Forming optimization control system of carbon fiber composite material

Abstract

The invention relates to the technical field of composite material thermocuring molding, in particular to a carbon fiber composite material molding optimization control system, which is characterized in that historical molding data of a carbon fiber composite material is acquired through a historical data acquisition unit, and current state information of a carbon fiber mold is acquired through a mold state acquisition unit. The process parameter determination unit generates process parameters based on the A-PPO model and the current state information. And the image acquisition unit applies the process parameters to the mold to acquire a carbon fiber composite material image. And the image detection unit extracts the image features of the carbon fiber composite material by using a ResNet model, and evaluates the quality of a finished product in combination with a random forest model. And the process parameter optimization unit dynamically adjusts process parameters through an A-PPO model according to a finished product quality result. And the parameter determination unit further optimizes the process parameters by using Bayesian optimization to obtain optimized process parameter settings. The finished product quality and the production efficiency of the carbon fiber composite material are effectively improved.

Description

Technical Field

[0001] The present invention relates to the technical field of thermosetting molding of composite materials, and specifically to a forming optimization control system for carbon fiber composite materials. Background Art

[0002] Carbon fiber composite materials are a type of high-performance materials with excellent mechanical properties and lightweight characteristics, and are widely used in fields such as aerospace, automotive manufacturing, sporting goods, and wind power generation. Due to their high strength, low density, corrosion resistance, and high temperature resistance, carbon fiber composite materials have gradually become key materials in many industries. However, the stamping forming process of carbon fiber composite materials is complex, especially during the thermosetting process, and small changes in process parameters such as temperature and pressure will significantly affect the quality of the finished product.

[0003] During the stamping forming process, common defects include cracks, bubbles, fiber shedding, and uneven material distribution. These defects not only reduce the mechanical properties of the material, but may also result in non-compliant finished products, increase production costs, and affect production efficiency. Traditional stamping forming processes mainly rely on empirical parameter settings or trial-and-error methods to optimize process parameters. However, due to the complexity and non-linearity of the stamping forming process, these methods are usually inefficient and difficult to achieve precise control of process parameters. Although various solutions have been proposed to improve the stamping forming process and increase production efficiency, existing solutions often fail to make full use of rich historical data and real-time monitoring information, making it difficult to adjust process parameters in a timely manner to reduce the generation of defects in carbon fiber composite materials, limiting the effect of further optimization, so there is still room for improvement in production efficiency.

[0004] Therefore, a forming optimization control system for carbon fiber composite materials is proposed. Summary of the Invention

[0005] The object of the present invention is to provide a forming optimization control system for carbon fiber composite materials. The historical forming data of carbon fiber composite materials is collected by a historical data acquisition unit, including historical operation data, template images and defect images; the mold state acquisition unit obtains the current state information of the carbon fiber mold in real time; the process parameter determination unit generates preliminary process parameters based on the A-PPO model and the current state information; the image acquisition unit applies these process parameters to the mold to obtain a preprocessed carbon fiber composite material image; the image detection unit extracts the carbon fiber composite material image features by using the ResNet model, and combines the random forest model to judge the finished product quality. The process parameter optimization unit calculates the reward value according to the finished product quality result, dynamically adjusts the process parameters through the A-PPO model, and obtains the preliminary process parameter setting. The final parameter optimization unit further adjusts the preliminary process parameters by using the Bayesian optimization method to obtain the final optimized process parameter setting. The system can achieve fine control of the stamping process, effectively improve the finished product quality and production efficiency of carbon fiber composite materials.

[0006] To achieve the above object, the present invention provides the following technical solutions:

[0007] A forming optimization control system for carbon fiber composite materials, comprising:

[0008] A historical data acquisition unit, configured to acquire historical forming data of carbon fiber composite materials, where the historical forming data includes historical operation data, defect images and template images;

[0009] A mold state acquisition unit, configured to acquire the current state information of the carbon fiber mold;

[0010] A process parameter determination unit, configured to establish an A-PPO model according to the current state information, where the A-PPO model is used to determine the current process parameters according to the attention mechanism;

[0011] An image acquisition unit, configured to acquire a carbon fiber composite material image according to the current process parameters, and preprocess the carbon fiber composite material image to obtain a first image;

[0012] An image detection unit, configured to extract the defect image features by using the ResNet model according to the historical forming data to obtain a carbon fiber feature vector, classify the defect types according to the template image and the defect image, and determine the defect level of the first image by using the random forest model according to the defect type and the carbon fiber feature vector to obtain the finished product quality result;

[0013] The process parameter optimization unit is used to calculate the finished product reward value using a reward function based on the finished product quality result, and re-determine the current process parameters using the A-PPO model based on the finished product reward value and the current state information to obtain the first process parameters;

[0014] The final parameter determination unit is used to adjust the first process parameters using Bayesian optimization to obtain the final process parameters.

[0015] Further, the current state information of the carbon fiber mold includes mold temperature, mold displacement, stamping speed, stamping pressure, and the finished product quality result.

[0016] Further, the process of determining the current process parameters includes:

[0017] Input the current state information into the A-PPO model. The A-PPO model performs feature enhancement based on the current state information and outputs normal distribution parameters. Random sampling is performed according to the normal distribution parameters to obtain the current process parameters including normal distribution noise terms;

[0018] Among them, the current process parameters include mold temperature adjustment value, mold displacement adjustment value, stamping speed adjustment value, and stamping pressure adjustment value.

[0019] Further, the process of obtaining the first image includes:

[0020] Perform grayscale conversion on the carbon fiber composite material image to obtain a second image;

[0021] Remove white fog from the second image using the dark channel prior method to obtain a third image;

[0022] Perform image denoising on the third image using Gaussian filtering, and perform image enhancement by adjusting contrast and brightness to obtain a fourth image;

[0023] Extract the edge contour of the carbon fiber composite material from the fourth image using the Canny edge detection algorithm to obtain the first image.

[0024] Further, the process of obtaining the carbon fiber feature vector includes:

[0025] Use convolutional layers and pooling layers to extract low-level features of the first image. The low-level features include the edge contour and texture direction of the carbon fiber composite material;

[0026] Extract high-level features using residual blocks based on the low-level features. The high-level features include the texture structure and local defects of the carbon fiber composite material;

[0027] The carbon fiber feature vector is obtained using the global average pooling layer according to the advanced feature.

[0028] Further, the A-PPO deep neural network includes:

[0029] An input layer for receiving the current state information;

[0030] A feature extraction layer for extracting the current state information features using a convolutional layer and a fully connected layer;

[0031] An attention mechanism layer for performing weighted processing on the current state information features to obtain a comprehensive feature representation;

[0032] A policy network layer for outputting the current process parameters according to the comprehensive feature representation;

[0033] A value network layer for outputting the current state value according to the comprehensive feature representation.

[0034] Further, the process of obtaining the finished product quality result includes:

[0035] Dividing the defect levels according to the severity of the defect types, and marking the defect images according to the defect levels and the template images to obtain a first data set;

[0036] Using the ResNet model to extract features from the defect images to obtain the carbon fiber feature vector;

[0037] Inputting the first data set and the carbon fiber feature vector into the random forest model, and the random forest model uses the method of majority voting to determine the defect levels to obtain a trained random forest model;

[0038] Using the ResNet model to extract features from the first image to obtain the carbon fiber feature vector;

[0039] Inputting the carbon fiber feature vector into the trained random forest model to obtain the defect level of the first image;

[0040] Obtaining the finished product quality result according to the defect level of the first image.

[0041] Further, the steps for obtaining the reward function include:

[0042] Multiplying the system energy consumption by the first weight factor to obtain an energy consumption penalty term;

[0043] Multiplying the production time by the second weight factor to obtain a time penalty term;

[0044] Subtract the energy consumption penalty term and the time penalty term from the finished product quality result to obtain the finished product reward value.

[0045] Further, the first process parameter acquisition process includes:

[0046] Step S10: Obtain the finished product reward value and the next moment state information according to the current state information and the current process parameters, and store the current state information, the current process parameters, the finished product reward value, and the next moment state information in the trajectory set;

[0047] Step S11: Determine whether the number of trajectories in the trajectory set reaches the first threshold;

[0048] Step S12: If the number of trajectories reaches the first threshold, calculate the cumulative return using the value function according to the trajectory set; otherwise, use the A-PPO model to determine the next moment process parameters according to the next moment state information, and execute Steps S10 to S11 according to the next moment process parameters and the next moment state information;

[0049] Wherein, the next moment state information is used as the current state information of the next trajectory, and the next moment process parameter is used as the current process parameter of the next trajectory;

[0050] Step S13: Input the current state information into the A-PPO model for feature enhancement to obtain the current state value;

[0051] Step S14: Calculate the advantage function according to the cumulative return and the current state value;

[0052] Step S15: Update the A-PPO model using the clip mechanism and the Adam optimizer according to the advantage function;

[0053] Step S16: If the determination times of the current process parameters reach the second threshold, output the first process parameter; otherwise, repeat Steps S10 to S16.

[0054] Further, the final process parameter acquisition step includes:

[0055] Step S20: Obtain the Bayesian objective function according to the reward function;

[0056] Step S21: Obtain a sample set according to the first process parameter and random sampling;

[0057] Step S22: According to the sample set, the Bayesian optimization uses the Gaussian process model to calculate the EI, and selects the sample point with the largest EI value as the sampling point;

[0058] Step S23: Obtain the Bayesian optimization value of the sampling points using the image acquisition unit and the image detection unit according to the Bayesian objective function;

[0059] Step S24: Compare the Bayesian optimization value of the first process parameter with the Bayesian optimization value of the sampling points to obtain the next sampling point, and update the Gaussian process model according to the next sampling point;

[0060] Step S25: If the Gaussian process model reaches the preset number of iterations, output the final process parameter; otherwise, repeat steps S21 to S25.

[0061] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0062] 1. During the stamping process of carbon fiber composite materials, defects such as cracks, bubbles, and fiber shedding are likely to occur in the finished products. Traditional detection methods rely on manual labor or simple equipment, making it difficult to achieve high-precision and automated detection. Through the image acquisition unit, the system can obtain the image of the formed carbon fiber composite material in real time, and use the image detection unit to analyze and classify the defects in the carbon fiber composite material image in combination with the ResNet model and the random forest model, so as to obtain the quality result of the finished product. This can not only identify and classify defects in a timely manner, but also improve the quality and production efficiency of the finished product.

[0063] 2. Traditional process parameter optimization methods usually rely on empirical settings or repeated trial and error, making it difficult to quickly respond to complex process changes and impossible to achieve precise control in a changing production environment. By introducing an attention mechanism into the PPO model to form an A-PPO model, it is possible to comprehensively consider real-time state information and dynamically adjust process parameters based on the reward function. Combining the A-PPO model with the quality result of the finished product to generate the current optimal process parameter realizes the adaptive control of complex non-linear stamping processes, not only reducing the generation of defects, but also improving the control accuracy and production efficiency.

[0064] 3. In order to further improve the process efficiency and the quality of the finished product during the forming process of carbon fiber composite materials, the final parameter determination unit introduces Bayesian optimization. Based on the first process parameter generated by A-PPO, Bayesian optimization models the process parameter space of carbon fiber composite materials through a Gaussian process model, and uses the expected improvement function to sample and evaluate new process parameters, gradually approaching the global optimal solution. It can achieve more fine-tuning in the complex multi-dimensional process parameter space of carbon fiber composite materials, reduce the uncertainty of process parameter adjustment, and improve the quality and process efficiency of the finished product. Description of the Drawings

[0065] Figure 1Schematic diagram of the structure of a forming optimization control system for a carbon fiber composite material provided by an embodiment of the present invention;

[0066] Figure 2 Schematic diagram of the process for obtaining the finished product quality result provided by an embodiment of the present invention;

[0067] Figure 3 Schematic diagram of the process for obtaining the first process parameters provided by an embodiment of the present invention;

[0068] Figure 4 Schematic diagram of the process for obtaining the final process parameters provided by an embodiment of the present invention. Detailed implementation manners

[0069] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0070] Carbon fiber composite materials are widely used in high-tech fields such as aerospace, automotive manufacturing, and wind power generation due to their excellent mechanical properties and lightweight characteristics. However, the forming process of carbon fiber composite materials is complex and requires extremely high precision. Especially during the stamping forming process, the die state and the selection of forming process parameters directly affect the quality of the finished product. For example, temperature and stamping pressure play crucial roles in the performance of carbon fiber composite materials. Excessive temperature will cause the material to melt excessively, reducing strength and stiffness; too low temperature will not fully melt the material, affecting the forming quality. Excessive stamping pressure may cause uneven deformation or cracks in the material, while insufficient pressure may result in incomplete forming. In addition, parameters such as stamping speed and die displacement also have a significant impact on the forming effect. Therefore, the reasonable selection of die temperature, die displacement, stamping speed, and pressure is the key to optimizing the forming process of carbon fiber composite materials.

[0071] Most traditional stamping forming processes rely on empirical settings or trial-and-error methods to adjust process parameters. This method not only fails to achieve automated precise control but also cannot respond in a timely manner to deviations in process parameters, resulting in the generation of product defects. Although existing optimization schemes can automatically adjust some process parameters according to the requirements of different products, most current defect detection methods rely on manual detection or simple automated equipment, with limited detection efficiency. They cannot effectively identify and classify complex defect types, nor can they further refine the adjustment of process parameters based on the detection results, leading to large fluctuations in product quality and difficulty in further improving production efficiency.

[0072] A certain factory currently uses traditional stamping forming technology. Although it has basic production capabilities, there are still problems of insufficient efficiency and unstable quality control. In this embodiment, the forming die of carbon fiber composite material in a certain factory is taken as the control object, aiming to improve production efficiency and finished product quality through a forming optimization control system for carbon fiber composite materials. Please refer to Figures 1 to 4 , and the technical solution is as follows:

[0073] As Figure 1 shown, a forming optimization control system for carbon fiber composite materials includes:

[0074] A historical data acquisition unit for acquiring the historical forming data of carbon fiber composite materials. The historical forming data includes historical operation data, template images, and defect images, providing basic support for subsequent process optimization and defect analysis.

[0075] Specifically, key process parameters during each forming process are recorded in real time through temperature sensors, displacement sensors, pressure sensors, and speed sensors. These sensors are distributed at key positions of the die and forming equipment to ensure the comprehensiveness and accuracy of the data; high-resolution industrial cameras are used or template images are obtained from previous practices, and the die state and process parameters are recorded. These images and parameters are used for subsequent defect comparison and process parameter adjustment; after each forming, high-resolution industrial cameras are also used to obtain defect images of the finished product. These images are preprocessed and stored through image processing software for training deep learning models to identify and classify defect types.

[0076] A die state acquisition unit for acquiring the current state information of the carbon fiber die;

[0077] Among them, the current state information of the carbon fiber die includes die temperature, die displacement, stamping speed, stamping pressure, and the finished product quality result, expressed as:

[0078] s(t) = {T(t), D(t), V(t), P(t), Q(t)};

[0079] Among them, s(t) is the state information of the carbon fiber die at time t, T(t) is the die temperature at time t, D(t) is the die displacement at time t, V(t) is the stamping speed at time t, P(t) is the stamping pressure at time t, and Q(t) is the finished product quality result at time t.

[0080] By acquiring the current state information of these dies, the system can comprehensively master the key process parameters during the stamping forming process, timely detect deviations and automatically adjust, effectively reduce material defects, and improve production efficiency and finished product quality. For example, when abnormal stamping pressure is detected, the system can automatically adjust process parameters to avoid quality problems caused by overpressure or underpressure.

[0081] Specifically, monitoring the mold temperature has an important impact on the forming effect of carbon fiber composites. Appropriate temperature control can avoid excessive melting or insufficient curing of the material, thereby improving the stability of the finished product quality. By monitoring the mold displacement information, the pressing accuracy during the forming process can be reflected, ensuring that the mold performs stamping operations at the correct position and avoiding deformation of the finished product caused by displacement deviation. The speed change of the mold during the stamping process, too fast or too slow stamping speed will affect the forming uniformity and surface quality of carbon fiber composites. Monitoring the stamping speed helps to improve the consistency and strength of the finished product. The stamping pressure directly affects the density and forming strength of the material. By monitoring the stamping pressure, the system can avoid overpressure or insufficiency of carbon fiber composites. The quality results of the finished product are used to judge whether the current process parameters are reasonable and provide a basis for subsequent parameter adjustment.

[0082] A process parameter determination unit for establishing an A-PPO model according to the current state information, where the A-PPO model is used to determine current process parameters according to the attention mechanism.

[0083] Furthermore, the process of determining the current process parameters includes:

[0084] Input the current state information into the A-PPO model, and the A-PPO model performs feature enhancement according to the current state information and outputs normal distribution parameters. The A-PPO model focuses on parameters that have a greater impact on the forming quality. For example, the model may assign higher weights to the mold temperature and stamping pressure because they have a greater impact on the material properties.

[0085] Among them, the normal distribution parameters of the current process parameters output by the A-PPO model are expressed as:

[0086] (avg(t), std(t)) = APPO(s(t));

[0087] Among them, avg(t) is the mean value of the process parameters at time t, std(t) is the standard deviation of the process parameters at time t, and APPO() is the A-PPO model;

[0088] Perform random sampling according to the normal distribution parameters to obtain the current process parameters including normal distribution noise terms, expressed as:

[0089] a(t) = avg(t) + u × std(t);

[0090] Among them, a(t) is the process parameter at time t, and u ∼ N(0, 1) is the standard normal distribution noise.

[0091] Among them, the process parameters at time t include the mold temperature adjustment value, the mold displacement adjustment value, the stamping speed adjustment value, and the stamping pressure adjustment value, which are expressed as:

[0092] a(t) = {ΔT(t), ΔD(t), ΔV(t), ΔP(t)};

[0093] Among them, ΔT(t) is the mold temperature adjustment value at time t, ΔD(t) is the mold displacement adjustment value at time t, ΔV(t) is the stamping speed adjustment value at time t, and ΔP(t) is the stamping pressure adjustment value at time t.

[0094] Specifically, assume that during a certain stamping process, the average value of the mold temperature output by the A-PPO model is 180 °C and the standard deviation is 5 °C. Then the ideal value of the mold temperature is 180 °C, but there is a range fluctuation of ±5 °C. The system samples from the normal distribution of the mold temperature and obtains the adjusted temperature of 182 °C. At the same time, it samples from the normal distributions of the mold displacement, stamping speed, and stamping pressure to obtain the corresponding mold displacement adjustment value, stamping speed adjustment value, and stamping pressure adjustment value. By introducing random sampling, the system can explore different process parameter settings within a certain range, effectively avoiding the problem of finished product quality caused by fixed parameters, thereby improving production efficiency.

[0095] The image acquisition unit is used to acquire the carbon fiber composite material image according to the current process parameters and preprocess the carbon fiber composite material image to obtain the first image.

[0096] Furthermore, the process of obtaining the first image includes:

[0097] Perform gray conversion on the carbon fiber composite material image to remove color information and obtain the second image;

[0098] Among them, the size of the carbon fiber composite material image is 2048×2048 pixels and it consists of three channels: red (R), green (G), and blue (B). The formula for gray conversion is:

[0099] I GRAY = 0.2989×R + 0.5870×G + 0.1140×B;

[0100] Among them, I GRAY is the pixel value of the gray image;

[0101] Some areas may form a "white fog" phenomenon with higher brightness due to the presence of white resin, which affects the recognition of the texture structure of the carbon fiber composite material. According to the second image, using the dark channel prior method and combining with the fog model, the white fog can be effectively removed to obtain the third image;

[0102] Perform image denoising on the third image using Gaussian filtering, and perform image enhancement by adjusting the contrast and brightness to obtain a fourth image;

[0103] Among them, the direct impact of contrast and brightness on the subsequent edge detection accuracy of the carbon fiber composite material image, and the adjustment formula is:

[0104] I enhanced = a×I filtered + b;

[0105] Among them, I enhanced is the pixel value of the fourth image; I filtered is the pixel value of the denoised third image; a is the contrast coefficient, set to 1.2; b is the brightness adjustment value, set to 10;

[0106] Extract the edge contour of the carbon fiber composite material using the Canny edge detection algorithm based on the fourth image to obtain the first image.

[0107] Specifically, during a stamping process of a carbon fiber composite material, assume that due to excessive mold temperature, the material is over-melted and cracks are generated. First, convert the original color image to a grayscale image, and the brightness of the crack area is significantly enhanced; subsequently, use the dark channel prior method to remove the white fog generated by high temperature and restore the true texture of the material; after denoising, by adjusting the contrast and brightness, make the crack edges clearer and highlight the texture of the carbon fiber and the resin distribution; finally, apply the Canny edge detection algorithm to extract the edge contour of the carbon fiber composite material, which helps the forming optimization control system of the carbon fiber composite material quickly locate crack defects, thereby improving the detection efficiency and accuracy.

[0108] The image detection unit is used to extract features of the defect image using the ResNet model based on the historical forming data to obtain a carbon fiber feature vector, classify the defect type according to the template image and the defect image, and determine the defect level of the first image using the random forest model according to the defect type and the carbon fiber feature vector to obtain the finished product quality result.

[0109] Furthermore, the process of obtaining the carbon fiber feature vector includes:

[0110] Input the first image into the ResNet model. To ensure that the input image meets the training requirements of the model, standardize the pixel values of the image at the input layer of the model and scale the pixel values to the range of 0 to 1;

[0111] Perform layer-by-layer convolution operations on the first image using convolutional layers. Each layer of convolution uses different convolutional kernels to extract low-level features of different scales. The low-level features include the edge contour and texture direction of the carbon fiber composite material;

[0112] Among them, the first convolutional layer uses a 7×7 convolutional kernel with a stride of 2 to extract the basic edges and geometric structures of the image. The subsequent convolutional layers use 3×3 convolutional kernels to gradually extract the edge contours and texture directions of the carbon fiber composite material;

[0113] After several convolutional layers, the ResNet model uses a pooling layer to reduce the size of the feature map and retain the most significant feature information;

[0114] According to the low-level features, residual blocks are used to extract high-level features. The residual blocks use small 3×3 convolutional kernels. The high-level features include the texture structure and local defects of the carbon fiber composite material;

[0115] Among them, the local defects are tiny surface defects, cracks, and flaws, which can help detect the quality problems of the carbon fiber composite material;

[0116] According to the high-level features, a global average pooling layer is used to integrate all the feature maps extracted by the convolutional layer into a fixed-length carbon fiber feature vector, and the carbon fiber feature vector is output.

[0117] Among them, the length of the carbon fiber feature vector is set to 512 dimensions, which contains the texture structure and tiny defect information of the carbon fiber composite material, helps the detection system quickly identify quality problems, and reduces errors.

[0118] Specifically, when processing a batch of carbon fiber composite material samples, by obtaining the carbon fiber feature vector, the system successfully and automatically detected tiny surface cracks and texture unevenness problems and issued a warning in a timely manner, which not only helps improve the detection accuracy and efficiency but also effectively improves the quality control level in the production process.

[0119] Furthermore, as Figure 2 shown, the process of obtaining the finished product quality result includes:

[0120] Dividing the defect levels according to the severity of the defect types, and marking the defect images according to the defect levels and the template image to obtain the first data set;

[0121] Among them, the defect types include cracks, bubbles, fiber shedding, and uneven distribution;

[0122] Among them, according to expert experience, if a slender notch in the defect image has a length greater than 5 mm and a width greater than 0.2 mm, it is determined as a crack; if any area in the defect image has abnormal brightness and a transparent or semi-transparent circular area with a diameter exceeding 2 mm, it is marked as a bubble; if the visible fiber loosening or shedding area in the defect image exceeds 1 square millimeter, it is marked as fiber shedding; if the standard deviation of the carbon fiber density in the defect image is greater than 5%, it is marked as uneven distribution;

[0123] The defect level is divided into minor, medium, and severe according to the severity of the defect type;

[0124] Among them, if the number of cracks per unit area is within 1, the bubble diameter is less than 5 mm, the fiber shedding does not exceed 5, there is no obvious uneven distribution and the standard deviation does not exceed 3%, it is marked as grade A; if the number of cracks per unit area is within 3, the bubble diameter is less than 10 mm, the fiber shedding does not exceed 10, there is local uneven distribution but the standard deviation does not exceed 5%, it is marked as grade B; if the number of cracks per unit area is more than 3, or the bubble diameter exceeds 10 mm, or the fiber shedding exceeds 10, or the standard deviation of the significant uneven distribution exceeds 5%, it is marked as grade C;

[0125] Use the ResNet model to extract features from the defect image to obtain the carbon fiber feature vector;

[0126] Input the first data set and the carbon fiber feature vector into the random forest model, and the random forest model uses majority voting to determine the defect level to obtain a trained random forest model;

[0127] Specifically, the first data set is divided into a training set and a test set. The training set contains 600 images, and the test set contains 150 images. The carbon fiber feature vector is used as the feature set, and the defect level label Y_severity corresponding to each image is used as the label set. Use the feature vector and defect level label Y_severity of the training set to train the random forest model to obtain a trained RF_severity model.

[0128] The accuracy, precision, recall, and F1-score of RF_severity were evaluated using the test set, and a confusion matrix was generated, as shown in Table 1, which shows the evaluation results of the RF_severity model. Table 2 shows the confusion matrix of the defect levels. The numbers in each cell represent the matching situation between the model's prediction results and the true labels. The numbers on the diagonal represent the number of correct predictions, and the numbers off the diagonal represent incorrect predictions. Combining Table 1 and Table 2, it can be seen that the RF_severity model can well predict the defect levels of carbon fiber composites, thus laying a good foundation for subsequent first-image defect prediction.

[0129] Table 1. Evaluation Results of the RF_severity Model

[0130] Indicator Accuracy Precision Recall F1 Score Result 0.88 0.87 0.88 0.87

[0131] Table 2. Confusion Matrix Results

[0132] True Class / Predicted Class Grade A Grade B Grade C Grade A 45 3 2 Grade B 5 40 5 Grade C 0 5 45

[0133] Next, defect detection on the production line was carried out. The ResNet model was used to extract features from the first image to obtain the carbon fiber feature vector.

[0134] The carbon fiber feature vector was input into the trained RF_severity random forest model. Based on the carbon fiber feature vector, the trained RF_severity random forest model predicted the defect level of the first image.

[0135] Based on the defect level of the first image, the finished product quality result was obtained.

[0136] Among them, the finished product quality result is the defect level value. The defect level values include 1.0 for grade A, 0.8 for grade B, and 0.5 for grade C.

[0137] With the carbon fiber feature vector extracted by the ResNet model and the trained RF_severity random forest model, not only can the defect level be accurately identified, enabling the production line to quickly identify finished product quality problems, but also a quantitative index for the finished product quality can be provided, facilitating subsequent optimization control and improvement of the finished product, thereby improving the overall production efficiency and product quality.

[0138] As Figure 3 shown, the process parameter optimization unit is used to calculate the finished product reward value using a reward function according to the finished product quality result, and re-determine the current process parameters using the A-PPO model based on the finished product reward value and the current state information to obtain the first process parameters.

[0139] Furthermore, the deep neural network of the A-PPO includes:

[0140] An input layer for receiving the current state information, monitoring the die state in real time through sensors, and inputting the die temperature, displacement, stamping speed, stamping pressure, and the finished product quality result into the input layer;

[0141] A feature extraction layer for extracting the current state information features using convolutional layers and fully connected layers;

[0142] An attention mechanism layer for weighting the current state information features to obtain a comprehensive feature representation. For example, since the die temperature and stamping pressure have a significant impact on the die performance, this mechanism will enhance the weights of these two features, while relatively weakening the impact on other features;

[0143] A policy network layer for outputting the current process parameters according to the comprehensive feature representation, and a value network layer for outputting the current state value according to the comprehensive feature representation.

[0144] The A-PP0 deep neural network combined with the attention mechanism can adapt to changes in production conditions and effects in real time, dynamically adjust process parameters, and can pay more attention to features that have a greater impact on die performance. For example, if the output value is 0.9, indicating that the current process parameters are appropriate, the system will decide to continue using the current process parameters, thereby improving the production efficiency and flexibility of carbon fiber composite materials.

[0145] Furthermore, the steps for obtaining the reward function include:

[0146] Multiplying the system energy consumption by the first weight factor to obtain an energy consumption penalty term;

[0147] Multiplying the production time by the second weight factor to obtain a time penalty term; where the production time is the time required for the defect level to reach grade A;

[0148] Subtracting the energy consumption penalty term and the time penalty term from the finished product quality result to obtain the finished product reward value, which is expressed by the formula:

[0149] R = Q - (w 1 × E) - (w 2 × T);

[0150] where R is the finished product reward value, Q is the finished product quality result, E is the system energy consumption, T is the production time, w 1 and w 2 are the first weight factor and the second weight factor.

[0151] where w 1 and w 2They are set to 0.2 and 0.1 respectively. Suppose that in a certain production process, the finished product quality result of the carbon fiber composite material is 0.9, the system energy consumption is 500 units, and the production time is 10 hours. After normalizing the three, the finished product reward value is approximately 0.666. This calculation process can ensure that the finished product quality is always the core goal during the optimization process. And by setting the energy consumption penalty term and the time penalty term, the A-PPO model can be effectively guided to achieve multi-objective optimization during the production process, including reducing energy consumption, shortening production time, and improving the quality of the finished product, avoiding the negative impacts caused by single-objective optimization, thereby improving the overall production efficiency.

[0152] Further, the first process parameter acquisition process includes:

[0153] Step S10: Obtain the finished product reward value and the next moment state information according to the current state information and the current process parameters, and store the current state information, the current process parameters, the finished product reward value, and the next moment state information into the trajectory set;

[0154] Step S11: Determine whether the number of trajectories in the trajectory set reaches the first threshold, such as 100 trajectories;

[0155] Step S12: If the number of trajectories reaches the first threshold, calculate the cumulative return using the value function according to the trajectory set, expressed as:

[0156] G(t) = R(t) + u × R(t + 1) + u 2 × R(t + 2) +...;

[0157] Among them, G(t) is the cumulative return at time t, R(t) is the reward function at time t, u is the discount factor of the future reward function, set to 0.9, and R(t + 1), R(t + 2) are all future reward functions predicted based on the trajectory set and the current state information;

[0158] Otherwise, use the A-PPO model to determine the next moment process parameters according to the next moment state information, and execute steps S10 to S11 according to the next moment process parameters and the next moment state information;

[0159] Among them, the next moment state information is used as the current state information of the next trajectory, and the next moment process parameters are used as the current process parameters of the next trajectory;

[0160] Step S13: Input the current state information into the A-PPO model for feature enhancement to obtain the current state value V(s(t));

[0161] Step S14: Calculate the advantage function according to the cumulative return and the current state value. The advantage function is expressed as:

[0162] A(s,a) = G(t) - V(s(t));

[0163] Wherein, A(s,a) is the advantage function, which is used to measure the advantages of process parameters and help the A-PPO model make better decisions;

[0164] Step S15, update the A-PPO model according to the advantage function using the clip mechanism and the Adam optimizer;

[0165] Among them, the clip mechanism is a strategy update method, which is widely used especially in the PPO model. By introducing the Clip mechanism, over-update is prevented, ensuring the stability of the learning process of the A-PPO model and avoiding performance degradation caused by improper update;

[0166] Step S16, if the current number of times of determining process parameters reaches the second threshold, such as 10 adjustments, output the first process parameter; otherwise, add 1 to the number of determinations and execute steps S10 to S16.

[0167] Specifically, for example, if the mold temperature is adjusted too high, resulting in excessive melting of the material and defects, the reward value of the A-PPO model decreases after detecting the change in the mold temperature. At this time, the feature weight of the temperature parameter will be enhanced through the attention mechanism, and the adjusted value of the mold temperature will be output again. At the same time, the system automatically adjusts the stamping speed and pressure according to the adjusted mold temperature. After outputting the current process parameters, if the mold temperature still affects the material forming, the A-PPO model will adaptively adjust the current process parameters repeatedly based on the new current state information and historical data. Through the above steps, the process parameters can be adaptively adjusted according to real-time data and historical data, and can also meet the requirements of different production conditions, thus improving the product quality and production efficiency.

[0168] The final parameter determination unit, as Figure 4 shown, is used to further adjust the first process parameter using Bayesian optimization to obtain the final process parameter.

[0169] Furthermore, the steps for obtaining the final process parameter include:

[0170] Step S20: Obtain the Bayesian objective function f(x) according to the reward function;

[0171] Step S21: Obtain a sample set {x 1 , x 2 , x 3 ..., x n}, the Bayesian optimization uses a Gaussian process model as a surrogate model, and inputs the sample set into the Gaussian process model. The Gaussian process model provides the predicted mean and standard deviation of each set of process parameters, thereby quantifying uncertainty;

[0172] Step S22: Calculate the EI according to the Gaussian process model and select the sample point with the largest EI value in the sample set as the sampling point, which is expressed as:

[0173] x next = argmax x EI(x);

[0174] where x next is the optimal process parameter; EI(x) is the improvement expectation function at the given sample point x, which is expressed as:

[0175] EI(x) = Ε[max(0, B(x) - B * )];

[0176] where Ε is the expectation function, B(x) is the Bayesian objective function value at point x, and B * is the current known optimal value of the Bayesian objective function;

[0177] Step S23: Use the image acquisition unit and the image detection unit to obtain the Bayesian optimization value of the sampling point according to the Bayesian objective function;

[0178] Step S24: Compare the Bayesian optimization value of the first process parameter with the Bayesian optimization value of the sampling point to obtain the next sampling point, and update the Gaussian process model according to the next sampling point;

[0179] Step S25: Set the preset number of iterations to 10 times. If the Gaussian process model reaches the preset number of iterations, output the final process parameter; otherwise, repeat steps S21 to S25.

[0180] Specifically, assume that the first process parameter obtained in A-PPO is x initial = [150°C, 2.0 mm, 10 mm / s, 80 bar], and the corresponding finished product reward value after percentage conversion is Q(x initial ) = 80, then the objective function value f(x initial ) is also 80; the adjustment range of the mold temperature adjustment value is 140°C to 170°C, the adjustment range of the mold displacement adjustment value is 1.5 mm to 3.0 mm, the adjustment range of the stamping speed adjustment value is 8 mm / s to 15 mm / s, and the adjustment range of the stamping pressure adjustment value is 75 bar to 90 bar. Randomly select three groups of process parameter combinations according to the selection range of the adjustment values to obtain {x initial, x 1 , x 2 , x 3}; At the first iteration, B * is 75, calculate x 1 , x 2 , x 3 's EI value, and get EI(x 1 ) = 2, EI(x 2 ) = 5, EI(x 3 ) = 3, then select x 2 as x next , apply x 2 to the image acquisition unit and the image detection unit, and get the objective function value f(x 2 ) = 82. After comparing with f(x initial ), f(x 2 ) becomes the new B * . After 10 iterations, Bayesian optimization can further finely optimize the process parameters on the basis of the existing parameter adjustment, and can find the process parameters close to the global optimum with fewer test times, thereby improving the finished product quality and work efficiency.

[0181] The forming optimization control system of carbon fiber composite materials proposed by the present invention improves the forming efficiency and finished product quality of carbon fiber composite materials through the coordinated action of multiple functional units. The historical data acquisition unit can make full use of historical forming data to ensure that decisions are based on comprehensive information; the mold state acquisition unit monitors the mold state in real time, enhancing the real-time feedback of the production process; the process parameter determination unit realizes the precise adjustment of process parameters through the A-PPO model, optimizing the production efficiency; the image acquisition and detection unit combines deep learning technology to effectively identify the defect type and defect level, thereby improving the finished product quality; the process parameter optimization unit makes feedback adjustments according to the finished product quality results to ensure the self-adaptability of the production process; the final parameter determination unit further improves the accuracy of process parameters through the Bayesian optimization method, ensuring the high-quality output and production efficiency of products.

[0182] Although the embodiments of the present invention have been shown and described, for those of ordinary skill in the art, it can be understood that various changes, modifications, substitutions and variations can be made to these embodiments without departing from the principles and spirit of the present invention. The scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A carbon fiber composite material forming optimization control system, characterized in that: include: A historical data acquisition unit, used to acquire historical molding data of the carbon fiber composite material, wherein the historical molding data includes historical operation data, defect images and template images; A mold status acquisition unit is used to obtain the current status information of the carbon fiber mold; A process parameter determination unit, used to establish an A-PPO model according to the current state information, wherein the A-PPO model is used to determine the current process parameters according to the attention mechanism; an image acquisition unit, configured to acquire an image of the carbon fiber composite material according to the current process parameters, and preprocess the image of the carbon fiber composite material to obtain a first image; An image detection unit is used to extract features of the defect image using a ResNet model according to the historical molding data to obtain a carbon fiber feature vector, and to classify defect types according to the template image and the defect image, and to determine a defect level of the first image using a random forest model according to the defect type and the carbon fiber feature vector to obtain a finished product quality result; A process parameter optimization unit, configured to calculate a finished product reward value using a reward function according to the finished product quality result, and to re-determine the current process parameter using the A-PPO model according to the finished product reward value and the current state information to obtain a first process parameter; The final parameter determination unit is used to adjust the first process parameter using Bayesian optimization to obtain the final process parameter.

2. The carbon fiber composite material forming optimization control system according to claim 1, characterized in that: The current status information of the carbon fiber mold includes mold temperature, mold displacement, stamping speed, stamping pressure and the finished product quality result.

3. The carbon fiber composite material forming optimization control system according to claim 1, characterized in that: The current process parameter determination process includes: Inputting the current state information into the A-PPO model, the A-PPO model performs feature enhancement according to the current state information and outputs normal distribution parameters, and performs random sampling according to the normal distribution parameters to obtain the current process parameters including the normal distribution noise term; The current process parameters include a mold temperature adjustment value, a mold displacement adjustment value, a stamping speed adjustment value, and a stamping pressure adjustment value.

4. The carbon fiber composite material forming optimization control system according to claim 1, characterized in that: The first image acquisition process includes: Performing grayscale conversion on the carbon fiber composite material image to obtain a second image; According to the second image, a dark channel prior method is used to remove white fog to obtain a third image; Perform image denoising using Gaussian filtering on the third image, and perform image enhancement by adjusting contrast and brightness to obtain a fourth image; The first image is obtained by extracting the edge contour of the carbon fiber composite material using the Canny edge detection algorithm according to the fourth image.

5. The carbon fiber composite material forming optimization control system according to claim 1, characterized in that: The carbon fiber feature vector acquisition process includes: Extracting low-level features of the first image using convolutional layers and pooling layers, wherein the low-level features include edge contours and texture directions of the carbon fiber composite material; Extracting high-level features using a residual block based on the low-level features, wherein the high-level features include texture structures and local defects of the carbon fiber composite material; The carbon fiber feature vector is obtained using a global average pooling layer according to the high-level features.

6. The carbon fiber composite material forming optimization control system according to claim 1, characterized in that: The A-PPO deep neural network includes: An input layer, used for receiving the current state information; Feature extraction layer, used to extract current state information features using convolutional layers and fully connected layers; An attention mechanism layer is used to perform weighted processing on the current state information features to obtain a comprehensive feature representation; A strategy network layer, used for outputting the current process parameters according to the comprehensive feature representation; The value network layer is used to output the current state value according to the comprehensive feature representation.

7. The carbon fiber composite material forming optimization control system according to claim 1, characterized in that: The process of obtaining the finished product quality results includes: Classifying the defect level according to the severity of the defect type, and marking the defect image according to the defect level and the template image to obtain a first data set; Using the ResNet model to extract features from the defect image to obtain the carbon fiber feature vector; Inputting the first data set and the carbon fiber feature vector into the random forest model, wherein the random forest model determines the defect level by majority voting to obtain a trained random forest model; Using the ResNet model to extract features from the first image, to obtain the carbon fiber feature vector; Inputting the carbon fiber feature vector into the trained random forest model to obtain a defect level of the first image; The finished product quality result is obtained according to the defect level of the first image.

8. The carbon fiber composite material forming optimization control system according to claim 1, characterized in that: The step of obtaining the reward function includes: Multiply the system energy consumption by the first weight factor to obtain an energy consumption penalty term; Multiply the production time by the second weight factor to obtain the time penalty term; The finished product reward value is obtained by subtracting the energy consumption penalty item and the time penalty item from the finished product quality result.

9. The carbon fiber composite material forming optimization control system according to claim 1, characterized in that: The first process parameter acquisition process includes: Step S10: obtaining the finished product reward value and the next moment state information according to the current state information and the current process parameters, and storing the current state information, the current process parameters, the finished product reward value and the next moment state information into a trajectory set; Step S11: determining whether the number of trajectories in the trajectory set reaches a first threshold; Step S12: if the number of trajectories reaches the first threshold, the cumulative reward is calculated according to the trajectory set using the value function; otherwise, the next moment process parameters are determined according to the next moment state information using the A-PPO model, and steps S10 to S11 are executed according to the next moment process parameters and the next moment state information; The next moment state information is used as the current state information of the next track, and the next moment process parameters are used as the current process parameters of the next track; Step S13: inputting the current state information into the A-PPO model for feature enhancement to obtain the current state value; Step S14: Calculate an advantage function according to the accumulated rewards and the current state value; Step S15: updating the A-PPO model using the clip mechanism and the Adam optimizer according to the advantage function; Step S16: If the number of determinations of the current process parameter reaches a second threshold, the first process parameter is output; otherwise, steps S10 to S16 are repeatedly executed.

10. The carbon fiber composite material forming optimization control system according to claim 1, characterized in that: The final process parameter acquisition step comprises: Step S20: Obtaining a Bayesian objective function according to the reward function; Step S21: obtaining a sample set according to the first process parameters and random sampling; Step S22: According to the sample set, the Bayesian optimization uses a Gaussian process model to calculate the EI, and selects the sample point with the largest EI value as the sampling point; Step S23: using the image acquisition unit and the image detection unit to obtain the Bayesian optimization value of the sampling point according to the Bayesian objective function; Step S24: comparing the Bayesian optimization value of the first process parameter with the Bayesian optimization value of the sampling point to obtain the next sampling point, and updating the Gaussian process model according to the next sampling point; Step S25: If the Gaussian process model reaches a preset number of iterations, the final process parameters are output; otherwise, steps S21 to S25 are repeated.

Images