Method, device, storage medium and product for identifying damage of cable-supported bridge structure

Abstract

The present invention discloses a method, device, storage medium and product for identifying structural damage of a cable-supported bridge. The identification method includes constructing a sample data set, where the input quantity of each sample data is an image obtained by splicing the time-frequency domain images converted from the acceleration response signals of the main girder and the stay cables respectively; constructing a damage identification model, which includes an input layer, an initial convolutional layer, a max pooling layer, a first combination layer, a second combination layer, a third combination layer, a fourth combination layer, a first global average pooling layer, a first fully connected layer and an output layer connected in sequence. The initial convolutional layer has 6 input channels, and each combination layer includes a convolutional block, an SE attention mechanism and a residual block connected in sequence; using the sample data set to train and verify the damage identification model to obtain a target damage identification model; using the target damage identification model to identify damage to the actual spliced image. The present invention significantly improves the damage identification accuracy.

Description

Technical Field

[0001] The present invention belongs to the technical field of bridge health monitoring, and particularly relates to a method, device, storage medium and product for identifying structural damage of a cable-supported bridge. Background Art

[0002] Traditional bridge health monitoring methods mainly rely on the measurement of physical quantities such as structural stress, strain, and displacement. Although these monitoring methods can provide certain structural state information, they have the following deficiencies:

[0003] (1) High cost and low efficiency of manual inspection: Manual inspection requires a large amount of manpower and time, and the inspection results are easily affected by subjective factors, making it difficult to ensure the comprehensiveness and accuracy of the inspection.

[0004] (2) Limitations of static testing methods: The health state of the structure is mainly evaluated by measuring static responses such as displacement and strain of the structure, but these static response signals often cannot reflect the dynamic characteristics and potential damage of the structure during actual service.

[0005] (3) Limited detection accuracy and sensitivity: The detection accuracy and sensitivity of traditional sensors are limited, making it difficult to capture small structural changes, resulting in limited detection accuracy.

[0006] (4) Complex data processing: The traditional methods generate a large amount of data, and processing and analyzing these data require a large amount of time and human resources.

[0007] (5) Poor real-time performance: Some monitoring methods cannot achieve real-time monitoring, and sudden structural problems cannot be discovered and processed in a timely manner.

[0008] In recent years, with the development of sensor technology and data processing technology, bridge health monitoring methods based on vibration signals have gradually attracted attention. Vibration signals can reflect the dynamic response of a bridge under load, and by analyzing vibration signals, the dynamic characteristics of the structure can be obtained, thereby judging the health status of the structure. However, the existing bridge health monitoring methods based on vibration signals still face the following challenges in practical applications:

[0009] (1) Difficult feature extraction: Bridge vibration signals contain a large amount of noise and complex non-linear features, and traditional vibration feature extraction methods are difficult to effectively extract useful information.

[0010] (2) Inaccurate damage identification: The existing methods have a low accuracy rate for identifying different types of structural damage, especially for the identification of initial minor damage, and the identification effect is not ideal.

[0011] (3) By installing acceleration sensors at key parts of the bridge, the vibration signals of the bridge under external loads can be collected in real time, which can more comprehensively reflect the dynamic characteristics and damage conditions of the bridge structure. However, the analysis method of a single sensor signal has certain limitations and it is difficult to make full use of the advantages of multi-modal data.

[0012] (4) Multi-sensor data fusion problem: Bridge health monitoring usually requires multiple sensors to be arranged. The data collected by sensors at different positions have differences in time-domain and frequency-domain characteristics. How to effectively fuse this multi-modal data is a difficult problem. Summary of the Invention

[0013] The purpose of the present invention is to provide a method, device, storage medium and product for identifying damage to the cable-supported bridge structure, so as to solve the problems that traditional methods are difficult to effectively extract useful information, have poor recognition effect on minor damage, and are difficult to make full use of multi-modal data.

[0014] The present invention solves the above technical problems through the following technical solutions: A method for identifying damage to a cable-supported bridge structure includes:

[0015] Construct a sample data set; wherein, each sample data in the sample data set includes an input quantity and an output quantity, the input quantity is a spliced image, and the output quantity is the damage condition of the spliced image; the spliced image is obtained by splicing the time-frequency domain images converted from the acceleration response signals of the main girder and the cable-stayed cable respectively;

[0016] Construct a damage identification model; wherein, the damage identification model includes an input layer, an initial convolutional layer, a max pooling layer, a first combination layer, a second combination layer, a third combination layer, a fourth combination layer, a first global average pooling layer, a first fully connected layer and an output layer connected in sequence; the initial convolutional layer has 6 input channels; each combination layer includes a convolutional block, an SE attention mechanism and a residual block connected in sequence;

[0017] Use the sample data set to train and verify the damage identification model to obtain a target damage identification model;

[0018] Obtain the actual acceleration response signals of the main girder and the cable-stayed cable;

[0019] Convert the actual acceleration response signals of the main girder and the cable-stayed cable into time-frequency domain images respectively, and then splice the time-frequency domain images of the main girder and the cable-stayed cable to obtain an actual spliced image;

[0020] Use the target damage identification model to identify damage to the actual spliced image.

[0021] Further, the sample data set is constructed through on-site tests, specifically including:

[0022] Build a test model of a cable - stayed bridge, set a first sensor on the main girder of the cable - stayed bridge test model, and set a second sensor on the stay cables of the cable - stayed bridge test model;

[0023] Simulate different damage conditions on the cable - stayed bridge test model, use the first sensor to collect the acceleration response signals of the main girder under different damage conditions, and use the second sensor to collect the acceleration response signals of the stay cables under different damage conditions;

[0024] Pre - process the acceleration response signals of the main girder and stay cables under different damage conditions;

[0025] Convert the pre - processed acceleration response signals of the main girder and stay cables into time - frequency domain images;

[0026] Stitch the time - frequency domain images of the main girder and stay cables under the same damage condition to obtain a stitched image;

[0027] Construct the sample data set according to the stitched image and its corresponding damage condition.

[0028] Further, the damage conditions include damage location, vehicle speed, vehicle mass, and damage degree; the vehicle speed is 0.5 m / s, 0.75 m / s, or 1 m / s, the vehicle mass is 5 kg, 10 kg, or 20 kg; the stay cables are numbered in order from right to left;

[0029] The first damage condition: no damage, combinations of different vehicle speeds and different vehicle masses;

[0030] The second damage condition: the first stay cable is damaged, combinations of different vehicle speeds and different vehicle masses, and the damage degree is 10%, 20%, or 50%;

[0031] The third damage condition: the third stay cable is damaged, combinations of different vehicle speeds and different vehicle masses, and the damage degree is 10%, 20%, or 50%;

[0032] The fourth damage condition: the sixth stay cable is damaged, combinations of different vehicle speeds and different vehicle masses, and the damage degree is 10%, 20%, or 50%;

[0033] The fifth damage condition: the first and fourth stay cables are damaged, combinations of different vehicle speeds and different vehicle masses, and the damage degree is 10%, 20%, or 50%;

[0034] The sixth damage condition: the fifth and sixth stay cables are damaged, combinations of different vehicle speeds and different vehicle masses, and the damage degree is 10%, 20%, or 50%.

[0035] Further, the S - transform is used to achieve the conversion of time - frequency domain images, and the specific formula is:

[0036]

[0037] Among them, S(τ,f) represents the time - frequency domain image, x(t) represents the acceleration response signal, ω(τ - t,f) represents the Gaussian window function, t represents time, τ represents the position of the Gaussian window on the time axis t, f represents the frequency of the acceleration response signal, and i represents the imaginary part symbol.

[0038] Further, the convolution block includes a first convolution layer, a first activation layer, a second convolution layer, a second activation layer, a third convolution layer, and a batch normalization layer connected in sequence;

[0039] The number of residual blocks in the first combination layer is 2, the number of residual blocks in the second combination layer is 3, the number of residual blocks in the third combination layer is 5, and the number of residual blocks in the fourth combination layer is 2; each of the residual blocks includes a fourth convolution layer, a third activation layer, a fifth convolution layer, a fourth activation layer, and a sixth convolution layer connected in sequence.

[0040] Further, the SE attention mechanism includes a second global average pooling layer, a second fully - connected layer, a third fully - connected layer, and a re - weighting layer connected in sequence.

[0041] Further, training the damage identification model using the sample data set includes:

[0042] Using the initial convolution layer to extract features from the spliced image received by the input layer to obtain a first feature map;

[0043] Using the max - pooling layer to perform max - pooling on the first feature map to obtain a second feature map;

[0044] Using the first combination layer to extract features from the second feature map to obtain a third feature map;

[0045] Using the second combination layer to extract features from the third feature map to obtain a fourth feature map;

[0046] Using the third combination layer to extract features from the fourth feature map to obtain a fifth feature map;

[0047] Using the fourth combination layer to extract features from the fifth feature map to obtain a sixth feature map;

[0048] Using the first global average pooling layer to perform global average pooling on the sixth feature map to obtain a single feature;

[0049] Flatten the single feature using the first fully connected layer and map it to the classification space to obtain the predicted damage category;

[0050] Calculate the loss value based on the predicted damage category and the damage condition of the spliced image, and adjust the parameters of the damage recognition model according to the loss value until the training times are reached or the loss value is less than the set loss accuracy.

[0051] Based on the same concept, the present invention also provides an electronic device, including a memory, a processor, and a computer program / instructions stored on the memory. The processor executes the computer program / instructions to implement the cable-supported bridge structure damage recognition method as described above.

[0052] Based on the same concept, the present invention also provides a computer-readable storage medium, on which a computer program / instructions are stored. When the computer program / instructions are executed by a processor, the cable-supported bridge structure damage recognition method as described above is implemented.

[0053] Based on the same concept, the present invention also provides a computer program product, including a computer program / instructions. When the computer program / instructions are executed by a processor, the cable-supported bridge structure damage recognition method as described above is implemented.

[0054] Advantageous Effects

[0055] Compared with the prior art, the advantages of the present invention are as follows:

[0056] The present invention converts the acceleration response signals of the main girder and the stay cables into time-frequency domain images, fully retaining the time dependence and non-linear characteristics of the acceleration response signals, which is beneficial for the damage recognition model to more accurately capture and analyze the minute changes and early damage of the bridge structure; splicing the time-frequency domain images of the main girder and the stay cables to form a spliced image with high-level features and low-level details, which can provide richer features, further facilitating the damage recognition model to capture more feature information, and thus contributing to the recognition of minute damage; improving the initial convolutional layer of the traditional residual neural network model, enabling the model to reason about the spliced image; integrating the SE attention mechanism in each combination layer, improving the learning ability of the damage recognition model, alleviating the problem of gradient disappearance, and improving the recognition accuracy of bridge structure damage. Description of the Drawings

[0057] In order to more clearly illustrate the technical solutions of the present invention, the drawings required for the description of the embodiments will be briefly introduced below. Obviously, the drawings in the following description are only one embodiment of the present invention. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0058] Figure 1 It is a flowchart of the method for identifying structural damage of a cable-supported bridge in an embodiment of the present invention;

[0059] Figure 2 It is a schematic diagram of the cable numbering of the cable-supported bridge test model in an embodiment of the present invention;

[0060] Figure 3 It is a schematic diagram of converting the acceleration response signal of the main girder into a time-frequency domain image in an embodiment of the present invention; among them, the color scale numbers represent the amplitudes of the points in the image;

[0061] Figure 4 It is a schematic diagram of converting the acceleration response signal of the cable into a time-frequency domain image in an embodiment of the present invention;

[0062] Figure 5 It is an architecture diagram of the improved residual neural network model in an embodiment of the present invention;

[0063] Figure 6 It is a 6-channel image tensor in an embodiment of the present invention;

[0064] Figure 7 It is a 64-channel output feature map in an embodiment of the present invention;

[0065] Figure 8 It is a heat map of feature extraction visualization of the last residual block in the second combination layer, the third combination layer, and the fourth combination layer in an embodiment of the present invention;

[0066] Figure 9 It is an identification accuracy curve of the identification method of the present invention (stitched image + improved residual neural network model) and the traditional identification method (time-frequency domain image of the main girder + traditional residual neural network model) in an embodiment of the present invention;

[0067] Figure 10 It is a loss value curve of the identification method of the present invention (stitched image + improved residual neural network model) and the traditional identification method (time-frequency domain image of the main girder + traditional residual neural network model) in an embodiment of the present invention;

[0068] Figure 11 It is a visualization diagram of the damage category of the time-frequency domain image of the main girder + traditional residual neural network model in an embodiment of the present invention;

[0069] Figure 12 It is a visualization diagram of the damage category of the time-frequency domain image of the cable + traditional residual neural network model in an embodiment of the present invention

[0070] Figure 13 It is a visualization diagram of the damage category of the stitched image + improved residual neural network model in an embodiment of the present invention;

[0071] Figure 14It is a visualization diagram of the damage identification accuracy of the identification method of the present invention and the traditional identification method in the embodiments of the present invention. Detailed implementation manners

[0072] Combined with the accompanying drawings in the embodiments of the present invention, the technical solutions in the present invention will be clearly and completely described. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without making creative efforts belong to the scope of protection of the present invention.

[0073] The technical solutions of the present application will be described in detail below with specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be repeated in some embodiments.

[0074] As Figure 1 shown, a method for identifying damage to a cable-supported bridge structure provided by an embodiment of the present invention includes the following steps:

[0075] Step 1: Construct a sample data set.

[0076] The sample data set includes a plurality of sample data, each sample data includes an input quantity and an output quantity, the input quantity is a spliced image, and the output quantity is the damage condition of the spliced image; wherein, the spliced image is obtained by splicing the time-frequency domain images converted from the acceleration response signals of the main girder and the acceleration response signals of the stay cables respectively.

[0077] The sample data set can be constructed by a finite element simulation method or by a field test. In the specific implementation manner of the present invention, the sample data set is constructed by a field test, which specifically includes:

[0078] Step 1.1: Build a cable-supported bridge test model, set a first sensor on the main girder of the cable-supported bridge test model, and set a second sensor on the stay cables of the cable-supported bridge test model.

[0079] As Figure 2 shown, the cable-supported bridge test model includes a main girder and stay cables. The stay cables are numbered in sequence from right to left, and the numbers of the 8 stay cables are S1 to S8 in turn. A first sensor is set at the mid-span of the main girder, and a second sensor is set on any one of the stay cables. In this embodiment, the second sensor is set on the fifth stay cable S5. The first sensor collects the acceleration response signal of the main girder, and the second sensor collects the acceleration response signal of the stay cable. During the test, the installation positions of the first sensor and the second sensor remain unchanged.

[0080] In this embodiment, the first sensor and the second sensor are piezoelectric acceleration sensors, with a sensitivity of 100 mV / g, a frequency range of 0.5 - 7000 Hz, and a measurement range of ±50 g.

[0081] The present invention uses a dynamic data acquisition device to obtain the acceleration response signals collected by the first sensor and the second sensor under different damage conditions, exports the data from the dynamic data acquisition device, and saves it in an excel table or other data formats for subsequent data processing and analysis. The dynamic data acquisition device in this embodiment selects a DH3822 strain tester, on which a DHDAS dynamic signal acquisition and analysis system is installed, and the sampling rate is 500 Hz.

[0082] Step 1.2: Simulate different damage conditions on the cable-supported bridge test model, use the first sensor to collect the acceleration response signals of the main girder under different damage conditions, and use the second sensor to collect the acceleration response signals of the stay cables under different damage conditions.

[0083] Table 1 Different damage conditions

[0084] Operating condition serial number Damage location Vehicle speed Vehicle weight Damage degree Operating condition 1 Healthy 0.5 m / s, 0.75 m / s, 1 m / s 5 kg, 10 kg, 20 kg / Operating condition 2 S1 0.5 m / s, 0.75 m / s, 1 m / s 5 kg, 10 kg, 20 kg 10％,20％,50％ Operating condition 3 S3 0.5 m / s, 0.75 m / s, 1 m / s 5 kg, 10 kg, 20 kg 10％,20％,50％ Operating condition 4 S6 0.5 m / s, 0.75 m / s, 1 m / s 5 kg, 10 kg, 20 kg 10％,20％,50％ Operating condition 5 S1, S4 0.5 m / s, 0.75 m / s, 1 m / s 5 kg, 10 kg, 20 kg 10％,20％,50％ Operating condition 6 S5, S6 0.5 m / s, 0.75 m / s, 1 m / s 5 kg, 10 kg, 20 kg 10％,20％,50％

[0085] As shown in Table 1, the damage conditions in this embodiment include the damage location, vehicle speed, vehicle mass, and damage degree. The traveling vehicle serves as the external load of the bridge. The vehicle is driven by a motor, and the vehicle speed is controlled by controlling the rotation speed of the motor. The vehicle speed is 0.5 m / s, 0.75 m / s, or 1 m / s, and the vehicle mass is 5 kg, 10 kg, or 20 kg. The first damage condition: no damage, combinations of different vehicle speeds and different vehicle masses; the second damage condition: the first stay cable S1 is damaged, combinations of different vehicle speeds and different vehicle masses, and the damage degree is 10%, 20%, or 50%; the third damage condition: the third stay cable S3 is damaged, combinations of different vehicle speeds and different vehicle masses, and the damage degree is 10%, 20%, or 50%; the fourth damage condition: the sixth stay cable S6 is damaged, combinations of different vehicle speeds and different vehicle masses, and the damage degree is 10%, 20%, or 50%; the fifth damage condition: the first and fourth stay cables S1 / S4 are damaged, combinations of different vehicle speeds and different vehicle masses, and the damage degree is 10%, 20%, or 50%; the sixth damage condition: the fifth and sixth stay cables S5 / S6 are damaged, combinations of different vehicle speeds and different vehicle masses, and the damage degree is 10%, 20%, or 50%.

[0086] Each damage condition forms multiple sub-conditions through different combinations of vehicle speed, vehicle mass, and damage degree. Taking the second damage condition as an example, the sub-conditions formed by the second damage condition include:

[0087] (1) S1 damage, vehicle speed is 0.5 m / s, vehicle weight is 5 kg, damage degree is 10%;

[0088] (2) S1 damage, vehicle speed is 0.75 m / s, vehicle weight is 5 kg, damage degree is 10%;

[0089] (3) S1 damage, vehicle speed is 1 m / s, vehicle weight is 5 kg, damage degree is 10%;

[0090] (4) S1 damage, vehicle speed is 0.5 m / s, vehicle weight is 10 kg, damage degree is 10%;

[0091] (5) S1 damage, vehicle speed is 0.5 m / s, vehicle weight is 20 kg, damage degree is 10%;

[0092] (6) S1 damage, vehicle speed is 0.75 m / s, vehicle weight is 10 kg, damage degree is 10%;

[0093] (7) S1 damage, vehicle speed is 0.75 m / s, vehicle weight is 20 kg, damage degree is 10%;

[0094] (8) S1 damage, vehicle speed is 1 m / s, vehicle weight is 10 kg, damage degree is 20%;

[0095] (9) S1 damage, vehicle speed is 1 m / s, vehicle weight is 20 kg, damage degree is 20%;

[0096] (10) S1 damage, vehicle speed is 0.5 m / s, vehicle weight is 5 kg, damage degree is 20%;

[0097] (11) S1 damage, vehicle speed is 0.75 m / s, vehicle weight is 5 kg, damage degree is 20%;

[0098] (12) S1 damage, vehicle speed is 1 m / s, vehicle weight is 5 kg, damage degree is 20%;

[0099] (13) S1 damage, vehicle speed is 0.5 m / s, vehicle weight is 10 kg, damage degree is 20%;

[0100] (14) S1 damage, vehicle speed is 0.5 m / s, vehicle weight is 20 kg, damage degree is 20%;

[0101] (15) S1 damage, vehicle speed is 0.75 m / s, vehicle weight is 10 kg, damage degree is 20%;

[0102] (16) S1 damage, vehicle speed is 0.75 m / s, vehicle weight is 20 kg, damage degree is 20%;

[0103] (17) S1 damage, vehicle speed is 1 m / s, vehicle weight is 10 kg, damage degree is 20%;

[0104] (18) S1 damage, vehicle speed is 1 m / s, vehicle weight is 20 kg, damage degree is 20%;

[0105] (19) S1 damage, vehicle speed is 0.5 m / s, vehicle weight is 5 kg, damage degree is 50%;

[0106] (20) S1 damage, vehicle speed is 0.75 m / s, vehicle weight is 5 kg, damage degree is 50%;

[0107] (21) S1 damage, vehicle speed is 1 m / s, vehicle weight is 5 kg, damage degree is 50%;

[0108] (22) S1 damage, vehicle speed is 0.5 m / s, vehicle weight is 10 kg, damage degree is 50%;

[0109] (23) S1 damage, vehicle speed is 0.5 m / s, vehicle weight is 20 kg, damage degree is 50%;

[0110] (24) S1 damage, vehicle speed is 0.75 m / s, vehicle weight is 10 kg, damage degree is 50%;

[0111] (25) S1 damage, vehicle speed is 0.75 m / s, vehicle weight is 20 kg, damage degree is 50%;

[0112] (26) S1 damage, vehicle speed is 1 m / s, vehicle weight is 10 kg, damage degree is 50%;

[0113] (27) S1 damage, vehicle speed is 1 m / s, vehicle weight is 20 kg, damage degree is 50%.

[0114] Multiple acceleration response signals can be obtained under each damage condition through combinations of different vehicle speeds, vehicle weights, and damage degrees.

[0115] Step 1.3: Preprocess the acceleration response signals of the main girder and stay cables under different damage conditions.

[0116] In the specific implementation manner of the present invention, the preprocessing includes normalization processing, calculating the maximum and minimum values of the acceleration response signals, and normalizing the amplitude of each acceleration response signal so that it is mapped to the interval [-1, 1] to ensure consistency.

[0117] Step 1.4: Convert the preprocessed acceleration response signals of the main girder and stay cables into time-frequency domain images to obtain the time-frequency domain images of the main girder and stay cables.

[0118] In the specific implementation manner of the present invention, the acceleration response signals of the main girder and the stay cables after preprocessing are transformed into time-frequency domain images by using the S transform. The specific formula is as follows:

[0119]

[0120]

[0121] where S(τ, f) represents the time-frequency domain image, x(t) represents the acceleration response signal, ω(τ - t, f) represents the Gaussian window function, t represents time, τ represents the position of the Gaussian window on the time axis t, f represents the frequency of the acceleration response signal, and i represents the imaginary part symbol. The time-frequency domain images of the main girder and the stay cables are as shown in Figure 3 and Figure 4 shown.

[0122] Each point of the time-frequency domain image corresponds to the intensity value of the acceleration response signal at a specific time and frequency. The S transform converts the one-dimensional acceleration response signal into a two-dimensional image, retaining the time dependence and non-linear characteristics of the acceleration response signal. While retaining the characteristic information in the original time series, it better displays the time-frequency characteristics of the signal, which is beneficial for the damage identification model to more accurately capture and analyze the tiny changes and early damage of the bridge structure, and significantly improves the accuracy and sensitivity of monitoring.

[0123] Step 1.5: Stitch the time-frequency domain images of the main girder and the stay cables under the same damage condition to obtain a stitched image.

[0124] Use the PLL library to load the time-frequency domain images of the main girder and the stay cables under the same damage condition, and adjust their sizes to 224×224. Through the function torchvision.transforms.ToTensor(), the time-frequency domain images of the main girder and the stay cables after size adjustment are converted into tensor format, and the tensor size is C×224×224, where C represents the number of image channels and C is equal to 3. The time-frequency domain images of the main girder and the stay cables in tensor format are stitched in the channel dimension to form a stitched image with 6 channels in tensor format.

[0125] The specific process of stitching is as follows: Stack two images (the time-frequency domain images of the main girder and the stay cables) with a size of 3×224×224 along the channel dimension through the torch.concat function to form a stitched image with 6 channels, and the size of the stitched image is 6×224×224.

[0126] Step 1.6: Construct a sample data set according to the stitched image and its corresponding damage condition.

[0127] For each damage condition, 200 acceleration response signals of the main girder and 200 acceleration response signals of the stay cables are collected. A total of 2,400 sample data are formed for the 6 conditions, that is, the sample data set contains 2,400 sample data. The sample data set is divided into a training data set, a validation data set, and a test data set according to a ratio of 6:2:2. The damage identification model is trained using the training data set, and the trained damage identification model is verified using the validation data set to obtain the target damage identification model.

[0128] Step 2: Construct a damage identification model;

[0129] In a specific embodiment of the present invention, the damage identification model adopts an improved residual neural network model, as Figure 5 shown. The improved residual neural network model includes an input layer, an initial convolutional layer new_conv1, a max pooling layer Maxpool1, a first combination layer Stage1, a second combination layer Stage2, a third combination layer Stage3, a fourth combination layer Stage4, a first global average pooling layer Avgpool, a first fully connected layer FC, and an output layer connected in sequence; the initial convolutional layer new_conv1 has 6 input channels; each combination layer (i.e., the first combination layer Stage1, the second combination layer Stage2, the third combination layer Stage3, the fourth combination layer Stage4) includes a convolutional block Conv Block, an SE attention mechanism, and an identity block Identity Block connected in sequence.

[0130] When constructing the damage identification model, first load a traditional residual neural network model pre-trained on a large-scale data set, and extract the parameters of the initial convolutional layer conv1 of the traditional residual neural network model (i.e., transfer learning TL). The initial convolutional layer conv1 has 3 input channels and outputs 64 feature maps. Since the spliced image obtained by splicing the time-frequency domain images of the main girder and the stay cables is a tensor with 6 input channels (as Figure 6As shown, the initial convolutional layer conv1 of the residual neural network model is not applicable to the spliced image. In the present invention, the initial convolutional layer conv1 is improved, and a new initial convolutional layer new_conv1 is recreated. The initial convolutional layer new_conv1 has 6 input channels and outputs 64 feature maps. The parameters of the initial convolutional layer conv1 are copied to the first 3 channels and the last 3 channels of the initial convolutional layer new_conv1, that is, new_conv1.weight[:,:3,:,:]=conv1.weight, new_conv1.weight[:,3:6,:,:]=conv1.weight, where weight represents the weight parameter, and : means assigning values to all spatial positions and convolutional kernels. The gradient is not calculated during the copying process to ensure that the model parameters are not updated during the weight parameter copying process.

[0131] The parameters of the initial convolutional layer new_conv1 are as follows:

[0132] new_conv1 = nn.Conv2d(in_channels = 6, out_channels = 64, kernel_size = 7, stride = 2, padding = 3) (3)

[0133] Among them, nn.Conv2d is a class for defining a two-dimensional convolutional layer in PyTorch; n_channels = 6 indicates that the number of input channels is 6, that is, the number of channels of the spliced image is 6; out_channels = 64 indicates that the number of output channels is 64, that is, the number of channels of the feature map obtained after processing by the initial convolutional layer new_conv1 is 64; kernel_size = 7 indicates that the size of the convolutional kernel is 7×7; stride = 2 indicates that the stride is 2, that is, the step size of the convolutional kernel sliding on the spliced image each time is 2, that is, moving 2 pixels each time; padding = 3 indicates that the number of padding pixels is 3 pixels to prevent the image size from shrinking too much after the convolution operation.

[0134] Replacing the initial convolutional layer conv1 of the residual neural network model with the initial convolutional layer new_conv1 can process the spliced image. The initial convolutional layer new_conv1 uses 64 convolutional kernels of size 7×7, with a stride of 2 and a padding of 3, and the size of the output feature map is (64, 112, 112). The 6-channel spliced image is input into the initial convolutional layer new_conv1 for forward propagation to obtain the output feature map result with 64 channels (i.e., the first feature map), as Figure 7 shown.

[0135] The max - pooling layer Maxpool1 performs a max - pooling operation on the first feature map output by the initial convolutional layer new_conv1 to obtain a second feature map. The size of the second feature map is (64, 56, 56). The pooling kernel size of the max - pooling layer Maxpool1 is 3×3, the stride is 2, and the boundary padding is 0.

[0136] Each combined layer of the traditional residual neural network model only includes a convolutional block Conv Block and an identity block IdentityBlock. The present invention improves the combined layer by inserting an SE attention mechanism between the convolutional block Conv Block and the identity block Identity Block, that is, each combined layer includes a convolutional block Conv Block, an SE attention mechanism, and an identity block Identity Block connected in sequence.

[0137] In this embodiment, each convolutional block Conv Block includes a first convolutional layer, a first activation layer, a second convolutional layer, a second activation layer, a third convolutional layer, and a batch normalization layer connected in sequence. The specific formula of the convolutional block Conv Block is:

[0138] Xout = BN(Conv1×1(ReLU(Conv3×3(ReLU(Conv1×1(Xin)))))) (4)

[0139] In formula (4), Xin represents the input feature map of the convolutional block Conv Block, Xout represents the output feature map of the convolutional block ConvBlock, Conv1×1 represents that the convolutional kernels of the first convolutional layer and the third convolutional layer are 1×1, Conv3×3 represents that the convolutional kernel of the second convolutional layer is 3×3, ReLU represents the ReLU activation function of the first activation layer and the second activation layer, and BN represents the batch normalization layer.

[0140] Each identity block Identity Block includes a fourth convolutional layer, a third activation layer, a fifth convolutional layer, a fourth activation layer, and a sixth convolutional layer connected in sequence. The specific formula of the identity block Identity Block is:

[0141] Xout = Conv1×1(ReLU(Conv3×3(ReLU(Conv1×1(Xin))))) (5)

[0142] In formula (5), Xin represents the input feature map of the residual block Identity Block, Xout represents the output feature map of the residual block Identity Block, Conv1×1 represents that the convolution kernels of the fourth and sixth convolutional layers are 1×1, Conv3×3 represents that the convolution kernel of the fifth convolutional layer is 3×3, and ReLU represents the ReLU activation function of the third and fourth activation layers.

[0143] The SE attention mechanism, namely the Squeeze-and-Excitation attention mechanism, is an attention mechanism widely used in deep learning, aiming to improve the performance and accuracy of the model. The SE attention mechanism includes a second global average pooling layer, a second fully connected layer, a third fully connected layer, and a reweighting layer connected in sequence. The second global average pooling layer is used to perform global average pooling on the feature map input to the SE attention mechanism, compressing the feature map from (C, H, W) to (C, 1, 1). The specific formula is:

[0144]

[0145] where C represents the number of channels of the feature map, H represents the height of the feature map, W represents the width of the feature map, x c,i,j represents the feature value at the position (i, j) of the c-th channel in the feature map input to the SE attention mechanism, and z c represents the compressed feature of the c-th channel, where c = 1, 2,..., C.

[0146] The second fully connected layer (with ReLU activation function) is used to process the feature map output by the second global average pooling layer. The specific formula is:

[0147] e c = ReLU(W 1 ·z c ) (7)

[0148] where W 1 represents the weight matrix of the second fully connected layer, and e c represents the ReLU activation feature of the c-th channel.

[0149] The third fully connected layer (with Sigmoid activation function) is used to process the feature map output by the second fully connected layer. The specific formula is:

[0150] s c = Sigmoid(W 2 ·e c ) (8)

[0151] where W 2 represents the weight matrix of the third fully connected layer, and sc Denotes the Sigmoid activation feature of the c-th channel.

[0152] Use a reweighting layer to multiply the Sigmoid activation feature s of each channel c by the feature map input to the SE attention mechanism to obtain a weighted feature. The specific formula is:

[0153]

[0154] where, Denotes the weighted feature of the c-th channel.

[0155] In this embodiment, the number of residual blocks Identity Block in the first combination layer is 2, the number of residual blocks Identity Block in the second combination layer is 3, the number of residual blocks Identity Block in the third combination layer is 5, and the number of residual blocks Identity Block in the fourth combination layer is 2. Each combination layer includes a convolutional block Conv Block and multiple residual blocks Identity Block. The output feature map of each combination layer serves as the input feature map of the next combination layer and is downsampled according to a specific stride to achieve layer-by-layer feature extraction. The specific formulas for each combination layer are:

[0156] Xstage1 = IdentityBlock(IdentityBlock(SEModule(ConvBlock(Xin)))) (10)

[0157] Xstage2 = IdentityBlock(IdentityBlock(IdentityBlock(SEModule(ConvBlock(Xstage1))))) (11)

[0158] Xstage3 = IdentityBlock(IdentityBlock(IdentityBlock(IdentityBlock(IdentityBlock(SEModule(ConvBlock(Xstage2)))))) (12)

[0159] Xstage4 = IdentityBlock(IdentityBlock(SEModule(ConvBlock(Xstage3)))) (13)

[0160] Among them, Xin represents the feature map input to the first combination layer stage1, Xstage1 represents the output feature map of the first combination layer stage1, Xstage2 represents the output feature map of the second combination layer stage2, Xstage3 represents the output feature map of the third combination layer stage3, and Xstage4 represents the output feature map of the fourth combination layer stage4 (i.e., the sixth feature map). Figure 8 The heatmap visualization of feature extraction of the last residual block IdentityBlock in the second combination layer, the third combination layer, and the fourth combination layer is shown, which can intuitively display the areas that the model focuses on.

[0161] The first global average pooling layer is used to perform global average pooling on the output feature map of the fourth combination layer to obtain a single feature; the first fully connected layer is used to flatten the single feature and map it to the classification space to obtain the predicted damage category.

[0162] Step 3: Use the sample data set to train and validate the damage recognition model to obtain the target damage recognition model.

[0163] In the specific implementation manner of the present invention, training the damage recognition model (i.e., the improved residual neural network model) using the sample data set includes:

[0164] Step 3.1: Use the initial convolutional layer new_conv1 to extract features from the spliced image received by the input layer to obtain the first feature map;

[0165] Step 3.2: Use the max pooling layer Maxpool1 to perform max pooling on the first feature map to obtain the second feature map;

[0166] Step 3.3: Use the first combination layer Stage1 to extract features from the second feature map to obtain the third feature map;

[0167] Step 3.4: Use the second combination layer Stage2 to extract features from the third feature map to obtain the fourth feature map;

[0168] Step 3.5: Use the third combination layer Stage3 to extract features from the fourth feature map to obtain the fifth feature map;

[0169] Step 3.6: Use the fourth combination layer Stage4 to extract features from the fifth feature map to obtain the sixth feature map;

[0170] Step 3.7: Use the first global average pooling layer Avgpool to perform global average pooling on the sixth feature map to obtain a single feature;

[0171] Step 3.8: Flatten the single feature using the first fully connected layer FC and map it to the classification space to obtain the predicted damage category;

[0172] Step 3.9: Calculate the loss value according to the predicted damage category and the damage condition of the spliced image, and adjust the parameters of the damage recognition model according to the loss value until the number of training times is reached or the loss value is less than the set loss accuracy.

[0173] In this embodiment, cross-entropy is used as the loss function to measure the difference between the predicted damage category and the damage condition of the spliced image. In the embodiment of the present invention, there are 6 damage categories. For each sample data, the damage recognition model will generate a prediction vector to represent the predicted probabilities of 6 damage categories. The damage condition of the spliced image is represented by the actual vector y. Only the element corresponding to the damage condition of the spliced image in the actual vector y is 1, and the rest of the elements are 0. The specific formula of the loss function is:

[0174]

[0175] where L represents the cross-entropy loss, y i represents the actual label value (0 or 1) of the i-th category, represents the predicted probability of the i-th category, and N represents the number of damage categories. Calculate the cross-entropy loss of each sample data in each batch, and then calculate the derivative of the cross-entropy loss with respect to the model parameters. The cross-entropy loss function quantifies the error of the model by calculating the gap between the actual label and the predicted probability. The smaller the loss value, the closer the predicted probability of the model is to the actual label value.

[0176] Step 4: Obtain the actual acceleration response signals of the main girder and the stay cables.

[0177] Step 5: Convert the actual acceleration response signals of the main girder and the stay cables into time-frequency domain images respectively, and then splice the time-frequency domain images of the main girder and the stay cables to obtain the actual spliced image.

[0178] Step 5 is similar to Steps 1.4 and 1.5.

[0179] Step 6: Use the target damage recognition model to identify the damage of the actual spliced image.

[0180] To verify the accuracy of the target damage recognition model of the present invention, the damage recognition accuracy of the method of the present invention is compared with the damage recognition accuracy of the traditional residual neural network model. The comparison results are shown in Table 2 and Figure 9 、 Figure 10 as shown. It can be seen from Table 2, Figure 9 and Figure 10 that the present invention has high damage recognition accuracy and low loss value.

[0181] Table 2 Comparison of damage recognition accuracy for different input images

[0182]

[0183] To visualize the damage categories of time-frequency domain images, t-distributed stochastic neighbor embedding (t-SNE) is used to visualize the damage categories of the time-frequency domain images of the main girder, the time-frequency domain images of the stay cables, and the spliced images, as Figure 11 、 Figure 12 、 Figure 13 shown. T-distributed stochastic neighbor embedding (t-SNE) is a non-linear dimensionality reduction technique used to map high-dimensional data into a two-dimensional space for easy visualization. Among them, the abscissa (X-axis) represents the first new coordinate axis after dimensionality reduction, and the ordinate (Y-axis) represents the second new coordinate axis after dimensionality reduction. Figures 11 - 13 The solid dots in represent the damage categories corresponding to the damage conditions. That is, the abscissa (X-axis) represents the position of the data in the first main direction in the low-dimensional embedding space, which can also be interpreted as the distribution of the data on the first significant feature; the ordinate (Y-axis) represents the position of the data in the second main direction in the low-dimensional embedding space, which can also be interpreted as the distribution of the data on another significant feature. The abscissa and ordinate themselves have no specific physical meaning or unit, and these coordinates are generated when the high-dimensional data points are embedded into the low-dimensional space by the t-SNE algorithm.

[0184] As can be seen from Figure 13 , although the solid dots of the damage categories slightly overlap, the method of the present invention can correctly identify the clustering clusters of the same damage category, indicating that the method of the present invention has good damage recognition ability. In contrast, there are varying degrees of confusion in the damage categories of the time-frequency domain images of the main girder and the time-frequency domain images of the stay cables. The boundaries of the clustering clusters are blurred, and the data of different damage categories are relatively close or even overlap. Therefore, the damage recognition ability of the traditional recognition method is poor. The traditional recognition method is prone to incorrect damage categories, thereby reducing the damage recognition accuracy. The visualization of damage categories provides an intuitive result of the distribution of damage recognition results. The results show that the method of the present invention has superior damage recognition performance and can effectively distinguish various types of data.

[0185] Another form of Table 2 is as Figure 14 shown. It can be intuitively seen from Figure 14 the high damage recognition accuracy of the method of the present invention. Compared with the traditional recognition method, the present invention significantly improves the damage recognition accuracy, indicating the effectiveness of introducing the SE attention mechanism into each combined layer and splicing the time-frequency domain images of the main girder and the stay cables.

[0186] Example 2

[0187] An embodiment of the present invention further provides an electronic device, which includes: a memory, a processor, and a computer program / instructions stored on the memory, and the processor executes the computer program / instructions to implement the cable-supported bridge structure damage identification method in the embodiments of the present application.

[0188] Although not shown, the electronic device includes a processor, which can perform various appropriate operations and processes according to the programs and / or data stored in the read-only memory (ROM) and / or the programs and / or data loaded from the storage part into the random access memory (RAM). The processor can be a multi-core processor or can include multiple processors. In some embodiments, the processor can include a general main processor and one or more special coprocessors, for example, a central processing unit, a graphics processing unit (GPU), a neural network processing unit (NPU), a digital signal processing unit (DSP), and so on. In the RAM, various programs and data required for device operation are also stored. The processor, the ROM, and the RAM are connected to each other through a bus. The input / output (I / O) interface is also connected to the bus.

[0189] The above-mentioned processor and the memory are jointly used to execute the programs / instructions stored in the memory, and when the programs / instructions are executed by a computer, they can implement the methods, steps, or functions described in the above embodiments.

[0190] Although not shown, an embodiment of the present invention further provides a computer-readable storage medium, on which a computer program / instructions are stored, and when the computer program / instructions are executed by a processor, the cable-supported bridge structure damage identification method in the embodiments of the present application is implemented.

[0191] The storage medium in the embodiments of the present invention includes items that are permanent and non-permanent, removable and non-removable, and can store information by any method or technology. Examples of storage media include, but are not limited to, phase change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, compact disc read-only memory (CD-ROM), digital versatile disc (DVD) or other optical storage, magnetic cassette tapes, magnetic disk storage or other magnetic storage devices, or any other non-transmission medium that can be used to store information accessible by a computing device.

[0192] A readable storage medium includes permanent and non-permanent, removable and non-removable media, and information storage can be implemented by any method or technology. The information can be computer-readable instructions, data structures, program modules, or other data. Examples of computer storage media include, but are not limited to, phase change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, compact disc read-only memory (CD-ROM), digital versatile disc (DVD) or other optical storage, magnetic cassette tapes, disk storage or other magnetic storage devices, or any other non-transitory medium that can be used to store information that can be accessed by a computing device. As defined herein, computer-readable media do not include transitory computer-readable media, such as modulated data signals and carrier waves.

[0193] Although not shown, an embodiment of the present invention also provides a computer program product, including: a computer program / instructions, which when executed by a processor, implement the cable-supported bridge structure damage identification method in the embodiments of the present application.

[0194] The above-disclosed are only specific embodiments of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art within the technical scope disclosed by the present invention can easily think of changes or variations, which should all be covered within the protection scope of the present invention.

Claims

1. A method for identifying damage to a cable-supported bridge structure, characterized in that: The identification method comprises: Constructing a sample data set; wherein each sample data in the sample data set includes an input quantity and an output quantity, the input quantity is a spliced ​​image, and the output quantity is a damage condition of the spliced ​​image; the spliced ​​image is obtained by splicing the acceleration response signals of the main beam and the acceleration response signals of the cable after they are converted into time-frequency domain images respectively; Constructing a damage recognition model; wherein the damage recognition model includes an input layer, an initial convolution layer, a maximum pooling layer, a first combination layer, a second combination layer, a third combination layer, a fourth combination layer, a first global average pooling layer, a first fully connected layer and an output layer connected in sequence; the initial convolution layer has 6 input channels; each combination layer includes a convolution block, an SE attention mechanism and a residual block connected in sequence; Using the sample data set to train and verify the damage identification model to obtain a target damage identification model; Obtain actual acceleration response signals of the main beam and cables; The actual acceleration response signals of the main beam and the cable are converted into time-frequency domain images respectively, and then the time-frequency domain images of the main beam and the cable are spliced ​​to obtain an actual spliced ​​image; The target damage identification model is used to perform damage identification on the actual spliced ​​image.

2. The cable-supported bridge structure damage identification method according to claim 1, characterized in that: A sample data set was constructed through field experiments, including: Building a cable-supported bridge test model, setting a first sensor on the main beam of the cable-supported bridge test model, and setting a second sensor on the cable of the cable-supported bridge test model; Simulating different damage conditions on the cable-bearing bridge test model, using a first sensor to collect acceleration response signals of the main beam under different damage conditions, and using a second sensor to collect acceleration response signals of the cable under different damage conditions; Preprocess the acceleration response signals of the main beam and cables under different damage conditions; Convert the pre-processed acceleration response signals of the main beam and cables into time-frequency domain images; The time-frequency domain images of the main beam and the cable under the same damage condition are spliced ​​to obtain a spliced ​​image; The sample data set is constructed according to the spliced ​​images and the damage conditions to which they belong.

3. The cable-supported bridge structure damage identification method according to claim 1, characterized in that: The damage condition includes the damage position, vehicle speed, vehicle mass and damage degree; the vehicle speed is 0.5m / s, 0.75m / s or 1m / s, and the vehicle mass is 5kg, 10kg or 20kg; the cables are numbered from right to left; The first damage condition: no damage, a combination of different vehicle speeds and different vehicle masses; The second damage condition: the first cable is damaged, with different combinations of vehicle speeds and vehicle masses, and the damage degree is 10%, 20%, or 50%; The third damage condition: the third cable is damaged, with different combinations of vehicle speeds and vehicle masses, and the damage degree is 10%, 20% or 50%; The fourth damage condition: the sixth cable is damaged, with different combinations of vehicle speeds and vehicle masses, and the damage degree is 10%, 20%, or 50%; The fifth damage condition: the first and fourth cables are damaged, with different combinations of vehicle speeds and vehicle masses, and the damage degree is 10%, 20% or 50%; The sixth damage condition: the fifth and sixth cables are damaged, with a combination of different vehicle speeds and different vehicle masses, and the degree of damage is 10%, 20% or 50%.

4. The cable-supported bridge structure damage identification method according to claim 1 or 2, characterized in that: The S transform is used to realize the conversion of time-frequency domain images. The specific formula is: Among them, S(τ,f) represents the time-frequency domain image, x(t) represents the acceleration response signal, ω(τ-t,f) represents the Gaussian window function, t represents time, τ represents the position of the Gaussian window on the time axis t, f represents the frequency of the acceleration response signal, and i represents the sign of the imaginary part.

5. The cable-supported bridge structure damage identification method according to claim 1, characterized in that: The convolution block includes a first convolution layer, a first activation layer, a second convolution layer, a second activation layer, a third convolution layer and a batch normalization layer connected in sequence; The number of residual blocks of the first combination layer is 2, the number of residual blocks of the second combination layer is 3, the number of residual blocks of the third combination layer is 5, and the number of residual blocks of the fourth combination layer is 2; each of the residual blocks includes a fourth convolutional layer, a third activation layer, a fifth convolutional layer, a fourth activation layer and a sixth convolutional layer connected in sequence.

6. The cable-supported bridge structure damage identification method according to claim 1, characterized in that: The SE attention mechanism includes a second global average pooling layer, a second fully connected layer, a third fully connected layer and a reweighted layer connected in sequence.

7. The cable-supported bridge structure damage identification method according to claim 1, characterized in that: Using the sample data set to train the damage identification model includes: Using the initial convolutional layer to extract features from the spliced ​​image received by the input layer to obtain a first feature map; Performing maximum pooling on the first feature map using the maximum pooling layer to obtain a second feature map; Using the first combined layer to perform feature extraction on the second feature map to obtain a third feature map; Using the second combined layer to perform feature extraction on the third feature map to obtain a fourth feature map; Using the third combined layer to perform feature extraction on the fourth feature map to obtain a fifth feature map; Using the fourth combined layer to perform feature extraction on the fifth feature map to obtain a sixth feature map; Using the first global average pooling layer to perform global average pooling on the sixth feature map to obtain a single feature; Using the first fully connected layer to flatten the single feature and map it to the classification space to obtain the predicted damage category; A loss value is calculated according to the predicted damage category and the damage condition of the stitched image, and the parameters of the damage recognition model are adjusted according to the loss value until the number of training times is reached or the loss value is less than the set loss accuracy.

8. An electronic device comprising a memory, a processor, and a computer program / instruction stored in the memory, characterized in that: The processor executes the computer program / instructions to implement the cable-supported bridge structure damage identification method according to any one of claims 1 to 7.

9. A computer-readable storage medium having a computer program / instruction stored thereon, characterized in that: When the computer program / instruction is executed by a processor, the cable-supported bridge structure damage identification method as described in any one of claims 1 to 7 is implemented.

10. A computer program product comprising a computer program / instructions, characterized in that When the computer program / instruction is executed by a processor, the cable-supported bridge structure damage identification method as described in any one of claims 1 to 7 is implemented.

Images