Heat treatment method and device based on GA-ANN artificial neural network model

Abstract

The invention relates to a heat treatment method and device based on a GA-ANN artificial neural network model. The method comprises the following steps: constructing a GA-ANN artificial neural network model in which a neural network algorithm and a genetic algorithm are combined; collecting instance data of steel heat treatment in previous production, and establishing a material database; setting learning data of an input layer and an output layer of the GA-ANN model according to material information and performance requirements of the current steel to be subjected to heat treatment; learning and verifying the GA-ANN model by using data in the material database; and pre-judging heat treatment parameters required by the material to be subjected to heat treatment by using the qualified GA-ANN artificial neural network model. The device comprises a construction module, a database module, a basic parameter setting module, a training module and a pre-judgment module. According to the method and the device, the heat treatment parameters which should be adopted by the target performance indexes can be effectively predicted.

Description

[0001] This application is a divisional application of patent application No. 201810810834.7, filed on July 23, 2018, entitled "Heat Treatment Method and Apparatus Based on GA-ANN Artificial Neural Network Model". Technical Field

[0002] This invention relates to the field of steel processing technology, specifically to a heat treatment method and apparatus based on a GA-ANN artificial neural network model. Background Technology

[0003] Artificial Neural Networks (ANNs) are neural networks that learn from life, constructed using electronic, optical, or other biophysical and chemical methods to mimic biological neural networks. The goal of artificial neural networks learning from biological neural networks is not to reach or surpass their learning objects in all aspects of performance, but rather to learn and implement the intelligence that humans need based on an understanding and analysis of the structure, mechanism, and function of biological neural networks.

[0004] Genetic Algorithms (GA) are stochastic optimization methods that borrow from the laws of biological evolution and heredity, achieving survival of the fittest through reproduction, inheritance, mutation, and competition. Due to their openness and scalability, GA easily integrates with and combines with other intelligent computing methods such as Expert Systems, Fuzzy Logic, Neural Networks, Simulated Annealing, and Chaos Theory. GA, along with Simulated Annealing, Neural Networks, and Elastic Nets, are considered physical computing (computation based on natural laws) because they all simulate the results of certain natural laws.

[0005] Currently, scholars from many disciplines and professions in China have conducted research and application of artificial neural networks or genetic algorithms, achieving certain technological results. These physical calculation methods have also been extensively studied in the steel industry, such as predicting manganese and phosphorus levels at the endpoint of converter smelting. However, much of this research focuses on how to select functions or methods to reduce the number of convergence cycles and how to avoid local minima. Few studies have developed complete and practical tools after the algorithm is constructed and a full model is established.

[0006] To ensure that special-performance steels meet usage requirements, steel companies often subject them to heat treatment methods such as quenching and tempering or solution treatment after production. However, in recent years, as user demands have become increasingly stringent and the performance requirements after heat treatment have become more precise, the adjustable range for the strength and hardness of many heat-treated materials has become increasingly narrow, to the point that the heat treatment temperature setting range is only 10-30℃. Currently, there is no clear formula or theoretical model for the relationship between the heat treatment parameters of general materials and the mechanical properties they can achieve after treatment. In other words, there is a lack of technology that can effectively predict the relevant heat treatment parameters required to achieve the target mechanical properties. In the current production process, achieving the preset target mechanical properties often requires multiple experiments to actually determine the mechanical properties of the material to be heat-treated under various different heat treatment parameter conditions, resulting in a long production cycle and high costs.

[0007] Therefore, how to effectively predict the heat treatment parameters required for materials to achieve their target mechanical properties has become a technical problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0008] This invention proposes a heat treatment method and apparatus based on a GA-ANN artificial neural network model to solve the aforementioned technical problems in the prior art.

[0009] The technical solution of this invention is as follows:

[0010] This invention provides a heat treatment method based on a GA-ANN artificial neural network model, comprising the following steps:

[0011] A GA-ANN artificial neural network model combining neural network algorithm and genetic algorithm is constructed; case data of steel heat treatment in previous production are collected to establish a material database; based on the current information and performance requirements of the steel to be heat treated, the learning data of the input and output layers of the GA-ANN artificial neural network model are set; the GA-ANN artificial neural network model is learned and verified using the data in the material database; the verified GA-ANN artificial neural network model is used to predict the heat treatment parameters required for the material to be heat treated.

[0012] The construction of the GA-ANN artificial neural network model, which combines neural network algorithms and genetic algorithms, includes the following steps: establishing a multi-layer neural network topology using the BP algorithm; and using a combination of neural network algorithms and genetic algorithms, determining network connection weights based on the established neural network topology using the genetic algorithm to construct the GA-ANN artificial neural network model.

[0013] The construction of the GA-ANN artificial neural network model, which combines neural network algorithms and genetic algorithms, also includes the step of setting an activation function with nonlinearity, differentiability, and monotonicity to introduce nonlinear characteristics into the GA-ANN artificial neural network model.

[0014] This includes collecting data on past steel heat treatment examples, including:

[0015] According to the steel grade classification, the heat treatment information of the steel grades that have been produced is collected and recorded. The heat treatment information includes, but is not limited to, size specifications, main component content, forging ratio of the finished product, heat treatment parameters, performance indicators, and the model of the heat treatment furnace used. The heat treatment parameters include, but are not limited to, quenching and tempering parameters and solution treatment parameters.

[0016] When the steel to be heat-treated is martensitic stainless steel 2Cr13, the heat treatment information includes: dimensions, forging ratio, ingot type, furnace number, batch number, quenching and tempering temperature, composition information, and strength and hardness data to be tested.

[0017] The GA-ANN artificial neural network model is learned and validated using data from the materials database, including setting the evaluation criterion as mean square error RMSE < 0.001, whereby the mean square error is used to complete the construction of the GA-ANN artificial neural network model.

[0018]

[0019] Among them, T i For the actual test results, Y i The output value of the GA-ANN artificial neural network model is given by N, where N is the number of training sample groups. The absolute error ARE is used to describe the generalization ability of the GA-ANN artificial neural network model.

[0020]

[0021] This invention also provides a heat treatment device based on a GA-ANN artificial neural network model, comprising: a construction module for constructing a GA-ANN artificial neural network model combining neural network algorithms and genetic algorithms; a database module for collecting example data of steel heat treatment in previous production processes and establishing a material database; a basic parameter setting module for setting the learning data of the input and output layers of the GA-ANN artificial neural network model according to the information and performance requirements of the steel material to be heat treated; a training module for learning and validating the GA-ANN artificial neural network model using data from the material database; and a prediction module for predicting the required heat treatment parameters for the material to be heat treated based on the validated GA-ANN artificial neural network model.

[0022] The building module is used for:

[0023] A multi-layered neural network topology is established using the backpropagation (BP) algorithm. Then, a combination of neural network algorithm and genetic algorithm is used to determine the network connection weights based on the established neural network topology, thereby constructing a GA-ANN artificial neural network model.

[0024] The database module is used for:

[0025] According to the steel grade classification, the heat treatment information of the steel grades that have been produced is collected and recorded. The heat treatment information includes, but is not limited to, size specifications, main component content, forging ratio of the finished product, heat treatment parameters, performance indicators, and the model of the heat treatment furnace used. The heat treatment parameters include, but are not limited to, quenching and tempering parameters and solution treatment parameters.

[0026] The building module is also used for:

[0027] An activation function with nonlinearity, differentiability, and monotonicity is set to introduce nonlinear characteristics into the GA-ANN artificial neural network model.

[0028] The technical effects disclosed in this invention are as follows:

[0029] This invention proposes a heat treatment method and apparatus based on a GA-ANN artificial neural network model. It combines neural network algorithms and genetic algorithms to construct a GA-ANN artificial neural network model. Historical steel heat treatment data is used as training data for model learning and verification. The verified model predicts the required heat treatment parameters for steel. Experimental testing shows that this neural network can effectively predict the material properties achieved under different heat treatment parameters with high accuracy. It can ensure that the material meets performance requirements after a single heat treatment based on set parameters. Compared with existing technologies, this shortens the production cycle and reduces production costs, making it a relatively effective prediction technology solution with practical applications. For example, in existing technologies, when large-scale material tempering is carried out in mass production, the furnace number, specifications, and card number involved are different. Selecting the same temperature for heat treatment may result in some materials failing to meet the requirements and needing to be re-tempered. Using the neural network prediction method provided by this invention, the optimal temperature can be selected, and materials that cannot be grouped can be isolated and heat-treated separately under different temperature conditions, ensuring that all materials meet the requirements in one tempering, thereby saving time and reducing costs. Attached Figure Description

[0030] The accompanying drawings, which form part of this application, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an undue limitation of the invention. Wherein:

[0031] Figure 1 This is a schematic diagram of the topology of a multilayer neural network model according to an embodiment of the present invention;

[0032] Figure 2 This is a schematic diagram illustrating the basic process of neural network learning in an embodiment of the present invention;

[0033] Figure 3 This is a schematic diagram of the basic process of the genetic algorithm in an embodiment of the present invention;

[0034] Figure 4 This is a schematic diagram illustrating the basic process of combining artificial neural networks and genetic algorithms in an embodiment of the present invention. Detailed Implementation

[0035] The present invention will now be described in detail with reference to the accompanying drawings and embodiments. Various examples are provided by way of explanation and not by way of limitation. Indeed, those skilled in the art will recognize that modifications and variations can be made to the invention without departing from its scope or spirit. For example, a feature shown or described as part of one embodiment may be used in another embodiment to produce yet another embodiment. Therefore, it is desirable that the invention encompass such modifications and variations falling within the scope of the appended claims and their equivalents.

[0036] This invention discloses a heat treatment method and apparatus based on a GA-ANN artificial neural network model.

[0037] The heat treatment method based on the GA-ANN artificial neural network model of this invention includes the following steps:

[0038] Construct a GA-ANN artificial neural network model that combines neural network algorithms and genetic algorithms;

[0039] Collect data on past steel heat treatment examples in production and establish a materials database;

[0040] Based on the current information and performance requirements of the steel to be heat-treated, the learning data for the input and output layers of the GA-ANN artificial neural network model are set.

[0041] The GA-ANN artificial neural network model was learned and validated using data from a materials database;

[0042] The validated GA-ANN artificial neural network model will be used to predict the heat treatment parameters required for the material to be heat-treated.

[0043] The heat treatment mentioned in this invention refers to methods such as tempering and solution treatment.

[0044] In this embodiment of the invention, the BP algorithm (Error Back Propagation) is used to build a multi-layer neural network. The topology of the multi-layer neural network model is as follows: Figure 1 As shown, it consists of an input layer, intermediate layers, and an output layer. The intermediate layers, also known as hidden layers, can be one or more. For the basic learning process of a neural network, see [link to documentation]. Figure 2 As shown.

[0045] Genetic algorithms simulate the natural evolutionary process in the biological world, using selection, crossover, and mutation operations to optimize and solve complex problems. A basic genetic algorithm includes chromosome encoding, individual fitness, genetic operators, and runtime parameters.

[0046] In this embodiment of the invention, the genetic operators are defined as three types: selection, crossover, and mutation, and are defined as an 8-tuple.

[0047] S GA =<C,E,P0,M,Φ,Γ,Ψ,T>

[0048] In the formula: C is the individual encoding method; E is the individual fitness evaluation function; P0 is the initial population; M is the population size; Φ is the selection operator; Γ is the crossover operator; Ψ is the mutation operator; T is the termination condition; SGA stands for Simple Genetic Algorithm.

[0049] For the basic process of genetic algorithms, please refer to Figure 3 As shown.

[0050] In this embodiment of the invention, the GA-ANN model is established by combining neural networks and genetic algorithms. Based on the aforementioned neural network topology, the genetic algorithm determines the network connection weights. As one possible implementation method, the basic process of combining artificial neural networks and genetic algorithms is described below. Figure 4 As shown.

[0051] In order to achieve effective prediction of target performance, it is also necessary to select an appropriate activation function.

[0052] In this embodiment of the invention, activation functions play a very important role in the neural network model. They introduce nonlinear characteristics into the neural network constructed in this invention. Their main purpose is to transform the input signal of a node in the ANN model into an output signal, and the output signal is stacked to become the input of the next layer.

[0053] A suitable activation function needs to possess three properties:

[0054] Nonlinearity: Combinations of linear functions can only solve a very limited number of problems, and they are helpless when faced with nonlinear problems.

[0055] Differentiability: This property is essential when the optimization method is gradient-based.

[0056] Monotonicity: When the activation function is monotonic, a single-layer network can be guaranteed to be a convex function.

[0057] The activation functions provided by this invention for the GA-ANN artificial neural network model include the sigmoid function, tanh function, ReLU function, and Guassian function.

[0058] Without activation functions, the output signal would simply be a linear function. A linear function is a first-order polynomial. Linear equations are easy to solve, have limited complexity, and are less capable of learning complex function mappings from data. A neural network without activation functions would simply be a linear regression model, which has limited power and performs poorly in most cases.

[0059] Before, after, or simultaneously with the construction of the GA-ANN artificial neural network model, it is necessary to organize and establish a material database of heat-treated steel. Based on the material characteristics of the steel to be heat-treated and the corresponding target properties, the learning data of the input layer and output layer are constructed.

[0060] The settings of training sample data and the learning data for the input and output layers are also key factors affecting the effectiveness of predictions. Only by identifying the variables affecting the results can the most accurate weighted network be derived. This is also the key reason why some existing technologies can only remain at the theoretical stage and be published in papers, unable to be promoted for practical application. For example, some existing technologies use neural networks to explore the optimal tempering temperature, but the parameters used are only quenching, tempering, tensile strength, and yield strength, without considering variables such as the composition, specifications, and forging ratio of the tempered material each time. Therefore, the calculated results actually have low prediction accuracy and lack practicality.

[0061] In this embodiment of the invention, the GA-ANN artificial neural network model requires a large number of instances for learning. Without instances, the neural network cannot be constructed, let alone made predictions. Therefore, the steel grades used for neural network learning must have a large amount of basic data. Based on steel grade classification, records are collected on the specifications of already produced steel grades, the content of main components (including trace elements affecting the heat treatment process), the forging ratio of the finished product, heat treatment parameters, performance indicators, and the heat treatment furnace used. Depending on the characteristics of the steel grade, the heat treatment methods mainly include quenching and tempering (quenching + tempering) and solution aging. Taking martensitic stainless steel 2Cr13 as an example, the data for the input layer and output layer are selected from the collected records, as shown in Tables 1 and 2.

[0062] Table 1 2Cr13 Input Layer Settings

[0063]

[0064] Table 2 2Cr13 Output Layer Settings

[0065]

[0066] As one possible implementation method, the main process steps of the heat treatment method provided by the present invention may include:

[0067] a) Construct a GA-ANN artificial neural network model;

[0068] b) Organize the material database for heat-treated steel;

[0069] c) Select the input and output layers based on the materials;

[0070] d) Use a portion of the data in the learning process of the neural network;

[0071] e) Use another portion of the data in the network verification process;

[0072] f) Use a validated neural network to predict the materials that require heat treatment;

[0073] g) Group the materials for heat treatment according to the predicted temperature;

[0074] h) Compare and evaluate the actual heat treatment results with the predicted results.

[0075] This invention also provides a heat treatment device based on a GA-ANN artificial neural network model, comprising:

[0076] The building block is used to construct a GA-ANN artificial neural network model that combines neural network algorithms and genetic algorithms.

[0077] The database module is used to collect example data of steel heat treatment in past production and establish a materials database;

[0078] The basic parameter setting module is used to set the learning data of the input layer and output layer of the GA-ANN artificial neural network model according to the information and performance requirements of the steel material to be heat-treated.

[0079] The training module is used to learn and validate the GA-ANN artificial neural network model using data from the material database;

[0080] The prediction module is used to predict the heat treatment parameters required for the material to be heat-treated based on the verified GA-ANN artificial neural network model.

[0081] To facilitate a further understanding of the technical solution of the present invention, a specific embodiment of the present invention is listed below.

[0082] Currently, eight batches of 2Cr13 require heat treatment for quenching and tempering. The performance requirements for the 1 / 2R section after quenching and tempering are as follows:

[0083] Table 3 Requirements for Conditioning and Tempering Performance

[0084] Yield strength Rp0.2 Tensile strength Rm Hardness HBW 552-655Mpa ≥655Mpa ≤237

[0085] The known information for the current 8 batches of 2Cr13 is as follows:

[0086] Table 4 Information on 2Cr13 Quenched and Tempered Materials

[0087]

[0088] 1) Pre-build a GA-ANN artificial neural network model and organize the material database for heat-treated steel:

[0089] 500 sets of recently generated 2Cr13 quenching and tempering data were entered into the database. The information included specifications, forging ratio, ingot type, furnace number, batch number, quenching and tempering temperature, composition information, and strength and hardness data of the test results.

[0090] 2) Select the input layer, output layer, basic parameters for network construction, and evaluation criteria based on the materials.

[0091] Table 5 2Cr13 Input Layer Settings

[0092] Specification C content Mo content Forging ratio Quenching temperature Tempering temperature

[0093] Table 6 2Cr13 Output Layer Settings

[0094] Yield strength Rp0.2 Tensile strength Rm Hardness HBW

[0095] Preferably, the neural network structure is 6-9-3, wherein the input layer is 6; the hidden layer is one layer with a structure of 9; and the output layer is 3.

[0096] Evaluation criteria: Mean squared error RMSE < 0.001.

[0097]

[0098] Among them, T i For the actual test results, Y i This is the output value of the neural network.

[0099] 3) Use 80% of the data for the learning process of the neural network to form a neural network.

[0100] 4) Use 20% of the data for network verification.

[0101] The network's generalization ability is described using the absolute error (ARE). The maximum absolute error is 2.256% for yield strength, 3.174% for tensile strength, and 2.873% for hardness. This indicates that the network can be used for prediction.

[0102]

[0103] 5) Use a validated neural network to predict the materials that require heat treatment.

[0104] Based on the prediction, the quenching temperature can be the same as 960℃, and the tempering temperature of the eight groups of materials can be calculated between 650-750℃.

[0105] Table 7 Tempering temperatures corresponding to the predicted qualified performance data.

[0106] Batch number Able to meet the qualified tempering temperature range 1 675-695 2 665-685 3 670-690 4 670-690 5 665-685 6 690-710 7 665-685 8 665-685

[0107] 6) Perform heat treatment on the materials in groups according to the predicted temperature.

[0108] Batch number 6 was tempered separately at 700℃ within the tempering temperature range. The remaining materials were tempered together at 680℃.

[0109] 7) Compare and evaluate the actual heat treatment results with the predicted results.

[0110] Table 8 Comparison of 2Cr13 test data and predicted data

[0111]

[0112]

[0113] In summary, the heat treatment method and apparatus based on the GA-ANN artificial neural network model provided by this invention can effectively solve the technical problem in the prior art of not being able to effectively predict which heat treatment parameters can achieve the target performance. According to the above technical solution of this invention, and based on the test results:

[0114] This neural network model can effectively predict the material properties at different heat treatment temperatures, group the materials for tempering based on the predicted temperatures, improve the first-time tempering pass rate of steel, and provide assistance in developing optimized heat treatment processes.

[0115] The above are merely preferred embodiments of the present invention and are not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A heat treatment method based on a GA-ANN artificial neural network model, characterized in that, Including the following steps: Construct a GA-ANN artificial neural network model that combines neural network algorithms and genetic algorithms; Collect data on past steel heat treatment examples in production and establish a materials database; Based on the current information and performance requirements of the steel material to be heat-treated, the learning data for the input and output layers of the GA-ANN artificial neural network model are set. The GA-ANN artificial neural network model is learned and validated using data from the material database. The validated GA-ANN artificial neural network model will be used to predict the heat treatment parameters required for the material to be heat-treated.

2. The heat treatment method based on the GA-ANN artificial neural network model according to claim 1, characterized in that, The steps described above for constructing a GA-ANN artificial neural network model that combines neural network algorithms and genetic algorithms include the following steps: A multi-layered neural network topology is established using the backpropagation (BP) algorithm. By combining neural network algorithms and genetic algorithms, and based on the established neural network topology, the genetic algorithm is used to determine the network connection weights to construct a GA-ANN artificial neural network model.

3. The heat treatment method based on the GA-ANN artificial neural network model according to claim 2, characterized in that, The step of constructing a GA-ANN artificial neural network model that combines neural network algorithms and genetic algorithms also includes the following steps: An activation function with nonlinearity, differentiability, and monotonicity is set to introduce nonlinear characteristics into the GA-ANN artificial neural network model.

4. The heat treatment method based on the GA-ANN artificial neural network model according to claim 1, characterized in that, The steps described above involve collecting data from past steel heat treatment examples, including: According to the steel grade classification, the heat treatment information of the steel grades that have been produced is collected and recorded. The heat treatment information includes, but is not limited to, size specifications, main component content, forging ratio of the finished product, heat treatment parameters, performance indicators, and the model of the heat treatment furnace used. The heat treatment parameters include, but are not limited to, quenching and tempering parameters and solution treatment parameters.

5. The heat treatment method based on the GA-ANN artificial neural network model according to claim 4, characterized in that, When the steel to be heat-treated is martensitic stainless steel 2Cr13, the heat treatment information includes: dimensions, forging ratio, ingot type, furnace number, batch number, quenching and tempering temperature, composition information, and strength and hardness data of the test.

6. The heat treatment method based on the GA-ANN artificial neural network model according to claim 5, characterized in that, The steps described above utilize data from the materials database to learn and validate the GA-ANN artificial neural network model, including: The evaluation criterion is set as mean squared error (RMSE) < 0.001, where the mean squared error is used to complete the construction of the GA-ANN artificial neural network model. Among them, T i For the actual test results, Y i This represents the output value of the GA-ANN artificial neural network model, where N is the number of training sample groups. The generalization ability of the GA-ANN artificial neural network model is described using the absolute error ARE.

7. A heat treatment device based on a GA-ANN artificial neural network model, characterized in that, include: The building block is used to construct a GA-ANN artificial neural network model that combines neural network algorithms and genetic algorithms. The database module is used to collect example data of steel heat treatment in past production and establish a materials database; The basic parameter setting module is used to set the learning data of the input layer and output layer of the GA-ANN artificial neural network model according to the information and performance requirements of the steel material to be heat-treated. The training module is used to learn and validate the GA-ANN artificial neural network model using data from the material database; The prediction module is used to predict the heat treatment parameters required for the material to be heat-treated based on the verified GA-ANN artificial neural network model.

8. The heat treatment apparatus based on the GA-ANN artificial neural network model according to claim 7, characterized in that, The building module is used for: A multi-layered neural network topology is established using the backpropagation (BP) algorithm. By combining neural network algorithms and genetic algorithms, and based on the established neural network topology, the genetic algorithm is used to determine the network connection weights to construct a GA-ANN artificial neural network model.

9. The heat treatment apparatus based on the GA-ANN artificial neural network model according to claim 7, characterized in that, The database module is used for: According to the steel grade classification, the heat treatment information of the steel grades that have been produced is collected and recorded. The heat treatment information includes, but is not limited to, size specifications, main component content, forging ratio of the finished product, heat treatment parameters, performance indicators, and the model of the heat treatment furnace used. The heat treatment parameters include, but are not limited to, quenching and tempering parameters and solution treatment parameters.

10. The heat treatment apparatus based on the GA-ANN artificial neural network model according to any one of claims 7-9, characterized in that, The building module is also used for: An activation function with nonlinearity, differentiability, and monotonicity is set to introduce nonlinear characteristics into the GA-ANN artificial neural network model.

Images