A method of predicting carbide characteristics in the hot working of high speed steel

Abstract

The application discloses a method for predicting carbide characteristics in a high-speed steel hot working process, which comprises the following steps: ingredients are prepared according to the standard grade composition of the high-speed steel to obtain an ingot, and the high-speed steel is obtained by hot working; a plurality of groups of metallographic microstructure photos are obtained along diagonal directions in the high-speed steel to obtain carbide characteristics as target values; the high-speed steel hot working process is simulated by using a finite element software, and stress, strain and temperature data in the finite element are extracted as characteristic values; the characteristic values are processed in multiple ways for dimension reduction, and the optimal way is screened to obtain optimal dimension reduction data results; the dimension reduction data results are segmented, a plurality of machine learning algorithms are used to model and predict the segmented data, and error evaluation is performed on a plurality of prediction results based on the target values to obtain optimal prediction results. The application effectively utilizes finite element data for analysis, accurately predicts the characteristics of carbides in the high-speed steel, and significantly reduces the "trial and error" cost by 35-65%, and improves the optimization efficiency.

Description

Technical Field

[0001] This invention relates to the field of high-speed steel technology, and in particular to a method for predicting carbide characteristics during the hot working of high-speed steel. Background Technology

[0002] High-speed steel (HSS) contains high-quality alloying elements, and the morphology, quantity, and distribution of the numerous alloy carbides formed determine its performance. Under typical casting processes, due to compositional segregation, carbides are large in size and unevenly distributed, exhibiting a network-like structure. This network thickness, in particular, severely degrades the performance of HSS. To improve carbide distribution, hot working methods such as forging and rolling are typically used to eliminate eutectic structures. However, optimizing hot working processes is challenging. Trial-and-error optimization is costly, while finite element simulation (FEM) is a cost-effective method, but it still requires extensive computation, making it time-consuming and labor-intensive. Therefore, effectively utilizing FEM data to optimize HSS processes is crucial for improving carbide distribution.

[0003] Therefore, there is a need in the existing technology to improve the methods for predicting carbide characteristics during hot working of high-speed steel. Summary of the Invention

[0004] In view of this, the purpose of this invention is to provide a method for predicting carbide characteristics during the hot working of high-speed steel.

[0005] To achieve the above objectives, embodiments of the present invention provide a method for predicting carbide characteristics during hot working of high-speed steel, comprising the following steps:

[0006] S1. Based on the standard grade composition of high-speed steel, the raw materials are batched and smelted to obtain ingots, and the ingots are hot-processed to obtain high-speed steel.

[0007] S2. Obtain multiple sets of metallographic microstructure photographs along the diagonal direction in high-speed steel to obtain carbide characteristics, and use the carbide characteristics as the target value.

[0008] S3. Simulate the hot working process of high-speed steel using finite element software, and extract the stress, strain, and temperature data from the finite element as characteristic values.

[0009] S4. Perform dimensionality reduction on the feature values ​​in multiple ways and compare them to select the optimal method to obtain the best dimensionality reduction data result;

[0010] S5. Segment the dimensionality reduction data results, model and predict the segmented data using various machine learning algorithms, and evaluate the error of multiple prediction results based on the target value to obtain the optimal prediction result.

[0011] In some implementations, in S1, the smelting technology includes casting, electroslag remelting, powder metallurgy, and spray forming, and the hot working includes forging, rolling, and extrusion.

[0012] In some embodiments, in S2, multiple sets of metallographic micrographs are obtained at equal intervals with a measurement error ≤0.2μm, and the metallographic structure is etched at room temperature for 20-35s using a 5-10% nitric acid alcohol solution.

[0013] In some embodiments, in S2, the carbide characteristics include the network carbide wall thickness, carbide content, and maximum carbide size.

[0014] In some embodiments, obtaining multiple sets of metallographic micrographs along the diagonal direction in high-speed steel to obtain characteristics of carbides includes:

[0015] Multiple sets of metallographic microstructure images were acquired to obtain the wall thickness of the network carbides. The metallographic microstructure images were processed for grayscale and binarization using Python. Subsequently, the wall thickness was measured multiple times using Image-Pro plus software. The measured values ​​were cleaned to remove invalid data, and the average value was calculated as the target value.

[0016] In some implementations, in S4, the dimensionality reduction process includes removing any one of stress, strain, and temperature, or averaging the temperature data, or averaging the stress and temperature data.

[0017] In some implementations, segmenting the dimensionality reduction data results in S5 includes:

[0018] The dimensionality reduction data results are divided into a test set and a training set, with the test set accounting for 15%–30% and the training set accounting for 70%–85%.

[0019] In some implementations, modeling and predicting segmented data using various machine learning algorithms includes:

[0020] The training set data is used to train and model predictions using various machine learning algorithms, while the test set data is not used for training.

[0021] In some implementations, in S5, various machine learning algorithms include random forest, Bayesian algorithm, classification and regression tree algorithm, gradient boosting decision tree algorithm, minimum absolute shrinkage and selection algorithm, nearest neighbor algorithm, ridge regression algorithm and support vector machine algorithm.

[0022] In some implementations, in S5, error evaluation includes using parameter R. 2 The coefficient of determination and mean squared error (MSE) are used for evaluation.

[0023] The present invention has at least the following beneficial technical effects:

[0024] 1. This invention effectively utilizes limited metadata for analysis, accurately predicting the characteristics of carbides in high-speed steel;

[0025] 2. This invention constructs an accurate model by combining machine learning, finite element method, and experimentation. After process optimization, the carbide characteristics are significantly improved.

[0026] 3. By using this invention to optimize the hot working process, the cost of "trial and error" can be significantly reduced by 35% to 65%, the optimization efficiency can be improved, and good economic and social benefits can be achieved. Attached Figure Description

[0027] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other embodiments can be obtained based on these drawings without creative effort.

[0028] Figure 1 A schematic diagram illustrating an embodiment of the method for predicting carbide characteristics during hot working of high-speed steel provided by the present invention;

[0029] Figure 2 A diagram illustrating the method for measuring the wall thickness of network carbides provided by this invention;

[0030] Figure 3 A comparison chart of predicted values ​​and test results using the dimensionality reduction method provided by this invention;

[0031] Figure 4 A comparison chart of predicted values ​​and test results from different machine learning models provided in this invention;

[0032] Figure 5 R corresponding to the different machine learning models provided in this invention 2 And MSE result graph;

[0033] Figure 6 Further verification results of the support vector machine algorithm and random forest model provided in this invention are shown in the figure.

[0034] Figure 7 The image shows the verification results of the Bayesian algorithm provided by this invention. Detailed Implementation

[0035] To make the objectives, technical solutions, and advantages of the present invention clearer, the embodiments of the present invention will be further described in detail below with reference to specific examples and the accompanying drawings.

[0036] The terms "comprising" and "having," and any variations thereof, used in the specification, claims, and accompanying drawings of this invention are intended to cover non-exclusive inclusion; the terms "first," "second," etc., used in the specification, claims, and accompanying drawings are used to distinguish different objects, not to describe a particular order. "A plurality of" means two or more, unless otherwise explicitly specified.

[0037] Furthermore, the reference to "embodiment" herein means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of the invention. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment mutually exclusive with other embodiments. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0038] Given the existing technological challenges, effectively utilizing data such as finite element analysis to optimize the hot working process of high-speed steel has become crucial for improving carbide distribution. Machine learning is a powerful tool for processing large amounts of data. Analyzing finite element simulation data during the forging process using machine learning algorithms can not only accelerate data analysis but also predict target values ​​by changing the initial input conditions based on existing data, thereby guiding the optimization of the hot working process.

[0039] like Figure 1 The diagram shown is an embodiment of the method for predicting carbide characteristics during hot working of high-speed steel provided by the present invention, including the following steps:

[0040] S1. Based on the standard grade composition of high-speed steel, the raw materials are batched and smelted to obtain ingots, and the ingots are hot-processed to obtain high-speed steel.

[0041] S2. Obtain multiple sets of metallographic microstructure photographs along the diagonal direction in high-speed steel to obtain carbide characteristics, and use the carbide characteristics as the target value.

[0042] S3. Simulate the hot working process of high-speed steel using finite element software, and extract the stress, strain, and temperature data from the finite element as characteristic values.

[0043] S4. Perform dimensionality reduction on the feature values ​​in multiple ways and compare them to select the optimal method to obtain the best dimensionality reduction data result;

[0044] S5. Segment the dimensionality reduction data results, model and predict the segmented data using various machine learning algorithms, and evaluate the error of multiple prediction results based on the target value to obtain the optimal prediction result.

[0045] Furthermore, in S1, the smelting technology includes casting, electroslag remelting, powder metallurgy, and spray forming technology, and the hot working includes forging, rolling, and extrusion.

[0046] Furthermore, in S2, carbide characteristics include network carbide wall thickness, carbide content, and maximum carbide size. Obtaining multiple sets of metallographic microstructure images along the diagonal direction in high-speed steel to acquire carbide characteristics includes: acquiring multiple sets of metallographic microstructure images to obtain network carbide wall thickness; performing grayscale and binarization processing on the metallographic microstructure images using Python; subsequently performing multiple wall thickness measurements using Image-Pro plus software; cleaning the measured values ​​to remove invalid data; and then calculating the average value as the target value.

[0047] Furthermore, in S4, the dimensionality reduction process includes removing any one of stress, strain, or temperature, or averaging the temperature data, or averaging the stress and temperature data.

[0048] Further, in S5, segmenting the dimensionality reduction data results includes dividing the dimensionality reduction data results into a test set and a training set, with the test set accounting for 15%–30% and the training set accounting for 70%–85%. In some embodiments, modeling and predicting the segmented data using multiple machine learning algorithms includes training and modeling the prediction using the training set data using multiple machine learning algorithms, while the test set data is not used for training. Furthermore, the algorithms used include Random Forests (RF), Naive Bayes (Bayes), Classification and Regression Tree (CART), Gradient Boosting Decision Tree (GBDT), Least Absolute Shrinkage and Selection Operator (LASSO), K-Nearest Neighbor (KNN), Ridge Regression (Ridge), and Support Vector Regression (SVR). Error evaluation includes using parameters R... 2 The coefficient of determination and mean squared error (MSE) are used for evaluation.

[0049] The present invention will be further explained below with reference to specific embodiments.

[0050] Example 1:

[0051] The tools used in this embodiment mainly include ABAQUS, a metallographic microscope, Image-Pro plus, and Python. The material used is M42 high-speed steel obtained by electroslag remelting, and its composition is shown in Table 1. No. 1 represents the standard composition range, and No. 2 represents the actual measured composition. The hot working process is forging, with an initial forging temperature of 1050-1070℃ and a final forging temperature above 950℃. The carbide characteristic is a network carbide wall thickness.

[0052] Table 1. Chemical composition (wt.%) of specific embodiments of the present invention

[0053]

[0054] Two sets of samples were taken diagonally from the forged sample, each measuring 100mm × 10mm × 5mm. After sanding and polishing, the samples were etched with a 5-10% nitric acid alcohol solution at room temperature for 20-35 seconds, for a total of 60 sets, spaced 3.5mm apart, with 10 images per set. The wall thickness was then measured. The specific steps were as follows: the metallographic images were processed for grayscale and binarization using Python, followed by wall thickness measurement using Image-Pro Plus software. Ten images were measured per set, and the measurement was performed at least 100 times. The average value was then taken as the target value. Figure 2 This is a method for measuring wall thickness.

[0055] Due to the sample preparation process, if there is dirt on the sample surface, the data for that part needs to be deleted. Furthermore, the impact of wall thickness on performance is mainly on the thicker portions; thicknesses less than 8 μm are not considered.

[0056] The forging process of M42 high-speed steel was accurately simulated using the finite element software ABAQUS. 60 sets of stress, strain and temperature data (corresponding to 2 sets of 100mm×10mm×5mm samples) were extracted from the finite element data as feature values. Each set of data contains 66 features, which are the peak stress, strain and temperature in 22 passes.

[0057] The relationship between the data extracted from the qualitative analysis and the wall thickness is used to make predictions.

[0058] With 66 feature dimensions being too high, the following dimensionality reduction methods were adopted to effectively perform machine learning modeling:

[0059] (1) Removing stress can reduce the feature dimension to 44;

[0060] (2) Removing strain can reduce the feature dimension to 44;

[0061] (3) Removing the temperature can reduce the feature dimension to 44;

[0062] (4) The stress and temperature are averaged, and the strain of 22 passes is retained, reducing the characteristic dimension to 24;

[0063] (5) The average temperature is taken, the strain is taken as the full strain of 22 passes, the stress of 22 passes is retained, and the characteristic dimension is reduced to 24.

[0064] Figure 3 The results are the prediction and test results of the dimensionality reduction method (5), and the results show that the above scheme is feasible.

[0065] Machine learning modeling was performed for the five methods mentioned above. The data was randomly divided into a test set and a training set, with the test set accounting for 25% of the total data and the training set accounting for 75% of the total data.

[0066] Machine learning algorithm modeling includes RF algorithm, Bayes algorithm, CART algorithm, GBDT algorithm, LASSO algorithm, KNN algorithm, Ridge algorithm and SVR algorithm.

[0067] A Python program was written to perform modeling and prediction using test set data. The feasibility and accuracy of each model were analyzed, and the fifth dimensionality reduction method was finally selected, which is "taking the average temperature, taking the strain from 22 full-process strains, retaining the stress from 22 stresses, and reducing the feature dimension to 24". Figure 4 These are the results of modeling and prediction using different algorithms.

[0068] Through R 2 Comparing and evaluating the candidate algorithms with MSE, the candidate modeling algorithms were determined to be RF and SVR, such as... Figure 5 .

[0069] The prediction performance of the SVR and RF models was validated separately, such as... Figure 6 The SVR model's predictions better match the test results, with an error of less than 10%. Therefore, the SVR model is selected as the final prediction model for the forging process parameters.

[0070] Table 2 shows a search space of size 3,669,984 for stress, strain, and temperature, constructed by varying stress, strain, and average temperature. A linear kernel support vector regression model was used to predict all process combinations within the search space, yielding partial predicted wall thickness values. Taking a tested wall thickness of 8.6 μm as an example, the results show that under different stress combinations at a total strain of 0.8 and an average temperature of 1136.85 °C, the wall thickness decreased to 7.67 μm.

[0071] Table 2. Partial Predicted and Tested Values ​​within the Search Space of Specific Embodiment 1 of the Invention

[0072]

[0073] Example 2

[0074] The material used in Example 2 is electroslag remelted M42 high-speed steel, and the forging process is the same as in Example 1. Example 2 focuses on the rolling process, in which the square billet obtained by forging is rolled into a bar with a diameter of 65mm. The rolling process is simulated by the finite element software Deform. The initial rolling temperature is 1010-1030℃, the final rolling temperature is 850-1000℃, and the target parameter is the maximum carbide particle size.

[0075] Two sets of samples were taken along the diameter of the rolled sample, with sample dimensions of 30mm × 10mm × 5mm. After sanding and polishing, the samples were etched with a 5-10% nitric acid alcohol solution at room temperature for 20-35 seconds, for a total of 30 sets, spaced 1mm apart, with 10 images per set. Subsequently, the maximum carbide particle size was determined. The specific steps were as follows: the metallographic images were processed for grayscale and binarization using Python, and then the area of ​​all carbide particles was counted using Image-Pro plus software and converted to diameter according to the equivalent circle. The average of the maximum values ​​in the 10 sets of images was taken as the target value.

[0076] If there is dirt on the surface of the sample due to the sample preparation process, that part of the data needs to be deleted.

[0077] The rolling process of M42 high-speed steel was accurately simulated using the finite element software Deform. Sixty sets of finite element data (corresponding to two sets of 30mm×10mm×5mm samples) of stress, strain and temperature were extracted as feature values. Each set of data contains 39 features, which are the peak stress, strain and temperature in 13 passes.

[0078] Qualitative analysis is used to analyze the relationship between the extracted data and the maximum carbide particle size to make predictions.

[0079] With 39 feature dimensions being too high, the following dimensionality reduction methods were adopted to effectively perform machine learning modeling:

[0080] (1) Stress removal can reduce the feature dimension to 26;

[0081] (2) Removing strain can reduce the feature dimension to 26;

[0082] (3) Removing the temperature can reduce the feature dimension to 26;

[0083] (4) The stress and temperature are averaged, and the strain of 13 passes is retained, and the characteristic dimension is reduced to 15.

[0084] (5) The average temperature is taken, the strain is taken as the full strain of 13 passes, the stress of 13 passes is retained, and the characteristic dimension is reduced to 15.

[0085] Machine learning models were developed for the five methods described above, and the data was randomly divided into a test set and a training set, with the test set accounting for 25% of the total data and the training set accounting for 75%.

[0086] Machine learning algorithm modeling includes RF algorithm, Bayes algorithm, CART algorithm, GBDT algorithm, LASSO algorithm, KNN algorithm, Ridge algorithm and SVR algorithm.

[0087] A Python program was written to perform modeling and prediction using test set data. The feasibility and accuracy of each model were analyzed, and the dimensionality reduction method was finally selected as "averaging the temperature, taking the strain from 13 full-process strains, retaining the stress from 13 stresses, and reducing the feature dimension to 15". By comparing the predicted values ​​and test values ​​on the test set and the training set, Bayes was deemed the optimal candidate modeling algorithm.

[0088] Through R 2 Compared with MSE, the candidate algorithms were evaluated and Bayes algorithm was deemed the best.

[0089] Figure 7 The model's prediction performance was validated, and the predicted values ​​were very close to the test values, with an error of less than 6%.

[0090] By changing the stress, strain, and average temperature, a search space of stress, strain, and temperature with a size of 6377292 was constructed. The Bayes model was used to predict all process combinations in the search space, and the corresponding predicted values ​​of the maximum carbide particle size were obtained, as shown in Table 3 (partial).

[0091] Table 3. Partial predicted and tested values ​​within the search space of Specific Embodiment 2 of the Invention.

[0092]

[0093] The above are exemplary embodiments disclosed in this invention. However, it should be noted that various changes and modifications can be made without departing from the scope of the embodiments of this invention as defined by the claims. The functions, steps, and / or actions of the methods according to the disclosed embodiments described herein do not need to be performed in any particular order. Furthermore, although the elements disclosed in the embodiments of this invention may be described or claimed individually, they may be understood as multiple unless explicitly limited to a singular number.

[0094] It should be understood that, as used herein, the singular form “a” is intended to include the plural form as well, unless the context clearly supports an exception. It should also be understood that, as used herein, “and / or” refers to any and all possible combinations of one or more of the associated listed items.

[0095] The embodiment numbers disclosed in the above embodiments of the present invention are merely for description and do not represent the superiority or inferiority of the embodiments.

[0096] Those skilled in the art should understand that the discussion of any of the above embodiments is merely exemplary and is not intended to imply that the scope of the invention (including the claims) is limited to these examples. Within the framework of the invention, technical features of the above embodiments or different embodiments can be combined, and many other variations of different aspects of the invention exist, which are not provided in the details for the sake of brevity. Therefore, any omissions, modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the invention should be included within the protection scope of the invention.

Claims

1. A method for predicting carbide characteristics during hot working of high-speed steel, characterized in that, include: S1. Based on the standard grade composition of high-speed steel, batching and smelting are carried out to obtain ingots, and the ingots are hot-processed to obtain high-speed steel; S2. Obtain multiple sets of metallographic microstructure photographs along the diagonal direction in the high-speed steel to obtain carbide characteristics, and use the carbide characteristics as target values; S3. Simulate the hot working process of high-speed steel using finite element software, and extract the stress, strain, and temperature data from the finite element as characteristic values. S4. Perform dimensionality reduction processing on the feature values ​​in multiple ways and compare them to select the optimal method to obtain the optimal dimensionality reduction data result; S5. Segment the dimensionality reduction data result, model and predict the segmented data using multiple machine learning algorithms, and evaluate the error of multiple prediction results based on the target value to obtain the optimal prediction result. In S2, the carbide characteristics include the network carbide wall thickness, carbide content, and maximum carbide size; Multiple sets of metallographic micrographs were taken along the diagonal direction in the high-speed steel to obtain the characteristics of the carbides, including: Multiple sets of metallographic microstructure images were obtained to obtain the wall thickness of the network carbide. The metallographic microstructure images were processed for grayscale and binarization using Python. Then, the wall thickness was measured multiple times using Image-Pro plus software. The measured values ​​were cleaned to remove invalid data, and the average value was calculated as the target value. In S4, the dimensionality reduction process includes removing any one of stress, strain, and temperature, or averaging the temperature data, or averaging the stress and temperature data.

2. The method for predicting carbide characteristics during hot working of high-speed steel according to claim 1, characterized in that, In S1, the smelting technology includes casting, electroslag remelting, powder metallurgy, and spray forming technology, and the hot working includes forging, rolling, and extrusion.

3. The method for predicting carbide characteristics during hot working of high-speed steel according to claim 1, characterized in that, In S2, the multiple sets of metallographic microstructure images are obtained at equal intervals, with a measurement error ≤0.2μm. The metallographic structure is etched at room temperature for 20~35s using a 5~10% nitric acid alcohol solution.

4. The method for predicting carbide characteristics during hot working of high-speed steel according to claim 1, characterized in that, In S5, segmenting the dimensionality reduction data results includes: The dimensionality reduction data results are divided into a test set and a training set, with the test set accounting for 15% to 30% and the training set accounting for 70% to 85%.

5. The method for predicting carbide characteristics during hot working of high-speed steel according to claim 4, characterized in that, Modeling and predicting segmented data using various machine learning algorithms includes: The training set data is used to train and model predictions using various machine learning algorithms, while the test set data is not used for training.

6. The method for predicting carbide characteristics during hot working of high-speed steel according to claim 1, characterized in that, In S5, the various machine learning algorithms include random forest, Bayesian algorithm, classification and regression tree algorithm, gradient boosting decision tree algorithm, minimum absolute shrinkage and selection algorithm, nearest neighbor algorithm, ridge regression algorithm and support vector machine algorithm.

7. The method for predicting carbide characteristics during hot working of high-speed steel according to claim 1, characterized in that, In S5, the error assessment includes evaluation using the parameters R² coefficient of determination and MSE mean square error.

Images