A Method and System for Recommending Aluminum Material Processes Based on Knowledge Distillation
By optimizing the manufacturing process of aluminum alloy materials through hierarchical modeling based on knowledge distillation and attention strategies, the challenges of alloy element ratio and process parameter optimization in aluminum alloy material processing have been solved, achieving efficient prediction of aluminum material performance and improvement of production efficiency.
Patent Information
- Application Number
- CN202511263093.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2025-09-05
- Publication Date
- 2025-11-14
- Estimated Expiration
- 2045-09-05
AI Technical Summary
Existing technologies struggle to efficiently optimize alloy element ratios and process parameters in aluminum alloy material processing, leading to inconsistent product quality, low production efficiency, high model training costs, and poor fitting results.
A knowledge-based distillation approach is adopted. By constructing a performance prediction model, the influence of process path and alloy element ratio is modeled hierarchically. An attention strategy is used for distillation to optimize the process of aluminum materials and recommend the best processing configuration.
It improves the adaptability and reliability of aluminum material performance prediction, reduces computing resource requirements, is suitable for real-time industrial control scenarios, and enhances production efficiency and product quality consistency.
Smart Images

Figure CN120748586B_ABST
Abstract
Description
Technical Field
[0001] This application relates to the technical field of aluminum material processing, specifically to a method and system for recommending aluminum material processing procedures based on knowledge distillation. Background Technology
[0002] The material properties of aluminum alloys, such as strength, plasticity, and corrosion resistance, are mainly determined by several key factors, including the alloy element ratios, processing steps, and process parameters. Each step involves multi-dimensional parameter control. Current technologies have some shortcomings in optimizing and recommending alloy element ratios and process parameters during aluminum processing. Traditional methods rely mainly on experience and repeated experiments to determine alloy element ratios, which is inefficient and makes it difficult to obtain optimal ratios, thus affecting the research and development and production of high-performance aluminum alloys. Aluminum processing includes multiple steps such as melting, heat treatment, and casting. The lack of precise methods for determining and optimizing process parameters for each step makes it difficult to guarantee the consistency and stability of product quality.
[0003] In actual production, it is necessary to select the combination of components and processes that achieves all performance indicators from a high-dimensional space of element combinations and process parameters. Due to the huge solution space, strong nonlinearity, and complex interactions between variables, traditional methods are difficult to efficiently complete global optimization design. Although some schemes using nonlinear modeling such as machine learning have emerged for predicting aluminum material properties and recommending process flows, the high dimensionality of input variables and the large number of parameters lead to high model training costs and poor model convergence and fitting results.
[0004] For example, patent application CN118964992A discloses a method and system for optimizing aluminum alloy laser welding process parameters based on machine learning. The method includes: determining welding process parameters based on welding material parameters and obtaining the objective function to be optimized; constructing a machine learning model to establish the relationship between the welding process parameters and the objective function; generating experimental data through experiments under different welding process parameters and using the experimental data as a sample dataset; training the machine learning model using the sample dataset and evaluating the trained machine learning model using an independent test dataset to obtain an optimized machine learning model; and optimizing the objective function corresponding to the optimized machine learning model using a genetic algorithm to obtain the optimized welding process parameters. This technical solution can solve the problems of incomplete process parameter models and difficulty in optimizing process parameters in current applications.
[0005] The above technical solutions suffer from the problems mentioned in the background: the high dimensionality of the input variables and the large number of parameters make it difficult to guarantee the model's convergence and fitting effect.
[0006] The information disclosed in this background section is intended only to enhance the understanding of the overall background of this application and should not be construed as an admission or in any way implying that the information constitutes prior art known to those skilled in the art. Summary of the Invention
[0007] The technical problem to be solved by this application is to overcome the shortcomings of the prior art and provide a method and system for recommending aluminum material process technology based on knowledge distillation. Through knowledge distillation, the process technology of aluminum materials can be accurately optimized and the configuration can be recommended, thereby improving the product performance and production efficiency of aluminum materials.
[0008] To solve the above-mentioned technical problems, this application provides the following technical solution:
[0009] On the one hand, this application provides a method for recommending aluminum material processing techniques based on knowledge distillation, including the following steps:
[0010] Collect process data of aluminum materials and construct a training set; any process data in the training set includes alloy element ratio, processing information, process parameters of each process, and performance indicators of aluminum materials.
[0011] Based on the training set, a performance prediction model for aluminum materials is trained.
[0012] Based on the performance prediction model, a response matrix of performance indicators is constructed, and the sensitivity of each process parameter to each performance indicator is calculated based on the response matrix.
[0013] An attention strategy is set based on the sensitivity of each process parameter and each performance indicator;
[0014] Based on the attention strategy, knowledge distillation is performed on the performance prediction model to obtain a performance prediction engine for aluminum materials.
[0015] The performance prediction engine is used to optimize and recommend process data for aluminum materials.
[0016] As a preferred embodiment of the knowledge distillation-based aluminum material process recommendation method described in this application, the performance prediction model includes a first sub-model and a second sub-model; wherein the input of the first sub-model includes processing procedure information, process parameters of each procedure and element ratio labels, and the output is a performance modulation vector of the aluminum material;
[0017] The input to the second sub-model includes the performance modulation vector and the alloy element ratio, and the output includes each performance index of the aluminum material; the alloy element ratio includes the addition ratio of each alloy element; the processing information includes the execution order of each process.
[0018] The method for training the performance prediction model is as follows:
[0019] Calculate the element ratio label for each process data based on the alloy element ratio;
[0020] The performance prediction model is iteratively trained. In each iteration, the element ratio label, processing information, process parameters of each process, and alloy element ratio of a process data are input into the performance prediction model, and each performance index is output. The error between the performance index output by the performance prediction model and the performance index in the process data is calculated, and the model parameters of the performance prediction model are optimized based on the error.
[0021] As a preferred embodiment of the aluminum material process recommendation method based on knowledge distillation described in this application, the performance modulation vector is a vector of length N, where N is the number of performance indicators; each element in the performance modulation vector corresponds to a performance indicator, and the value of the element represents the modulation intensity of the performance indicator by the currently input element ratio label, processing information, and process parameters of each process.
[0022] For any given process data, the method for calculating the element ratio label is as follows:
[0023] The addition ratio of each alloying element in each process data is normalized and used as the proportion factor for each element.
[0024] The proportioning factor of each element is organized into a proportioning factor sequence of the corresponding process data;
[0025] Clustering is performed on the ratio factor sequence of each process data to obtain the clustering results of each process data;
[0026] The clustering results of each process data point are encoded into element ratio labels.
[0027] As a preferred embodiment of the knowledge distillation-based aluminum material process recommendation method described in this application, the response matrix of the performance index includes a first response matrix and a second response matrix.
[0028] The method for constructing the first response matrix and the second response matrix is as follows:
[0029] Initialize an M-row, N-column first response matrix and a K-row, N-column second response matrix; M is the number of process parameters, and K is the number of alloying elements in the alloying element ratio;
[0030] Input a set of element ratio labels, processing information, process parameters of each process, and alloy element ratios into the performance prediction model; extract the performance modulation vector output by the first sub-model, and extract each performance index output by the second sub-model.
[0031] Calculate the gradient of each element in the performance modulation vector with respect to each process parameter, and assign values to the elements in the first response matrix. The element in the i-th row and j-th column of the first response matrix corresponds to the gradient of the j-th element in the performance modulation vector with respect to the i-th process parameter.
[0032] Calculate the gradient of each performance index with respect to the addition ratio of each alloying element, and assign values to the elements in the second response matrix. The element in the p-th row and q-th column of the second response matrix corresponds to the gradient of the q-th performance index element with respect to the p-th alloying element.
[0033] As a preferred embodiment of the knowledge distillation-based aluminum material process recommendation method described in this application, the method for calculating the sensitivity of any process parameter is as follows: extract the row vector of the row containing any process parameter in the first response matrix; take the absolute value of each element in the row vector and calculate the mean to obtain the sensitivity of the corresponding process parameter;
[0034] The method for calculating the sensitivity of any performance indicator is as follows: extract the column vector of the column containing any performance indicator in the second response matrix; take the absolute value of each element in the column vector and calculate the mean to obtain the sensitivity of the corresponding performance indicator.
[0035] As a preferred embodiment of the knowledge distillation-based aluminum material process recommendation method described in this application, the performance prediction engine includes a first prediction unit; the attention strategy includes a first attention strategy; the first attention strategy is used to control the distillation of a first sub-model to obtain the first prediction unit; the first attention strategy specifically includes:
[0036] Calculate the attention mask for each process parameter; the attention mask for any process parameter is the ratio of the sensitivity of the process parameter to the sum of the sensitivities of all process parameters.
[0037] M input channels are set in the first prediction unit, and each input channel is used to map a process parameter into a feature matrix;
[0038] Deploy the attention mask for each process parameter into the corresponding input channel;
[0039] In the knowledge distillation process, the feature matrix of each process parameter is weighted by using an attention mask for each process parameter.
[0040] As a preferred embodiment of the knowledge distillation-based aluminum material process recommendation method described in this application, the performance prediction engine further includes a second prediction unit; the attention strategy further includes a second attention strategy; the second attention strategy is used to control the distillation of the second sub-model to obtain the second prediction unit; the second attention strategy specifically includes:
[0041] Calculate the loss weight for each performance metric; the loss weight for any performance metric is the ratio of the sensitivity of that performance metric to the sum of the sensitivities of all performance metrics.
[0042] In the knowledge distillation process, a weighted loss function is constructed using the loss weight of each performance metric;
[0043] The output error between the second prediction unit and the second sub-model is calculated based on the weighted loss function.
[0044] The model parameters of the second prediction unit are optimized based on the output error.
[0045] As a preferred embodiment of the knowledge distillation-based aluminum material process recommendation method described in this application, the step of calculating the output error between the second prediction unit and the second sub-model based on the weighted loss function specifically includes: calculating the square of the difference between any performance index output by the second prediction unit and the second sub-model to obtain the basic loss value of the corresponding performance index; multiplying the basic loss by the corresponding loss weight to obtain the weighted loss value of the corresponding performance index; and calculating the sum of the weighted loss values of each performance index to obtain the output error.
[0046] As a preferred embodiment of the knowledge distillation-based aluminum material process recommendation method described in this application, the optimization and recommendation of aluminum material process data based on the performance prediction engine specifically includes:
[0047] Set up processing configuration schemes; each processing configuration scheme includes a set of process data; calculate the element ratio labels of the corresponding processing configuration scheme based on the alloy element ratios in the process data;
[0048] The process data and element ratio labels of each processing configuration scheme are sequentially input into the performance prediction engine. The performance prediction engine calculates and outputs each performance index corresponding to each processing configuration scheme.
[0049] The optimal processing configuration is selected based on each performance index corresponding to each processing configuration scheme; the alloy element ratio, processing procedure information, and process parameters of each procedure in the optimal processing configuration scheme are recommended to the user.
[0050] Secondly, this application provides a knowledge distillation-based aluminum material process recommendation system, including a data acquisition module, a performance prediction module, a strategy module, a knowledge distillation module, and an optimization module; wherein:
[0051] The data acquisition module is used to collect process data and build a training set; the data acquisition module also calculates element ratio labels based on process data to supplement the training set.
[0052] The performance prediction module trains a performance prediction model for aluminum materials based on the training set;
[0053] The strategy module constructs a response matrix of performance indicators based on the performance prediction model, and sets an attention strategy for knowledge distillation based on the response matrix;
[0054] The knowledge distillation module performs knowledge distillation on the performance prediction model based on the attention strategy to obtain a performance prediction engine for aluminum materials.
[0055] The optimization module is used to deploy the performance prediction engine and optimize and recommend process data for aluminum materials based on the performance prediction engine.
[0056] Compared with the prior art, the beneficial effects achieved by this application are as follows:
[0057] This application constructs a performance prediction model, which can more effectively model the ability of process parameters to modulate the performance of aluminum materials and the response law of alloy element ratio to performance, thereby improving the adaptability and reliability of the performance prediction results of aluminum materials under complex processing conditions.
[0058] By constructing a hierarchical modeling structure and adopting differentiated distillation strategies for different sub-models, the high-dimensional input space can be effectively compressed without sacrificing prediction performance, thereby meeting the real-time and efficiency requirements of field-level model deployment and making it suitable for real-time industrial control scenarios with limited computing resources.
[0059] By introducing an attention strategy into the knowledge distillation process, the learning model's ability to identify highly sensitive variables in alloy element ratios and process parameters is enhanced, thereby improving the model's generalization and result interpretation capabilities. Attached Figure Description
[0060] To more clearly illustrate the technical solutions of the embodiments of this application, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. Wherein:
[0061] Figure 1A flowchart of the recommended process for aluminum material manufacturing based on knowledge distillation provided in this application;
[0062] Figure 2 A schematic diagram of the structure of the knowledge distillation-based aluminum material process recommendation system provided in this application. Detailed Implementation
[0063] The technical solution of this application will be described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the embodiments and specific features in the embodiments are detailed descriptions of the technical solution of this application, rather than limitations thereof. In the absence of conflict, the embodiments and technical features in the embodiments can be combined with each other.
[0064] Example 1
[0065] This embodiment introduces a method for recommending aluminum material processing procedures based on knowledge distillation, referring to... Figure 1 The method includes the following steps:
[0066] Collect process data of aluminum materials and construct a training set; any process data in the training set includes alloy element ratio, processing information, process parameters of each process, and performance indicators of aluminum materials.
[0067] The performance indicators include yield strength, ductility, electrical conductivity, and corrosion resistance. The alloy element ratio includes the proportion of each alloy element added. For example, alloy elements include magnesium, zinc, iron, and manganese. The proportion of each alloy element affects the performance of the final aluminum material. For example, adding magnesium can improve the strength of the aluminum alloy, but too much magnesium makes the aluminum alloy prone to fracture under stress. Adding zinc can improve the hardness of the aluminum alloy, but excessive addition will lead to a decrease in the heat resistance of the aluminum alloy, making it prone to deformation in high-temperature environments. The processing procedure information includes the execution sequence of each procedure. For example, the procedures include melting, heat preservation, casting, homogenization, sawing, extrusion, quenching, and stretching. Each procedure includes multiple configurable process parameters. The execution sequence of the procedures and the setting of the process parameters both affect the performance of the aluminum material. For example, if the cooling rate during quenching is too slow, the aluminum alloy cannot be fully dissolved, resulting in low strength; if the cooling rate is too fast, there is a risk of cracking. During the stretching process, a moderate stretching rate can eliminate internal stress and improve the dimensional stability of aluminum materials. However, if the stretching rate is too high, the plasticity of the aluminum alloy will decrease, making it prone to fracture.
[0068] Based on the training set, a performance prediction model for aluminum materials is trained.
[0069] The performance prediction model includes a first sub-model and a second sub-model; wherein, the input of the first sub-model includes processing procedure information, process parameters of each procedure and element ratio labels, and the output is a performance modulation vector of aluminum material;
[0070] The input to the second sub-model includes the performance modulation vector and the alloy element ratio, and the output includes each performance index of the aluminum material;
[0071] The method for training the performance prediction model is as follows:
[0072] Calculate the element ratio label for each process data based on the alloy element ratio;
[0073] The performance prediction model is iteratively trained. In each iteration, the element ratio label, processing information, process parameters of each process, and alloy element ratio of a process data are input into the performance prediction model, and each performance index is output. The error between the performance index output by the performance prediction model and the performance index in the process data is calculated, and the model parameters of the performance prediction model are optimized based on the error.
[0074] Specifically, in each iteration, the element ratio labels, processing information, and process parameters of each process are input into the first sub-model, and the first sub-model outputs a performance modulation vector as an intermediate variable. The performance modulation vector and the alloy element ratio are input into the second sub-model, and the second sub-model calculates and outputs each performance index. The error between the performance index output by the second sub-model and the performance index in the process data is calculated, and the gradient of this error with respect to each parameter in the first and second sub-models is calculated using the backpropagation algorithm. Then, the parameters in the first and second sub-models are updated based on the gradient descent method or its variants to minimize the error.
[0075] The performance modulation vector is a vector of length N, where N is the number of performance indicators. Each element in the performance modulation vector corresponds to a performance indicator, and the value of the element represents the modulation intensity of the performance indicator by the currently input element ratio label, processing procedure information, and process parameters of each procedure. In this embodiment, the first sub-model maps its input data to the degree of influence on each performance indicator, i.e., the modulation intensity, and encodes the above influence degree into a structured performance modulation vector. Optionally, the value range of each element in the performance modulation vector is scaled to [-1, 1], where the positive or negative sign indicates a tendency to enhance or weaken the performance indicator, and the absolute value of the element indicates the degree of enhancement or weakening. The performance modulation vector is an intermediate expression of the performance prediction model and is input into the second sub-model along with the alloy element ratio to participate in performance indicator prediction. Both the first and second sub-models are trained and optimized using the performance indicator prediction loss of the second sub-model. Optionally, the performance modulation vector can be used as a bias factor for predicting each performance indicator in the second sub-model, thereby establishing a clear supervision path and facilitating unified backpropagation training.
[0076] For any given process data, the method for calculating the element ratio label is as follows:
[0077] The addition ratio of each alloying element in each process data is normalized and used as the proportion factor for each element.
[0078] The proportioning factor of each element is organized into a proportioning factor sequence of the corresponding process data;
[0079] Clustering is performed on the ratio factor sequence of each process data to obtain the clustering results of each process data;
[0080] The clustering results of each process data point are encoded into element ratio labels.
[0081] Optionally, the proportioning factor sequence can be clustered using expert rules or unsupervised clustering. This embodiment preferably uses the K-means clustering algorithm to cluster the proportioning factor sequence for each process data point, and determines the optimal number of clusters (i.e., the number of clustering results) using the elbow rule. The element proportioning labels obtained through clustering allow alloy element proportions with similar characteristics to be grouped into the same category, simplifying the input data structure of the first sub-model and reducing the complexity of model training.
[0082] For the performance of aluminum alloy materials, the process path and alloying elements are two semantically distinct influencing factors. Process path-related factors, such as processing step information and process parameters for each step, affect material performance by controlling material evolution and deformation. The proportion of alloying elements directly determines the material composition and directly affects its various properties; however, the coupling effect of these two factors on material performance is not significant. This application reduces a large number of ineffective input coupling dimensions by splitting the performance prediction model into two sub-models to model the influence of process path and alloying elements on performance indicators separately. Compared to the traditional approach of using a large model with a unified input for multi-indicator prediction, the performance prediction model of this application is easier to train and converge. The first sub-model mainly models the influence of the process path on material performance. For the alloying element proportion, feature dimensionality reduction is performed, processing only the proportion category label to learn the degree of influence of the process path on material performance under typical proportion patterns. The second sub-model mainly models the influence of element proportion on material performance and uses the influence of the process path as a stable, structured partial input, making it easier to fit without exploring the complex influence patterns of the process path on material performance.
[0083] Based on the performance prediction model, a response matrix of performance indicators is constructed, and the sensitivity of each process parameter to each performance indicator is calculated based on the response matrix.
[0084] The response matrix of the performance index includes a first response matrix and a second response matrix;
[0085] The method for constructing the first response matrix and the second response matrix is as follows:
[0086] Initialize an M-row, N-column first response matrix and a K-row, N-column second response matrix; M is the number of process parameters, and K is the number of alloying elements in the alloying element ratio;
[0087] Input a set of element ratio labels, processing information, process parameters of each process, and alloy element ratios into the performance prediction model; extract the performance modulation vector output by the first sub-model, and extract each performance index output by the second sub-model.
[0088] Calculate the gradient of each element in the performance modulation vector with respect to each process parameter, and assign values to the elements in the first response matrix. The element in the i-th row and j-th column of the first response matrix corresponds to the gradient of the j-th element in the performance modulation vector with respect to the i-th process parameter.
[0089] Calculate the gradient of each performance index with respect to the addition ratio of each alloying element, and assign values to the elements in the second response matrix. The element in the p-th row and q-th column of the second response matrix corresponds to the gradient of the q-th performance index element with respect to the p-th alloying element.
[0090] Optionally, multiple sets of element ratio labels, processing procedure information, process parameters of each procedure, and alloy element ratios are repeatedly input into the performance prediction model to obtain multiple first response matrices and second response matrices; the average value of the elements at the same position in the multiple first response matrices is calculated, and the average value of the elements at each position forms the final first response matrix; similarly, the average value of the elements at the same position in the multiple second response matrices is calculated, and the average value of the elements at each position forms the final second response matrix.
[0091] In commonly used deep learning frameworks such as TensorFlow, automatic differentiation mechanisms can be used to calculate the gradient of each element in the performance modulation vector with respect to each process parameter, as well as the gradient of each performance index with respect to the addition ratio of each alloying element. The first response matrix quantifies the sensitivity of the modulation intensity of each performance index in the performance modulation vector to the process parameter, while the second response matrix quantifies the sensitivity of each performance index output by the second sub-model to the alloying element ratio. Through the joint construction and gradient calculation of the two response matrices, sensitivity extraction can be achieved in both the process parameter and element ratio dimensions, providing data support and a foundation for subsequent knowledge distillation and attention setting.
[0092] The method for calculating the sensitivity of any process parameter is as follows: extract the row vector of the row containing any process parameter in the first response matrix; take the absolute value of each element in the row vector and calculate the mean to obtain the sensitivity of the corresponding process parameter;
[0093] In this implementation, the sensitivity of process parameters is used to identify key parameters that play a major role in regulating performance indicators in the current process configuration. By extracting the average response intensity of each process parameter in the first response matrix, the influence of each process parameter on multiple performance indicators can be determined, providing direction for setting distillation weights and model distillation compression in the process parameter dimension.
[0094] The method for calculating the sensitivity of any performance indicator is as follows: extract the column vector of the column containing any performance indicator in the second response matrix; take the absolute value of each element in the column vector and calculate the mean to obtain the sensitivity of the corresponding performance indicator.
[0095] In this embodiment, the sensitivity of performance indicators is used to identify key performance indicators that are most sensitive to changes in elemental ratios. By extracting the response intensity of each performance indicator in the second response matrix for each alloying element, its dependence on alloy composition perturbations can be quantified, thereby guiding the distillation strategy on the model output side and ensuring the learning model's predictive ability for key performance indicators.
[0096] An attention strategy is set based on the sensitivity of each process parameter and each performance indicator;
[0097] Based on the attention strategy, knowledge distillation is performed on the performance prediction model to obtain a performance prediction engine for aluminum materials.
[0098] The performance prediction engine includes a first prediction unit and a second prediction unit; the attention strategy includes a first attention strategy and a second attention strategy; the first attention strategy is used to control the distillation of the first sub-model to obtain the first prediction unit; the first attention strategy specifically includes:
[0099] Calculate the attention mask for each process parameter; the attention mask for any process parameter is the ratio of the sensitivity of the process parameter to the sum of the sensitivities of all process parameters.
[0100] M input channels are set in the first prediction unit, and each input channel is used to map a process parameter into a feature matrix;
[0101] Deploy the attention mask for each process parameter into the corresponding input channel;
[0102] In the knowledge distillation process, the feature matrix of each process parameter is weighted by using an attention mask for each process parameter.
[0103] During the distillation process of the first sub-model, the prediction capability of the first prediction unit gradually approaches that of the first sub-model by gradually reducing the error between the performance modulation vector output by the first prediction unit and the performance modulation vector output by the first sub-model. In this process, the feature matrix of each process parameter is weighted using an attention mask, so that the feature matrix of each process parameter mapped by the first prediction unit is multiplied by the corresponding attention mask before subsequent feature extraction and performance modulation vector calculation. This allows the first prediction model to prioritize learning the influence of highly sensitive process parameters on the performance modulation vector during the distillation process, ensuring high-fidelity inheritance of the prediction capability of the first sub-model. Thus, the knowledge distillation process achieves a reasonable allocation of training resources, avoiding excessive waste of computational resources on relatively unimportant process parameters.
[0104] In this embodiment, in addition to having M input channels for processing process parameters, the first prediction unit also has an input channel for processing element ratio labels and processing procedure information.
[0105] The second attention strategy is used to control the distillation of the second sub-model to obtain the second prediction unit; the second attention strategy specifically includes:
[0106] Calculate the loss weight for each performance metric; the loss weight for any performance metric is the ratio of the sensitivity of that performance metric to the sum of the sensitivities of all performance metrics.
[0107] In the knowledge distillation process, a weighted loss function is constructed using the loss weight of each performance metric;
[0108] The output error between the second prediction unit and the second sub-model is calculated based on the weighted loss function. Specifically, this includes: squaring the difference between any performance index output by the second prediction unit and the second sub-model to obtain the basic loss value of the corresponding performance index; multiplying the basic loss value by the corresponding loss weight to obtain the weighted loss value of the corresponding performance index; and calculating the sum of the weighted loss values of each performance index to obtain the output error.
[0109] The model parameters of the second prediction unit are optimized based on the output error.
[0110] During knowledge distillation, the performance prediction model and the performance prediction engine process the same input data and both output each performance metric. The model parameters of the second prediction unit are adjusted using optimization algorithms such as gradient descent to minimize the output error, allowing the prediction accuracy of the second prediction unit to gradually approach that of the second sub-model, thus completing the knowledge distillation of the second sub-model.
[0111] By controlling the second attention strategy, the base loss values of performance metrics are weighted according to their loss weights. Performance metrics with higher sensitivity will account for a larger proportion of the output error, ensuring that the second prediction unit prioritizes optimizing the prediction of high-sensitivity performance metrics during the distillation process, reducing overfitting of low-sensitivity metrics. Process parameters and performance metrics are both high-dimensional features; this embodiment uses an attention strategy to achieve feature dimensionality reduction in the distillation process, simplifying the model complexity of the performance prediction engine, improving training efficiency, and ensuring the model's stability and interpretability.
[0112] The performance prediction engine is a lightweight prediction model derived from the performance prediction model through knowledge distillation. Through knowledge distillation, the performance prediction model is compressed into a lightweight performance prediction engine for deployment in industrial settings and for large-scale, efficient computation. In an aluminum alloy processing industrial setting, the performance prediction engine rapidly calculates multiple performance parameters for each process configuration to compare the advantages and disadvantages of different options, thereby finding the optimal solution from a vast space of feasible solutions. Optionally, the performance prediction model can be any of a Transformer encoder, a residual network model, or a deep feedforward network model, and the performance prediction engine can be any of a sparse attention network model or a lightweight residual network model.
[0113] The performance prediction engine optimizes and recommends process data for aluminum materials; specifically including:
[0114] Set up processing configuration schemes; each processing configuration scheme includes a set of process data; calculate the element ratio labels of the corresponding processing configuration scheme based on the alloy element ratios in the process data;
[0115] The process data and element ratio labels of each processing configuration scheme are sequentially input into the performance prediction engine. The performance prediction engine calculates and outputs each performance index corresponding to each processing configuration scheme.
[0116] The optimal processing configuration is selected based on each performance index corresponding to each processing configuration scheme; the alloy element ratio, processing procedure information, and process parameters of each procedure in the optimal processing configuration scheme are recommended to the user.
[0117] Optionally, the method for selecting the optimal processing configuration scheme is as follows: assign a weight coefficient to each performance indicator; normalize and dimensionlessly process each performance indicator; perform a weighted summation of each performance indicator based on the weight coefficient to obtain the optimization objective value of each processing configuration scheme; and select the processing configuration scheme with the largest optimization objective value as the optimal processing configuration scheme. Preferably, the setting of the processing configuration scheme and the selection of the optimal processing configuration scheme can be implemented based on optimization algorithms such as genetic algorithms and constrained particle swarm optimization.
[0118] Example 2
[0119] This embodiment is the second embodiment of this application; it is based on the same inventive concept as Embodiment 1, and refers to... Figure 2 This embodiment introduces a knowledge distillation-based aluminum material process recommendation system, including a data acquisition module, a performance prediction module, a strategy module, a knowledge distillation module, and an optimization module; wherein:
[0120] The data acquisition module collects aluminum processing data and constructs a training set. It also supplements the training set by calculating element ratio labels based on the aluminum processing data. Each data point includes alloy element ratios, processing steps, process parameters for each step, and performance indicators of the aluminum alloy product. This module calculates element ratio labels based on the addition ratio of each alloy element in each aluminum processing data point.
[0121] The performance prediction module trains a performance prediction model for aluminum alloy products based on the training set. The performance prediction model includes a first sub-model and a second sub-model. The first sub-model takes processing information, process parameters for each process, and element ratio labels as inputs and outputs a performance modulation vector. The second sub-model takes the performance modulation vector and alloy element ratios as inputs and outputs various performance indicators.
[0122] The strategy module constructs a response matrix of performance indicators based on the performance prediction model, and sets an attention strategy for knowledge distillation based on the response matrix; the response matrix includes a first response matrix and a second response matrix; the strategy module calculates the sensitivity of each process parameter based on the first response matrix, and sets a first attention strategy based on the sensitivity of each process parameter; the strategy module also calculates the sensitivity of each performance indicator based on the second response matrix, and sets a second attention strategy based on the sensitivity of each performance indicator.
[0123] The knowledge distillation module performs knowledge distillation on the performance prediction model based on the attention strategy to obtain a performance prediction engine for aluminum materials. The performance prediction engine includes a first prediction unit and a second prediction unit. The first prediction unit, under the control of the first attention strategy, prioritizes learning the influence of highly sensitive process parameters in the first sub-model. The second prediction unit, under the control of the second attention strategy, prioritizes learning the prediction of highly sensitive performance indicators by the second sub-model.
[0124] The optimization module is used to deploy the performance prediction engine and optimize and recommend process data for aluminum materials based on the performance prediction engine. The optimization module is configured with an optimization algorithm that selects the optimal processing configuration by setting multiple processing configuration schemes and calculating each performance index of each processing configuration scheme based on the performance prediction engine, thereby optimizing the alloy element ratios and process parameters for aluminum processing.
[0125] The specific functions of each module described above are explained in the relevant content of the knowledge-based distillation aluminum material process recommendation method described in Example 1, and will not be repeated here.
[0126] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.
[0127] The embodiments of this application have been described above with reference to the accompanying drawings. However, this application is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of this application without departing from the spirit and scope of protection of this application, and these forms are all within the protection scope of this application.
Claims
1. A method for recommending aluminum material processing techniques based on knowledge distillation, characterized in that: Includes the following steps: Collect process data of aluminum materials and construct a training set; any process data in the training set includes alloy element ratio, processing information, process parameters of each process, and performance indicators of aluminum materials. Based on the training set, a performance prediction model for aluminum materials is trained. The performance prediction model includes a first sub-model and a second sub-model; wherein, the input of the first sub-model includes processing procedure information, process parameters of each procedure and element ratio labels, and the output is a performance modulation vector of aluminum material; The input to the second sub-model includes the performance modulation vector and the alloy element ratio, and the output includes each performance index of the aluminum material; the alloy element ratio includes the addition ratio of each alloy element; the processing information includes the execution order of each process. The performance modulation vector is a vector of length N, where N is the number of performance indicators. Each element in the performance modulation vector corresponds to a performance indicator, and the value of the element represents the modulation intensity of the performance indicator by the currently input element ratio label, processing procedure information, and process parameters of each procedure. For any given process data, the method for calculating the element ratio label is as follows: The addition ratio of each alloying element in each process data is normalized and used as the proportion factor for each element. The proportioning factor of each element is organized into a proportioning factor sequence of the corresponding process data; Clustering is performed on the ratio factor sequence of each process data to obtain the clustering results of each process data; The clustering results of each process data point are encoded into element ratio labels; Based on the performance prediction model, a response matrix of performance indicators is constructed, and the sensitivity of each process parameter to each performance indicator is calculated based on the response matrix. An attention strategy is set based on the sensitivity of each process parameter and each performance indicator; Based on the attention strategy, knowledge distillation is performed on the performance prediction model to obtain a performance prediction engine for aluminum materials. The performance prediction engine is used to optimize and recommend process data for aluminum materials.
2. The aluminum material processing recommendation method based on knowledge distillation as described in claim 1, characterized in that: The method for training the performance prediction model is as follows: Calculate the element ratio label for each process data based on the alloy element ratio; The performance prediction model is iteratively trained. In each iteration, the element ratio label, processing information, process parameters of each process, and alloy element ratio of a process data are input into the performance prediction model, and each performance index is output. The error between the performance index output by the performance prediction model and the performance index in the process data is calculated, and the model parameters of the performance prediction model are optimized based on the error.
3. The aluminum material processing recommendation method based on knowledge distillation as described in claim 2, characterized in that: The response matrix of the performance index includes a first response matrix and a second response matrix; The method for constructing the first response matrix and the second response matrix is as follows: Initialize an M-row, N-column first response matrix and a K-row, N-column second response matrix; M is the number of process parameters, and K is the number of alloying elements in the alloying element ratio; Input a set of element ratio labels, processing information, process parameters of each process, and alloy element ratios into the performance prediction model; extract the performance modulation vector output by the first sub-model, and extract each performance index output by the second sub-model. Calculate the gradient of each element in the performance modulation vector with respect to each process parameter, and assign values to the elements in the first response matrix. The element in the i-th row and j-th column of the first response matrix corresponds to the gradient of the j-th element in the performance modulation vector with respect to the i-th process parameter. Calculate the gradient of each performance index with respect to the addition ratio of each alloying element, and assign values to the elements in the second response matrix. The element in the p-th row and q-th column of the second response matrix corresponds to the gradient of the q-th performance index element with respect to the p-th alloying element.
4. The aluminum material processing recommendation method based on knowledge distillation as described in claim 3, characterized in that: The method for calculating the sensitivity of any process parameter is as follows: extract the row vector of the row containing any process parameter in the first response matrix; take the absolute value of each element in the row vector and calculate the mean to obtain the sensitivity of the corresponding process parameter; The method for calculating the sensitivity of any performance indicator is as follows: extract the column vector of the column containing any performance indicator in the second response matrix; take the absolute value of each element in the column vector and calculate the mean to obtain the sensitivity of the corresponding performance indicator.
5. The aluminum material processing recommendation method based on knowledge distillation as described in claim 4, characterized in that: The performance prediction engine includes a first prediction unit; the attention strategy includes a first attention strategy. The first attention strategy is used to control the distillation of the first sub-model to obtain the first prediction unit; The first attention strategy specifically includes: Calculate the attention mask for each process parameter; the attention mask for any process parameter is the ratio of the sensitivity of the process parameter to the sum of the sensitivities of all process parameters. M input channels are set in the first prediction unit, and each input channel is used to map a process parameter into a feature matrix; Deploy the attention mask for each process parameter into the corresponding input channel; In the knowledge distillation process, the feature matrix of each process parameter is weighted by using an attention mask for each process parameter.
6. The aluminum material processing recommendation method based on knowledge distillation as described in claim 5, characterized in that: The performance prediction engine further includes a second prediction unit; the attention strategy further includes a second attention strategy; The second attention strategy is used to control the distillation of the second sub-model, thereby obtaining the second prediction unit; The second attention strategy specifically includes: Calculate the loss weight for each performance metric; the loss weight for any performance metric is the ratio of the sensitivity of that performance metric to the sum of the sensitivities of all performance metrics. In the knowledge distillation process, a weighted loss function is constructed using the loss weight of each performance metric; The output error between the second prediction unit and the second sub-model is calculated based on the weighted loss function. The model parameters of the second prediction unit are optimized based on the output error.
7. The aluminum material processing recommendation method based on knowledge distillation as described in claim 6, characterized in that: The step of calculating the output error between the second prediction unit and the second sub-model based on the weighted loss function specifically includes: calculating the square of the difference between any performance index output by the second prediction unit and the second sub-model to obtain the basic loss value of the corresponding performance index; multiplying the basic loss by the corresponding loss weight to obtain the weighted loss value of the corresponding performance index; and calculating the sum of the weighted loss values of each performance index to obtain the output error.
8. The aluminum material processing recommendation method based on knowledge distillation as described in claim 7, characterized in that: Based on the performance prediction engine, the process data for aluminum materials is optimized and recommended, specifically including: Set up processing configuration schemes; each processing configuration scheme includes a set of process data; calculate the element ratio labels of the corresponding processing configuration scheme based on the alloy element ratios in the process data; The process data and element ratio labels of each processing configuration scheme are sequentially input into the performance prediction engine. The performance prediction engine calculates and outputs each performance index corresponding to each processing configuration scheme. The optimal processing configuration is selected based on each performance index corresponding to each processing configuration scheme; the alloy element ratio, processing procedure information, and process parameters of each procedure in the optimal processing configuration scheme are recommended to the user.
9. A knowledge-based distillation-based aluminum material process recommendation system, used to implement the knowledge-based distillation-based aluminum material process recommendation method according to any one of claims 1-8, characterized in that: It includes a data acquisition module, a performance prediction module, a strategy module, a knowledge distillation module, and an optimization module; among which: The data acquisition module is used to collect process data and build a training set; the data acquisition module also calculates element ratio labels based on process data to supplement the training set. The performance prediction module trains a performance prediction model for aluminum materials based on the training set; The strategy module constructs a response matrix of performance indicators based on the performance prediction model, and sets an attention strategy for knowledge distillation based on the response matrix; The knowledge distillation module performs knowledge distillation on the performance prediction model based on the attention strategy to obtain a performance prediction engine for aluminum materials. The optimization module is used to deploy the performance prediction engine and optimize and recommend process data for aluminum materials based on the performance prediction engine.
Citation Information
Patent Citations
Method and system for optimizing aluminum alloy laser welding process parameters based on machine learning
CN118964992A
System and method for optimizing and improving performance of manufacturing and processing material based on aluminum alloy parts
CN120562114A