A method and system for evaluating the stability of steep slope bridges based on deep learning

Abstract

This invention discloses a method and system for evaluating the stability of steep-slope bridges based on deep learning, comprising: S1: acquiring bridge region DEM data, bridge displacement data, and bridge stress data; S2: constructing a bridge stability evaluation network and extracting bridge region DEM features based on a convolutional neural network; S3: extracting internal structural features of the bridge based on a recurrent neural network; S4: fusing bridge region DEM features and internal structural features of the bridge based on an attention mechanism; S5: setting the network's input, output, and optimization objective, and optimizing the parameters of the bridge stability evaluation network based on an adaptive learning rate optimization algorithm. This solution addresses the problem of insufficient accuracy in bridge stability evaluation.

Description

Technical Field

[0001] This invention relates to the field of bridge monitoring technology, and in particular to a method and system for evaluating the stability of steep slope bridges based on deep learning. Background Technology

[0002] Bridges are vital transportation arteries, and the safety of bridge structures involves significant property and even life safety. As bridges age and traffic volume increases, localized damage to bridge structures is inevitable; therefore, monitoring the health of bridges is of paramount importance.

[0003] However, traditional deep learning-based bridge stability monitoring technology evaluates bridge stability by analyzing internal structural parameters, without taking into account the complex terrain of steep mountain slopes, making it difficult to achieve good results in steep mountain slope scenarios.

[0004] In mountainous steep slope scenarios, the terrain is characterized by large elevation differences, frequent changes, and steep slopes. Therefore, bridges on mountainous steep slopes often feature high adjacent piers and frequent interactions between the bridge and the road. Compared to traditional bridge monitoring scenarios, bridge monitoring in mountainous steep slope scenarios requires careful consideration of the interaction between the bridge's various structural details and the terrain. Summary of the Invention

[0005] In view of this, the present invention provides a method and system for evaluating the stability of bridges on steep slopes based on deep learning, aiming to solve the problem that the accuracy of bridge stability monitoring is greatly affected by terrain in mountainous steep slope scenarios. Traditional data fusion methods can only fuse various sensor data, but cannot fuse multimodal heterogeneous data such as sensor data and DEM data. The present invention performs feature representation on the DEM data of the bridge area and performs multimodal fusion with the internal structural data of the bridge, thereby improving the accuracy of bridge stability identification.

[0006] A deep learning-based method for evaluating the stability of steep slope bridges includes the following steps:

[0007] S1: Acquire bridge area DEM data, bridge displacement data, and bridge stress data;

[0008] S2: Construct a bridge stability evaluation network and extract DEM features of the bridge area based on a convolutional neural network;

[0009] S3: Extracting internal structural features of bridges based on recurrent neural networks;

[0010] S4: Fusion of bridge region DEM features and bridge internal structural features based on attention mechanism;

[0011] S5: Define the network's input, output, and optimization objectives, and optimize the network parameters based on the adaptive learning rate optimization algorithm;

[0012] Furthermore, the acquisition of bridge area DEM data, bridge displacement data, and bridge stress data in step S1 includes:

[0013] S11: Acquire bridge displacement and stress data through sensors, and acquire remote sensing image data of the bridge area through satellite;

[0014] S12: Perform image correction, geometric correction, and atmospheric correction on the remote sensing image data of the bridge area, and register the image with ground control points to generate a stereo image;

[0015] S13: Using stereo imagery and a stereo matching algorithm, extract the DEM data for the bridge area. DEM (Digital Elevation Model) is a data structure representing terrain elevation. The stereo matching algorithm extracts the disparity value of each pixel from the stereo imagery. Then, based on camera parameters and geometric relationships, the disparity values ​​are converted into elevation values, thus obtaining the DEM data for the bridge area.

[0016] Further, step S2 extracts DEM features of the bridge region based on a convolutional neural network, including:

[0017] S21: Construct a bridge stability evaluation network using convolutional neural networks, GRU networks, and attention mechanisms;

[0018] S22: Two sets of convolutional neural networks, each consisting of convolutional layers, activation functions, and pooling layers, are used to convolve the DEM data of the bridge area. The structure of the first set of convolutional neural networks is as follows:

[0019] The first group of convolutional neural networks uses a 3×3 kernel with a stride of 1 and padding of 1 to convolve the input image, resulting in 32 output channels. The convolution is calculated as follows:

[0020]

[0021] Among them, O i,j,k Let x be the (i, j)th pixel value of the k-th channel of the output feature map after convolution, where i and j are the row and column indices of the matrix after convolution, respectively, and k is the channel index. u,v c is the (u, v)th element of the original matrix. i-u,j-vis the (iu, jv)th element of the convolution kernel, where u and v are the row and column indices of the original matrix; in the first group of convolutional neural networks, a 3x3 convolution kernel with a stride of 1 and padding of 1 is selected, which can keep the image size unchanged while capturing high-frequency information such as edges and corners that correspond to changes in image details;

[0022] The number of channels determines the number of output feature maps of the convolutional layer, which in turn determines the types of features the convolutional layer can extract. More channels mean richer features can be extracted, but also increase computational and memory consumption. Therefore, a channel count of 32 is appropriate, ensuring the convolutional layer can extract spatial and elevation information of the bridge area while avoiding excessive computation and memory consumption.

[0023] The first group of convolutional neural networks uses the ReLU function as the activation function to perform a non-linear transformation on the output of the convolutional layer. The calculation method is as follows:

[0024] A i,j,k =max(0,O) i,j,k )

[0025] Among them, A i,j,k The output value is the activation function's output value. Since ReLU only needs to determine if the input is greater than 0 without performing complex calculations, it is computationally fast. Furthermore, ReLU effectively solves the gradient vanishing problem and enhances the network's sparsity by including more zero values ​​in the network's output, reducing network parameters and computational cost, and improving the network's generalization ability.

[0026] The pooling layers of the first convolutional neural network use max pooling to downsample the output of the activation function. The pooling kernel size is 2×2, the stride is 2, and the number of output channels is 32. The calculation method is as follows:

[0027]

[0028] Among them, P i,j,k The output value of the pooling layer is denoted by m and n, which are the counting indices. After pooling, the output value of the pooling layer is input into the second convolutional neural network. Max pooling can divide the input feature map into several small regions, and then select the maximum value in each small region as the output. This is beneficial for focusing on the spatial and elevation changes of various parts of the bridge area. At the same time, it increases the translation invariance of the feature map, that is, when the satellite image of the input bridge area changes slightly in shooting angle, the output of max pooling will not change significantly, thereby improving the robustness of the network.

[0029] The structure of the second group of convolutional neural networks is as follows:

[0030] The second set of convolutional neural networks uses a 5×5 kernel with a stride of 1 and padding of 2 to convolve the output of the pooling layer of the first set of convolutional neural networks, resulting in an output channel count of 64.

[0031] The activation functions and pooling layers of the second convolutional neural network are set in the same way as those of the first convolutional neural network; the calculation methods of the convolutional layers, activation functions, and pooling layers of the second convolutional neural network are also the same as those of the first convolutional neural network.

[0032] In the second set of convolutional neural networks, selecting a 5x5 convolutional kernel with a stride of 1 and padding of 2 can increase the receptive field, that is, the range of pixels covered by the convolutional kernel, and at the same time capture low-frequency information such as texture and shape corresponding to smooth changes in the image.

[0033] Meanwhile, since the second set of convolutional neural networks needs to extract higher-level bridge region DEM features, and is based on the first set of convolutional neural networks, it is necessary to increase the number of output channels to ensure that the convolutional layers have sufficient feature extraction capabilities. The first set of convolutional neural networks has already performed pooling, which reduces the size of the feature map. Therefore, the second set of convolutional neural networks has 64 output channels, which will not bring excessive computational load and memory consumption.

[0034] S23: Using fully connected layers and activation functions, the output of the second set of convolutional layers is compressed into a one-dimensional vector as a representation of the DEM features of the bridge region. The calculation method of the fully connected layer is as follows:

[0035]

[0036] F = [F k＝1 ,F k＝2 ,…,F k＝k ]

[0037] Among them, F k W represents the pixel value of the k-th channel of the output feature map of the fully connected layer. i,j,k b represents the weights of the fully connected layer. i,j,k is the bias term of the fully connected layer, N is the length of the vector after the pooling layer is flattened, F is the concatenation result of the pixel values ​​of each channel of the fully connected layer, and [] is the concatenation operation. The fully connected layer compresses the output feature map into a one-dimensional vector, realizing the dimensionality reduction and abstraction of the DEM features of the bridge area. Then, the features of different channels are combined through the concatenation operation, thereby enhancing the feature discrimination and laying the groundwork for subsequent feature fusion based on the Attention mechanism.

[0038] Furthermore, step S3 extracts the internal structural features of the bridge based on a recurrent neural network, including:

[0039] S31: The bridge displacement data and bridge stress data are concatenated into a two-dimensional matrix, where each row represents the data of a time step and each column represents the data of a measurement point, which serves as the input to the recurrent neural network.

[0040] S32: A GRU network is used to model the sequence of the input vector, and the output hidden state vector is used as a representation of the internal structural features of the bridge. The GRU is calculated as follows:

[0041] Renew t =sigmoid(W z ·[hidden t-1 input t ])

[0042] Reset t =sigmoid(W r ·[hidden t-1 ,inpu t ])

[0043]

[0044] Among them, Renew t This is the update gate for the current time step, where t is the current time step, sigmoid() is the sigmoid function, and W is the update gate for the current time step. z The weight matrix is ​​hidden. t-1 The input is the hidden state from the previous time step. t Reset is the input for the current time step. t W is the reset gate for the current time step. r This is the weight matrix. This represents the candidate hidden states at the current time step, where tanh() is the tanh function, W is the weight matrix, ⊙ represents element-wise multiplication, and hidden is the hidden state. t The hidden state is the current time step. GRU dynamically adjusts the flow of information through update gates and reset gates, transforming the temporal correlation of bridge displacement data and bridge stress data into a hidden state vector. It adaptively learns the complexity and diversity of the bridge's internal structure, thereby capturing the long-term dependence and short-term changes of the input sequence while avoiding the problems of gradient vanishing or exploding.

[0045] Furthermore, step S4, based on an attention mechanism, fuses the DEM features of the bridge region with the internal structural features of the bridge, including:

[0046] S41: Combine the DEM features of the bridge area with the internal structural features of the bridge to form a two-dimensional matrix;

[0047] S42: Using a fully connected layer and a softmax function, an attention weight is calculated for each row of the matrix, representing the importance of the features in that row. The attention weight is calculated as follows:

[0048] a t =softmax(W a [hidden t ,F]+b a )

[0049] Among them, a t The attention weight vector is W, where softmax() is the softmax function. a Let b be the weight matrix. a It is the bias vector;

[0050] S43: Using another fully connected layer and a sigmoid function, each column of the matrix is ​​transformed to obtain a new feature matrix. Then, the matrix is ​​weighted and summed according to the attention weights to obtain a one-dimensional vector, which serves as the fusion representation of the bridge area DEM features and the bridge's internal structural features. The calculation methods involved are as follows:

[0051]

[0052] Where I is the fusion representation vector and T is the total step size of the time step; the Attention mechanism can adaptively weight the input features, highlight important features, suppress irrelevant features, and integrate different dimensions of bridge area DEM features and bridge internal structural features to obtain a more comprehensive and accurate bridge feature representation, so as to reflect the overall state and local changes of the bridge and provide richer and more reliable information for subsequent bridge health assessment.

[0053] Further, step S5 sets the network's input, output, and optimization objective, and optimizes the network parameters based on an adaptive learning rate optimization algorithm, including:

[0054] S51: The fusion representation of bridge area DEM features and bridge internal structural features is used as the input of the network, and the bridge stability rating is used as the output of the network.

[0055] S52: Select the cross-entropy loss function L CNN As the loss function of CNN, the smoothing loss function L GRU The loss function for GRU is calculated as follows:

[0056]

[0057] Where H is the sample size, D is the number of categories for bridge stability rating, and y hdp is the true label of the d-th class for the h-th sample. hd Let be the predicted probability of the d-th class for the h-th sample. Let y be the predicted value of the h-th sample at the t-th time step. ht Let σ be the true value of the h-th sample at the t-th time step, and let σ be the smoothing function, calculated as follows:

[0058]

[0059] Where f is the input of the smoothing function; the smoothing function can approximate the original function with a quadratic function near the inflection point, reducing the sensitivity of the GRU loss function to outliers and improving the model's iterative stability;

[0060] S53: The weighted sum of the loss functions of CNN and GRU is used as the optimization objective to obtain the total loss function of the network. The final loss function is:

[0061] L total =L CNN +L GRU

[0062] S54: Optimize the parameters of the bridge stability evaluation network based on the adaptive learning rate optimization algorithm. The calculation method is as follows:

[0063] q p =β1·q p-1 +(1-β1)·g p

[0064]

[0065] Where, q p This is a first-order moment estimate, where p is the number of iterations and q is the value of q. p-1 This is the first-order moment estimate from the previous iteration, where β1 and β2 are decay coefficients, and g... p Let s be the gradient of the current iteration step. p For second-order moment estimation, s p-1 This is the second-order moment estimate from the previous iteration. and To correct for the biases in the first and second moments, η is the learning rate, and θ is the learning rate. p Let θ be the parameter of the bridge stability evaluation network. p+1 ∈ represents the parameters of the bridge stability evaluation network after iteration, where ∈ is a constant.

[0066] This invention also provides a deep learning-based stability evaluation system for steep slope bridges, comprising:

[0067] Data acquisition module: Acquires bridge area DEM data, bridge displacement data, and bridge stress data;

[0068] DEM Feature Extraction Module: Extracts DEM features of the bridge area based on a convolutional neural network;

[0069] Structural feature extraction module: Extracts internal structural features of bridges based on recurrent neural networks;

[0070] Feature fusion module: Fusion of bridge region DEM features and bridge internal structural features based on attention mechanism;

[0071] Network training module: Sets the network's input, output, and optimization objectives, and optimizes the network parameters based on an adaptive learning rate optimization algorithm.

[0072] Compared with the prior art, the advantages of the present invention are:

[0073] (1) In terms of the overall scheme, the present invention integrates the DEM data of the bridge area and the internal structure data of the bridge into multimodal heterogeneous data, and comprehensively considers the influence of the terrain and the internal structural features of the bridge to achieve the assessment of the stability of the bridge.

[0074] (2) In terms of algorithm model, the present invention first uses CNN to extract the DEM features of the bridge area and converts the image data into vector data to complete the structural unification of heterogeneous data; secondly, it uses GRU to extract the internal structural features of the bridge, and finally uses the attention mechanism to judge the importance of features and perform adaptive fusion to realize the fusion of multi-dimensional heterogeneous data. While ensuring a low number of parameters, it improves the accuracy and efficiency of bridge stability evaluation.

[0075] (3) In terms of network parameter optimization, this invention adopts a new loss function, combining the cross-entropy loss function and the smoothing loss function to optimize each part of the bridge stability evaluation network, thereby improving the model training effect and fitting effect. Attached Figure Description

[0076] Figure 1 A flowchart illustrating a deep learning-based method for evaluating the stability of steep slope bridges provided by this invention.

[0077] Figure 2 This is a schematic diagram of the algorithm flow for a deep learning-based method for evaluating the stability of steep slope bridges, provided by the present invention. Detailed Implementation

[0078] The present invention will be further described below with reference to the accompanying drawings, but this is not intended to limit the present invention in any way. Any modifications or substitutions made based on the teachings of the present invention shall fall within the protection scope of the present invention.

[0079] Example 1: A deep learning-based method for evaluating the stability of steep slope bridges, such as... Figure 1 As shown, the following steps are included:

[0080] S1: Acquire bridge area DEM data, bridge displacement data, and bridge stress data;

[0081] S11: Acquire bridge displacement and stress data through sensors, and acquire remote sensing image data of the bridge area through satellite;

[0082] S12: Perform image correction, geometric correction, and atmospheric correction on the remote sensing image data of the bridge area, and register the image with ground control points to generate a stereo image;

[0083] S13: Using stereo imagery and a stereo matching algorithm, extract the DEM data for the bridge area. DEM (Digital Elevation Model) is a data structure representing terrain elevation. The stereo matching algorithm extracts the disparity value of each pixel from the stereo imagery. Then, based on camera parameters and geometric relationships, the disparity values ​​are converted into elevation values, thus obtaining the DEM data for the bridge area.

[0084] S2: Construct a bridge stability evaluation network and extract DEM features of the bridge area based on a convolutional neural network;

[0085] S21: Construct a bridge stability evaluation network using convolutional neural networks, GRU networks, and attention mechanisms;

[0086] S22: Two sets of convolutional neural networks, each consisting of convolutional layers, activation functions, and pooling layers, are used to convolve the DEM data of the bridge area. The structure of the first set of convolutional neural networks is as follows:

[0087] The first group of convolutional neural networks uses a 3×3 kernel with a stride of 1 and padding of 1 to convolve the input image, resulting in 32 output channels. The convolution is calculated as follows:

[0088]

[0089] Among them, O i,j,k Let x be the (i, j)th pixel value of the k-th channel of the output feature map after convolution, where i and j are the row and column indices of the matrix after convolution, respectively, and k is the channel index. u,v c is the (u, v)th element of the original matrix. i-u,j-v is the (iu, jv)th element of the convolution kernel, where u and v are the row and column indices of the original matrix; in the first group of convolutional neural networks, a 3x3 convolution kernel with a stride of 1 and padding of 1 is selected, which can keep the image size unchanged while capturing high-frequency information such as edges and corners that correspond to changes in image details;

[0090] The number of channels determines the number of output feature maps of the convolutional layer, which in turn determines the types of features the convolutional layer can extract. More channels mean richer features can be extracted, but also increase computational and memory consumption. Therefore, a channel count of 32 is appropriate, ensuring the convolutional layer can extract spatial and elevation information of the bridge area while avoiding excessive computation and memory consumption.

[0091] The first group of convolutional neural networks uses the ReLU function as the activation function to perform a non-linear transformation on the output of the convolutional layer. The calculation method is as follows:

[0092] A i,j,k =max(0,O) i,j,k )

[0093] Among them, A i,j,k The output value is the activation function's output value. Since ReLU only needs to determine if the input is greater than 0 without performing complex calculations, it is computationally fast. Furthermore, ReLU effectively solves the gradient vanishing problem and enhances the network's sparsity by including more zero values ​​in the network's output, reducing network parameters and computational cost, and improving the network's generalization ability.

[0094] The pooling layers of the first convolutional neural network use max pooling to downsample the output of the activation function. The pooling kernel size is 2×2, the stride is 2, and the number of output channels is 32. The calculation method is as follows:

[0095]

[0096] Among them, P i,j,k The output value of the pooling layer is denoted by m and n, which are the counting indices. After pooling, the output value of the pooling layer is input into the second convolutional neural network. Max pooling can divide the input feature map into several small regions, and then select the maximum value in each small region as the output. This is beneficial for focusing on the spatial and elevation changes of various parts of the bridge area. At the same time, it increases the translation invariance of the feature map, that is, when the satellite image of the input bridge area changes slightly in shooting angle, the output of max pooling will not change significantly, thereby improving the robustness of the network.

[0097] The structure of the second group of convolutional neural networks is as follows:

[0098] The second set of convolutional neural networks uses a 5×5 kernel with a stride of 1 and padding of 2 to convolve the output of the pooling layer of the first set of convolutional neural networks, resulting in an output channel count of 64.

[0099] The activation functions and pooling layers of the second convolutional neural network are set in the same way as those of the first convolutional neural network; the calculation methods of the convolutional layers, activation functions, and pooling layers of the second convolutional neural network are also the same as those of the first convolutional neural network.

[0100] In the second set of convolutional neural networks, selecting a 5x5 convolutional kernel with a stride of 1 and padding of 2 can increase the receptive field, that is, the range of pixels covered by the convolutional kernel, and at the same time capture low-frequency information such as texture and shape corresponding to smooth changes in the image.

[0101] Meanwhile, since the second set of convolutional neural networks needs to extract higher-level bridge region DEM features, and is based on the first set of convolutional neural networks, it is necessary to increase the number of output channels to ensure that the convolutional layers have sufficient feature extraction capabilities. The first set of convolutional neural networks has already performed pooling, which reduces the size of the feature map. Therefore, the second set of convolutional neural networks has 64 output channels, which will not bring excessive computational load and memory consumption.

[0102] S23: Using fully connected layers and activation functions, the output of the second set of convolutional layers is compressed into a one-dimensional vector as a representation of the DEM features of the bridge region. The calculation method of the fully connected layer is as follows:

[0103]

[0104] F = [F k＝1 ,F k＝2 ,…,F k＝k ]

[0105] Among them, F k W represents the pixel value of the k-th channel of the output feature map of the fully connected layer. i,j,k b represents the weights of the fully connected layer. i,j,k is the bias term of the fully connected layer, N is the length of the vector after the pooling layer is flattened, F is the concatenation result of the pixel values ​​of each channel of the fully connected layer, and [] is the concatenation operation. The fully connected layer compresses the output feature map into a one-dimensional vector, realizing the dimensionality reduction and abstraction of the DEM features of the bridge area. Then, the features of different channels are combined through the concatenation operation, thereby enhancing the feature discrimination and laying the groundwork for subsequent feature fusion based on the Attention mechanism.

[0106] For example, assuming the length and width of the DEM data are 256, after the first convolutional layer, the output feature map has a size of 256x256x32 and a dimension of 3; after the first pooling layer, the output feature map has a size of 128x128x32 and a dimension of 3; after the second convolutional layer, the output feature map has a size of 128x128x64 and a dimension of 3; after the second pooling layer, the output feature map has a size of 64x64x64 and a dimension of 3; after the fully connected layer, the output feature map has a size of 1x1x4096 and a dimension of 3, which serves as a representation of the bridge region DEM features.

[0107] S3: Extracting internal structural features of bridges based on recurrent neural networks;

[0108] S31: The bridge displacement data and bridge stress data are concatenated into a two-dimensional matrix, where each row represents the data of a time step and each column represents the data of a measurement point, which serves as the input to the recurrent neural network.

[0109] Assuming the bridge has X measuring points, and each measuring point has two data points, displacement and stress, then the data at each time step is a 2X-dimensional vector. By concatenating the data from T time steps, we obtain a T×2X matrix, which serves as the input to the recurrent neural network.

[0110] For example, assuming a bridge has four measuring points, and the displacement and stress data for each measuring point are shown in the table below, then the input to GRU at each time step would be:

[0111] Time step Displacement of measuring point 1 Stress at measuring point 1 Displacement at measuring point 2 Stress at measuring point 2 Displacement of measuring point 3 Stress at measuring point 3 Displacement at measuring point 4 Stress at measuring point 4 1 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 2 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 3 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0

[0112] S32: A GRU network is used to model the sequence of the input vector, and the output hidden state vector is used as a representation of the internal structural features of the bridge. The GRU is calculated as follows:

[0113] Renew t =sigmoid(W z ·[hidden t-1 input t ])

[0114] Reset t =sigmoid(W r ·[hidden t-1 input t ])

[0115]

[0116] Among them, Renew tThis is the update gate for the current time step, where t is the current time step, sigmoid() is the sigmoid function, and W is the update gate for the current time step. z The weight matrix is ​​hidden. t-1 The input is the hidden state from the previous time step. t Reset is the input for the current time step. t W is the reset gate for the current time step. r This is the weight matrix. This represents the candidate hidden states at the current time step, where tanh() is the tanh function, W is the weight matrix, ⊙ represents element-wise multiplication, and hidden is the hidden state. t The hidden state is the current time step. GRU dynamically adjusts the flow of information through update gates and reset gates, transforming the temporal correlation of bridge displacement data and bridge stress data into a hidden state vector. It adaptively learns the complexity and diversity of the bridge's internal structure, thereby capturing the long-term dependence and short-term changes of the input sequence while avoiding the problems of gradient vanishing or exploding.

[0117] S4: Fusion of bridge region DEM features and bridge internal structural features based on an attention mechanism, such as... Figure 2 As shown;

[0118] S41: Combine the DEM features of the bridge area with the internal structural features of the bridge to form a two-dimensional matrix;

[0119] S42: Using a fully connected layer and a softmax function, an attention weight is calculated for each row of the matrix, representing the importance of the features in that row. The attention weight is calculated as follows:

[0120] a t =softmax(W a [hidden t ,F]+b a )

[0121] Among them, a t The attention weight vector is W, where softmax() is the softmax function. a Let b be the weight matrix. a It is the bias vector;

[0122] S43: Using another fully connected layer and a sigmoid function, each column of the matrix is ​​transformed to obtain a new feature matrix. Then, the matrix is ​​weighted and summed according to the attention weights to obtain a one-dimensional vector, which serves as the fusion representation of the bridge area DEM features and the bridge's internal structural features. The calculation methods involved are as follows:

[0123]

[0124] Where I is the fusion representation vector and T is the total step size of the time step; the Attention mechanism can adaptively weight the input features, highlight important features, suppress irrelevant features, and integrate different dimensions of bridge area DEM features and bridge internal structural features to obtain a more comprehensive and accurate bridge feature representation, so as to reflect the overall state and local changes of the bridge and provide richer and more reliable information for subsequent bridge health assessment.

[0125] For example, given two sets of DEM data, each with a length and width of 256, representing different terrain features of the bridge area, after passing through a convolutional neural network and a fully connected layer, two one-dimensional vectors of length 4096 are obtained, which serve as representations of the DEM features of the bridge area, denoted as F. (1) and F (2) Similarly, after processing the two sets of bridge displacement and stress data using GRU, two one-dimensional vectors of length 1024 are generated, representing the characteristics of the bridge's internal structure, denoted as H. (1) and H (2) ;

[0126] The DEM features of the bridge area and the internal structural features of the bridge are stitched together to form two two-dimensional matrices of size 5120x2, denoted as M. (1) and M (2) ;

[0127] After attention calculation, two sets of attention weights are obtained, namely a (1) = [0.38, 0.62], a (2) = [0.17, 0.83]; it can be seen that a (2) The first element is greater than a (1) The smaller the first element, the less the DEM features have an impact on the second set of data; while the internal structural features of the bridge have a greater impact and are more important for the health assessment of the bridge.

[0128] The classification results of the bridge health assessment are as follows:

[0129] I<0.7: Bridge health;

[0130] 0.7 <= I < 0.8: The bridge is slightly damaged;

[0131] 0.8 <= I < 0.85: Moderate damage to the bridge;

[0132] 0.85 <= I < 0.9: The bridge is severely damaged;

[0133] 0.9 <= I < 1: The extent of bridge damage is unacceptable;

[0134] S5: Define the network's input, output, and optimization objectives, and optimize the network parameters based on the adaptive learning rate optimization algorithm;

[0135] S51: The fusion representation of bridge area DEM features and bridge internal structural features is used as the input of the network, and the bridge stability rating is used as the output of the network.

[0136] S52: Select the cross-entropy loss function L CNN As the loss function of CNN, the smoothing loss function L GRU The loss function for GRU is calculated as follows:

[0137]

[0138] Where H is the sample size, D is the number of categories for bridge stability rating, and y hd p is the true label of the d-th class for the h-th sample. hd Let be the predicted probability of the d-th class for the h-th sample. Let y be the predicted value of the h-th sample at the t-th time step. ht Let σ be the true value of the h-th sample at the t-th time step, and let σ be the smoothing function, calculated as follows:

[0139]

[0140] Where f is the input of the smoothing function; the smoothing function can approximate the original function with a quadratic function near the inflection point, reducing the sensitivity of the GRU loss function to outliers and improving the model's iterative stability;

[0141] S53: The weighted sum of the loss functions of CNN and GRU is used as the optimization objective to obtain the total loss function of the network. The final loss function is:

[0142] L total =L CNN +L GRU

[0143] S54: Optimize the parameters of the bridge stability evaluation network based on the adaptive learning rate optimization algorithm. The calculation method is as follows:

[0144] q p =β1·q p-1 +(1-β1)·g p

[0145]

[0146] Where, q p This is a first-order moment estimate, where p is the number of iterations and q is the value of q. p-1This is the first-order moment estimate from the previous iteration, where β1 and β2 are decay coefficients, and g... p Let s be the gradient of the current iteration step. p For second-order moment estimation, s p-1 This is the second-order moment estimate from the previous iteration. and To correct for the biases in the first and second moments, η is the learning rate, and θ is the learning rate. p Let θ be the parameter of the bridge stability evaluation network. p+1 ∈ represents the parameters of the bridge stability evaluation network after iteration, where ∈ is a constant.

[0147] Example 2: This invention also discloses a deep learning-based stability evaluation system for steep slope bridges, comprising:

[0148] Data acquisition module: Acquires bridge area DEM data, bridge displacement data, and bridge stress data;

[0149] DEM Feature Extraction Module: Extracts DEM features of the bridge area based on a convolutional neural network;

[0150] Structural feature extraction module: Extracts internal structural features of bridges based on recurrent neural networks;

[0151] Feature fusion module: Fusion of bridge region DEM features and bridge internal structural features based on attention mechanism;

[0152] Network training module: Sets the network's input, output, and optimization objectives, and optimizes the network parameters based on an adaptive learning rate optimization algorithm.

[0153] It should be noted that the sequence numbers of the above embodiments of the present invention are merely for descriptive purposes and do not represent the superiority or inferiority of the embodiments. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, apparatus, article, or method that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, apparatus, article, or method. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, apparatus, article, or method that includes that element.

[0154] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk) as described above, and includes several instructions to cause a terminal device (which may be a mobile phone, computer, server, or network device, etc.) to execute the methods described in the various embodiments of the present invention.

[0155] The above are merely preferred embodiments of the present invention and do not limit the scope of the patent. Any equivalent structural or procedural transformations made based on the description and drawings of the present invention, or direct or indirect applications in other related technical fields, are similarly included within the scope of patent protection of the present invention.

Claims

1. A method for evaluating the stability of steep slope bridges based on deep learning, characterized in that, Includes the following steps: S1: Acquire bridge area DEM data, bridge displacement data, and bridge stress data; S2: Construct a bridge stability evaluation network and extract DEM features of the bridge area based on a convolutional neural network; Step S2 includes the following steps: S21: Construct a bridge stability evaluation network using convolutional neural networks, GRU networks, and attention mechanisms; S22: Two sets of convolutional neural networks, each consisting of convolutional layers, activation functions, and pooling layers, are used to convolve the DEM data of the bridge area. The structure of the first set of convolutional neural networks is as follows: The first group of convolutional neural networks uses a 3×3 kernel with a stride of 1 and padding of 1 to convolve the input image, resulting in 32 output channels. The convolution is calculated as follows: Among them, O i,j,k Let x be the (i, j)th pixel value of the k-th channel of the output feature map after convolution, where i and j are the row and column indices of the matrix after convolution, respectively, and k is the channel index. u,v c is the (u, v)th element of the original matrix. i-u,j-v Let be the (iu, jv)th element of the convolution kernel, where u and v are the row and column indices of the original matrix; The first group of convolutional neural networks uses the ReLU function as the activation function to perform a non-linear transformation on the output of the convolutional layer. The calculation method is as follows: A i,j,k =max(0,O i,j,k ) Among them, A i,j,k The output value of the activation function; The pooling layers of the first convolutional neural network use max pooling to downsample the output of the activation function. The pooling kernel size is 2×2, the stride is 2, and the number of output channels is 32. The calculation method is as follows: Among them, P i,j,k The output value of the pooling layer is , and m and n are the counting indices. After pooling is completed, the output value of the pooling layer is input into the second convolutional neural network. The structure of the second group of convolutional neural networks is as follows: The second set of convolutional neural networks uses a 5×5 kernel with a stride of 1 and padding of 2 to convolve the output of the pooling layer of the first set of convolutional neural networks, resulting in an output channel count of 64. The activation functions and pooling layers of the second convolutional neural network are set in the same way as those of the first convolutional neural network; the calculation methods of the convolutional layers, activation functions, and pooling layers of the second convolutional neural network are also the same as those of the first convolutional neural network. S23: Using fully connected layers and activation functions, the output of the second set of convolutional layers is compressed into a one-dimensional vector as a representation of the DEM features of the bridge region. The calculation method of the fully connected layer is as follows: F=[F k＝1 ,F k＝2 ,…,F k＝k ] Among them, F k W represents the pixel value of the k-th channel of the output feature map of the fully connected layer. i,j,k b represents the weights of the fully connected layer. i,j,k is the bias term of the fully connected layer, N is the vector length after the pooling layer is flattened, F is the concatenation result of the pixel values ​​of each channel of the fully connected layer, and [] is the concatenation operation; S3: Extracting internal structural features of bridges based on recurrent neural networks; S4: Fusion of bridge region DEM features and bridge internal structural features based on attention mechanism; S5: Define the network's input, output, and optimization objectives, and optimize the network parameters based on an adaptive learning rate optimization algorithm.

2. The deep learning-based method for evaluating the stability of steep slope bridges according to claim 1, characterized in that, Step S1 includes the following steps: S11: Acquire bridge displacement and stress data through sensors, and acquire remote sensing image data of the bridge area through satellite; S12: Perform image correction, geometric correction, and atmospheric correction on the remote sensing image data of the bridge area, and register the image with ground control points to generate a stereo image; S13: Using stereo imagery and a stereo matching algorithm, extract the DEM data of the bridge area.

3. The deep learning-based method for evaluating the stability of steep slope bridges according to claim 2, characterized in that, Step S3 includes the following steps: S31: Concatenate the bridge displacement data and bridge stress data into a two-dimensional matrix, where each row represents the data of a time step and each column represents the data of a measurement point, which serves as the input vector for the recurrent neural network. S32: A GRU network is used to model the sequence of the input vector, and the output hidden state vector is used as a representation of the internal structural features of the bridge. The GRU is calculated as follows: Renew t =sigmoid(W z ·[hidden t-1 ,input t ]) Reset t =sigmoid(W r ·[hidden t-1 ,input t ]) Among them, Renew t This is the update gate for the current time step, where t is the current time step, sigmoid() is the sigmoid function, and W is the update gate for the current time step. z The weight matrix is ​​hidden. t-1 The input is the hidden state from the previous time step. t Reset is the input for the current time step. t W is the reset gate for the current time step. r This is the weight matrix. This represents the candidate hidden states at the current time step, where tanh() is the tanh function, W is the weight matrix, ⊙ represents element-wise multiplication, and hidden is the hidden state. t The hidden state at the current time step.

4. The deep learning-based method for evaluating the stability of steep slope bridges according to claim 3, characterized in that, Step S4 includes the following steps: S41: Combine the DEM features of the bridge area with the internal structural features of the bridge to form a two-dimensional matrix; S42: Using a fully connected layer and a softmax function, attention weights are calculated for each row of the matrix. The attention weights are calculated as follows: a t =softmax(W a [hidden t ,F]+b a ) Among them, a t The attention weight vector is W, where softmax() is the softmax function. a Let b be the weight matrix. a It is the bias vector; S43: Using another fully connected layer and a sigmoid function, each column of the matrix is ​​transformed to obtain a new feature matrix. Then, the matrix is ​​weighted and summed according to the attention weights to obtain a one-dimensional vector, which serves as the fusion representation of the bridge area DEM features and the bridge's internal structural features. The calculation methods involved are as follows: Where I is the fusion representation vector and T is the total step size of the time step.

5. The deep learning-based method for evaluating the stability of steep slope bridges according to claim 4, characterized in that, Step S5 includes the following steps: S51: The fusion representation of bridge area DEM features and bridge internal structural features is used as the input of the network, and the bridge stability rating is used as the output of the network. S52: Select the cross-entropy loss function L CNN As the loss function of CNN, the smoothing loss function L GRU The loss function for GRU is calculated as follows: Where H is the sample size, D is the number of categories for bridge stability rating, and y hd p is the true label of the d-th class for the h-th sample. hd Let be the predicted probability of the d-th class for the h-th sample. Let y be the predicted value of the h-th sample at the t-th time step. ht Let σ be the true value of the h-th sample at the t-th time step, and let σ be the smoothing function, calculated as follows: Where f is the input to the smoothing function; S53: The weighted sum of the loss functions of CNN and GRU is used as the optimization objective to obtain the total loss function of the network. The final loss function is: L total =L CNN +L GRU S54: Optimize the parameters of the bridge stability evaluation network based on the adaptive learning rate optimization algorithm. The calculation method is as follows: q p =β1·q p-1 +(1-β1)·g p Where, q p This is a first-order moment estimate, where p is the number of iterations and q is the value of q. p-1 This is the first-order moment estimate from the previous iteration, where β1 and β2 are decay coefficients, and g... p Let s be the gradient of the current iteration step. p For second-order moment estimation, s p-1 This is the second-order moment estimate from the previous iteration. and To correct for the biases in the first and second moments, η is the learning rate, and θ is the learning rate. p Let θ be the parameter of the bridge stability evaluation network. p+1 ∈ represents the parameters of the bridge stability evaluation network after iteration, where ∈ is a constant.

6. A deep learning-based stability evaluation system for steep slope bridges, characterized in that, include: Data acquisition module: Acquires bridge area DEM data, bridge displacement data, and bridge stress data; DEM Feature Extraction Module: Extracts DEM features of the bridge area based on a convolutional neural network; Structural feature extraction module: Extracts internal structural features of bridges based on recurrent neural networks; Feature fusion module: Fusion of bridge region DEM features and bridge internal structural features based on attention mechanism; Network training module: Sets the network's input, output, and optimization objectives, and optimizes the network parameters based on an adaptive learning rate optimization algorithm; To achieve the deep learning-based stability evaluation method for steep slope bridges as described in any one of claims 1-5.

Images