A hot-rolled bar mechanical property prediction method and system based on deep learning

Abstract

The application provides a hot-rolled bar mechanical property prediction method and system based on deep learning. The method acquires and pre-processes process parameter data of the hot-rolled bar; inputs an effective data set into a pre-trained mechanical property prediction model, acquires a time-space feature matrix of the process parameters based on a time-space attention mechanism, and multiplies the time-space feature matrix by the original input to obtain a weighted input sequence; performs deep feature extraction on the weighted sequence through a Transformer module to obtain an output vector; and synchronously outputs prediction results including yield strength, tensile strength and elongation by using a multi-task learning module. Through deep integration of the time-space attention mechanism, the Transformer and the multi-task learning, the application can fully mine complex time-space correlations among the process parameters, realize high-precision and high-efficiency parallel prediction of multiple mechanical property indexes, and effectively solve the deficiencies of traditional mechanism models and machine learning methods in terms of precision and generalization ability.

Description

Technical Field

[0001] This invention relates to the field of metal material performance prediction technology, and in particular to a method and system for predicting the mechanical properties of hot-rolled bars based on deep learning. Background Technology

[0002] Hot-rolled bars, as a fundamental and crucial material in modern industrial systems, possess excellent mechanical properties that are a core guarantee for the development of fields such as construction engineering, bridge construction, transportation, machinery manufacturing, and energy equipment. Currently, the market has increasingly stringent requirements for their load-bearing limits, durability, and performance stability; however, the industry still widely relies on traditional quality control models based on "trial and error" and "destructive testing." This model not only consumes large amounts of raw materials, has long testing cycles, and is costly, but also severely restricts the rapid improvement and continuous optimization of production efficiency and product quality. Therefore, researching methods that can quickly and accurately predict the mechanical properties of hot-rolled bars is of great significance for achieving online monitoring of product quality, guiding the optimization of process parameters, and promoting the efficient and intelligent development of the industry.

[0003] In recent years, the rapid development of artificial intelligence and big data technologies has provided new technical pathways for predicting the mechanical properties of hot-rolled bars. Existing research focuses mainly on two types of models: one is a mechanistic model based on metallurgical physicochemical principles, and the other is a machine learning model based on historical data.

[0004] However, both technical approaches have significant limitations. Metallurgical mechanism models rely heavily on complex mechanistic assumptions and empirical parameters, making it difficult to accurately describe and uncover the nonlinear dynamic changes of various parameters in the complex hot rolling process. Furthermore, these models have long iteration cycles, poor versatility, and cannot effectively cover the varied production scenarios in modern bar production lines. Traditional machine learning models (such as linear regression and support vector machines) can utilize data-driven modeling, but their ability to automatically mine the deep, multi-granular features inherent in massive, high-dimensional, and strongly coupled process data is limited. They cannot fully capture the complex interactions and correlations between various parameters such as rolling temperature, rolling rate, and alloy composition, resulting in a trade-off between prediction accuracy and model generalization ability, failing to meet the high requirements of prediction accuracy and stability in industrial settings.

[0005] Therefore, there is an urgent need in this field for a method that can deeply integrate process characteristics, deeply mine the complex relationships inherent in data, and simultaneously predict multiple key mechanical performance indicators with high accuracy. Summary of the Invention

[0006] This invention provides a deep learning-based device for predicting the mechanical properties of hot-rolled bars, which can achieve high-precision and high-efficiency parallel prediction of the mechanical property indicators of hot-rolled bars.

[0007] The first aspect of this invention provides a method for predicting the mechanical properties of hot-rolled bars based on deep learning, comprising: Obtain process parameter data during the hot-rolled bar production process and preprocess the data to obtain a valid dataset; The effective dataset is input into a pre-trained mechanical performance prediction model, which obtains the spatiotemporal feature matrix of the process parameter data based on a spatiotemporal attention mechanism. Multiplying the spatiotemporal feature matrix with the original input data yields a weighted input sequence; The weighted input sequence is fed into the Transformer module for feature extraction to obtain the output vector; The output vector is input into the multi-task learning module for prediction, and the predicted mechanical properties of hot-rolled bars are obtained. The predicted results include the yield strength, tensile strength and elongation of hot-rolled bars.

[0008] Optionally, the spatiotemporal feature matrix of the process parameter data obtained based on the spatiotemporal attention mechanism includes: Based on the spatial attention mechanism, the softmax function is used to calculate the spatial weight of a single variable, and the spatial weights of all time series are combined into a spatial attention matrix. Based on the time attention mechanism, the softmax function is used to calculate the time weight of a single variable, and the time weights of each sequence are concatenated into a time attention matrix. The spatial attention matrix and the temporal attention matrix are multiplied element-wise to obtain the spatiotemporal feature matrix.

[0009] Optionally, the spatial attention mechanism constructs the attention matrix by calculating the normalized weights of each variable in the input sequence at all time steps; the temporal attention mechanism constructs the attention matrix by calculating the normalized weights of all variables at each time point.

[0010] Optionally, the weighted input sequence is fed into the Transformer module for feature extraction to obtain an output vector, specifically including: The weighted input sequence is computed through a multi-head self-attention layer; The output of the multi-head self-attention layer is summed with the weighted input sequence, and then the layer is normalized. The normalized intermediate vector is input into the feedforward network, and features are filtered through the dropout mechanism. The output of the feedforward network is again subjected to residual summation and layer normalization. Global average pooling is performed on the final output feature sequence to obtain the output vector.

[0011] Optionally, the output vector is input into a multi-task learning module for prediction to obtain the predicted mechanical properties of hot-rolled bars, specifically including: The features in the output vector corresponding to the three major tasks of yield strength, tensile strength and elongation are shared and unified to obtain a shared output. The shared output is input to a linear layer, and synchronous decoding is used to obtain the predicted mechanical properties of hot-rolled bars, including yield strength, tensile strength, and elongation.

[0012] Optionally, when training the mechanical performance prediction model, a multi-task loss function based on homoscedasticity uncertainty is used for model optimization. The loss function is used to collaboratively minimize the combined error of all prediction tasks.

[0013] Optionally, the process parameter data includes the furnace tapping / continuous casting hot delivery temperature, roughing rolling temperature, intermediate rolling temperature, pre-finishing rolling temperature, finishing rolling temperature, upper cooling bed temperature, rolling speed, and alloy chemical composition content.

[0014] A second aspect of the present invention provides a deep learning-based device for predicting the mechanical properties of hot-rolled bars, comprising: The data processing unit is used to acquire process parameter data during the hot-rolled bar production process and preprocess the data to obtain a valid dataset. The spatiotemporal feature extraction unit is used to input the effective dataset into the pre-trained mechanical performance prediction model. The mechanical performance prediction model obtains the spatiotemporal feature matrix of the process parameter data based on the spatiotemporal attention mechanism. The spatiotemporal feature matrix is ​​multiplied with the original input data to obtain a weighted input sequence. The deep feature extraction unit is used to input the weighted input sequence into the Transformer module for feature extraction and to obtain the output vector; The prediction unit is used to input the output vector into the multi-task learning module for prediction, and obtain the prediction results of the mechanical properties of hot-rolled bars. The prediction results include the yield strength, tensile strength and elongation of hot-rolled bars.

[0015] A third aspect of the present invention provides an electronic device, comprising: One or more processors; A memory on which one or more programs are stored; When the one or more programs are executed by the one or more processors, the one or more processors implement the deep learning-based method for predicting the mechanical properties of hot-rolled bars as described in any of the preceding claims.

[0016] A fourth aspect of the present invention provides a computer storage medium for storing a program, which, when executed, is used to implement the deep learning-based method for predicting the mechanical properties of hot-rolled bars as described in any of the preceding claims.

[0017] Beneficial Effects: This invention provides a method and system for predicting the mechanical properties of hot-rolled bars based on deep learning. The method involves acquiring and preprocessing process parameter data of the hot-rolled bars; inputting the effective dataset into a pre-trained mechanical property prediction model; obtaining the spatiotemporal feature matrix of the process parameters based on a spatiotemporal attention mechanism; multiplying this matrix with the original input to obtain a weighted input sequence; performing deep feature extraction on the weighted sequence using a Transformer module to obtain an output vector; and simultaneously outputting prediction results including yield strength, tensile strength, and elongation using a multi-task learning module. This invention, through the deep integration of spatiotemporal attention mechanism, Transformer, and multi-task learning, can fully explore the complex spatiotemporal correlations between process parameters, achieving high-precision and high-efficiency parallel prediction of multiple mechanical property indicators, effectively solving the shortcomings of traditional mechanistic models and machine learning methods in terms of accuracy and generalization ability. The spatiotemporal attention mechanism effectively solves the adaptability defects of single-dimensional attention in hot-rolled bar production scenarios by capturing the dynamic temporal and spatial correlations of data, and is more robust in dynamic and complex scenarios. The multi-task learning mechanism can better discover the correlations between tasks, helping to improve the prediction accuracy of the model. This invention combines spatiotemporal attention mechanism and multi-task mechanism to fully explore the potential relationships and mutual influences in the data, thereby achieving accurate prediction of the mechanical properties of hot-rolled bars. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 A schematic diagram of the module structure of a deep learning-based method for predicting the mechanical properties of hot-rolled bars, provided in an embodiment of the present invention; Figure 2 A flowchart illustrating a method for predicting the mechanical properties of hot-rolled bars based on deep learning, provided for an embodiment of the present invention; Figure 3 A schematic diagram of a deep learning-based device for predicting the mechanical properties of hot-rolled bars provided in an embodiment of the present invention; Figure 4 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present invention. Detailed Implementation

[0020] This invention provides a deep learning-based device for predicting the mechanical properties of hot-rolled bars, which can achieve high-precision and high-efficiency parallel prediction of the mechanical property indicators of hot-rolled bars.

[0021] Figure 1 This is a schematic diagram of the module structure of a deep learning-based method for predicting the mechanical properties of hot-rolled bars, provided in an embodiment of the present invention. Figure 1 As shown, the core processing flow of this method involves three core parts: spatiotemporal attention mechanism, Transformer module, and multi-task learning module.

[0022] The spatiotemporal attention mechanism receives the original process parameter input and contains two parallel computing units: temporal attention and spatial attention, which are used to extract feature weights from the input data from the time and spatial dimensions, respectively.

[0023] The Transformer module receives the input sequence weighted by spatiotemporal attention. Its internal processing flow includes: calculation through a multi-head attention layer, followed by layer normalization and residual connection; the processed features are fed into a feedforward neural network for nonlinear transformation, followed by layer normalization and residual connection again; finally, the feature sequence is integrated through a global average pooling layer to output a fixed-length feature vector.

[0024] The multi-task learning module receives feature vectors. First, the features are processed and concatenated using shared parameters in the Share & Concat section. Then, the integrated features are input into three independent task branches: Task1, Task2, and Task3, which correspond to the predicted outputs of yield strength, tensile strength, and elongation of hot-rolled bars, respectively.

[0025] Figure 1 The invention clearly demonstrates the data flow path and the compositional relationship of each core processing step in the method of this invention, reflecting the complete technical process from feature enhancement, deep coding to multi-task prediction.

[0026] See Figure 2 This figure is a flowchart illustrating a deep learning-based method for predicting the mechanical properties of hot-rolled bars according to an embodiment of the present invention. The deep learning-based method for predicting the mechanical properties of hot-rolled bars provided in this embodiment of the present invention combines... Figure 1 For example, this can be achieved through the following steps S201-205.

[0027] S201: Obtain process parameter data during the hot-rolled bar production process and preprocess the data to obtain a valid dataset.

[0028] In this embodiment of the invention, the process parameter data includes the furnace tapping / continuous casting hot delivery temperature, roughing rolling temperature, intermediate rolling temperature, pre-finishing rolling temperature, finishing rolling temperature, upper cooling bed temperature, rolling speed, and alloy chemical composition content.

[0029] S202: Input the effective dataset into the pre-trained mechanical performance prediction model. The mechanical performance prediction model obtains the spatiotemporal feature matrix of the process parameter data based on the spatiotemporal attention mechanism.

[0030] In this embodiment of the invention, based on the spatial attention mechanism, the softmax function is used to calculate the spatial weight of a single variable, and the spatial weights of all time series are combined into a spatial attention matrix; based on the temporal attention mechanism, the softmax function is used to calculate the temporal weight of a single variable, and the temporal weights of each sequence are concatenated into a temporal attention matrix; the spatial attention matrix and the transpose of the temporal attention matrix are multiplied element-wise to obtain the spatiotemporal feature matrix. The spatial attention mechanism constructs the attention matrix by calculating the normalized weight of each variable in the input sequence at all time steps; the temporal attention mechanism constructs the attention matrix by calculating the normalized weight of all variables at each time point.

[0031] Specifically, the temporal features of hot-rolled bar process data are obtained based on the temporal attention mechanism module, and the spatial features are obtained based on the spatial attention mechanism module. The temporal and spatial features are then fused to obtain the spatiotemporal features of the hot-rolled bar process data. The input data is... ,in and This represents the features and variables of the input data. The time step of sequence t is... The ReLU() function is used as the activation function. The function generates a spatial attention matrix. It then calculates the spatial weights of individual variables. : ; Combine the spatial weights of all time series in the input sequence into a spatial attention matrix A for each input: ; Time attention can focus on the relationship between hot-rolled bar process parameters at consecutive time points; similarly... The time weight of a single variable: ; The time weights of each sequence are concatenated to obtain the time attention matrix B for each input: ; Finally, the spatial and temporal attention matrices are multiplied to obtain the spatiotemporal attention matrix STA of the input data. ; The spatial attention matrix and the temporal attention matrix are multiplied by matrix element-wise multiplication to obtain the spatiotemporal feature matrix of the input data.

[0032] In one implementation of this invention, the invention establishes a deep learning-based prediction model for the mechanical properties of hot-rolled bars, comprising the following steps: Hot rolling process data was obtained from a hot-rolled bar mill. Each data point contained 15 items (heating furnace tapping / continuous casting hot delivery temperature, roughing rolling temperature, intermediate rolling temperature, pre-finishing rolling temperature, finishing rolling temperature, upper cooling bed temperature, alloy composition C, alloy composition Si, alloy composition V, alloy composition Mn, etc.). The data conformed to international standards and had practical significance, and the experimental results could be used in real-world production environments. After data preprocessing operations such as removing outliers and missing values, valid data was obtained and used for training tests at a 9:1 ratio.

[0033] The advantage of the hot-rolled bar mechanical property prediction model in this invention lies in its consideration of not only the correlation and influence between tasks, but also its in-depth exploration of domain expertise and integration with process flow stages. This helps to analyze how features affect mechanical properties and improves the model's interpretability. This method proposes a spatiotemporal attention mechanism to simultaneously capture the correlation information of features in both time and space dimensions, thereby improving model performance. Furthermore, through a multi-task learning mechanism, information sharing and collaborative training among tasks simultaneously enhance the overall performance of multiple tasks, balancing efficiency and generalization ability.

[0034] S203: Multiply the spatiotemporal feature matrix with the original input data to obtain a weighted input sequence.

[0035] S204: Input the weighted input sequence into the Transformer module for feature extraction to obtain the output vector.

[0036] In this embodiment of the invention, the weighted input sequence is computed through a multi-head self-attention layer; the output of the multi-head self-attention layer and the weighted input sequence are summed by residual, and then normalized by layer; the normalized intermediate vector is input into a feedforward network, and features are filtered by the dropout mechanism; the output of the feedforward network is summed by residual and normalized by layer again; the final output feature sequence is subjected to global average pooling to obtain the output vector.

[0037] Specifically, the Transformer module extracts features from the weighted input sequence obtained in the previous step, as follows: First, the output of the multi-head attention is summed with the input sequence using residuals, followed by layer normalization; then, the resulting intermediate vector is input into the feedforward network, where features are filtered through network transformation and dropout mechanisms, followed by another round of residual summation and layer normalization; finally, the output vector of the feature extraction layer is obtained from the global average pooling layer. The self-attention mechanism can be determined by calculating the query matrix Q, the key matrix K, and the value matrix V. ; Each self-attention head captures different positions in the sequence. Then, the multi-head attention concatenates the output matrices and feeds them into a linear layer to obtain the multi-head attention result. : ; ; S205: Input the output vector into the multi-task learning module for prediction to obtain the predicted mechanical properties of hot-rolled bars. The predicted results include the yield strength, tensile strength and elongation of the hot-rolled bars.

[0038] In this embodiment of the invention, the features corresponding to the three main tasks of yield strength, tensile strength, and elongation in the output vector are shared and unified to obtain a shared output. The shared output is then input into a linear layer, and synchronous decoding is performed to obtain the predicted mechanical properties of hot-rolled bars, including yield strength, tensile strength, and elongation. When training the mechanical property prediction model, a multi-task loss function based on homoscedastic uncertainty is used for model optimization. The loss function is used to collaboratively minimize the combined error of all prediction tasks.

[0039] Specifically, a multi-task learning mechanism is used to predict the mechanical properties of hot-rolled bars, including yield strength, tensile strength, and elongation. To achieve multi-task learning, this layer shares and unifies the feature learning for the three main tasks: yield strength, tensile strength, and elongation. The prediction results are as follows: ; ; Following the shared parameter module, an independent decoder is designed for each prediction task. Multi-task prediction results are obtained through a task-specific output layer, resulting in three separate prediction results: ; For multi-output problems, a dedicated loss function is typically required for each output to assess error. This means the model needs to collaboratively minimize the combined error of all outputs, rather than optimizing individual outputs in isolation. To capture the inherent correlation between different outputs and avoid the cumbersome weight trade-offs and adjustments between multiple loss functions, a multi-task loss function based on homoscedastic uncertainty is adopted. This loss function can characterize the inherent uncertainty and relative confidence of different tasks. It is constructed by maximizing Gaussian likelihood, based on the premise of homoscedasticity of the data, as shown in the following formula: ; in, For noise scalar, For model parameters, For the true value vector, Output vectors for the model.

[0040] In one implementation of this invention, training and testing experiments are conducted at a ratio of 9:1. The formula for evaluating the mechanical properties of hot-rolled bars is as follows: ; in, Indicates the first Predicted values ​​for each sample; Indicates the first The true value of each sample; This represents the true mean of all samples.

[0041] This invention addresses the nonlinear, time-varying, strongly coupled, and multi-parameter characteristics of hot-rolled bar production data by proposing a deep learning-based method for predicting the mechanical properties of hot-rolled bars, primarily for use in hot-rolled bar production lines. This method predicts the mechanical properties of the finished rolled bars, including yield strength (YS), tensile strength (TS), and elongation (EL), serving as a crucial guiding and research tool for hot-rolled bar production lines. The production process of hot-rolled bars is complex, involving numerous production parameters related to mechanical properties, with interactions and correlations among these parameters. To accurately predict mechanical properties, this method uses deep learning to extract key features from production data such as roughing temperature, intermediate rolling temperature, pre-finishing rolling temperature, finishing rolling temperature, cooling bed temperature, rolling speed, and alloy composition, thereby establishing a predictive model between process parameters and mechanical properties. The main purpose of this method is to accurately predict the mechanical properties of hot-rolled bars in the hot rolling process, reduce safety hazards in the production process, promote continuous improvement of the production process, make manufacturers more competitive, and achieve high-quality development. Accurate and comprehensive prediction of product mechanical properties can serve as an important indicator for quality control, helping manufacturers to identify and solve problems in the production process in a timely manner and ensure stable product quality. In terms of versatility, this method is a deep learning model, capable of deeply mining the intrinsic relationships between data, performing well on unfamiliar data, exhibiting strong compatibility, moderate hardware requirements, and fast real-time inference speed. In terms of innovation, compared to physical models and machine learning, the deep learning model established by this method can extract deeper, more dimensional, and more granular data features from large amounts of data. It can more accurately capture important features between process parameters and effectively model nonlinear relationships. It can better extract the mutual influence and impact relationships between process parameters, automatically learn feature representations, better utilize data information, improve model prediction accuracy, and provide strong support for the comprehensive and accurate analysis of the mechanical properties of hot-rolled bars.

[0042] Based on the methods provided in the above embodiments, this invention also provides a deep learning-based device for predicting the mechanical properties of hot-rolled bars. The following description of the deep learning-based device for predicting the mechanical properties of hot-rolled bars is in conjunction with the accompanying drawings.

[0043] See Figure 3 The figure is a schematic diagram of a deep learning-based device for predicting the mechanical properties of hot-rolled bars, provided in an embodiment of the present invention.

[0044] The hot-rolled bar mechanical property prediction device 300 based on deep learning provided in this embodiment of the invention includes: a data processing unit 301, a spatiotemporal feature extraction unit 302, a deep feature extraction unit 303, and a prediction unit 304.

[0045] The data processing unit 301 is used to acquire process parameter data during the hot-rolled bar production process and preprocess the data to obtain a valid dataset. The spatiotemporal feature extraction unit 302 is used to input the effective dataset into the pre-trained mechanical performance prediction model. The mechanical performance prediction model obtains the spatiotemporal feature matrix of the process parameter data based on the spatiotemporal attention mechanism. The spatiotemporal feature matrix is ​​multiplied by the original input data to obtain a weighted input sequence. The deep feature extraction unit 303 is used to input the weighted input sequence into the Transformer module for feature extraction and obtain the output vector; The prediction unit 304 is used to input the output vector into the multi-task learning module for prediction, and obtain the prediction results of the mechanical properties of hot-rolled bars. The prediction results include the yield strength, tensile strength and elongation of hot-rolled bars.

[0046] In one possible implementation, the spatiotemporal feature extraction unit 302 is specifically used for: Based on the spatial attention mechanism, the softmax function is used to calculate the spatial weight of a single variable, and the spatial weights of all time series are combined into a spatial attention matrix. Based on the time attention mechanism, the softmax function is used to calculate the time weight of a single variable, and the time weights of each sequence are concatenated into a time attention matrix. The spatial attention matrix and the temporal attention matrix are multiplied element-wise to obtain the spatiotemporal feature matrix.

[0047] In one possible implementation, the spatiotemporal feature extraction unit 302 is specifically used for: Spatial attention mechanisms construct attention matrices by calculating the normalized weights of each variable in the input sequence at all time steps; temporal attention mechanisms construct attention matrices by calculating the normalized weights of all variables at each time point.

[0048] In one possible implementation, the deep feature extraction unit 303 has the following functions: The weighted input sequence is computed through a multi-head self-attention layer; The output of the multi-head self-attention layer is summed with the weighted input sequence, and then the layer is normalized. The normalized intermediate vector is input into the feedforward network, and features are filtered through the dropout mechanism. The output of the feedforward network is again subjected to residual summation and layer normalization. Global average pooling is performed on the final output feature sequence to obtain the output vector.

[0049] In one possible implementation, the prediction unit 304 has the following functions: The features in the output vector corresponding to the three major tasks of yield strength, tensile strength and elongation are shared and unified to obtain a shared output. The shared output is input to a linear layer, and synchronous decoding is used to obtain the predicted mechanical properties of hot-rolled bars, including yield strength, tensile strength, and elongation.

[0050] In one possible implementation, when training the mechanical performance prediction model, a multi-task loss function based on homoscedastic uncertainty is used for model optimization. The loss function is used to collaboratively minimize the combined error of all prediction tasks.

[0051] In one possible implementation, the process parameter data includes the furnace tapping / continuous casting hot delivery temperature, roughing rolling temperature, intermediate rolling temperature, pre-finishing rolling temperature, finishing rolling temperature, upper cooling bed temperature, rolling speed, and alloy chemical composition content.

[0052] Since the device 300 is the same as the deep learning-based method for predicting the mechanical properties of hot-rolled bars provided in the above method embodiments, the specific implementation of each unit of the device 300 is based on the same concept as the above method embodiments. Therefore, for the specific implementation of each unit of the device 300, please refer to the description of the deep learning-based method for predicting the mechanical properties of hot-rolled bars in the above method embodiments, and it will not be repeated here.

[0053] This invention also provides an electronic device, the device comprising: a processor and a memory; The memory is used to store instructions; The processor is used to execute the instructions in the memory to perform the deep learning-based method for predicting the mechanical properties of hot-rolled bars mentioned in the above embodiments.

[0054] It should be noted that the hardware structure of the electronic devices provided in the embodiments of the present invention can be as follows. Figure 4 The structure shown, Figure 4 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present invention.

[0055] Please see Figure 4 As shown, device 400 includes: processor 410, communication interface 420, and memory 430. The number of processors 410 in device 400 can be one or more. Figure 4 Taking a processor as an example, in this embodiment of the invention, the processor 410, communication interface 420, and memory 430 can be connected via a bus system or other means. Figure 4 Taking the connection between China and Israel via the bus system 440 as an example.

[0056] Processor 410 may be a central processing unit (CPU), a network processor (NP), or a combination of a CPU and an NP. Processor 410 may further include hardware chips. These hardware chips may be application-specific integrated circuits (ASICs), programmable logic devices (PLDs), or combinations thereof. The PLD may be a complex programmable logic device (CPLD), a field-programmable gate array (FPGA), generic array logic (GAL), or any combination thereof.

[0057] The memory 430 may include volatile memory, such as random-access memory (RAM); the memory 430 may also include non-volatile memory, such as flash memory, hard disk drive (HDD) or solid-state drive (SSD); the memory 430 may also include a combination of the above types of memory.

[0058] Optionally, the memory 430 stores an operating system and programs, executable modules, or data structures, or subsets thereof, or extended sets thereof. The programs may include various operation instructions for implementing various operations. The operating system may include various system programs for implementing various basic business processes and handling hardware-based tasks. The processor 410 can read the programs in the memory 430 to implement the deep learning-based method for predicting the mechanical properties of hot-rolled bars provided in this embodiment of the invention.

[0059] The bus system 440 can be a Peripheral Component Interconnect (PCI) bus or an Extended Industry Standard Architecture (EISA) bus, etc. The bus system 440 can be divided into address bus, data bus, control bus, etc. For ease of representation, Figure 4The bus is represented by a single thick line, but this does not mean that there is only one bus or one type of bus.

[0060] This invention also provides a computer-readable storage medium including instructions that, when executed on a computer, cause the computer to perform the deep learning-based method for predicting the mechanical properties of hot-rolled bars mentioned in the above embodiments.

[0061] This invention also provides a computer program product containing instructions that, when run on a computer, cause the computer to execute the deep learning-based method for predicting the mechanical properties of hot-rolled bars mentioned in the above embodiments.

[0062] Although the invention has been specifically shown and described in conjunction with preferred embodiments, those skilled in the art should understand that various changes in form and detail may be made to the invention without departing from the spirit and scope of the invention as defined in the appended claims, all of which shall be within the scope of protection of the invention.

Claims

1. A method for predicting the mechanical properties of hot-rolled bars based on deep learning, characterized in that, The method includes: The process parameter data of hot-rolled bar production process is obtained, and the data is preprocessed to obtain a valid dataset; The effective dataset is input into a pre-trained mechanical performance prediction model. During training, a multi-task loss function based on homoscedasticity uncertainty is used for model optimization. This loss function collaboratively minimizes the combined error of all prediction tasks. The mechanical performance prediction model obtains the spatiotemporal feature matrix of the process parameter data based on a spatiotemporal attention mechanism. This includes calculating the spatial weights of individual variables using a softmax function based on a spatial attention mechanism, and combining the spatial weights of all time series into a spatial attention matrix; calculating the temporal weights of individual variables using a softmax function based on a temporal attention mechanism, and concatenating the temporal weights of each sequence into a temporal attention matrix; and performing element-wise multiplication between the spatial attention matrix and the transpose of the temporal attention matrix to obtain the spatiotemporal feature matrix. Multiplying the spatiotemporal feature matrix with the original input data yields a weighted input sequence; The weighted input sequence is fed into the Transformer module for feature extraction to obtain the output vector; The output vector is input to a multi-task learning module for prediction to obtain the predicted mechanical properties of hot-rolled bars. The predicted results include the yield strength, tensile strength, and elongation of the hot-rolled bars. The process of inputting the output vector to the multi-task learning module for prediction to obtain the predicted mechanical properties of hot-rolled bars includes: sharing and unifying the features in the output vector corresponding to the three tasks of yield strength, tensile strength, and elongation to obtain a shared output; and inputting the shared output to a linear layer for synchronous decoding to obtain the predicted mechanical properties of hot-rolled bars including yield strength, tensile strength, and elongation.

2. The method according to claim 1, characterized in that, The spatial attention mechanism constructs an attention matrix by calculating the normalized weights of each variable in the input sequence at all time steps; the temporal attention mechanism constructs an attention matrix by calculating the normalized weights of all variables at each time point.

3. The method according to claim 1, characterized in that, The step of inputting the weighted input sequence into the Transformer module for feature extraction to obtain the output vector specifically includes: The weighted input sequence is computed through a multi-head self-attention layer; The output of the multi-head self-attention layer is summed with the weighted input sequence using residual summation, followed by layer normalization. The normalized intermediate vector is input into the feedforward network, and features are filtered through the dropout mechanism. The output of the feedforward network is again subjected to residual summation and layer normalization. The final output feature sequence is subjected to global average pooling to obtain the output vector.

4. The method according to claim 1, characterized in that, The process parameter data includes the furnace tapping / continuous casting hot delivery temperature, roughing rolling temperature, intermediate rolling temperature, pre-finishing rolling temperature, finishing rolling temperature, upper cooling bed temperature, rolling speed, and alloy chemical composition content.

5. A device for predicting the mechanical properties of hot-rolled bars based on deep learning, characterized in that, include: The data processing unit is used to acquire process parameter data during the hot-rolled bar production process and preprocess the data to obtain a valid dataset. A spatiotemporal feature extraction unit is used to input the effective dataset into a pre-trained mechanical performance prediction model. When training the mechanical performance prediction model, a multi-task loss function based on homoscedasticity uncertainty is used for model optimization. The loss function is used to collaboratively minimize the comprehensive error of all prediction tasks. The mechanical performance prediction model obtains the spatiotemporal feature matrix of the process parameter data based on a spatiotemporal attention mechanism. The spatiotemporal feature matrix is ​​multiplied by the original input data to obtain a weighted input sequence. The spatiotemporal feature extraction unit is specifically used to calculate the spatial weight of a single variable using the softmax function based on the spatial attention mechanism, and combine the spatial weights of all time series into a spatial attention matrix; and to calculate the temporal weight of a single variable using the softmax function based on the temporal attention mechanism, and concatenate the temporal weights of each sequence into a temporal attention matrix. The spatial attention matrix and the transpose of the temporal attention matrix are multiplied element by element to obtain the spatiotemporal feature matrix; The deep feature extraction unit is used to input the weighted input sequence into the Transformer module for feature extraction and to obtain an output vector; The prediction unit is used to input the output vector into the multi-task learning module for prediction to obtain the predicted mechanical properties of hot-rolled bars, including the yield strength, tensile strength and elongation of the hot-rolled bars. The prediction unit is specifically used to share and unify the features in the output vector corresponding to the three major tasks of yield strength, tensile strength and elongation to obtain a shared output; the shared output is input to a linear layer and synchronously decoded to obtain the prediction results of the mechanical properties of hot-rolled bars including yield strength, tensile strength and elongation.

6. An electronic device, characterized in that, The device includes: a processor and a memory; The memory is used to store instructions; The processor is configured to execute the instructions in the memory to perform the method according to any one of claims 1-4.

7. A computer-readable storage medium, characterized in that, Including instructions that, when run on a computer, cause the computer to perform the method described in any one of claims 1-4 above.

Images