Method for predicting hvo f coating performance optimization of spraying parameters based on hml

Abstract

The present application relates to a kind of methods for optimizing spraying parameters based on HML to predict HVOF coating performance, belong to coating technical field.The method includes: S1, preparation experimental data: carry out supersonic flame spraying experiment, obtain spraying parameter and coating performance result;S2, establish data set: spraying parameter is regarded as input, and coating performance result is regarded as output, forms data set;S3, construct layered machine learning model, by training, intermediate variable is linked with coating performance, and only containing input performance prediction formula is obtained;S4, model evaluation is carried out to the layered machine learning model of training completion, judge whether it meets preset accuracy;S5, the spraying parameter to be measured is input to the layered machine learning model of training completion, and coating performance result is output by the training of model.The method requires less number of data set, can obtain good prediction quality under limited data set, effectively reduces experimental cost.

Description

Technical Field

[0001] This invention belongs to the field of coating technology, and in particular relates to a method for optimizing spraying parameters based on HML prediction of HVOF coating performance. Background Technology

[0002] Hypersonic flame spraying (HVOF) is a thermal spraying technology used to protect or improve the surface properties of workpiece substrates. It offers advantages such as good wear resistance, high deposition efficiency, and low porosity, and is widely used for surface protection of equipment. The spraying process involves chemical reactions, turbulent compressible flow with multiphase interactions, subsonic / supersonic transitions, droplet deformation, and solidification. In this process, a completely mixed fuel—oxygen gas—and microparticles are injected into the gun chamber, undergoing a combustion reaction that generates a thermosonic / supersonic multiphase gas flow. Microparticles of metals, alloys, and / or ceramics are accelerated, heated, and ultimately deposited onto the substrate at high speed. The coating formation process includes stages such as sputtering of molten or semi-molten particles, sintering of deposited particles on the substrate, solidification, and deformation. Due to the complexity of the spraying process, exploring strategies to optimize coating performance remains an ongoing challenge.

[0003] With advancements in materials and machinery, experts and scholars have begun to focus on machine learning in order to design optimal coating performance parameters for spraying processes. Compared to traditional methods, it can significantly reduce human resources. Furthermore, machine learning can accurately predict the wear resistance of coatings and adjust spraying process parameters according to the application scenario to obtain coatings with the desired performance. Hierarchical machine learning (HML) is a hybrid physical-statistical machine learning method developed on small experimental datasets. In this method, the predictor is connected to the system response through an intermediate layer whose variables are parameterized by known physicochemical relationships from domain knowledge related to the system. Regression techniques are used to connect the intermediate layer to the system response for prediction and optimization.

[0004] Currently, existing patents have improved coating performance by optimizing spraying parameters through experimental comparison. Patent CN 101713059 A utilizes plasma spraying technology to spray a working coating and employs orthogonal design to optimize spraying parameters. Based on the different properties of the sprayed materials, it obtains the optimal parameter combination, resulting in a wear-resistant coating with higher hardness and lower porosity. However, this invention only proposes the optimal parameters for Ni / Al coatings, and the obtained parameters are derived from a large number of experimental comprehensive evaluations. The method has a limited scope of application, high experimental costs, and lacks practical application value. Introducing machine learning, patent CN 105117599 A constructs a BP neural network model and uses test samples to test the BP neural network model optimized by a genetic algorithm, obtaining a high-alumina bronze coating behavior prediction model. This model has high prediction accuracy and improves the research efficiency of high-alumina bronze coatings. However, the BP neural algorithm used in this invention has low correlation with the process itself, and the training samples are few, making the model prone to overfitting and limiting its practical application value. Therefore, existing methods for predicting coating performance and optimizing spraying parameters suffer from problems such as high cost, low application value, and easy overfitting of models. There is an urgent need to develop a new method for predicting HVOF coating performance and optimizing spraying parameters. Summary of the Invention

[0005] Therefore, the technical problem to be solved by the present invention is to overcome the problems of high cost, low application value and easy overfitting of existing methods for predicting coating performance and optimizing spraying parameters.

[0006] To address the aforementioned technical problems, this invention provides a method for optimizing spraying parameters based on HML prediction of HVOF coating performance. Utilizing hierarchical machine learning (HML) technology, a method is proposed that can predict the performance of supersonic flame spraying (HVOF) coatings and thereby optimize spraying parameters. This reduces the time spent manually exploring spraying parameters, improves work efficiency, and achieves the desired coating performance.

[0007] The purpose of this invention is to provide a method for optimizing spraying parameters based on HML prediction of HVOF coating performance, comprising the following steps:

[0008] S1. Preparation of experimental data: Supersonic flame spraying experiments were conducted, and spraying parameters and coating performance results were obtained using the controlled variable method.

[0009] S2. Create a dataset: Use the spraying parameters described in S1 as input and the coating performance results as output to form a dataset.

[0010] S3. Based on the dataset described in S2, construct a hierarchical machine learning model:

[0011] S31. Using MATLAB software, construct the first layer of a hierarchical machine learning model;

[0012] S32. Parameterize the input quantities in the dataset described in S2 to obtain intermediate variables;

[0013] S33. Construct the second layer of a hierarchical machine learning model using regression techniques;

[0014] By training, the intermediate variables are linked to the coating performance, resulting in a performance prediction formula that only includes the input quantities.

[0015] S4. Evaluate the hierarchical machine learning model trained in S3 to determine whether it meets the preset accuracy.

[0016] If the error does not exceed the preset accuracy range, the training is complete.

[0017] If the error exceeds the preset accuracy range, adjust the parameters of the hierarchical machine learning model, continue training, and repeat S3 until the model converges to achieve the preset accuracy range of the hierarchical machine learning model, then the training is complete.

[0018] S5. Input the spraying parameters to be tested into the hierarchical machine learning model trained in S4. Through model training, output the coating performance results.

[0019] In one embodiment of the present invention, in S1, the spraying parameters are: spraying distance of 150-400 mm; oxygen flow rate of 150-300 slpm; and methane flow rate of 100-250 slp.

[0020] In one embodiment of the present invention, in S1, the number of groups of spraying parameters is 40-50.

[0021] In one embodiment of the present invention, in S3, the input quantities include spraying distance, oxygen flow rate, and methane flow rate.

[0022] In one embodiment of the present invention, in S3, the intermediate variable is a set of generalized physical equations used to correlate the physical interactions and processes between the experimental predictor and the system output, including particle flight velocity and particle flight temperature; the particle flight velocity is 200-600 m / s; the particle flight temperature is 1500-3000 K.

[0023] In one embodiment of the present invention, the parameterization method for particle flight velocity and particle flight temperature is as follows:

[0024] (1) Formula for calculating particle flight speed:

[0025]

[0026] Where, m p v is the particle mass. pv is the particle velocity. g ρ is the gas velocity. g A is the gas density; p C is the projected area of ​​the particle on a plane perpendicular to the flow direction; D The drag coefficient is the effect of particle shape.

[0027] (2) Formula for calculating particle flight temperature:

[0028]

[0029] Where, m p T is the particle mass. p A' represents the particle temperature. p c is the particle surface area; p For particle heat capacity, T g Let t be the gas temperature and h be the heat transfer coefficient.

[0030] In one embodiment of the present invention, in S3, the regression technique is the Lasso regression method.

[0031] In one embodiment of the present invention, the Lasso regression is performed using a linear function f(x) = ω. T To fit a set of data D = {(x1,y1),(x2,y2),...,(x+b}, use x+b to fit the data. n ,y n And cause losses Minimum. Lasso regression modifies J based on standard linear regression, specifically by adding L1 regularization to minimize it. The goal is to minimize the weights, ultimately constructing a model with relatively small parameters. ω represents the coefficients of the features (parameters of x). The regularization term restricts these coefficients; L1 regularization refers to the sum of the absolute values ​​of all elements in the weight vector ω, typically represented as ||ω||1. A coefficient λ is added before the regularization term; this coefficient is an initial value set beforehand and then determined by parameter tuning based on feedback.

[0032] In one embodiment of the present invention, in S3, the second layer of the hierarchical machine learning model represents the response of the complex system to be optimized, which is the output quantity; the output quantity is a performance index, including hardness, porosity and wear resistance.

[0033] In one embodiment of the present invention, in S4, the model evaluation is performed by comparing the R-values ​​of the model predictions with the observed data. 2 The value is used to evaluate the accuracy of the prediction.

[0034] In one embodiment of the present invention, R 2 The calculation formula is as follows:

[0035]

[0036] Among them, y i Represents the actual observed value, y 2 This represents the average of the actual observed values. This represents the predicted value.

[0037] In one embodiment of the present invention, in S4, the adjustment of the model parameters includes: refining the supersonic flame spraying parameters, adjusting the physical equations of the hierarchical machine learning model, and increasing the number of training iterations, thereby further adjusting the HML model parameters until the error is within the preset accuracy range, at which point the training is complete.

[0038] In one embodiment of the invention, in S5, the coating performance results include hardness, porosity, and abrasion resistance.

[0039] The technical solution of the present invention has the following advantages compared with the prior art:

[0040] (1) The method described in this invention has high prediction accuracy and low cost. It utilizes the physical equations of supersonic flame spraying to construct the first layer of a hierarchical machine learning (HML) model. Since the HML model can connect system variables and reactions using physical relationship equations under limited dataset conditions, it can effectively reduce the need for iterative testing and improve prediction accuracy and efficiency. In addition, the established HML model requires less data and can obtain good prediction quality under limited dataset conditions, effectively reducing experimental costs.

[0041] (2) The method described in this invention can predict coating performance more efficiently. By using the spraying parameters and coating performance results obtained from supersonic flame spraying as HML model dataset, the spraying parameters are linked to the coating performance to construct an HML model. By inputting the spraying parameters (spraying distance, oxygen flow rate, methane flow rate) into the trained HML model, the time for manually exploring the spraying parameters can be effectively reduced, thereby improving work efficiency and achieving better coating performance. Attached Figure Description

[0042] To make the content of this invention easier to understand, the invention will be further described in detail below with reference to specific embodiments and accompanying drawings, wherein:

[0043] Figure 1 This is a flowchart of the method for optimizing spraying parameters based on HML prediction of HVOF coating performance according to the present invention.

[0044] Figure 2 R is the prediction data for Embodiment 1 and Comparative Example 1 of the present invention. 2 Value comparison chart. Detailed Implementation

[0045] The present invention will be further described below with reference to the accompanying drawings and specific embodiments, so that those skilled in the art can better understand and implement the present invention. However, the embodiments described are not intended to limit the present invention.

[0046] Example 1

[0047] Reference Figure 1 As shown, a method for optimizing spraying parameters based on HML prediction of HVOF coating performance specifically includes the following steps:

[0048] S1. Experimental data preparation: Using high-velocity oxygen fuel (HVOF) spraying, nickel-chromium composite powder was sprayed onto the surface of a nickel substrate and tested to obtain 40 sets of spraying parameters and coating performance results.

[0049] The supersonic flame spraying parameters for the nickel-chromium coating, i.e., the input parameters, are: spraying distance of 150-400 mm; oxygen flow rate of 150-300 slpm; and methane flow rate of 100-250 slp. The coating performance results are hardness, porosity, and wear resistance.

[0050] S2. Dataset Establishment: Using the 40 sets of spraying parameters obtained in S1 as inputs and the coating performance results as outputs, a dataset is formed. The dataset mainly includes spraying parameters (spraying distance, oxygen flow rate, methane flow rate) and coating performance results (coating hardness, porosity, abrasion resistance). 70%, 15%, and 15% of these parameters are randomly selected as the training set, validation set, and test set for the model, respectively; that is, the training set consists of 28 sets, the validation set consists of 6 sets, and the test set consists of 6 sets.

[0051] S3. Based on the dataset described in S2, construct an HML model:

[0052] S31, First Layer of HML Model (Input Variables → Intermediate Variables): Using MATLAB software, construct the first layer of a hierarchical machine learning HML model;

[0053] S32. Intermediate Variables (Intermediate Layer Design): The spraying parameters in the dataset are used as inputs to parameterize the physicochemical relationships, resulting in intermediate variables. These intermediate variables include particle velocity and particle temperature. The particle velocity is 200-600 m / s; the particle temperature is 1500-3000 K. The parameterization methods for particle velocity and particle temperature are as follows:

[0054] (1) Formula for calculating particle flight speed:

[0055] Where, m p v is the particle mass. p v is the particle velocity. g ρ is the gas velocity.g A is the gas density; p C is the projected area of ​​the particle on a plane perpendicular to the flow direction; D The drag coefficient is the effect of particle shape.

[0056] (2) Formula for calculating particle flight temperature:

[0057] Where, m p T is the particle mass. p A' represents the particle temperature. p c is the particle surface area; p For particle heat capacity, T g Let t be the gas temperature and h be the heat transfer coefficient.

[0058] S33, HML Model Second Layer (Intermediate Variable → Output): Construct the second layer of the hierarchical machine learning model using the Lasso regression method;

[0059] Forty sets of nickel-chromium datasets were imported into the HML model for training. Through training, intermediate variables were linked to coating performance, resulting in a performance prediction formula that only includes the input quantity.

[0060] S4. Evaluate the HML model trained in S3 using the R-squared value between the model predictions and the observed data. 2 The value is used to evaluate the prediction accuracy and determine whether it meets the preset accuracy. To avoid overfitting, the preset accuracy is 0.9.

[0061] If the error does not exceed the preset accuracy range, the training is complete.

[0062] If the error exceeds the preset accuracy range, adjust the parameters of the hierarchical machine learning model, continue training, and repeat S3 until the model converges to achieve the preset accuracy range of the hierarchical machine learning model, then the training is complete.

[0063] After 10,000 iterations of training, an HML model that meets the convergence accuracy requirement is finally obtained.

[0064] R 2 The calculation formula is as follows:

[0065] Among them, y i Represents the actual observed value, y 2 This represents the average of the actual observed values. This represents the predicted value.

[0066] S5. Coating performance test: Input the supersonic flame spraying parameters (spraying distance, oxygen flow rate, methane flow rate) of the nickel-chromium coating to be tested into the trained HML model. After iterative training of the HML model, the coating performance results (hardness, porosity, wear resistance) are obtained.

[0067] Comparative Example 1

[0068] The method for optimizing spraying parameters based on predicting HVOF coating performance using ordinary artificial neural networks (ANN) specifically includes the following steps:

[0069] S1. Experimental data preparation: Using high-velocity oxygen fuel (HVOF) spraying, nickel-chromium composite powder was sprayed onto the surface of a nickel substrate and tested to obtain 40 sets of spraying parameters and coating performance results.

[0070] The supersonic flame spraying parameters for the nickel-chromium coating, i.e., the input parameters, are: spraying distance of 150-400 mm; oxygen flow rate of 150-300 slpm; and methane flow rate of 100-250 slp. The coating performance results are hardness, porosity, and wear resistance.

[0071] S2. Dataset Establishment: The 40 sets of spraying parameters and coating performance obtained in S1 are used as the dataset. The dataset mainly includes spraying parameters (spraying distance, oxygen flow rate, methane flow rate) and coating performance results (coating hardness, porosity, abrasion resistance). 70%, 15%, and 15% of these parameters are randomly selected as the training set, validation set, and test set for the model, respectively; that is, the training set consists of 28 sets, the validation set consists of 6 sets, and the test set consists of 6 sets.

[0072] S3. Constructing the ANN Model: Using deep learning methods, construct an Artificial Neural Network (ANN) model. Import the nickel-chromium coating dataset into the ANN model and perform 10,000 iterations of training, the same as in the previous example. Test the prediction accuracy of the trained ANN model to obtain its accuracy at this point.

[0073] S4. Material Catalytic Performance Test: By inputting the supersonic flame spraying parameters of the nickel-chromium coating into the trained ANN model, the hardness, porosity, and wear resistance of the coating can also be predicted.

[0074] Test case

[0075] The prediction data from Embodiment 1 and Comparative Example 1 of the present invention were used with R... 2 The value is used to evaluate the prediction quality. The prediction results of HML and ANN models are compared. 2 Value test results are as follows Figure 2 As shown, by Figure 2The data shows that this invention can predict high-performance coatings. By utilizing the intermediate variables (intermediate layers) of the HML model and applying parameter relationships and physical equations, high accuracy can be achieved even with a relatively small dataset, demonstrating that the HML model is capable of predicting coating performance for supersonic flame spraying. In contrast, the R... (The sentence is incomplete and requires further context to translate accurately.) 2 The value is lower than that in Example 1, indicating that the HML model can avoid the overfitting problem of artificial neural networks (ANN) by associating the physical interaction and process between the predictor and the system output under the same amount of dataset and the same number of training iterations. It is more suitable for supersonic flame spraying and effectively improves prediction accuracy and efficiency.

[0076] Obviously, the above embodiments are merely illustrative examples for clear explanation and are not intended to limit the implementation. Those skilled in the art will recognize that other variations or modifications can be made based on the above description. It is neither necessary nor possible to exhaustively list all possible implementations here. However, obvious variations or modifications derived therefrom are still within the scope of protection of this invention.

Claims

1. A method for optimizing spraying parameters based on HML prediction of HVOF coating performance, characterized in that, Includes the following steps: S1. Preparation of experimental data: Supersonic flame spraying experiments were conducted, and spraying parameters and coating performance results were obtained using the controlled variable method. S2. Create a dataset: Use the spraying parameters described in S1 as input and the coating performance results as output to form a dataset. S3. Based on the dataset described in S2, construct a hierarchical machine learning model: S31. Using MATLAB software, construct the first layer of a hierarchical machine learning model; the first layer of the hierarchical machine learning model is the input quantity, which includes spraying distance, oxygen flow rate and methane flow rate; S32. Parameterize the input quantities in the dataset described in S2 to obtain intermediate variables; the intermediate variables include particle flight velocity and particle flight temperature; the parameterization methods for particle flight velocity and particle flight temperature are as follows: Formula for calculating particle flight speed: ； in, For particle mass; The particle velocity; For gas velocity; The density of the gas; The area of ​​the particle projected onto a plane perpendicular to the flow direction; The drag coefficient is the effect of particle shape. (2) Formula for calculating particle flight temperature: ； in, For particle mass; The particle temperature; The particle surface area; For particle heat capacity, For gas temperature, h The heat transfer coefficient; S33. Construct the second layer of a hierarchical machine learning model using regression techniques; the second layer of the hierarchical machine learning model is the output quantity; the output quantity is a performance index, including hardness, porosity, and wear resistance; By training, the intermediate variables are linked to the coating performance, resulting in a performance prediction formula that only includes the input quantities. S4. Evaluate the hierarchical machine learning model trained in S3 to determine whether it meets the preset accuracy. If the error does not exceed the preset accuracy range, the training is complete. If the error exceeds the preset accuracy range, adjust the parameters of the hierarchical machine learning model, continue training, and repeat S3 until the model converges to achieve the preset accuracy range of the hierarchical machine learning model, then the training is complete. S5. Input the spraying parameters to be tested into the hierarchical machine learning model trained in S4. Through model training, output the coating performance results.

2. The method for optimizing spraying parameters based on HML prediction of HVOF coating performance according to claim 1, characterized in that, In S1, the spraying parameters are: spraying distance of 150-400 mm; oxygen flow rate of 150-300 slpm; and methane flow rate of 100-250 slp.

3. The method for optimizing spraying parameters based on HML prediction of HVOF coating performance according to claim 1, characterized in that, In S1, the number of spraying parameters is 40-50.

4. The method for optimizing spraying parameters based on HML prediction of HVOF coating performance according to claim 1, characterized in that, In S3, the particle flight speed is 200-600 m / s; the particle flight temperature is 1500-3000 K.

5. The method for optimizing spraying parameters based on HML prediction of HVOF coating performance according to claim 1, characterized in that, In S4, the model evaluation is performed by comparing the R-values ​​of the model predictions with the observed data. 2 The value is used to evaluate the accuracy of the prediction.

6. The method for optimizing spraying parameters based on HML prediction of HVOF coating performance according to claim 1, characterized in that, In S4, the adjustment of the model parameters includes: refining the supersonic flame spraying parameters, adjusting the physical equations of the hierarchical machine learning model, and increasing the number of training iterations.

7. The method for optimizing spraying parameters based on HML prediction of HVOF coating performance according to claim 1, characterized in that, In S5, the coating performance results include hardness, porosity, and abrasion resistance.

Images