Damage detection method based on deep learning

Abstract

To provide a method for predicting a strain distribution map and determining damage based on deep learning.SOLUTION: The present invention comprises: a step S1 of establishing an image dataset of finite element analysis results for strain distribution map prediction; a step S2 of building a deep learning model for strain distribution map prediction based on a DeepLabv3+ network and performing learning and validation; a step S3 of establishing a dataset for damage determination of a structural analysis model and performing preliminary processing and data enhancement operations; a step S4 of building a binary classification deep learning model for damage determination based on a convolutional neural network and performing learning and validation for transition learning; and a step S5 of performing an interpretability analysis on the learned binary classification structural analysis model and outputting a region in an image that more contributes to classification.SELECTED DRAWING: Figure 1

Description

[Technical Field]

[0001] The present invention belongs to the field of computer vision-aided mechanical property analysis, and in particular to a method for strain map prediction and damage assessment based on deep learning. [Background technology]

[0002] Numerical simulation is an important tool used to analyze and evaluate the mechanical properties of various structures and components. To obtain accurate structural analysis results, an expert in finite element analysis typically needs to execute a workflow of "modeling-computer analysis-result output-manual evaluation." As the number of combinations of structural analysis model parameters increases, the time required for finite element analysis doubles and the demands on computer performance gradually increase. At the same time, modeling and condition evaluation require repeated work by experienced engineers, which poses many limitations to the use of finite element analysis for mechanical analysis of multi-parameter combined structural analysis models. Deep learning methods can fully utilize existing finite element analysis data to obtain highly accurate mechanical analysis results and condition evaluation results. Therefore, in this study, strain distribution map prediction and damage evaluation are performed based on deep learning and neural network methods.

[0003] Currently, research into using deep learning techniques as a rapid replacement for finite element analysis (FEA) is still in its early stages of development. Some researchers have used methods such as fully convolutional neural networks to predict stress distribution in two-dimensional linear elastic material components, achieving a mean relative error (MRE) of only 2.04%. However, these methods only output strain and strain distribution results, requiring manual determination of the strain and strain distribution state, and lacking a subsequent automated analysis step. More research has been done on component damage assessment using neural network methods. Typically, image data of various types of damage is collected and information about each state is assigned. Then, analysis and research are performed using image classification neural networks, ultimately yielding the probability that the input image belongs to each damage state. However, the assessment is often considered a "black box," leaving engineers uncertain about the assessment results and structural principles, and further improvement is needed.

[0004] To solve the above problems, the present invention proposes a fast method for strain distribution map prediction, component state assessment analysis, and its interpretability analysis based on deep learning. This method inputs structural analysis model data for a new parameter combination and outputs the resulting strain distribution map. The strain distribution map under that state is then input into a damage assessment neural network, which outputs the damage state under that load. Finally, a heat map is used to identify areas with high contribution to the assessment of the structural analysis model. The results demonstrate that this method can quickly and accurately obtain the strain distribution and failure assessment results of a component under a certain load, and output an analytical heat map that can be interpreted according to the theory of fracture mechanics. Summary of the Invention [Problem to be solved by the invention]

[0005] The present invention provides an automated method for strain distribution imaging and damage assessment based on deep learning to reduce the consumption of time and labor resources in the finite element analysis process of similar structural analysis models with many parameter combinations and to provide efficient and accurate analysis results of mechanical properties. [Means for solving the problem]

[0006] The technical problem to be solved by the present invention is achieved by using the following technical solutions: A damage determination method for predicting a strain distribution map based on deep learning and determining damage based on the strain distribution map, It comprises the following steps: Step S1: Using finite element analysis software to calculate structural analysis models with multiple variable parameters, obtaining strain distribution maps of each structural analysis model under all loading steps, and dividing the sample data set into a training set, a validation set, and a test set; Step S2: Using a convolutional neural network for area classification (e.g., DeepLabv3+ network), a deep learning model for strain distribution map prediction is constructed, and training and validation are performed. The geometric information, material information, and load information of the structural analysis model are input, and the strain distribution map is output. Step S3: Using the strain distribution map obtained by the finite element analysis in step S1 as an input, and adding information on the destruction or non-destruction of the structural analysis model in the state of the strain distribution map, thereby forming a damage assessment data set, performing a preliminary balancing operation and a data strengthening operation on the data set, and then dividing the sample data set into a training set, a validation set, and a test set; Step S4: constructing a damage determination deep learning model using a convolutional neural network (for example, a VGG-16 network, which is a convolutional neural network with a depth of 16 layers), performing learning and validation, inputting a strain distribution map of a structural analysis model in a certain state, and outputting the probability that damage is currently determined; The method includes step S5 of performing an interpretability analysis of the gradient-weighted class activation mapping (Grad-CAM) method on the deep learning model for two-class classification obtained by learning in step S4, obtaining the contribution distribution of the neural network model to the prediction results, and highlighting image regions with a large classification contribution.

[0007] In the deep learning-based strain distribution map prediction and damage assessment method, in step S1, the process of acquiring and processing a strain distribution map data set includes the following steps: Step S1.1: outputting the geometrical figure and strain distribution map of the structural analysis model through finite element analysis software, and outputting the geometrical figure and strain distribution map in a uniform, proportional square image; Step S1.2, keeping the size of the image constant, and showing the non-geometric information, i.e., material information and load information, as a grayscale image filled with a certain multiple value, thereby achieving the purpose of characterizing the non-geometric information with grayscale values; Step S1.3 includes dividing the dataset into a training set, a validation set, and a test set.

[0008] In the method for strain distribution map prediction and damage assessment based on deep learning, in step S2, the process of constructing a deep learning model for strain distribution map prediction includes the following steps: Step S2.1: constructing a deep learning model based on a convolutional neural network for area classification, which inputs geometric information, material information, and load information of a structural analysis model and outputs a plane strain distribution diagram; In the strain distribution map prediction problem, the following process is performed: S2.2, in which the cross-entropy function is used as the loss function:

number

number

[0009] In the method for strain distribution map prediction and damage assessment based on deep learning, in step S3, the step of strengthening the unbalanced data set of the damage assessment category and pre-balancing includes the following steps: Step S3.1, performing data enhancement operations on the image dataset, specifically including fuzzification, sharpening, horizontal flipping, vertical flipping, 90° rotation, and 270° rotation; and step S3.2 of randomly sampling images of categories with more samples to pre-balance the difference in numbers between different categories in the dataset.

[0010] In the method for strain distribution map prediction and damage assessment based on deep learning, in step S4, the process of constructing a two-class classification deep learning model for damage assessment includes the following steps: A step S4.1 of constructing a deep learning model that inputs a strain distribution map of a target to be judged based on a convolutional neural network and outputs a judgment result; For the problem of two-class classification with imbalanced categories in the training data, the Focal Loss function is used as the loss function in the training process, and the following formula is used in step S4.2:

number

number

[0011] In the method for strain distribution map prediction and damage assessment based on deep learning, in step S5, the interpretability of the deep learning model for damage assessment is analyzed as follows: using the gradient-weighted class activation mapping (Grad-CAM) method, the distribution of the contribution of the neural network model to the predicted output is obtained; that is, for strain distribution maps determined to be "damage-occurring," the locations having the characteristics of the strain distribution map are highlighted by color in step S5.1. [Effects of the Invention]

[0012] Compared with the prior art, the advantages and beneficial effects of the present invention are as follows: (1) In the strain distribution map prediction module, the present invention introduces a method of using grayscale images to characterize non-geometric information such as the material and load of the structural analysis model. By integrating a theoretical structural analysis model based on semantic segmentation with the mechanical analysis of components, the present invention realizes the output of strain distribution results by inputting the initial operating state of the structural analysis model to be analyzed. (2) In the present invention, in the processing of the damage pattern determination module, a transition learning method is used for learning, a focal loss function is introduced, and F1 value, recall rate, and accuracy are introduced as quantitative evaluation indices. (3) Based on image data and deep learning methods, this invention proposes an automated complete workflow including strain distribution map prediction, damage mode determination, and structural analysis model interpretability analysis, which improves the efficiency and reliability when facing multi-parameter combination tasks compared with traditional finite element analysis methods. [Brief explanation of the drawings]

[0013] [Figure 1] 1 is a specific flowchart of strain distribution map prediction and damage determination based on deep learning. [Figure 2] Schematic diagrams of a typical finite element structural analysis model and loading method, where (a) is a front view, (b) is a side view, (c) is a plan view, and (d) is a perspective view. [Figure 3] 10A and 10B show an embodiment of a data image set in a dataset, where (a) is mortar, (b) is rebar, (c) is load, (d) is an internal cross-sectional strain distribution map, and (e) is a front surface strain distribution map. [Figure 4] This is a diagram of the network architecture of DeepLabv3+. It consists of an encoder module and a decoder. Here, the main body of the encoder module is a deep convolutional neural network (DCNN) with astronomical convolution, followed by a multi-scale pyramid pooling module that enables feature pooling at different scales. Finally, the decoder restores the resolution of the initial image and fuses low-level features, which are usually upsampled using techniques such as bilinear interpolation or transposed convolution, to finally obtain the output image. [Figure 5A]This is a curve showing the change in loss value during the training and validation processes of the strain distribution prediction network. Figure 5A shows the predicted relationship between the initial conditions and the strain distribution on the front surface of the structural analysis model. In the figure, the abscissa (Epoch) represents the iteration round, and the ordinate (Cross Entropy Loss) represents the cross-entropy loss. FCN-Train is the training set for the fully convolutional neural network, FCN-val is the validation set for the fully convolutional neural network training process, DeepLabv3+-Train is the training set for the DeepLabv3+ neural network, and DeepLabv3+-val is the validation set for the DeepLabv3+ neural network training process. [Figure 5B] This shows the loss curves during the strain distribution prediction network training and validation processes. Figure 5B shows the initial conditions and the predicted internal cross-sectional strain distribution of the structural analysis model. In the figure, the abscissa (Epoch) represents the iteration round, and the ordinate (Cross Entropy Loss) represents the cross-entropy loss. FCN-Train is the training set for the fully convolutional neural network, FCN-val is the validation set for the fully convolutional neural network training process, DeepLabv3+-Train is the training set for the DeepLabv3+ neural network, and DeepLabv3+-val is the validation set for the DeepLabv3+ neural network training process. [Figure 6] This figure shows the strain distribution prediction results for three conditions of samples randomly selected from the test set. The load magnitude is indicated by the displacement load. FEA(S) refers to the finite element analysis results of the sample's internal cross-sectional strain distribution, DeepLabv3+(S) refers to the DeepLabv3+ prediction results of the sample's internal cross-sectional strain distribution, FCN-8s(S) refers to the fully convolutional neural network prediction results of the sample's internal cross-sectional strain distribution, FEA(F) refers to the finite element analysis results of the sample's front surface strain distribution, DeepLabv3+(F) refers to the DeepLabv3+ prediction results of the sample's front surface strain distribution, and FCN-8s(F) refers to the fully convolutional neural network prediction results of the sample's front surface strain distribution. [Figure 7] This is a diagram of the VGG-16 network architecture suitable for two-class classification. The input is a frontal distortion distribution map of a sample with a height of 224 and a width of 224 in a certain state. Five modules undergo convolution-max pooling operations. The first and second modules each contain two convolution layers and a max pooling layer. The third to fifth modules each contain three convolution layers and a max pooling layer. These are followed by three fully connected layers. Finally, there is a SoftMax classification output layer, whose output is the posterior probability of classifying the sample in that state as corrupt or non-corrupt. [Figure 8] This shows the results of an interpretability analysis of the Grad-CAM method for a damage assessment structural analysis model. DETAILED DESCRIPTION OF THE INVENTION

[0014] In order to facilitate understanding of the technical means, creative features, and objectives and effects achieved by the present invention, the present invention will be described in more detail below in conjunction with specific embodiments and drawings.

[0015] In a specific embodiment, the present invention provides a deep learning-based method for strain distribution map prediction and damage assessment, which includes the following steps: (1) establishing an image dataset of finite element analysis results for strain distribution map prediction; (2) building a deep learning model for strain distribution map prediction based on a DeepLabv3+ network (a convolutional neural network for region classification) and conducting training and validation; (3) establishing a dataset for sample damage assessment and conducting preliminary processing and data enhancement operations; (4) building a two-class classification deep learning model for damage assessment based on a VGG-16 network (a 16-layer deep convolutional neural network) and conducting training and validation for transition learning; (5) conducting an interpretability analysis of the trained two-class classification model and outputting the regions in the image that contribute to classification; and (6) rapidly and accurately obtaining the strain distribution and failure assessment results for a component subjected to a certain load and outputting interpretable analysis results consistent with fracture mechanics theory.

[0016] PyTorch is an open-source Python machine learning library. Both the DeepLabv3+ and VGG-16 neural networks in this invention were built on a server running the Ubuntu system. All training, validation, and interpretable analysis were completed on a graphics processing unit (GPU). The specific hardware and software version models are shown in Table 1 (environment configuration table), i.e., [Table 1]

[0017] The analytical embodiment is a mechanical property analysis problem based on finite element analysis of the mechanical anchorage method of reinforcing steel bars. To study the factors affecting the tensile strength of the samples, the mechanical properties of 72 samples with different parameter combinations were analyzed under a displacement loading rule. The basic finite element structural analysis model and loading locations are shown in Figure 2. The basic finite element structural analysis model is formed by combining three reinforcing steel bars with anchor plates 1 arranged parallel from top to bottom with high-performance fiber reinforced mortar 2 (HPFRM). Each reinforcing steel bar with anchor plates 1 is an integrally connected structure consisting of a reinforcing steel rod section 11 and a reinforcing steel anchor plate 12. The reinforcing steel rod section 11 of the top and bottom two reinforcing steel bars with anchor plates 1 extends to the left of the HPFRM 2, and the reinforcing steel rod section 11 of the middle reinforcing steel bar with anchor plate 1 extends to the right of the HPFRM 2. The boundary conditions of the finite element model (structural analysis model) are that the x, y, and z directions of both extension ends of the two reinforcing bars on the left side are constrained, the z direction of the top and bottom of the structural analysis model is constrained, and the y direction of the cross section within the structural analysis model is constrained.

[0018] As shown in FIG. 1, the present invention provides a method for strain distribution map prediction and damage determination based on deep learning, and the specific steps are as follows:

[0019] Step 1: Acquire and process the strain distribution dataset (1) Using the finite element analysis software DIANA, the geometrical shape and strain distribution diagram of the sample of the mechanical anchorage of the rebar are output. As shown in Table 2 (embodiment of sample parameter combinations), there are 72 cases with combinations of four types of parameters (rebar diameter, rebar spacing, rebar lap length, and filling mortar fracture energy). Each parameter combination of the sample in the loading process has strain distribution diagrams for approximately 20 to 50 loading steps, and is output uniformly as a square image with the same proportions. [Table 2]

[0020] (2) For one set of data in the dataset, the information on the rebar diameter, rebar spacing, and rebar lap length in Table 2 can be displayed as geometric figures. For the non-geometric information of the mortar fracture energy, the value is multiplied by 10 and converted into a grayscale value, which is then filled into the mortar position for characterization. For the load magnitude information (load information), the value is multiplied by 10 according to the displacement load value to generate a solid grayscale image. Finally, 3,239 sets of images are generated, including mortar information, rebar information, load information, front strain distribution maps, and internal cross-section strain distribution maps, as shown in Figure 3.

[0021] (3) The image size is 704 x 704, and the images are divided into a training set, a validation set, and a test set in a ratio of 8:1:1.

[0022] Step 2: Building a deep learning model for strain distribution map prediction (1) A network construction architecture based on DeepLabv3+ is selected, and the sample's geometric information, material information, and load information are input to output a plane strain distribution map. As shown in Figure 4, the network structure diagram of the algorithm shows that the overall architecture of the DeepLabv3+ model consists of an encoder module and a decoder module. The encoder module itself is a deep convolutional neural network with astrace convolution, and features are pooled at different scales by a multi-scale pyramid pooling module. Finally, the decoder module restores the resolution of the original image.

[0023] (2) The backbone network Xception extracts multi-scale features from part of the results by parallel dilated convolution, and the other part is sent directly to the decoder. In the decoder, the multi-scale features are bilinearly upsampled and fused with backbone low-level features of the same scale. The fused feature information is further processed by a 3x3 convolution kernel and bilinear upsampling to select the optimal feature and restore the original image size. In the strain distribution map prediction problem, the cross-entropy function is used as the loss function in the following equation:

number

number

[0024] (3) Training the neural network model. Two neural network models were trained in parallel to establish the relationship between the initial conditions and the strain distributions on the front and inner cross sections of the sample, respectively, and the change in the loss value during the training process. See Figures 5A and 5B. The initial learning rate was 0.001 and was adjusted to 0.0001 after 50 iterations, for a total of 100 iterations. As shown in Figure 6, the strain distribution prediction results for three conditions of a randomly selected sample from the test set showed that both DeepLabv3+ and the fully convolutional neural network achieved prediction results similar to those of the finite element analysis, with DeepLabv3+ demonstrating a clear advantage over the fully convolutional neural network.

[0025] (4) By diversifying the evaluation indexes, the evaluation situation of the structural analysis model can be reflected from various perspectives. In addition to the average cross-linking, pixel accuracy and weighted average cross-linking are also introduced and evaluated together. A fully convolutional neural network (FCN) is also introduced for comparison. The final verification results are shown in Table 3 (Quantitative analysis of the strain distribution map prediction effect), i.e., [Table 3] As can be seen from Table 3, compared to the fully convolutional neural network model, the average cross-connection index predicted using the DeepLabv3+ network can be improved by up to 0.2135. This shows that the use of the method proposed by the present invention can significantly improve the performance of the structural analysis model, proving the effectiveness of the method proposed by the present invention.

[0026] Step 3: Processing and enriching the unbalanced dataset of damage assessment categories (1) The information for damage assessment was mainly derived from the strain distribution map on the front surface of the sample. Information was assigned according to the conical damage state of the sample corresponding to the strain distribution map. A total of 355 samples were assigned the information of “damage occurred” and 2,891 samples were assigned the information of “no damage occurred.”

[0027] (2) The image data is augmented by fuzzification, sharpening, horizontal flipping, vertical flipping, 90° rotation, and 270° rotation, resulting in a total of 22,722 sets of strain distribution map data, and the size of all images is unified to 224 × 224.

[0028] (3) Because this dataset has an obvious category imbalance problem, in order to give the neural network reliable classification ability, the number of “non-damaged” samples, which has a large number of samples, was reduced to 11,564, while the number of “damaged” samples remained at 2,485 pairs, achieving a preliminary balance of the amount of category data.

[0029] Step 4: Building a two-class classification deep learning model for damage detection (1) Based on the VGG-16 network, a deep learning model is constructed that inputs the distortion distribution map of the object to be judged and outputs the judgment result. The structure of the VGG-16 network used for the two-class classification task is shown in Figure 7. As can be seen from the structural diagram of the VGG-16 network suitable for two-class classification, the input to the deep learning model is the front distortion distribution map of a 224 x 224 sample in a certain state, which passes through five modules including convolution and pooling operations and three fully connected layers, and is finally classified using SoftMax, and the output is the posterior probability of belonging to destructive or non-destructive.

[0030] (2) Because the problem of category imbalance in the training data still exists, the Focal Loss (FL) function is used as the loss function in the training process, and the following formula is used:

number

[0031] (3) The initial learning rate is 0.001, and a total of 20 iteration rounds are performed using pre-running weights on ImageNet. The neural network model is trained using the transition learning method, which significantly reduces data collection costs and improves the model fitting rate.

[0032] (4) By diversifying the evaluation indices, the evaluation status of the structural analysis model can be reflected from various angles, and the following formula is used with the F1 value, recall rate, and accuracy as the evaluation indices:

number

[0033] Step 5: The interpretability analysis process of damage determination using deep learning model is as follows: (1) Extract the frontal strain distribution map in the test set that is not involved in the learning and validation in step 4 and perform an interpretability analysis test.

[0034] (2) As shown in Figure 8, the Gradient Weighted Class Activation Mapping (Grad-CAM) method can be used to determine the contribution distribution of the neural network model to the predicted output. That is, for strain distribution maps that are determined to be damaged, the locations of the features of such strain distribution maps are highlighted in red, and areas of the image that contribute less to the classification are filled in with a bluish color.

[0035] Furthermore, the interpretability of the two-class classification neural network model was analyzed using the Grad-CAM method.The regions that contributed most to the judgment results were consistent with the fracture mechanics principle in the axial tension process of the mechanical anchor for rebar, that is, conical damage is likely to occur when the base angle of the conical strain region is close to the mortar surface.

[0036] The results show that, compared with existing numerical simulation methods, the deep learning-based strain distribution map prediction and damage assessment method provided by the present invention can save a lot of time and human resources for repeated analysis, and at the same time, can obtain analysis results for the strain distribution map and damage assessment of the sample that are highly useful as reference.

[0037] The basic principles and processes, main features and advantages of the present invention have been shown and described above. The present invention is not limited to the above embodiments. The above embodiments and the description of the specification are merely illustrative of the principles of the present invention. It should be understood that various modifications and alterations within the scope of the protection claims of the present invention exist without departing from the spirit and scope of the present invention. The scope of the protection claims of the present invention is defined by the appended claims and their equivalents. [Explanation of symbols]

[0038] 1 Reinforced concrete with anchor plate 11 Reinforced concrete rod 12 Rebar anchor plate 2. Ultra-high performance fiber reinforced mortar

Claims

1. A damage determination method for predicting a strain distribution map based on deep learning and determining damage based on the strain distribution map, It comprises the following steps: Step S1: Using finite element analysis software to calculate structural analysis models with multiple variable parameters, obtaining strain distribution maps of each structural analysis model under all loading steps, and dividing the sample data set into a training set, a validation set, and a test set; Step S2: constructing a deep learning model for predicting a strain distribution map using a convolutional neural network for region classification, performing learning and validation, inputting diagrams of geometric information, material information, and load information of a structural analysis model, and outputting the strain distribution map; Step S3: using the strain distribution map obtained by the finite element analysis in step S1 as an input, adding information on whether the structural analysis model is destructive or non-destructive in the state of the strain distribution map, thereby forming a damage assessment dataset, performing a preliminary balancing operation and a data strengthening operation on the dataset, and then dividing the sample dataset into the training set, the validation set, and the test set; Step S4: constructing a deep learning model for damage determination using a convolutional neural network, performing learning and verification, inputting a strain distribution diagram of a structural analysis model in a certain state, and outputting the probability that damage is currently determined; and step S5 of performing an interpretability analysis of the gradient-weighted class activation mapping (Grad-CAM) method on the two-class classification model obtained by learning in step S4, obtaining a contribution distribution of the neural network model to the prediction result, and highlighting image regions with a large classification contribution. A damage assessment method characterized by:

2. In step S1, the process of acquiring and processing the strain distribution data set includes the following steps: Step S1.1: outputting a geometric information diagram and a strain distribution diagram of the structural analysis model through the finite element analysis software, and outputting the geometric information diagram and the strain distribution diagram in a unified square image of the same proportion; Step S1.2, keeping the size of the image constant, and showing the non-geometric information, i.e., material information and load information, as a grayscale image filled with a certain multiple value, thereby achieving the purpose of characterizing the non-geometric information with grayscale values; S1.3, dividing the dataset into the training set, the validation set, and the test set. The damage determination method according to claim 1 .

3. In step S2, the construction process of deep learning for predicting the strain distribution map includes the following steps: Step S2.1: constructing a deep learning model based on a convolutional neural network for area classification, which inputs geometric information, material information, and load information of a structural analysis model and outputs a plane strain distribution diagram; In the strain distribution map prediction problem, the following process is performed: S2.2, in which the cross-entropy function is used as the loss function: [Equation 1] where p t is the probability that the prediction is true, W is the width of the image, H is the height of the image, and p t , i , j is the probability that the prediction is correct at image position (i, j), In the strain distribution map prediction problem, step S2.3 uses the average cross-connection MIoU as the evaluation index in the following formula: [Equation 2] where k is the number of color gradations in the strain distribution map, i is the color ordinal number of the finite element analysis result of the pixel, j is the color ordinal number of the neural network prediction result of the pixel, and p ij is the number of pixels that incorrectly predicted color i as color j, and step S2.4 of adjusting the learning hyperparameters, training the neural network model, and verifying a better combination based on the prediction results. The damage determination method according to claim 1 .

4. In step S3, the step of strengthening and pre-balancing the unbalanced data set of the damage determination category includes the following steps: Step S3.1, performing data enhancement operations on the image dataset, specifically including fuzzification, sharpening, horizontal flipping, vertical flipping, 90° rotation, and 270° rotation; and step S3.2 randomly sampling images of categories with more structural analysis models to pre-balance the difference in numbers between different categories in the dataset. The damage determination method according to claim 1 .

5. In step S4, the process of constructing a two-class classification deep learning model for damage determination includes the following steps: A step S4.1 of constructing a deep learning model that inputs a strain distribution map of a target to be judged based on a convolutional neural network and outputs a judgment result; In the problem of classification with imbalanced categories in the training data, the Focal Loss function is used as the loss function in the training process, and the following equation is used in step S4.2: [Equation 3] where p t is the probability that the prediction is true, and α t are category weights that can be used to balance the contributions of different categories, γ is a focusing parameter that controls the rate at which the loss is concentrated on difficult-to-classify samples, Step S4.3 uses precision, recall, and F1 score as evaluation indices in the following formula: [Equation 4] Here, TP is the number of times that a strain distribution map with damage was correctly predicted as a strain distribution map with damage, FP is the number of times that a strain distribution map with no damage was correctly predicted as a strain distribution map with no damage, and FN is the number of times that a strain distribution map with damage was incorrectly predicted as a strain distribution map with no damage, and step S4.4 of adjusting the learning hyperparameters, training the neural network model using a transition learning method, and verifying a better combination based on the prediction results. The damage determination method according to claim 1 .

6. In step S5, the process of analyzing the interpretability of the deep learning model for damage determination is as follows: The damage assessment method according to claim 1, characterized in that step S5.1 uses the gradient-weighted class activation mapping (Grad-CAM) method to obtain the distribution of the contribution of the neural network model to the predicted output, i.e., for a strain distribution map in which damage is determined to have occurred, positions having characteristics of the strain distribution map are highlighted by coloring.

Images