Small sample frame structure damage detection method based on meta learning and DCCMN model

Abstract

The invention discloses a small sample frame structure damage detection method based on meta-learning and a DCCMN model, and belongs to the technical field of structure damage detection, and the method comprises the following steps: S1, collecting a vibration signal, and dividing data into a training task and a test task through data processing; s2, randomly selecting a task from the training tasks, and inputting a support set of the selected task into the DCCMN model for training to complete internal circulation; s3, evaluating the model by using the query set, calculating a loss function, and updating model parameters; and S4, finally obtaining a detection result through model testing. According to the small sample frame structure damage detection method based on meta learning and the DCCMN model, the problem of data scarcity in structure damage detection is solved, the accuracy and generalization ability of small sample damage detection are improved, and a more reliable and efficient structure health monitoring solution is brought to the engineering field.

Description

Technical Field

[0001] The present invention relates to the technical field of structural damage detection, and in particular to a small sample frame structure damage detection method based on meta-learning and a DCCMN model. Background Art

[0002] Engineering structures are susceptible to human and environmental influences, including gradual failures caused by corrosion and fatigue, as well as sudden failures caused by earthquakes and fires. These long-term and short-term damages will reduce the service life of the structure and reduce its reliability and safety. Therefore, it is necessary to monitor and detect the structure.

[0003] With the rapid development of machine learning and deep learning technologies, deep learning-based methods provide new possibilities and efficiencies for structural damage detection, which is of great significance for solving damage problems with complex nonlinear characteristics. Deep learning has shown many advantages in structural damage detection, but it also faces some challenges. For example, deep learning methods usually require a large number of labeled samples for training, while in civil structure damage detection, the available labeled samples are often limited, which may cause deep learning methods to perform poorly in generalizing to new samples. In addition, the complexity of deep learning models may require more computing resources and time. Summary of the invention

[0004] The purpose of this invention is to provide a small sample frame structure damage detection method based on meta-learning and DCCMN model, to solve the problem of data scarcity in structural damage detection, to improve the accuracy and generalization ability of small sample damage detection, and to bring more reliable and efficient structural health monitoring solutions to the engineering field.

[0005] To achieve the above object, the present invention provides a small sample frame structure damage detection method based on meta-learning and DCCMN model, comprising the following steps: Step S1, in the data processing stage, vibration signals are collected and stored according to the sensor number, and then the data is normalized. After data processing, the data is divided into training tasks and test tasks; Step S2: In the model training phase, a task is randomly selected from the training tasks, and the support set of the selected task is input into the DCCMN model for training to complete the inner loop; Step S3: Use the query set to evaluate the model, calculate the loss function, and update the model parameters to complete the outer loop; In the outer loop, the model parameters are further updated by calculating the average loss function and gradient of multiple tasks, and the updated DCCMN model is saved after the set number of loops. Step S4: In the model testing phase, the support set is used to fine-tune the parameters of the saved DCCMN model, and then the query set is used to test the DCCMN model after parameter fine-tuning, and finally the detection result is obtained.

[0006] Preferably, the calculation process of the inner loop parameter update in step S2 is as follows: ; in, Indicates the parameters updated by the inner loop; represents the initial parameters; Represents the learning rate of the inner loop; represents the loss function of the support set in the training task. The loss function adopts the cross entropy loss function, and its formula is as follows: ; in, Indicates the number of samples in the support set; Indicates The true labels of samples; Indicates The predicted label of the sample, that is, the model's prediction of The probability distribution of samples belonging to each category.

[0007] Preferably, the parameter update calculation process of the outer loop in step S3 is as follows: ; in, The updated parameters of the outer loop; Represents the learning rate of the outer loop; represents the meta-loss function; Represents the gradient accumulation of all training tasks; Indicates the number of times the outer loop is executed.

[0008] Preferably, the calculation process of the parameter fine-tuning stage in step S4 is as follows: ; in, represents the fine-tuned parameters in the test task; Represents the learning rate for parameter fine-tuning; represents the loss function of the support set in the test task.

[0009] Preferably, in the DCCMN model, the convolution kernel size of the first convolution layer is set to 16, the step size is set to 4, the convolution kernel size of the second convolution layer is 8, the step size is set to 2, and a BN layer and a maximum pooling layer are added after each convolution layer, and then a cross-mixing module is gradually introduced for feature fusion. Four cross-mixing modules are added to the DCCMN model, and a global average pooling layer is added to the last mixing module to reduce the dimension of the features. Finally, the softmax function is used to obtain the output of the DCCMN model.

[0010] Preferably, the global average is calculated as follows: ; in, yes c The output of each channel; is the input feature map in The value at H and W are the height and width of the feature map respectively; The calculation of the softmax function is as follows: ; in, represents the probability of each output, The sum of all is 1; K is the number of categories for the multi-classification problem; Represents the first Perform exponential operation on each element; Represents the first The meaning of the above formula is to find the Elements account for all The percentage of elements.

[0011] Therefore, the present invention adopts the above-mentioned small sample frame structure damage detection method based on meta-learning and DCCMN model, and uses the DCCMN model to perform feature learning and damage detection of vibration signals. During the training and testing process, through the use of support sets and query sets, the model can perform effective parameter updates and fine-tuning to adapt to the data characteristics of different tasks, thereby achieving accurate damage detection.

[0012] The technical solution of the present invention is further described in detail below through the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0013] Figure 1 It is a damage detection flow chart of an embodiment of a small sample frame structure damage detection method based on meta-learning and DCCMN model of the present invention; Figure 2It is a DCCMN model structure diagram of an embodiment of a small sample frame structure damage detection method based on meta-learning and a DCCMN model of the present invention; Figure 3 It is a frame structure and sensor position diagram of an embodiment of a small sample frame structure damage detection method based on meta-learning and DCCMN model of the present invention; Figure 3 (a) in the figure shows the 3D framework structure diagram; Figure 3 (b) in the figure shows the sensor location diagram; Figure 4 It is a 2-way K-shot training task process diagram of an embodiment of a small sample frame structure damage detection method based on meta-learning and DCCMN model of the present invention; Figure 4 (a) shows the change of the accuracy of the support set and query set as the number of outer loops changes; Figure 4 (b) shows the change of loss function value of support set, query set and meta-loss as the number of outer loops changes; Figure 5 This is a 2-way K-shot test task result diagram of an embodiment of a small sample frame structure damage detection method based on meta-learning and DCCMN model of the present invention; Figure 5 (a) in the figure shows a bar graph of the query set accuracy; Figure 5 (b) in the figure shows a bar graph of the loss function value of the query set; Figure 6 It is a 9-way K-shot training task process diagram in an embodiment of a small sample frame structure damage detection method based on meta-learning and DCCMN model of the present invention; Figure 6 (a) shows the change of the accuracy of the support set and query set as the number of outer loops changes; Figure 6 (b) shows the change of loss function value of support set, query set and meta-loss as the number of outer loops changes; Figure 7 This is a 9-way K-shot test task result diagram of an embodiment of a small sample frame structure damage detection method based on meta-learning and DCCMN model of the present invention; Figure 7 (a) in the figure shows a bar graph of the query set accuracy; Figure 7 (b) in the figure shows a bar graph of the loss function value of the query set; Figure 8 Comparison chart of the average model accuracy in the embodiment of the small sample frame structure damage detection method based on meta-learning and DCCMN model of the present invention; Fig. 9 A comparison diagram of model confusion matrices in an embodiment of a small sample frame structure damage detection method based on meta-learning and DCCMN model of the present invention; Fig. 9 (a) in the figure represents the confusion matrix diagram of the SVM model; Fig. 9(b) in the figure shows the confusion matrix diagram of the ResNet-18 model; Fig. 9 (c) in the figure shows the confusion matrix diagram of the DCCMN model; Fig. 9 (d) in the figure represents the confusion matrix diagram of the Meta+RestNet-18 model; Fig. 9 (e) in the figure represents the confusion matrix diagram of Meta+DCCMN model; Fig.10 TSNE visualization comparison diagram of the model in the embodiment of the small sample frame structure damage detection method based on meta-learning and DCCMN model of the present invention; Fig.10 (a) shows the TSNE visualization diagram of the original data set; Fig.10 (b) shows the TSNE visualization diagram of the SVM model; Fig.10 (c) shows the TSNE visualization diagram of the ResNet-18 model; Fig.10 (d) shows the TSNE visualization diagram of the DCCMN model; Fig.10 (e) in the figure shows the TSNE visualization diagram of the Meta+RestNet-18 model; Fig.10 (f) in the figure shows the TSNE visualization diagram of the Meta+DCCMN model. DETAILED DESCRIPTION

[0014] The technical solution of the present invention is further described below through the accompanying drawings and embodiments.

[0015] Unless otherwise defined, technical or scientific terms used in the present invention shall have the common meanings understood by one having ordinary skills in the field to which the present invention belongs.

[0016] Embodiment 1 like Figure 1 As shown, the present invention provides a small sample frame structure damage detection method based on meta-learning and DCCMN model, comprising the following steps: Step S1: In the data processing stage, vibration signals are collected and stored according to the sensor number, and then the data is normalized so as to better process and analyze the data in subsequent steps. After data processing, the data is divided into training tasks and test tasks; Step S2: In the model training phase, firstly, a task is randomly selected from the training tasks, and the support set of the selected task is input into the DCCMN model for training to complete the inner loop.

[0017] The calculation process of the inner loop parameter update is as follows: ; in, Indicates the parameters updated by the inner loop; represents the initial parameters; Represents the learning rate of the inner loop; represents the loss function of the support set in the training task. The loss function adopts the cross entropy loss function, and its formula is as follows: ; in, N Indicates the number of samples in the support set; Indicates The true labels of samples; Indicates The predicted label of the sample, that is, the model's prediction of The probability distribution of samples belonging to each category.

[0018] Step S3: Use the query set to evaluate the model, calculate the loss function, and update the model parameters to complete the outer loop; During the outer loop, the model parameters are further updated by calculating the average loss function and gradient of multiple tasks, and the updated DCCMN model is saved after the set number of loops.

[0019] The parameter update calculation process of the outer loop is as follows: ; in, The updated parameters of the outer loop; Represents the learning rate of the outer loop; represents the meta-loss function; Represents the gradient accumulation of all training tasks; E Indicates the number of times the outer loop is executed.

[0020] Step S4: In the model testing phase, a test task is randomly selected, and the number of support set samples of the task can be small. First, the saved DCCMN model is fine-tuned using the support set, and then the DCCMN model with fine-tuned parameters is tested using the query set to obtain the detection result.

[0021] The calculation process of the parameter fine-tuning stage is as follows: ; in, represents the fine-tuned parameters in the test task; Represents the learning rate for parameter fine-tuning; represents the loss function of the support set in the test task.

[0022] The DCCMN model structure is as follows Figure 2As shown in the figure, in the initial stage of the DCCMN model, large convolution kernels and simple operations are used, such as setting the convolution kernel size of the first convolution layer to 16 and the step size to 4, the convolution kernel size of the second convolution layer to 8 and the step size to 2, and adding a BN layer and a maximum pooling layer after each convolution layer. The purpose of this is to help the model quickly learn the low-level features of the input data, reduce excessive parameters and calculations, make the model easier to converge, and reduce the possibility of overfitting.

[0023] Then, cross-mixing modules are gradually introduced for feature fusion. Four cross-mixing modules are added to the DCCMN model, which helps improve the performance and training efficiency of the model to capture more advanced feature representations. A global average pooling layer is added to the last mixing module to reduce the dimension of the features. The softmax function is used at the end of the DCCMN model to obtain the output.

[0024] The global average is calculated as follows: ; in, yes c The output of each channel; is the input feature map in The value at H and W are the height and width of the feature map respectively; The calculation of the softmax function is as follows: ; in, represents the probability of each output, The sum of all is 1; K is the number of categories for the multi-classification problem; Represents the first Perform exponential operation on each element; Represents the first Perform exponential operation on the elements.

[0025] A small sample experimental study was conducted on the damage detection method. The small sample experimental study mainly includes the following parts: experimental objects and data, damage detection experiments and model comparison experiments. The deep learning framework used in the experiment is Tensorflow2.6.1 version, the computer is Intel Core i5-7300HQ, and the GPU is NVIDIA 1050Ti.

[0026] 1.1 Experimental subjects and data The present invention uses the floor frame structure built by Columbia University as the experimental object. The 3D schematic diagram of the frame structure is shown in FIG. Figure 3As shown in (a), it includes a main body with a bottom size of 2.5m×2.5m and a height of 3.6m. The entire frame structure is divided into four sides: east, south, west and north. Each side is composed of beams and columns with the same structural dimensions. In order to facilitate the identification of components in different directions, the same coding is used in the figure, such as "East 1", "North 1", etc.

[0027] A total of 15 acceleration sensors were placed on the frame structure. Three acceleration sensors were installed at the connection of each floor, located on the west side, center, and east side. Three sensors were placed close to the ground, and the remaining sensors were placed at the top of each floor. The specific sensor locations are as follows: Figure 3 (b) As shown. By placing sensors on the frame structure, vibration signal data can be obtained for damage diagnosis and experimental evaluation. These data will be used to train and test the proposed Meta+DCCMN method, and its performance will be verified and analyzed to verify the effectiveness and applicability of the Meta+DCCMN method in structural damage detection.

[0028] Through Figure 3 The diagonal supports and bolted connections numbered 1-12 in (a) were removed or loosened. The experiment simulated 9 different damage conditions, each of which was quite different from the other. The operation process is shown in Table 1.

[0029] Table 1 Specific damage conditions ;

[0030] In the experiment, the frame was destroyed in sequence according to the process shown in Table 1, and a 200Hz impact was applied to the frame to collect vibration signals. According to the data published by Columbia University, a total of 135 sets of data were obtained from 15 sensors under 9 different damage conditions. Among them, the signal length of Case 1-Case 5 is 24,000, the signal length of Case 6 is 60,000, and the signal length of Case 7-Case 9 is 72,000.

[0031] Since the length of the signal in each case is inconsistent, and in order to make the samples closer, the signal lengths of the above 9 cases were truncated during the experiment to make their lengths uniformly 24000. Then, in order to increase the number of samples, the sliding window overlapping sampling method was used to produce 100 samples in each case.

[0032] 1.2 Damage detection experiment In this experiment, 15 sensors in the framework structure were selected for research. Since the data of different sensor positions are quite different, in order to balance the characteristics of data at different positions, 10 sensors were selected as training tasks and the remaining 5 sensors were selected as test tasks. Specifically, the sensors of the test tasks are Figure 3 Numbers 4, 6, 8, 10, and 12 in (b).

[0033] In small sample learning, the classification problem is usually summarized as N-way K-shot, where N represents the number of categories and K represents the number of samples in each category, such as 5-way 3-shot means there are 5 categories, each of which contains 3 samples. The framework structure simulates 9 damage conditions, where case 1 and case 9 represent the intact state and the most severely damaged condition, respectively. For these two conditions, a 2-way K-shot experiment is designed, using the data of case 1 and case 9 as the support set and query set. In addition, a 9-way K-shot experiment is designed for all conditions, using the data of all conditions as the support set and query set. In this way, the small sample learning ability under different conditions is comprehensively evaluated, and the performance of the DCCMN model under different classification tasks is comprehensively analyzed. Through the above experimental design, we can have a more comprehensive understanding of the performance and applicability of the DCCMN model in small sample learning.

[0034] 1.2.1 2-way K-shot experiment In the training phase of the 2-way K-shot experiment, the number of inner and outer loops is set to 10, and the cross entropy loss function is used as the standard. The training task process is as follows Figure 4 As shown. Figure 4 As can be seen from (a), when the number of outer loops reaches 3, the accuracy of the support set and query set of the training task reaches more than 99%. Figure 4 As can be seen from (b), when the number of outer loops reaches 4, the loss function values ​​of the support set and the query set are 9×10 -3 and 6×10 -3 After each outer loop, the calculated average meta-loss function value stabilizes below 0.1. Overall, the DCCMN model shows a faster fitting speed and higher accuracy in the training task. This shows that the model can effectively learn and generalize when dealing with 2-classification problems, achieving better training results.

[0035] In the testing phase, take the results of sensor 6 in the test task as an example. First, the support set of sensor 6 is used to train the saved model for parameter fine-tuning. Here, 2-way 1-shot, 2-way 3-shot, and 2-way 5-shot experiments are used respectively. Then, the query set is used for testing. When the number of samples in the query set is too small, the accuracy may not truly reflect the performance and generalization ability of the model. The model may only perform well on the test set by chance or coincidentally predict a part of the samples correctly, rather than truly having a high accuracy. Relying solely on the accuracy of small-scale samples will bring uncertainty and bias. Therefore, for the query set of the test task, 20, 40, and 80 samples were used for detection, respectively, and 5 tests were performed in each case. The results are as follows: Figure 5 As shown in the figure, K1 represents the number of samples of each category in the support set of the test task (K1 = shot), K2 represents the number of samples in the query set of the test task, and the red values ​​represent the mean and standard deviation of the corresponding positions.

[0036] Depend on Figure 5 (a) It can be seen that the average accuracy of the support set is stable at 100% in multiple tests, which is Figure 5 (b) It can be seen that the loss function value of the support set can reach as low as 1×10-7. When K1=1, K2=20, the average accuracy of the query set is 98%; when K1=1, K2=80, the average accuracy of the query set is 95%. When K1=5, the average accuracy of the query set reaches 100%. From an overall analysis, as the number of K1 increases, the accuracy of the query set tends to increase, and the loss function value tends to decrease.

[0037] 1.2.2 9-way K-shot experiment In the training phase of the 9-way K-shot experiment, the number of inner loops is set to 15, the number of outer loops is set to 10, and the cross entropy loss function is used as the standard. The training task process is as follows Figure 6 As shown. Figure 6 As can be seen from (a), when the number of outer loops reaches 2, the accuracy of the query set in the training task reaches about 90%. Figure 6 As can be seen from (b), when the number of outer loops reaches 4, the loss function values ​​of the support set and the query set are 0.01 and 0.03 respectively. After each outer loop, the calculated average meta-loss function value is stable below 0.2. In general, the DCCMN model can achieve good training results when dealing with 9-classification problems.

[0038] In the test phase, the results of sensor 6 in the test task are also taken as an example. First, the saved model is trained with fine-tuning parameters using the support set of sensor 6. Here, 9-way 1-shot, 9-way 3-shot, 9-way 5-shot, and 9-way 10-shot experiments are used. Then, the query set is used for testing. The query set uses 20, 40, and 80 samples for detection, and 5 tests are performed in each case. K1 represents the number of samples of each category in the support set of the test task (K1 = shot), K2 represents the number of samples in the query set in the test task, and the red values ​​represent the mean and standard deviation of the corresponding positions.

[0039] Depend on Figure 7 (a) It can be seen that the average accuracy of the support set is basically stable at 100% in multiple tests, with the lowest being 97.8%. When K2=40, as the value of K1 increases, the average accuracy of the query set is 94.5%, 95.5%, 96.3%, and 97.75% respectively. When K2=80, as the value of K1 increases, the average accuracy of the query set is 90%, 93.25%, 97%, and 99% respectively. From an overall analysis, Figure 7 (b) It can be seen that as the number of K1 increases, the accuracy of the query set increases and the loss function value decreases.

[0040] 1.3 Model comparison experiment The experimental results of 2-way K-shot and 9-way K-shot show that the combination of meta-learning and DCCMN model method (Meta+DCCMN) has a high and stable accuracy, and can achieve damage detection with very few samples. In order to further verify the superiority of this method, this section conducts comparative experiments with a variety of other methods, including SVM, ResNet−18, DCCMN, Meta + ResNet−18, and Meta + DCCMN methods. Through comparative experiments, the performance of different methods can be comprehensively evaluated, and it can be verified whether the Meta+DCCMN method has better performance and accuracy in small sample learning.

[0041] The above experimental methods were used in the same conditions using the 9-way K-shot experimental steps. Since the ResNet-18 model is more suitable for larger data sets, the experimental data used two cases: K1 = 10, K2 = 80 and K1 = 50, K2 = 80. Each model was trained 20 times and tested 5 times in the training phase. The experiment recorded their accuracy on the test set or the test task query set, and the results are shown in Tables 2 and 3.

[0042] Table 2 K1 = 10, K2 = 80 model accuracy ;

[0043] Table 3. Model accuracy when K1 = 50 and K2 = 80 ;

[0044] Plot the average values ​​in the above table into a graph, such as Figure 8 As shown in the figure. Combined with the chart, it can be seen that when K1 = 10, the accuracy of SVM is 46%, and both ResNet-18 and DCCMN models are underfitting, with accuracy rates of 11.2% and 20.5% respectively. Compared with ResNet-18, the accuracy of Meta + ResNet-18 has little change, only increasing by 2%, while the accuracy of Meta + DCCMN can reach 99%. When K1 = 50, as the number of training samples increases, the accuracy of each model improves. The accuracy of SVM is 71%, the accuracy of ResNet-18 is 15.8%, and the accuracy of the DCCMN model increases by 76%, reaching 96.5%. The accuracy of Meta + ResNet-18 is 17.7%, while the accuracy of Meta + DCCMN is 100%.

[0045] By selecting one of the experimental results for each model under the experimental conditions of K1 = 50 and K2 = 80, the confusion matrix and TSNE visualization diagram are plotted to further analyze the performance of the model. These results are Fig. 9 and Fig.10 Displayed in. Fig. 9 (a)-(e) show the confusion matrix diagrams of the models SVM, ResNet-18, DCCMN, Meta+RestNet-18, and Meta+DCCMN respectively; Fig.10 (a)-(f) show the TSNE visualizations of the original dataset, model SVM, ResNet-18, DCCMN, Meta+RestNet-18, and Meta+DCCMN, respectively.

[0046] The confusion matrix provides a detailed evaluation of the model's classification performance. It shows the relationship between the true label and the model's predicted label, helping to understand the model's accuracy and error on different categories. Through the confusion matrix, you can intuitively observe the performance differences of the model on each category.

[0047] TSNE visualization can help present the clustering of data in two-dimensional space. It reduces the dimension of high-dimensional features and visualizes data points of different categories on a plane. By observing the TSNE graph, you can intuitively understand the distribution of samples of different categories and the similarities between samples. This helps to determine whether the model can effectively distinguish samples in a specific task.

[0048] By analyzing the confusion matrix and TSNE visualization graph, we can further evaluate the performance of each model under the conditions of K1 = 50 and K2 = 80, and compare their differences and advantages in classification tasks.

[0049] From the confusion matrix, it can be concluded that the SVM, DCCMN, and Meta + DCCMN methods perform well in classification. However, due to the complex structure of the ResNet−18 model, it does not perform well in the classification of small sample detection of vibration signals. The TSNE visualization results show that the original data is messy and the SVM model performs better in the classification of a few categories (such as label 1 and label 5). The clustering of the ResNet−18 model is poor both before and after adding meta-learning. In contrast, the DCCMN model clusters very well both before and after adding meta-learning, and is able to classify the data into clusters.

[0050] Therefore, the present invention adopts the above-mentioned small sample frame structure damage detection method based on meta-learning and DCCMN model to solve the problem of data scarcity in structural damage detection, improve the accuracy and generalization ability of small sample damage detection, and bring more reliable and efficient structural health monitoring solutions to the engineering field.

[0051] It is worth noting that the contents not elaborated in detail in the present invention are all prior art and are well known to those skilled in the art.

[0052] Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention rather than to limit it. 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 solution of the present invention with equivalents, and these modifications or equivalent replacements cannot cause the modified technical solution to deviate from the spirit and scope of the technical solution of the present invention.

Claims

1. A small sample frame structure damage detection method based on meta-learning and DCCMN model, characterized in that: The following steps are involved: Step S1, in the data processing stage, vibration signals are collected and stored according to the sensor number, and then the data is normalized. After data processing, the data is divided into training tasks and test tasks; Step S2: In the model training phase, a task is randomly selected from the training tasks, and the support set of the selected task is input into the DCCMN model for training to complete the inner loop; Step S3: Use the query set to evaluate the model, calculate the loss function, and update the model parameters to complete the outer loop; In the outer loop, the model parameters are further updated by calculating the average loss function and gradient of multiple tasks, and the updated DCCMN model is saved after the set number of loops. Step S4: In the model testing phase, the support set is used to fine-tune the parameters of the saved DCCMN model, and then the query set is used to test the DCCMN model after parameter fine-tuning, and finally the detection result is obtained.

2. According to claim 1, the small sample frame structure damage detection method based on meta-learning and DCCMN model is characterized in that: The calculation process of the inner loop parameter update in step S2 is as follows: ； in, Indicates the parameters updated by the inner loop; represents the initial parameters; Represents the learning rate of the inner loop; Represents the loss function of the support set in the training task, and the loss function adopts the cross entropy loss function , the formula is as follows: ； in, N Indicates the number of samples in the support set; Indicates The true labels of samples; Indicates The predicted label of the sample, that is, the model's prediction of The probability distribution of samples belonging to each category.

3. The small sample frame structure damage detection method based on meta-learning and DCCMN model according to claim 2 is characterized in that: The parameter update calculation process of the outer loop in step S3 is as follows: ； in, The updated parameters of the outer loop; Represents the learning rate of the outer loop; represents the meta-loss function; Represents the gradient accumulation of all training tasks; E Indicates the number of times the outer loop is executed.

4. The small sample frame structure damage detection method based on meta-learning and DCCMN model according to claim 3 is characterized in that: The calculation process of the parameter fine-tuning stage in step S4 is as follows: ； in, represents the fine-tuned parameters in the test task; Represents the learning rate for parameter fine-tuning; represents the loss function of the support set in the test task.

5. The small sample frame structure damage detection method based on meta-learning and DCCMN model according to claim 1 is characterized in that: In the DCCMN model, the convolution kernel size of the first convolution layer is set to 16, the step size is set to 4, the convolution kernel size of the second convolution layer is set to 8, the step size is set to 2, and the BN layer and the maximum pooling layer are added after each convolution layer. Then the cross-mixing module is gradually introduced for feature fusion. Four cross-mixing modules are added to the DCCMN model, and a global average pooling layer is added to the last mixing module to reduce the dimension of the features. Finally, the softmax function is used to obtain the output of the DCCMN model.

6. The small sample frame structure damage detection method based on meta-learning and DCCMN model according to claim 5 is characterized in that: The global average is calculated as follows: ； in, yes c The output of each channel; is the input feature map in The value at and are the height and width of the feature map respectively; The calculation of the softmax function is as follows: ； in, Represents the probability of each element output, The sum of all is 1; K is the number of categories for the multi-classification problem; Represents the first Perform exponential operation on each element; Represents the first Perform exponential operation on the elements.

Images