Convolutional neural network-based low-pressure cast aluminum alloy hub structure prediction method

Abstract

The invention discloses a low-pressure cast aluminum alloy hub structure prediction method based on a convolutional neural network. The method comprises the following steps: step 1, obtaining and establishing a casting process parameter and product microstructure morphology database; 2, parameter data screening and data preprocessing; 3, building a convolutional neural network prediction model architecture; 4, training the established convolutional neural network prediction model through database data; and 5, verifying the reliability and accuracy of the prediction capability of the model. The defect that in the aluminum alloy hub production process, after molten aluminum is solidified in a metal mold, the internal organization structure is difficult to observe is overcome, the cooling rate, the pouring temperature, the alloy components and the heat treatment technology in the casting process are adjusted, intervention is conducted in time to obtain the optimal organization structure, the performance of a casting is improved through optimized technological parameters, and the service life of the casting is prolonged. Potential defects can be found in advance in the casting production process through accurate structure prediction, the rejection rate and the production cost are reduced, and energy consumption is reduced.

Description

Technical Field

[0001] The present invention belongs to aluminum alloy casting technology and is applied to the field of low-pressure casting of aluminum alloy wheels. The present invention is a prediction model obtained by using a convolutional neural network, and a method for predicting the microstructure of low-pressure cast aluminum alloy wheels by using the model. Background Art

[0002] Aluminum alloy materials are widely used in aviation, automobile, construction, electronics and other fields due to their advantages of light weight, high strength, corrosion resistance, easy processing and forming, etc. At the same time, facing the current hot issue of automobile lightweighting in the industry, the use of aluminum alloys to manufacture automobile parts such as frames, bodies, wheels, engine parts, etc. can effectively reduce the overall weight of the car, reduce energy consumption, and improve environmental protection and economy, thus becoming one of the important ways to achieve lightweighting of the car body.

[0003] Nowadays, most aluminum alloy wheel hubs are produced using low-pressure casting (LPDC), including mold closing, pressure increasing, filling, pressure holding, pressure relief, mold opening and other steps. Specifically, it means applying gas pressure on the surface of the aluminum liquid in the sealed insulation furnace after mold closing. As the pressure in the furnace increases, the aluminum liquid rises from the riser under the action of gas pressure and enters the mold cavity above. When the aluminum liquid completely fills the cavity, this process is called filling. Then the pressure is increased to solidify the aluminum liquid according to the shape of the mold. When the aluminum liquid has completed filling and the gate position is completely solidified, the pressure is stopped, and finally the pressure is released to demold and remove the parts.

[0004] Aluminum alloy low-pressure casting is widely used and is suitable for producing aluminum alloy castings with complex shapes and thin walls. Compared with high-pressure casting and gravity casting, low-pressure casting has a higher material utilization rate. The aluminum liquid has good fluidity in the mold, the outline of the casting is clearer, the surface roughness is lower, and it also has the advantages of effectively preventing the oxidation inclusions generated during the pouring process.

[0005] Invention patent CN116384159A discloses a method and system for temperature simulation and macroscopic structure prediction of continuous casting process, which is characterized by obtaining parameters and storing the parameters in MongoDB database, setting the model according to the parameters stored in MongoDB database, performing temperature simulation calculation, and then using the macroscopic structure prediction submodel to perform macroscopic structure prediction calculation, and storing the calculation results in MongoDB database. The purpose of temperature control calibration and macroscopic structure control in the continuous casting process is achieved through the calculation results.

[0006] Invention patent CN113240096A discloses a method for predicting the microstructure of a casting cylinder head based on rough sets and neural networks. The rough set theory is used to simplify the various index attributes of casting and heat treatment processes that affect the microstructure of the material, select the index with a larger weight, and reduce the input dimension of the neural network at the same time, thereby enhancing the learning efficiency of the neural network and improving the accuracy and efficiency of the microstructure prediction to a certain extent. The purpose of reducing the metal consumption rate and the casting scrap rate is achieved, and the design and production costs are reduced.

[0007] Invention patent CN106055844A discloses a method for predicting and controlling the microstructure of a nickel-based high-temperature alloy based on a BP neural network. The method is characterized in that a neural network model is predicted and controlled based on offline training of historical die forging process parameters and microstructures, and the prediction neural network model is used to predict the microstructure at the next moment, and the predicted microstructure is feedback-corrected and the target value of the microstructure at the next moment is planned. Finally, based on the feedback correction value and target value of the microstructure, the control neural network model gives the die forging process parameters of the system at the current moment.

[0008] Although cast wheels are widely used in the automotive industry, especially for their cost efficiency and design flexibility, during the low-pressure casting process of aluminum alloy wheels, the mold is non-transparent and a series of solidification, crystallization and phase change behaviors of the aluminum liquid after mold closing are all carried out in the closed mold cavity, so it is impossible to directly monitor and evaluate the internal structure in real time. Temperature control, metal flow and cooling rate in the casting process have a great influence on the product structure and performance, and the microstructure of the cast wheel may contain segregation and inclusions, pores, microcracks, second phase particles, etc., which will reduce the overall strength and fatigue performance of the wheel. Therefore, accurate analysis of microstructural characteristics is crucial to optimize the performance of cast aluminum alloy wheels. By precisely controlling the casting and heat treatment processes, the quality and durability of the wheel can be significantly improved.

[0009] At present, in the actual production field, the acquisition of casting structure data usually uses traditional physical simulation to make samples to simulate the cooling and solidification process in the actual casting process. By observing and testing samples, reflecting the structure information, or based on historical data and experience, the developed formula can predict certain properties of castings, such as grain size, segregation degree, etc. These formulas are usually based on simplified physical principles and experimental data. In addition, the purpose of structure prediction can also be achieved through computer software simulation analysis, but various unstable factors in the actual production process usually limit real-time application. Summary of the invention

[0010] The purpose of the present invention is to provide a method for predicting the structure of low-pressure cast aluminum alloy wheels based on convolutional neural networks, which can adapt to different production conditions and changes and respond quickly to real-time production data, in order to solve the problem that the structure characteristics of various parts of low-pressure cast aluminum alloy wheels cannot be directly observed and predicted in real time. Compared with product trial production and simulation, it can predict products quickly and accurately. At the same time, through accurate prediction, potential quality problems can be foreseen before casting. Through timely intervention, the flexibility of the production process can be increased, the overall product quality can be improved, the scrap rate can be reduced, and the material and energy consumption can be reduced.

[0011] A specific implementation process of a low-pressure cast aluminum alloy wheel structure prediction method based on a convolutional neural network includes the following steps:

[0012] Step 1: Acquire and establish a database of casting process parameters and product microstructure morphology;

[0013] The casting process parameters include: air cooling channel opening time, water cooling channel opening time, aluminum liquid casting temperature, mold temperature, filling time, filling pressure, filling speed, holding pressure, pressurization time, holding time and pressure relief time.

[0014] The microstructure morphology data include: metallographic photographs of the corresponding data collection points and the use of image recognition software to obtain the aluminum phase grain size, the average size of primary silicon, the average size of eutectic silicon, the area ratio of eutectic silicon, the area ratio of Mg2Si phase, the secondary dendrite arm spacing, and the planar porosity.

[0015] Step 2: Parameter data screening and data preprocessing;

[0016] Perform a quality check on the database data to remove erroneous or unclear samples to ensure the quality of the training data, and give priority to selecting parameters with higher characteristics as subsequent learning training data to ensure that the selected data can represent various situations that will be encountered in actual applications. Normalize the pixel values ​​of the image and adjust the image to meet the network requirements.

[0017] Step 3: Establish the convolutional neural network prediction model architecture;

[0018] To build a convolutional neural network, first select a suitable function to train the model. The network architecture includes input layer, convolution layer, pooling layer, fully connected layer and output layer. Configure the layer parameters. According to the complexity of the task and the characteristics of the data, you can adjust the parameters of each layer, such as the number of filters in the convolution layer and the number of neurons in the fully connected layer.

[0019] Step 4: Train the established convolutional neural network prediction model using database data.

[0020] Step 5: Compare the predicted results under any randomly selected process parameters with the actual workpiece samples produced under the process to verify the actual organizational morphology and the reliability and accuracy of the model prediction ability. If the relative error requirements are met, it proves that the prediction method is effective.

[0021] Beneficial effects: It overcomes the difficulty of observing the internal organizational structure after the aluminum liquid solidifies in the metal mold during the production of aluminum alloy wheels, and the defects such as non-uniformly distributed grains, poor second phases, and pores are easily generated inside the casting. By analyzing the relationship between the prediction results and the actual production parameters, the cooling rate, pouring temperature, alloy composition and heat treatment process in the casting process can be adjusted, so that timely intervention can be made in the production process to obtain the best organizational structure, and then the optimized process parameters can be obtained to improve the performance and life of the casting. At the same time, accurate organizational prediction can detect potential defects in advance during the production of castings, reduce scrap rates and production costs, and save energy consumption. The neural network model can analyze and provide feedback in real time, making production adjustments faster and more accurate. Compared with traditional trial production methods, this rapid response shortens the time for testing and error correction, significantly optimizes the production process, and improves overall production efficiency. It can bring substantial economic benefits and technological progress to the casting industry. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] Figure 1 It is a technical route flow chart of the present invention;

[0023] Figure 2 Schematic diagram of the neural network structure of the present invention. DETAILED DESCRIPTION

[0024] In order to facilitate the understanding of those skilled in the art, the present invention is further described below in conjunction with embodiments, and the contents mentioned in the implementation modes are not intended to limit the present invention.

[0025] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0026] The specific implementation method of step one is as follows:

[0027] (1) First, obtain the process parameters and metallographic photos of sensitive positions of the wheel from the historical database of the production workshop. Usually, cooling pipes and thermocouples are arranged in the wheel core, spokes, inner and outer rims, and other sensitive parts that are prone to defects. The microstructure and properties of these parts play an important role in the overall performance of the wheel hub. Therefore, the casting process parameters include: air cooling channel opening time, water cooling channel opening time, cooling channel diameter, aluminum liquid casting temperature, mold temperature, filling time, filling pressure, filling speed, holding pressure, pressurization time, holding time, and pressure relief time.

[0028] The organizational data are derived from scanning electron microscopy and optical microscopy, including metallographic photographs of the corresponding data collection points and calculation software to obtain the aluminum phase grain size, average primary silicon size, average eutectic silicon size, eutectic silicon area ratio, Mg2Si phase area ratio, secondary dendrite arm spacing, planar porosity, etc.

[0029] The specific implementation method of step 2 is as follows:

[0030] The metallographic photos in the casting database are annotated in detail using image recognition software, and the production process parameters and the metallographic photos of each part under the parameters are matched with index values ​​through database numbers.

[0031] According to the requirements of the convolutional neural network, the image is processed by the weighted average grayscale method. The weight formula is as follows:

[0032] Fg(i,j)=0.299×R(i,j)+0.587×G(i,j)+0.114×B(i,j)

[0033] In order to prevent overfitting of a single network when the amount of data is small, the mirroring and rotation operations are used to increase the diversity of training samples, and the pixels of the image are adjusted to 224×224×3. When adjusting the size, padding or cropping methods are selected to maintain the aspect ratio of the image to prevent image distortion. It is also necessary to normalize the pixel values ​​of the image from [0,255] to [0,1] to avoid numerical instability during the training process. The purpose of data preprocessing is to make the data set meet the definition domain and range of the model transfer function, so that the model training can be carried out smoothly and the accuracy of the model prediction results can be ensured. The final expanded data set reached 1266, and the data set was randomly divided into a training set and a validation set in a ratio of 8:2.

[0034] The specific implementation method of step three is as follows:

[0035] Convolutional neural network prediction model architecture is established. First, the size of the convolution kernel of the convolutional neural network is selected to be 3×3. During the convolution operation, the step size of the convolution kernel is set to 1. Each convolution block is connected by a 2×2 maximum pooling layer, which slides on the image with a moving rule of step size 2 to downsample the image. The data features extracted by convolution are input into the batch normalization BN layer, and the output value of each convolution layer is normalized to make the distribution of input data of two adjacent convolution layers relatively stable, making the network training process more stable and improving the generalization ability of the network. The normalization operation of the current input data ensures that the output of each layer is limited to a fixed range. The ReLu activation function performs nonlinear calculations. The activation function formula is as follows:

[0036] f(x)=max(0,x)

[0037] The ReLU function converts the negative part of the input x to zero, while the positive part remains unchanged. Graphically, the ReLU function is linear in the range x>0, and x It is always zero in the interval ≤0.

[0038] In addition, the SE module is added after the convolution layer to calculate the weights of each feature channel of the input image, thereby correcting the feature channels, highlighting important information, removing redundant information, and improving feature representation capabilities. The convolution formula is as follows:

[0039] z=AvgPool(X)

[0040] The result is then nonlinearly transformed using the ReLu activation function, and another fully connected layer is used to map the result back to the original dimension, and the Sigmoid activation function is used to limit the weights between 0 and 1. Two fully connected layers and activation functions are usually used to implement this excitation step to indicate the importance of the feature. The excitation operation process is as follows:

[0041] s=σ(ReLu(W2·ReLu(W1·z)))

[0042] Finally, by adding this parameter to the original features of each channel, the formula for recalibrating the original feature weights is as follows:

[0043] xc~=Fscale(uc,sc)=sc·uc

[0044] The specific implementation method of step 4 is as follows:

[0045] The training process includes the following steps: obtaining data and corresponding index values ​​from the training data set, obtaining the model's predicted output of the alloy structure through forward propagation, calculating the loss value between the predicted result and the true label, performing backpropagation to obtain the gradient information of each parameter, and using the gradient information to update the model's parameters. During the training process, the loss values ​​are accumulated to output the current average loss. In addition, the built-in optimizer of PyTorch is used to update the parameters, and the cross entropy loss function is used to calculate the loss value.

[0046] The specific implementation method of step five is as follows:

[0047] To verify the effectiveness of the prediction model, the remaining 20% ​​of the original data is selected as the test set as mentioned above. After training, if the prediction error of the prediction model is within a stable range, proving that the model training is successful, it means that the model can continue to be used for wheel hub organization prediction.

[0048] After the actual training, the model was tested on the remaining test set. By comparing the original organization image, the prediction results of the convolutional neural network for α-Al grain size, porosity and secondary dendrite arm spacing were compared with the experimental results. The predicted values ​​were close to the actual values. The maximum prediction error of porosity was 19.35%, the minimum was 5.8%, and the average error was 12.58%. The maximum prediction error of secondary dendrite arm spacing was 8.25%, the minimum was 3.73%, and the average error was 5.99%. The maximum prediction error of α-Al grain size was 8.98%, the minimum was 3.65%, and the average error was 6.32%. It is proved that the model can play a role in predicting the microstructure of low-pressure cast aluminum alloy wheels.

[0049] It will be understood by those skilled in the art that, unless otherwise defined, all terms (including technical and scientific terms) used herein have the same meaning as generally understood by those skilled in the art in the art to which the present invention belongs. It should also be understood that terms such as those defined in general dictionaries should be understood to have meanings consistent with the meanings in the context of the prior art, and will not be interpreted with idealized or overly formal meanings unless specifically defined as herein.

[0050] It should be understood that the detailed description of the technical solutions of the present invention by means of the preferred embodiments is illustrative rather than restrictive. A person skilled in the art may modify the technical solutions described in the embodiments, or replace some of the technical features by equivalents, based on reading the specification of the present invention; and these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for predicting the microstructure of low-pressure cast aluminum alloy wheel hub based on convolutional neural network, characterized in that: The following steps are involved: Step 1: Acquire and establish a database of casting process parameters and product microstructure morphology; Step 2: Parameter data screening and data preprocessing; Perform quality checks on the database data to remove erroneous or unclear samples, ensure the quality of the training data, normalize the pixel values ​​of the images and adjust the images to meet the network requirements; Step 3: Establish the convolutional neural network prediction model architecture; To build a convolutional neural network, first select a suitable function to train the model. The network architecture includes input layer, convolution layer, pooling layer, fully connected layer and output layer; configure layer parameters; Step 4: Train the established convolutional neural network prediction model using database data; Step 5: Compare the predicted results under any randomly selected process parameters with the actual workpiece samples produced under the process to verify the actual organizational morphology and the reliability and accuracy of the model's prediction ability.

2. The method according to claim 1, characterized in that The casting process parameters in step 1 include: air cooling channel opening time, water cooling channel opening time, aluminum liquid casting temperature, mold temperature, filling time, filling pressure, filling speed, holding pressure, pressurization time, holding time and pressure relief time; The microstructure morphology data include: metallographic photographs of the corresponding data collection points and the use of image recognition software to obtain the aluminum phase grain size, the average size of primary silicon, the average size of eutectic silicon, the area ratio of eutectic silicon, the area ratio of Mg2Si phase, the secondary dendrite arm spacing, and the planar porosity.

3. The method according to claim 1, characterized in that: In step 2, the metallographic photos in the casting database are annotated using image recognition software, and the production process parameters and the metallographic photos of each part under the parameters are mapped to index values ​​through database numbers; According to the requirements of the convolutional neural network, the image is processed by the weighted average grayscale method; To prevent overfitting of a single network when the amount of data is small, mirroring and rotation operations are used to increase the diversity of training samples, and the image pixels are adjusted to 224×224×3. When adjusting the size, padding or cropping methods are selected to maintain the aspect ratio of the image; the pixel values ​​of the image also need to be scaled from [0,255] to [0,1] for normalization.

4. The method according to claim 1, characterized in that In step 2, the final expanded data set reaches 1266 images, and the data set is randomly divided into training set and validation set in a ratio of 8:

2.

5. The method according to claim 1, characterized in that In step three, the parameters of each layer can be adjusted according to the complexity of the task and the characteristics of the data, such as the number of filters in the convolutional layer and the number of neurons in the fully connected layer.

6. The method according to claim 1, characterized in that In step 3, the convolutional neural network prediction model architecture is established. First, the size of the convolutional kernel of the selected convolutional neural network is 3×3. During the convolution operation, the step size of the convolutional kernel is set to 1. Each convolution block is connected by a 2×2 maximum pooling layer, which slides on the image with a moving rule of step size 2 to downsample the image; the data features extracted by convolution are input into the batch normalization BN layer, and the output value of each convolution layer is normalized. The normalization operation of the current input data and the ReLu activation function perform nonlinear calculations. The activation function formula is: f(x)=max(0,x); x is the input value; max is the maximum value; f(x) is the output result of the ReLu activation function. The ReLU function converts all negative values ​​in the input x to zero, while the positive value remains unchanged. Graphically, the ReLU function is linear in the interval x>0, and is always zero in the interval x≤0.

7. The method according to claim 1, characterized in that In step 3, the SE module is added after the convolution layer to calculate the weights of each feature channel of the input image, thereby correcting the feature channels, highlighting important information, removing redundant information, and improving feature representation capabilities; the convolution formula is as follows: z=AvgPool(X); where X is the input tensor, AvgPool is the average pooling operation; z is the output result after the pooling operation. The result is then nonlinearly transformed using the ReLu activation function, and another fully connected layer is used to map the result back to the original dimension, and the Sigmoid activation function is used to limit the weights between 0 and 1. Two fully connected layers and activation functions are usually used to implement this excitation step. The incentive operation process is as follows: s=σ(ReLu(W2·ReLu(W1·z))); where s is the final output value; σ is the activation function; ReLu is the activation function, representing the linear rectifier unit; W1 and W2 represent weight matrices. Finally, by adding this parameter to the original features of each channel, the formula for recalibrating the original feature weights is as follows: x c ~=F scale (u c ,s c )=s c ·u c ; where x c ~ is the scaled output; F scale (u c ,s c ) is the scaling function; s c is the scaling factor; u c For input data.

8. The method according to claim 1, characterized in that: In step 4, data and corresponding index values ​​are obtained from the training data set, and the model's predicted output for the alloy structure is obtained through forward propagation. The loss value between the predicted result and the true label is calculated, and back propagation is performed to obtain the gradient information of each parameter, and the gradient information is used to update the model parameters. During the training process, the loss value is accumulated to output the current average loss.

9. The method according to claim 1, characterized in that: In step five, in order to verify the effectiveness of the prediction model, the remaining original data is selected as the test set. After training, if the prediction error of the prediction model is within a stable range, proving that the model training is successful, the model will continue to be used to continue the hub organization prediction.

Images