feature extraction module have the same structure, and both include a short-term feature extraction module, a medium-term feature extraction module, and a long-term feature extraction module. Based on three different time scales of short-term, medium-term, and long-term, three parallel convolutional neural networks are used for feature extraction; A feature fusion module for inputting the strain feature and the vibration feature into a fusion feature extraction module for fusion to obtain a reconstructed deflection.

9. A computer device, comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, characterized in that, When the processor executes the program, the steps of the method according to any one of claims 1-7 are implemented.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, the steps of the method according to any one of claims 1-7 are executed.

Images