A method for predicting forming quality of a multi-element material connector based on a hierarchical cascading generative adversarial network

Abstract

The application discloses a kind of multivariate material connector forming quality prediction method based on hierarchical cascade generative adversarial network, constructs topography prediction GAN network dataset, stress / variable field prediction GAN network dataset;Build initial topography prediction GAN network, utilize topography prediction GAN network dataset to carry out training and precision verification to it, obtain final topography prediction GAN network, and input without process parameter used for training, obtain the predicted connector profile topography chart;Based on stress / variable field prediction GAN network dataset, initial stress / variable field prediction GAN network is built respectively, and is trained and verified using stress / variable field prediction GAN network dataset, obtain final stress / variable field prediction GAN network, and input connector profile topography chart, output predicted connector profile stress field nephogram and strain field nephogram.The profile topography characteristics and stress / variable field information of connector obtained by the application can be quickly and accurately obtained.

Description

Technical Field

[0001] This invention belongs to the field of connector forming quality analysis technology, specifically relating to a connector forming quality prediction method based on hierarchical cascaded generative adversarial networks. Background Technology

[0002] Lightweighting is a crucial means of achieving energy conservation and carbon reduction in automobiles, while the hybridization of multiple materials is a key path to balancing vehicle safety, lightweighting, and cost. Traditional spot welding processes are insufficient for reliable connections between multiple materials, especially between high-strength steel and aluminum alloys, magnesium alloys, plastics, and composite materials. Self-piercing riveting (SPR), adhesive bonding, glued riveting, and hot-melt self-tapping screws (FDS) are effective solutions for joining dissimilar materials.

[0003] The forming quality of connectors directly determines their connection performance, and the cross-sectional morphology and stress-strain field distribution are important indicators for evaluating the forming quality of connectors. Typically, cross-sectional metallographic observation tests or numerical simulation methods of the forming process are used to obtain information on the cross-sectional morphology and stress-strain field distribution of the connector after forming, thereby assessing its connection quality. While these methods are effective, they suffer from high costs and long processing times.

[0004] In recent years, using deep learning methods to predict the forming quality of connectors has become an important technical approach. Chinese Patent (CN112100753A) discloses a system and method for predicting key geometric parameters of self-piercing riveting joints using big data, achieving rapid prediction of forming quality indicators. However, this method predicts data indicators but fails to predict the cross-sectional morphology and stress-strain field distribution of the self-piercing riveting joint. Chinese Patent (CN116757068A) discloses a deep learning-based prediction method for the forming process of CFRP self-piercing riveting, achieving prediction of the cross-sectional morphology features of CFRP self-piercing riveting joints. Chinese Patent (CN116758394A) discloses a connector cross-sectional prediction method based on conditional generative adversarial networks, achieving prediction of both the cross-sectional morphology features and forming quality indicators of the connector. The stress-strain field distribution of the connector is strongly correlated with its forming morphology features, which is crucial for assessing the reliability of the connector connection. However, no method has yet been able to simultaneously predict the forming morphology features and stress-strain field distribution of the connector, limiting its forming quality assessment. Summary of the Invention

[0005] To address the problems existing in the background technology, this invention provides a method for predicting the forming quality of multi-material connectors based on hierarchical cascaded generative adversarial networks. This method enables rapid prediction of the cross-sectional forming morphology, stress-strain field, and automatic reading of the forming quality indicators of multi-material connectors under given process parameters.

[0006] The present invention achieves the above-mentioned technical objectives through the following technical means.

[0007] A method for predicting the forming quality of multi-material connectors based on hierarchical cascaded generative adversarial networks:

[0008] Obtain cross-sectional topography images of connectors and corresponding process parameters, and construct a topography prediction GAN network dataset.

[0009] The cross-sectional topography of the connector and the corresponding stress and strain fields were obtained by numerical simulation model, and stress field prediction GAN network dataset and strain field prediction GAN network dataset were constructed respectively.

[0010] An initial shape prediction GAN network is constructed; the initial shape prediction GAN network includes a generator and a discriminator; the generator includes a conditional coding branch, an initial projection layer, a progressive upsampling path and an output layer, and the discriminator includes a conditional embedding branch, a main convolutional path and a projection-based splitting layer;

[0011] The initial shape prediction GAN network was trained and its accuracy was verified using the shape prediction GAN network dataset to obtain the final shape prediction GAN network.

[0012] Inputting no process parameters for training into the final morphology prediction GAN network yields the predicted cross-sectional morphology of the connector.

[0013] The accuracy of the forming quality indicators of the predicted results is verified by using experimental results to ensure the reliability of the predicted morphology results;

[0014] An initial stress field prediction GAN network and an initial strain field prediction GAN network were built based on the stress field prediction GAN network dataset and the strain field prediction GAN network dataset, respectively. Both the initial stress field prediction GAN network and the initial strain field prediction GAN network include a generator and a discriminator. The generator adopts the U-NET architecture and includes an encoder path and a decoder path. The discriminator adopts the Patch GAN architecture and includes multiple convolutional blocks.

[0015] The initial stress field prediction GAN network is trained and validated using the stress field prediction GAN network dataset to obtain the final stress field prediction GAN network; the initial strain field prediction GAN network is trained and validated using the strain field prediction GAN network dataset to obtain the final strain field prediction GAN network.

[0016] Input the cross-sectional topography of the connector into the final stress field prediction GAN network and the final strain field prediction GAN network respectively, and output the predicted stress field cloud map and strain field cloud map of the connector cross-section.

[0017] The accuracy of the predicted stress field cloud map and strain field cloud map of the connector section is verified by using simulation results to ensure the reliability of the prediction results.

[0018] Furthermore, in the generator of the initial shape prediction GAN network, the conditional coding branch consists of multiple fully connected layers and activation function layers; the initial projection layer includes fully connected layers and activation function layers; the progressive upsampling path includes multiple upsampling blocks, each upsampling block including a nearest neighbor upsampling layer, a convolutional layer, an instance normalization layer, and an activation function layer; the output layer includes a transition layer and a final convolutional layer, the transition layer including a convolutional layer, an instance normalization layer, and an activation function layer, and the final convolutional layer including a convolutional layer and an activation function layer.

[0019] Furthermore, the objective function of the initial topography prediction GAN network generator is: In the formula, For the generator value function, To help the generator combat loss, for Reconstruction losses, For feature matching loss, , For the corresponding weighting coefficients, For the generator in noise With conditional information The predicted topography map generated under constraints This is the corresponding real-world morphological image. For the first Feature extraction function of layer This is a conditional discriminator.

[0020] Furthermore, in the discriminator of the initial shape prediction GAN network, the conditional embedding branch consists of multiple fully connected layers; the main convolutional path consists of multiple convolutional blocks, each convolutional block including a convolutional layer and an activation function layer, and each convolutional layer uses spectral normalization; the projection-based scoring layer calculates the final score through global average pooling and conditional projection.

[0021] Furthermore, the objective function of the initial shape prediction GAN network discriminator is: In the formula, Let be the discriminator value function. To provide adversarial loss for real samples, To generate samples for adversarial loss, For the generator in noise With conditional information The predicted topography map generated under constraints This is the corresponding real-world morphological image. This is a conditional discriminator.

[0022] Furthermore, the initial shape prediction GAN network is trained and its accuracy is verified using the aforementioned shape prediction GAN network dataset, specifically including:

[0023] Step 1: Construct the initial dataset for the topography prediction GAN network using cross-sectional topography images of connectors under multiple combinations of different process parameters;

[0024] Step 2: Preprocess the initial dataset of the shape prediction GAN network and divide it into training set and test set;

[0025] Step 3: Train the initial shape prediction GAN network using the training set, and verify the accuracy using K-fold cross-validation based on the test set to obtain the final shape prediction GAN network.

[0026] Furthermore, in the generators of the initial stress field prediction GAN network and the initial strain field prediction GAN network, the encoder path includes multiple downsampling blocks, each downsampling block including a double convolutional block and a max pooling layer, and the decoder path includes multiple upsampling blocks, each upsampling block including a transposed convolutional layer and a double convolutional block, the double convolutional block including a convolutional layer, a batch normalization layer, a rectifier layer, a convolutional layer, a batch normalization layer, and a rectifier layer connected in sequence.

[0027] Furthermore, the objective functions of the generators for the initial stress field prediction GAN network and the initial strain field prediction GAN network are: In the formula, For the generator value function, For conditional discriminators, For conditional information images, To generate the desired image, for Reconstruction losses, for Reconstruction loss weighting coefficient Is the real image as Conditional information image is Expectations at that time.

[0028] Furthermore, in the discriminator of the initial stress field prediction GAN network and the initial strain field prediction GAN network, the convolutional block includes a convolutional layer, a batch normalization layer and an activation function layer, with the last layer being a convolutional layer;

[0029] The objective functions of the discriminators in the initial stress field prediction GAN network and the initial strain field prediction GAN network are: In the formula, Let be the discriminator value function. Is the real image as Conditional information image is Expectations at that time To generate the desired image, This is a conditional discriminator.

[0030] Furthermore, the initial stress / variable field prediction GAN network is trained and validated using the aforementioned stress / variable field prediction GAN network dataset, specifically including:

[0031] Step 1: Construct the initial dataset for the initial stress / variable field prediction GAN network using multiple cross-sectional topographic images of steel-aluminum glued riveted connectors and corresponding simulated stress / variable field cloud maps.

[0032] Step 2: Preprocess the initial dataset of the stress / variable field prediction GAN network and divide it into training set and test set;

[0033] Step 3: Train the stress / variable field prediction GAN network using the training set, and verify the accuracy using K-fold cross-validation based on the test set to obtain the final stress / variable field prediction GAN network.

[0034] The beneficial effects of this invention are as follows:

[0035] (1) This invention overcomes the limitations of existing deep learning methods that can only predict single morphological features or discrete quality indicators, and realizes the synchronous prediction of the forming morphological features and stress-strain field distribution of connectors and the coupled characterization of multiple physics fields. By constructing a hierarchical generative adversarial network model of process parameters-section morphology-stress-strain field, the technical bottleneck of traditional evaluation methods being unable to simultaneously obtain the cross-sectional geometry and internal stress-strain state is solved, filling the technical gap in the comprehensive evaluation of the forming quality of dissimilar material connectors in multiple dimensions, and significantly improving the accuracy and comprehensiveness of connector reliability prediction.

[0036] (2) This invention overturns the traditional quality assessment model that relies on cross-sectional metallographic observation tests and numerical simulation of forming processes, effectively overcoming its key defects of high cost and long cycle. Based on the efficient prediction framework of deep learning, it can quickly obtain the cross-sectional morphology features and stress-strain field distribution information of the connector, greatly reducing the test cost and high-performance computing resource consumption, and is particularly suitable for rapid parameter iteration and multi-objective optimization of multi-material connection processes in automotive lightweight design. Attached Figure Description

[0037] Figure 1 This is a flowchart of the prediction method described in this invention;

[0038] Figure 2This is a metallographic cross-sectional image of the steel-aluminum adhesive riveting connector used in the embodiments of the present invention.

[0039] Figure 3 This is a simulated cross-sectional morphology of the steel-aluminum adhesive riveting connector used in the embodiments of the present invention;

[0040] Figure 4 This is a simulation stress field cloud diagram of the steel-aluminum glued riveting connector used in the embodiments of the present invention.

[0041] Figure 5 This is a simulation strain field cloud diagram of the steel-aluminum adhesive riveting connector used in the embodiments of the present invention.

[0042] Figure 6 A schematic diagram of the GAN network structure for initial topography prediction;

[0043] Figure 7 This is a generator network structure diagram of the initial topography prediction GAN network used in the embodiments of the present invention;

[0044] Figure 8 This is a diagram of the discriminator network structure of the initial topography prediction GAN network used in the embodiments of the present invention;

[0045] Figure 9 This is a schematic diagram of the key point positioning of the automatic extraction process of forming quality indicators used in the embodiments of the present invention;

[0046] Figure 10 This is a network structure diagram of the initial stress field prediction GAN network and the initial strain field prediction GAN network generator used in the embodiments of the present invention;

[0047] Figure 11 This is a diagram showing the network structure of the GAN network for predicting the final stress field and the GAN network for predicting the final strain field used in the embodiments of the present invention.

[0048] Figure 12 These are samples of the predicted topography images from the final topography prediction GAN network in the embodiments of the present invention.

[0049] Figure 13 These are real morphological image samples from the embodiments described in this invention;

[0050] Figure 14 This is a schematic diagram showing the selection of sample points for stress and strain field comparison in the embodiments of the present invention;

[0051] Figure 15 This is a sample of the stress field cloud map predicted by the final stress field prediction GAN network in the embodiments of the present invention;

[0052] Figure 16These are real stress field cloud map samples from the embodiments described in this invention;

[0053] Figure 17 This is a sample strain field cloud map predicted by the final strain field prediction GAN network in the embodiment of the present invention;

[0054] Figure 18 These are sample cloud maps of the actual strain field in the embodiments described in this invention. Detailed Implementation

[0055] The present invention will be further described below with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the scope of the invention.

[0056] Figure 1 The present invention provides a flowchart of a method for predicting the forming quality of multi-material connectors based on hierarchical cascaded generative adversarial networks. The overall technical route can be summarized into eight steps:

[0057] Step 1: Obtain the cross-sectional topography of the connector and the corresponding process parameters, and construct the topography prediction GAN network dataset.

[0058] In this embodiment, the selected connector is a steel-aluminum adhesive-fit connector, and its cross-sectional morphology includes cross-sectional morphology obtained from metallographic observation experiments (see...). Figure 2 ), and the cross-sectional topography obtained from the numerical simulation model (see Figure 3 ).

[0059] Process parameters include, but are not limited to, upper plate thickness, lower plate thickness, rivet length, mold diameter, and mold depth.

[0060] Step 2: Obtain the cross-sectional morphology of the connector and the corresponding stress and strain fields through numerical simulation models, and construct stress field prediction GAN network datasets and strain field prediction GAN network datasets respectively. The stress field cloud maps and strain field cloud maps of the steel-aluminum glued riveting connector forming simulation are shown below. Figure 4 , 5 As shown.

[0061] Step 3: Build the initial morphology prediction GAN network.

[0062] The principle of the initial topography prediction GAN network described in this embodiment is as follows: Figure 6 As shown, it consists of two parts: a generator G and a discriminator D. The generator generates images from random noise, while the discriminator tries to distinguish between real images and generated images. The two networks "play a game" with each other until the generator can generate images that are realistic enough.

[0063] The generator G of the initial morphology prediction GAN network consists of a conditional coding branch, an initial projection layer, a progressive upsampling path, and an output layer. The network architecture diagram is shown in Figure 7.

[0064] The conditional coding branch consists of three fully connected layers, each using Leaky ReLU as the activation function to encode the 5-dimensional process parameters into a 512-dimensional feature vector.

[0065] The initial projection layer consists of a fully connected layer and a Leaky ReLU activation function layer. By reshaping, the 128-dimensional random noise (the initial input of the network) and the 512-dimensional feature vector are concatenated and projected into a 2×2×512 initial feature map.

[0066] The progressive upsampling path includes 8 upsampling blocks. Each upsampling block contains a nearest neighbor upsampling layer, a 3×3 convolutional layer, an instance normalization layer, and a ReLU activation function layer. The nearest neighbor upsampling layer enlarges the feature map size, the convolutional layer reconstructs the features, the instance normalization layer stabilizes the training, and the ReLU activation function introduces non-linearity to improve the feature representation capability.

[0067] The output layer consists of a transition layer and a final convolutional layer. The transition layer contains a 3×3 convolutional layer, an instance normalization layer, and a ReLU activation function layer. The final convolutional layer contains a 3×3 convolutional layer and a tanh activation function layer. The transition layer is used to smooth features and perform channel integration, and then the target image is output through the final convolutional layer.

[0068] The discriminator D of the initial topography prediction GAN network distinguishes between the real topography image and the topography image generated by the generator G. At the same time, it introduces conditional information to discriminate and constrain the features related to the physical process in the generated topography image, so as to enhance the correlation between the generated result and the input conditions and improve the prediction accuracy. The discriminator includes a conditional embedding branch, a main convolutional path and a projection-based layering. The network architecture diagram is shown in Figure 8.

[0069] The conditional embedding branch consists of three fully connected layers that embed the 5-dimensional process parameters into a 512-dimensional feature vector.

[0070] The main convolutional path consists of 5 convolutional blocks, each containing a 3×3 convolutional layer and a Leaky ReLU activation function layer. Each convolutional layer uses spectral normalization to extract and discriminate multi-scale discriminative features of the image layer by layer, thereby enhancing the discriminative stability.

[0071] Projective layer scoring evaluates the cross-sectional topography of connectors from two aspects: one is based on the learnable vectors. inner product To assess the realism of an image, another part involves conditional coding. inner product To evaluate the degree of matching between the image and the conditions, the final score is the sum of these two items: Where h represents the features of the cross-sectional topography of the connector. This allows the network to learn the correspondence of the conditions more accurately, while also preventing the generator from generating images that appear realistic but do not match the conditions. Projective hierarchical scoring calculates the final score through global average pooling and conditional projection.

[0072] The objective function of the generator G in the initial topography prediction GAN network is: .

[0073] In the formula, The value function of the generator. For the generator's adversarial loss, for Reconstruction losses, For feature matching loss, , For the corresponding weighting coefficients, For the generator in noise With conditional information The predicted topography map generated under constraints This is the corresponding real-world morphological image. For the first Feature extraction function of layer It serves as a conditional discriminator, used to evaluate the authenticity of the generated topographic map under given conditional constraints.

[0074] The objective function of the discriminator D in the initial topography prediction GAN network is:

[0075] .

[0076] In the formula, The value function of the discriminator, For adversarial loss of real samples, Adversarial loss for generating samples.

[0077] Adversarial loss is a loss function in generative adversarial networks defined by the "game" relationship between the generator G and the discriminator D. It enables the generator to generate more "realistic" samples to "deceive" the discriminator, while the discriminator learns to better distinguish between real and fake samples through it.

[0078] Reconstruction loss is the loss of the L1 distance between the generated image and the real image in pixel space, which gradually reduces the pixel-level difference between the generated image and the real image.

[0079] The feature matching loss compares the differences between the generated image and the real image in the intermediate layer features of the discriminator D, so that the discriminator D can not only distinguish between real and fake images, but also distinguish between the internal features of the image.

[0080] Step 4: Train and validate the initial shape prediction GAN network using the shape prediction GAN network dataset to obtain the final shape prediction GAN network. This includes the following steps:

[0081] Step 1: Construct the initial dataset for the morphology prediction GAN network using cross-sectional morphology images of connectors under multiple combinations of different process parameters. In this embodiment, the connector selected is a steel-aluminum glued joint, and its cross-sectional morphology images include cross-sectional morphologies obtained from metallographic observation experiments under different combinations of process parameters. Figure 10 Zhang, used to verify the accuracy of the model prediction, and 70 cross-sectional topography images with different combinations of process parameters obtained from the numerical simulation model, of which 60 images are used as the training set for training the model, and 10 images are used as the test set for comparing and testing the model's prediction performance.

[0082] Step 2: Perform data preprocessing on the initial dataset of the shape prediction GAN network.

[0083] The data preprocessing process involves batch cropping the initial dataset images to 512×512 pixels.

[0084] Step 3: Train the initial shape prediction GAN network using the training set.

[0085] The training parameters include: training optimizer, learning rate, batch size, training epochs, L1 loss weights, and feature matching loss weights.

[0086] In this embodiment, the Adam optimizer is used as the training optimizer for the initial shape prediction GAN network. It is suitable for training GAN models and is used for a smooth training process. The optimizer parameters are set to beta1 of 0.5 and beta2 of 0.999.

[0087] In this embodiment, a Step LR learning rate decay strategy is adopted, where the learning rate decays by 50% every 100 epochs of training.

[0088] In this embodiment, a learning rate warm-up strategy is adopted, which performs linear warm-up in the first 5 epochs of training to prevent the model from becoming unstable due to an excessively high initial learning rate.

[0089] In this embodiment, a mixed precision training strategy is adopted, which automatically switches between FP16 and FP32 to improve memory utilization, accelerate training, and maintain numerical stability.

[0090] In this embodiment, a Tensor Board is used to monitor various losses, learning rate changes, and discriminator accuracy during the training process, for real-time visualization of the training process.

[0091] In this embodiment, a strategy of simultaneously augmenting real and generated images during training is adopted. The data augmentation methods used are: random cropping, rotation, and horizontal flipping.

[0092] In this embodiment, the hyperparameters during the training phase are set as follows: 400 training epochs, 4 batch size, 0.0002 generator learning rate, and 0.0001 discriminator learning rate. Step 5: Input the simulation-obtained process parameters (not used for training) into the final morphology prediction GAN network after training to obtain the predicted cross-sectional morphology of the steel-aluminum glued riveting connector.

[0093] In this embodiment, structural similarity (SSIM), peak signal-to-noise ratio (PSNR), mean absolute error (MAE), and mean squared error (MSE) are used to evaluate the difference between the predicted topography map and the real images in the test set. Based on the test set, this embodiment uses K-fold cross-validation to verify the accuracy of the initial topography prediction GAN network after training. The average evaluation metrics in the test set are: SSIM (structural similarity): 0.9390, PSNR (peak signal-to-noise ratio): 20.9732dB, MAE (mean absolute error): 0.0116, and MSE (mean squared error): 0.0089, all of which reach a high level.

[0094] Step 6: Construct an automatic identification process for the forming quality index of connector cross-section based on OpenCV, perform error analysis on the predicted forming quality index of connector cross-section, and compare the efficiency of using this method to predict the cross-sectional morphology and forming quality index of connector with the efficiency of manual experiment / measurement and finite element simulation methods.

[0095] In this embodiment, the forming quality indicators of the glued joint include head height, interlocking value, bottom thickness, and remaining thickness;

[0096] The automatic identification process for the cross-sectional forming quality indicators of connectors includes the following steps:

[0097] Step 1: Load the topography image predicted in Step 5, and convert it into a category index mask according to a predefined color scale, which serves as the basis for subsequent edge detection and feature point extraction;

[0098] Step 2: For each category of connector (rivet, upper plate, lower plate, adhesive layer), construct a binary mask, remove noise through morphological closing operation and Gaussian smoothing, extract the outer contour, merge the contour points and store them as boundary point coordinate sets for each category, thereby obtaining a structured boundary coordinate dictionary to support subsequent geometric point calculations.

[0099] Step 3: Using the top left corner of the image as the origin, move to the right... The positive direction of the axis is downward. A coordinate system is constructed along the positive axis. Based on the boundary point coordinates, key points for positioning forming quality indicators are automatically extracted, including:

[0100] Key points in the rivet area:

[0101] r1: The topmost edge point in the rivet area located on the upper left side of the image;

[0102] r2: The lowest edge point in the rivet area;

[0103] r3: The rightmost edge point in the horizontal direction in the lower half of the rivet area;

[0104] Key points of the upper plate area:

[0105] t1: The lowest edge point in the upper plate area;

[0106] t2: The uppermost edge point of the right quarter area in the upper plate region;

[0107] Key points in the lower plate area:

[0108] b1: Filter the leftmost edge point in the horizontal direction after cutting off the upper right area by x range and y in the lower plate area;

[0109] b2: The lowest edge point in the lower plate area;

[0110] Locating key points such as Figure 9 The result is a set of coordinates for 7 key points, which provides input for index calculation.

[0111] Step 4: Calculate the pixel thickness at point t1 in the upper plate area. Combined with the known actual thickness, convert the pixel distance of the measured quality index into the actual forming quality index.

[0112] Step 5: Calculate the following forming quality indicators based on the geometric relationship between key points:

[0113] Head height: the vertical distance between r1 and t2;

[0114] Interlock value: the horizontal distance between r3 and b1;

[0115] Bottom thickness: the vertical distance between r2 and b2;

[0116] Remaining thickness: the vertical distance between t1 and b2;

[0117] All distances are converted from pixel distance to actual physical distance using the formula: Actual distance = Pixel difference × (Actual thickness / Thickness pixel difference);

[0118] Step 6: To evaluate the reliability of the final morphology prediction results from the GAN network, the automatically identified forming quality indicators were compared with the manually measured results of the cross-sectional metallographic experiment. Five sets of automatically identified quality indicators from the predicted cross-sectional morphology images were randomly selected and compared with the experimental measurement results of the corresponding process parameters. The relative error was calculated, and the comparison results are shown in Table 1. Furthermore, to evaluate the efficiency improvement of this method in engineering applications, the average time taken to obtain the joint cross-sectional morphology image and forming quality indicators based on a set of process parameters using this method was compared with the time taken by manual experiment / measurement and finite element simulation methods. The comparison results are shown in Table 2.

[0119] Table 1 Comparison of GAN network prediction accuracy results for final morphology prediction

[0120]

[0121] Table 2 Comparison of prediction efficiency of the three methods

[0122]

[0123] The above comparison results show that the present invention achieves a leapfrog improvement from the "minute-level" to the "second-level" response of traditional manual experiments and finite element simulations in predicting the cross-sectional joint morphology and forming quality indicators of connectors, with an overall efficiency improvement of nearly 100 times. In five randomly selected test samples, the relative error of each forming quality indicator identified by the present invention compared with manual measurement was less than 10%, and the prediction accuracy exceeded 90%.

[0124] Step 7: Construct the initial stress field prediction GAN network and the initial strain field prediction GAN network based on the stress field prediction GAN network dataset and the strain field prediction GAN network dataset, respectively.

[0125] In this embodiment, the initial stress field prediction GAN network and the initial strain field prediction GAN network are conditional generative adversarial network models.

[0126] Both prediction networks use a U-NET architecture for their generators, including encoder and decoder paths, with the network architecture as follows: Figure 10 As shown;

[0127] The encoder path includes five downsampling blocks. Each downsampling block includes a double convolutional block and a max pooling layer. The double convolutional block is used to extract and fuse local features to improve feature representation. The max pooling layer downsamples the feature map to expand the receptive field, retain salient features, and reduce computation.

[0128] The decoder path includes 5 upsampling blocks, each of which includes a transposed convolutional layer and a double convolutional block. The transposed convolutional block upsamples the feature map, restores the spatial resolution, and learns a trainable reconstruction map.

[0129] A double convolutional block consists of a convolutional layer, a batch normalized layer, a rectifier layer, a convolutional layer, a batch normalized layer, and a rectifier layer connected in sequence.

[0130] Each upsampling layer has a skip connection that concatenates the feature map of the corresponding layer in the encoder to the decoder.

[0131] The discriminator D of the initial stress field prediction GAN network and the initial strain field prediction GAN network adopts the PatchGAN architecture, and the network architecture is as follows: Figure 11 As shown, it includes 4 convolutional blocks, each of which includes a convolutional layer, a batch normalization layer, and a Leaky ReLU activation layer, with the last layer being a convolutional layer;

[0132] The objective function of the generator G in the initial stress field prediction GAN network and the initial strain field prediction GAN network is:

[0133] ;

[0134] In the formula, The value function of the generator. This is a conditional information image, i.e., a simulated topographic image. To generate the desired image, for Reconstruction losses, for Reconstruction loss weighting coefficient;

[0135] The objective function of the discriminator D in the initial stress field prediction GAN network and the initial strain field prediction GAN network is:

[0136] ;

[0137] In the formula, Let be the discriminator value function. For real images Conditional information image is Expectations at that time.

[0138] Step 8: Train and validate the initial stress field prediction GAN network based on the stress field prediction GAN network dataset to obtain the final stress field prediction GAN network; train and validate the initial strain field prediction GAN network based on the strain field prediction GAN network dataset to obtain the final strain field prediction GAN network. Specifically, this includes the following steps:

[0139] Step 1: Construct the initial dataset for the initial stress field prediction GAN network using multiple cross-sectional topographic images of steel-aluminum glued riveted connectors and the corresponding simulated stress field cloud maps. Construct the initial dataset for the initial strain field prediction GAN network using multiple cross-sectional topographic images of steel-aluminum glued riveted connectors and the corresponding simulated strain field cloud maps.

[0140] The stress and strain field cloud maps of the steel-aluminum glued riveted joint cross-section used in this embodiment were obtained through forming simulations of the glued riveted joint under a large number of different process parameter combinations. The stress and strain field cloud maps of the forming simulation results were exported as initial datasets; each initial dataset for stress and strain field cloud maps included 60 training images and 10 test images.

[0141] Step 2: Perform data preprocessing on the initial datasets of the initial stress field prediction GAN network and the initial strain field prediction GAN network, that is, batch crop the image size of the initial dataset to 512×512 pixels;

[0142] Step 3: Train the initial stress field prediction GAN network and the initial strain field prediction GAN network using the training set from Step 1, and verify the accuracy using the 5-fold cross-validation method based on the test set to obtain the final stress field prediction GAN network and the final strain field prediction GAN network that predict stress field and strain field cloud maps based on the cross-sectional topography of the connector, respectively.

[0143] The training parameters for the initial stress field prediction GAN network and the initial strain field prediction GAN network include: training optimizer, batch size, learning rate, training epochs, and L1 reconstruction loss weights.

[0144] In this embodiment, the Adam optimizer is used for training both the initial stress field prediction GAN network and the initial strain field prediction GAN network, with the optimizer parameters set to beta1 of 0.5 and beta2 of 0.999 respectively.

[0145] In this embodiment, the Reduce LR On Plateau learning rate scheduling strategy is adopted. When the monitored metrics do not improve significantly after 10 consecutive epochs, the scheduler will reduce the learning rate by 50% to optimize the training process.

[0146] In this embodiment, a mixed precision training strategy is adopted, which automatically switches between FP16 and FP32 to improve memory utilization, accelerate training, and maintain numerical stability.

[0147] In this embodiment, a Tensor Board is used to monitor various losses, learning rate changes, and discriminator accuracy during the training process, for real-time visualization of the training process.

[0148] In this embodiment, a strategy of simultaneously augmenting real and generated images during training is adopted. The data augmentation methods used are: random cropping, rotation, and horizontal flipping.

[0149] In this embodiment, the hyperparameters during the training phase are set as follows: 400 training epochs, 2 batch sizes, 0.0001 generator learning rate, and 0.0001 discriminator learning rate.

[0150] Step 9: Input the cross-sectional topography of the steel-aluminum hybrid adhesive riveting connector into the final stress field prediction GAN network and the final strain field prediction GAN network respectively, and output the corresponding predicted stress field and strain field cloud maps to realize multi-level mapping prediction from process parameters to the cross-sectional topography and stress and strain fields of the steel-aluminum adhesive riveting connector.

[0151] In this embodiment, structural similarity (SSIM), peak signal-to-noise ratio (PSNR), mean absolute error (MAE), and mean squared error (MSE) are used to evaluate the difference between the predicted topography map and the real topography map in the test set. The average evaluation metrics in this embodiment on the test set are: SSIM (structural similarity): 0.9950, PSNR (peak signal-to-noise ratio): 16.4098 dB, MAE (mean absolute error): 0.0446, and MSE (mean squared error): 0.0239. The prediction performance is excellent, and the topography map samples predicted by the model are very similar to the real topography map samples. Figure 12 , 13 As shown.

[0152] Step 10: Construct an automatic numerical identification process for the stress and strain field of the connector profile based on OpenCV, perform error analysis on the predicted stress and strain field contour maps of the connector profile, and compare the efficiency of using this method to predict the stress and strain field contour maps of the connector profile with the efficiency of the finite element simulation method. Specifically, this includes the following steps:

[0153] Step 1: Read the R / G / B components and corresponding numerical ranges (stress / strain value ranges) of each pixel from the color scale (CSV file) exported from the finite element software, and take the average value of each numerical range as the representative value of that pixel.

[0154] Step 2: For any pixel in the stress field cloud map and strain field cloud map, determine its value according to the following mechanism: If the pixel's RGB values ​​are strictly consistent with the colors in the color scale, then directly take the corresponding values; if there is no perfect match, calculate the Euclidean distance between the pixel and all pixels, and select the pixel with the smallest Euclidean distance as the recognition result; if the pixel is close to pure black, then set the value to zero to avoid identifying false regions.

[0155] Step 3: Calculate the stress / strain values ​​pixel by pixel for the input image, construct a two-dimensional numerical matrix, and export it as a .txt file;

[0156] Step 4: To evaluate the reliability of the predicted stress and strain field contour maps of the connector profile using this method, five sample points from the simulated stress and strain fields, along with the maximum value point, are compared with the corresponding predicted sample values. The locations of the sample points are as follows: Figure 14 As shown, the comparison results between the simulated actual values ​​and model predictions of the obtained stress and strain fields are shown in Tables 3 and 4, respectively. Furthermore, to verify the efficiency improvement of this method in engineering applications, the average time taken to obtain the stress and strain fields of the connector cross-sectional morphology using the finite element simulation method with a set of process parameters was compared with the time taken using this method. The comparison results are shown in Table 5.

[0157] Table 3 Comparison of GAN network prediction results for final stress field prediction

[0158]

[0159] Table 4 Comparison of GAN network prediction results for final strain field prediction

[0160]

[0161] Table 5 Comparison of prediction efficiency between finite element simulation and the method of this invention

[0162]

[0163] The above comparison results show that, compared with the finite element simulation method, the present invention improves the overall efficiency of predicting the stress and strain field contour maps of the connector cross-section by nearly 100 times. In the numerical comparison of five randomly selected sample points, the relative errors of the stress and strain field values ​​identified by the present invention compared with the actual simulated values ​​are all less than 10%, with an accuracy exceeding 90%. The stress field contour map samples predicted by the model and the actual stress field contour map samples are shown below. Figure 15 , 16 As shown, the predicted strain field contour map samples and the actual strain field contour map samples are as follows: Figure 17 , 18 As shown.

[0164] The above embodiments are only used to illustrate the design concept and features of the present invention, and their purpose is to enable those skilled in the art to understand the content of the present invention and implement it accordingly. The protection scope of the present invention is not limited to the above embodiments. Therefore, all equivalent changes or modifications made based on the principles and design ideas disclosed in the present invention are within the protection scope of the present invention.

Claims

1. A method for predicting the forming quality of multi-material connectors based on hierarchical cascaded generative adversarial networks, characterized in that: Obtain cross-sectional topography images of connectors and corresponding process parameters, and construct a topography prediction GAN network dataset. The cross-sectional topography of the connector and the corresponding stress and strain fields were obtained by numerical simulation model, and stress field prediction GAN network dataset and strain field prediction GAN network dataset were constructed respectively. An initial shape prediction GAN network is constructed; the initial shape prediction GAN network includes a generator and a discriminator; the generator includes a conditional coding branch, an initial projection layer, a progressive upsampling path and an output layer, and the discriminator includes a conditional embedding branch, a main convolutional path and a projection-based splitting layer; The initial shape prediction GAN network was trained and its accuracy was verified using the shape prediction GAN network dataset to obtain the final shape prediction GAN network. Inputting no process parameters for training into the final morphology prediction GAN network yields the predicted cross-sectional morphology of the connector. The accuracy of the forming quality indicators of the predicted results is verified by using experimental results to ensure the reliability of the predicted morphology results; An initial stress field prediction GAN network and an initial strain field prediction GAN network were built based on the stress field prediction GAN network dataset and the strain field prediction GAN network dataset, respectively. Both the initial stress field prediction GAN network and the initial strain field prediction GAN network include a generator and a discriminator. The generator adopts the U-NET architecture and includes an encoder path and a decoder path. The discriminator adopts the Patch GAN architecture and includes multiple convolutional blocks. The initial stress field prediction GAN network is trained and validated using the stress field prediction GAN network dataset to obtain the final stress field prediction GAN network; the initial strain field prediction GAN network is trained and validated using the strain field prediction GAN network dataset to obtain the final strain field prediction GAN network. Input the cross-sectional topography of the connector into the final stress field prediction GAN network and the final strain field prediction GAN network respectively, and output the predicted stress field cloud map and strain field cloud map of the connector cross-section. The accuracy of the predicted stress field cloud map and strain field cloud map of the connector section is verified by using simulation results to ensure the reliability of the prediction results.

2. The method for predicting the forming quality of multi-material connectors according to claim 1, characterized in that, In the generator of the initial shape prediction GAN network, the conditional coding branch consists of multiple fully connected layers and activation function layers; the initial projection layer includes a fully connected layer and an activation function layer; the progressive upsampling path includes multiple upsampling blocks, each upsampling block including a nearest neighbor upsampling layer, a convolutional layer, an instance normalization layer, and an activation function layer; the output layer includes a transition layer and a final convolutional layer, the transition layer including a convolutional layer, an instance normalization layer, and an activation function layer, and the final convolutional layer including a convolutional layer and an activation function layer.

3. The method for predicting the forming quality of multi-material connectors according to claim 2, characterized in that, The objective function of the initial topography prediction GAN network generator is: In the formula, For the generator value function, To help the generator combat loss, for Reconstruction losses, For feature matching loss, , For the corresponding weighting coefficients, For the generator in noise With conditional information The predicted topography map generated under constraints This is the corresponding real-world morphological image. For the first Feature extraction function of layer This is a conditional discriminator.

4. The method for predicting the forming quality of multi-material connectors according to claim 1, characterized in that, In the discriminator of the initial shape prediction GAN network, the conditional embedding branch consists of multiple fully connected layers; the main convolutional path consists of multiple convolutional blocks, each convolutional block including a convolutional layer and an activation function layer, and each convolutional layer uses spectral normalization. The projection-based tiered scoring calculates the final score through global average pooling and conditional projection.

5. The method for predicting the forming quality of multi-material connectors according to claim 4, characterized in that, The objective function of the initial topography prediction GAN network discriminator is: In the formula, Let the discriminator value function be... To provide adversarial loss for real samples, To generate samples to counteract loss, For the generator in noise With conditional information The predicted topography map generated under constraints This is the corresponding real-world morphological image. This is a conditional discriminator.

6. The method for predicting the forming quality of multi-material connectors according to claim 1, characterized in that, The initial shape prediction GAN network is trained and its accuracy is verified using the aforementioned shape prediction GAN network dataset, specifically including: Step 1: Construct the initial dataset for the topography prediction GAN network using cross-sectional topography images of connectors under multiple combinations of different process parameters; Step 2: Preprocess the initial dataset of the shape prediction GAN network and divide it into training set and test set; Step 3: Train the initial shape prediction GAN network using the training set, and verify the accuracy of the trained initial shape prediction GAN network using the K-fold cross-validation method based on the test set to obtain the final shape prediction GAN network.

7. The method for predicting the forming quality of multi-material connectors according to claim 1, characterized in that, In the generators of the initial stress field prediction GAN network and the initial strain field prediction GAN network, the encoder path includes multiple downsampling blocks, each downsampling block including a double convolutional block and a max pooling layer, and the decoder path includes multiple upsampling blocks, each upsampling block including a transposed convolutional layer and a double convolutional block, the double convolutional block including a convolutional layer, a batch normalization layer, a rectifier layer, a convolutional layer, a batch normalization layer, and a rectifier layer connected in sequence.

8. The method for predicting the forming quality of multi-material connectors according to claim 7, characterized in that, The objective functions of the generators for the initial stress field prediction GAN network and the initial strain field prediction GAN network are: In the formula, For the generator value function, For conditional discriminators, For conditional information images, To generate the desired image, for Reconstruction losses, for Reconstruction loss weighting coefficient Is the real image as Conditional information image is Expectations at that time.

9. The method for predicting the forming quality of multi-material connectors according to claim 1, characterized in that, In the discriminator of the initial stress field prediction GAN network and the initial strain field prediction GAN network, the convolutional block includes a convolutional layer, a batch normalization layer and an activation function layer, and the last layer is a convolutional layer. The objective functions of the discriminators in the initial stress field prediction GAN network and the initial strain field prediction GAN network are: In the formula, Let the discriminator value function be... Is the real image as Conditional information image is Expectations at that time To generate the desired image, This is a conditional discriminator.

10. The method for predicting the forming quality of multi-material connectors according to claim 1, characterized in that, The initial stress / variable field prediction GAN network is trained and validated using the aforementioned stress / variable field prediction GAN network dataset, specifically including: Step 1: Construct the initial dataset for the initial stress / variable field prediction GAN network using multiple cross-sectional topographic images of steel-aluminum glued riveted connectors and corresponding simulated stress / variable field cloud maps. Step 2: Preprocess the initial dataset of the stress / variable field prediction GAN network and divide it into training set and test set; Step 3: Train the stress / variable field prediction GAN network using the training set, and verify the accuracy of the trained stress / variable field prediction GAN network using the K-fold cross-validation method based on the test set to obtain the final stress / variable field prediction GAN network.

Images