CN116882585A

CN116882585A - Genetic algorithm and neural network coupled aluminum alloy hub low-pressure casting process optimization method

Info

Publication number: CN116882585A
Application number: CN202310941971.5A
Authority: CN
Inventors: 毕江; 贾四放; 董国疆; 李世德; 王佶; 郭世威
Original assignee: Yanshan University
Current assignee: Yanshan University
Priority date: 2023-07-28
Filing date: 2023-07-28
Publication date: 2023-10-13

Abstract

The invention discloses an optimization method of an aluminum alloy hub low-pressure casting process by coupling a genetic algorithm and a neural network, which aims at the problems of lower design accuracy of an aluminum alloy hub basic process and larger fluctuation of quality of a finished product, and combines a GA algorithm with a BP neural network to obtain a more excellent prediction model GA-BP regression prediction model. The weight and the threshold of the BP neural network are optimized through a genetic algorithm, so that the prediction accuracy of the BP neural network is improved, and the problems of excessive parameters, overfitting and the like can be avoided. Under the condition of multiple input and single output, the GA-BP regression prediction model is used, so that the process window prediction precision of the neural network can be further improved.

Description

Genetic algorithm and neural network coupled aluminum alloy hub low-pressure casting process optimization method

Technical Field

The invention relates to an aluminum alloy hub low-pressure casting process optimization method, in particular to an aluminum alloy hub low-pressure casting process optimization method based on genetic algorithm and neural network coupling.

Background

The aluminum alloy material has the advantages of low density, high specific strength and good corrosion resistance, and is widely applied to the field of automobiles. The aluminum alloy is adopted to manufacture the automobile parts, so that the weight of the automobile body can be effectively reduced, the energy consumption of the automobile can be reduced, and the maneuvering performance of the automobile can be improved.

Low pressure casting (low pressure die casting, LPDC) refers to a casting method in which molten aluminum is charged into a cavity of an aluminum mold and solidified under a relatively low pressure to obtain a high quality casting. And (3) rapidly lifting the aluminum liquid to the pouring gate along the liquid lifting pipe, then carrying out constant-current stable mold filling, carrying out pressure maintaining treatment after the mold filling is completed so as to reduce shrinkage cavities, cooling and solidifying, and then carrying out pressure relief to eject the casting.

The invention patent CN201811099788.0 discloses a low-pressure casting production data processing and process optimizing method. According to a large amount of data generated in the production process, the mold filling temperature, the mold filling time and the production cycle number are selected, and the processes of collection, processing, analysis training and the like are combined with a machine learning method, so that the relation between each parameter and the product yield is analyzed, main influencing factors are found, and the process is optimized to a certain extent.

The invention patent CN202110843647.0 discloses an aluminum alloy engine cylinder casting process design optimization method based on BP neural network and shoal algorithm, which is characterized by comprising process design and numerical simulation analysis model establishment, process design improvement, determination of optimization variables and test design, BP neural network model establishment, shoal algorithm process parameter optimization and production inspection.

The invention patent CN115470591A discloses an integrated aluminum alloy precision casting engine hood structure optimization method based on a multi-island genetic algorithm, which comprises the following steps: extracting sample points from the optimal Latin hypercube, and constructing an RBF neural network model; carrying out deterministic optimization on the constrained first-order mode and quality by using a multi-island genetic algorithm; the results of deterministic optimization were analyzed and optimized based on 6Sigma reliability. The invention only aims at the multi-objective optimization design of the integral aluminum alloy precision casting engine hood morphology and size, and effectively improves the rigidity performance of the engine hood in the constrained first-order mode, the dent resistance working condition, the local compression working condition, the forward bending working condition, the lateral bending working condition and the torsion working condition.

The invention patent CN115034142A discloses a design method of temperature parameters of an extrusion casting process, which is characterized in that the design method of temperature parameters is firstly based on a KNN algorithm to refer to temperature process parameters of similar material castings to perform initial design on pouring temperature and mould preheating temperature from a material composition layer. And then, the mould preheating temperature with larger design error is corrected by combining the casting mould parameters and a Generalized Regression Neural Network (GRNN), so that an error correction model is established. The design method is based on the existing research data, fully refers to and utilizes the existing related research results, does not need to carry out complicated experimental research, and has higher design precision of technological parameters.

The above method is not suitable for low pressure casting of aluminum alloy hubs.

In the low-pressure casting process of the aluminum alloy hub, filling solidification of aluminum liquid is performed in a cavity of a closed mold, and visual characterization cannot be performed. The determination of the low pressure casting process parameters often depends on the experience of workers, and the unstable casting quality of the hub and the generation of waste products are easily caused. In the casting process, if the shape of the casting or the casting process is unreasonable, casting defects such as air holes, insufficient casting, shrinkage porosity, shrinkage cavity, flash and the like can be generated, so that the yield of the hub is reduced, and quality improvement and efficiency improvement are difficult.

At present, a BP neural network is mostly adopted as a casting process parameter prediction algorithm, and although the process window of each parameter in the casting process can be optimized to a certain extent, the accuracy of the neural network algorithm is to be improved, the learning speed is slow, and even a simple problem is solved, hundreds of times or even thousands of times of learning are generally required to be converged; and the local minimum value is easy to fall into, and the selection of the network layer number and the neuron number has no corresponding theoretical guidance.

Disclosure of Invention

The invention aims to provide an aluminum alloy hub low-pressure casting process design optimization method based on a genetic algorithm and a BP neural network, aiming at the problems of low design accuracy of an aluminum alloy hub basic process and large quality fluctuation of a finished product. The method realizes accurate prediction of the casting process parameters of the aluminum alloy wheel, reduces the overall defect rate and meets the appearance requirement and the mechanical property index of the product.

A low-pressure casting aluminum alloy hub process optimization method combining a genetic algorithm and a neural network comprises the following steps:

step one: establishing a casting process parameter database;

establishing a casting process parameter database for casting the aluminum alloy hub; and comprehensively analyzing the temperature curve change and defects in the filling solidification process, and determining the initial process parameter design.

Step two: screening and optimizing variable parameters and preprocessing data;

and selecting part of process parameters to be optimized from the casting pseudo process parameters and the temperature curve in the database as optimization variables, and processing the selected data through data as the input of the neural network. And outputting the characteristic parameters of the surface morphology of the reaction casting.

Step three: establishing a BP neural network model;

and (3) designing a proper BP neural network framework according to the data analysis obtained from the process database.

Step four: establishing a GA-BP neural network prediction model;

and establishing a GA-BP neural network regression model, optimizing the weight and the threshold of the BP neural network by using a genetic algorithm, and continuously updating the structural parameters of the BP neural network.

Step five: training the optimized GA-BP neural network by using data samples in the database, and comparing and verifying the accuracy of the model by using test set data in the database and the neural network prediction casting process parameters after training is completed.

Step six: and manufacturing and verifying the actual performance of the optimized wheel-shaped casting parameter process window, and carrying out mass production if the actual performance meets the requirements.

Term interpretation:

genetic Algorithm (GA) is an optimization algorithm initiated by biological evolution theory, and the optimal solution of the problem is found by optimizing parameter values through the processes of gene crossing, mutation, natural selection and the like. The BP neural network is a feedforward artificial neural network, and regression or classification prediction can be performed through training. The GA-BP algorithm combines a genetic algorithm and a BP neural network, takes the weight and the threshold value of the BP neural network as parameters of the GA algorithm, optimizes the BP neural network through the genetic algorithm, and improves the performance of the BP neural network.

Advantageous effects

And combining the GA algorithm with the BP neural network to obtain a more excellent prediction model GA-BP regression prediction model. The weight and the threshold of the BP neural network are optimized through a genetic algorithm, so that the prediction accuracy of the BP neural network can be improved, and the problems of excessive parameters, overfitting and the like can be avoided. Under the condition of multiple input and single output, the prediction precision can be further improved by using the GA-BP regression prediction model. Compared with the common BP neural network, the design method can optimize the weight and the threshold of the BP neural network through a genetic algorithm, so that the prediction accuracy of the BP neural network can be improved, and the problems of excessive parameters, overfitting and the like can be avoided. Under the condition of multiple input and single output, the prediction precision can be improved by using the GA-BP regression prediction model. Further improves the defect of finished products, further reduces the rejection rate in the trial production process, improves the forming quality of the aluminum alloy hub, shortens the working period from process design to trial production qualification of new products, reduces the workload and the energy consumption cost of workers, and improves the overall production efficiency of the production line.

Drawings

FIG. 1 is a flow chart of a technical route of the present invention;

FIG. 2 is a flow chart of a principle of implementing genetic algorithm optimization neural network

FIG. 3 is a schematic diagram of a neural network according to the present invention;

FIG. 4 is an X-ray radiographic inspection of a hub produced in accordance with the teachings of the present invention.

Detailed Description

The invention will be further illustrated by the following examples, which are not intended to limit the scope of the invention, in order to facilitate the understanding of those skilled in the art.

The specific implementation method of the first step is as follows:

(1) A large number of hub casting historical process parameters are obtained through a factory casting workshop, and main parameters include aluminum liquid casting temperature, mold filling time, mold filling pressure, water cooling time, air cooling time, mold filling speed, pressure maintaining pressure, pressurizing time, pressure maintaining time and pressure relief time, and the appearance positions of corresponding internal pores, shrinkage cavities and shrinkage porosity of the casting are extracted.

(2) And quantifying the defects obtained by simulation, characterizing the defects by using numerical values, and dividing the data into two groups, wherein one group is used as a training set of the neural network, and the other group is used as a detection set.

The specific implementation method of the second step is as follows:

(1) Because the units of input data are different, the range of some data may be particularly large, the effect of the input with large data range in pattern classification may be larger, and the effect of the input with small data range may be smaller, so that the neural network converges slowly and the training time is long. Gradient descent and other algorithms are more beneficial to solving under data normalization. Therefore, we will generally normalize the data before training the BP neural network with gradient descent or other algorithms. The training phase therefore requires that all data be normalized before training. Wherein, the normalization formula is as follows:

where X is the raw data, xn is the normalized data, xmin is the minimum value of the dataset, and Xmax is the maximum value of the dataset.

(2) Because the range of the activation function of the neural network output layer is limited, it is necessary to map the target data of the network training to the range of the activation function. For example, if the output layer of the neural network adopts an S-shaped activation function, the value range of the S-shaped function is limited to (0, 1), that is, the output of the neural network is limited to (0, 1), so that the output of the training data is normalized to the [0,1] interval. The sigmoid activation function is very gentle in the region outside the (0, 1) interval, and the degree of discrimination is too small.

The specific implementation method of the third step is as follows:

BP neural networks are typically composed of an input layer, a hidden layer, and an output layer. The input layer passes the input data to the hidden layer, which calculates the output and passes it to the output layer. The output layer gives the prediction of the network according to the input and the output of the hidden layer, and each layer is connected with weight.

According to the data analysis obtained from the process database, a proper BP neural network frame is designed, the neural network is of a three-layer structure, and the node numbers of an input layer, an hidden layer and an output layer of the neural network are set; the first column of neurons is referred to as input layer i, the middle neurons are referred to as middle layer j, and the last neuron is referred to as output layer k. Each neuron is provided with a bias b and an activation function f, except for the input layer. The links between each neuron are called weights w,

computing inputs and outputs of hidden layers: ,

input and output of the calculation output layer: ,

the loss function expression equation in the forward propagation calculation loss is: ,

the specific implementation method of the fourth step is as follows:

the GA-BP neural network implementation flow is as follows in FIG. 2:

initializing a population: an initial population is generated using binary encoding. The method comprises the steps of linking weights of an input layer and a hidden layer, linking weights of the hidden layer and an output layer, and linking all weights and threshold codes into an individual code by using M-bit binary codes for each weight and threshold. Within the range of the connection weights and the threshold values, a population M of such chromosomes is generated, constituting the initial population.

Fitness function: and taking the norm of an error matrix of the predicted value and the expected value of the predicted sample as the output of an objective function, and taking the fitness function as the function of the error. Individuals with a relatively high fitness can be more likely to be inherited into the next generation. The mean square error is used as a standard of judgment accuracy in the BP neural network. Therefore, when the fitness function takes the minimum value, the weight and the threshold value of the BP network are optimized.

Selecting an operator: the selection operator uses random traversal samples.

Crossover operator: a simple single point crossover operator is used.

Mutation operator: a random method is used to select mutated genes, the mutation generates mutation base factors with a certain probability, and if the code of the selected genes is 1, the genes are converted into 0, and otherwise, the genes are converted into 1.

And then calculating the fitness value of each part in the current population, and repeatedly iterating the population individuals with the optimal fitness value until the set initial condition is met. If the initial condition cannot be met, the operation is carried out until the set maximum iteration times are reached, and the initial weight and the threshold value with smaller errors and more complete and accurate operation are obtained through the genetic algorithm operation.

After the neural network training is completed, the model established in the steps is adopted to input initial parameters to be optimized to execute optimizing calculation, and a parameter data optimizing window reflecting the current optimal production process is obtained according to the output result. And (3) carrying out actual production trial production on the obtained technological parameters, evaluating a prediction result according to the appearance and the performance of the actual produced hub finished product, analyzing the prediction result, and generating a technological recommendation window if the quality is qualified.

Embodiment one:

step one: and extracting main parameters including aluminum liquid casting temperature, mold temperature, filling time, filling pressure, water cooling time, air cooling time, filling speed, pressure maintaining pressure, pressurizing time, pressure maintaining time and pressure relief time and the appearance positions of the castings corresponding to the internal pores, shrinkage cavities and shrinkage porosity according to the casting historical data of the aluminum alloy hub of a certain brand A356 on the existing casting production line. The chemical composition of the a356 aluminum alloy is shown in table 1 below.

Table 1a356 aluminum alloy chemistry

Element(s)

Al

Si

Fe

Mg

Cu

Ti

Mn

Zn

Content of

Allowance of

6.74

0.08

0.35

0.04

0.09

0.02

0.05

Step two: and (3) establishing a data set of the aluminum alloy casting process parameters corresponding to the molding quality according to the historical process data of the model hub extracted in the step one.

Step three: and determining the node number of an input layer, an hidden layer and an output layer of the BP neural network according to the input and the output, and giving the operation rate and the neuron excitation function. And initializing the connection weight and threshold of the neurons among the three layers of structures.

Step four: and (3) optimizing the BP neural network model established in the step (III) by utilizing a genetic algorithm, establishing a machine learning model of technological parameters required by the die casting process of the aluminum alloy hub, and verifying the accuracy of the machine learning model. The surface roughness and air hole defects during model verification are required, and the indexes are within the allowable range of the enterprise production standard.

Step five: according to the basic process requirements of the casting of this example, the required process parameters are input into a machine learning model, as shown in the following table 2, the casting temperature is 680 ℃, the mold temperature is 400 ℃, the filling time is 12s, the water cooling time is 110s, the air cooling time is 100s, the pressure maintaining time is 80s, and the pressure maintaining pressure is 900MPa.

Step six: if the casting temperature is 690 ℃, the die temperature is 415 ℃, the filling time is 14s, the water cooling time is 120s, the air cooling time is 105s, the pressure maintaining time is 90s and the pressure maintaining pressure is 950MPa after the parameters are optimized.

The performance indexes after the actual production trial production by the process are shown in the following table 3, and the tensile strength of the rim part of the finished hub reaches 237MPa, the yield strength reaches 179MPa, and the porosity is 0.81%.

By comparing the tensile properties of castings manufactured by the original non-optimized process method, the tensile strength and the yield strength are improved, and the air hole defect ratio is reduced. As a result of X-ray flaw detection (FIG. 4), it was found that the aluminum alloy pressing hub manufactured by the method was excellent in internal quality.

Table 2 input process parameters in machine learning model

TABLE 3 hub Performance index after trial manufacture

The above embodiments are not exhaustive of the possible implementations of the invention. Any obvious modifications thereof, which would be apparent to those skilled in the art without departing from the principles and spirit of the present invention, should be considered to be included within the scope of the appended claims.

It will be understood by those skilled in the art that all terms (including technical and scientific terms) used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention belongs unless defined otherwise. It will be further understood that terms, such as those defined in commonly used dictionaries, should be interpreted as having a meaning that is consistent with their meaning in the context of the prior art and will not be interpreted in an idealized or overly formal sense unless expressly so defined herein.

It should be understood that the detailed description of the technical solution of the present invention, given by way of preferred embodiments, is illustrative and not restrictive. Modifications of the technical solutions described in the embodiments or equivalent substitutions of some technical features thereof may be performed by those skilled in the art on the basis of the present description; such modifications and substitutions do not depart from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. The low-pressure casting aluminum alloy hub process optimization method combining a genetic algorithm and a neural network is characterized by comprising the following steps of:

step one: establishing a casting process parameter database;

establishing a casting process parameter database for casting the aluminum alloy hub; analyzing the temperature curve change and defects in the filling solidification process, and determining an initial process and a parameter database;

step two: screening and optimizing variable parameters and preprocessing data;

selecting partial process parameters to be optimized from casting process parameters in a database, including aluminum liquid casting temperature, mold filling time, mold filling pressure, water cooling time, air cooling time, mold filling speed, pressure maintaining pressure, pressurizing time, pressure maintaining time and pressure relief time, as optimization variables, and performing data processing on the selected data to convert the data into input conforming to the type of the neural network; outputting characteristic parameters of the surface morphology of the reaction casting;

step three: establishing a BP neural network model;

determining a BP neural network framework according to data obtained from a process database;

step four: establishing a GA-BP neural network prediction model;

establishing a GA-BP neural network regression model, optimizing the weight and the threshold of the BP neural network by utilizing a genetic algorithm, and updating the structural parameters of the BP neural network;

step five: training the optimized GA-BP neural network by using a data sample in a database, and comparing and verifying the accuracy of a model by using test set data in the database and the neural network prediction casting process parameters after training is completed;

step six: and manufacturing and verifying the optimized casting parameter process, and carrying out mass production if the manufacturing and verifying requirements are met.

2. The method of claim 1, wherein the process parameters in step one mainly comprise the casting temperature of the molten aluminum, the mold temperature, the mold filling time, the mold filling pressure, the water cooling time, the air cooling time, the mold filling speed, the pressure maintaining pressure, the pressurizing time, the pressure maintaining time and the pressure relief time, and the appearance positions of the casting corresponding to the internal pores, shrinkage cavities and shrinkage porosity.

3. The method of claim 1, wherein in step one, the defects obtained by simulation are quantified and characterized by numerical values, and the data are divided into two groups, one group being a training set of the neural network and the other group being a detection set.

4. The method of claim 1, wherein the data is normalized prior to training the BP neural network in step two; the normalization formula is as follows:

X _n ＝(X-X _min )/(X _max -X _min )

5. The method of claim 1, wherein in step two, the target data of the network training is mapped to a value range of the activation function.

6. The method of claim 1, wherein the neural network in the third step is a three-layer structure, and the node numbers of the input layer, the hidden layer and the output layer of the neural network are set; the first column of neurons is called input layer i, the middle neuron is called middle layer j, and the last neuron is called output layer k; each neuron is self-contained with a bias b, and is self-contained with an activation function f, except for the input layer; the connection between each neuron is called a weight w.

7. The method according to claim 1, wherein the GA-BP neural network implementing step in step four includes:

initializing a population: generating an initial population by adopting binary codes; the method comprises the steps of linking weights of an input layer and a hidden layer, linking weights of the hidden layer, weights of the hidden layer and an output layer, and linking all weights and threshold codes into an individual code by using M-bit binary codes for each weight and threshold; within the range of the junction weights and the threshold values, a population M of such chromosomes is generated, constituting the initial population.

8. The method according to claim 1, wherein the GA-BP neural network implementing step in step four includes:

fitness function: taking the norm of an error matrix of a predicted value and an expected value of a predicted sample as the output of an objective function, and taking an adaptive function as an error function; individuals with relatively high fitness can be more likely to be inherited into the next generation; taking the mean square error as a standard of judgment accuracy in the BP neural network;

and/or

Selecting an operator: the selection operator uses random traversal sampling;

and/or

Crossover operator: using a simple single-point crossover operator;

and/or

Mutation operator: selecting mutated genes by using a random method, wherein mutation generates mutation base factors with a certain probability, and if the code of the selected genes is 1, the genes are converted into 0, and otherwise, the genes are converted into 1;

calculating the fitness value of each part in the current population, and repeatedly iterating the population individuals with the optimal fitness value until the set initial condition is met; and/or if the initial condition cannot be met, the method runs until the set maximum iteration number is reached, and the accurate initial weight and the threshold value are obtained through the genetic algorithm operation.

9. The method of claim 1 wherein the data of step one is process data in existing past actual production and simulation software.