Method for steel product quality prediction based on evolutionary strategy-based neural architecture search

Abstract

The application provides a steel product quality prediction method based on evolutionary strategy-based neural architecture search, and the method comprises the following steps: acquiring historical data and related parameters affecting steel quality under a target production line or scene to construct a feature set required for model training and verification; based on the feature set, a search space containing network structure and hyperparameters is constructed, and a variable-length coding strategy is adopted to encode network layer number, neuron number, activation function, optimizer type and training parameters to generate an initial candidate model population; an improved ES algorithm is adopted to automatically search the structure and hyperparameters of the neural network, and the performance of each generation of candidate models is evaluated by using back propagation training; and the trained optimal model is used for quality prediction of steel production data. The application improves the accuracy of steel product quality prediction.

Description

Technical Field

[0001] This invention relates to the field of steel product quality inspection technology, and in particular to a method for predicting steel product quality based on neural architecture search using evolutionary strategies. Background Technology

[0002] With the rapid development of artificial intelligence technology, intelligent prediction and decision-making methods provide crucial support for the optimization and upgrading of key production processes in the steel industry. For a long time, steel product quality assurance has relied heavily on post-production control, which often struggles to identify substandard products in a timely manner, leading to resource waste and economic losses. Artificial neural networks, through systematic analysis of multiple parameters in the production process, can achieve accurate quality predictions, enabling companies to adjust processes and optimize solutions at an early stage. This effectively reduces the impact of adverse factors and improves the control level of the production process.

[0003] However, with increasingly fierce competition in the steel market, customer demands for steel composition and performance are becoming more personalized, diversified, and produced in smaller batches. This change presents new challenges to the application of artificial neural networks in the steel industry. Since neural network structures are typically designed for specific problems, changes in data characteristics or production conditions often necessitate adjustments or even complete reconstruction of the original network. For steel companies with diverse product lines and frequently changing customer needs, fixed predictive models struggle to maintain high accuracy. Relying on frequent manual adjustments or redesigns of models would consume significant manpower and time. Furthermore, manual model design often involves repeated trial and error, and the inherent "black box" nature of neural networks limits the effectiveness of traditional structural adjustments and hyperparameter optimization experience when dealing with complex or weakly correlated data.

[0004] Neural network architecture search, as a technique that automatically finds the optimal network structure through intelligent algorithms, provides a feasible solution to the aforementioned problems. This method can automatically explore and generate highly adaptable and accurate network architectures for steel quality prediction tasks. This not only improves the accuracy and efficiency of the model but also significantly shortens the model development cycle, thereby better meeting the urgent needs of the steel industry for rapid and accurate quality prediction and supporting agile product development and production adjustments.

[0005] Currently, common neural network architecture search methods mainly fall into two categories: reinforcement learning and gradient descent. While reinforcement learning can achieve automatic structure search, it typically requires significant computational resources and time, often proving costly even for medium-sized datasets. Gradient descent methods generally require building and training a supernetwork, and its implementation demands a high level of expertise and experience. Therefore, how to efficiently and adaptively complete network structure search under resource-constrained conditions remains a key issue that needs further exploration in the current intelligent prediction practices of the steel industry. Summary of the Invention

[0006] To address the above issues, this invention utilizes variable-length coding to jointly optimize the network structure and hyperparameters, combined with an improved ES algorithm, to create a neural network architecture and hyperparameter tuning search for steel product quality prediction. This improves the accuracy of steel product quality prediction and is suitable for steel production scenarios involving multiple varieties, small batches, and fast pace.

[0007] According to embodiments of the present invention, a method for predicting the quality of steel products based on neural architecture search using an evolutionary strategy is provided.

[0008] In a first aspect of the invention, a method for predicting the quality of steel products based on neural architecture search using an evolutionary strategy is provided. The method includes: Step S01: Obtain historical data and relevant parameters affecting steel quality in the target production line or scenario to construct the feature set required for model training and validation; Step S02: Based on the feature set, construct a search space containing network structure and hyperparameters, and use a variable-length encoding strategy to encode the number of network layers, number of neurons, activation function, optimizer type and training parameters to generate an initial candidate model population; Step S03: The improved ES algorithm is used to automatically search for the structure and hyperparameters of the neural network, and backpropagation is used to train and evaluate the performance of each generation of candidate models; Step S04: Use the trained optimal model to predict the quality of steel production data.

[0009] Furthermore, the historical data and related parameters mentioned in step S01 include sensor data and parameters that affect product quality during the quality measurement and analysis process.

[0010] Furthermore, after obtaining historical data and relevant parameters affecting steel quality in the target production line or scenario as described in step S01, the data is preprocessed as follows: for features with missing values, mean or interpolation methods are used to fill them in according to their characteristics; data with a large number of duplicate values ​​in a specific time period are removed; outliers are detected and deleted if any are found; numerical features are normalized; and categorical data are one-hot encoded.

[0011] Furthermore, the encoding described in step S02 first determines the number of hidden layers in the neural network, and then determines the specific hyperparameters required for each layer.

[0012] Furthermore, the specific steps of step S03 are as follows: Step S031: Initialize the parameters of the improved ES algorithm: initial population Population size Number of elites Maximum number of iterations ; Step S032: Perform iterations, first for the population All individuals are randomly selected to undergo two mutation operations using a mutation operator, resulting in... Each individual is recorded as a descendant population. ; Step S033: For the population Perform the selection operation, denoted as ; Step S034: Population After individuals are selected using the roulette wheel operator, a single-point crossover operator is used to perform a crossover operation, randomly selecting the genotype boundary point on the chromosome as the crossover point, and the offspring population is denoted as […]. ; Step S035: For the population and Conduct joint selection operation; Step S036: For the parent population P Use an elite retention strategy; Step S037: Decompose the elite set Q Merge with the best-performing individuals of the present generation to form a new population; Step S038: End After several iterations, the optimal individual is selected as the final solution.

[0013] Furthermore, the two mutation operations performed by selecting the mutation operator in step S032 specifically refer to: genotype mutation and multi-point mutation; Genotype mutations randomly select genotype segments to be mutated; for the variable-length part of the code, i.e. the hidden layer network structure, the number of layers is first randomly re-determined to determine the new network architecture in the offspring individuals; as the number of network layers increases, the hyperparameters of the added network structure are randomly generated; if the selected genotype is a fixed part of the code, the mutation will be randomly generated in its corresponding coding space.

[0014] Furthermore, the multi-point mutation randomly selects several genes in the chromosome for mutation, and does not include the number of network layers. The number of mutation points is randomly generated within an integer range.

[0015] In a second aspect of the invention, an apparatus for predicting the quality of steel products based on neural architecture search using an evolutionary strategy is provided. The apparatus includes: Feature set construction module: used to acquire historical data and relevant parameters affecting steel quality in the target production line or scenario, in order to build the feature set required for model training and validation; Model building module: Based on the feature set, it constructs a search space containing network structure and hyperparameters, and uses a variable-length encoding strategy to encode the number of network layers, number of neurons, activation function, optimizer type and training parameters to generate an initial candidate model population. Model optimization module: Used to automatically search for the structure and hyperparameters of neural networks using an improved ES algorithm, and to evaluate the performance of each generation of candidate models using backpropagation training; Model Inference Module: Used to predict the quality of steel production data using the trained optimal model.

[0016] In a third aspect of the invention, an electronic device is provided. The electronic device includes a memory and a processor, the memory storing a computer program, the processor executing the program to implement the method according to a first aspect of the invention.

[0017] In a fourth aspect of the invention, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the method according to a first aspect of the invention.

[0018] This invention improves the accuracy of steel product quality prediction by jointly optimizing the network structure and hyperparameters through variable-length coding and combining it with an improved ES algorithm. It is suitable for steel production scenarios with multiple varieties, small batches, and fast pace.

[0019] It should be understood that the description in the Summary of the Invention is not intended to limit the key or essential features of the embodiments of the present invention, nor is it intended to restrict the scope of the invention. Other features of the invention will become readily apparent from the following description.

[0020] The beneficial effects of this invention are: 1. An improved evolutionary strategy (ES) algorithm is used to automatically search for the optimal neural network structure and hyperparameters, overcoming the shortcomings of manual network structure design, which relies on experience and is prone to getting trapped in local optima. Through a variable-length encoding strategy, the algorithm can flexibly adjust the number of network layers and the parameters of each layer to generate customized models for specific steel production data characteristics.

[0021] 2. Addressing the challenges of multi-variety, small-batch production in the steel industry, and the frequent changes in data characteristics caused by diverse customer demands for steel performance, an automated Neural Architecture Search (NAS) replaces fixed, manual model design. When production conditions or data distribution change, there's no need for manual model reconstruction; the algorithm automatically optimizes and generates a network architecture adapted to the new environment. This effectively solves the pain point of traditional fixed models struggling to maintain high accuracy, supporting agile product development and production adjustments. Attached Figure Description

[0022] The above and other features, advantages, and aspects of the various embodiments of the present invention will become more apparent from the accompanying drawings and the following detailed description. Wherein: Figure 1 A flowchart of a method for predicting the quality of steel products based on neural architecture search using an evolutionary strategy, according to an embodiment of the present invention, is shown. Figure 2 A schematic diagram of an encoding method according to an embodiment of the present invention is shown; Figure 3 A flowchart of the improved ES algorithm according to an embodiment of the present invention is shown; Figure 4 A block diagram of an apparatus for predicting the quality of steel products based on neural architecture search using an evolutionary strategy, according to an embodiment of the present invention, is shown. Figure 5 A schematic diagram of a device for predicting the quality of steel products based on neural architecture search using an evolutionary strategy, according to an embodiment of the present invention, is shown. Detailed Implementation

[0023] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0024] According to an embodiment of the present invention, a method for predicting the quality of steel products based on neural architecture search using an evolutionary strategy is proposed. By jointly optimizing the network structure and hyperparameters through variable-length encoding and combining it with an improved ES algorithm, a neural network architecture and hyperparameter tuning search for steel product quality prediction is used, which improves the accuracy of steel product quality prediction and is suitable for steel production scenarios with multiple varieties, small batches and fast pace.

[0025] The principles and spirit of the present invention will be explained in detail below with reference to several representative embodiments.

[0026] Figure 1 This is a schematic flowchart of a method for predicting the quality of steel products based on neural architecture search using an evolutionary strategy, according to an embodiment of the present invention. The method includes: Step S01: Obtain historical data and relevant parameters affecting steel quality in the target production line or scenario to construct the feature set required for model training and validation; Step S02: Based on the feature set, construct a search space containing network structure and hyperparameters, and use a variable-length encoding strategy to encode the number of network layers, number of neurons, activation function, optimizer type and training parameters to generate an initial candidate model population; Step S03: The improved ES algorithm is used to automatically search for the structure and hyperparameters of the neural network, and backpropagation is used to train and evaluate the performance of each generation of candidate models; Step S04: Use the trained optimal model to predict the quality of steel production data.

[0027] It should be noted that although the operation of the method of the present invention has been described in a specific order in the above embodiments and figures, this does not require or imply that the operations must be performed in that specific order, or that all the operations shown must be performed to achieve the desired result. Additionally or alternatively, certain steps may be omitted, multiple steps may be combined into one step, and / or one step may be broken down into multiple steps.

[0028] To provide a clearer explanation of the above-described method for predicting steel product quality based on neural architecture search using evolutionary strategies, a specific embodiment is described below. However, it is worth noting that this embodiment is only for better illustrating the present invention and does not constitute an undue limitation on the present invention.

[0029] The following example will further illustrate the method of predicting steel product quality based on neural architecture search using an evolutionary strategy.

[0030] Step S01: Obtain historical data and relevant parameters affecting steel quality in the target production line or scenario to construct the feature set required for model training and validation.

[0031] Relevant data was collected based on the steel product type whose quality was to be predicted, including sensor data and parameters that affect product quality during the quality measurement and analysis process. Historical data and parameters of factors that may affect steel product quality underwent data preprocessing: for features with missing values, mean or interpolation methods were used to fill in the missing values; data with many duplicate values ​​within a specific time period were removed; outliers were detected and deleted if found; numerical features such as tapping and converter temperatures, and casting time of process parameters were normalized; and categorical data such as the addition or non-addition of auxiliary materials and the use of certain processes were subjected to one-hot encoding.

[0032] Step S02: Based on the feature set, construct a search space containing network structure and hyperparameters, and use a variable-length encoding strategy to encode the number of network layers, number of neurons, activation function, optimizer type and training parameters to generate an initial candidate model population.

[0033] To avoid the NAS search space becoming too large, the range of parameters to be optimized is set, and the range and type of encoding parameters are set according to the size of the dataset.

[0034] To more flexibly and efficiently express the most important network structure components when optimizing using the improved evolutionary strategy (ES) algorithm proposed in this invention, a variable-length encoding method based on network layer structure is used to encode hyperparameters such as the number of hidden layers, the number of neurons in each layer, the type of activation function, the type of optimizer, and training parameters into chromosomes. During encoding, the number of hidden layers in the neural network is determined first, and then the specific hyperparameters required for each layer, such as the number of neurons and the type of activation function, are determined. Figure 2 As shown.

[0035] Step S03: Use the improved ES algorithm to automatically search for the structure and hyperparameters of the neural network (BPNN), and use backpropagation to train and evaluate the performance of each generation of candidate models.

[0036] like Figure 3 As shown, the specific steps are as follows, where the population... and Elite Collection The quantities are 3 and 1 respectively.

[0037] Step S031: Initialize the parameters of the improved ES algorithm: initial population Population size Number of elites Maximum number of iterations .

[0038] Step S032: Perform algorithm iteration, first on the population All individuals are randomly selected to undergo two mutation operations using a mutation operator, resulting in... Each individual is recorded as a descendant population. .

[0039] Among them, the mutation operation does not use the traditional ( + )-ES Evolutionary Strategy Strength Vector Instead of controlling the mutation, two different mutation operators are used for mutation operations: genotypic mutation and multipoint mutation.

[0040] Gene segments that correspond to different functional types in the encoding are called genotypes, such as hidden layer structure, optimizer type, and training parameters. Figure 2 Different parameter types in the text.

[0041] For genotype mutations, the genotype segment to be mutated is first randomly selected. For the variable-length portion of the encoding, i.e., the hidden layer network structure, the number of layers is first randomly re-determined to determine the new network architecture in the offspring. As the number of network layers increases, the hyperparameters (number of neurons and activation function type) of the increased network structure are randomly generated. If the selected genotype is a fixed part of the encoding (optimizer type and training parameters), the mutation will be randomly generated within its corresponding encoding space.

[0042] For multi-point mutation, several genes on the chromosome are randomly selected for mutation, excluding the number of network layers. The number of mutation points is randomly generated within an integer range. Genotypic mutation aims to explore better architectures, while multi-point mutation is used to reduce the probability of the population getting trapped in local optima.

[0043] Step S033: For the population Perform the selection operation, denoted as .

[0044] Step S034: Population After individuals are selected using the roulette wheel operator, a single-point crossover operator is used to perform a crossover operation, randomly selecting the genotype boundary point on the chromosome as the crossover point, and the offspring population is denoted as […]. .

[0045] Step S035: For the population and Perform joint selection operation and record .

[0046] Step S036: For the parent population P Using an elite retention strategy, remember... .

[0047] Step S037: Decompose the elite set Q Merging with contemporary superior individuals to form a new population, .

[0048] Step S038: End After the second iteration, the optimal individual is selected. This is the final solution.

[0049] In the classic ( + In the )-ES algorithm, Each parent generation first produces through mutation. Each offspring, and then from the whole... + Select the individual with the best fitness value Each individual proceeds to the next iteration. In the improved ES algorithm of this invention, to better explore the solution space, Each parent will randomly select a mutation operator to perform two mutations (the mutation operators are different in the two mutation operations), and start from 2 individual offspring populations ( Choose the optimal one. Each offspring. Then, in order to promote the transmission of superior gene fragments within the population, the population... The offspring population is generated through roulette wheel selection and crossover operations. In algorithm step S035, the merged population is... Select the optimal offspring population. Furthermore, the improved ES algorithm uses an elitist strategy to avoid losing high-quality solutions and allows elite solutions to participate in the evolutionary process, ensuring that the optimal individual can evolve fully. In steps S036 and S037, the elite solution set... Q Preservation of the previous generation population P The optimal solution is then compared with the optimal offspring population of this generation. The populations are merged into a new population. Finally, the improved ES algorithm has 2×10^2 evaluations. The values ​​are in lines 3 and 5 of the algorithm, respectively. It is worth noting that each evaluation requires sufficient training of the network represented by the individual in the population before it can be performed.

[0050] Step S04: Use the trained optimal model to predict the quality of steel production data.

[0051] The optimal BPNN network structure and hyperparameters of the individuals obtained through the iterative ES algorithm proposed in this invention are used to train the prediction model. The trained optimal model is then used to predict the quality of steel production data.

[0052] To demonstrate the effectiveness of the method presented in this application, a dataset provided by steel companies is used. This dataset includes various factors affecting the quality of steel products, such as the main uses of the target steel grade (carbon structural steel, alloy tool steel, and precision alloys), the proportions of auxiliary materials and smelting batching methods, tapping temperature, casting time of process parameters, and various final chemical compositions of furnace steel. These are used as feature data in the dataset. Two key stage indicators in the production and R&D phases, namely ingot yield and rolled product yield, are used to reflect the product quality at different stages. Therefore, two datasets are provided, each corresponding to one of these two indicators.

[0053] The comparison algorithms were selected based on applicability and effectiveness. First, classic prediction models, including Convolutional Neural Networks (CNNs), Long Short-Term Memory Networks (LSTMs), and Support Vector Machines (SVMs), were compared; the architectures and hyperparameters of these models were manually tuned. Second, a manually tuned Backpropagation Neural Network (BPNN) was used to verify the effectiveness of the proposed method.

[0054] During network training, data from the dataset will be used for training, validation, and testing in a 6:2:2 ratio. The Mean Absolute Error (MAE) can be used as the model evaluation metric, and can be expressed as: ,in Represents the actual label value, These are the model's predicted values. m The size of the dataset is specified. Algorithm parameter settings are as described in Table 1.

[0055] Table 1

[0056] The experimental results of each algorithm on the spindle yield and rolled product yield test sets are shown in Table 2. The MAE (Maximum Averaging Value) was used as the error performance index, and the mean and standard deviation of five independent tests were recorded. Furthermore, the number of parameters in each model network reflects its complexity to some extent.

[0057] Table 2

[0058] As shown in Table 2, on the spindle yield dataset, LSTM outperforms CNN in terms of average error, but on the yield dataset, CNN outperforms LSTM. SVM slightly outperforms both CNN and LSTM, while BPNN performs the worst on both datasets. This indicates that for the actual prediction problem in this embodiment, it is difficult to achieve good performance for BPNN by manually designing a suitable network architecture and adjusting hyperparameters. However, compared to manually designed BPNN, the algorithm proposed in this invention can significantly improve the performance of BPNN. Although, as shown in Table 2, the model of this invention has more network parameters than manually designed BPNN, this does not mean that its performance is worse. Rather, it is because the algorithm tends to design more complex structures to improve prediction accuracy. The standard deviation of the test results reflects the stability of the model's prediction performance to some extent. As shown in the standard deviation experimental results in Table 2, SVM performs best in terms of standard deviation because, for a fixed test set, SVM can obtain the same prediction result for each test after determining the hyperplane. Except for SVM, among the neural network-based prediction models, the model of this invention performs best in terms of standard deviation across different datasets.

[0059] Based on the same inventive concept, this invention also proposes a device for predicting the quality of steel products based on neural architecture search using an evolutionary strategy. The implementation of this device is similar to the implementation of the method described above, and repeated details will not be elaborated further. Figure 4 As shown, the device 100 includes: Feature set construction module 101: used to obtain historical data and related parameters affecting steel quality in the target production line or scenario, in order to construct the feature set required for model training and validation; Model building module 102: Based on the feature set, it constructs a search space containing network structure and hyperparameters, and uses a variable-length encoding strategy to encode the number of network layers, number of neurons, activation function, optimizer type and training parameters to generate an initial candidate model population. Model optimization module 103: Used to automatically search for the structure and hyperparameters of the neural network using the improved ES algorithm, and to evaluate the performance of each generation of candidate models by training with backpropagation; Model Inference Module 104: Used to predict the quality of steel production data using the trained optimal model.

[0060] Those skilled in the art will understand that, for the sake of convenience and brevity, the specific working process of the described module can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0061] like Figure 5 As shown, the device includes a central processing unit (CPU), which can perform various appropriate actions and processes based on computer program instructions stored in read-only memory (ROM) or loaded from storage units into random access memory (RAM). The RAM can also store various programs and data required for device operation. The CPU, ROM, and RAM are interconnected via a bus. Input / output (I / O) interfaces are also connected to the bus.

[0062] Multiple components in the device are connected to the I / O interface, including: input units such as keyboards and mice; output units such as various types of displays and speakers; storage units such as disks and optical discs; and communication units such as network interface cards (NICs), modems, and wireless transceivers. The communication unit allows the device to exchange information / data with other devices through computer networks such as the Internet and / or various telecommunications networks.

[0063] The processing unit executes the various methods and processes described above, such as method steps S01 to S04. For example, in some embodiments, method steps S01 to S04 may be implemented as a computer software program tangibly contained in a machine-readable medium, such as a storage unit. In some embodiments, part or all of the computer program may be loaded and / or installed on the device via ROM and / or a communication unit. When the computer program is loaded into RAM and executed by the CPU, one or more steps of method steps S01 to S04 described above may be performed. Alternatively, in other embodiments, the CPU may be configured to execute method steps S01 to S04 by any other suitable means (e.g., by means of firmware).

[0064] The functions described above in this document can be performed at least in part by one or more hardware logic components. For example, exemplary types of hardware logic components that can be used, without limitation, include: field programmable gate arrays (FPGAs), application-specific integrated circuits (ASICs), application-specific standard products (ASSPs), systems-on-a-chip (SoCs), payload programmable logic devices (CPLDs), and so on.

[0065] The program code used to implement the methods of the present invention can be written in any combination of one or more programming languages. This program code can be provided to a processor or controller of a general-purpose computer, special-purpose computer, or other programmable data processing device, such that when executed by the processor or controller, the program code causes the functions / operations specified in the flowcharts and / or block diagrams to be implemented. The program code can be executed entirely on the machine, partially on the machine, as a standalone software package partially on the machine and partially on a remote machine, or entirely on a remote machine or server.

[0066] In the context of this invention, a machine-readable medium can be a tangible medium that may contain or store a program for use by or in conjunction with an instruction execution system, apparatus, or device. A machine-readable medium can be a machine-readable signal medium or a machine-readable storage medium. Machine-readable media can include, but are not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatus, or devices, or any suitable combination of the foregoing. More specific examples of machine-readable storage media include electrical connections based on one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fibers, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination of the foregoing.

[0067] Furthermore, although the operations are described in a specific order, this should be understood as requiring that such operations be performed in the specific order shown or in sequential order, or requiring that all illustrated operations be performed to achieve the desired result. In certain environments, multitasking and parallel processing may be advantageous. Similarly, although several specific implementation details are included in the above discussion, these should not be construed as limiting the scope of the invention. Certain features described in the context of individual embodiments may also be implemented in combination in a single implementation. Conversely, various features described in the context of a single implementation may also be implemented individually or in any suitable sub-combination in multiple implementations.

[0068] Although the subject matter has been described using language specific to structural features and / or methodological logic, it should be understood that the subject matter defined in the appended claims is not necessarily limited to the specific features or actions described above. Rather, the specific features and actions described above are merely illustrative examples of implementing the claims.

Claims

1. A method for predicting the quality of steel products based on neural architecture search using an evolutionary strategy, characterized in that, The method includes: Step S01: Obtain historical data and relevant parameters affecting steel quality in the target production line or scenario to construct the feature set required for model training and validation; Step S02: Based on the feature set, construct a search space containing network structure and hyperparameters, and use a variable-length encoding strategy to encode the number of network layers, number of neurons, activation function, optimizer type and training parameters to generate an initial candidate model population; Step S03: The improved ES algorithm is used to automatically search for the structure and hyperparameters of the neural network, and backpropagation is used to train and evaluate the performance of each generation of candidate models; Step S04: Use the trained optimal model to predict the quality of steel production data.

2. The method for predicting the quality of steel products based on neural architecture search using an evolutionary strategy, as described in claim 1, is characterized in that... The historical data and related parameters mentioned in step S01 include sensor data and parameters that affect product quality during the quality measurement and analysis process.

3. The method for predicting the quality of steel products based on neural architecture search using an evolutionary strategy according to claim 1, characterized in that, After obtaining historical data and relevant parameters affecting steel quality in the target production line or scenario as described in step S01, the data is preprocessed as follows: for features with missing values, the mean or interpolation method is used to fill them according to their characteristics; data with a large number of duplicate values ​​in a specific time period is removed; outliers are detected and deleted if any are found; and numerical features are normalized. Perform one-hot encoding on categorical data.

4. The method for predicting the quality of steel products based on neural architecture search using an evolutionary strategy according to claim 1, characterized in that, The encoding described in step S02 first determines the number of hidden layers in the neural network, and then determines the specific hyperparameters required for each layer.

5. The method for predicting the quality of steel products based on neural architecture search using an evolutionary strategy according to claim 1, characterized in that, The specific steps of step S03 are as follows: Step S031: Initialize the parameters of the improved ES algorithm: initial population Population size Number of elites Maximum number of iterations ; Step S032: Perform iterations, first for the population All individuals are randomly selected to undergo two mutation operations using a mutation operator, resulting in... Each individual is recorded as a descendant population. ; Step S033: For the population Perform the selection operation, denoted as ; Step S034: Population After individuals are selected using the roulette wheel operator, a single-point crossover operator is used to perform a crossover operation, randomly selecting the genotype boundary point on the chromosome as the crossover point, and the offspring population is denoted as […]. ; Step S035: For the population and Conduct joint selection operation; Step S036: For the parent population P Use an elite retention strategy; Step S037: Decompose the elite set Q Merge with the best-performing individuals of the present generation to form a new population; Step S038: End After several iterations, the optimal individual is selected as the final solution.

6. The method for predicting the quality of steel products based on neural architecture search using an evolutionary strategy according to claim 5, characterized in that, The two mutation operations described in step S032 are: genotypic mutation and multi-point mutation. Genotype mutations randomly select genotype segments to be mutated; for the variable-length part of the code, i.e. the hidden layer network structure, the number of layers is first randomly re-determined to determine the new network architecture in the offspring individuals; as the number of network layers increases, the hyperparameters of the added network structure are randomly generated; if the selected genotype is a fixed part of the code, the mutation will be randomly generated in its corresponding coding space.

7. The method for predicting the quality of steel products based on neural architecture search using an evolutionary strategy according to claim 6, characterized in that, The multi-point mutation randomly selects several genes in the chromosome for mutation, and does not include the number of network layers. The number of mutation points is randomly generated within an integer range.

8. An apparatus for predicting the quality of steel products based on neural architecture search using an evolutionary strategy, characterized in that, The device implements the method as described in any one of claims 1 to 7, comprising: Feature set construction module: used to acquire historical data and relevant parameters affecting steel quality in the target production line or scenario, in order to build the feature set required for model training and validation; Model building module: Based on the feature set, it constructs a search space containing network structure and hyperparameters, and uses a variable-length encoding strategy to encode the number of network layers, number of neurons, activation function, optimizer type and training parameters to generate an initial candidate model population. Model optimization module: Used to automatically search for the structure and hyperparameters of neural networks using an improved ES algorithm, and to evaluate the performance of each generation of candidate models using backpropagation training; Model Inference Module: Used to predict the quality of steel production data using the trained optimal model.

9. An electronic device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the program, it implements the method as described in any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the method as described in any one of claims 1 to 7.

Images