Structure low-cycle fatigue life prediction method based on image and deep learning method

Abstract

The invention belongs to the technical field of material life prediction, and particularly relates to a structure low-cycle fatigue life prediction method based on an image and a deep learning method, which comprises the following steps of: 1, selecting a target structure material, preparing DIC speckles on the surface of the material, and setting N1 groups of low-cycle fatigue test loading conditions; and continuously shooting surface deformation images of the selected area by using an industrial camera to obtain speckle images. According to the structure low-cycle fatigue life prediction method based on the image and the deep learning method, structure low-cycle fatigue life prediction based on any strain image can be realized according to the prediction model provided by the invention only through image data set construction and convolutional neural network model training. According to the method, the structural fatigue life information can be obtained directly through image acquisition, the method is simple and convenient, the method has the advantages of intelligent calculation, real-time result output and the like, and the prediction accuracy is high.

Description

Technical Field

[0001] The present invention relates to the technical field of material life prediction, and in particular to a method for predicting structural low-cycle fatigue life based on image and deep learning methods. Background Art

[0002] In the engineering field, structural materials for heavy-load, high-speed, and lightweight service are very prone to low-cycle fatigue damage. To prevent sudden failure, accurate description of their fatigue life is the key issue and difficulty to ensure reliable service. In traditional methods, for the evaluation of the low-cycle fatigue life of materials, fatigue tests are required to establish the strain-life (ε-N curve) relationship, or to use empirical formulas such as Manson-Coffin for estimation. The former is often time-consuming and costly, while the latter has the disadvantages of numerous parameters and hysteresis. In the context of big data and artificial intelligence, machine learning methods spawned by data science have been applied in the field of low-cycle fatigue life prediction. Machine learning can quickly and automatically process large amounts of data and adaptively learn the relationship between input and target output parameters, thereby achieving accurate prediction of target parameters. The deep learning framework developed from this performs better in processing multi-source data tasks.

[0003] In the prior art, for example, a method for predicting low-cycle fatigue life based on deep learning, which is published on the Chinese patent website with the announcement number CN 117725846 B, is used to obtain the stress-strain data of metal materials and then process the stress data using a convolutional neural network to achieve material fatigue strength prediction. However, in actual application, the collection of structural material strain data is time-consuming and costly, the difficulty of collection is greatly affected by the service environment, and the accuracy is difficult to guarantee. Therefore, the low-cycle fatigue life prediction of materials in the prior art is prone to the problems of insufficient accuracy, poor economy and reliability due to inaccurate collection of stress-strain data and insufficient sample size. Summary of the invention

[0004] Based on the technical problems of insufficient accuracy, economy and reliability in describing the low-cycle fatigue life of existing materials, the present invention proposes a method for predicting the low-cycle fatigue life of structures based on image and deep learning methods.

[0005] The present invention proposes a structural low-cycle fatigue life prediction method based on image and deep learning methods, including step one, selecting a target structural material, then preparing DIC speckles on the surface of the material, and setting N1 group low-cycle fatigue test loading conditions.

[0006] Step 2: During the low-cycle fatigue test, an industrial camera is used to continuously capture the surface deformation images of the selected area to obtain a speckle image.

[0007] Step 3: Use the DIC calculation module VIC-2D to calculate the image. The calculation parameters set include the sub-area radius r, the sub-area step size s, and calculate the average strain ε of each speckle image. ij , which is the average strain value corresponding to the jth image under the i-th group of loading conditions.

[0008] Step 4: For the images obtained under each set of loading conditions, sample and stack the images according to the DIC calculation results, construct an "image-fatigue life" data set, and divide it into a training set and a test set.

[0009] Step 5. Build a three-dimensional convolutional neural network model based on the Pytorch framework.

[0010] Step 6: Call the model, process its unknown arbitrary speckle image sequence according to the data format required by its input layer, and then input it into the model for low-cycle fatigue life prediction.

[0011] Preferably, the loading condition settings of the low cycle fatigue test in step 1 include stress ratio R, loading waveform, frequency F (Hz), and stress amplitude range σ1-σ2 (MPa).

[0012] Assuming the shooting interval of the industrial camera is T1 (ms), the number of deformation speckle images that can be collected in each test cycle is

[0013] Considering the device memory and data processing task volume, it is set to collect deformation speckle images for N5 consecutive test cycles every N4 test cycles.

[0014] Through the above technical solution, an industrial camera is used for image capture, and the setting parameters include camera exposure time, number of test cycles and intervals for capture.

[0015] Preferably, in step 3, under each set of loading conditions, the final fatigue life N of the material is i The system is divided into I equal parts, and based on the result of the average strain calculation, a strain peak value at a certain point of the material is randomly selected in each equal part of the life span interval, and N6 speckle images corresponding to both sides of the peak value are obtained.

[0016] I*N6 deformation speckle images were sampled under each group of loading conditions, and a total of N1*I*N6 deformation speckle images were sampled under all loading conditions. All sampled images were cropped using a Python script and uniformly processed into L*W pixels to ensure consistent sample size during subsequent data processing.

[0017] Preferably, in step 4, by defining the tensor sequence to be connected and the stacking dimension, the N6 images sampled from each equal life span interval and the 1 original image are stacked according to the time dimension of the experimental process, and the obtained tensor shape is (I*N6+1)*L*W.

[0018] Preferably, for each set of loading conditions, the low cycle fatigue life N of the material is i After normalization, the formula is:

[0019]

[0020] Among them, N f is the dimensionless low-cycle fatigue life characteristic value after normalization, and its value range is 0-1; N i 、N min and N max They are the material low cycle fatigue life test value, minimum value and maximum value respectively.

[0021] After the normalization, the “image-fatigue life” data set is constructed (i=1,2,…,N1), divide the training set into proportions and test set

[0022]

[0023] Where M and N are the number of training and testing samples respectively, and M+K=N1.

[0024] Through the above technical scheme, the original undeformed image of the structural material under the same loading condition and the sampled deformed image are combined into a sequence to obtain the same number of image sequences as the number of loading condition groups, and then the low-cycle fatigue life data corresponding to each group are normalized.

[0025] Based on the aforementioned image sequence and normalized life data, an "image-fatigue life" data set is constructed and divided into a training set and a test set according to a ratio of 4:1.

[0026] Preferably, in step five, the model architecture includes an input layer, a convolutional layer, a pooling layer, a fully connected layer and an output layer.

[0027] Through the above technical solution, the model hyperparameters are initially configured, including the number of network layers and units in each layer, convolution kernel size, Dropout ratio, activation function, learning rate, number of iterations, optimizer, and loss function. The determination coefficient can be used to evaluate the accuracy of the model in predicting low-cycle fatigue life. Combined with the data set, the model iteration is completed, thereby obtaining the best convolutional neural network parameter configuration results and saving the model.

[0028] Preferably, the data received by the input layer is a color digital image based on an RGB (red, green, and blue) color model, so the number of channels of the model input layer is set to 3*(I*N6+1)*L*W.

[0029] Preferably, the convolution layer is set to N C Layer, each layer has a convolution kernel size of K ci and step length S ci (i=1,2,…,N c ), at the same time, the number of channels in each layer can be doubled layer by layer according to the rule of thumb, and the output of each convolutional layer is:

[0030]

[0031] Among them, Q i-1 is the output of the previous layer, that is, the input of this layer. For the first convolutional layer, its Q i-1 It is the input tensor of the entire model, that is, 3*(I*N6+1)*L*W.

[0032] Preferably, the feature information processed by the convolution layer is passed to the pooling layer, and the global average pooling torch.mean() function is called to pool the Nth C The features extracted by the convolutional layer are converted into 1*1*1 dimensions, and then input into the fully connected layer. Set Q f Layers of fully connected layers, each layer has M fi (i=1,2,…,Q f ) neurons, the output of each fully connected layer is:

[0033] O i =f(O i-1 *W+b)

[0034] Among them, f is the activation function, commonly used activation functions such as ReLU, Tanh or Sigmoid. W is the weight matrix and b is the bias vector, both of which are determined by continuous iteration during the model training process.

[0035] Q f The output of the fully connected layer is the output of the convolutional neural network, that is, the fatigue life characteristic value predicted by the model.

[0036] Preferably, in the process of training the convolutional neural network model using the training set, the Matplotlib library is called to visually monitor the convergence of the loss function. When the loss function converges to the expected result, that is, there is no obvious improvement or the improvement is very slow, the ideal number of model training iterations (epochs) is obtained. This parameter usually needs to be obtained through multiple debugging.

[0037] After that, the test set data is called, the image tensor is input into the model, and the determination coefficient is used to evaluate the life prediction performance of the model on the test set. The determination coefficient calculation formula is:

[0038]

[0039] Among them, N f is the fatigue life characteristic value obtained from the test, is the model prediction result, is the mean of the experimental characteristic values, and the coefficient of determination R 2 The value range is 0-1.

[0040] The beneficial effects of the present invention are:

[0041] A method for predicting the low-cycle fatigue life of materials under arbitrary strain images is provided. Only by building an image data set and training a convolutional neural network model, a prediction model provided by the present invention can be used to predict the low-cycle fatigue life of structures under arbitrary strain images. The method can obtain structural fatigue life information directly by collecting images, is simple and convenient, has the advantages of intelligent calculation, real-time output results, and has high prediction accuracy.

[0042] The proposed model combines DIC image information and deep learning models. It does not require the input of material life cycle data when predicting the low-cycle fatigue life of structures, and has the advantages of convenience and economy. By calling the model, accurate and reliable low-cycle fatigue life can be performed, which can be used as a general image-based life prediction method. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] Figure 1 This is a flow chart of a method for predicting low-cycle fatigue life of a structure based on image and deep learning methods proposed by the present invention;

[0044] Figure 2 This is a schematic diagram of the prediction results of a structural low-cycle fatigue life prediction method based on image and deep learning methods proposed in the present invention. DETAILED DESCRIPTION

[0045] The technical solutions in the embodiments of the present invention will be described clearly and completely below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments.

[0046] Reference Figure 1-Figure 2, a method for predicting the low-cycle fatigue life of a structure based on image and deep learning methods. For a certain high-alloy steel structural material, a low-cycle fatigue test was carried out on it and deformation speckle images were taken. The material strain speckle image obtained by DIC technology has the key features to describe the low-cycle fatigue life of the material. Through image data processing and convolutional neural network training, the "image-fatigue life" data set obtained based on the low-cycle fatigue test was divided, and the prediction of the low-cycle fatigue life of the material was realized. The specific prediction is as follows:

[0047] Step 1: Select the target structural material, prepare DIC speckle on the surface of the material, clamp it on the fatigue testing machine, and turn on the external light source to supplement the light on the surface of the material. The external light source can be illuminated by LED lamps. Install the industrial camera on the microscopic high temperature observation system, use the level to adjust the camera level, and adjust the position of the camera relative to the testing machine and the camera focal length, set the N1 group low-cycle fatigue test loading conditions, until a clear material surface speckle image is captured.

[0048] The loading conditions of low-cycle fatigue test include stress ratio R, loading waveform, frequency F (Hz), and stress amplitude range σ1-σ2 (MPa).

[0049] Step 2: During the low-cycle fatigue test, an industrial camera is used to continuously capture the surface deformation images of the selected area to obtain a speckle image. Before each test, the original undeformed image of the material surface is captured as a reference for subsequent DIC strain calculation and data set construction. During each test, the aforementioned image acquisition system is used to capture images of the material surface, and the camera shooting interval is set to T1 (ms). The number of deformation speckle images that can be captured in each test cycle is Considering the device memory and data processing task volume, it is set to collect deformation speckle images for N5 consecutive test cycles every N4 test cycles.

[0050] Step 3: Use the DIC calculation module VIC-2D to calculate and process the image. DIC calculation module: DIC (Digital Image Correlation) calculation technology is used to analyze and measure the deformation, displacement and strain of the object during the loading process.

[0051] The calculation parameters set include the sub-area radius r, the sub-area step size s, and the average strain ε of each speckle image. ij , which is the average strain value corresponding to the jth image under the i-th group of loading conditions. Under each group of loading conditions, the final fatigue life of the material N iDivide into I equal parts, and combine the above average strain calculation results to randomly select N6 speckle images corresponding to the strain peak of a certain point of the material and the peak on both sides in each equal life interval. Thus, I*N6 deformation speckle images are sampled under each group of loading conditions, and N1*I*N6 deformation speckle images are sampled under all loading conditions. All the sampled images are cropped using Python scripts and uniformly processed into L*W pixels to ensure consistent sample size during subsequent data processing.

[0052] Step 4: For the images obtained under each set of loading conditions, sample and stack the images according to the DIC calculation results, construct the "image-fatigue life" data set, and divide it into training set and test set.

[0053] Using the torch.stack() function in the Pytorch framework, by defining the tensor sequence to be connected and the stacking dimension, the N6 images sampled from each equal life span interval can be stacked with 1 original image according to the time dimension of the experimental process, and the obtained tensor shape is (I*N6+1)*L*W.

[0054] In this step, the series of deformation speckle images and original images sampled under the same group of loading conditions are stacked in the time dimension. This is because the strain information on the material surface under the low-cycle fatigue test is time-series, and the order information of the image sequence needs to be retained through the Stack function; at the same time, when the Stack function is called, the input tensor must have the same shape, which is the reason why all images are uniformly cropped into L*W pixels at the end of step three.

[0055] Then the low cycle fatigue life N of the material under each set of loading conditions is calculated. i After normalization, the formula is:

[0056]

[0057] Among them, N f is the dimensionless fatigue life after normalization (low-cycle fatigue life characteristic value), and its value range is 0-1; N i 、N min and N max They are the material low cycle fatigue life test value, minimum value and maximum value respectively.

[0058] This constructs the "image-fatigue life" dataset (i=1,2,…,N1), divide the training set into proportions and test set Where M and N are the number of training and testing samples respectively, and M+K=N1.

[0059] Step 5: Build a 3D convolutional neural network model based on the Pytorch framework. The model architecture includes input layer, convolution layer, pooling layer, fully connected layer and output layer. The data received by the input layer is a color digital image based on the RGB (red, green and blue) color model, so the number of channels of the model input layer is set to 3*(I*N6+1)*L*W. Combined with the image tensor obtained in step 4, the final input tensor shape of the input layer is. It is worth noting that the architecture parameters of the neural network can be optimized by comparing the performance of the model under various parameter configurations. In this example, N is set. C Convolutional layers, each with a convolution kernel size of K ci and step length S ci (i=1,2,…,N c ), and at the same time, the number of channels in each layer can be doubled layer by layer according to the rule of thumb, such as 64, 128, and 256 respectively. The output of each convolutional layer is:

[0060]

[0061] Among them, Q i-1 is the output of the previous layer, which is also the input of this layer. For the first convolutional layer, its Q i-1 It is the input tensor of the entire model, that is, 3*(I*N6+1)*L*W.

[0062] After that, the feature information is passed to the pooling layer, and the global average pooling torch.mean() function is called to pool the Nth C The features extracted by the convolutional layer are converted into 1*1*1 dimensions. Then the features are input into the fully connected layer and Q is set. f Layers of fully connected layers, each layer has M fi (i=1,2,…,Q f ) neurons, the output of each fully connected layer is:

[0063] O i =f(O i-1 *W+b)

[0064] Among them, f is the activation function, commonly used activation functions such as ReLU, Tanh or Sigmoid. W is the weight matrix and b is the bias vector, both of which are determined by continuous iteration during the model training process.

[0065] Q f The output of the fully connected layer is the output of the convolutional neural network in this example, that is, the fatigue life characteristic value predicted by the model.

[0066] In this example, the optimizer is Adam, which combines the advantages of momentum and adaptive learning. It finds adaptive learning rates for different parameters by calculating the first-order moment estimate and the second-order moment estimate of the gradient, so as to accelerate the convergence speed of model training and improve model performance. When calling the Adam optimizer, first initialize and define the learning rate α, the exponential decay rate β1 of the first-order moment estimate, the exponential decay rate β2 of the second-order moment estimate, a small constant term, and the loss function. Among them, the learning rate controls the speed of model learning. A larger value can achieve a faster learning speed, while a smaller value can achieve better convergence. It is usually set to 0.001; the exponential decay rate of the first-order moment estimate is set to a number close to 1, such as 0.999; the exponential decay rate of the second-order moment estimate is 0.999, and the constant term is a very small number, usually 10 -8 , to prevent the divisor from being 0 during optimization calculation; the loss function is a function that evaluates the error between the model prediction result and the actual result. During model training, the optimization algorithm will update the model parameters according to the value of the loss function to minimize the model prediction error. Commonly used loss functions include mean square error (MSE), mean absolute error (MAE) and cross entropy loss (CEL). Since fatigue life prediction in this example is a regression task, the mean square error loss function is selected, and the calculation formula is:

[0067]

[0068] Among them, Loss represents the model mean square error loss, N f is the fatigue life characteristic value based on the test, is the model prediction result, and N1 is the number of life eigenvalues.

[0069] In the process of training the convolutional neural network model using the training set, the Matplotlib library is called to visualize the convergence of the loss function. When the loss function converges to the expected result, that is, there is no obvious improvement or the improvement is very slow, the ideal number of model training iterations (epochs) is obtained. This parameter usually needs to be obtained through multiple debugging. After that, the test set data is called, the image tensor is input into the model, and the determination coefficient is used to evaluate the life prediction performance of the model on the test set. The determination coefficient calculation formula is:

[0070]

[0071] Among them, N f is the fatigue life characteristic value obtained from the test, is the model prediction result, is the mean of the characteristic values ​​obtained from the experiment. 2The value range is 0-1. The closer the value is to 1, the higher the reliability and generalizability of the obtained model. The optimal model parameter configuration is saved accordingly. At the same time, the plot function in the Matplotlib library can be called to combine the error band evaluation model in the training set D train and the test set D test The predicted performance on .

[0072] Step 6: Call the model, process its unknown arbitrary speckle image sequence according to the data format required by its input layer, and then input it into the model for low-cycle fatigue life prediction.

[0073] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can make equivalent replacements or changes according to the technical scheme and inventive concept of the present invention within the technical scope disclosed by the present invention, which should be covered by the protection scope of the present invention.

Claims

1. A method for predicting low-cycle fatigue life of a structure based on image and deep learning method, characterized in that: The method comprises the following steps: first, selecting a target structural material, preparing a DIC speckle pattern on the surface of the material, and setting a N1 group low cycle fatigue test loading condition; Step 2: During the low cycle fatigue test, an industrial camera is used to continuously capture the surface deformation image of the selected area to obtain a speckle image; Step 3: Use the DIC calculation module VIC-2D to calculate the image. The calculation parameters set include the sub-area radius r, the sub-area step size s, and calculate the average strain ε of each speckle image. ij , that is, the average strain value corresponding to the jth image under the i-th group of loading conditions; Step 4: For the images obtained under each set of loading conditions, sample and stack the images according to the DIC calculation results, construct an "image-fatigue life" data set, and divide it into a training set and a test set; Step 5: Build a 3D convolutional neural network model based on the Pytorch framework; Step 6: Call the model, process its unknown arbitrary speckle image sequence according to the data format required by its input layer, and then input it into the model for low-cycle fatigue life prediction.

2. The method for predicting low cycle fatigue life of a structure based on image and deep learning method according to claim 1 is characterized in that: The low cycle fatigue test loading condition settings in step 1 include stress ratio R, loading waveform, frequency F (Hz), and stress amplitude range σ1-σ2 (MPa); Assuming the shooting interval of the industrial camera is T1 (ms), the number of deformation speckle images that can be collected in each test cycle is Considering the device memory and data processing task volume, it is set to collect deformation speckle images for N5 consecutive test cycles every N4 test cycles.

3. The method for predicting structural low cycle fatigue life based on image and deep learning method according to claim 1 is characterized in that: In step 3, under each set of loading conditions, the final fatigue life N of the material is i Divide into I equal parts, and randomly select N6 speckle images corresponding to a strain peak at a certain point of the material and both sides of the peak in each equal life interval in combination with the result of the average strain calculation; I*N6 deformation speckle images were sampled under each group of loading conditions, and a total of N1*I*N6 deformation speckle images were sampled under all loading conditions. All sampled images were cropped using a Python script and uniformly processed into L*W pixels to ensure consistent sample size during subsequent data processing.

4. The method for predicting low cycle fatigue life of a structure based on image and deep learning method according to claim 1, characterized in that: In the step 4, by defining the tensor sequence to be connected and the stacking dimension, the N6 images sampled from each equal life span interval and the 1 original image are stacked according to the time dimension of the test process, and the obtained tensor shape is (I*N6+1)*L*W.

5. The method for predicting low cycle fatigue life of a structure based on image and deep learning method according to claim 4 is characterized in that: The low cycle fatigue life N of the material under each set of loading conditions i After normalization, the formula is: Among them, N f is the dimensionless low-cycle fatigue life characteristic value after normalization, and its value range is 0-1; N i 、N min and N max They are the numerical value, minimum value and maximum value of the material low cycle fatigue life test; After the normalization, the "image-fatigue life" data set is constructed Divide the training set into proportions and test set Where M and N are the number of training and testing samples respectively, and M+K=N1.

6. The method for predicting structural low cycle fatigue life based on image and deep learning method according to claim 1, characterized in that: In step five, the model architecture includes an input layer, a convolutional layer, a pooling layer, a fully connected layer and an output layer.

7. The method for predicting low cycle fatigue life of a structure based on image and deep learning method according to claim 6 is characterized in that: The data received by the input layer is a color digital image based on the RGB (red, green, blue) color model, so the number of channels of the model input layer is set to 3*(I*N6+1)*L*W.

8. The method for predicting low cycle fatigue life of a structure based on image and deep learning method according to claim 7 is characterized in that: The convolutional layer is set to N C Layer, each layer has a convolution kernel size of K ci and step length S ci (i=1,2,…,N c ), at the same time, the number of channels in each layer can be doubled layer by layer according to the rule of thumb, and the output of each convolutional layer is: Among them, Q i-1 is the output of the previous layer, that is, the input of this layer. For the first convolutional layer, its Q i-1 It is the input tensor of the entire model, that is, 3*(I*N6+1)*L*W.

9. The method for predicting structural low cycle fatigue life based on image and deep learning method according to claim 8, characterized in that: The feature information processed by the convolution layer is passed to the pooling layer, and the global average pooling torch.mean() function is called to pool the Nth C The features extracted by the convolutional layer are converted into 1*1*1 dimensions, and then input into the fully connected layer. Set Q f Layers of fully connected layers, each layer has M fi (i=1,2,…,Q f ) neurons, the output of each fully connected layer is: O i =f(O i-1 *W+b) Among them, f is the activation function, W is the weight matrix, and b is the bias vector, both of which are determined by continuous iteration during the model training process; Q f The output of the fully connected layer is the output of the convolutional neural network, that is, the fatigue life characteristic value predicted by the model.

10. The method for predicting structural low cycle fatigue life based on image and deep learning method according to claim 9, characterized in that: When using the training set to train the convolutional neural network model, call the Matplotlib library to visualize the convergence of the loss function. When the loss function converges to the expected result, that is, there is no obvious improvement or the improvement is very slow, the ideal number of model training iterations (epochs) is obtained. This parameter usually needs to be obtained through multiple debugging. After that, the test set data is called, the image tensor is input into the model, and the determination coefficient is used to evaluate the life prediction performance of the model on the test set. The determination coefficient calculation formula is: Among them, N f is the fatigue life characteristic value obtained from the test, is the model prediction result, is the mean of the experimental characteristic values, and the coefficient of determination R 2 The value range is 0-1.

Images