A multi-node damage diagnosis method for a frame structure

Abstract

The present invention discloses a method for diagnosing multi-node damage of a frame structure, comprising the following steps: S1) Collect the acceleration data of each node of the frame structure under seismic excitation according to the set damage conditions and sampling frequency; S2) Establish a damage database of the frame structure; S3) Divide the sample data in the damage database into a training set and a test set according to a ratio, and perform multi-label annotation on the samples; S4) Build a multi-output residual convolutional neural network model; S5) Input the training set and test set data into the built multi-output residual convolutional neural network model respectively to complete the training and dynamic testing of the network; S6) The trained multi-output residual convolutional neural network model can automatically extract sensitive information on the damage location and damage degree from the acceleration data and give the damage diagnosis results of each node of the frame structure. The method of the present invention can directly extract damage features from the original data and simultaneously complete the diagnosis of the damage conditions of all nodes of the frame structure.

Description

Technical Field

[0001] The present invention relates to the field of structural damage identification, and particularly to a multi-node damage diagnosis method for a frame structure. Background Art

[0002] During the long-term service of a frame structure, damage is inevitably caused by factors such as environmental corrosion and material aging. When the damage accumulates to a certain extent, it may cause the overall failure of the structure, posing a serious threat to people's lives and property safety. Therefore, it is very necessary to timely detect the location of structural damage and actively take repair measures.

[0003] Damage detection methods based on machine learning are mainly divided into damage detection methods based on traditional machine learning and current damage detection methods based on deep learning. The damage detection method based on traditional machine learning mainly uses an artificial neural network (ANN) or a support vector machine (SVM) for damage identification. The idea of this method is to first perform signal processing on the original data to preliminarily extract the features of the data, then select the features most sensitive to structural damage as the input, and finally use ANN or SVM to classify the damage; this method requires manual selection and extraction of features, that is, the professional knowledge and experience of experts, and it is more difficult to apply in engineering. The damage detection method based on deep learning mainly uses time series data and a convolutional neural network for damage identification. The idea of this method is to use the acceleration data or displacement data collected by the structure in various damaged states and undamaged states to establish a structural damage identification data set, and then use the convolutional neural network to automatically extract damage-sensitive features from the data set to complete the identification or classification of damage.

[0004] There are a large number of existing studies on structural damage identification based on vibration data and convolutional neural networks, but most of them adopt a single-output network model, that is, they can only classify various damage conditions, and it is difficult to intuitively judge the damage location, and they are mainly aimed at beam structures. Therefore, there is a need to provide a method that can simultaneously diagnose the damage of multiple nodes of a complex frame structure. Summary of the Invention

[0005] The technical problem to be solved by the present invention is to provide a multi-node damage diagnosis method for a frame structure in view of the defects in the prior art.

[0006] The technical solution adopted by the present invention to solve its technical problems is: a multi-node damage diagnosis method for a frame structure, including the following steps:

[0007] S1) Perform dynamic analysis of the frame in finite element software, determine the target frame structure, damage simulation method, and form of seismic excitation, and collect the vibration acceleration data of each node under different damage conditions of the frame structure under seismic excitation according to the set frame structure damage conditions and sampling frequency;

[0008] S2) Perform data augmentation on the collected vibration acceleration data to establish a frame structure damage database; the data in the frame structure damage database is saved as (damage condition, node acceleration, node position number, form of seismic excitation);

[0009] S3) Divide the samples in the damage database into a training set and a test set, and perform multi-label annotation on the samples in the training set and the test set;

[0010] The label is a vector composed of multiple elements, and the position numbers of each element correspond to the position numbers of each node;

[0011] When the value of the element is 1, it represents that the node is damaged, and when the value of the element is 0, it represents that the node is intact;

[0012] S4) Build a multi-output residual convolutional neural network model for node damage diagnosis;

[0013] The structure of the multi-output residual convolutional neural network model is as follows:

[0014] It includes an input layer, and the input layer is a convolutional layer with a convolutional kernel size set to s×3, where s corresponds to the total number of nodes in the frame structure;

[0015] The initial damage feature extraction layer includes a BN layer and a Relu activation function to perform initial extraction of damage features;

[0016] The deep damage feature extraction layer is composed of three convolutional layer groups connected in sequence, each of which is composed of residual blocks; the specific structure of each residual block is: composed of two convolutional layers, and a BN layer is connected after each convolutional layer, and a Relu activation function is used between the two modules;

[0017] An average pooling layer to compress the data and the number of parameters and reduce the risk of overfitting;

[0018] A flattening layer to expand the previous output into a one-dimensional vector;

[0019] Finally, a fully connected layer is constructed for each node respectively, that is, s fully connected layers are used;

[0020] S5) Input the training set and test set sample data into the built multi-output residual convolutional neural network model respectively to complete the training and dynamic testing of the network model;

[0021] For the calculation of the training loss function of the multi-output network model, it is necessary to calculate the loss function values of each label separately first, and then sum the loss function values of all labels to obtain the overall loss function value of the multi-output network model;

[0022] Train the network through the Adam algorithm. When the loss function value is less than the set threshold, save the network model;

[0023] When dynamically testing the network, the absolute matching rate is used to calculate the prediction accuracy. That is, when the label predictions of all nodes are correct, the sample prediction is correct;

[0024] S6) The trained multi-output residual convolutional neural network model can automatically extract sensitive information on the damage location and damage degree from the acceleration data, and give the damage diagnosis results of each node of the frame structure.

[0025] According to the above scheme, the calculation of the prediction accuracy of the multi-output network model includes:

[0026] The absolute matching rate is used to calculate the prediction accuracy. That is, only when all labels are predicted correctly is the sample prediction considered correct;

[0027] The calculation formula is

[0028]

[0029] where n represents the total number of samples, I(·) is the indicator function, which takes 1 when the true label Y i and the predicted label Z i are exactly equal, and 0 otherwise.

[0030] The beneficial effects produced by the present invention are:

[0031] (1) The present invention uses the original data as the input, reduces complex data processing operations, and reduces the difficulty of obtaining training samples.

[0032] (2) By adopting multi-label annotation and a multi-output residual convolutional neural network model, the problem of difficult to simultaneously perform damage diagnosis on multiple parts of a complex frame can be well solved, that is, it can simultaneously judge whether there is damage to multiple nodes of the frame.

[0033] (3) By introducing the ideas of residual learning, multi-output classification and the step-by-step extraction design of features into the convolutional neural network model, the network model can have extremely high prediction accuracy, extremely small training error values, a fast convergence speed and good stability. Brief Description of the Drawings

[0034] The present invention will be further described below in conjunction with the accompanying drawings and embodiments. In the accompanying drawings:

[0035] Figure 1 is a flowchart of a multi-node damage diagnosis method for a frame structure of the present invention;

[0036] Figure 2 is a diagram of a frame structure model and a schematic diagram of node damage adopted by the present invention;

[0037] Figure 3 is a schematic diagram of the structure of a multi-output residual convolutional neural network model considering residual learning adopted by the present invention;

[0038] Figure 4 is a loss function curve graph of the network of the present invention on the training set;

[0039] Figure 5 is an accuracy curve graph of the network of the present invention on the test set;

[0040] Figure 6 is a damage probability prediction result graph of 5-8 node damages of the present invention;

[0041] Figure 7 is a comparison graph of the test accuracy of the network model adopted by the present invention and other network models. Detailed implementation manners

[0042] In order to make the objectives, technical solutions and advantages of the present invention clearer and more understandable, the present invention will be further described in detail below in conjunction with embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0043] As Figure 1 shown, a multi-node damage diagnosis method for a frame structure includes the following steps:

[0044] S1) Perform dynamic analysis of the frame in finite element software, determine the target frame structure, damage simulation method and form of seismic excitation, and collect the vibration acceleration data of each node under different damage conditions of the frame structure under seismic excitation according to the set frame structure damage conditions and sampling frequency;

[0045] Collect the acceleration data of each node of the frame structure under seismic excitation by the following method:

[0046] Simulate damage by loosening the bolts at the beam-column connection points (actual structure) or reducing the stiffness near the beam-column connection points (numerical simulation structure);

[0047] Select white noise excitation to simulate seismic loads;

[0048] S2) Perform data augmentation on the collected vibration acceleration data to establish a frame structure damage database; the data in the frame structure damage database is saved as (damage condition, node acceleration, node position number, seismic excitation form);

[0049] The data augmentation process is as follows:

[0050] Arrange the acceleration response data of each node under the same condition according to the node number to form a sample matrix;

[0051] Perform sliding window sampling on the sample matrix along the time axis to generate multiple samples for each condition;

[0052] Store all samples in a folder to construct a frame structure damage database;

[0053] S3) Divide the samples in the damage database into a training set and a test set, and perform multi-label annotation on the samples in the training set and the test set;

[0054] Divide the samples in the damage database into a training set and a test set according to a ratio of 9:1;

[0055] Perform multi-label annotation on the samples in the training set and the test set respectively. The label is a vector composed of multiple elements, and the position numbers of each element correspond to the position numbers of each node;

[0056] When the value of the element is 1, it means the node is damaged; when the value of the element is 0, it means the node is intact;

[0057] S4) Build a multi-output residual convolutional neural network model for node damage diagnosis;

[0058] The structure of the multi-output residual convolutional neural network model is as follows:

[0059] It includes an input layer, and the input layer is a convolutional layer with a convolutional kernel size set to s×3, where s corresponds to the total number of nodes in the frame structure;

[0060] The initial damage feature extraction layer includes a BN layer and a Relu activation function to perform initial extraction of damage features;

[0061] The deep damage feature extraction layer is composed of three convolutional layer groups connected in sequence, each of which is composed of residual blocks;

[0062] An average pooling layer to compress the data and the number of parameters and reduce the risk of overfitting;

[0063] A flattening layer to expand the previous output into a one-dimensional vector;

[0064] Finally, a fully connected layer is constructed for each node respectively, that is, s fully connected layers are adopted;

[0065] Fully connected layer: For each node of the frame structure, a fully connected layer is constructed, and the output size is set to 2, that is, multiple feature classification layers are adopted to diagnose whether there is damage for each node of the frame structure respectively.

[0066] S5) Input the training set and test set sample data into the constructed multi-output residual convolutional neural network model respectively to complete the training and dynamic testing of the network model;

[0067] For the calculation of the training loss function of the multi-output network model, the loss function values of each label need to be calculated separately first, and then the loss function values of all labels are summed to obtain the overall loss function value of the multi-output network model;

[0068] The network is trained by the Adam algorithm. When the loss function value is less than the set threshold, the network model is saved; the purpose of training is to make the loss function value as close to 0 as possible;

[0069] When dynamically testing the network, the absolute matching rate is used to calculate the prediction accuracy, that is, when the labels of all nodes are predicted correctly, the sample is predicted correctly;

[0070] During the training process, the test set is input into the network every several cycles to dynamically verify the accuracy of network prediction.

[0071] The calculation of the loss function value of the multi-output network model is as follows:

[0072] Calculate the loss function values of each label separately;

[0073] Then add the loss function values of each label to obtain the overall loss function value of the network model.

[0074] The calculation of the prediction accuracy of the multi-output network model is as follows:

[0075] The absolute matching rate is used to calculate the prediction accuracy, that is, only when all labels are predicted correctly, the sample is considered to be predicted correctly;

[0076] The calculation formula is

[0077] where n represents the total number of samples, I(·) is the indicator function, which takes 1 when the true label Y i and the predicted label Z i are exactly equal, and 0 otherwise.

[0078] S6) After the multi-output residual convolutional neural network model is trained, it can automatically extract sensitive information about the damage location and degree from the acceleration data and give the damage diagnosis results of each node of the frame structure.

[0079] Set the damage conditions according to the number of damaged nodes, the location of the damaged nodes, and the degree of damage.

[0080] Now, combined with an example, the specific steps of the inventive method are demonstrated:

[0081] Here, a numerical simulation of a 4-story and 1-span plane symmetric frame is taken as an example:

[0082] First, use the SAP2000 software to establish Figure 2 the undamaged model of the frame structure shown; then use the MATLAB software to generate white noise excitation with a duration of 2000 seconds, a sampling frequency of 200 HZ, and an amplitude of 6 mm / s 2 ; then apply the white noise excitation to the bottom of the column and conduct a time history analysis under the undamaged state; finally, export the acceleration data of each node.

[0083] The damage simulation method is to reduce the elastic modulus in the area of 0.4 m near the beam-column connection points, as Figure 2 shown. Here, three representative degrees of damage are considered, 20%, 50%, and 80%, which are represented by multiplying the original elastic modulus by 80%, 50%, and 20% respectively.

[0084] According to the number of damaged nodes, the location of the damaged nodes, and the degree of damage, a total of 30 conditions (undamaged + 12 single-node damages + 17 multi-node damages) are established. The length of the acceleration data of each node under each condition is 400000. Then, store the acceleration data of each node under the same damage condition in a folder and name it according to the location and degree of the damaged nodes.

[0085] Arrange the acceleration data of all nodes under each condition in ascending order of node numbers to form an 8×400000 sample matrix; perform non-cyclic sampling of the sliding window along the time axis for the sample matrix so that 800 8×500 sample matrices are generated under each condition; then store all the samples in a folder to construct a frame structure damage database, and the damage database contains 24000 samples.

[0086] Divide the samples in the damage database into a training set and a test set according to a ratio of 9:1. Then, the training set has 21,600 samples and the test set has 2,400 samples. Perform multi-label annotation on the samples. The label is a vector composed of 8 elements. The positions of the 8 elements represent the numbers of 8 nodes respectively. When the value of an element is 1, it means the node at that place is damaged; when the value of an element is 0, it means the node at that place is not damaged. For example, [0, 0, 1, 1, 0, 0, 0, 0] represents that the third node and the fourth node are damaged.

[0087] Design a multi-output residual convolutional neural network model. This network model aims to automatically extract sensitive features of the damage location and damage degree from the original acceleration data of a complex framework, and then perform damage diagnosis on multiple nodes of the framework structure simultaneously based on this feature and multiple classifiers. The designed network model is as Figure 3 shown. First, there is an input layer, and the size of its convolution kernel is set to 8×3 (8 corresponds to the number of nodes in the framework structure). Then, a preliminary damage feature extraction layer is used, which includes a BN layer and a Relu activation function to preliminarily extract damage features from the acceleration sample matrix. Next, a deep damage feature extraction layer is used, which includes three sequentially connected convolutional layer groups. Each convolutional layer group is composed of different numbers of residual blocks, and each residual block includes two convolutional layers, and a BN layer is connected after each convolutional layer to further extract deep damage features from the acceleration sample matrix. Then, an average pooling layer is used to compress the data and the number of parameters and reduce the risk of overfitting. Then, a flattening layer is used to expand the previous output into a one-dimensional vector. Finally, a fully connected layer is constructed for each node respectively, that is, 8 fully connected layers are adopted. This network model is established based on the Pytorch deep learning framework and uses the Python language. For the calculation of the loss function value of the multi-output network model, first calculate the loss function value of each label separately, and then sum the loss function values of all labels to obtain the overall loss function value of the network model. Here, the cross-entropy loss function is used to calculate the error value for each label. The network is trained by the Adam algorithm. When the loss function value is less than the set threshold, the network model is saved. When dynamically testing the network, the absolute matching rate is used to calculate the prediction accuracy, that is, when the label predictions of all nodes are correct, the sample is considered to be predicted correctly.

[0088] Input the training set into the built network model, and use the Adam method to train the network. The initial learning rate is taken as 0.005, the batch size is taken as 200, and it is trained for 30 generations. During the training process, the test set is input into the network every several loops to dynamically verify the accuracy of the network prediction. Figure 4 The training loss function curve of the network is given, Figure 5 The prediction accuracy curve of the network is given, Figure 6The comparison chart of the predicted values and the true values of the damage probabilities of each node of a certain sample is given. From Figures 4 to 6 it can be seen that by using the multi-output residual convolutional neural network model, sensitive features of the damage location and degree can be automatically extracted from the original acceleration data, realizing the simultaneous damage diagnosis of multiple nodes of a complex frame, and having extremely high accuracy. Figure 6 When the value of Damage conditions in

[0089] is 0, it represents that the node is not damaged, and when the value is 1, it represents that the node is damaged. Figure 7 To further verify the effectiveness of the method for simultaneously diagnosing the damage of multiple nodes of a frame using the multi-output residual convolutional neural network model proposed in the present invention, the present invention also compares its test accuracy with the test accuracies of the single-output convolutional neural network model and the single-output residual convolutional neural network model. The comparison is as shown in Figure 7 and Table 1. From

[0090] Table 1 Comparison table of test accuracies

[0091]

[0092] It should be pointed out that the above examples have preliminarily verified the effectiveness of the method of the present invention, that is, the method of the present invention can be extended to practical structures for application and can obtain good prediction results.

[0093] It should be understood that for those of ordinary skill in the art, improvements or transformations can be made according to the above description, and all such improvements and transformations should fall within the protection scope of the appended claims of the present invention.

Claims

1. A multi-node damage diagnosis method for a frame structure, characterized in that, It includes the following steps: S1) Conduct dynamic analysis of the frame in finite element software, determine the target frame structure, damage simulation method, and form of seismic excitation. According to the set frame structure damage conditions and sampling frequency, collect the vibration acceleration data of each node under different damage conditions of the frame structure under the action of seismic excitation; S2) Perform data augmentation processing on the collected vibration acceleration data to establish a frame structure damage database; The data in the frame structure damage database is saved as (damage condition, node acceleration, node position number, form of seismic excitation); S3) Divide the samples in the damage database into a training set and a test set, and perform multi-label annotation on the samples in the training set and the test set; The label is a vector composed of multiple elements, and the position numbers of each element correspond to the position numbers of each node respectively; When the value of the element is 1, it represents that the node is damaged, and when the value of the element is 0, it represents that the node is intact; S4) Build a multi-output residual convolutional neural network model for node damage diagnosis; The structure of the multi-output residual convolutional neural network model is as follows: It includes an input layer, and the input layer is a convolutional layer with a convolutional kernel size set to s×3, where s corresponds to the total number of nodes in the frame structure; The preliminary damage feature extraction layer includes a BN layer and a Relu activation function to perform preliminary extraction of damage features; The deep damage feature extraction layer is composed of three convolutional layer groups connected in sequence, each of which is composed of residual blocks; An average pooling layer to compress the data and the number of parameters and reduce the risk of overfitting; A flattening layer to expand the previous output into a one-dimensional vector; Finally, a fully connected layer is constructed for each node respectively, that is, s fully connected layers are used; S5) Input the training set and test set sample data into the built multi-output residual convolutional neural network model respectively to complete the training and dynamic testing of the network model; For the calculation of the training loss function of the multi-output network model, first calculate the loss function value of each label respectively, and then sum the loss function values of all labels to obtain the overall loss function value of the multi-output network model; Train the network through the Adam algorithm, and save the network model when the loss function value is less than the set threshold; When performing dynamic testing on the network, the absolute matching rate is used to calculate the prediction accuracy, that is, when the label predictions of all nodes are correct, the sample prediction is correct; S6) The multi-output residual convolutional neural network model after training can automatically extract sensitive information about the damage location and damage degree from the acceleration data and give the damage diagnosis results of each node of the frame structure.

2. The multi-node damage diagnosis method for a frame structure according to claim 1, characterized in that, In step S4), the loss function value of each label is calculated using the cross-entropy loss function.

3. The multi-node damage diagnosis method for a frame structure according to claim 1, characterized in that The specific structure of the residual block in step S4) is: composed of two convolutional layers, and a BN layer is connected after each convolutional layer, and a Relu activation function is used between the two modules.

4. The method for diagnosing multi-node damage of a frame structure according to claim 1, characterized in that The calculation of the prediction accuracy of the multi-output network model in step S4) includes: The absolute matching rate is used to calculate the prediction accuracy, that is, only when all labels are predicted correctly is the sample prediction considered correct; The calculation formula is Where n represents the total number of samples, I(·) is the indicator function, and when the true label Y i and predicted label Z i If they are completely equal, it is 1; otherwise, it is 0.

Images