Part surface treatment oxidation process data enhancement and key parameter prediction method based on small sample data

Abstract

The invention provides a part surface treatment oxidation process data enhancement and key parameter prediction method based on small sample data. Data enhancement and key parameter prediction are carried out on anodic oxidation data, and finally a prediction result is obtained. Firstly, data enhancement is carried out through quadratic B-spline interpolation and a generative adversarial network (GAN), and the problem that the weight data of an oxide film is insufficient to establish a prediction model is effectively solved; then, in order to evaluate the quality of the data in the enhanced anodic oxidation process, a comprehensive evaluation index (CEI) using Kullback-Leibler (KL) divergence and a mean square error (MSE) is provided; and finally, in consideration of complex coupling characteristics of a plurality of process parameters in the anodic oxidation process, an attention mechanism (AM) and a random forest (RF) are utilized to construct a film weight prediction model of the oxidation film, so that enhanced anodic oxidation process data are fully utilized, and the prediction performance is enhanced. According to the method, the problem that an accurate prediction model is difficult to establish due to insufficient sample data and complex coupling characteristics of multiple influence factors is avoided to a certain extent. The prediction precision of the key parameters of the small sample data in the industrial process is improved.

Description

Technical Field

[0001] The present invention relates to component industry data prediction, and in particular to a component surface treatment oxidation process data enhancement and key parameter prediction method based on small sample data. Background Art

[0002] With the deep integration of industrialization and informatization, the aviation industry process is developing towards high efficiency, intelligence and integration. The surface treatment process of aviation parts, such as anodizing, directly affects the performance and life of the parts. However, due to insufficient sample data and the complex coupling characteristics of multiple process parameters, it is difficult to accurately predict the key parameters of the oxidation process, which in turn affects the quality and production efficiency of the product. Traditional surface treatment oxidation process data analysis methods are often difficult to establish effective prediction models due to insufficient sample size. In addition, there is a complex coupling relationship between multiple process parameters in the anodizing process, making it difficult for a single parameter prediction method to accurately reflect the actual situation. These problems limit the optimization and improvement of the oxidation process, which in turn affects the quality and reliability of aviation parts. Summary of the Invention

[0003] The present invention provides a method for data enhancement and key parameter prediction of the surface treatment oxidation process of parts based on small sample data. By performing data enhancement on the anodizing data and predicting key parameters, a prediction result is finally obtained. For the prediction of the weight of the oxide film of a small sample, the present invention proposes a method based on data enhancement and key parameter prediction, which solves the problem of decreased accuracy of key parameter prediction caused by insufficient sample data and the complex coupling characteristics of multiple process parameters. We first interpolate the film weight of the original data using the quadratic B-spline interpolation method, and use GAN to further increase the data volume of the interpolated data. The comprehensive evaluation indicators of KL divergence and MSE are used to evaluate the quality of the enhanced anodizing process data. Then, the attention mechanism is used to extract important influencing factor information, which helps to avoid the interference of unimportant and unnecessary influencing factors on the film weight prediction model. Finally, based on the extraction of important influencing factor information, the random forest method is used to establish an oxide film weight prediction model. Through the above method, we have enhanced the prediction performance of key parameters of small sample data.

[0004] The technical solution adopted by the present invention to achieve the above-mentioned purpose is: a method for data enhancement and key parameter prediction of the surface treatment oxidation process of parts based on small sample data, comprising the following steps:

[0005] Step 1: Collect data on the surface treatment oxidation process of parts;

[0006] Step 2: interpolate the anodizing process data and expand the data points;

[0007] Step 3: Use generative adversarial networks to further enhance the data;

[0008] Step 4: Evaluate the quality of the enhanced anodizing process data;

[0009] Step 5: Use the attention mechanism to extract important influencing factor information to avoid interference of non-influencing factors on the membrane weight prediction model;

[0010] Step 6: Build a random forest model to predict the oxide film weight and output the prediction results.

[0011] The component surface treatment oxidation process data includes impact data, non-impact data, and film weight data at each sampling time point during the surface treatment oxidation process; wherein, the impact data includes sulfuric acid concentration, temperature, voltage, and time; and the non-impact data includes sodium borate concentration, malic acid concentration, and cerium inorganic salt aqueous solution concentration.

[0012] The influencing data is expanded according to the preset step size, and the film weight data of the anodizing process is interpolated using the quadratic B-spline interpolation method. The formula is:

[0013]

[0014] in, Represents the tth value of the interpolated data, x t-1,m , x t,m and x t+1,m They represent the (t-1), t and (t+1)th membrane weight values before interpolation respectively.

[0015] The method of further enhancing the data by using a generative adversarial network includes: iteratively enhancing and outputting the input oxidation process impact data with an optimization function as a goal;

[0016] The optimization objective function is:

[0017]

[0018] in, Indicates the interpolated impact data and film weight data, represents the expectation of sampling from the real data distribution, Z represents random noise, G(Z) represents the data generated by the generator using Z, and D(G(z)) represents the probability that the discriminator judges that the data generated by the generator is real. represents the expectation of sampling from the noisy data distribution, P Z (Z) and They respectively represent the data distribution of noise, the data affected by the anodic oxidation process after network enhancement, and the data distribution of film weight.

[0019] The KL divergence and mean square error (MSE) are used to establish a comprehensive evaluation index for evaluation, which includes the following steps:

[0020] Step 4-1: Calculate the MSE and KL divergence values of the enhanced data and the data before and after enhancement respectively. The formula is:

[0021]

[0022] in, Represents the enhanced data The i-th sample in M represents The number of samples in x j represents the jth sample in the data X before enhancement;

[0023]

[0024] Among them, P X and Represents the data before enhancement X and the data after enhancement respectively The probability distribution, P X (x i )and Represents X and Medium x i The probability distribution of

[0025] Step 4-2: Calculate the comprehensive evaluation index CEI of MSE and KL divergence, formula:

[0026]

[0027] Where α∈0,1] represents the balance coefficient between MSE and KL divergence;

[0028] Step 4-3, Settings If it is not within the threshold range, return to step 3 to enhance the data again until The threshold requirement is met.

[0029] The method of extracting important influencing factors using the attention mechanism includes the following steps:

[0030] Step 5-1. Define the attention mechanism network model including query, key and value, define the dataset as input, and output the data X calculated by the attention mechanism. weight ;

[0031] Step 5-2: Select mean square error (MSE) as the loss function and use the optimization algorithm to train the model.

[0032] The attention mechanism network model calculation includes:

[0033] Will be from The feature matrix F extracted from [1] is converted into a key matrix K and a value matrix V through linear transformation. The attention score is calculated using the query vector q and the key matrix K according to formula (6):

[0034] e l =score(q,k l )(6)

[0035] Among them, e l represents the lth attention score, k l represents the lth key vector;

[0036] The attention score is processed by the alignment layer, and the attention weight is calculated according to formula (7):

[0037] a l =align(e l ;e)(7)

[0038] Among them, e represents the attention score vector, a l Represents the attention weight corresponding to the l-th value vector;

[0039] Use the attention weights and value vectors to calculate the weighted feature X according to formula (8) weight ,formula:

[0040]

[0041] Among them, v l represents the lth value vector, n f is the number of eigenvectors extracted from the original data.

[0042] The random forest method is used to predict the oxide film weight, which includes the following steps:

[0043] Step 6-1: Given training data (X weight ,y train ), when training the t-th decision tree, select a sample subset X t and a target membrane weight y t , use these data to build a decision tree h t ;

[0044] Step 6-2: Given new input data X, each decision tree predicts X and obtains the predicted film weight according to formula (9): formula:

[0045]

[0046] Step 6-3: The final prediction value of the random forest model is calculated as the average of all decision tree prediction results according to formula (10) Output as the predicted result of the final film weight. Formula:

[0047]

[0048] Where T is the number of decision trees in the random forest.

[0049] The present invention has the following beneficial effects and advantages:

[0050] The proposed small-sample data augmentation method based on quadratic B-spline interpolation and generative adversarial networks can effectively address the difficulty in establishing prediction models due to insufficient data. The proposed prediction model based on an attention mechanism and random forests enhances the prediction performance of key parameters. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] Figure 1 is a flow chart of the method of the present invention;

[0052] Figure 2(a) is a comparison of the original data and the interpolated data;

[0053] Figure 2(b) shows the comparison between the original data and the data after interpolation + GAN enhancement;

[0054] Figure 3 This is a comparison chart of the prediction results of the present invention;

[0055] Table 1 shows the CEI indicators of different data augmentation methods;

[0056] Table 2 shows the performance comparison of different prediction models. DETAILED DESCRIPTION

[0057] To make the above-mentioned objects, features, and advantages of the present invention more readily apparent, the specific implementation methods of the present invention are described in detail below with reference to the accompanying drawings. The following description sets forth many specific details to facilitate a full understanding of the present invention. However, the present invention can be implemented in many other ways than those described herein, and those skilled in the art may make similar modifications without departing from the scope of the invention. Therefore, the present invention is not limited to the specific implementation methods disclosed below.

[0058] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by those skilled in the art of the art to which the present invention pertains. The terms used in the specification of the invention herein are for the purpose of describing specific embodiments only and are not intended to limit the present invention.

[0059] like Figure 1 FIG. 2 is a flow chart of the method of the present invention.

[0060] The method of the present invention uses the data of the surface treatment oxidation process of aviation parts as an example to perform data processing, enhance and predict key parameters. The programming language used in the program execution steps of the present invention is not limited to MATLAB, Python, etc.

[0061] The specific steps are as follows:

[0062] Step 1: Collect data from the surface treatment oxidation process of aviation parts. Data on impact, non-impact, and film weight at each sampling time point. Impact data includes sulfuric acid concentration, temperature, voltage, and time; non-impact data includes sodium borate concentration, malic acid concentration, and the concentration of an inorganic cerium salt solution.

[0063] Step 2: Use the quadratic B-spline interpolation method to interpolate the film weight during the anodizing process to generate more data points.

[0064] The quadratic B-spline interpolation method is used to interpolate the anodizing process data to obtain The interpolation formula is as follows:

[0065]

[0066] in, represents the tth value of the interpolated membrane weight data, x t-1,m , x t,m and x t+1,m They represent the (t-1), t and (t+1)th values of membrane weight in the original data respectively.

[0067] The interpolation object is the film weight, and sulfuric acid concentration, temperature, and other variables are the independent variables of the interpolation process. Taking the experimental data on the effect of sulfuric acid concentration on the performance of the oxide film during the anodization process of 2024 aluminum alloy as an example, the sulfuric acid concentration is expanded by a step size of 0.1 using formula (1), and the film weight values are interpolated to generate the film weight at the new sulfuric acid concentration with a sample interval of 0.1. Figure 2(a) shows a comparison of the original data and the interpolated data. The calculated KL divergence value is 0.047, indicating that the interpolated data has very little fluctuation compared to the original data.

[0068] Step 3: Taking the experimental data of the effect of temperature on the performance of the oxide film during the anodizing process of 2024 aluminum alloy as an example, for the new data after interpolation, a generative adversarial network is used to further enhance the data to meet the objective function formula (2).

[0069] The generative adversarial network augments four data pairs: sulfuric acid concentration-membrane weight, temperature-membrane weight, voltage-membrane weight, and time-membrane weight. The network outputs the augmented data. The original data represents the effects of varying sulfuric acid concentration, temperature, voltage, and time on membrane weight. When examining the effect of a single variable on membrane weight, the other variables are assumed to be constant.

[0070] The optimization goal of further enhancing data using a generative adversarial network is as follows:

[0071]

[0072] in, represents the interpolated impact data and membrane weight data, For the generator, represents the expectation of sampling from the real data distribution, Z represents random noise, G(Z) represents the data generated by the generator using Z, and D(G(z)) represents the probability that the discriminator judges that the data generated by the generator is real. represents the expectation of sampling from the noisy data distribution, P Z (Z) and They represent the data distribution of noise, the data affected by the anodic oxidation process after network enhancement, and the data distribution of film weight, respectively. The learning rate is set to 0.0002 and the adam optimizer is set to 0.5.

[0073] Figure 2(b) shows the comparison between the original data and the data after interpolation + GAN enhancement. The KL divergence value is calculated to be 0.022.

[0074] Step 4: In order to evaluate the quality of the enhanced anodizing process data, the comprehensive evaluation indicators of KL divergence and MSE are used.

[0075] Step 4-1: Calculate the MSE and KL divergence values of the enhanced data and the data before and after enhancement respectively. The formula is as follows:

[0076]

[0077] in, Represents the enhanced data The i-th sample in M represents The number of samples in x j represents the jth sample in the data X before enhancement.

[0078]

[0079] Among them, P X and Represents the data before enhancement X and after enhancement The probability distribution, P X (x i )and Represents X and Medium x i The probability distribution of .

[0080] Step 4-2: Calculate the MSE and KL divergence comprehensive evaluation index (CEI) formula as follows:

[0081]

[0082] Where α∈[0,1] represents the balance coefficient between MSE and KL divergence. Given that the mean square error reflects the degree of difference between the enhanced anodized data and the original data and is more important for data augmentation, CEI chooses a balance coefficient of 0.2.

[0083] Step 4-3, Settings If it is not within the threshold range, return to step 3 to enhance the data again until The threshold requirement is met.

[0084] Step 5: Use the attention mechanism to extract important influencing factors to fully extract the features of the enhanced anodizing process data.

[0085] Prepare a data set: Use important influencing data, such as sulfuric acid concentration, temperature, voltage, time, and membrane weight data at each sampling time point, as well as non-influencing data such as constants that do not change numerically during the oxidation process (such as sodium borate concentration, malic acid concentration, and cerium inorganic salt aqueous solution concentration) as the data set for this step 5.

[0086] The input layer receives the standardized input data; the processing layer performs a series of linear transformations and attention mechanism calculations on the input data to highlight important features; the output layer maps the processed features to the target output, which is the final predicted value.

[0087] Step 5-1: Define an attention network model that includes a query, a key, and a value. Query, key, and value are matrices obtained by linearly transforming the input data X.

[0088] Step 5-2: Select mean square error (MSE) as the loss function and use an optimization algorithm (such as Adam optimizer) to train the model.

[0089] Step 5-3, will The feature matrix F extracted from is converted into a key matrix K and a value matrix V through linear transformation. The attention score is calculated using the query vector q and the key matrix K according to formula (6). The formula is as follows:

[0090] e l =score(q,k l ) (6)

[0091] Among them, e l represents the lth attention score, k lRepresents the lth key vector.

[0092] Step 5-4: The attention score is processed by the alignment layer and the attention weight is calculated according to formula (7). The formula is as follows:

[0093] a l =align(e l ;e) (7)

[0094] Among them, e represents the attention score vector, a l Represents the attention weight corresponding to the l-th value vector.

[0095] Step 5-5: Use the attention weight and value vector to calculate the weighted input feature X according to formula (8) weight The formula is as follows:

[0096]

[0097] Among them, v l represents the lth value vector, n f is the number of eigenvectors extracted from the original data.

[0098] Step 6: For the extracted weighted input feature X weight As input, the random forest method is used to predict the oxide film weight.

[0099] Step 6-1: Given training data (X weight ,y train ), when training the t-th decision tree, select a sample subset X t and a target membrane weight y t , use these data to build a decision tree h t .

[0100] Step 6-2: Given new input data X, each decision tree predicts X and obtains the predicted film weight according to formula (9): The following formula:

[0101]

[0102] Step 6-3: The final prediction value of the random forest model is calculated as the average of all decision tree prediction results according to formula (10) Output as the predicted result of the final film weight. The formula is as follows:

[0103]

[0104] Where T is the number of decision trees in the random forest. The parameters in this method include the learning rate set to 0.0001 and the epoch set to 300.

[0105] Comparison results of different prediction methods Figure 3 As shown, it can be seen that the prediction curve of the AM-RF method proposed in the present invention is closest to the original data, indicating that the prediction effect is the best and the prediction performance of key parameters is enhanced.

[0106] As shown in Table 1, the CEI value for the interpolation + GAN data augmentation method is lower than that for the GAN method. Furthermore, the CEI values for the interpolation + CtabGAN and interpolation + TimeGAN methods are also lower than those for the CtabGAN and TimeGAN methods, indicating that quadratic B-spline interpolation helps improve the performance of GAN data augmentation methods with limited data samples. Furthermore, for the anodizing process of aviation parts, more complex GAN methods do not necessarily produce better data augmentation results.

[0107] Table 1

[0108]

[0109]

[0110] Table 2 shows the final prediction results of the oxide film weight by each prediction model. It can be seen that the AM-RF method achieves the best prediction performance in terms of mean absolute error (MAE), mean square error (MSE), and root mean square error (RMSE) of the oxide film weight.

[0111] Table 2

[0112]

[0113]

[0114] This paper proposes a method for predicting oxide film weight on aviation components based on data augmentation and a random forest (RF) algorithm. This method is suitable for small sample data. Based on various experiments investigating various factors affecting the performance of anodic oxide films, only 27 samples, including film weight, and corresponding influencing factors such as sulfuric acid concentration, temperature, voltage, and time, were collected. First, a data augmentation method using quadratic B-spline interpolation and a generative adversarial network (GAN) was proposed to enhance the quantity and characterization of anodic oxidation process data. This method also fully utilizes the mechanistic analysis information of the influencing factors from different experiments, addressing the scarcity of original data. The quality of the enhanced anodic oxidation process data was then evaluated using the CEI (Consolidation Index)—a comprehensive metric combining KL divergence and mean squared error (MSE)—to evaluate the quality of the enhanced anodic oxidation process data. Finally, considering the complex coupling characteristics of multiple influencing factors in anodic oxidation, an oxide film weight prediction model was constructed using an attention mechanism (AM) and RF to improve prediction performance. The effect of data augmentation is clearly demonstrated in Figure 2, where the sulfuric acid concentration, film weight, and temperature are normalized. After interpolation and GAN enhancement, the data distribution more closely resembles the actual data distribution. Figure 3 The film weight was normalized in the experiment. The results of comparison with other methods showed that the AM-RF method of the present invention has better performance in predicting oxide film weight, which further verified the advantage of this method in processing small sample data.

[0115] The embodiments described above will help those skilled in the art further understand the present invention, but are not intended to limit the present invention in any way. It should be noted that a person skilled in the art may make various modifications and improvements without departing from the scope of the present invention. Such modifications and improvements are all within the scope of protection of the present invention.

Claims

1. A method for data enhancement and key parameter prediction of component surface treatment oxidation process based on small sample data, characterized by: The steps include: Step 1: Collect data on the surface treatment oxidation process of parts; Step 2: interpolate the anodizing process data and expand the data points; Step 3: Use generative adversarial networks to further enhance the data; Step 4: Evaluate the quality of the enhanced anodizing process data; Step 5: Use the attention mechanism to extract important influencing factor information to avoid interference of non-influencing factors on the membrane weight prediction model; Step 6: Build a random forest model to predict the oxide film weight and output the prediction results.

2. The method for data enhancement and key parameter prediction of component surface treatment oxidation process based on small sample data according to claim 1 is characterized in that: The component surface treatment oxidation process data includes impact data, non-impact data, and film weight data at each sampling time point during the surface treatment oxidation process; wherein, the impact data includes sulfuric acid concentration, temperature, voltage, and time; and the non-impact data includes sodium borate concentration, malic acid concentration, and cerium inorganic salt aqueous solution concentration.

3. The method for data enhancement and key parameter prediction of component surface treatment oxidation process based on small sample data according to claim 1 is characterized in that: The influencing data is expanded according to the preset step size, and the film weight data of the anodizing process is interpolated using the quadratic B-spline interpolation method. The formula is: in, represents the t-th membrane weight value of the interpolated data, x t-1,m , x t,m and x t+1,m They represent the (t-1), t and (t+1)th membrane weight values before interpolation respectively.

4. The method for data enhancement and key parameter prediction of component surface treatment oxidation process based on small sample data according to claim 1 is characterized in that: The method of further enhancing the data by using a generative adversarial network includes: iteratively enhancing and outputting the input oxidation process impact data with an optimization function as a goal; The optimization objective function is: in, Indicates the interpolated impact data and film weight data, represents the expectation of sampling from the real data distribution, Z represents random noise, G(Z) represents the data generated by the generator using Z, and D(G(z)) represents the probability that the discriminator judges that the data generated by the generator is real. represents the expectation of sampling from the noisy data distribution, P Z (Z) and They respectively represent the data distribution of noise, the data affected by the anodic oxidation process after network enhancement, and the data distribution of film weight.

5. The method for data enhancement and key parameter prediction of component surface treatment oxidation process based on small sample data according to claim 1 is characterized in that: The KL divergence and mean square error (MSE) are used to establish a comprehensive evaluation index for evaluation, which includes the following steps: Step 4-1: Calculate the MSE and KL divergence values of the enhanced data and the data before and after enhancement respectively. The formula is: in, Represents the enhanced data The i-th sample in M represents The number of samples in x j represents the jth sample in the data X before enhancement; Among them, P X and Represents the data before enhancement X and the data after enhancement respectively The probability distribution, P X (x i )and Represents X and Medium x i The probability distribution of Step 4-2: Calculate the comprehensive evaluation index CEI of MSE and KL divergence, formula: Where α∈[0,1] represents the balance coefficient between MSE and KL divergence; Step 4-3, Settings If it is not within the threshold range, return to step 3 to enhance the data again until The threshold requirement is met.

6. The method for data enhancement and key parameter prediction of component surface treatment oxidation process based on small sample data according to claim 1 is characterized in that: The method of extracting important influencing factors using the attention mechanism includes the following steps: Step 5-1. Define the attention mechanism network model including query, key and value, define the dataset as input, and output the data X calculated by the attention mechanism. weight ; Step 5-2: Select mean square error (MSE) as the loss function and use the optimization algorithm to train the model.

7. The method for data enhancement and key parameter prediction of component surface treatment oxidation process based on small sample data according to claim 6 is characterized in that: The attention mechanism network model calculation includes: Will be from The feature matrix F extracted from [1] is converted into a key matrix K and a value matrix V through linear transformation. The attention score is calculated using the query vector q and the key matrix K according to formula (6): e l =score(q,k l )(6) Among them, e l represents the lth attention score, k l represents the lth key vector; The attention score is processed by the alignment layer, and the attention weight is calculated according to formula (7): a l =align(e l ;e)(7)where e represents the attention score vector, a l Represents the attention weight corresponding to the l-th value vector; Use the attention weights and value vectors to calculate the weighted feature X according to formula (8) weight ,formula: Among them, v l represents the lth value vector, n f is the number of eigenvectors extracted from the original data.

8. The method for data enhancement and key parameter prediction of component surface treatment oxidation process based on small sample data according to claim 1 is characterized in that: The random forest method is used to predict the oxide film weight, which includes the following steps: Step 6-1: Given training data (X weight ,y train ), when training the t-th decision tree, select a sample subset X t and a target membrane weight y t , use these data to build a decision tree h t ; Step 6-2: Given new input data X, each decision tree predicts X and obtains the predicted film weight according to formula (9): formula: Step 6-3: The final prediction value of the random forest model is calculated as the average of all decision tree prediction results according to formula (10) Output as the predicted result of the final film weight. Formula: Where T is the number of decision trees in the random forest.

Images