Cross-domain bridge damage identification method based on deep learning

Abstract

The cross-domain bridge damage identification method based on deep learning disclosed in the present invention includes the following steps: establishing a vehicle-bridge finite element model; simulating the real structure of the bridge by adding uncertainty to the bridge finite element model in the vehicle-bridge finite element model; data pre-processing; constructing a generative adversarial neural network; training and generation phase; constructing a dynamic domain adversarial adaptive network; training phase; using the trained feature extractor F q and label classifier F y The target domain dataset is tested to obtain corresponding test results. The present invention utilizes the aforementioned deep learning-based cross-domain bridge damage identification method, which uses a deep generative adversarial network to simulate the expansion of the bridge displacement response signal generated by a two-axle vehicle traveling at a constant speed on the bridge. A dynamic adversarial adaptive network is then used to project the source and target domain features into the same feature space, enabling cross-domain damage feature extraction and adaptation, and completing cross-domain bridge damage identification.

Description

Technical Field

[0001] The present invention relates to a bridge detection technology, and in particular to a cross-domain bridge damage identification method based on deep learning. Background Art

[0002] Due to the powerful and efficient ability of deep learning to learn and predict large amounts of data, data-driven data mining techniques have been developed for structural damage detection in the field of structural health monitoring. Using structural responses as input for feature mining, deep learning can mine damage-sensitive features from massive amounts of data without knowing the specific structural information, and is more effective than many traditional methods.

[0003] However, research on data-driven deep learning methods for structural damage detection is far from mature. A major challenge is the lack of labeled damage data from actual structures, as the structural condition is unknown in advance. Some researchers have built finite element models of real structures to generate labeled damage data, which can consider all possible damage scenarios for network training. However, finite element models are affected by the real environment and parameters such as boundary conditions, which are difficult to determine and therefore difficult to model.

[0004] When a deep learning model trained based on a finite element model is applied to a real structure, the differences between the finite element model and the real structure may lead to performance degradation.

[0005] To address this issue, existing technologies have developed unsupervised domain adaptation to address data distribution differences between source and target domains. This approach intelligently applies knowledge learned from a labeled source domain to an unlabeled target domain. However, unsupervised domain adaptation presents another major challenge: it requires a large amount of data in both the target domain and the target domain. This lack of data in practice is a major bottleneck that hinders the application of unsupervised domain adaptation in real-world scenarios. Summary of the Invention

[0006] To solve the above problems, the purpose of the present invention is to provide a cross-domain bridge damage identification method based on deep learning. Based on dynamic domain adaptation after data generalization, it not only solves the problem that the actual scene in bridge detection cannot provide a large amount of damage data in the target domain, but also can dynamically migrate the knowledge of the source domain to the target domain, solving the gap between the finite element model of the bridge and the actual structure of the bridge due to environmental or modeling errors, thereby obtaining a damage identification model with good recognition accuracy and can be applied to actual situations.

[0007] To achieve the above objectives, the present invention provides a cross-domain bridge damage identification method based on deep learning, comprising the following steps:

[0008] Step 1: Establish a vehicle-bridge finite element model;

[0009] Step 2: Simulate the real bridge structure by adding uncertainty to the bridge finite element model in the vehicle-bridge finite element model;

[0010] Step 3: Data pre-processing

[0011] The source domain data obtained and target domain Perform normalization and interpolation processing to keep the spatial dimensions of all samples consistent and obtain the processed source domain data and target domain data

[0012] Step 4: Construct the θ and critic D ω Generative adversarial neural network composed of;

[0013] Step 5: Training and generation phase;

[0014] Step 6: Construct the feature extractor F q , label predictor F y , global domain discriminator F g and local domain discriminator F l Dynamic domain adversarial adaptive network composed of;

[0015] Step 7: Training phase;

[0016] Step 8: Use training to get the optimal parameters The feature extractor F q and label classifier F y For the target domain dataset Perform detection to obtain corresponding test results.

[0017] Preferably, step 1 specifically includes the following steps:

[0018] Step 1.1: Determine the bridge parameters, use the bridge parameters to construct the bridge finite element model, and divide the bridge finite element model into units and number them in sequence [1, 2, 3, …, C];

[0019] Step 1.2: Determine the two-axle vehicle parameters, build a vehicle finite element model based on the two-axle vehicle parameters, and set the uncertainty vehicle weight m v =m v0 +a0×sin(t0) and uncertain vehicle speed v=v0+a1×sin(t1), where a0, a1 are the amplitudes of change, t0, t1∈[0,2π];

[0020] Step 1.3: When the bridge is intact, the vehicle is loaded with an uncertain weight m. vThe Newmark-β method is used to calculate the displacement response of the bridge. Repeat step 1.3n times to get the sample set and build the tag in

[0021] Step 1.4: Simulate bridge damage by reducing the element stiffness δ×E0×I0 in the bridge finite element model, where δ is the reduction factor;

[0022] Perform stiffness reduction on unit 1 of the bridge division and repeat the method of obtaining the bridge displacement response in step 1.3 n times to obtain the sample set. and build the tag The sample label is the damaged unit number, that is,

[0023] Step 1.5: Repeat step 1.4 until all units of the bridge are divided and the source domain dataset is finally obtained. and label sets

[0024] Preferably, the bridge parameters in step 1.1 include the bridge moment of inertia I0, elastic modulus E0, density per linear meter ρ0 and length L b ;

[0025] The two-axle vehicle parameters in step 1.2 include the total weight m v0 , two wheelbases d1, d2 and driving speed v0.

[0026] Preferably, the moment of inertia I0 in step 1.1 is 1.3901, and the elastic modulus E0 is 3.5×10 10 pa, density per linear meter ρ0=18358 and length L b =25m, the number of division units C = 10, the bridge finite element model is divided into units and numbered in sequence [1, 2, 3, ..., 10];

[0027] The change amplitudes in step 1.2 are a0=50, a1=0.1;

[0028] The sampling frequency in step 1.3 is 500Hz;

[0029] The reduction factor δ in step 1.4 is 0.75.

[0030] Preferably, step 2 specifically includes the following steps:

[0031] Step 2.1: Add five types of uncertainty:

[0032] (1) The influence of temperature on the bridge is simulated by changing the elastic modulus of the bridge finite element model, that is, E' = E0 × (1 + ζ1), where ζ1∈(-0.05, 0.05);

[0033] (2) By setting the vertical stiffness E of the boundary nodes of the bridge finite element model v and angular stiffness E r To simulate the boundary condition form of bridge elastic support;

[0034] (3) The geometric error of bridge modeling is simulated by changing the moment of inertia I0 of the bridge finite element model, that is, I'=I0×(1+ζ2), where ζ2∈(-0.03, 0.03); at the same time, the material error of bridge modeling is simulated by changing the density ρ0 of the bridge finite element model, that is, ρ'=ρ0×(1+ζ3), where ζ3∈(-0.02, 0.02);

[0035] (4) Simulating the bridge deck roughness by setting the pavement roughness level A0 of the bridge finite element model;

[0036] After adding the above four uncertainties, the bridge finite element model is used to obtain the bridge displacement response by the method in step 1.3. And add to the obtained bridge displacement response: (5) obeys the mean of 0 and the variance of σ 2 Gaussian distribution noise γ~N(0,σ 2 ),get

[0037] Step 2.2: Repeat steps 1.3 to 1.4 to obtain the displacement response sample sets of the intact bridge and each unit with damage. Finally, the target domain dataset is obtained

[0038] Preferably, in step 4:

[0039] The generator G θ It includes l1 linear layers and c1 transposed convolutional layers in sequence. Each linear layer and each transposed convolutional layer except the last one is followed by a linear activation layer and a normalization layer. The convolution kernel size of the transposed convolutional layer is k1, the number of convolution kernels is h1, and the stride is s1.

[0040] The critic D ω It includes c2 convolutional layers and l2 linear layers in sequence. An activation layer, a regularization layer, and a maximum pooling layer are added after each convolutional layer. A regularization layer and an activation layer are added after each linear layer except the last layer. The convolution kernel size of the convolution layer is k2, the number of convolution kernels is h2, and the stride is s2. The convolution kernel size of the maximum pooling layer is k3, and the stride is s3.

[0041] Preferably, step 5 specifically includes the following steps:

[0042] Step 5.1: Training phase

[0043] The input of the generative adversarial network is the target domain data And randomly generate a noise vector z that obeys the Gaussian distribution, and then input the noise vector z into the generator G θ (z) to obtain the generated sample

[0044] Step 5.2: Calculate interpolation data

[0045]

[0046] Among them, ε obeys the uniform distribution U[0,1], and the target domain samples

[0047] Step 5.3: Put the target domain sample x and the interpolation sample and generate samples Input Critic D ω , calculate the loss function L:

[0048]

[0049] Among them, λ is the penalty weight, is the 2-norm after differentiation of the critic;

[0050] Step 5.4: Use the adaptive optimizer Adam to perform back-propagation gradient descent on the loss function L to solve the critic D in the loss function ω The current optimal parameters

[0051] Step 5.5: Randomly generate a batch of p noise sets that follow a Gaussian distribution Calculate the loss function F:

[0052]

[0053] Step 5.6: Use the adaptive optimizer Adam to perform back-propagation gradient descent on the loss function F to solve the generator G in the loss function θ The current optimal parameters

[0054] Step 5.7: Repeat steps 5.3-5.6 until the loss function L and F converge to the optimal state and the critic D is obtained. ω and generator G θ The optimal parameters and

[0055] Step 5.8: Generate Phase

[0056] Randomly generate n noises that obey Gaussian distribution Input the trained generator , we get the expanded target domain sample set

[0057] Step 5.9: Repeat steps 5.1 to 5.8 for target domain data As input, we get the trained generators for each category And the expanded datasets for each category Finally, the expanded dataset is obtained

[0058] Preferably, in step 6:

[0059] The feature extractor F q It includes c3 convolutional layers in sequence, and each convolutional layer is followed by an activation layer, a regularization layer, and a maximum pooling layer. The convolutional kernel size of the convolutional layer is k4, the stride is s4, and the convolution kernel size of the maximum pooling layer is k5, the stride is s5;

[0060] The label predictor F y It includes l3 linear layers in sequence;

[0061] The global domain identifier F g It includes 14 linear layers, two linear layers and an activation layer in the middle in sequence;

[0062] The local domain discriminator F l It includes 15 linear layers, two linear layers and an intermediate activation layer in sequence.

[0063] Preferably, step 7 specifically includes the following steps:

[0064] Step 7.1: From the source domain dataset Randomly select a batch containing n s samples Pass through the feature extractor F q and label classifier F y Calculate the label loss function L y :

[0065]

[0066] in, for 's label;

[0067] Step 7.2: Set Domain Labels

[0068] The source data field label is set to Expand the target domain data field label to

[0069] Step 7.3: From the source domain dataset Randomly select a batch containing n s samples and from the expanded target domain dataset Randomly select a batch containing n t samples Pass through the feature extractor F q and the global domain discriminator F g Calculate the global domain loss function L g :

[0070]

[0071] Among them, L d is the cross entropy function, d i is x i The domain label of

[0072] Step 7.4: Separately from the source domain dataset Randomly select a batch containing n s samples and from the expanded target domain dataset Randomly select a batch containing n t samples Pass through the feature extractor F q and local domain discriminator F l , and calculate the local domain loss function of each category through formula (6) and formula (7) And the local domain total loss function L c :

[0073]

[0074]

[0075] in, is the local domain discriminator for the c-th category, is the cross entropy function of the c-th category;

[0076] Step 7.5: Use Equation (8) and Equation (9) to calculate the A-distance of the global domain discriminator and the local domain discriminator, and get d g d l ,

[0077] d g =2(1-2(Lg )) (8)

[0078]

[0079] Step 7.6: Calculate the dynamic factor κ:

[0080]

[0081] Step 7.7: Combining the above loss functions, calculate the objective function M:

[0082]

[0083] Where: θ q ,θ y ,θ g , They are feature extractors F q , label classifier F y , global domain discriminator F g , local domain discriminator F l Parameters;

[0084] Step 7.8: Set up the stochastic gradient descent SGD optimizer and perform back-propagation gradient descent on the objective function M to solve the feature extractor F in the objective function M. q , label classifier F y , global domain discriminator F g , local domain discriminator F l The current optimal parameters

[0085] Step 7.9: Repeat steps 7.1 to 7.8 until the objective function M converges to the optimal state and the feature extractor F is obtained. q , label classifier F y , global domain discriminator F g , local domain discriminator F l The optimal parameters

[0086] Preferably, in step 4, l1 is 1, c1 is 3, the convolution kernel size k1=3, the number of convolution kernels h1 is 64, 16, and 3 respectively, the step size s1=2, the first two activation functions use the LeakyRule function, and the last layer activation function is the Sigmoid function;

[0087] c2 is 2, the convolution kernel size k2=16, the number of convolution kernels h2 is 128, 64 respectively, the step size s2=2, the maximum pooling layer convolution kernel size k3=4, the step size s3=4, l2 is 2, the activation function uses the LeakyRule function, and the regularization layer is the Dropout layer;

[0088] In step 6, c3 is set to 5, the convolution kernel size of the convolution layer is k4, which takes values ​​of 16, 64, 128, 256, 512, and the step size is s4 = 1. The convolution kernel size of the maximum pooling layer is k5 = 4, and the step size is s5 = 2. The activation function uses the LeakyRule function.

[0089] Therefore, the present invention adopts the above-mentioned cross-domain bridge damage identification method based on deep learning, which has the following beneficial effects:

[0090] 1. The samples generated by the generative adversarial network are diverse and not easily converted into noise. They have the same dimension and distribution as the target samples, which makes up for the defect that deep learning networks cannot provide large amounts of data in real scenarios.

[0091] 2. The dynamic domain adaptation network adopted can dynamically align the joint distribution of source domain and target domain data, and can adapt to various application scenarios of bridge damage identification in reality.

[0092] 3. There is no need to deploy a large number of sensors to obtain multiple position responses. Only a small number of position responses are needed to obtain the damage characteristics of the bridge, which greatly reduces the cost of bridge damage detection.

[0093] The technical solution of the present invention is further described in detail below through the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0094] Figure 1 is a flow chart of the present invention;

[0095] Figure 2 A diagram of a vehicle-bridge finite element model of the present invention;

[0096] Figure 3 This is a diagram of the generative adversarial network structure of the present invention;

[0097] Figure 4 T-SNE diagram of the generated data and target domain data of the present invention;

[0098] Figure 5 This is a diagram of the dynamic domain adversarial adaptation network structure of the present invention;

[0099] Figure 6 This is a diagram of the bridge damage detection results of the present invention. DETAILED DESCRIPTION

[0100] The present invention will be further described below in conjunction with the accompanying drawings. It should be noted that this embodiment is based on the technical solution and provides a detailed implementation method and specific operation process, but the protection scope of the present invention is not limited to this embodiment.

[0101] The cross-domain bridge damage identification method based on deep learning includes the following steps:

[0102] Step 1: Establish a vehicle-bridge finite element model;

[0103] Preferably, step 1 specifically includes the following steps:

[0104] Step 1.1: Determine the bridge parameters, use the bridge parameters to construct the bridge finite element model, and divide the bridge finite element model into units and number them in sequence [1, 2, 3, …, C];

[0105] Preferably, the bridge parameters in step 1.1 include the bridge moment of inertia I0, elastic modulus E0, density per linear meter ρ0 and length L b ;

[0106] Preferably, the moment of inertia I0 in step 1.1 is 1.3901, and the elastic modulus E0 is 3.5×10 10 pa, density per linear meter ρ0=18358 and length L b = 25m, the number of division units C = 10, the bridge finite element model is divided into units and numbered in sequence [1, 2, 3, ..., 10];

[0107] Step 1.2: Determine the two-axle vehicle parameters, build a vehicle finite element model based on the two-axle vehicle parameters, and set the uncertainty vehicle weight m v =m v0 +a0×sin(t0) and the uncertain vehicle speed v=v0+a1×sin(t1), where a0 and a1 are the amplitudes of change, t0, t1∈[0,2π]; the amplitudes of change in step 1.2 are a0=50 and a1=0.1;

[0108] The two-axle vehicle parameters in step 1.2 include the total weight m v0 , two wheelbases d1, d2 and driving speed v0.

[0109] The total weight m in this embodiment v0 = 18000kg, two wheelbases d1 = 1.95m, d2 = 1.05m and driving speed v0 = 10m / s,

[0110] Step 1.3: When the bridge is intact, the vehicle is loaded with an uncertain weight m. v The Newmark-β method is used to calculate the displacement response of the bridge when the uncertain velocity v passes through the bridge. The present invention extracts the displacement responses of unit nodes 1, 5, and 9 as samples, repeats step 1.3n times (n=1250 in this embodiment), and obtains the sample set and build the tag in

[0111] The sampling frequency in step 1.3 is 500Hz;

[0112] Step 1.4: Simulate bridge damage by reducing the element stiffness divided in the bridge finite element model by δ×E0×I0, where δ is the reduction factor; the reduction factor δ in step 1.4 is 0.75.

[0113] Perform stiffness reduction on unit 1 of the bridge division and repeat the method of obtaining the bridge displacement response in step 1.3 n times (n=1250 in this embodiment) to obtain the sample set and build the tag The sample label is the damaged unit number, that is,

[0114] Step 1.5: Repeat step 1.4 until all units of the bridge are divided and the source domain dataset is finally obtained. and label sets

[0115] Step 2: Simulate the real bridge structure by adding uncertainty to the bridge finite element model in the vehicle-bridge finite element model;

[0116] Preferably, step 2 specifically includes the following steps:

[0117] Step 2.1: Add five types of uncertainty:

[0118] (1) The influence of temperature on the bridge is simulated by changing the elastic modulus of the bridge finite element model, that is, E'=E0×(1+ζ1), where ζ1∈(-0.05, 0.05); in this embodiment, ζ1=0.04;

[0119] (2) By setting the vertical stiffness E of the boundary nodes of the bridge finite element model v and angular stiffness E r To simulate the boundary condition of the bridge elastic support; in this embodiment, the vertical stiffness E of the boundary node v =1.95×10 11 N / m and angular stiffness E r =1800N□m;

[0120] (3) The geometric error of bridge modeling is simulated by changing the moment of inertia I0 of the bridge finite element model, that is, I' = I0 × (1 + ζ2), where ζ2∈(-0.03, 0.03), in this embodiment ζ2 = -0.09; at the same time, the material error of bridge modeling is simulated by changing the density ρ0 of the bridge finite element model, that is, ρ' = ρ0 × (1 + ζ3), where ζ3∈(-0.02, 0.02), in this embodiment ζ3 = 0.02;

[0121] (4) The bridge surface roughness is simulated by setting the road surface roughness level A0 of the bridge finite element model; in this embodiment, the road surface roughness level A0 = 16

[0122] After adding the above four uncertainties, the bridge finite element model is used to obtain the bridge displacement response by the method in step 1.3. And add to the obtained bridge displacement response: (5) obeys the mean of 0 and the variance of σ 2 Gaussian distribution noise γ~N(0,σ 2 ),get

[0123] Step 2.2: Repeat steps 1.3 to 1.4 to obtain the displacement response sample sets of the intact bridge and each unit with damage. Finally, the target domain dataset is obtained Since there are only a few samples in reality, in this embodiment, the number of target domain samples is m=250.

[0124] Step 3: Data pre-processing

[0125] The source domain data obtained and target domain Perform normalization and interpolation processing to keep the spatial dimensions of all samples consistent and obtain the processed source domain data and target domain data

[0126] Step 4: Construct the θ and critic D ω Generative adversarial neural network composed of;

[0127] Preferably, in step 4:

[0128] The generator G θ It includes l1 linear layers and c1 transposed convolutional layers in sequence. Each linear layer and each transposed convolutional layer except the last one is followed by a linear activation layer and a normalization layer. The convolution kernel size of the transposed convolutional layer is k1, the number of convolution kernels is h1, and the stride is s1.

[0129] Preferably, in step 4, l1 is 1, c1 is 3, the convolution kernel size k1=3, the number of convolution kernels h1 is 64, 16, and 3 respectively, the step size s1=2, the first two activation functions use the LeakyRule function, and the last layer activation function is the Sigmoid function;

[0130] The critic D ωIt includes c2 convolutional layers and l2 linear layers in sequence. An activation layer, a regularization layer, and a maximum pooling layer are added after each convolutional layer. A regularization layer and an activation layer are added after each linear layer except the last layer. The convolution kernel size of the convolution layer is k2, the number of convolution kernels is h2, and the stride is s2. The convolution kernel size of the maximum pooling layer is k3, and the stride is s3.

[0131] c2 is 2, the convolution kernel size k2=16, the number of convolution kernels h2 is 128, 64 respectively, the step size s2=2, the maximum pooling layer convolution kernel size k3=4, the step size s3=4, l2 is 2, the activation function uses the LeakyRule function, and the regularization layer is the Dropout layer;

[0132] Step 5: Training and generation phase;

[0133] Preferably, step 5 specifically includes the following steps:

[0134] Step 5.1: Training phase

[0135] The input of the generative adversarial network is the target domain data And randomly generate a noise vector z that obeys Gaussian distribution. In this embodiment, the noise vector length is z=400, and the Gaussian distribution obeys the mean of 0 and the variance of 1. Then input the noise vector z into the generator G θ (z) to obtain the generated sample

[0136] Step 5.2: Calculate interpolation data

[0137]

[0138] Among them, ε obeys the uniform distribution U[0,1], and the target domain samples

[0139] Step 5.3: Put the target domain sample x and the interpolation sample and generate samples Input Critic D ω , calculate the loss function L:

[0140]

[0141] Among them, λ is the penalty weight, is the 2-norm after differentiation of the critic;

[0142] Step 5.4: Use the adaptive optimizer Adam to perform back-propagation gradient descent on the loss function L to solve the critic D in the loss function ω The current optimal parameters

[0143] Step 5.5: Randomly generate a batch of p noise sets that follow a Gaussian distribution In this embodiment, p=16, and the loss function F is calculated:

[0144]

[0145] Step 5.6: Use the adaptive optimizer Adam to perform back-propagation gradient descent on the loss function F to solve the generator G in the loss function θ The current optimal parameters The optimizer learning rate in this embodiment is set to 0.0001;

[0146] Step 5.7: Repeat steps 5.3-5.6 until the loss function L and F converge to the optimal state and the critic D is obtained. ω and generator G θ The optimal parameters and

[0147] Step 5.8: Generate Phase

[0148] Randomly generate n noises that obey Gaussian distribution Input the trained generator , we get the expanded target domain sample set In this example, the same number of samples as the source domain data is generated;

[0149] Step 5.9: Repeat steps 5.1 to 5.8 for target domain data As input, we get the trained generators for each category And the expanded datasets for each category Finally, the expanded dataset is obtained

[0150] Step 6: Construct the feature extractor F q , label predictor F y , global domain discriminator F g and local domain discriminator F l Dynamic domain adversarial adaptive network composed of;

[0151] Preferably, in step 6:

[0152] The feature extractor F q It includes c3 convolutional layers in sequence, and each convolutional layer is followed by an activation layer, a regularization layer, and a maximum pooling layer. The convolutional kernel size of the convolutional layer is k4, the stride is s4, and the convolution kernel size of the maximum pooling layer is k5, the stride is s5;

[0153] In step 6, c3 is set to 5, the convolution kernel size of the convolution layer is k4, which takes values ​​of 16, 64, 128, 256, 512, and the step size is s4 = 1. The convolution kernel size of the maximum pooling layer is k5 = 4, and the step size is s5 = 2. The activation function uses the LeakyRule function.

[0154] The label predictor F y It includes l3 linear layers in sequence, where l3=2 in this example;

[0155] The global domain identifier F g The sequence includes l4 linear layers, two linear layers, and an activation layer in the middle. In this example, l4=2, and the activation function uses the LeakyRule function.

[0156] The local domain discriminator F l The sequence includes l5 linear layers, two linear layers, and an activation layer added in the middle. In this example, l5=2, and the activation function uses the LeakyRule function.

[0157] Step 7: Training phase;

[0158] Preferably, step 7 specifically includes the following steps:

[0159] Step 7.1: From the source domain dataset Randomly select a batch containing n s samples Pass through the feature extractor F q and label classifier F y Calculate the label loss function L y :

[0160]

[0161] in, for 's label;

[0162] Step 7.2: Set Domain Labels

[0163] The source data field label is set to Expand the target domain data field label to

[0164] Step 7.3: From the source domain dataset Randomly select a batch containing n s samples and from the expanded target domain dataset Randomly select a batch containing n t samples Pass through the feature extractor F q and the global domain discriminator Fg Calculate the global domain loss function L g :

[0165]

[0166] Among them, L d is the cross entropy function, d i is x i The domain label of

[0167] Step 7.4: Separately from the source domain dataset Randomly select a batch containing n s samples and from the expanded target domain dataset Randomly select a batch containing n t samples In this embodiment, n t =32, and then pass through the feature extractor F q and local domain discriminator F l , and calculate the local domain loss function of each category through formula (6) and formula (7) And the local domain total loss function L c :

[0168]

[0169]

[0170] in, is the local domain discriminator for the c-th category, is the cross entropy function of the c-th category;

[0171] Step 7.5: Use Equation (8) and Equation (9) to calculate the A-distance of the global domain discriminator and the local domain discriminator, and get d g d l ,

[0172] d g =2(1-2(L g )) (8)

[0173]

[0174] Step 7.6: Calculate the dynamic factor κ:

[0175]

[0176] Step 7.7: Combining the above loss functions, calculate the objective function M:

[0177]

[0178] Where: θ q ,θ y ,θ g , They are feature extractors F q , label classifier F y , global domain discriminator F g , local domain discriminator F l Parameters;

[0179] Step 7.8: Set up the stochastic gradient descent SGD optimizer. In this embodiment, the SGD optimizer learning rate is set to 0.0001 and the momentum is set to 0.9. Back-propagation gradient descent is performed on the objective function M to solve the feature extractor F in the objective function M. q , label classifier F y , global domain discriminator F g , local domain discriminator F l The current optimal parameters

[0180] Step 7.9: Repeat steps 7.1 to 7.8 until the objective function M converges to the optimal state and the feature extractor F is obtained. q , label classifier F y , global domain discriminator F g , local domain discriminator F l The optimal parameters

[0181] Step 8: Use training to get the optimal parameters The feature extractor F q and label classifier F y For the target domain dataset To test, the corresponding test results are obtained. It should be noted that the final results are as follows Figure 6 The confusion matrix shown shows that the average accuracy of the detection results is 83.60%.

[0182] Therefore, the present invention adopts the above-mentioned cross-domain bridge damage identification method based on deep learning, which can perform data expansion on the target domain data and generate a large amount of pseudo data that is similar to the target domain data and has the same distribution as the target domain data, thereby providing the required target domain data for the unsupervised domain adaptation method, and then participating in learning, providing a basis for the application of unsupervised domain adaptation to actual scenarios.

[0183] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit the same. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that they can still modify or replace the technical solutions of the present invention with equivalents, and these modifications or equivalent replacements cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A cross-domain bridge damage identification method based on deep learning, characterized by: The following steps are involved: Step 1: Establish a vehicle-bridge finite element model; Step 1 specifically includes the following steps: Step 1.1: Determine the bridge parameters, use the bridge parameters to build the bridge finite element model, and divide the bridge finite element model into units and number them in sequence. ; Step 1.2: Determine the two-axle vehicle parameters, build a vehicle finite element model based on the two-axle vehicle parameters, and set the uncertainty vehicle weight and uncertain speed ,in 、 is the amplitude of change, ; Step 1.3: When the bridge is intact, the vehicle is loaded with an uncertain weight. and uncertain speed Cross the bridge using Method to calculate the displacement response of the bridge Repeat step 1.3 times, get the sample set , and build the label ,in ; Step 1.4: Reduce the element stiffness in the bridge finite element model to simulate bridge damage, where is the reduction factor; Perform stiffness reduction on unit 1 of the bridge division and repeat the method in step 1.3 to obtain the bridge displacement response. times, get the sample set , and build the label , where the sample label is the damaged unit number, i.e. ; Step 1.5: Repeat step 1.4 until all units of the bridge are divided and the source domain dataset is finally obtained. and label sets ; Step 2: Simulate the real bridge structure by adding uncertainty to the bridge finite element model in the vehicle-bridge finite element model; Step 2 specifically includes the following steps: Step 2.1: Add five types of uncertainty: (1) The influence of temperature on the bridge is simulated by changing the elastic modulus of the bridge finite element model, that is, ,in ; (2) By setting the vertical stiffness of the boundary nodes of the bridge finite element model and angular stiffness To simulate the boundary condition form of bridge elastic support; (3) By changing the moment of inertia of the bridge finite element model To simulate the geometric error of bridge modeling, that is ,in ; At the same time, by changing the density of the bridge finite element model To simulate the material error of bridge modeling, that is ,in ; (4) By setting the road surface roughness level of the bridge finite element model To simulate the roughness of the bridge surface; After adding the above four uncertainties, the bridge finite element model is used to obtain the bridge displacement response by the method in step 1.

3. , and add to the resulting bridge displacement response: (5) obeys the mean of 0 and the variance of Gaussian noise ,get ; Step 2.2: Repeat steps 1.3 to 1.4 to obtain the displacement response sample sets of the intact bridge and each unit with damage. , and finally obtain the target domain dataset ; Step 3: Data pre-processing The source domain data obtained and target domain Perform normalization and interpolation processing to keep the spatial dimensions of all samples consistent and obtain the processed source domain data and target domain data ; Step 4: Build the Builder and critics Generative adversarial neural network composed of; Step 5: Training and generation phase; Step 6: Build a feature extractor , label predictor , global domain discriminator and local domain discriminator Dynamic domain adversarial adaptive network composed of; Step 7: Training phase; Step 8: Use training to get the optimal parameters Feature extractor and label classifier For the target domain dataset Perform detection to obtain corresponding test results.

2. The cross-domain bridge damage identification method based on deep learning according to claim 1 is characterized by: The bridge parameters in step 1.1 include the bridge moment of inertia , elastic modulus , density per linear meter and length ; The two-axle vehicle parameters in step 1.2 include the gross weight , two wheelbases 、 and driving speed .

3. The cross-domain bridge damage identification method based on deep learning according to claim 2 is characterized by: The moment of inertia in step 1.1 , elastic modulus , density per linear meter and length , the number of division units , divide the bridge finite element model into units and number them in sequence ; The magnitude of the change in step 1.2 、 ; The sampling frequency in step 1.3 is ; Reduction factor in step 1.4 .

4. The cross-domain bridge damage identification method based on deep learning according to claim 3 is characterized by: In step 4: The generator Include in order linear layers and Transposed convolution layers are added, and each linear layer and each transposed convolution layer except the last layer are followed by a linear activation layer and a normalization layer. The convolution kernel size of the transposed convolution layer is , the number of convolution kernels is , the step size is ; The critic Include in order convolutional layers and Linear layers, activation layer, regularization layer, and maximum pooling layer are added after each convolution layer. Regularization layer and activation layer are added after each linear layer except the last layer. The convolution kernel size of the convolution layer is , the number of convolution kernels is , the step size is , the convolution kernel size of the maximum pooling layer is , the step size is .

5. The cross-domain bridge damage identification method based on deep learning according to claim 4 is characterized by: Step 5 specifically includes the following steps: Step 5.1: Training phase The input of the generative adversarial network is the target domain data and randomly generate a noise vector that follows a Gaussian distribution , and then the noise vector Input to the generator Get the generated sample ; Step 5.2: Calculate interpolation data : (1) in, Uniform distribution , target domain samples ; Step 5.3: Target domain samples , interpolation samples and generate samples Input Criter , calculate the loss function : (2) in, is the penalty weight, is the 2-norm after differentiation of the critic; Step 5.4: Use the adaptive optimizer Adam to adjust the loss function Perform back propagation gradient descent to solve the critic in the loss function The current optimal parameters ; Step 5.5: Randomly generate a batch of A noise set that follows a Gaussian distribution , calculate the loss function : (3) Step 5.6: Use the adaptive optimizer Adam to adjust the loss function Perform back propagation gradient descent to solve the generator in the loss function The current optimal parameters ; Step 5.7: Repeat steps 5.3 to 5.6 until the loss function and When it converges to the optimal value, the critic is obtained and generators The optimal parameters and ; Step 5.8: Generate Phase Random Generation Gaussian noise , input the trained generator , we get the expanded target domain sample set ; Step 5.9: Repeat steps 5.1 to 5.8 for target domain data As input, we get the trained generators for each category , and expanded datasets for each category , and finally get the expanded data set .

6. The cross-domain bridge damage identification method based on deep learning according to claim 5 is characterized by: In step 6: The feature extractor Include in order Convolutional layers are constructed, and activation layers, regularization layers, and maximum pooling layers are added after each convolutional layer. The convolution kernel size of the convolutional layer is , the step size is , the convolution kernel size of the maximum pooling layer is , the step size is ; The label predictor Include in order linear layers; The global domain identifier Include in order linear layers, two linear layers and an activation layer added in the middle; The local domain identifier Include in order Linear layers, two linear layers and an activation layer added in the middle.

7. The cross-domain bridge damage identification method based on deep learning according to claim 6 is characterized by: Step 7 specifically includes the following steps: Step 7.1: From the source domain dataset Randomly select a batch containing samples , through the feature extractor and label classifier Calculate label loss function : (4) in, for 's label; Step 7.2: Set Domain Labels The source data field label is set to , expand the target domain data domain label to ; Step 7.3: From the source domain dataset Randomly select a batch containing samples and from the expanded target domain dataset Randomly select a batch containing samples , through the feature extractor and global domain discriminator Calculate the global domain loss function : (5) in, is the cross entropy function, , for The domain label of Step 7.4: Separately from the source domain dataset Randomly select a batch containing samples and from the expanded target domain dataset Randomly select a batch containing samples , through the feature extractor and local domain discriminator , and calculate the local domain loss function of each category through formula (6) and formula (7) And the local domain total loss function : (6) (7) in, For the local domain discriminator for each category, For the Cross entropy function of categories; Step 7.5: Use Equations (8) and (9) to calculate the global domain discriminator and the local domain discriminator respectively. , respectively , (8) (9) Step 7.6: Calculate the dynamic factor : (10) Step 7.7: Combining the above loss functions, calculate the objective function : (11) in: Feature extractors , label classifier , global domain discriminator , local domain discriminator Parameters; Step 7.8: Set up the stochastic gradient descent SGD optimizer and optimize the objective function Perform back propagation gradient descent to solve the objective function Feature Extractor , label classifier , global domain discriminator , local domain discriminator The current optimal parameters ; Step 7.9: Repeat steps 7.1 to 7.8 until the objective function When converged to the optimal, the feature extractor is obtained , label classifier , global domain discriminator , local domain discriminator The optimal parameters .

8. The cross-domain bridge damage identification method based on deep learning according to claim 7 is characterized by: In step 4 Take 1, Take 3, the convolution kernel size , the number of convolution kernels Take 64, 16, 3 in turn, step size The first two activation functions use the LeakyRule function, and the last activation function is the Sigmoid function; Take 2, the convolution kernel size , the number of convolution kernels Take 128, 64, and step length , the maximum pooling layer convolution kernel size , the step size is , Take 2, the activation function uses the LeakyRule function, and the regularization layer is the Dropout layer; In step 6 Take 5, the convolution kernel size of the convolution layer is , its value is 16, 64, 128, 256, 512, and the step size is , the convolution kernel size of the maximum pooling layer is , the step size is , the activation function uses the LeakyRule function.

Images