Method and equipment for predicting hardenability of metal material

Abstract

The invention relates to the field of material performance prediction, in particular to a metal material hardenability prediction method and equipment. The method comprises the following steps: each client obtains a respective standard hardenability data set; each client constructs a respective MRAN-hard model, and sends the respective MRAN-hard model and the standard hardenability data set to the server; the server optimizes the MRAN-hard model according to the received data, and sends an optimization parameter to a corresponding client; the client obtains an optimal loss value and an optimal parameter of the model according to the received parameters, and sends the optimal loss value and the optimal parameter to the server; and the server carries out adaptive global model aggregation and lightweight transformation on the optimal loss value and the optimal parameter of the model of each client, the transformed model is sent to each client, and each client carries out metal material hardenability prediction according to the transformed model. In this way, an efficient, intelligent and engineering-oriented metal material hardenability prediction system can be constructed.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of material performance prediction, and more particularly, to a method and device for predicting the hardenability of a metal material. BACKGROUND

[0002] The hardenability of a metal material is a key indicator for evaluating the safety of a material in service. Traditional Jominy end quenching tests have problems such as long detection cycle, high cost, and low efficiency. Machine learning prediction methods based on data-driven have become an efficient alternative. However, in industrial practical applications, there is a serious data island phenomenon among metal enterprises, and a single enterprise cannot obtain sufficient and diversified training data, which leads to the inability to train a high-precision and high-generalization hardenability prediction model.

[0003] Currently, transverse federated learning can realize collaborative training of a global regression model by multiple clients without sharing the original data of each participant, and has become the mainstream technical path to solve this problem. However, the existing transverse federated learning method still has many technical bottlenecks in the prediction of the hardenability of metal materials, including: (1) The hyperparameters are highly dependent on human experience and cannot be dynamically adjusted according to the real-time data distribution and model convergence state of the client, which easily leads to overfitting or underfitting and reduces the prediction accuracy; (2) The feature engineering capability is weak, relying only on manually set chemical composition features without deeply integrating material field knowledge, which has problems such as feature redundancy, insufficient low-order feature mining, and the like, and the federated standardization is not optimized in combination with material characteristics, further limiting the model performance; (3) The heterogeneous data adaptation capability is poor, and there is a lack of client-level intelligent scheduling, and the participation of low-quality and slow-convergence clients in global aggregation will lower the overall model effect, and the unified training strategy cannot adapt to the heterogeneous differences of different enterprise data; (4) The existing solutions are disconnected from the material research and development process, have strong operation professionalism, and only output a single prediction value, which cannot form an executable research and development decision-making suggestion. SUMMARY

[0004] According to the present application, a metal material hardenability prediction scheme is provided. This scheme can construct an efficient, intelligent, and engineering-landable metal material hardenability prediction system.

[0005] In a first aspect of the invention, a method for predicting the hardenability of metallic materials is provided, applied to a client. The method includes: each client acquiring a corresponding initial hardenability dataset through a data interface, and sequentially preprocessing and federated normalization the initial hardenability dataset to obtain a standard hardenability dataset for each client. Each client constructs its own MRAN-hard model and sends its standard hardenability dataset and MRAN-hard model to a server. In response to the model optimization parameters of each client, the MRAN-hard model of the corresponding client is adjusted based on the model optimization parameters of each client; the adjusted MRAN-hard model of the corresponding client is trained according to the standard hardenability dataset of each client; wherein the model optimization parameters of each client are obtained by the server optimizing the MRAN-hard model of the corresponding client based on the standard hardenability dataset of each client. The standard hardenability dataset of each client is input into the trained MRAN-hard model of the corresponding client to obtain the optimal model loss value and optimal model parameters of each client; and the optimal model loss value and optimal model parameters of each client are sent to the server. In response to the lightweight aggregation model, each client performs hardenability prediction of metallic materials based on the lightweight aggregation model; wherein, the lightweight aggregation model is obtained by lightweighting the aggregated adaptive global model; the aggregated adaptive global model is obtained by the server aggregating the adaptive global model based on the optimal loss value and optimal parameters of each client's model.

[0006] In a second aspect of the invention, a method for predicting the hardenability of metallic materials is provided, applied to a server. The method includes: responding to a standard hardenability dataset and an MRAN-hard model for each client; optimizing the MRAN-hard model for the corresponding client based on the standard hardenability dataset for each client to obtain model optimization parameters for each client; and sending the model optimization parameters for each client to the corresponding client; wherein the standard hardenability dataset for each client is obtained by each client through a data interface after acquiring the corresponding initial hardenability dataset, and then sequentially performing preprocessing and federated standardization on the initial hardenability dataset. In response to the optimal model loss and optimal model parameters of each client, adaptive global model aggregation is performed based on the optimal model loss and optimal model parameters of each client to obtain an aggregated model. The aggregated model is then lightweighted to obtain a lightweight aggregated model, which is sent to the clients so that each client can predict the hardenability of metallic materials based on the lightweight aggregated model. Specifically, the optimal model loss and optimal model parameters of each client are obtained by inputting the standard hardenability dataset of each client into the corresponding trained MRAN-hard model. The trained MRAN-hard model is obtained by training the corresponding adjusted MRAN-hard model based on the standard hardenability dataset of each client. The adjusted MRAN-hard model is obtained by each client adjusting its corresponding MRAN-hard model according to the model optimization parameters.

[0007] In a third aspect of the invention, an electronic device is provided. The electronic device includes at least one processor; and a memory communicatively connected to said at least one processor; the memory stores instructions executable by said at least one processor to enable said at least one processor to perform the method of the first aspect of the invention.

[0008] Compared with the prior art, the present invention has the following beneficial technical effects: (1) It adopts a federated learning architecture, combined with data preprocessing and federated standardization processing technology, to realize secure collaborative modeling of multi-source heterogeneous hardenability data. The original data does not need to leave the local client, effectively avoiding the risk of data leakage. At the same time, it breaks the data silo and improves the adaptability of the model to multiple scenarios and batches of metal materials. (2) Through the MRAN-hard dedicated model, it accurately fits the complex nonlinear relationship between hardenability of metal materials and factors such as composition and process. Compared with the general model, it significantly improves the hardenability prediction accuracy and model training convergence efficiency, and reduces redundant computational overhead. (3) The server adopts adaptive global model aggregation technology, dynamically allocates aggregation weights according to the optimal loss value and optimal parameters of each client model, avoids the accuracy loss caused by simple average aggregation, and makes the aggregation model take into account both universality and local optima. The overall prediction performance is better than that of a single client independent modeling. (4) The aggregation model is lightweighted. Under the premise of ensuring that the prediction accuracy remains basically unchanged, the model size and computational requirements are reduced, and the model can be deployed in engineering on ordinary clients, industrial control computers and other equipment to meet the real-time and low-latency prediction requirements of industrial sites. (5) Relying on the closed-loop iteration mechanism of end-cloud collaboration, the model can be dynamically updated as new material data accumulates, without the need to retrain from scratch, thus reducing long-term usage costs. At the same time, through standardized data interface and modular architecture, the system can be improved to enhance compatibility with existing equipment and systems, making it easier to extend to other metal material performance prediction scenarios and build an efficient, intelligent, engineering-applicable and sustainably evolving hardenability prediction system. Attached Figure Description

[0009] Figure 1 A flowchart of a method for predicting the hardenability of metallic materials applied to a client, according to an embodiment of the present invention, is shown. Figure 2 A flowchart of a method for predicting the hardenability of metallic materials applied to a server, according to an embodiment of the present invention, is shown. Detailed Implementation

[0010] In this invention, each client constructs its own MRAN-hard (Minimal Resource Allocation Network with Hard Constraints) model and sends its MRAN-hard model and collected dataset to the server. The server optimizes the MRAN-hard model based on the received data and sends the optimized parameters to the corresponding client. The client obtains the optimal loss value and optimal parameters of the model based on the received parameters and sends them to the server. The server performs lightweight modifications to the MRAN-hard model based on the received parameters and sends the modified model to each client. Each client then uses the modified model to predict the hardenability of metallic materials. In this way, an efficient, intelligent, engineering-applicable, and sustainably evolving hardenability prediction system can be constructed. Example 1

[0011] Figure 1 A flowchart illustrating a method for predicting the hardenability of metallic materials applied to a client-side application, according to an embodiment of the present invention, is shown. The method includes steps S101-S105: S101. Each client obtains the corresponding initial hardenability dataset through the data interface, and performs preprocessing and federated standardization on the initial hardenability dataset to obtain the standard hardenability dataset for each client. In this embodiment, the preprocessing includes: data format uniformity verification (converting non-numerical features into numerical features), feature column consistency verification (ensuring that the feature dimensions of each client are consistent with the definition), removal of obvious abnormal data (removing outliers caused by detection errors based on the 3σ principle, with a removal ratio of <1%), and reasonable completion of missing values ​​(for a small number of missing process parameters, the average of the same batch of data is used for completion; for chemical composition features, missing data is directly removed, and the final number of valid data for each client is between 800 and 1200).

[0012] In this embodiment, the federal standardization process includes S201~S202: S201. Based on the preprocessed dataset of each client's initial hardenability dataset, calculate the total number of global data points. Then, calculate the global mean and global variance based on the preprocessed dataset and the total number of global data points. Specifically, the total number of global data points refers to the sum of the number of data points in the initial hardenability dataset obtained by each client. The formula for calculating the global mean is: The formula for calculating the global variance is: ,in, This represents the total number of global data. This represents the total number of clients. The dataset after preprocessing the initial hardenability dataset for the client; This is the global mean. This represents the total number of clients. For each data feature vector; For the first Sample indexes in a client dataset; This represents the global variance.

[0013] S202. Construct the federated normalization formula. Input the preprocessed initial hardenability dataset, global mean, and global variance of each client into the federated normalization formula to obtain the standard hardenability dataset for each client. Specifically, the federated normalization formula is as follows: ,in, A standard hardenability dataset for each client; For each data feature vector; This is the global mean. The global variance; To avoid division by zero errors caused by a square root of 0, set to a minimum positive value. =10, [the value less than 10 in the global standard deviation] Replace the value with In this embodiment, after constructing the federalized formula, the process further includes optimizing the federalized formula. The specific optimized formula is as follows: ,in The domain weights are obtained by normalizing the physical fusion score.

[0014] Through the aforementioned federated standardization process, global feature normalization can be achieved without the original hardened data remaining locally on each client, effectively protecting data privacy and meeting data security compliance requirements. Simultaneously, it eliminates differences in data dimensions and distribution across different clients, providing a unified feature foundation for subsequent federated modeling and improving model convergence speed, prediction accuracy, and generalization ability. This is further enhanced by introducing constants. This avoids division-by-zero anomalies caused by zero variance, improves the stability of the algorithm, and this basic standardization method can provide support for subsequent weight optimization combined with the physical domain, making feature processing more in line with the characteristics of material hardenability data.

[0015] In this embodiment, to improve model generalization, after obtaining the standard hardened dataset for each client, the method further includes: constructing a label smoothing formula, using the label smoothing formula to smooth the standard hardened dataset for each client to avoid model overconfidence, and using the smoothed standard hardened dataset as the final standard hardened dataset for each client. The label smoothing calculation formula is as follows: ,in, The label smoothing coefficient. =0.01; The label mean; Smooth the label results; This is standard hardenability dataset data.

[0016] S102. Each client builds its own MRAN-hard model and sends its standard hardenability dataset and MRAN-hard model to the server.

[0017] In this embodiment, the MRAN-hard model module, serving as the core global / local prediction model, integrates an efficient channel attention mechanism, a pre-activated residual structure, a multi-dimensional regularization strategy, and a customized compatibility-oriented activation function to construct an end-to-end deep learning regression model. By effectively capturing the complex nonlinear distribution of hardenability performance characteristics, the MRAN-hard model can achieve high-precision and stable prediction of hardenability material properties. Specifically, the overall architecture of the MRAN-hard model consists of six core functional modules connected in a hierarchical manner: a GELU (Gaussian Error Linear Unit) module, an ECA (Efficient Channel Attention) layer, Gaussian noise + gradient noise, a pre-activated residual block, a feature aggregation layer, and a regression network. These modules work collaboratively to form a complete prediction chain: "feature preprocessing → deep feature extraction → feature aggregation → regression output". The activation function of the GELU module is as follows: ,in This is the output of the Gaussian error linear unit activation function; These are the input features of neurons; is the error function; 0.5 is the output scaling factor; 2 is the input normalization factor. In addition, the model includes an evaluation and early stopping module, which uses the coefficient of determination R² and root mean square error RMSE as the core metric to achieve global / local early stopping.

[0018] S103. In response to the model optimization parameters of each client, adjust the MRAN-hard model of the corresponding client based on the model optimization parameters of each client; train the adjusted MRAN-hard model of the corresponding client according to the standard hardenability dataset of each client; wherein, the model optimization parameters of each client are obtained by the server optimizing the MRAN-hard model of the corresponding client according to the standard hardenability dataset of each client.

[0019] In this embodiment, the optimization of the MRAN-hard model for the corresponding client based on the standard hardenability dataset for each client includes steps S301-S302: S301. Constructing the Intelligent Agent Model. Specifically, the intelligent agent model is deployed on the central server, serving as the system's "intelligent brain." It consists of one central meta-agent and five dedicated sub-agents, forming a two-tiered intelligent agent architecture of "centralized coordination and distributed execution." All agents are built upon a large model fine-tuned based on knowledge from the metallurgical / materials science domain, possessing domain knowledge reasoning, autonomous decision-making, real-time perception, and interaction capabilities. The agents exchange status information and decision commands in real time through standardized communication protocols. The central meta-agent continuously optimizes its collaborative strategies through reinforcement learning, enabling the entire intelligent agent layer to possess self-learning and adaptive capabilities.

[0020] In this embodiment, the agent model includes the following five agents: (1) The central meta-agent is used to calculate the agent reward value. Specifically, the central meta-agent acts as the "decision coordination center" of the system. Based on the multi-agent reinforcement learning framework, it perceives the global state of federated learning in real time, dynamically coordinates the optimization strategies of each sub-agent, and achieves globally optimal collaborative decision-making. In addition, the central meta-agent is the core coordinator of the entire agent layer; it collects the local state of each sub-agent in real time (such as the current value of hyperparameters, feature generation status, client scheduling strategy, model distillation progress, etc.) and the global state of federated learning (global rounds, average loss, convergence speed, evaluation indicators uploaded by each client, etc.). Through the multi-agent reinforcement learning framework, it dynamically generates collaborative instructions to guide each sub-agent to adjust its decision-making strategy in order to avoid decision conflicts and improve global optimization efficiency.

[0021] In this embodiment, the formula for calculating the agent's reward value is as follows: .in, The reward value for the agent is a composite reward that takes into account the improvement of model accuracy, the acceleration of convergence, and the physical consistency. The first weighting coefficient for reward value; This represents the increase in R² for this round, where R² is the coefficient of determination. This is the second weighting coefficient for the reward value; The number of cumulative cycles required to achieve the current level of accuracy (the smaller the better); This is the third weighting coefficient for the reward value; Predict the number of times the physical laws are violated for this round of modeling (calculated based on the physical constraint checker); This is the fourth weighting coefficient for the reward value; A penalty term to encourage agents to explore diverse strategies.

[0022] (2) A hyperparameter optimization agent is used to dynamically adjust the data of the standard hardenability dataset for each client based on the standard hardenability dataset of each client, the agent's reward value, and the model parameters of the MRAN-hard model. Specifically, the hyperparameter optimization agent dynamically adjusts the hyperparameters of the federated learning and MRAN-hard models under the coordination of the central meta-agent.

[0023] In this embodiment, based on the standard hardened dataset of each client, the agent's reward value, and the model parameters of the MRAN-hard model, the data of the standard hardened dataset of each client is dynamically adjusted. This includes: under the coordination of the central meta-agent, using the federated learning core parameter set (client aggregation rounds, learning rate, proximal term penalty coefficient, etc.) and the MRAN-hard model parameter set (network layer dimension, attention coefficient, regularization parameters, etc.) as the optimization space, and integrating reinforcement learning and materials science cognitive reasoning to construct an optimization strategy, thereby achieving real-time dynamic adaptive adjustment of hyperparameters. Specifically, the formula for dynamic adaptive adjustment is: ,in, for; To dynamically and adaptively adjust the first weight coefficient; The coefficient of determination; To dynamically and adaptively adjust the second weighting coefficient; Let t be the convergence speed of the model. To dynamically and adaptively adjust the third weighting coefficient; Let be the standard deviation of the training loss over t rounds. The agent iteratively updates the hyperparameters by maximizing the reward function. To dynamically and adaptively adjust the fourth weighting coefficient; This is an additional reward given by the central meta-agent based on current collaborative needs.

[0024] (3) A feature engineering augmentation agent is used to generate phase transition features based on the data output by the hyperparameter optimization agent; a fusion score is calculated based on the phase transition features, and phase transition features with a fusion score greater than the fusion score threshold are used as the output of the feature engineering augmentation agent; the fusion score threshold is a preset fixed value of 0.25. Among them, the feature engineering augmentation agent dynamically generates and selects features based on material physics knowledge under the coordination of the central meta-agent.

[0025] In this embodiment, the feature engineering enhancement agent acts as the global feature engineer in federated learning, possessing the ability to generate dynamic features driven by physical information. The agent has a built-in component-process phase diagram knowledge base, which allows it to perceive the component distribution and process parameter ranges of each client in real time. When it detects that the content of a certain element or a process parameter crosses a physical critical point, it automatically triggers the physical feature generator to generate the corresponding phase transition feature; when the content of a certain element or a process parameter does not cross a physical critical point, the original data is retained, and the enhancement operation of the feature engineering enhancement agent is no longer executed.

[0026] As some optional implementation methods in this embodiment, examples of physical feature generation rules are as follows: , ,in, This is the temperature at which the martensitic phase transformation begins. This represents the mass fraction of carbon. This represents the mass fraction of manganese. This represents the mass fraction of chromium. This represents the mass fraction of nickel. This represents the mass fraction of molybdenum. Carbon equivalent; This represents the mass fraction of vanadium. This represents the mass fraction of copper.

[0027] In this embodiment, the formula for calculating the fusion score is: , in, For integration scoring; The first fusion coefficient, ∈[0,1]; For feature importance based on random forest (normalized to [0,1]); The second fusion coefficient, ∈[0,1]; The domain correlation between feature f and hardenability index (based on the materials science knowledge base, normalized to [0,1]); The third fusion coefficient, ∈[0,1]; The physical significance score for a feature is calculated based on the frequency of its occurrence in the physical formula and its theoretical correlation strength with the hardenability index. Furthermore, .

[0028] (4) A heterogeneous data scheduling agent is used to calculate the comprehensive client quality score based on the output data of the feature engineering enhancement agent, and to take the data with a comprehensive client quality score greater than the quality score threshold (0.5) as the output data of the heterogeneous data scheduling agent; the quality score threshold is a preset fixed value. Among them, the heterogeneous data scheduling agent evaluates the client quality and implements differentiated scheduling under the coordination of the central meta-agent.

[0029] In this embodiment, the heterogeneous data scheduling agent acts as a global scheduling center, sensing the core metrics (data scale) of each client in real time. Data distribution similarity Local training loss Convergence Cycle This approach dynamically grades and differentiates clients, enabling client quality grading, differentiated aggregation weight allocation, and differentiated local training configuration, while also establishing an early warning mechanism for abnormal clients. Furthermore, semantic alignment is introduced on this basis. A four-dimensional evaluation system is constructed. Among these, semantic alignment is... The calculation method is as follows: The client's local feature names are mapped to the federated knowledge graph maintained by the central server; the semantic distance with standard materials science terminology entities is calculated; and this distance is used as the semantic alignment score, normalized to... Furthermore, regarding semantic alignment... Below the threshold For clients with a resolution of (0.8), the scheduling agent triggers the semantic correction module, which automatically generates feature alignment rules (such as temperature unit conversion, measurement location compensation formulas, etc.) based on the large model. These rules are then used in federated aggregation only after semantic correction is complete, or their aggregation weight is reduced before correction is finished. Specifically, the comprehensive client quality scoring formula is: ,in, The quality score for the kth client (normalized to [0,1]); The first indicator weight; For data scale; The total number of data; As the weight of the second indicator; The distribution similarity between client k data and global data is calculated based on MMD (Maximum Mean Discrepancy) distance and normalized to [0,1]; As the third indicator weight; This represents the greatest loss during local training. This is due to local training losses; As the weight of the fourth indicator; The minimum number of convergence rounds for all clients; For local training rounds, according to Divide the client into high ( ≥ ),middle( ≤ < ),Low( < There are three levels, and differentiated aggregation weight coefficients are assigned to clients of different levels. Local training rounds , The threshold for high-level client quality scoring. This is the lower limit threshold for the quality score of medium-level clients; The weight of the fifth indicator; This represents semantic alignment. Among them, the index weight... satisfy .

[0030] (5) Model fusion lightweight agent, used to optimize the MRAN-hard model based on the output data of the scheduling agent according to the heterogeneous data. Among them, the model fusion lightweight agent realizes causal knowledge distillation and lightweight transformation under the coordination of the central meta-agent.

[0031] In this embodiment, the lightweight agent, coordinated by the central meta-agent, achieves three core functions: knowledge fusion and transformation of the MRAN-hard model, distillation of the large model's inference capabilities, and hybrid prediction of the large model and MRAN-hard model, thereby optimizing the MRAN-hard model. Through a causal knowledge distillation framework guided by physical metallurgical mechanisms, the lightweight model not only fits the prediction output of the large model but also learns its internal causal decision-making paths, ensuring that the model maintains physical plausibility even in data-sparse and extrapolation scenarios. Specifically, the optimization process of the MRAN-hard model includes S401~S404: S401. Based on the output data of the heterogeneous data scheduling agent, causal path extraction is performed to obtain the causal path vector. Specifically, through large-model causal path extraction (causal path vector calculation), a causal path extractor is deployed in the large model; based on the structural causal model, a lightweight approximation of the formula is adjusted through a backdoor to calculate the causal effect weights of each input feature on the prediction result, ultimately forming the causal path vector of the large model. The specific calculation formula is as follows: ;in, For the causal path vector of the large model, The weights for the causal effects corresponding to the first input feature. The weights for the causal effects corresponding to the second input feature. For the first Causal effect weights corresponding to each input feature The input feature dimension.

[0032] S402. The causal awareness attention module is used to perform prior processing on the output data of the heterogeneous data scheduling agent to obtain prior data; a causal distillation loss function is constructed; and consistency constraints are applied to the prior data based on the causal distillation loss function to obtain a lightweight feature vector. In this embodiment, a causal awareness attention module is deployed in the lightweight MRAN-hard model, injecting causal prior weights on top of the ECA channel attention to guide the model to focus on key physical features in the early stages of training, thus contributing to the feature importance distribution. Provides structural support for learning and enhances... The physical plausibility of the causal perception attention module is explained by the following formula: ,in, These are causal prior weights pre-generated based on knowledge in the materials field. For causal perception attention weights; Use the Sigmoid activation function; Input feature vectors into the model; This is a one-dimensional convolution operation.

[0033] Specifically, consistency constraints can enforce the feature importance distribution in lightweight MRAN-hard models. The large model causal path vector obtained in approximation step S401 Specifically, the causal distillation loss function is: ： ,in, for Loss function used to fit the predicted output of a lightweight model Compared with the true value of hardenability of metallic materials Deviation; KL divergence is used to measure the distribution of feature importance in lightweight models. With large model causal path vector Differences; The causal consistency coefficient is dynamically adjusted by the central meta-agent according to the current federated learning stage, and is used to balance the weights of prediction fitting and causal consistency constraints. The feature importance distribution learned by the lightweight MRAN-hard model; The loss function is the causal distillation function. This represents the true value of hardenability for metallic materials. The hardenability prediction value output by the lightweight MRAN-hard model; This is the causal path vector.

[0034] S403. Calculate the causal confidence of the large model based on the causal path vector; calculate the causal confidence of the lightweight model based on the lightweight feature vector; construct the model fusion weights based on the causal confidence of the large model and the causal confidence of the lightweight model.

[0035] In this embodiment, the formula for calculating the causal confidence score of the large model is as follows: The formula for calculating the causal confidence score of a lightweight model is: ;in, For large-scale causal confidence; This represents the causal path vector; This serves as a standard causal template in the field of materials science. The denominator should be a very small positive number to avoid being zero. The feature importance distribution for lightweight models; This refers to the causal confidence score of the lightweight model. Specifically, the formula for calculating the model fusion weights is: ;in, For model fusion weights; Predict confidence levels for large models; Predict confidence levels for lightweight MRAN-hard models.

[0036] S404. The MRAN-hard models of each client are fused according to the model fusion weights to obtain the optimized MRAN-hard model.

[0037] By optimizing the MRAN-hard model, the causal knowledge of materials and metallurgy contained in the large model can be efficiently distilled into a lightweight model. Relying on causal path extraction and consistency constraints, the model learns feature distributions that conform to physical mechanisms, avoiding noise and spurious associations, and significantly improving the physical rationality and extrapolation reliability of the prediction results. At the same time, by adopting causal awareness attention and causal distillation loss, prediction accuracy and generalization ability are greatly improved while maintaining the model's lightweight nature and fast inference, effectively solving the problem of insufficient accuracy in traditional lightweight models. In addition, based on adaptive calculation of fusion weights based on causal confidence and combined with heterogeneous data scheduling optimization, model performance can be dynamically balanced, the influence of low-quality clients can be weakened, the convergence speed can be accelerated, and the robustness, stability, and engineering applicability of the federated learning global model can be improved.

[0038] S302. Input the standard hardenability dataset and MRAN-hard model of each client into the agent model, and then optimize the MRAN-hard model; based on the optimized MRAN-hard model, obtain the model optimization parameters for each client.

[0039] By constructing a collaborative architecture between a central meta-agent and five specialized sub-agents, full-process adaptive optimization of MRAN-hard models can be achieved. This architecture integrates materials science knowledge with reinforcement learning strategies, enabling dynamic adjustment of hyperparameters, automatic generation of physically significant phase transition features, and precise scheduling of high-quality client data. This effectively addresses issues such as inefficiency in manual parameter tuning, feature redundancy, and poor adaptation to heterogeneous data. Furthermore, relying on collaborative decision-making among agents and causal knowledge distillation, the model maintains high accuracy, fast convergence, and strong physical consistency within the federated learning framework, significantly improving the generalization ability, robustness, and engineering performance of the global model.

[0040] S104. Input the standard hardenability dataset of each client into the MRAN-hard model trained on the corresponding client to obtain the optimal loss value and optimal parameters of the model for each client; and send the optimal loss value and optimal parameters of the model for each client to the server.

[0041] In this embodiment, the step of training the corresponding client-adjusted MRAN-hard model based on the standard hardenability dataset of each client includes S501~S504: S501. Divide the standard hardenability dataset of each client into training set data and validation set data in a 7:3 ratio.

[0042] S502. Each client constructs its own loss function. The total loss during local training consists of two parts: the regression loss (HuberLoss) and the proximal term penalty. The specific formula is as follows: , ;in, It uses Huber Loss, which is highly robust to outliers; This represents the true value of hardenability for metallic materials. The hardenability prediction value output by the lightweight MRAN-hard model; The segmentation threshold for Huber loss is used to distinguish between squared loss and linear loss; This represents the total training loss on the client's local machine. This is the FedProx proximal term penalty coefficient, which controls the degree of deviation between the local model and the global model; These are the current parameters of the client's local model; This is the complete set of parameters for the local model. These are the parameters for the global model of federated learning.

[0043] S503. Based on the training set data, validation set data and loss function of each client, the MRAN-hard model is iteratively trained using a gradient accumulation strategy until the iteration termination condition is met, thereby obtaining the target MRAN-hard model for each client.

[0044] In this embodiment, based on the training set data, validation set data, and loss function of each client, a gradient accumulation strategy is used to iteratively train the MRAN-hard model until the iteration termination condition is met, including S601~S603: S601. Based on the training set data, the MRAN-hard model is trained using a gradient accumulation strategy. Specifically, the gradient accumulation strategy includes: (1) Gradient accumulation: The gradient accumulation step is set to 4. The gradient is accumulated through multiple forward propagations before updating the parameters to simulate the effect of large-batch training and improve training stability. (2) Gradient clipping: The maximum norm of the gradient is set to 0.5. The gradient is clipped using torch.nn.utils.clip_grad_ The norm_ function is used to avoid gradient explosion. (3) Dual learning rate scheduling: ① Global scheduling: CosineAnnealingLR is used, and the learning rate is decayed to 10 with each global round. ② Local scheduling: OneCycleLR is used, and the learning rate is increased and then decreased in the local training round to balance the convergence speed and generalization. (4) Local early stopping: The local early stopping patience value is set to 10. If the local loss does not decrease for 10 consecutive rounds, the local training is terminated early to avoid overfitting and ineffective training.

[0045] S602. Input the validation set data into the trained MRAN-hard model and calculate the model loss and evaluation metrics. Specifically, the evaluation metrics include R² (coefficient of determination), which characterizes the model's fit to the data and its predictive explanatory power; RMSE (root mean square error), which measures the average deviation between the predicted and true values; and MAE (mean absolute error), which reflects the absolute average level of the prediction error. The calculation formulas are as follows: , , ;in, This represents the actual hardenability test value. These are the model's predicted values. This represents the average of the actual measured values. For the number of data items; Index for the number of data items.

[0046] S603. If the model loss value and evaluation metric meet the iteration termination condition, then terminate the training and use the trained MRAN-hard model as the target MRAN-hard model; otherwise, adjust the learning rate of the trained MRAN-hard model using a dual learning rate scheduling strategy, and use the adjusted MRAN-hard model as the MRAN-hard model for the next training iteration, repeating the above steps.

[0047] Specifically, the termination condition is as follows: if the local loss does not decrease for 10 consecutive rounds or the evaluation metric meets the target, local early stop is triggered, terminating local training. The dual learning rate scheduling strategy is as follows: the global learning rate follows the CosineAnnealingLR strategy and decays with each global round (initially 0.005, decaying to 1e every 10 rounds). -5 The local learning rate follows the OneCycleLR strategy, first increasing and then decreasing within the local training epochs (maximum learning rate 0.01, minimum learning rate 1e). -6 This balances convergence speed and generalization.

[0048] Iterative training of the MRAN-hard model significantly improves its training stability and convergence efficiency. Gradient accumulation and pruning effectively prevent gradient explosion, simulating large-batch training effects with limited computing power and enhancing training stability. Combining dual-learning-rate scheduling and local early stopping strategies accelerates convergence while preventing overfitting and ineffective iterations, optimizing generalization ability. Furthermore, multi-metric evaluation ensures high prediction accuracy, low error, and strong physical consistency after training, making it more suitable for the hardening prediction needs of industrial scenarios.

[0049] S504. Based on the target MRAN-hard model for each client, obtain the optimal loss value and optimal parameters of the model for each client.

[0050] By obtaining the optimal loss values ​​and parameters for each client model, high-quality local updates can be provided for federated learning while ensuring client data privacy. Reasonable partitioning of the training and validation sets, combined with robust Huber loss and FedProx proximal constraints, can suppress outlier interference and constrain the deviation between local and global models, improving the stability and convergence speed of federated aggregation. The optimal parameters obtained through iterative optimization have high accuracy and strong generalization, providing a reliable foundation for subsequent global model fusion and lightweight deployment, effectively improving the overall hardenability prediction performance.

[0051] S105. In response to the lightweight aggregation model, each client performs hardenability prediction of metallic materials based on the lightweight aggregation model; wherein, the lightweight aggregation model is obtained by lightweighting the aggregated adaptive global model; the aggregated adaptive global model is obtained by the server by aggregating the adaptive global model based on the optimal loss value and optimal parameters of each client's model.

[0052] In this embodiment, adaptive global model aggregation is performed based on the optimal loss value and optimal parameters of each client, including S601~S604: S601. Calculate the sample number weight for each client based on the standard hardenability dataset for each client, where the sample number is the number of data in the client's standard hardenability dataset.

[0053] Specifically, the formula for calculating the sample size weight is as follows: ,in, Weights are based on the number of samples; The number of samples in the k-th client standard hardenability dataset; For client indexing (a counter variable that iterates through all clients); The total number of clients participating in federated learning; For the first The number of samples in a client-side standard hardenability dataset.

[0054] S602. Based on the optimal model loss value and optimal model parameters for each client, calculate the loss weight for each client. Specifically, a strategy of "the greater the loss, the lower the weight" is adopted, first performing inverse normalization on the loss, and then standardizing it. In this embodiment, the formula for calculating the loss weight is: ,in, The local optimal loss for the k-th client This represents the maximum local loss across all clients. Traverse the index for the client; The total number of clients participating in federated learning; Let the local optimal loss be for the m-th client; This represents the local optimal loss for the k-th client; For loss weights; It is a very small positive number, used to prevent the denominator from being 0 and to ensure calculation stability.

[0055] S603. Calculate the combined weights based on the sample size weights and loss weights for each client. The specific calculation formula is as follows: ,in, For combined weights; Weights are based on the number of samples; This is the balance coefficient for the combined weights; This is the loss weight.

[0056] In this embodiment, the sample number weight and the loss weight each account for 50%; finally, the combined weight of all clients is normalized to ensure that the sum of the weights is 1.

[0057] S604. Based on the combined weights, the MRAN-hard model of each client is weighted and summed to obtain the aggregated model. Specifically, the formula for weighted summation is: ,in, For the first Local model parameters for each client. These are the parameters for the new global model after aggregation. Traverse the index for the client; The total number of clients participating in federated learning; For the first The combined weights corresponding to each client.

[0058] This embodiment adopts a two-factor adaptive weighting strategy that integrates the number of samples and the training loss, and considers both the client data scale and the model training quality during aggregation, so as to achieve a more reasonable and robust global model update. The specific advantages are as follows: (1) Upgrading from single weighting to two-factor adaptive weighting abandons the traditional FedAvg method of weighting only by the number of samples, and introduces the training loss as the second weighting dimension, so that high-quality clients with more samples and lower loss can obtain higher aggregation weights, and suppress the negative impact of noisy data and insufficiently trained clients. (2) Layered normalization and adaptive weight integration: Sample number normalization weight: retains the sample size information and reflects the contribution of the client data scale; Loss normalization weight: adopts the reverse normalization mechanism to realize that the greater the loss, the lower the weight, and automatically weakens the clients with poor training effect; Combined weight: through hyperparameter balancing, the sample weight and loss weight are integrated to achieve an adaptive balance between data scale and training quality, and finally unified normalization ensures that the weights are legal and valid. (3) It can be embedded into the end-to-end federated learning framework. Based on adaptive weights, it performs weighted summation of the local model parameters of each client and directly outputs the updated global model. It can be seamlessly embedded into various federated learning architectures without modifying the local training process. It has low computational overhead and is easy to implement in engineering. (4) It simultaneously models the data scale and training loss. It achieves priority aggregation of high-quality clients through two-factor weighting, which improves the convergence speed and generalization ability in non-independent and heterogeneous data scenarios.

[0059] As some optional implementation methods of this embodiment, an industrial application interface layer is deployed on the central server and each client. The industrial application interface layer is the connection layer between the system and the industrial scenario, as follows: (1) Natural language interaction interface: It supports enterprise R&D personnel to issue various instructions such as metal material hardenability prediction, formula optimization, and process parameter adjustment in the form of code-free natural language. The system can be operated without mastering algorithm professional knowledge. The interface has a built-in natural language understanding and instruction conversion module, which can automatically parse the core requirements in natural language and convert them into structured input features and task instructions that the model can recognize. At the same time, it supports instruction feedback and sends the model execution status and prediction results back to the R&D personnel in the form of natural language to realize two-way natural interaction. (2) Data docking interface: A standardized data interaction channel is built based on the industrial general data protocol to realize seamless docking with the enterprise material R&D management system, production data acquisition system, and laboratory information management system. It supports automatic synchronization and incremental upload of local material data. It can also automatically push model prediction results and R&D decision suggestions to the enterprise business system to realize the two-way flow of data between the algorithm system and the industrial system, ensuring data timeliness and consistency, and eliminating the need for manual import and export of data. (3) Visualization Interface: Relying on big data visualization technology, a multi-dimensional visualization report of the entire federated learning process is generated, covering three core dimensions: training process, model performance, and prediction results. Abstract algorithm indicators are transformed into intuitive charts. The core content that can be displayed includes: federated training rounds-loss change curves, comparison of model evaluation indicators of each client, scatter plot of the fitting of predicted values ​​and experimental values, heat map of the contribution of each client to the global model, distribution trend of the hardening prediction results, etc. At the same time, it supports real-time refresh of indicators and historical data backtracking, so that developers can intuitively grasp the model training status and prediction effect.

[0060] (4) Decision Recommendation Output Interface: As the hub for transforming model prediction results into industrial R&D decisions, the interface transforms the pure numerical hardenability prediction results output by the model into actionable, evidence-based, and multi-form material R&D formulation and process optimization decision recommendations by combining knowledge from the metallurgical / materials science field with the actual R&D needs of enterprises. The interface supports two output formats: natural language description and standardized forms. The form format can be directly connected to the enterprise R&D management system and includes core optimization suggestions (such as the proportion of alloy element addition / reduction, quenching temperature / cooling rate adjustment values), theoretical basis (correlation law between material composition, process and hardenability), and data support (deviation between prediction results and target values, expected improvement in hardenability after optimization), etc. Example 2

[0061] Figure 2 A flowchart illustrating a method for predicting the hardenability of metallic materials applied to servers according to an embodiment of the present invention is shown. The method includes steps S701-S702: S701, responding to the standard hardenability dataset and MRAN-hard model of each client; optimizing the MRAN-hard model of the corresponding client based on the standard hardenability dataset of each client to obtain the model optimization parameters of each client; and sending the model optimization parameters of each client to the corresponding client.

[0062] In this embodiment, the optimization of the MRAN-hard model for each client is performed based on the standard hardenability dataset of each client to obtain the model optimization parameters for each client, including S801~S802: S801, Construct the intelligent agent model.

[0063] In this embodiment, the intelligent agent model includes: (1) Central meta-agent, used to calculate agent reward value.

[0064] (2) Hyperparameter optimization agent, used to dynamically adjust the data of the standard hardenability dataset of each client based on the standard hardenability dataset of each client, the agent reward value and the model parameters of the MRAN-hard model.

[0065] (3) Feature engineering enhances the agent, which is used to generate phase transition features based on the data output by the agent according to the hyperparameter optimization; calculates the fusion score based on the phase transition features, and takes the phase transition features with the fusion score greater than the fusion score threshold as the output of the feature engineering enhances the agent; the fusion score threshold is a preset fixed value.

[0066] (4) A heterogeneous data scheduling agent is used to calculate the comprehensive quality score of the client based on the output data of the feature engineering enhancement agent, and to take the data with a comprehensive quality score of the client greater than the quality score threshold as the output data of the heterogeneous data scheduling agent; the quality score threshold is a preset fixed value; (5) Model fusion lightweight agent, used to optimize the MRAN-hard model based on the output data of the heterogeneous data scheduling agent, including S901~S904: S901, extract causal paths based on the output data of the heterogeneous data scheduling agent to obtain causal path vectors. S902, use the causal awareness attention module to perform prior analysis on the output data of the heterogeneous data scheduling agent to obtain prior data; construct a causal distillation loss function; apply consistency constraints to the prior data based on the causal distillation loss function to obtain lightweight feature vectors. S903, calculate the causal confidence of the large model based on the causal path vectors; calculate the causal confidence of the lightweight model based on the lightweight feature vectors; construct model fusion weights based on the causal confidence of the large model and the causal confidence of the lightweight model. S904, fuse the MRAN-hard models of each client based on the model fusion weights to obtain the optimized MRAN-hard model.

[0067] S802. Input the standard hardenability dataset and MRAN-hard model of each client into the agent model, and then optimize the MRAN-hard model; based on the optimized MRAN-hard model, obtain the model optimization parameters for each client.

[0068] Specifically, the standard hardenability dataset for each client is obtained by preprocessing and federated normalization of the initial hardenability dataset obtained by each client through the data interface. In this embodiment, the federated normalization process includes S1001~S1002: S1001, Calculate the total number of global data based on the preprocessed dataset of each client's initial hardenability dataset, and then calculate the global mean and global variance based on the preprocessed dataset and the total number of global data. S1002, Construct a federated normalization formula, input the preprocessed initial hardenability dataset of each client, the global mean, and the global variance into the federated normalization formula to obtain the standard hardenability dataset for each client.

[0069] S702. In response to the optimal model loss value and optimal model parameters of each client, perform adaptive global model aggregation based on the optimal model loss value and optimal model parameters of each client to obtain an aggregated model; perform lightweight transformation on the aggregated model to obtain a lightweight aggregated model, and send the lightweight aggregated model to the client so that each client can predict the hardenability of metal materials based on the lightweight aggregated model.

[0070] In this embodiment, the adaptive global model aggregation based on the optimal loss value and optimal parameters of each client to obtain the aggregated model includes S1101~S1104: S1101. Calculate the sample number weight for each client based on the standard hardenability dataset for each client.

[0071] S1102. Calculate the loss weight for each client based on the optimal loss value and optimal parameters of the model for each client.

[0072] S1103 Calculate the combined weights based on the sample number weights and loss weights for each client. S1104 Calculate the weighted sum of the MRAN-hard models for each client based on the combined weights to obtain the aggregated model.

[0073] Specifically, the optimal loss value and optimal parameters of the model for each client are obtained by inputting the standard hardened dataset of each client into the corresponding trained MRAN-hard model; the trained MRAN-hard model is obtained by training the corresponding adjusted MRAN-hard model based on the standard hardened dataset of each client; the adjusted MRAN-hard model is obtained by each client adjusting the corresponding MRAN-hard model according to the model optimization parameters.

[0074] In this embodiment, the training of the corresponding client-adjusted MRAN-hard model based on the standard hardenability dataset of each client includes steps 1201 to 1203: S1201. Divide the standard hardenability dataset of each client into training set data and validation set data.

[0075] S1202, Each client constructs its own loss function.

[0076] S1203. Based on the training set data, validation set data, and loss function for each client, the MRAN-hard model is iteratively trained using a gradient accumulation strategy until the iteration termination condition is met, including steps 1301-1303: S1301. Based on the training set data, the MRAN-hard model is trained using a gradient accumulation strategy.

[0077] S1302. Input the validation set data into the trained MRAN-hard model, and calculate the model loss value and evaluation metric.

[0078] S1303. If the model loss value and evaluation metric meet the iteration termination condition, then terminate the training and use the trained MRAN-hard model as the target MRAN-hard model; otherwise, adjust the learning rate of the trained MRAN-hard model using a dual learning rate scheduling strategy, and use the adjusted MRAN-hard model as the MRAN-hard model for the next training iteration, repeating the above steps. Example 3

[0079] Based on the aforementioned method for predicting the hardenability of metallic materials, a transfer learning mechanism is introduced to transfer knowledge from a pre-trained model for the hardenability of general metallic materials to the J9 value prediction task for gear steel, addressing the issues of insufficient dedicated data for gear steel and weak model generalization ability. The system is deployed across one central server and four client servers belonging to gear steel production and R&D enterprises. The local data on each client server includes 25 chemical composition features and corresponding J9 value labels, ensuring complete protection of original data privacy. Through a combined strategy of pre-trained model transfer and federated learning fine-tuning, high-precision prediction of the J9 value of gear steel and engineering R&D decision-making are achieved. The specific process is as follows: Step 1: System Deployment and Initialization Configuration: (1) Build a distributed federated learning environment, deploy the GPU central server and 4 enterprise clients, and adopt a three-layer architecture: the federated learning base layer is deployed on the server and all clients; the large model agent layer is deployed on the central server; the industrial application interface layer is deployed on the server and clients, and an encrypted communication channel is established. The clients only upload model parameters and feature information and do not share the original data to ensure data privacy and security. (2) Complete the deployment and configuration of the large model agent on the central server, load the knowledge base in the field of metallurgical materials, and deploy five agents: hyperparameter optimization, feature engineering enhancement, heterogeneous data scheduling, model fusion lightweighting, and industrial implementation empowerment. With the central meta-agent as the core, global collaborative decision-making is achieved based on the MADDPG multi-agent reinforcement learning framework, and the permissions, interaction logic and communication protocol of each agent are clarified. (3) Deploy the five modules of the federated learning base layer: data preprocessing, FedProx local training, adaptive global aggregation, MRAN-hard model, model evaluation and early stopping. For the prediction of J9 value of gear steel, a customized MRAN-hard model structure is used: 25-dimensional input layer, 64 channels in ECA attention layer, 128 neurons in pre-activated residual block, and 1 neuron in output layer; the activation function is GELU, the weight is initialized with He, and the L2 regularization coefficient is 1e-4. (4) Deploy four standardized interfaces in the industrial application interface layer: natural language interaction interface, OPC UA / MQTT protocol data docking interface (compatible with PDM, SCADA, LIMS system), visualization display interface, and decision suggestion output interface (supports natural language reports and Excel forms) to achieve code-free operation, automatic data communication and result visualization. (5) The central server uniformly configures and distributes global initialization parameters: number of clients K=4, gradient accumulation steps 4, maximum gradient norm 0.5, local early stopping patience value 10, label smoothing coefficient 0.01, ε=10⁻ 6 Huber threshold δ=1.0, FedProx proximal term μ=0.01. Load the parameters of the general hardenability pre-trained model, encrypt and distribute them to each client as initial values ​​for local training. Step 2: Data Preprocessing and Federated Standardization: (1) Local Data Retrieval and Basic Preprocessing. Each client automatically obtains the local gear steel hardenability dataset through the data interface, which includes 25-dimensional chemical composition features, 8-dimensional rolling / quenching process parameters, and J9 value detection labels. The data preprocessing module completes basic cleaning and standardization. (2) Each client performs information statistics on the local dataset to form basic data features. (3) The central server calculates the global mean and global variance based on the global statistical results and distributes them to each client. If the global standard deviation is less than 10, it is replaced with 10 to avoid division by zero errors in the calculation. (4) Based on the global statistics, each client uses the federated standardization formula to process the local feature vectors, eliminate the differences in the dimensions and distribution bias of the data from different enterprises, and improve the model's cross-client generalization ability. (5) Each client performs data augmentation and label smoothing operations on the training set to enhance the robustness of the model and complete the construction of the local training dataset. (6) Each client divides the dataset into training and validation sets in a 7:3 ratio and uploads the division results, feature dimensions, sample distribution and other information to the server feature engineering augmentation agent to provide support for subsequent feature optimization.

[0080] Step 3: Global optimization configuration of the large model agent layer: (1) Feature engineering enhances the dynamic feature generation and screening of the agent. Based on the knowledge base of gear steel composition-hardenability law and phase diagram, the client features are dynamically generated and screened by physical information. The composition distribution is perceived in real time. It is detected that the Cr content (1.0%~1.5%) of client 1 crosses the eutectoid point and the Mo content of client 3 is too high. The physical feature generation is automatically triggered to obtain the carbon equivalent C. eq Twelve derived features, including martensitic transformation temperature (Ms), Cr equivalent, Mo equivalent, and elemental interaction terms, were identified. Physical significance was calculated based on a fusion score according to feature importance. Low-contribution features were removed with a threshold of 0.25, ultimately retaining 28 core features (20 of the original 25 dimensions were retained, with 8 new physical features added). The domain weights of each feature were calculated based on the physical fusion score. (Normalized to [0,1]), used for federated standardization weighting, to strengthen high physical significance features such as Ms and Ceq. (2) Quality grading and semantic correction of heterogeneous data scheduling agent clients. Client 2's "quenching temperature" feature deviates from the standard semantics, with semantic alignment A2=0.6, which is lower than the threshold A0=0.8. The scheduling agent initiates semantic correction, and the large model generates correction rules adapted to the enterprise's equipment: quenching temperature = austenitizing temperature + 10℃. After correction, A2 increases to 0.95, restoring normal aggregation weights. A four-dimensional evaluation system is constructed based on sample size, data distribution similarity, training loss, convergence rounds, and semantic alignment, and client quality is scored. The clients are divided into three levels: high (Qk≥0.8), medium (0.5≤Qk<0.8), and low (Qk<0.5), and differentiated aggregation weights and training rounds are assigned: high level weight ×1.2, 5 training rounds; medium level according to standard configuration, 3 training rounds; low level weight ×0.8, 1 training round. In this embodiment, clients 1 and 3 are high level, and clients 2 and 4 are medium level. (3) Multi-agent collaborative decision optimization. The three agents of hyperparameter optimization, feature engineering enhancement, and heterogeneous data scheduling make joint decisions under the central meta-agent collaborative framework. The central meta-agent calculates the collaborative reward based on the training state of the 5th round of federation (R²=0.76, convergence slows down). The value was lower than the historical average, so an adjustment strategy was immediately issued: the hyperparameter agent was instructed to reduce the learning rate to 0.001, the feature engineering agent was instructed to strengthen the generation of physical features of Cr element, and the scheduling agent was instructed to increase the number of local training rounds for high-quality clients. After the adjustment, the R² of the 6th round increased to 0.81, and the collaborative reward rebounded to 0.45. (4) Model fusion and lightweight agent model transformation. Based on the MRAN-hard basic architecture, knowledge fusion and lightweight preparation were completed for the prediction of J9 value of gear steel: a gear steel domain knowledge embedding layer was implanted after the input layer to transfer the pre-trained knowledge of hardenability of general materials; the lightweight core parameters such as knowledge distillation temperature coefficient and distillation loss weight were determined; the modified model architecture parameters were distributed to each client to complete the local model update.

[0081] Step 4: FedProx Local Model Training: (1) Each client initializes its local MRAN-hard model using the global model parameters issued by the server. The first round uses pre-trained parameters, and subsequent rounds use the parameters updated in Step 5. The pre-processed and optimized local training and validation sets are also loaded. (2) Each client constructs a local loss function with proximal term penalty based on FedProx, and uses Huber loss segmentation thresholds. delta =1.0, proximal term penalty coefficient muThe first 5 rounds are 0.01, and the 6th round onwards are adjusted to 0.02. (3) The local training adopts the gradient accumulation strategy. The model parameters are updated uniformly after every 4 forward propagations, and the gradient norm is controlled within 0.5 through gradient clipping to prevent gradient explosion. (4) The training process adopts a dual learning rate scheduling mechanism: the global learning rate decays according to the CosineAnnealingLR strategy, with an initial value of 0.005, and decreases to 1e-5 every 10 rounds; the local learning rate is dynamically adjusted according to the OneCycleLR strategy, with a maximum value of 0.01 and a minimum value of 1e-6, taking into account both convergence speed and model generalization ability. (5) After each round of local training, the model is evaluated using the validation set, and indicators such as R², RMSE, and MAE are calculated; if the local loss does not decrease for 10 consecutive rounds, early stopping is automatically triggered and training is terminated. (6) After the client completes the training, the local optimal model parameters and optimal loss are uploaded to the global aggregation module, and the training logs, indicator changes, and other data are uploaded to the industrial implementation empowerment intelligent agent.

[0082] Step 5: Adaptive Global Model Aggregation: (1) The central server's adaptive global aggregation module receives the local optimal model parameters, optimal loss, and sample count uploaded by each client, and calculates the maximum local loss value for each client. (2) Calculate the sample count weight and loss weight for each client according to the corresponding formulas. (3) Calculate the loss weight according to the loss normalization formula, following the principle of "the greater the loss, the lower the weight," and introduce... epsilon To avoid division by zero error. In this embodiment, the local optimal losses of clients 1–4 are 0.32, 0.45, 0.28, and 0.41 respectively, and the maximum loss Lmax = 0.45. The resulting loss weights are 0.30, 0.10, 0.35, and 0.25 respectively. (4) By combining the weight formula, the sample number weight and the loss weight are fused in proportion α = 0.5 to obtain the initial combined weights of 0.2905, 0.173, 0.3065, and 0.230 respectively. After normalization, the final adaptive weights are 0.290, 0.173, 0.307, and 0.230. (5) The adaptive global aggregation module calculates the updated global model parameters by weighted summation of the local optimal model parameters of each client according to the final adaptive weights. (6) The model evaluation and early termination module uses a third-party publicly available J9 value dataset for gear steel (300 samples) as an independent external validation set, and evaluates the global model with R² ≥ 0.85 as the core threshold. If the threshold is not met or the model does not converge (loss fluctuation > 5%), the new global model parameters are encrypted and sent to the client to continue the next round of federated training; when the threshold is met and convergence is achieved, the global model training is completed, and the optimal gear steel J9 value hardenability prediction model is obtained. In this embodiment, the global model in the 12th round achieved R² of 0.88 and RMSE = 0.99 HRC on the external validation set, which met the requirements, and training was stopped.

[0083] Step 6: Model Fusion Lightweighting and Industrial Adaptation: (1) Server Model Fusion Lightweight Agent performs knowledge distillation and lightweight transformation on the globally optimal MRAN-hard model: using the metallurgical field large model as the teacher model and MRAN-hard as the student model, extracting domain knowledge and reasoning ability through distillation loss, and constructing a large model + MRAN-hard hybrid prediction architecture with weights automatically assigned by prediction confidence. (2) Encrypt and distribute the lightweight model and hybrid architecture to the client to complete the local model replacement adaptation; the single-sample inference time on the edge device is reduced from 12ms to 5ms, with an efficiency improvement of 58%, meeting the requirements of low computing power deployment. (3) Industrial Adaptation Empowering Agent performs structured analysis on training logs, model indicators, client status and global aggregated data, and generates reports such as training curves, indicator comparisons, fitting scatter points, and contribution heatmaps through a visualization interface, which are pushed to the terminal in real time and support historical backtracking.

[0084] Step 7: Hardenability prediction and intelligent R&D decision output of metal materials: (1) Enterprise R&D personnel input information such as gear steel composition, process, and target J9 value in a no-code manner through the client's natural language interaction interface and issue hardenability prediction instructions; the industrial implementation empowerment intelligent agent parses the instructions, extracts the core parameters and converts them into structured inputs that the model can recognize. (2) The client inputs the structured features into the lightweight MRAN-hard model to complete the rapid prediction; for high-precision requirements such as new product R&D, the large model and MRAN-hard hybrid prediction architecture are automatically enabled to improve the reliability of the prediction, and the results are encrypted and uploaded to the server. (3) The server's industrial implementation empowerment intelligent agent combines gear steel material science knowledge and R&D needs, compares the predicted value with the target value, calculates the formula and process adjustment amount according to the deviation, generates optimization schemes including alloy element ratio, quenching temperature, cooling rate, etc., and provides theoretical basis and data support. (4) Decision suggestions are output in two forms: natural language reports and standardized forms. The forms can be directly connected to the enterprise's PDM R&D management system, and the predictions and suggestions are synchronized to the SCADA production acquisition system to provide a basis for process adjustment. (5) The system supports continuous iterative optimization: the client can upload new production and testing data in real time, and the server automatically triggers incremental federated training without having to train from scratch; the large model intelligent agent layer continuously senses the state and dynamically optimizes it, so that the model can continuously adapt to the actual needs of gear steel research and development and production.

[0085] According to the embodiments of the present invention, the present invention has the following advantages compared with the prior art: (1) The present invention constructs a collaborative architecture of "central meta-intelligent agent + five special sub-intelligent agents", integrates knowledge of the field of materials metallurgy and causal reasoning, guides the MRAN-hard model to learn feature associations that conform to physical mechanisms, avoids false associations and data noise interference, and combines Huber loss, causal distillation and other strategies to significantly improve the accuracy, stability and extrapolation reliability of hardenability prediction, and adapt to the accurate prediction needs in industrial scenarios. (2) Through optimizations such as causal knowledge distillation and lightweight feature extraction, while ensuring that the MRAN-hard model is small in size and fast inference speed, the reasoning ability and decision logic of the large model are retained, solving the pain points of decreased accuracy and weak generalization of traditional lightweight models, reducing client computing power overhead, and facilitating deployment and application in industrial sites. (3) Relying on the hierarchical scheduling and semantic correction mechanism of the heterogeneous data scheduling intelligent agent, the negative impact of low-quality clients and data is weakened; through the weighted aggregation of sample number weight and loss weight, the contribution of each client is balanced, avoiding the global model being dominated by extreme samples or low-quality data, and improving the efficiency and stability of global aggregation of federated learning. (4) From dynamic adjustment of hyperparameters and automatic generation of physical features to iterative training of models and global aggregation, the entire process is autonomously decided by the intelligent agent without human intervention, solving problems such as inefficiency of manual parameter tuning, cumbersome feature engineering, and poor adaptation to heterogeneous scenarios, improving the automation level of model training and optimization, and lowering the threshold for industrial applications. (5) Adopting a federated learning framework, each client does not need to upload data to the central server, but only shares model parameters, effectively protecting the privacy of core enterprise data; at the same time, through mechanisms such as near-end penalty and client quality assessment, it takes into account both data privacy and model collaboration effect, and adapts to industrial scenarios of multi-enterprise and multi-dataset collaboration.

[0086] According to embodiments of the present invention, an electronic device is also provided. The electronic device is intended to represent various forms of digital computers, such as laptop computers, desktop computers, workstations, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. The electronic device can also represent various forms of mobile devices, such as personal digital processors, cellular phones, smartphones, wearable devices, and other similar computing devices.

Claims

1. A method for predicting the hardenability of metallic materials, applied to a client, characterized in that, include: Each client obtains the corresponding initial hardenability dataset through the data interface, and performs preprocessing and federated standardization on the initial hardenability dataset in sequence to obtain the standard hardenability dataset for each client. Each client builds its own MRAN-hard model and sends its standard hardenability dataset and MRAN-hard model to the server. In response to the model optimization parameters of each client, the MRAN-hard model of the corresponding client is adjusted based on the model optimization parameters of each client; the adjusted MRAN-hard model of the corresponding client is trained according to the standard hardenability dataset of each client; wherein, the model optimization parameters of each client are obtained by the server optimizing the MRAN-hard model of the corresponding client according to the standard hardenability dataset of each client; Input the standard hardenability dataset of each client into the MRAN-hard model trained on the corresponding client to obtain the optimal loss value and optimal parameters of the model for each client; and send the optimal loss value and optimal parameters of the model for each client to the server. In response to the lightweight aggregation model, each client performs hardenability prediction of metallic materials based on the lightweight aggregation model; wherein, the lightweight aggregation model is obtained by lightweighting the aggregated adaptive global model; the aggregated adaptive global model is obtained by the server aggregating the adaptive global model based on the optimal loss value and optimal parameters of each client's model.

2. The method according to claim 1, characterized in that, The federal standardization process includes: Based on the preprocessed dataset of each client's initial hardenability dataset, calculate the total number of global data, and then calculate the global mean and global variance based on the preprocessed dataset and the total number of global data. A federated standardization formula is constructed by inputting the preprocessed initial hardenability dataset, global mean, and global variance of each client into the federated standardization formula to obtain the standard hardenability dataset for each client.

3. The method according to claim 1, characterized in that, The step of training the corresponding client-adjusted MRAN-hard model based on the standard hardenability dataset for each client includes: The standard hardenability dataset for each client is divided into training set data and validation set data; Each client constructs its own loss function; Based on the training set data, validation set data, and loss function for each client, the MRAN-hard model is iteratively trained using a gradient accumulation strategy until the iteration termination condition is met.

4. The method according to claim 3, characterized in that, Based on the training set data, validation set data, and loss function for each client, the MRAN-hard model is iteratively trained using a gradient accumulation strategy until the iteration termination condition is met, including: Based on the training set data, the MRAN-hard model is trained using a gradient accumulation strategy; Input the validation set data into the trained MRAN-hard model, and calculate the model loss value and evaluation metrics; If the model loss value and evaluation metric meet the iteration termination condition, training is terminated, and the trained MRAN-hard model is used as the target MRAN-hard model; otherwise, the learning rate of the trained MRAN-hard model is adjusted using a dual learning rate scheduling strategy, and the adjusted MRAN-hard model is used as the MRAN-hard model for the next training iteration, and the above steps are repeated.

5. A method for predicting the hardenability of metallic materials, applied to servers, characterized in that, include: In response to each client's standard hardenability dataset and MRAN-hard model; based on each client's standard hardenability dataset, optimize the corresponding client's MRAN-hard model to obtain the model optimization parameters for each client; and send the model optimization parameters for each client to the corresponding client; wherein, the standard hardenability dataset for each client is obtained by each client through the data interface after obtaining the corresponding initial hardenability dataset, and then performing preprocessing and federated standardization on the initial hardenability dataset in sequence; In response to the optimal model loss and optimal model parameters of each client, adaptive global model aggregation is performed based on the optimal model loss and optimal model parameters of each client to obtain an aggregated model. The aggregated model is then lightweighted to obtain a lightweight aggregated model, which is sent to the clients so that each client can predict the hardenability of metallic materials based on the lightweight aggregated model. Specifically, the optimal model loss and optimal model parameters of each client are obtained by inputting the standard hardenability dataset of each client into the corresponding trained MRAN-hard model. The trained MRAN-hard model is obtained by training the corresponding adjusted MRAN-hard model based on the standard hardenability dataset of each client. The adjusted MRAN-hard model is obtained by each client adjusting its corresponding MRAN-hard model according to the model optimization parameters.

6. The method according to claim 5, characterized in that, The process involves optimizing the MRAN-hard model for each client based on the standard hardenability dataset for that client, resulting in optimized model parameters for each client, including: Construct an intelligent agent model; The standard hardenability dataset and MRAN-hard model of each client are input into the agent model, and the MRAN-hard model is then optimized. Based on the optimized MRAN-hard model, the model optimization parameters of each client are obtained.

7. The method according to claim 6, characterized in that, The agent model includes: The central meta-agent is used to calculate the agent's reward value; The hyperparameter-optimized agent is used to dynamically adjust the data of the standard hardenability dataset for each client based on the standard hardenability dataset for each client, the agent reward value, and the model parameters of the MRAN-hard model. A feature-engineered augmented agent is used to generate phase transition features based on the data output by the agent through hyperparameter optimization; a fusion score is calculated based on the phase transition features, and phase transition features with a fusion score greater than a fusion score threshold are used as the output of the feature-engineered augmented agent; the fusion score threshold is a preset fixed value. A heterogeneous data scheduling agent is used to calculate a comprehensive client quality score based on the output data of the feature engineering enhancement agent, and to use data with a comprehensive client quality score greater than a quality score threshold as the output data of the heterogeneous data scheduling agent; the quality score threshold is a preset fixed value. The model incorporates a lightweight agent to optimize the MRAN-hard model by scheduling the agent's output data based on heterogeneous data.

8. The method according to claim 7, characterized in that, The optimization of the MRAN-hard model based on the output data of the heterogeneous data scheduling agent includes: Causal path extraction is performed on the output data of the heterogeneous data scheduling agent to obtain the causal path vector; The output data of the heterogeneous data scheduling agent is processed using a causal awareness attention module to obtain prior data; a causal distillation loss function is constructed; and a consistency constraint is applied to the prior data based on the causal distillation loss function to obtain a lightweight feature vector. Calculate the causal confidence score of the large model based on the causal path vector; calculate the causal confidence score of the lightweight model based on the lightweight feature vector; construct model fusion weights based on the causal confidence scores of the large model and the lightweight model. The MRAN-hard models of each client are fused according to the model fusion weights to obtain the optimized MRAN-hard model.

9. The method according to claim 5, characterized in that, The adaptive global model aggregation based on the optimal loss value and optimal parameters of each client's model yields an aggregated model, including: Calculate the sample number weight for each client based on the standard hardenability dataset for each client; Calculate the loss weight for each client based on the optimal loss value and optimal parameters of the model for each client; Calculate the combined weights based on the sample size weights and loss weights for each client; The aggregated model is obtained by weighted summation of the MRAN-hard model for each client based on the combined weights.

10. An electronic device, comprising at least one processor; and a memory communicatively connected to said at least one processor; characterized in that, The memory stores instructions executable by the at least one processor, which, when executed by the at least one processor, enables the at least one processor to perform the method of any one of claims 1-4 or 5-9.

Images