A method and system for real-time prediction of mechanical properties of nickel-cobalt alloy based on big data

Abstract

The present application relates to a kind of real-time prediction method and system of nickel-cobalt alloy mechanical properties based on big data, comprising: multi-source in-situ big data acquisition, alloy composition data, big data preprocessing, mechanism-constrained hybrid prediction model construction, model training and optimization, real-time prediction and edge deployment, prediction result feedback and adaptive update etc., the real-time prediction system of nickel-cobalt alloy mechanical properties based on big data includes, acquisition module, processing module, calculation module, optimization module, prediction module and feedback module: the present application realizes the online real-time prediction of nickel-cobalt alloy mechanical properties, response time ≤100ms, without offline sampling, solve the pain point of traditional detection mode lag, destructive, can realize production process real-time quality control, reduce the batch production of unqualified products, reduce production cost, improve production efficiency.

Description

Technical Field

[0001] This invention relates to the field of nickel-cobalt alloy performance testing technology, and in particular to a method and system for real-time prediction of the mechanical properties of nickel-cobalt alloys based on big data. Background Technology

[0002] Nickel-cobalt alloys are widely used in high-end fields such as aerospace, new energy, and high-end equipment manufacturing due to their excellent high-temperature strength, corrosion resistance, and wear resistance. Their mechanical properties directly determine the reliability and service life of end products. Therefore, accurate and real-time testing of the mechanical properties of nickel-cobalt alloys is a core part of production quality control.

[0003] Currently, the mechanical property testing of nickel-cobalt alloys mainly adopts offline sampling testing, which involves cutting alloy samples on the production line and sending them to the laboratory for routine tests such as tensile and hardness tests. This method has inherent defects such as long testing cycle (usually 24-72 hours), destructive sampling, inability to achieve real-time quality control of the production process, and test results lagging behind the production process. It is easy to lead to the mass production of unqualified products, which will significantly increase production costs, reduce production efficiency, and make it difficult to adapt to the quality control requirements of modern continuous production.

[0004] In existing technologies, some studies have attempted to use machine learning methods to predict the mechanical properties of nickel-cobalt alloys, but most of them have obvious limitations: First, they rely on a single data source, mostly using only alloy composition or static process parameters, without integrating dynamic signals and microstructure characteristics during the production process, resulting in insufficient feature dimensions; second, they use traditional regression or general machine learning models, without combining the mechanical strengthening mechanism of nickel-cobalt alloys, which are purely data-driven "black box" models, resulting in low prediction accuracy, weak generalization ability, and poor interpretability; third, most of them are offline prediction modes, which cannot adapt to the real-time data acquisition and rapid prediction needs on the production line, and are difficult to meet the real-time quality control requirements of industrial production.

[0005] To address the above problems, this invention proposes a method and system for real-time prediction of the mechanical properties of nickel-cobalt alloys based on big data. Summary of the Invention

[0006] In order to improve the accuracy of predicting the mechanical properties of nickel-cobalt alloys, this invention provides a method and system for real-time prediction of the mechanical properties of nickel-cobalt alloys based on big data.

[0007] In a first aspect, the present invention provides a real-time prediction method for the mechanical properties of nickel-cobalt alloys based on big data, employing the following technical solution: A real-time prediction method for the mechanical properties of nickel-cobalt alloys based on big data includes: S1: Multi-source in-situ data acquisition: On the nickel-cobalt alloy production line, multi-source in-situ data of the alloy production process is collected synchronously by multiple sensors to construct a big data set for nickel-cobalt alloy production; the multi-source in-situ data includes alloy composition data, process data, in-situ physical field signal data and microstructure characteristic data, realizing full coverage of data of the entire production process and providing data support for subsequent accurate prediction. S2: Data preprocessing: Cleaning, standardizing, feature extraction and feature screening of the multi-source in-situ big data collected in step S1, eliminating invalid data, eliminating the influence of units, and extracting key features to obtain an effective feature set for mechanical performance prediction, thereby improving model training efficiency and prediction accuracy. S3: Construction of a Hybrid Prediction Model with Mechanism Constraints: Based on the mechanical strengthening mechanisms of nickel-cobalt alloys (dislocation strengthening, solid solution strengthening, precipitation strengthening, grain boundary strengthening), a hybrid prediction model of "physical mechanism module + AI prediction module" is constructed. The physical mechanism module is used to provide theoretical constraints on mechanical properties and solve the problem of poor interpretability of pure AI models. The AI ​​prediction module is used to learn the nonlinear coupling relationship between multi-source features and mechanical properties, correct the theoretical errors of the physical mechanism module, and improve accuracy. S4: Model Training and Optimization: Using the effective feature set obtained in step S2 and the corresponding measured data of the mechanical properties of nickel-cobalt alloy, the hybrid prediction model is trained, verified and optimized to determine the optimal parameters of the model and ensure that the prediction accuracy and generalization ability of the model meet the requirements of industrial production. S5: Real-time prediction and edge deployment: The optimized hybrid prediction model is lightweighted to reduce the model's computing power requirements and deployed to edge computing nodes on the production line; during the production process, multi-source in-situ data is collected in real time, and after preprocessing in step S2, it is input into the lightweight model to realize real-time prediction of the mechanical properties of nickel-cobalt alloy and output the results. S6: Prediction Result Feedback and Adaptive Update: Compare the predicted mechanical properties with the actual offline test results, calculate the prediction error, and when the error exceeds the preset threshold, adaptively update the parameters of the hybrid prediction model to ensure the long-term stability of the prediction accuracy and adapt to the fluctuations in working conditions during the production process.

[0008] Optionally, in step S1, the alloy composition data includes Ni content, Co content, Cr content, Mo content, Al content, and Ti content, which are obtained in real time by inductively coupled plasma optical emission spectrometry (ICP-OES) with a detection accuracy of ≤0.01wt% to ensure the accuracy of the composition data. The process data includes rolling temperature, rolling force, strain rate, cooling rate, heat treatment temperature, and holding time. These data are collected synchronously using temperature sensors, pressure sensors, and strain sensors at a frequency of 100Hz to ensure the real-time and continuous nature of the process data. The in-situ physical field signal data includes ultrasonic backscattered signals, EBSD grain orientation signals, and infrared thermal imaging temperature field signals, which are acquired by an ultrasonic sensor, an EBSD detector, and an infrared thermal imager, respectively. The ultrasonic sensor has a frequency of 5MHz, the EBSD detection step size is 0.5μm, and the infrared thermal imager has a resolution of 320×240 pixels to ensure the clarity and effectiveness of the physical field signals. The microstructure feature data includes grain size, dislocation density, and precipitate volume fraction, which are obtained by inversion using EBSD and ultrasonic signals. This eliminates the need for offline sampling and detection, enabling online acquisition of microstructure features.

[0009] Optionally, in step S2, the big data preprocessing specifically includes: S21. Data cleaning: Use the 3σ criterion to remove outliers (data that deviates from the mean by more than 3 times the standard deviation is considered outlier and removed), and fill in missing data (linear interpolation is used when the missing rate is ≤5%, and K-nearest neighbor interpolation is used when the missing rate is >5%) to ensure the integrity and reliability of the data. S22. Data Standardization: All data are standardized using the Z-score standardization method to eliminate the influence of dimensions, allowing different types of data to be directly integrated for calculation. The standardization formula is: ; in, The data is standardized and has no unit; x represents the original data (corresponding to the original unit of different types of data, such as wt% for composition data and ℃ for temperature data). This is the mean of this type of data, and its unit is consistent with the original data. This is the standard deviation of this type of data, in the same units as the original data. and The mean is calculated from all the original data of the corresponding category collected in the big data set in statistical step S1, i.e., the formula for calculating the mean is: The formula for calculating standard deviation is: n is the number of samples of this type of data (unitless), which is obtained by counting the number of data of the corresponding category in the big data set; S23. Feature Extraction: Targeted feature extraction is performed on the in-situ physical field signal data, including extracting peak intensity (unit: V), signal amplitude variance (unit: V²), and propagation time (unit: ...) of the ultrasonic backscatter signal. 3 core features; EBSD grain orientation signal extraction extracts three core features: grain orientation difference (unit: °), grain size distribution (unit: μm), and grain boundary density (unit: m / m³); The infrared thermal imaging temperature field signal extracts three core features: average temperature (unit: °C), temperature gradient (unit: °C / mm), and temperature fluctuation amplitude (unit: °C). All features are calculated using corresponding signal analysis software (such as EBSD analysis software and ultrasonic signal processing software). S24. Feature Selection: The Random Forest algorithm is used to calculate the importance of each feature, and features with an importance score ≥ 0.05 are selected as the effective feature set. Redundant features are eliminated to reduce the computational cost of the model. The formula for calculating the feature importance score is as follows: ; in, Importance(f) Features f Importance score (unitless); K K is the number of decision trees in the random forest (unitless), with a value of 100, determined by 5-fold cross-validation (the dataset is divided into 5 parts, and 4 parts are used as the training set and 1 part as the validation set in turn to verify the model performance corresponding to different K values ​​and select the optimal K value); OOBk is the out-of-bag error of the k-th decision tree (unitless), and OOBk,f is the out-of-bag error of the k-th decision tree after randomly shuffling the values ​​of feature f (unitless); the out-of-bag error is calculated from the samples that did not participate in the training of the k-th decision tree, that is, the proportion of samples that the model predicted incorrectly among the samples that did not participate in the training.

[0010] Optionally, in step S3, the physical mechanism module is constructed based on the four major strengthening mechanisms of nickel-cobalt alloys (dislocation strengthening, solid solution strengthening, precipitation strengthening, and grain boundary strengthening), and the theoretical calculation expression for mechanical properties is as follows: ; in, The theoretical calculated values ​​of mechanical properties (unit: MPa) serve as the theoretical benchmark for characterizing the mechanical properties of nickel-cobalt alloys. The intrinsic lattice friction stress (unit: MPa) of the nickel-cobalt alloy is obtained by weighting the intrinsic stress of pure nickel and pure cobalt according to the alloy composition after tensile testing. The calculation formula is as follows: ,in , They are respectively Ni, Co The mass fraction (unitless) is obtained from the alloy composition data detected by ICP-OES in step S1; (The intrinsic lattice friction stress of pure nickel was obtained through tensile testing of pure nickel.) (The intrinsic lattice friction stress of pure cobalt was obtained through tensile testing of pure cobalt.) The contribution value for dislocation strengthening (unit: MPa) is calculated using the Taylor formula. The formula is as follows: Where M is the Taylor factor (unitless), and 3.06 is taken for FCC structure nickel-cobalt alloys (determined based on the crystal structure of nickel-cobalt alloys and referring to existing known techniques); α is an empirical constant (unitless), taken as 0.2 (obtained by fitting a large amount of tensile test data of nickel-cobalt alloys); G is the alloy shear modulus (unit: GPa), calculated from the elastic modulus E, the calculation formula is: Where v is Poisson's ratio (unitless), 0.31 for nickel-cobalt alloys (determined with reference to existing known technology); E is the elastic modulus of nickel-cobalt alloys (unit: GPa), taken as 196 GPa (referring to the elastic modulus range of Kovar alloys (nickel-cobalt alloys) 196-215 GPa, taking the median value, or it can be obtained through tensile testing); b is the Burgers vector (unit: m), taken as... (Determined with reference to existing known techniques); ρ is the dislocation density (unit: m). -2 The dislocation density is obtained by inverting the EBSD signal in step S1. Specifically, the grain orientation signal is processed by EBSD analysis software to calculate the dislocation density. The contribution value for solid solution strengthening (unit: MPa) is calculated using the solid solution strengthening model. The calculation formula is as follows: Where K is the solid solution strengthening coefficient (unitless), and 1200 is taken for the Ni-Co-Cr system (obtained by fitting a large number of solid solution strengthening test data of Ni-Co-Cr system alloys); ci is the mole fraction of each solid solution element (Cr, Mo, Al, Ti) in the alloy (unitless), which is obtained by converting the mass fraction detected by ICP-OES in step S1 (conversion formula: mole fraction = (mass fraction of a certain element / relative atomic mass of the element) / sum of (mass fraction / relative atomic mass) of all elements (Ni, Co, Cr, Mo, Al, Ti); n is the solid solution strengthening index (unitless), which is taken as 0.5 (determined with reference to existing solid solution strengthening theory and experimental data of nickel-cobalt alloys); To determine the precipitation enhancement contribution value (unit: MPa), the Orowan formula was used for calculation. The formula is as follows: ,in f The volume fraction of precipitated phases (unitless, expressed as a decimal) is obtained through EBSD signal inversion in step S1. The grain orientation signal is processed by EBSD analysis software to statistically determine the proportion of precipitated phases. r is the average radius of precipitated phases (unit: m), obtained by scanning electron microscopy (SEM). Specifically, SEM images of the alloy microstructure are captured, and the radii of at least 50 precipitated phases are measured using image analysis software, and the average value is taken. The meanings and acquisition methods of G and b are consistent with those in the above calculation of dislocation strengthening contribution value. The contribution value for grain boundary strengthening (unit: MPa) is calculated using the Hall-Petch formula. The formula is as follows: Where kHP is the Hall-Petch coefficient (unit: MPa·m¹ / ²), and for nickel-cobalt alloys it is 17.5 MPa·m¹ / ² (obtained by fitting a large amount of experimental data on the grain size and mechanical properties of nickel-cobalt alloys); d is the average grain size (unit: m), which is obtained by statistical analysis of EBSD signals in step S1. The grain orientation signal is processed by EBSD analysis software, and the size of at least 100 grains is statistically analyzed and the average value is taken.

[0011] The AI ​​prediction module is constructed using an attention-enhanced graph neural network (GNN) to correct theoretical calculation errors in the physical mechanism module, accurately learn the nonlinear coupling relationship between multi-source features and mechanical properties, and overcome the limitations of the "black box" model. The final output of the hybrid prediction model is: ; in, The final predicted values ​​of the mechanical properties of nickel-cobalt alloy (unit: MPa) are the predicted values ​​of yield strength and tensile strength. The predicted value of microhardness needs to be converted into HV hardness value according to industry standards. ω1 and ω2 are the output values ​​(unit: MPa) of the AI ​​prediction module, used to correct the theoretical errors of the physics mechanism module; ω1 and ω2 are weighting coefficients (unitless), satisfying... The values ​​are obtained by fitting the training data in step S4, with the initial values ​​all set to 0.5. During the fitting process, the values ​​are adjusted according to the prediction accuracy of the validation set to minimize the overall prediction error of the model.

[0012] Optionally, in step S4, model training and optimization specifically include: S41. Dataset Division: The effective feature set obtained in step S2 and the corresponding measured data of nickel-cobalt alloy mechanical properties (yield strength, tensile strength, and microhardness) are divided into a training set, a validation set, and a test set in a ratio of 7:2:1. The measured mechanical property data are obtained through offline tensile tests and microhardness tests. The tensile tests are conducted using a universal testing machine with a loading rate of 2 mm / min, according to GB / T 228.1-2010 standard, and the yield strength and tensile strength are recorded. The microhardness tests are conducted using a Vickers hardness tester with a load of 100 g and a holding time of 15 s, according to GB / T 4340.1-2009 standard, and the microhardness values ​​are recorded. S42. Model Training: Using the training set as input and measured mechanical performance data as labels, the Adam optimizer is used to train the hybrid prediction model. The initial learning rate is 0.001, the decay coefficient is 0.95 (the learning rate decays to 0.95 every 10 iterations), and the number of iterations is 500. The mean squared error (MSE) loss function is used to measure the deviation between the predicted and measured values. The loss function formula is: ; Where L is the loss value (unitless), the smaller the loss value, the higher the prediction accuracy of the model; m is the number of training set samples (unitless), which is obtained by counting the number of samples in the training set. This is the predicted value (in MPa) for the i-th sample. The measured value (unit: MPa) of the i-th sample is taken from the dataset in step S41. S43. Model Validation and Optimization: Validate the model performance using the validation set, calculate the prediction accuracy (R²) and mean absolute error (MAE). When R² < 0.92 or MAE > 5 MPa, adjust the network structure of the AI ​​prediction module (increase / decrease the number of layers in the graph neural network, with the number of neurons per layer adjusted from 64 to 256) and the weighting coefficients. , Retrain the model until it meets the performance requirements (R² ≥ 0.92 and MAE ≤ 5MPa); the formulas for calculating prediction accuracy and mean absolute error are: ; ; in, The prediction accuracy (unitless) ranges from [0,1]. The closer it is to 1, the higher the model's prediction accuracy. The mean of the measured values ​​in the validation set (unit: MPa) is obtained by averaging all measured values ​​in the validation set; MAE is the mean absolute error (unit: MPa), the smaller the value, the smaller the model prediction bias; m, , The meaning is consistent with that in the above loss function formula.

[0013] Optionally, in step S5, the model lightweighting process adopts a feature distillation + model pruning approach. Specifically, through feature distillation, 10-15 key mechanically sensitive features (i.e., features with an importance score ≥ 0.08) are extracted from the effective feature set in step S2, and secondary features are removed. The graph neural network of the AI ​​prediction module is pruned to remove redundant neurons and connections, retain the core network structure, reduce the number of model parameters by more than 60%, and ensure that the model can achieve millisecond-level inference (response time ≤ 100ms) on edge computing nodes (such as embedded chips). The edge computing node deployment adopts Docker containerization to ensure the stability and portability of the model operation and adapt to the equipment environment of different production lines.

[0014] Optionally, in step S6, the prediction error is calculated using absolute error, and the calculation formula is as follows: , where Error is the absolute prediction error (unit: MPa). The model predicted value (unit: MPa). The values ​​are measured offline (unit: MPa). The preset error thresholds are set according to different mechanical performance indicators, with the error thresholds for yield strength and tensile strength set at 5 MPa and the error threshold for microhardness set at 3 HV. When the error exceeds the corresponding threshold, an incremental learning method is adopted to add the effective features and measured values ​​of the sample to the training set and adaptively update the hybrid prediction model. The update frequency is once every 24 hours to ensure that the model can adapt to changes in operating conditions during the production process (such as fluctuations in raw material composition and adjustments to process parameters) and maintain long-term prediction accuracy.

[0015] Secondly, the present invention provides a real-time prediction system for the mechanical properties of nickel-cobalt alloys based on big data, characterized in that it includes: Acquisition module: The output end is electrically connected to the input end of the processing module, and is used to acquire composition data, process data, in-situ physical field signal data and microstructure characteristic data of nickel-cobalt alloy; Processing module: The output end is electrically connected to the input end of the calculation module. It is used to clean, standardize, extract features, and filter features of the big data collected in the acquisition module to obtain an effective feature set for mechanical performance prediction. The calculation module is electrically connected to the input of the optimization module at its output end, and is used to construct a hybrid prediction model of "physical mechanism module + AI prediction module". Optimization module: The output end is electrically connected to the input end of the prediction module. It is used to train, verify and optimize the hybrid prediction model based on the effective feature set obtained from the processing module and the corresponding measured data of the mechanical properties of nickel-cobalt alloy, and to determine the optimal parameters of the model. Prediction module: The output end is electrically connected to the input end of the feedback module. It is used to deploy the optimized hybrid prediction model to the edge computing node after lightweight processing. During the production process, multi-source in-situ data is collected in real time, preprocessed, and then input into the lightweight model to output the prediction results of the mechanical properties of nickel-cobalt alloy. Feedback module: Compares the predicted mechanical properties with the actual measured results, calculates the prediction error, and adaptively updates the parameters of the hybrid prediction model when the error exceeds the preset threshold to ensure the long-term stability of the prediction accuracy.

[0016] In summary, the present invention has at least one of the following beneficial technical effects: It creatively breaks through the limitations of "single data source + black box model" and adopts a hybrid prediction model that integrates multi-source in-situ big data (composition, process, physical field, microstructure) and mechanism constraints. This solves the problems of poor interpretability and weak generalization ability of pure data-driven models, as well as the problem of insufficient accuracy of pure mechanism models.

[0017] It enables online real-time prediction of the mechanical properties of nickel-cobalt alloys with a response time of ≤100ms and eliminates the need for offline sampling. This solves the problems of traditional testing methods, such as lag and high destructiveness. It can achieve real-time quality control in the production process, reduce the batch production of defective products, lower production costs by more than 30%, and improve production efficiency by more than 50%.

[0018] The big data preprocessing process is comprehensive. Redundant information is removed through feature extraction and screening. The model is lightweight and can be deployed on edge nodes to meet the real-time operation requirements of the production line without relying on cloud computing power, thus reducing industrial deployment costs.

[0019] It has the functions of prediction result feedback and adaptive update, which can adapt to the fluctuation of working conditions in the production process. The long-term prediction accuracy is stable (R²≥0.92, MAE≤5MPa), and it has strong generalization ability, which can be adapted to different grades and processes of nickel-cobalt alloy production scenarios. Attached Figure Description

[0020] Figure 1 This is an overall flowchart of Embodiment 1 of this application; Figure 2 This is a schematic diagram of the system structure of Embodiment 2 of this application. Detailed Implementation

[0021] The following combination Figures 1 to 2 The present invention will be described in further detail below.

[0022] Example 1: This application provides a method for real-time prediction of the mechanical properties of nickel-cobalt alloys based on big data, referring to... Figure 1 A real-time prediction method for the mechanical properties of nickel-cobalt alloys based on big data includes the following steps: S1: Multi-source in-situ big data acquisition: On the nickel-cobalt alloy hot working production line, ICP-OES spectrometer, temperature sensor, pressure sensor, strain sensor, ultrasonic sensor, EBSD detector, and infrared thermal imager are deployed simultaneously to collect multi-source in-situ data and build a big dataset (sample number n=1000).

[0023] Alloy composition data: Real-time ICP-OES analysis revealed the following contents: Ni 61.00 wt%, Co 30.00 wt%, Cr 15.00 wt%, Mo 5.00 wt%, Al 2.00 wt%, and Ti 1.00 wt%, with a detection accuracy of 0.01 wt%. Process data: acquired through corresponding sensors, rolling temperature 1150℃, rolling force 3500kN, strain rate 0.5s. -1 Cooling rate 20℃ / s, heat treatment temperature 850℃, holding time 2h, sampling frequency 100Hz.

[0024] In-situ physical field signal data: Ultrasonic backscattered signals were acquired by an ultrasonic sensor (5MHz), grain orientation signals were acquired by an EBSD detector (detection step size 0.5μm), and temperature field signals were acquired by an infrared thermal imager (320×240 pixels). Microstructure characteristics: obtained through EBSD signal inversion, grain size d = 8.5 × 10⁻⁶ -6 m, dislocation density ρ = 2.5 × 10 12 m -2 The volume fraction of the precipitated phase was f = 0.12; the average radius of the precipitated phase was obtained by SEM observation as r = 1.2 × 10⁻⁶. -7 m.

[0025] S2: Big Data Preprocessing Data cleaning: Outliers were removed using the 3σ criterion (a total of 32 outlier samples were removed), and missing data (missing rate 3.2%) were filled using linear interpolation, resulting in 968 valid samples; Data standardization: All data were processed using the Z-score standardization method. Taking rolling temperature as an example, the mean of the raw data was... Given a standard deviation σ = 12℃ and a sample rolling temperature x = 1150℃, the standardized data... Other data are standardized using the same formula. Feature extraction: Peak intensity of 2.5V and signal amplitude variance of 0.8V were extracted from the ultrasonic backscatter signal. 2 The propagation time was 12.3 μs; the grain orientation difference was 15.2°, the grain size distribution was 8.5 ± 1.2 μm, and the grain boundary density was 1.8 × 10⁻⁶ from the EBSD grain orientation signal. 6m / m³; The average temperature extracted from the infrared thermal image temperature field signal is 1145℃, the temperature gradient is 0.5℃ / mm, and the temperature fluctuation amplitude is 3.2℃. Feature selection: The Random Forest algorithm (K=100) was used to calculate the importance of each feature, and 12 features with an importance score ≥0.05 were selected as the effective feature set. Among them, the importance scores of grain size, dislocation density and precipitate volume fraction were 0.12, 0.10 and 0.09, respectively, and they were the core features.

[0026] S3: Construction of Hybrid Prediction Models with Mechanism Constraints Physical mechanism module construction: Based on the four strengthening mechanisms of nickel-cobalt alloys, calculate the theoretical values ​​of mechanical properties. The specific calculation process is as follows: Intrinsic lattice friction stress : ; Shear modulus G: ; Dislocation enhancement contribution value : ; Solid solution strengthening contribution value First, convert the mass fraction of each solid-solid element to mole fraction (Cr: 0.148, Mo: 0.052, Al: 0.074, Ti: 0.021). ; Extraction of enhanced contribution value : ; Grain boundary strengthening contribution value : ; Theoretical calculation value : ; AI prediction module construction: Employs an attention-enhanced graph neural network (GNN), taking an effective feature set as input and outputting an error correction value. ; Hybrid prediction model: final output formula Initial weighting coefficients .

[0027] S4: Model Training and Optimization Dataset partitioning: The 968 valid samples were divided into a training set (678 samples), a validation set (194 samples), and a test set (96 samples) in a ratio of 7:2:1. The measured mechanical properties data were obtained through offline testing. The average measured yield strength of this batch of alloys was 7850 MPa, the average measured tensile strength was 8200 MPa, and the average measured microhardness was 225 HV. 4.2 Model Training: The Adam optimizer was used with an initial learning rate of 0.001, a decay coefficient of 0.95, 500 iterations, and MSE as the loss function. The loss value gradually decreased during training, and the final training set loss value was 18.6. 4.3 Model Validation and Optimization: The validation set prediction accuracy R² = 0.93 and MAE = 4.2 MPa, meeting the performance requirements (R² ≥ 0.92, MAE ≤ 5 MPa); the optimal weighting coefficients were obtained through fitting. , .

[0028] S5: Real-time prediction and edge deployment: The optimized hybrid prediction model was lightweighted (12 key features were extracted by feature distillation, and the parameters were reduced by 65% ​​after model pruning), and deployed to edge computing nodes using Docker containerization, with a response time of 85ms. During production, multi-source in-situ data was collected in real time, preprocessed, and then input into the model to output the prediction results: yield strength 7845MPa, tensile strength 8192MPa, and microhardness 223HV.

[0029] S6: Prediction Result Feedback and Adaptive Update: The predicted results are compared with the actual offline test results, and the prediction error is calculated: yield strength error. (Not exceeding the threshold of 5MPa), tensile strength error (Exceeding the threshold of 5MPa), microhardness error (Not exceeding the threshold of 3HV); the sample was added to the training set, and the model was adaptively updated. After the update, the tensile strength prediction error of the model was reduced to 4.8MPa, which met the requirements.

[0030] The results of this embodiment show that the method of the present invention has high accuracy in predicting the mechanical properties of nickel-cobalt alloys (R²=0.93, MAE=4.2MPa), fast response speed (85ms), and can realize real-time prediction. It can effectively solve the pain points of the existing technology and meet the quality control needs of industrial production.

[0031] Example 2: This application provides a real-time prediction system for the mechanical properties of nickel-cobalt alloys based on big data, referring to... Figure 2 A real-time prediction system for the mechanical properties of nickel-cobalt alloys based on big data includes the following: Acquisition module: The output end is electrically connected to the input end of the processing module, and is used to acquire composition data, process data, in-situ physical field signal data and microstructure characteristic data of nickel-cobalt alloy; Processing module: The output end is electrically connected to the input end of the calculation module. It is used to clean, standardize, extract features, and filter features of the big data collected in the acquisition module to obtain an effective feature set for mechanical performance prediction. The calculation module is electrically connected to the input of the optimization module at its output end, and is used to construct a hybrid prediction model of "physical mechanism module + AI prediction module". Optimization module: The output end is electrically connected to the input end of the prediction module. It is used to train, verify and optimize the hybrid prediction model based on the effective feature set obtained from the processing module and the corresponding measured data of the mechanical properties of nickel-cobalt alloy, and to determine the optimal parameters of the model. Prediction module: The output end is electrically connected to the input end of the feedback module. It is used to deploy the optimized hybrid prediction model to the edge computing node after lightweight processing. During the production process, multi-source in-situ data is collected in real time, preprocessed, and then input into the lightweight model to output the prediction results of the mechanical properties of nickel-cobalt alloy. Feedback module: Compares the predicted mechanical properties with the actual measured results, calculates the prediction error, and adaptively updates the parameters of the hybrid prediction model when the error exceeds the preset threshold to ensure the long-term stability of the prediction accuracy.

[0032] The above are all preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Therefore, all equivalent changes made in accordance with the structure, shape and principle of the present invention should be covered within the scope of protection of the present invention.

Claims

1. A method for real-time prediction of the mechanical properties of nickel-cobalt alloys based on big data, characterized in that, Includes the following steps: S1. Multi-source in-situ big data acquisition: On the nickel-cobalt alloy production line, multi-source in-situ data of the alloy production process are collected synchronously by multiple sensors to construct a nickel-cobalt alloy production big data set; the multi-source in-situ data includes alloy composition data, process data, in-situ physical field signal data, and microstructure characteristic data; S2. Big data preprocessing: Clean, standardize, extract and filter the multi-source in-situ big data collected in step S1 to obtain an effective feature set for mechanical performance prediction. S3. Construction of a hybrid prediction model with mechanistic constraints: Based on the mechanical strengthening mechanism of nickel-cobalt alloys, a hybrid prediction model of "physical mechanism module + AI prediction module" is constructed. The physical mechanism module is used to provide theoretical constraints on mechanical properties, and the AI ​​prediction module is used to learn the nonlinear coupling relationship between multi-source features and mechanical properties. S4. Model Training and Optimization: Using the effective feature set obtained in step S2 and the corresponding measured data of the mechanical properties of nickel-cobalt alloy, the hybrid prediction model is trained, validated and optimized to determine the optimal parameters of the model. S5. Real-time prediction and edge deployment: After the optimized hybrid prediction model is lightweighted, it is deployed to edge computing nodes; during the production process, multi-source in-situ data is collected in real time, preprocessed and input into the lightweight model, and the predicted mechanical properties of nickel-cobalt alloy are output. S6. Prediction Result Feedback and Adaptive Update: Compare the predicted mechanical properties with the actual measured results, calculate the prediction error, and when the error exceeds the preset threshold, adaptively update the parameters of the hybrid prediction model to ensure the long-term stability of the prediction accuracy.

2. The method for real-time prediction of the mechanical properties of nickel-cobalt alloys based on big data according to claim 1, characterized in that, In step S1, the alloy composition data includes Ni, Co, Cr, Mo, Al, and Ti content, which are obtained in real time by inductively coupled plasma optical emission spectrometry (ICP-OES) with a detection accuracy ≤0.01wt%. The process data includes rolling temperature, rolling force, strain rate, cooling rate, heat treatment temperature, and holding time, which are simultaneously acquired by temperature sensors, pressure sensors, and strain sensors at a frequency of 100Hz. The in-situ physical field signal data includes ultrasonic backscattered signal, EBSD grain orientation signal, and infrared thermal imaging temperature field signal, which are acquired by ultrasonic sensor, EBSD detector, and infrared thermal imager, respectively. The ultrasonic sensor frequency is 5MHz, the EBSD detection step size is 0.5μm, and the infrared thermal imager resolution is 320×240 pixels. The microstructure characteristic data includes grain size, dislocation density, and precipitate volume fraction, which are obtained by inversion of EBSD and ultrasonic signals.

3. The method for real-time prediction of the mechanical properties of nickel-cobalt alloys based on big data according to claim 1, characterized in that, In step S2, the big data preprocessing specifically includes: S21. Data cleaning: Use the 3σ criterion to remove outliers and fill in missing data. When the missing rate is ≤5%, use linear interpolation; when the missing rate is >5%, use K-nearest neighbor interpolation. S22. Data Standardization: The Z-score standardization method is adopted, and the standardization formula is as follows: ,in The data is standardized, and x represents the original data. Let σ be the mean of this type of data, and σ be the standard deviation of this type of data. σ is obtained by calculating all the original data of the corresponding category in the big data set collected in statistical step S1; S23. Feature Extraction: Feature extraction is performed on in-situ physical field signal data. For ultrasonic backscattered signals, three features are extracted: peak intensity, signal amplitude variance, and propagation time. For EBSD grain orientation signals, three features are extracted: grain orientation difference, grain size distribution, and grain boundary density. For infrared thermal imaging temperature field signals, three features are extracted: average temperature, temperature gradient, and temperature fluctuation amplitude. S24. Feature Selection: The Random Forest algorithm is used to calculate the importance of each feature, and features with an importance score ≥ 0.05 are selected as the effective feature set. The formula for calculating the feature importance score is as follows: Where K is the number of decision trees in the random forest, and its value is 100. Let the out-of-bag error of the k-th decision tree be . This refers to the out-of-bag error of the k-th decision tree after randomly shuffling the values ​​of feature f.

4. The method for predicting the mechanical properties of nickel-cobalt alloys based on big data according to claim 1, characterized in that, In step S3, the physical mechanism module is constructed based on the dislocation strengthening, solid solution strengthening, precipitation strengthening, and grain boundary strengthening mechanisms of nickel-cobalt alloys. The theoretical calculation expression for mechanical properties is as follows: ; in, These are theoretically calculated values ​​of mechanical properties, in MPa. The intrinsic lattice friction stress of the nickel-cobalt alloy, in MPa, is obtained through... Calculated , These are the mass fractions of Ni and Co, respectively. , ; The value represents the dislocation strengthening contribution, expressed in MPa, using Taylor's formula. Calculations were performed, where M is the Taylor factor (taken as 3.06), α is an empirical constant (taken as 0.2), and G is the alloy shear modulus in GPa. Calculate, where ν is Poisson's ratio, taken as 0.31, and b is the Burgers vector, taken as... ρ is the dislocation density, in meters. -2 ; The value represents the contribution of solid solution strengthening, expressed in MPa. Calculate, where K is the solid solution strengthening coefficient, taken as 1200. denoted as the mole fraction of each solid solution element, and n as the solid solution strengthening index, which is taken as 0.5; To extract the enhancement contribution value in MPa, the Orowan formula was used. Calculate, where f is the volume fraction of the precipitated phase and r is the average radius of the precipitated phase, in meters; The value represents the contribution to grain boundary strengthening, expressed in MPa, and is calculated using the Hall-Petch formula. calculate, The Hall-Petch coefficient is set to 17.5 MPa·m. 1 / 2 d represents the average grain size in meters (m).

5. The method for predicting the mechanical properties of nickel-cobalt alloys based on big data according to claim 1, characterized in that, In step S3, the AI ​​prediction module is constructed using an attention-enhanced graph neural network (GNN), and the final output of the hybrid prediction model is: ,in These are the final predicted values ​​of mechanical properties, in MPa. The output value of the AI ​​prediction module, in MPa. , For the weighting coefficients, satisfying The values ​​were obtained by fitting the training data, with the initial values ​​all set to 0.

5.

6. The method for predicting the mechanical properties of nickel-cobalt alloys based on big data according to claim 1, characterized in that, In step S4, model training and optimization specifically include: dividing the effective feature set and the measured mechanical performance data into training set, validation set, and test set in a 7:2:1 ratio; training the model using the Adam optimizer with an initial learning rate of 0.001, a decay coefficient of 0.95, 500 iterations, and mean squared error (MSE) as the loss function; calculating the prediction accuracy (R²) and mean absolute error (MAE) using the validation set; and adjusting the AI ​​prediction module network structure and weighting coefficients when R² < 0.92 or MAE > 5 MPa, and retraining until the performance requirements are met.

7. The method for predicting the mechanical properties of nickel-cobalt alloys based on big data according to claim 1, characterized in that, In step S5, the model lightweighting process uses feature distillation and model pruning to extract 10-15 key mechanically sensitive features, reducing the number of model parameters by more than 60%, and deploying it to edge computing nodes with a response time of ≤100ms. In step S6, the prediction error is calculated using absolute error, with preset error thresholds of yield strength and tensile strength ≤5MPa and microhardness ≤3HV. When the error exceeds the threshold, incremental learning is used to adaptively update the model parameters.

8. The method for predicting the mechanical properties of nickel-cobalt alloys based on big data according to claim 4, characterized in that, In the calculation of the alloy shear modulus G, the elastic modulus E is taken as 196 GPa, or obtained by tensile testing; the average radius r of the precipitated phase is obtained by scanning electron microscopy (SEM), the average grain size d is obtained by statistical analysis of EBSD signals, and the dislocation density ρ is obtained by inversion of EBSD signals.

9. A big data-based system for predicting the mechanical properties of nickel-cobalt alloys, employing the big data-based method for predicting the mechanical properties of nickel-cobalt alloys as described in any one of claims 1-8, characterized in that: include: Acquisition module: The output end is electrically connected to the input end of the processing module, and is used to acquire composition data, process data, in-situ physical field signal data and microstructure characteristic data of nickel-cobalt alloy; Processing module: The output end is electrically connected to the input end of the calculation module. It is used to clean, standardize, extract features, and filter features of the big data collected in the acquisition module to obtain an effective feature set for mechanical performance prediction. Calculation module: The output end is electrically connected to the input end of the optimization module, and is used to construct a hybrid prediction model of "physical mechanism module + AI prediction module"; Optimization module: The output end is electrically connected to the input end of the prediction module. It is used to train, verify and optimize the hybrid prediction model based on the effective feature set obtained from the processing module and the corresponding measured data of the mechanical properties of nickel-cobalt alloy, and to determine the optimal parameters of the model. Prediction module: The output end is electrically connected to the input end of the feedback module, and is used to deploy the optimized hybrid prediction model to the edge computing node after lightweight processing. During the production process, multi-source in-situ data are collected in real time, pre-processed, and then input into the lightweight model to output the predicted mechanical properties of nickel-cobalt alloys. Feedback module: Compares the predicted mechanical properties with the actual measured results, calculates the prediction error, and adaptively updates the parameters of the hybrid prediction model when the error exceeds the preset threshold to ensure the long-term stability of the prediction accuracy.

Images