Wind power main shaft process parameter design method, system and device based on physical constraint machine learning

Abstract

The present application relates to the technical field of wind power large forging manufacturing, and particularly provides a wind power main shaft process parameter design method, system and equipment based on physical constraint machine learning, comprising: obtaining data samples containing process parameters, microstructure parameters and mechanical property parameters, and extracting physical metallurgical features therefrom; constructing a physical constraint forward mapping model based on the features, establishing a mapping relationship among the process parameters, microstructure and mechanical properties; generating a lookup table offline by exhaustive enumeration based on the forward mapping model; inversely deducing a target microstructure interval according to a target mechanical property parameter input by a user, searching for a candidate process parameter combination in the lookup table; performing multi-objective Pareto sorting and three-layer feasibility filtering on the candidate scheme, and outputting a recommended process scheme in an early sorting position. The present application ensures the physical rationality of the prediction result by embedding physical metallurgical rules, and realizes a second-level reverse optimization response by an offline table building strategy.

Description

Technical Field

[0001] This invention belongs to the field of wind power large forging manufacturing technology, specifically involving a method, system and equipment for designing wind turbine main shaft process parameters based on physical constraint machine learning. Background Technology

[0002] The wind turbine main shaft is a key load-bearing component of a wind turbine generator set. Its mechanical properties (tensile strength, yield strength, impact energy at -40℃, and fatigue life) directly determine the safety and service life of the wind turbine generator set.

[0003] Currently, the design of wind turbine main shaft processes mainly relies on the following two methods: the experience-based trial-and-error method, which is highly dependent on personal experience and results in significant differences in process schemes designed by different engineers; and the physical metallurgical model method, which has limited ability to describe the multi-factor coupling effects under complex process conditions and has a bottleneck in prediction accuracy.

[0004] In recent years, machine learning methods have been introduced into the field of materials processing optimization. However, existing solutions often directly map process parameters to mechanical properties, skipping the physical carrier of microstructure. This results in a lack of interpretability in the models and can lead to serious errors due to dimensional generalization collapse and confusion between statistical correlation and physical causality. Furthermore, the models establish a peak mapping between process parameters and mechanical properties. As long as the absolute value of temperature changes, the performance prediction will fluctuate drastically, resulting in poor process robustness. Summary of the Invention

[0005] To address the aforementioned shortcomings of existing technologies, this invention provides a method, system, and equipment for designing wind turbine main shaft process parameters based on physical constraint machine learning, in order to solve the aforementioned technical problems.

[0006] In a first aspect, the present invention provides a method for designing process parameters of wind turbine main shafts based on physical constraint machine learning, comprising: Acquire data samples, which include process parameters, microstructure parameters, and mechanical property parameters with mapping relationships; extract physical metallurgical characteristics from the data samples; A forward mapping model with physical constraints is constructed. The forward mapping model takes the process parameters and the physical metallurgical characteristics as inputs and the microstructure parameters and mechanical property parameters as multi-task outputs. The physical constraints include applying a hard physical boundary truncation to the predicted values ​​of the microstructure parameters and adding a causal consistency penalty term to the model loss function. Based on the forward mapping model, an offline exhaustive sampling is performed on the preset process parameter space to generate a lookup table; the lookup table stores the predicted values ​​of microstructure parameters and mechanical property parameters corresponding to each set of process parameters. Based on the target mechanical property parameters input by the user, the target microstructure range is deduced, and candidate process parameter combinations that simultaneously satisfy the target mechanical property parameters and the target microstructure range are retrieved from the lookup table. The candidate process parameter combinations are ranked in a multi-objective Pareto order and then filtered for feasibility. The top-ranked process parameters after filtering are then output as recommended process schemes.

[0007] Secondly, the present invention provides a wind turbine main shaft process parameter design system based on physical constraint machine learning, comprising: The data acquisition module is used to acquire data samples, which include process parameters, microstructure parameters, and mechanical property parameters with mapping relationships; and to extract physical metallurgical characteristics from the data samples. The model building module is used to construct a forward mapping model with physical constraints. The forward mapping model takes the process parameters and the physical metallurgical characteristics as inputs and the microstructure parameters and mechanical property parameters as multi-task outputs. The physical constraints include applying a hard physical boundary truncation to the predicted values ​​of the microstructure parameters and adding a causal consistency penalty term to the model loss function. The exhaustive sampling module is used to perform offline exhaustive sampling of the preset process parameter space based on the forward mapping model and generate a lookup table; the lookup table stores the predicted values ​​of microstructure parameters and mechanical property parameters corresponding to each set of process parameters. The data lookup module is used to infer the target microstructure range based on the target mechanical property parameters input by the user, and to retrieve candidate process parameter combinations that simultaneously satisfy the target mechanical property parameters and the target microstructure range from the lookup table. The solution output module is used to perform multi-objective Pareto ranking on the candidate process parameter combinations, and then perform feasibility filtering in sequence. The top-ranked process parameters after filtering are output as recommended process solutions.

[0008] Thirdly, a device is provided, comprising: The memory is used to store the design program for the process parameters of the wind turbine main shaft based on physical constraint machine learning; The processor is configured to implement the steps of the wind turbine main shaft process parameter design method based on physical constraint machine learning as provided in the first aspect when executing the wind turbine main shaft process parameter design program based on physical constraint machine learning.

[0009] Fourthly, a computer-readable medium is provided, on which a wind turbine main shaft process parameter design program based on physical constraint machine learning is stored. When the wind turbine main shaft process parameter design program based on physical constraint machine learning is executed by a processor, it implements the steps of the wind turbine main shaft process parameter design method based on physical constraint machine learning provided in the first aspect.

[0010] The beneficial effects of this invention are as follows: The wind turbine main shaft process parameter design method, system, and equipment based on physical constraint machine learning provided by this invention construct a physically interpretable forward mapping model by using physical metallurgical features as mandatory inputs, applying hard truncation of output boundaries, and causal consistency penalties. This ensures that the prediction results always conform to the basic metallurgical laws such as grain refinement strengthening and dynamic recrystallization, eliminating the prediction risk of pure data-driven models violating physical common sense. The phased training strategy effectively reduces the risk of overfitting under small sample conditions, improving the model's generalization ability and engineering reliability. Combining offline exhaustive table building and a three-layer feasibility filtering reverse optimization strategy, while ensuring online response speed, it ensures that the recommended solutions meet the triple hard constraints of equipment capability, industry standards, and process stability, achieving a unity of physical rationality, prediction accuracy, and engineering feasibility. This can effectively guide the development of new wind turbine main shaft processes and the optimization of existing processes. Attached Figure Description

[0011] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0012] Figure 1 This is a schematic flowchart of a method according to an embodiment of the present invention.

[0013] Figure 2 This is a schematic flowchart illustrating the construction of a forward mapping model of physical constraints according to an embodiment of the present invention.

[0014] Figure 3 This is a schematic block diagram of a system according to an embodiment of the present invention.

[0015] Figure 4 This is a schematic diagram of the structure of a device provided in an embodiment of the present invention. Detailed Implementation

[0016] To enable those skilled in the art to better understand the technical solutions of this invention, the technical solutions of the embodiments of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this invention, and not all embodiments. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of this invention.

[0017] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used herein in the description of the invention is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention.

[0018] The wind turbine main shaft process parameter design method based on physical constraint machine learning provided in this embodiment of the invention is executed by computer equipment, and correspondingly, the wind turbine main shaft process parameter design system based on physical constraint machine learning runs in the computer equipment.

[0019] Figure 1 This is a schematic flowchart illustrating a method according to an embodiment of the present invention. Wherein, Figure 1 The implementing entity can be a wind turbine main shaft process parameter design system based on physical constraint machine learning. Depending on different requirements, the order of steps in this flowchart can be changed, and some steps can be omitted.

[0020] like Figure 1 As shown, the method includes: S1. Obtain data samples, the data samples including process parameters, microstructure parameters and mechanical property parameters with mapping relationships; extract physical metallurgical characteristics from the data samples; S2. Construct a forward mapping model with physical constraints, wherein the forward mapping model takes the process parameters and the physical metallurgical characteristics as inputs and the microstructure parameters and mechanical property parameters as multi-task outputs; the physical constraints include applying a hard physical boundary truncation to the predicted values ​​of the microstructure parameters and adding a causal consistency penalty term to the model loss function; S3. Based on the forward mapping model, perform offline exhaustive sampling of the preset process parameter space to generate a lookup table; the lookup table stores the predicted values ​​of microstructure parameters and mechanical property parameters corresponding to each set of process parameters; S4. Based on the target mechanical property parameters input by the user, the target microstructure range is deduced, and candidate process parameter combinations that simultaneously satisfy the target mechanical property parameters and the target microstructure range are retrieved from the lookup table; S5. Perform multi-objective Pareto ranking on the candidate process parameter combinations, and then perform feasibility filtering in sequence. Output the top-ranked process parameters after filtering as recommended process schemes.

[0021] In one embodiment of the present invention, based on step S1, the following will provide a possible embodiment and describe its specific implementation in a non-limiting manner.

[0022] S101. Data Acquisition This step involves collecting production and testing data for wind turbine main shafts from multiple data sources to establish a basic dataset.

[0023] The data collection scope covers 42CrMo4 material wind turbine main shafts produced within the past three years, with complete records exported from the enterprise's quality management system and manufacturing execution system. Each record contains at least the following three types of information: (1) Process parameters: including steel ingot smelting batch number, forging heating temperature (°C), initial forging temperature (°C), final forging temperature (°C), strain rate (s). -1 The parameters include: forging ratio, post-forging cooling method, quenching temperature (°C), quenching holding time (h), quenching medium type and temperature, tempering temperature (°C), and tempering holding time (h). The forging temperature range is 1050–1200°C, and the strain rate range is 0.001–1 s². -1 .

[0024] (2) Microstructure parameters: including ASTM grain size grade, carbide equiaxedness rate (%), dynamic recrystallization percentage (%), and grain size range (grade). The above parameters are obtained by metallographic testing of samples taken from the spindle body, with sampling locations including at least 1 / 3 of the distance from the surface and the core.

[0025] (3) Mechanical property parameters: including tensile strength (MPa), yield strength (MPa), elongation after fracture (%), impact energy at -40℃ (J), and minimum single impact energy (J). The mechanical properties were obtained by tensile and impact tests on test bars that were heat-treated in the same furnace as the spindle.

[0026] For records with missing microstructure and mechanical property parameters, if the number of missing items exceeds 30% of the total number of fields, the entire record will be discarded. For each retained record, quality is tagged according to the following rules: process parameters from the Manufacturing Execution System (MES) automatic recorder are Grade A, while those from manual entry are Grade B; microstructure and mechanical property parameters from third-party testing reports are Grade A, while those from in-plant testing are Grade B. Grade A data is prioritized during model training, with Grade B data used only as supplementary training samples.

[0027] Outlier removal was performed on all collected raw data: records with process parameters exceeding the physical feasible range were directly deleted; records with microstructure parameters where the grain size exceeded the ASTM 5-8 range or the dynamic recrystallization percentage exceeded 0-100% were marked as outliers and deleted; records with mechanical property parameters where the tensile strength exceeded the 600-1300MPa range were marked as outliers and deleted.

[0028] After the above collection and cleaning, no fewer than 200 valid samples were obtained.

[0029] S102. Feature Extraction in Physical Metallurgy This step extracts four types of physical metallurgical features from the data samples obtained in step S101, transforming prior metallurgical knowledge into structured numerical features that can be used as input to the model.

[0030] (1) Zener-Hollomon parameter calculation The deformation temperature T (unit: K) and strain rate in the process parameters are used to determine the relationship between these parameters. Substituting the definitions of the Zener-Hollomon parameters, the Z-parameters, which comprehensively characterize the dynamic conditions of thermal deformation, are calculated. The calculation formula is as follows:

[0031] Where Q is the activation energy of hot deformation of 42CrMo4 steel, in J / mol. In this embodiment... J / mol; R is the ideal gas constant. J mol -1 K -1 T represents absolute temperature (K), which is converted from Celsius temperature t. .

[0032] The Z-parameter comprehensively reflects the synergistic effect of deformation temperature and strain rate on dynamic recrystallization behavior: a higher Z-value indicates that the deformation conditions are less favorable for dynamic recrystallization, and the recrystallization driving force is greater. The magnitude of the Z-parameter is typically in the range of... Within the range.

[0033] (2) Calculation of theoretical value of dynamic recrystallization grain size Based on the Z parameter calculated in step (1), the theoretical grain size after dynamic recrystallization is calculated by using the recrystallization grain size evolution model obtained by fitting the Gleeble thermal simulation experimental data in advance.

[0034] Gleeble thermal simulation experiments were conducted at deformation temperatures of 950–1200 °C and strain rates of 0.001–1 s⁻¹. -1Isothermal compression experiments were conducted within a certain range to obtain measured values ​​of recrystallized grain size under different deformation conditions. Nonlinear regression fitting was performed on the experimental data to obtain the dynamic recrystallized grain size. The relationship model between (unit: μm) and Z parameters:

[0035] Among them, A and is a material constant, obtained by fitting thermal simulation experimental data. For 42CrMo4 steel, the fitting result is: , Goodness of fit Not less than 0.95.

[0036] The grain size in this theory represents the grain refinement limit that can be achieved by dynamic recrystallization under given deformation conditions, reflecting the physical control of deformation conditions on the evolution of microstructure.

[0037] (3) Calculation of the quantitative value of the contribution of grain boundaries to strength Based on the measured grain size d (unit: μm) in the microstructure parameters obtained in step S101, the quantitative value of the contribution of grain boundaries to yield strength is calculated according to the principle of grain refinement strengthening. .

[0038] Grain size d is obtained by converting from ASTM grain size grade G:

[0039] Where G is the ASTM grain size grade number, and the value ranges from 5 to 8.

[0040] Based on the Hall-Petch grain refinement strengthening relation, the grain boundary strengthening contribution value (Unit: MPa) is:

[0041] in, Grain boundary strengthening factor (unit: MPa) mm 1 / 2 For 42CrMo4 steel, MPa mm 1 / 2 This coefficient reflects the ability of grain boundaries to impede dislocation movement and is an inherent strengthening parameter of the material.

[0042] This quantitative value reflects the physical law that the finer the grains and the larger the total area of ​​grain boundaries, the stronger the resistance to dislocation movement and the higher the material strength. It is a direct numerical mapping of the fine-grain strengthening mechanism.

[0043] (4) Calculation of the corrected quantitative value of the effect of carbide morphology on toughness Carbide isoaxiality based on the microstructure parameters obtained in step S101 (Unit: %), the morphology correction quantification value is calculated based on the empirical rules of the influence of carbide morphology on toughness. (Unit: J)

[0044] Carbide isometric ratio Obtained through metallographic image analysis, it is defined as the proportion of equiaxed carbide area to the total carbide area within the field of view, ranging from 0 to 100%. A higher equiaxation rate indicates a greater degree of spherical or equiaxed dispersion of carbides, and a weaker cutting effect on the matrix.

[0045] Morphological correction quantization value The calculation method is as follows:

[0046] in, and The morphological sensitivity coefficient was obtained through linear regression fitting. Based on the statistical relationship between the morphology of quenched and tempered carbides in 42CrMo4 steel and the impact energy at -40℃, the coefficient was taken as... J / %, J / % 2 This quadratic function relationship reflects the physical fact that the marginal contribution of the equiaxed ratio to toughness decreases as the equiaxed ratio increases.

[0047] (5) Combine to generate input feature set The Z-parameters and theoretical values ​​of dynamic recrystallized grain size obtained from steps (1) to (4) are used as the basis for calculation. Quantitative value of grain boundary strengthening contribution and carbide morphology correction quantization value Together with the process parameters (forging temperature, strain rate, quenching temperature, quenching holding time, and tempering temperature) obtained in step S101, these physical metallurgical features form the input feature set for the subsequent forward mapping model. The total dimension of this input feature set is no less than 9 dimensions, of which the original process parameters are 5 dimensions and the physical metallurgical features are 4 dimensions.

[0048] In one embodiment of the present invention, based on step S2, the following will provide a possible embodiment and its specific implementation plan in a non-limiting manner. Please refer to [the relevant documentation]. Figure 2 .

[0049] S201. Model Framework Selection and Multitasking Settings XGBoost (Extreme Gradient Boosting) ensemble tree was adopted as the basic learner and extended into a multi-task learning framework. The reasons for choosing XGBoost are as follows: (1) The ensemble tree model has low requirements for the number of samples and can obtain stable prediction results with 200 to 500 sets of effective samples, which is suitable for the actual scenario where the metallographic data of wind turbine main shaft is limited; (2) XGBoost naturally supports feature importance analysis and visualization of single tree structure, and has strong interpretability; (3) Physical constraint terms can be flexibly embedded by custom loss function.

[0050] The model has three output task branches that share the same set of input features. Each task branch shares the leaf node splitting path of the preorder tree within the model, and task separation only occurs in the final output layer. Task 1: Microstructure parameter prediction task, output nodes include ASTM grain size level prediction values. Predicted values ​​of carbide isometric ratio (%), predicted percentage of dynamic recrystallization (%) Task 2: Strength performance prediction task, output nodes include predicted tensile strength values. (MPa) and predicted yield strength (MPa); Task 3: Toughness performance prediction task, output nodes include predicted impact energy at -40℃. (J).

[0051] The three task branches share the underlying tree integration structure, and leverage the physical correlation between tasks (grain refinement simultaneously improves strength and toughness) to enhance generalization ability under small sample conditions.

[0052] S202. Forced Injection of Physical Metallurgical Characteristics and Process Parameters The four physical metallurgical features extracted in step S102 are: Zener-Hollomon parameters. Theoretical value of dynamic recrystallization grain size Quantitative value of grain boundary strengthening contribution Quantitative value of carbide speciation correction —Compared with the original process parameters (forging temperature) strain rate Quenching temperature quenching and holding time Tempering temperature Together, they form the complete input feature set of the basic model.

[0053] The above four physical metallurgical features are used as mandatory input features in the node splitting of each decision tree during model training. They are also given a prior weight higher than the original process parameters in the information gain calculation of each feature, so that the tree model will preferentially split along the direction of physical features each time.

[0054] During the data preprocessing stage of model training, for each set of samples, the physical baseline values ​​of microstructure and mechanical properties are calculated using the above-mentioned physical metallurgical characteristics through the following physical model: The physical reference value of the microstructure index is obtained by using the theoretical value of the dynamic recrystallized grain size calculated in step S102. By combining recrystallization kinetics and grain growth models, the theoretical evolution of grain size under given process conditions is calculated. ; Physical baseline values ​​for mechanical properties: Substitute the measured grain size into the Hall-Petch relationship to calculate the baseline contribution of grain refinement to yield strength. Simultaneously, the morphological strengthening contribution corresponding to the equiaxed ratio of carbides is introduced to calculate the physical reference value of strength. .

[0055] The training objective of the base model is defined as follows: taking the complete input feature set as input, the training objective is the deviation of the measured value corresponding to each output task from the physical benchmark value. The model is trained as a learning target. It only needs to fit the residual error that the physical model fails to describe. This residual error usually comes from higher-order interaction effects and random fluctuations that the physical model fails to cover, and the numerical magnitude is significantly smaller than the predicted quantity itself.

[0056] S203. Output Boundary Hard Truncation A hard truncation module is set after the model output layer. For predicted microstructure parameters, a forced correction is performed based on metallurgical physical thresholds. If the predicted grain size value Level, then forced Level; if Level, then forced The grade is determined based on the metallographic inspection standard for 42CrMo4 wind turbine main shaft forgings; If the predicted value of carbide equiaxylation Then forced ;like Then forced ; If the predicted value of dynamic recrystallization percentage Then forced ;like Then forced .

[0057] The aforementioned hard truncation takes effect simultaneously during the model inference phase and forward propagation, ensuring that the model will not output physically absurd microscopic organization parameters under any input conditions. Simultaneously, during backpropagation, gradients of predicted values ​​exceeding the boundary are set to zero to prevent anomalous gradients at the boundary from interfering with the update of model weights.

[0058] S204. Causality Consistency Penalty Item Three causal consistency penalty terms are introduced into the loss function of the base model.

[0059] XGBoost's base loss function is mean squared error:

[0060] Based on this, define the causal consistency penalty term. It consists of three sub-penalty terms:

[0061] in, , , The weighting coefficient for each penalty item is 0.15 in this embodiment.

[0062] The three sub-penalty terms are defined as follows: First item —Causal consistency between forging temperature and grain size: Based on the theory of recrystallization and grain growth, under the same strain rate, as the forging temperature increases, the atomic diffusion rate increases, the driving force for recrystallization and grain growth strengthens, and the grain size should tend to increase or remain unchanged. There should be no situation where the grains refine with increasing temperature. If the model predicts a smaller grain size for a high-temperature sample than for a low-temperature sample at the same strain rate, a penalty is applied.

[0063] in, This represents the predicted grain size value for samples in the low-temperature region. This represents the predicted grain size of the sample in the high-temperature region.

[0064] Second item Causal consistency between strain rate and dynamic recrystallization percentage: According to the dynamic recrystallization kinetics theory, as the strain rate increases, the dislocation multiplication rate increases, the recrystallization driving force strengthens, and the dynamic recrystallization percentage should tend to increase or remain unchanged. If the model predicts that increasing the strain rate will instead lead to a decrease in the dynamic recrystallization percentage, a penalty is imposed.

[0065] Third item Causal consistency between grain size and strength: According to the Hall-Petch grain refinement strengthening theory, yield strength and tensile strength should increase synchronously when grain size decreases (grain size level increases). If the model predicts that grain refinement will lead to a decrease in strength, a penalty is applied.

[0066] The three sub-penalty terms are only included in the loss function calculation during the training phase; they are not included during the model inference phase. The total training loss function is defined as:

[0067] In each iteration of model training, the total loss is minimized, guiding the model to conform to metallurgical causal laws while fitting the data.

[0068] S205. Phased Training Strategy A three-stage progressive training strategy is adopted, and the data sources, training objectives, and parameter update ranges for each stage are as follows: Phase 1: Pre-training. Using the Gleeble thermal simulation experimental data and historical metallographic data obtained in step S101 as the training set, the output layer of the mechanical property prediction task branch in the multi-task learning framework is frozen, and only the process parameter and microstructure mapping channel is trained. The goal of this phase is to prioritize enabling the model to master the regulatory laws of process parameters on microstructure evolution. The Gleeble experimental data covers 950–1200℃ and 0.001–1s... -1 The broad process range provides scarce extreme condition samples from historical production data, helping to broaden the model's understanding of process boundaries. After pre-training, the tree structure weights of the process evolution channels are saved as initial parameters.

[0069] Phase Two: Joint Fine-Tuning. All model parameters are released, and the measured mechanical property data obtained in step S101 are used to jointly fine-tune all three task branches. The process evolution channel weights saved in Phase One are used as initial values, and the accuracy of microstructure prediction and mechanical property prediction is simultaneously optimized based on these values. In this phase, the boundary hard truncation from step S203 and the causal consistency penalty term from step S204 are applied simultaneously to ensure that the multi-task joint optimization does not deviate from physical laws.

[0070] Phase 3: Online Incremental Update. After the model is deployed online, each new batch of spindle production and inspection data triggers the incremental update process. The new samples undergo quality control using the same cleaning rules as in step S101, are then merged with the historical training set, and the model is lightly fine-tuned. During fine-tuning, L2 regularization constraints are applied to the splitting thresholds of the first-level tree nodes corresponding to each physical metallurgical feature to prevent noise features in the new data from altering the model's fundamental physical understanding. The incremental update is triggered when the cumulative number of new valid samples reaches 20 or more than 30 days have passed since the last update.

[0071] In this embodiment, the hardware environment for model training is as follows: CPU Intel Xeon Gold 6248 (2.5GHz), 128GB memory, operating system Ubuntu 20.04LTS, Python version 3.8, and XGBoost version 1.5.0. The pre-training phase has 200 rounds, the joint fine-tuning phase has 100 rounds, and the early stopping phase has 20 rounds, meaning training terminates if the validation set loss does not decrease for 20 consecutive rounds.

[0072] To verify the prediction accuracy and physical rationality of the forward mapping model constructed in this embodiment, the following verifications were performed: Take the test set samples that did not participate in the training in step S101 (accounting for 20% of the total samples), and use the following three models for comparative prediction: (a) the physical constraint XGBoost multi-task model of this embodiment; (b) the pure data-driven XGBoost model without physical metallurgy features; (c) the ablation version of this model without causal consistency penalty terms.

[0073] Validation metrics include the coefficient of determination (R²) for each output task. 2 The results include the root mean square error (RMSE) and the physical consistency pass rate, which is the proportion of predictions in the test set that do not violate the three metallurgical causal laws.

[0074] The model in this embodiment has better prediction accuracy than the control group on each output task, and the physical consistency pass rate is not less than 95%, indicating that the constructed positive mapping model has physical rationality while maintaining prediction accuracy.

[0075] After model training is complete, the contribution of each input feature to the model's prediction is calculated using XGBoost's built-in gain importance analysis, outputting a feature importance ranking table. This ranking table can intuitively show whether physical features such as Zener-Hollomon parameters and Hall-Petch reinforcement terms are dominant, serving as a supplementary verification basis for the physical rationality of the model.

[0076] In one embodiment of the present invention, based on step S3, the following will provide a possible embodiment and describe its specific implementation in a non-limiting manner.

[0077] S301. Determination of Key Process Parameter Types and Their Physical Feasibility Domains This step determines the types of key process parameters involved in offline exhaustive sampling throughout the entire forging and heat treatment process of wind turbine main shafts, as well as the physical feasible domain of each parameter.

[0078] The key process parameters include the following five categories, covering the entire process from forging heating to final tempering: Forging temperature This refers to the surface temperature of the steel ingot during the initial forging process. The lower limit of the physically feasible region is based on the austenitizing temperature of 42CrMo4 steel. The temperature is determined by allowing sufficient temperature drop margin above (approximately 800℃), with a lower limit of 1050℃ to ensure the entire forging process is completed within the austenitic single-phase region; the upper limit is determined based on a combination of the grain coarsening temperature (approximately 1200℃) and the highest operating temperature of the heating furnace, with an upper limit of 1200℃. ℃.

[0079] strain rate This refers to the nominal strain rate during the forging deformation process. The lower limit is determined based on the minimum deformation rate of industrial forging equipment during normal operation, and is taken as 0.001s. -1 The upper limit is determined based on the equipment's maximum capacity and forging efficiency requirements, and is taken as 1.0s. -1 .Right now s -1 .

[0080] Quenching temperature This refers to the immersion temperature in water (or oil) after the austenitizing and heat treatment process. The lower limit is determined based on sufficient superheat above the complete austenitizing temperature of 42CrMo4 steel, and is set at 850℃; the upper limit is determined based on the grain coarsening temperature and the temperature control accuracy of the heat treatment furnace, and is set at 880℃. ℃.

[0081] Quenching and holding time This refers to the holding time of the workpiece at the quenching temperature, which directly affects the degree of austenitization homogenization and carbide dissolution. The lower limit is determined based on the shortest time required for the core to reach the quenching temperature and complete austenite homogenization, and is taken as 1 hour; the upper limit is determined based on oxidation and decarburization control requirements and production cycle constraints, and is taken as 6 hours. h.

[0082] Tempering temperature This refers to the set temperature during the tempering and holding process, which directly determines the final carbide precipitation morphology and the degree of dislocation recovery. The lower limit is determined based on the minimum temperature required to eliminate quenching stress and obtain tempered sorbite structure, and is set at 600℃; the upper limit is determined comprehensively based on the upper boundary of the second type of tempering brittleness-sensitive temperature range (approximately 550~650℃) and the lower limit requirement of tensile strength, and is set at 700℃. ℃.

[0083] The physical feasible domains of the above parameters are stored in tabular form in the system configuration file, allowing process engineers to manually modify them based on equipment upgrades or new standard releases. For wind turbine main shafts of different power ratings (e.g., 5MW and 15MW), the quenching and holding time... The feasible region can be scaled and adjusted proportionally according to the cross-sectional dimensions using a square relationship.

[0084] S302. Sampling strategy combining stratified sampling and localized dense sampling In this step, within the five-dimensional process parameter space determined in step S301, a strategy combining stratified sampling and localized dense sampling is used to generate a set of sampling points.

[0085] (1) Stratified sampling The five process parameter dimensions are divided into layers: Forging temperature Divided into 3 layers based on 50°C intervals, namely [1050, 1100), [1100, 1150), and [1150, 1200]. strain rate Divided by order of magnitude, there are three layers: [0.001, 0.01), [0.01, 0.1), and [0.1, 1.0]. Quenching temperature Divided into 3 layers at 10°C intervals: [850, 860), [860, 870), and [870, 880]. Quenching and holding time Divided into 5 layers based on 1-hour intervals: [1,2), [2,3), [3,4), [4,5), [5,6]. Tempering temperature Divided into 20℃ intervals, there are 5 layers in total: [600,620), [620,640), [640,660), [660,680), and [680,700].

[0086] Combining the five dimensions mentioned above using a Cartesian product, we get a total of Each process subspace is divided into two subspaces. Within each subspace, two sample points are extracted using the Latin hypercube sampling method to ensure a uniform distribution of sample points across all subspaces. The stratified sampling stage generates a total of [number missing] process subspaces. Each sampling point covers the entire process space.

[0087] (2) Local encrypted sampling Based on prior physical knowledge, the following three types of performance-sensitive regions are identified: Region 1: Near the critical temperature for dynamic recrystallization (forging temperature) In the temperature range of 1050~1100℃, the strain rate Between 0.01 and 0.1 seconds -1 Within this range, the recrystallization driving force and storage energy are in a critical state, and the tissue evolution is extremely sensitive to changes in temperature and rate. Zone 2: Tempering brittleness sensitive temperature range (tempering temperature) Within the range of 620~660℃, alloying elements show a significant tendency to segregate along grain boundaries, and the impact energy may undergo abrupt changes. Region 3: Quenching temperature boundary corresponding to the critical section of hardenability (quenching temperature) Within the range of 855~865℃, the degree of austenitization has a significant impact on the hardenability of the core.

[0088] Within the three sensitive regions mentioned above, the stratification intervals of the corresponding dimensions are reduced to half of the original intervals (e.g., forging temperatures are spaced at 25°C intervals, and tempering temperatures at 10°C intervals). The same Latin hypercube method is used for encrypted sampling, with an additional 3 sample points added to each encrypted subspace. Approximately 800 additional sample points are generated during the local encryption stage.

[0089] All sampling points from the stratified sampling stage and the local densification stage are merged and deduplicated (neighboring points whose Euclidean distance is less than 5% of the interlayer spacing in each dimension are removed), resulting in approximately 2150 candidate sampling points. In this embodiment, the total number of sampling points is in the range of 2000-3000 groups, which balances the resolution requirements of the lookup table for the process space with the computational cost of subsequent forward prediction.

[0090] To verify the coverage integrity of the sampling strategy, the coverage index of sampling points in each dimension (the proportion of the length of the sampled interval in each dimension to the total length of the feasible region) is calculated, and the coverage of each dimension is required to be no less than 95%.

[0091] S303. Call the forward mapping model to calculate the predicted value. This step inputs each set of sampled process parameters generated in step S302 into the forward mapping model constructed in step S103, and calculates the predicted values ​​of microstructure parameters and mechanical property parameters corresponding to each set of process parameters in batches.

[0092] The specific operation process is as follows: Approximately 2150 sets of sampled process parameters are organized into batches of 100 sets each, and input into the forward mapping model service interface in CSV format. After receiving the input vector containing five process parameters, the model service interface executes the following sequentially: The feature engineering module calculates four physical metallurgical features (Zener-Hollomon parameters, theoretical value of dynamic recrystallization grain size, quantified value of grain boundary strengthening contribution, and quantified value of carbide morphology correction), inputs them into the XGBoost multi-task learning model for forward inference, outputs the predicted values ​​of three task branches, and finally returns the prediction results after correction by the output boundary hard truncation module.

[0093] For each set of process parameter input The prediction results output by the forward mapping model include: Predicted values ​​of microstructure parameters: (ASTM grain size grade) (Carbide isoaxialization rate, %) (Percentage of dynamic recrystallization, %) Predicted values ​​of mechanical performance parameters: (Tensile strength, MPa) (Yield strength, MPa) (Impact energy at -40℃, J).

[0094] After the prediction calculations for all approximately 2150 sets of process parameters are completed, the input vectors of each set are merged with the corresponding prediction results to form a complete original prediction record set. The inference time for a single set of process parameters does not exceed 50 milliseconds, and the total calculation time for all sampling points is approximately 2 minutes.

[0095] S304. Structured storage generates lookup tables. This step involves structuring and storing the original prediction record set generated in step S303 according to the correspondence between the process parameter combination and the prediction results, and generating a lookup table.

[0096] The lookup table uses a relational data table structure, where each record corresponds to a set of process parameter combinations and all predicted results. The data table fields are designed as follows: Primary key field: Record number (auto-incrementing integer, used as a unique identifier).

[0097] Process parameter fields: Forging temperature (floating point, unit °C), Strain rate (floating point, unit s) - ¹), Quenching temperature (floating point, unit °C), quenching holding time (floating point, unit h), tempering temperature (floating point, unit °C).

[0098] Microstructure prediction fields: grain size prediction (floating point, one decimal place), carbide equiaxed ratio prediction (floating point, one decimal place, unit %), dynamic recrystallization percentage prediction (floating point, one decimal place, unit %).

[0099] Mechanical property prediction fields: tensile strength prediction (floating point, integer retained, unit MPa), yield strength prediction (floating point, integer retained, unit MPa), -40℃ impact energy prediction (floating point, integer retained, unit J).

[0100] Metadata fields: sampling batch number (identifying whether the record belongs to a stratified sampling batch or a locally encrypted batch), record generation timestamp.

[0101] The data table is persisted to the cloud storage system in Parquet columnar storage format. This format supports efficient compression and fast column reading, making it suitable for subsequent retrieval operations that filter by performance indicator fields. Furthermore, to accelerate the query efficiency of subsequent steps S104 and S105, which involve searching by performance thresholds and filtering by process parameter ranges, B-tree indexes are created on the process parameter fields (forging temperature, quenching temperature, tempering temperature) and the mechanical property prediction fields (tensile strength, -40℃ impact energy).

[0102] S305. Outlier Removal Processing This step performs outlier removal on the data records in the lookup table generated in step S304 to ensure that all records in the lookup table meet the physical rationality requirements.

[0103] The rules for identifying outliers are as follows: If the predicted grain size value If the value exceeds the ASTM level 5-8 range, it is considered abnormal and the record is deleted. The physical basis for this range is the acceptable grain size requirement for quenched and tempered state in the metallographic inspection standard for 42CrMo4 wind turbine main shaft forgings. If the predicted value of carbide equiaxylation If the percentage exceeds 90-100%, it is considered abnormal and the record is deleted. The physical basis for this lower limit is the minimum degree of carbide spheroidization required to meet the requirement of ≥110J impact energy at -40℃. If the predicted value of dynamic recrystallization percentage If the percentage exceeds 90-100%, it is considered abnormal and the record is deleted. The physical basis for this lower limit is the minimum degree of recrystallization to ensure sufficient refinement of the forging structure; If the predicted tensile strength value If the pressure exceeds the range of 600~1300MPa, it is considered abnormal and the record is deleted. The physical basis for this range is the theoretical limit of the tensile strength of 42CrMo4 steel in the quenched and tempered state; If the predicted impact energy at -40℃ If the value is less than 20J, it is considered abnormal and the record is deleted. The physical basis for this lower limit is the minimum acceptable impact toughness in engineering.

[0104] The above judgment rules are stored in configuration files, allowing for parameter adjustments based on material standard updates or specific customer requirements. After outlier removal, the number of remaining valid records is counted. If all records in a process subspace are removed, the system issues a warning and recommends re-sampling points in that subspace.

[0105] After removing outlier records, the lookup table retains at least 1800 valid records. The cleaned lookup table is then released as the official version, with the version number identified by a date and timestamp (in the format YYYYMMDD_HHMMSS). A new version is generated for each lookup table update (due to incremental model training or sampling range adjustments), and older versions are archived and stored for at least one year, supporting historical tracking and version rollback.

[0106] To verify the effectiveness and coverage completeness of the lookup table generated in this embodiment, the following verification was performed: Leave-one-out cross-validation was used. 5% of the records (approximately 90 records) were randomly selected from the lookup table, and their corresponding process parameters were re-input into the forward mapping model for prediction. The prediction results were then compared with the corresponding records stored in the lookup table. The requirement was that in over 95% of the sampled records, the predicted values ​​of microstructure parameters should not deviate by more than ±5%, and the predicted values ​​of mechanical property parameters should not deviate by more than ±3%, thus verifying the consistency between the lookup table data and the model predictions.

[0107] In one embodiment of the present invention, based on step S4, the following will provide a possible embodiment and describe its specific implementation in a non-limiting manner.

[0108] S401. Inferring the target microstructure region based on the target mechanical performance parameters input by the user. (1) Obtain the target mechanical performance parameters input by the user The target mechanical performance parameters input by the process designer are obtained through the user interface. These target mechanical performance parameters are submitted in a structured data format and include at least the following three items: Target value of tensile strength The unit is MPa. Users can enter a single point value (such as "≥850") or a range value (such as "850~1000"). -40℃ Impact Energy Target Value The unit is J. Users can enter a single point value (such as "≥110") or a range value (such as "110~150"). Fatigue life target value The unit is the number of loops, and the user can input a single value (e.g., "≥1×10"). 8 ” or lower limit value.

[0109] For numerical values ​​entered in the format ≥X, the lower limit value is automatically used as the target value for subsequent back-calculation; for numerical values ​​entered in the format of an interval, the lower limit value of the interval is taken as the target value for back-calculation to ensure that the micro-organism requirements derived from the back-calculation cover the entire target interval.

[0110] In addition, users can choose to set multi-objective optimization priorities, including but not limited to: maximizing mechanical performance margins, minimizing production costs, and minimizing overall energy consumption. These priority settings take effect in the subsequent multi-objective Pareto ranking in step S501.

[0111] (2) Based on the principle of fine grain strengthening, the target grain size range is deduced. According to the Hall-Petch fine-grain strengthening theory, there is a quantitative relationship between grain size and yield strength. For 42CrMo4 steel in the quenched and tempered state, the yield strength... (MPa) and grain size The relationship between (μm) can be expressed as:

[0112] in, It represents the intracrystalline frictional stress (MPa). Grain boundary strengthening factor (MPa) ).

[0113] The target tensile strength value input by the user Based on experience ( For quenched and tempered 42CrMo4 steel, as a work hardening allowance, (MPa) converted to target yield strength value .

[0114] Back-engineering the maximum grain size required :

[0115] For 42CrMo4 steel MPa MPa .Will According to the ASTM grain size conversion formula ( (Unit: mm) Convert to the corresponding lower limit of grain size level This refers to the minimum grain size level required to meet the tensile strength requirement.

[0116] Simultaneously, based on the industry standard documents' requirements for grain size (ASTM 5~8), the target grain size range is determined as follows:

[0117] That is, the lower limit of the target grain size level is taken as the larger value between the required grain size calculated by reverse calculation and the standard grain size requirement of level 5, and the upper limit is taken as the upper limit of the standard requirement of level 8.

[0118] Example: User inputs target tensile strength value MPa, converted to MPa, substituting into the above formula, yields... μm, corresponding to ASTM grain size The final target grain size range is ASTM 7.0~8.0.

[0119] (3) Based on the physical laws of the influence of carbide morphology on toughness, the target isoaxiality range is deduced. Based on the physical laws governing the influence of carbide morphology on low-temperature impact toughness, the equiaxedness of carbides... (%) and impact energy at -40℃ There is a positive correlation between (J) and [the other two]. For quenched and tempered 42CrMo4 steel, this relationship can be expressed as:

[0120] in, The intrinsic impact energy (J) is the energy generated when the carbides are distributed entirely in rod or plate form. The coefficient of equiaxed ratio for impact energy (J / %) is given. This is based on the fitting results of statistical data from quenched and tempered 42CrMo4 steel. J, J / %.

[0121] The user-inputted target value for impact energy at -40℃ Substituting the above relationships, we can deduce the minimum required carbide equiaxation rate. :

[0122] Furthermore, based on basic metallurgical knowledge, the equiaxed ratio of carbides has a physical upper limit of 100%, and in actual engineering, the equiaxed ratio of quenched and tempered 42CrMo4 steel can typically reach over 92%. Therefore, the target equiaxed ratio range is determined as follows:

[0123] That is, the lower limit is taken as the greater of the required isoaxialization rate and the engineering experience value of 90%, and the upper limit is taken as the physical upper limit of 100%.

[0124] Example: User inputs target value for impact energy at -40℃ J, substituting into the above formula, yields The final target is an isoaxialization rate of 92.0% to 100%.

[0125] (4) Inversely deduce grain size uniformity constraints based on the physical laws governing the influence of grain uniformity on fatigue performance. Based on the accelerating effect of grain inhomogeneity on fatigue crack initiation and propagation, extremely poor grain size... (Grade) and fatigue life There is a negative correlation between (number of cycles). The grain size range is defined as the difference between the largest and smallest grain size levels within the same field of view. The larger the value, the worse the microstructure uniformity, the more significant the tendency for local stress concentration, and the lower the fatigue life.

[0126] For 42CrMo4 steel in the quenched and tempered state, it meets the following requirements. The upper limit of the grain size range required for the fatigue life of the number of cycles is determined based on industry experience. Grade. The physical basis for this empirical value is: for every 0.5 grade increase in grain size range, the local stress concentration factor increases by approximately 8-12%, and at 10... 8 The decrease in fatigue strength can reach 15-20% with each cycle.

[0127] The fatigue life target value input by the user and Compare the number of iterations:

[0128] That is, the target fatigue life reaches 10. 8 During cycling, the grain size difference is required to be no more than level 1; the target fatigue life is relaxed to 10. 7 During cycling, the grain size tolerance requirement can be relaxed to level 1.5.

[0129] Example: User inputs target fatigue life value The upper limit of the grain size uniformity index is... class.

[0130] (5) Combine to generate target micro-organization intervals The three micro-organization constraints obtained from steps (2) to (4) are summarized to form a complete target micro-organization interval. :

[0131] The target micro-organization interval is passed to the subsequent step S402 in JSON structured data format as a hard constraint for lookup table retrieval.

[0132] S402. Retrieve candidate process parameter combinations from the lookup table. This step uses the target mechanical performance parameters and target microstructure intervals determined in step S401 as joint constraints, and performs multi-condition retrieval in the lookup table generated in step S305.

[0133] Search criteria are divided into two categories: Mechanical property hard constraints: Look up the predicted tensile strength value in the table. Not lower than the user-input tensile strength target value -40℃ Impact Energy Predicted Value Not lower than the user-inputted impact energy target value The predicted fatigue life value (when the lookup table contains this field) shall not be lower than the user-inputted target fatigue life value. .

[0134] Microstructure hard constraints: Look up the predicted grain size value in the table. Falling into the target grain size range determined in step S401 Predicted carbide isoaxialization rate Falling into the target isoaxialization range Predicted grain size range Not greater than the upper limit of the grain size uniformity index determined in step S401 .

[0135] The search criteria are expressed in a structured query statement (SQL) and executed with the assistance of a B-tree index on the lookup table. The search returns all candidate records that simultaneously satisfy all the above hard constraints. If the search results are empty, a prompt message is returned to the user, suggesting that the user relax the target mechanical performance requirements (such as reducing the target tensile strength value by 20 MPa or the target impact energy value by 5 J), and providing adjustment suggestions to guide the user to interactively relax the constraints and search again.

[0136] Each candidate record in the search results should include at least: a complete combination of process parameters, the corresponding predicted values ​​of microstructure parameters, the corresponding predicted values ​​of mechanical property parameters, and a unique number for that record in the lookup table.

[0137] The search results are used as a set of candidate process parameter combinations and passed to the subsequent step S501 for multi-objective Pareto ranking and feasibility filtering.

[0138] To verify the effectiveness of the search results in this step, 10 candidate process schemes with the middle ranking in the search results were randomly selected for Gleeble thermal simulation verification experiments. The measured mechanical properties were compared with the predicted values ​​stored in the lookup table. At least 8 of the 10 schemes were required to achieve measured values ​​exceeding 90% of the user's target requirements. Only after successful verification could the search logic be deployed online. For schemes that failed verification, the reasons for deviations were analyzed and fed back to the forward mapping model for incremental training or adjustment of the physical metallurgical feature extraction parameters.

[0139] In one embodiment of the present invention, based on step S5, a possible embodiment will be given below, and its specific implementation will be described in a non-limiting manner.

[0140] S501. Perform multi-objective Pareto ranking on candidate process parameter combinations and then perform feasibility filtering accordingly. (1) Obtain candidate process parameter combinations and their corresponding predicted values Extract all candidate process parameter combinations from the search results of step S402. Each candidate record contains the following information: a complete five-dimensional process parameter combination (forging temperature). strain rate Quenching temperature quenching and holding time Tempering temperature ); corresponding predicted values ​​of microstructure parameters (grain size) equiaxed carbide Percentage of dynamic recrystallization ); the corresponding predicted values ​​of mechanical property parameters (tensile strength) Yield strength -40℃ impact energy The candidate record set is denoted as ); and the unique ID of the record in the lookup table. ,in The total number of candidate solutions.

[0141] (2) Determine the weight allocation for multi-objective optimization Based on the multi-objective optimization priority set by the user in step S401, the weight allocation of the following three optimization objectives is determined: Optimization Objective 1: Maximize the margin of mechanical properties. The margin of mechanical properties is defined as the excess of the predicted value of each mechanical property parameter relative to the user's target value. Taking tensile strength as an example, the margin... This objective reflects the safety margin of the candidate solutions, and its weight is denoted as... The value range is 0~1, and the default value is 0.6; Optimization Objective Two: Minimize Production Costs. Production costs are quantitatively estimated based on the contribution of each process parameter, including heating energy costs (positively correlated with forging and quenching temperatures), time costs (positively correlated with quenching holding time and tempering time), and equipment wear and tear costs (positively correlated with strain rate). This objective value is calculated using a pre-defined cost estimation model, with weights denoted as follows: The value range is 0~1, and the default value is 0.3; Optimization Objective 3: Minimize overall energy consumption. Overall energy consumption is estimated by summing the heating and holding power consumption of each heat treatment process, with weights denoted as follows: The value range is 0~1, and the default value is 0.1.

[0142] The weights of the three optimization objectives satisfy the normalization constraint. Users can customize the values ​​of each weight using the interface sliders, and three preset weight templates are provided for quick selection: performance-first template ( Cost-first template ), Equalization template ( ).

[0143] (3) Multi-objective Pareto non-dominated ranking For the candidate solution set All solutions are subjected to Pareto non-dominated sorting. The Pareto dominance relation is defined as follows: For the two candidate solutions and A scheme is called a scheme if it meets all three of the following conditions. Domination Plan (recorded as) ): Condition 1: Plan The value of the mechanical performance margin target is not less than that of the scheme. Corresponding value (i.e., a higher or equal surplus); Condition 2: Plan The value in terms of production cost target is no greater than the plan. Corresponding value (i.e., lower or equal cost); Condition 3: Plan The value for the overall energy consumption target is no greater than that of the scheme. Corresponding value (i.e., lower or equal energy consumption); Condition 4: At least one of the above three conditions is a strict inequality (i.e., the two schemes do not completely overlap in the three-dimensional target space).

[0144] If candidate solutions Not included in the candidate solution set Dominated by any other scheme in it, Classified as the first Pareto frontier .Will The plan in the middle After removing the middle option, repeat the above process for the remaining options to obtain the second Pareto front. And so on, until all candidate solutions are sorted hierarchically.

[0145] After Pareto sorting is completed, the schemes are labeled with their hierarchical levels. The solutions at the first frontier are Pareto optimal solutions, and the smaller the level value, the higher the overall optimization degree of the solution. All candidate solutions are initially ranked according to the frontier level: all solutions at the first frontier are ranked before those at the second frontier, and so on.

[0146] (4) Perform multi-level feasibility filtering in sequence. For the candidate process parameter combinations after initial ranking, a three-layer progressive feasibility filtering is performed sequentially. The three layers of filtering are executed in order of increasing filtering intensity (i.e., increasing computational cost). Any scheme that fails to pass any filtering layer is eliminated and will not be evaluated in subsequent layers.

[0147] First filtration level: Equipment capability feasibility filtration Verify one by one whether each process parameter in the candidate process parameter combination falls within the equipment capacity range of the current production workshop.

[0148] The determination of equipment capacity constraints is based on the following: reading the current available equipment ledger from the enterprise equipment management system, and recording the following capacity parameters for each piece of equipment: the maximum operating temperature and temperature control accuracy of the heating furnace, the maximum rated load and operating frequency range of the forging press, the maximum loading capacity and heating rate of the heat treatment furnace, the type of cooling medium and maximum cooling rate of the quenching tank, and the temperature uniformity index of the tempering furnace.

[0149] Each process parameter in the candidate scheme was verified item by item: Forging temperature The temperature must not exceed the maximum operating temperature of the workshop heating furnace (1250℃ in this example), and should be within the temperature control accuracy range of the heating furnace (usually ±10℃). strain rate The strain rate must not exceed the maximum strain rate corresponding to the rated operating frequency of the forging press (1.5s in this embodiment). -1 ); Quenching temperature The temperature must not exceed the maximum operating temperature of the heat treatment furnace (950°C in this embodiment), and should be within the range guaranteed by the thermocouple accuracy. Quenching and holding time The maximum allowable holding time must not exceed the furnace capacity of the heat treatment furnace (due to the risk of oxidation and decarburization, the upper limit in this embodiment is 8 hours). Tempering temperature The temperature must not exceed the maximum operating temperature of the tempering furnace (750°C in this embodiment).

[0150] Any solution whose process parameters exceed the capability range of the corresponding equipment is rejected, marked as equipment infeasible, and will not proceed to subsequent filtration stages.

[0151] Second filtration stage: Process specification feasibility filtration Each candidate process parameter combination was verified to ensure it met the relevant industry standards and manufacturing specifications for wind turbine main shaft forgings.

[0152] The determination of process specification constraints is based on reading the system's pre-set process specification knowledge base, which contains at least the process control requirements specified in the following standards.

[0153] Specific verification items include, but are not limited to: whether the ratio of forging heating and holding time to cross-sectional dimensions meets the standard requirements (e.g., holding time for no less than 1 hour for every 100mm cross-sectional thickness); whether the quenching cooling medium temperature is within the range specified in the standard (water quenching 18~40℃, oil quenching 20~80℃); whether the tempering holding time meets the standard requirements for cross-sectional dimensions (e.g., holding time for no less than 1 hour for every 25mm cross-sectional thickness); and whether the post-forging cooling method meets the material's sensitivity to white spots.

[0154] Any solution that does not meet any of the above specifications will be rejected, marked as specification infeasible, and will not proceed to the third filtering level. For constraint changes caused by specification version updates, the system provides a version management function for the specification knowledge base, supporting process engineers to manually update constraint parameters.

[0155] Third filtration stage: Process stability and feasibility filtration For each process parameter in the candidate process parameter combination, simulate the operating conditions when each process parameter fluctuates within the actual production control accuracy range, and evaluate the process stability of the scheme.

[0156] The specific operating procedure for process stability filtration is as follows: First, the typical fluctuation range of each process parameter in actual production is determined. Based on historical control chart data from the workshop, the fluctuation ranges of each parameter are: forging temperature ±10℃, strain rate ±5%, quenching temperature ±5℃, quenching holding time ±10%, and tempering temperature ±5%. These fluctuation ranges are stored in configuration files, supporting personalized adjustments based on the actual control capabilities of the workshop equipment.

[0157] Then, disturbance simulations were performed on the candidate schemes: the setpoints of each process parameter were superimposed with positive and negative fluctuation limits to generate multiple sets of disturbance conditions. For each of the five process parameters, single-factor disturbance tests were conducted one by one, i.e., only one parameter was fluctuated at a time, while the other parameters remained unchanged, generating a total of [number missing] disturbance conditions. Group disturbance conditions.

[0158] For each set of disturbance conditions, the forward mapping model constructed in step S103 is called again for prediction, and the predicted values ​​of the mechanical performance parameters after the disturbance are calculated. , and predicted values ​​of microstructure parameters , .

[0159] The criteria for determining whether a candidate solution passes the process stability filter are: under all 10 disturbance conditions, the predicted tensile strength is not lower than the target tensile strength input by the user. 95%; the predicted impact energy at -40℃ is not lower than the target impact energy input by the user. 95%; the predicted grain size values ​​are all no coarser than the lower limit of the user's target grain size range. (i.e., grain size level not lower than) The predicted carbide equiaxedness ratios are all no lower than the lower limit of the user's target equiaxedness ratio range. 95%.

[0160] If any predicted value under any disturbance condition is lower than the above threshold, the scheme is determined to be unstable and eliminated, and marked as unstable and infeasible.

[0161] After all three layers of filtering are completed, the pass rate of each layer of filtering is calculated, a feasibility filtering report is generated, showing the layers at which each candidate solution was rejected and the reasons therefor, for process designers to trace and refer to.

[0162] S502. Output the top-ranked sets of process parameters after filtering as recommended process schemes. This step will sequentially process the candidate process parameter combinations through three layers of feasibility filtering, and output the final recommended order according to the hierarchical order determined by the Pareto sort in step S501.

[0163] In some embodiments, the wind turbine main shaft process parameter design system based on physical constraint machine learning may include multiple functional modules composed of computer program segments. The computer programs for each program segment in the wind turbine main shaft process parameter design system based on physical constraint machine learning may be stored in the memory of a computer device and executed by at least one processor to perform (see details). Figure 1 (Description) Functionality for designing wind turbine main shaft process parameters based on physical constraint machine learning.

[0164] In this embodiment, the wind turbine main shaft process parameter design system based on physical constraint machine learning can be divided into multiple functional modules according to its functions, such as... Figure 3 As shown. The module referred to in this invention is a series of computer program segments that can be executed by at least one processor and perform a fixed function, and is stored in memory. In this embodiment, the functions of each module will be described in detail in subsequent embodiments.

[0165] The data acquisition module is used to acquire data samples, which include process parameters, microstructure parameters, and mechanical property parameters with mapping relationships; and to extract physical metallurgical characteristics from the data samples. The model building module is used to construct a forward mapping model with physical constraints. The forward mapping model takes the process parameters and the physical metallurgical characteristics as inputs and the microstructure parameters and mechanical property parameters as multi-task outputs. The physical constraints include applying a hard physical boundary truncation to the predicted values ​​of the microstructure parameters and adding a causal consistency penalty term to the model loss function. The exhaustive sampling module is used to perform offline exhaustive sampling of the preset process parameter space based on the forward mapping model and generate a lookup table; the lookup table stores the predicted values ​​of microstructure parameters and mechanical property parameters corresponding to each set of process parameters. The data lookup module is used to infer the target microstructure range based on the target mechanical property parameters input by the user, and to retrieve candidate process parameter combinations that simultaneously satisfy the target mechanical property parameters and the target microstructure range from the lookup table. The solution output module is used to perform multi-objective Pareto ranking on the candidate process parameter combinations, and then perform feasibility filtering in sequence. The top-ranked process parameters after filtering are output as recommended process solutions.

[0166] Figure 4 The wind turbine main shaft process parameter design method based on physical constraint machine learning provided in this application embodiment can be applied to equipment. Those skilled in the art will understand that the equipment structure involved in the embodiments of this invention does not constitute a limitation on the equipment. The equipment may include more or fewer components than shown in the figures, or combine certain components, or have different component arrangements. Specifically, the equipment 400 may include: a processor 410, a memory 420, and a communication unit 430. These components communicate through one or more buses. Those skilled in the art will understand that the server structure shown in the figures does not constitute a limitation on the invention; it may be a bus topology or a star topology, and may include more or fewer components than shown in the figures, or combine certain components, or have different component arrangements.

[0167] Although the present invention has been described in detail with reference to the accompanying drawings and preferred embodiments, the present invention is not limited thereto. Various equivalent modifications or substitutions can be made to the embodiments of the present invention by those skilled in the art without departing from the spirit and essence of the invention, and such modifications or substitutions should all be within the scope of the present invention. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should also be covered within the protection scope of the present invention.

Claims

1. A method for designing wind turbine main shaft process parameters based on physical constraint machine learning, characterized in that, include: Acquire data samples, which include process parameters, microstructure parameters, and mechanical property parameters with mapping relationships; extract physical metallurgical characteristics from the data samples; A forward mapping model with physical constraints is constructed. The forward mapping model takes the process parameters and the physical metallurgical characteristics as inputs and the microstructure parameters and mechanical property parameters as multi-task outputs. The physical constraints include applying a hard physical boundary truncation to the predicted values ​​of the microstructure parameters and adding a causal consistency penalty term to the model loss function. Based on the aforementioned forward mapping model, an offline exhaustive sampling is performed on the preset process parameter space to generate a lookup table; The lookup table stores the predicted values ​​of microstructure parameters and mechanical property parameters corresponding to each set of process parameters. Based on the target mechanical property parameters input by the user, the target microstructure range is deduced, and candidate process parameter combinations that simultaneously satisfy the target mechanical property parameters and the target microstructure range are retrieved from the lookup table. The candidate process parameter combinations are ranked in a multi-objective Pareto order and then filtered for feasibility. The top-ranked process parameters after filtering are then output as recommended process schemes.

2. The method according to claim 1, characterized in that, Extracting physical metallurgical features from the data sample includes: Substituting the deformation temperature and strain rate in the process parameters into the definition of the Zener-Hollomon parameters, the Z parameters, which comprehensively characterize the thermal deformation kinetics, are calculated. Based on the Z parameters, the theoretical grain size after dynamic recrystallization is calculated using a recrystallization grain size evolution model obtained by fitting thermal simulation experimental data in advance. Based on the measured grain size in the microstructure parameters, the quantitative value of the contribution of grain boundaries to strength is calculated according to the principle of fine grain strengthening. Based on the carbide equiaxation rate in the microstructure parameters, the morphology correction quantification value is calculated according to the empirical law of the influence of carbide morphology on toughness. The Z parameter, the theoretical grain size, the quantified value of the grain boundary strengthening contribution, and the quantified value of the carbide morphology correction are used together as physical metallurgical features, and combined with the process parameters to form the input feature set of the forward mapping model.

3. The method according to claim 1, characterized in that, Constructing a forward mapping model of physical constraints includes: An ensemble tree multi-task learning framework is used as the basic model, and multiple output tasks are set. The output tasks include at least a microstructure parameter prediction task and a mechanical property parameter prediction task. The extracted physical metallurgical features and the process parameters are used together as the forced input features of the basic model, so that the basic model can perform residual learning within the feasible domain determined by the physical prior. The predicted values ​​of microstructure parameters output by the basic model are hard-truncated at the boundary based on metallurgical physical thresholds, and the predicted values ​​that exceed the reasonable range are forcibly corrected to the corresponding boundary values. A causal consistency penalty term is introduced into the loss function of the basic model. When the direction of change of process parameters violates the preset physical causal relationship with the direction of change of microstructure parameters or mechanical property parameters, an additional penalty is applied to the loss function. A phased training strategy is adopted. First, the basic model is pre-trained with thermal simulation experimental data and microstructure measured data. Then, the basic model is jointly fine-tuned with mechanical property measured data. During the fine-tuning stage, boundary hard truncation and causal consistency penalty terms are applied simultaneously.

4. The method according to claim 3, characterized in that, The extracted physical metallurgical features and the process parameters are used together as the forced input features of the basic model, enabling the basic model to perform residual learning within the feasible domain determined by physical priors, including: Based on the Z parameter and the theoretical value of dynamic recrystallization grain size in the physical metallurgical characteristics, the physical reference values ​​of microstructure indicators are calculated through the dynamic recrystallization kinetic model. Based on the grain refinement strengthening term and carbide morphology correction term in the aforementioned physical metallurgical characteristics, the physical benchmark values ​​of mechanical property indicators are calculated through a composite model of grain refinement strengthening and precipitation strengthening. The physical reference values ​​of the microstructure index and the physical reference values ​​of the mechanical property index are used as prior reference features, and combined with the process parameters and the physical metallurgical features to form the complete input feature set of the basic model; The basic model is trained using the complete input feature set as input and the deviation of the measured values ​​of the microstructure parameters and mechanical property parameters from the physical benchmark values ​​as the learning target, so that the model only needs to learn the residual errors that the physical model cannot describe.

5. The method according to claim 1, characterized in that, Based on the aforementioned forward mapping model, an offline exhaustive sampling is performed on the preset process parameter space to generate a lookup table, including: The key process parameters and their physical feasible domains in the entire process of forging and heat treatment of wind turbine main shafts are determined. The key process parameters include at least forging temperature, strain rate, quenching temperature, quenching holding time and tempering temperature. The physical feasible domains of each parameter are determined jointly based on the material phase transformation temperature, equipment limit capacity and industry standards. A strategy combining stratified sampling and localized dense sampling is adopted to sample the process parameter space. First, uniform stratified sampling is performed according to the physical feasible region of each process parameter to cover the entire area of ​​the complete process space. Then, based on physical prior knowledge, performance-sensitive areas are identified, and the sampling density is increased in the performance-sensitive areas. Each set of sampled process parameters is input into the forward mapping model one by one, and the forward mapping model is called to calculate the predicted values ​​of microstructure parameters and mechanical property parameters corresponding to each set of process parameters. The predicted values ​​of each set of process parameters and their corresponding microstructure parameters and mechanical property parameters are stored in a structured manner according to the correspondence between the combination of process parameters and the prediction results, and the lookup table is generated. Outlier removal is performed on the data records in the lookup table, removing records whose predicted values ​​of microstructure parameters or mechanical property parameters exceed the corresponding physical thresholds.

6. The method according to claim 1, characterized in that, Based on the target mechanical performance parameters input by the user, the target microstructure range is deduced, including: Obtain the target mechanical performance parameters input by the user, which include at least the target value of tensile strength, the target value of impact energy at -40℃, and the target value of fatigue life; Based on the physical principle of fine grain strengthening, the target tensile strength value is substituted into the quantitative relationship model between grain size and yield strength to deduce the upper limit of grain size required to meet the tensile strength requirement. The target grain size range is determined according to the allowable grain size range specified in the standard for wind turbine main shaft forgings. Based on the physical laws governing the influence of carbide morphology on toughness, the target value of impact energy at -40℃ is substituted into the quantitative relationship model between carbide equiaxedness and impact toughness to deduce the lower limit of carbide equiaxedness required to meet the impact energy requirement, and the target equiaxedness range is determined based on metallurgical knowledge. Based on the physical laws governing the influence of grain uniformity on fatigue performance, the target fatigue life value is substituted into the quantitative relationship model between grain size difference and fatigue life, and the upper limit of the grain size uniformity index required to meet the fatigue life requirement is derived. The target grain size range, the target equiaxed ratio range, and the upper limit of the grain size uniformity index are combined to form the target microstructure range.

7. The method according to claim 1, characterized in that, Candidate process parameter combinations are ranked using a multi-objective Pareto order, and then subjected to feasibility filtering, including: Obtain the retrieved candidate process parameter combinations and their corresponding predicted values ​​of mechanical property parameters and microstructure parameters; Based on the user-defined multi-objective optimization priority, the weight allocation of each optimization objective is determined. The optimization objectives include at least maximizing the mechanical performance margin, minimizing production costs, and minimizing energy consumption. A multi-objective Pareto non-dominated ranking is performed on each candidate process parameter combination. The schemes that are not simultaneously dominated by other candidate schemes on all optimization objectives are included in the Pareto front, and the candidate schemes are initially ranked according to the front level. For the candidate process parameter combinations after preliminary sorting, multi-level feasibility filtering is performed in sequence. The feasibility filtering is performed in order of increasing filtering intensity, and any scheme that fails to pass any filtering level is eliminated.

8. The method according to claim 7, characterized in that, For the candidate process parameter combinations after initial ranking, multi-level feasibility filtering is performed sequentially, including: Each process parameter in the candidate process parameter combination is verified to see if it falls within the equipment capacity range of the current production workshop. The equipment capacity includes the maximum operating temperature and temperature control accuracy of the heating furnace, the maximum rated load of the forging press, the furnace loading capacity of the heat treatment furnace, and the maximum cooling rate of the cooling system. Any scheme with any process parameter exceeding the corresponding equipment capacity range is eliminated. Each candidate process parameter combination is verified to ensure it meets the relevant industry standards and manufacturing specifications for wind turbine main shaft forgings. Any scheme that does not meet any specification requirements is eliminated. For the set values ​​of each process parameter in the candidate process parameter combination, simulate the working conditions when each process parameter fluctuates within the actual production control accuracy range. After superimposing the positive fluctuation limit value and the negative fluctuation limit value on the set values ​​of each parameter, call the positive mapping model again for prediction. If the predicted value of the mechanical performance parameter or the predicted value of the microstructure parameter under any fluctuation condition is lower than the requirement corresponding to the target mechanical performance parameter or the target microstructure interval input by the user, the scheme is judged to be unstable and is eliminated.

9. A wind turbine main shaft process parameter design system based on physical constraint machine learning, characterized in that, include: The data acquisition module is used to acquire data samples, which include process parameters, microstructure parameters, and mechanical property parameters with mapping relationships; and to extract physical metallurgical characteristics from the data samples. The model building module is used to construct a forward mapping model with physical constraints. The forward mapping model takes the process parameters and the physical metallurgical characteristics as inputs and the microstructure parameters and mechanical property parameters as multi-task outputs. The physical constraints include applying a hard physical boundary truncation to the predicted values ​​of the microstructure parameters and adding a causal consistency penalty term to the model loss function. The exhaustive sampling module is used to perform offline exhaustive sampling of the preset process parameter space based on the forward mapping model and generate a lookup table; The lookup table stores the predicted values ​​of microstructure parameters and mechanical property parameters corresponding to each set of process parameters. The data lookup module is used to infer the target microstructure range based on the target mechanical property parameters input by the user, and to retrieve candidate process parameter combinations that simultaneously satisfy the target mechanical property parameters and the target microstructure range from the lookup table. The solution output module is used to perform multi-objective Pareto ranking on the candidate process parameter combinations, and then perform feasibility filtering in sequence. The top-ranked process parameters after filtering are output as recommended process solutions.

10. A device for designing process parameters of wind turbine main shafts based on physical constraint machine learning, characterized in that, include: The memory is used to store the design program for the process parameters of the wind turbine main shaft based on physical constraint machine learning; A processor is configured to implement the steps of the wind turbine main shaft process parameter design method as described in any one of claims 1-8 when executing the wind turbine main shaft process parameter design program based on physical constraint machine learning.

Images