A terahertz quantitative assessment method for protective coating thickness based on machine learning

Abstract

This invention discloses a terahertz quantitative assessment method for the thickness of protective coatings based on machine learning, comprising the following steps: Constructing a dataset based on the finite-difference time-domain method and extracting time-domain and frequency-domain correlation features for each signal group; establishing regression models using multiple machine learning algorithms, selecting the optimal prediction model through root mean square error and coefficient of determination, and training the model by selecting the optimal hyperparameters using a grid search cross-validation method; training multiple regression models by using thickness as the output parameter and key feature parameters and refractive index combination as input parameters; and obtaining the terahertz detection signal of the protective coating sample through a terahertz time-domain spectroscopy system to achieve quantitative assessment of the protective coating thickness using terahertz time-domain spectroscopy. This invention can be used for quantitative assessment of the thickness of protective coatings on ceramic matrix composite surfaces, improving the effectiveness and accuracy of terahertz time-domain spectroscopy quantitative assessment of the thickness of protective coatings on ceramic matrix composite surfaces.

Description

Technical Field

[0001] This invention is a terahertz quantitative evaluation method for the thickness of protective coatings based on machine learning, belonging to the field of nondestructive testing and evaluation. Background Technology

[0002] Ceramic matrix composites (CMCs) have a wide range of applications, including key components in aerospace fields such as missile air intakes, aircraft flaps, and rocket thrust chambers. As a crucial material used in high-temperature components, CMCs inherently possess non-uniform characteristics and inevitably encounter harsh environments such as high temperatures and exhaust gases during service. To ensure efficient and long-term service, these materials typically have a protective surface coating. However, due to various factors, coating thinning damage is prone to occur during the manufacturing and use of these materials. Therefore, improving the reliability of CMCs during service is a pressing issue, specifically requiring an effective method for detecting the thickness of the protective surface coating. Appropriate non-destructive testing methods are needed at every stage of CMC processing, production, and use, including raw material manufacturing, component assembly, and service life, to obtain information such as the thickness of the protective surface coating. This allows for effective quality control of the material, which is crucial for ensuring product quality and improving reliability during use.

[0003] For non-destructive testing of the thickness of protective coatings on ceramic matrix composite surfaces, various methods are available, including ultrasound, X-ray, and infrared thermography. Terahertz time-domain spectroscopy (TDS) offers unique advantages in this area. Compared to traditional ultrasonic or thermographic methods, it requires no coupling agent and is unaffected by the low thermal conductivity of CMC (ceramic matrix composites), allowing for deeper penetration into non-conductive, non-polar materials, a wider detection range, and better spatial resolution. Compared to X-ray methods, it has lower energy (only in the millielectron volt range), is harmless to human tissue, avoids the health effects of ionization, and avoids the limitations of X-ray shielding requirements. Therefore, existing methods face significant limitations in practical applications and struggle to meet the quantitative assessment requirements for protective coating thickness on ceramic matrix composite surfaces. On the other hand, terahertz waves exhibit good penetration in the detection of many non-metallic materials, including composites, ceramics, and rubber. Therefore, TDS is particularly suitable for comprehensive, rapid, and non-contact non-destructive testing of ceramic matrix composites. Based on this, developing a machine learning-based terahertz quantitative assessment method for protective coating thickness is essential.

[0004] This invention, based on the finite-difference time-domain method, constructs terahertz simulation models of protective coatings with different refractive indices and thicknesses as a dataset, extracting time-domain and frequency-domain correlation features for each signal group. After normalization using the min-max normalization method, important feature parameters with F-scores greater than preset standard values ​​are selected. Multiple machine learning algorithms are used to establish regression models, and the optimal prediction model is selected through root mean square error and coefficient of determination. The optimal hyperparameter result is then selected for model training using a grid search cross-validation method. Thickness is used as the output parameter, and the feature parameters combining important feature parameters and refractive index are used as input parameters to train multiple regression models, selecting the optimal machine learning model. Terahertz detection signals of the protective coating samples are obtained through a terahertz time-domain spectroscopy system. The signal features and refractive index feature parameters are input together into the trained optimal machine learning model, with thickness serving as the output parameter predicted by the optimal machine learning model, thus achieving quantitative assessment of the protective coating thickness using terahertz time-domain spectroscopy. This method can be used for quantitative assessment of the thickness of protective coatings on ceramic matrix composites, solving the problem that traditional non-destructive testing methods are insufficient to effectively obtain information on the thickness of protective coatings on ceramic matrix composites. It provides solid technical support for the detection of protective coating thickness and service condition assessment of ceramic matrix composites, and this innovative method will open up new directions for the development of non-destructive testing technology in the field of ceramic matrix composites. A terahertz-based quantitative assessment method for protective coating thickness based on machine learning has not been reported domestically or internationally. Summary of the Invention

[0005] The purpose of this invention is to address the shortcomings of existing technologies by developing a terahertz time-domain spectroscopy method for quantitatively assessing coating thickness based on machine learning algorithms. This method solves the problems of difficulty in obtaining information on the thickness of protective coatings on ceramic matrix composite surfaces and the difficulty in quantitatively assessing coating thickness using traditional non-destructive testing methods. The proposed machine learning-based terahertz quantitative assessment method for protective coating thickness can improve the quality and accuracy of quantitative assessment of protective coating thickness.

[0006] This method primarily utilizes the finite-difference time-domain (FDTD) method to construct terahertz simulation models of protective coatings with different refractive indices and thicknesses as a dataset, extracting time-domain and frequency-domain correlation features for each signal group. After normalization using the min-max normalization method, important feature parameters with F-scores greater than preset standard values ​​are selected. Multiple machine learning algorithms are employed to establish regression models, and the optimal prediction model is selected using root mean square error and coefficient of determination. The optimal hyperparameters are then selected for model training using a grid search cross-validation method. Thickness is used as the output parameter, while the combination of important feature parameters and refractive index is used as the input parameters. Multiple regression models are trained, and the optimal machine learning model is selected. Terahertz detection signals of the protective coating samples are obtained using a terahertz time-domain spectroscopy system. Signal features and refractive index features are input into the trained optimal machine learning model, with thickness serving as the output parameter predicted by the optimal machine learning model. This achieves quantitative assessment of the protective coating thickness using terahertz time-domain spectroscopy. The flowchart of this method is shown below. Figure 11 As shown.

[0007] To achieve the above objectives, the present invention adopts the following design scheme:

[0008] Based on the finite-difference time-domain method, a terahertz simulation model for protective coating detection was established using simulation software. Simulation parameters were set, and simulation results of terahertz detection signals with different media structural parameters and optical parameters were obtained. Multiple simulation models of protective coatings on ceramic matrix composite surfaces with different refractive indices were constructed, with each set of data corresponding to multiple sets of different thicknesses. Two types of features were extracted from each signal set: multiple parameters of time-domain correlation and multiple parameters of frequency-domain correlation. All feature parameters were selected for signal feature extraction. The min-max normalization method was used to normalize the feature values ​​of different dimensions. The F-scores of all feature parameters were calculated and ranked based on feature importance ranking results. Important feature parameters with F-scores greater than a preset standard value were selected as inputs for model training. Multiple machine learning algorithms were used to establish regression models for training. These regression models were trained and evaluated, and the optimal prediction model was selected using root mean square error and coefficient of determination as model evaluation metrics. A grid search cross-validation method is used to arrange and combine all hyperparameters of the model. Each combination is evaluated according to model evaluation metrics, and the best parameter combination is output. Parameter optimization is then performed on multiple machine learning models, and the optimal hyperparameter result is selected for model training. Thickness is used as the output parameter, and key feature parameters with F-scores greater than a preset standard value and feature parameters of the refractive index combination are used as input parameters to train multiple regression models. A scatter plot of the model predictions is generated, and the optimal model evaluation metric and the predicted scatter plot are selected. Figure 1The model with the highest similarity is designated as the optimal machine learning model. Terahertz detection signals of the protective coating samples are obtained using a terahertz time-domain spectroscopy system. Features are ranked and selected based on their F-scores, and important feature parameters with F-scores greater than a preset standard value are extracted from the terahertz time-domain spectral detection signals. After min-max normalization of the selected signal features, they are input along with refractive index feature parameters into the trained optimal machine learning model. Thickness is used as the output parameter predicted by the optimal machine learning model, achieving quantitative assessment of the protective coating thickness using terahertz time-domain spectroscopy.

[0009] The aforementioned terahertz simulation model for protective coating detection based on the finite-difference time-domain method and simulation software is characterized by: when simulating the protective coating on the surface of ceramic matrix composite materials, a protective coating detection model with arbitrary thickness and refractive index is designed according to the specific characteristics of the object being detected.

[0010] The aforementioned simulation model for constructing multiple sets of protective coatings on the surface of ceramic matrix composite materials with different refractive indices, each set of data corresponding to multiple sets of different thicknesses, is characterized in that: all sets of data are used as datasets for model training and testing, wherein the refractive index is used as an input parameter and the thickness is used as a predicted output parameter.

[0011] The method of extracting feature parameters for each group of signals is characterized by: extracting two types of features for each group of signals, namely multiple parameters of time-domain correlation features and multiple parameters of frequency-domain correlation features, and selecting all feature parameters for signal feature extraction.

[0012] The method of normalizing feature values ​​and calculating and sorting the F-scores of feature parameters is characterized by: normalizing feature values ​​of different dimensions by adopting the minimum-maximum normalization method; calculating and sorting the F-scores of all feature parameters according to the feature importance ranking results; and selecting important feature parameters with F-scores greater than a preset standard value as input for model training.

[0013] The method of using machine learning algorithms to build and train regression models is characterized by employing multiple machine learning algorithms, including extreme gradient boosting regression, support vector regression, random forest regression, and multilayer perceptron regression. These various regression models are trained and evaluated, and the optimal prediction model is selected using two model evaluation metrics: root mean square error and coefficient of determination.

[0014] The grid search cross-validation method described herein arranges and combines all hyperparameters of the model. Its features include: merging all parameter combinations to generate a "grid", using an exhaustive search method to search for all hyperparameters for model training, evaluating each combination according to model evaluation metrics, outputting the best parameter combination result, optimizing parameters for multiple machine learning models, and selecting the optimal hyperparameter result for model training.

[0015] The method of training multiple regression models and generating scatter plots of model predictions is characterized by: using thickness as the output parameter, and using key feature parameters with F-scores greater than a preset standard value and feature parameters combining refractive index as input parameters, and training multiple regression models such as limit gradient boosting regression, support vector regression, random forest regression, and multilayer perceptron regression respectively. A scatter plot of model predictions is generated, and the optimal model evaluation metric and the predicted scatter plot are selected. Figure 1 The model with the highest degree of similarity is denoted as the optimal machine learning model.

[0016] The aforementioned feature parameters are input into a pre-trained optimal machine learning model to achieve quantitative assessment of the thickness of the protective coating using terahertz time-domain spectroscopy. The key feature is that: terahertz detection signals of the protective coating sample are obtained through a terahertz time-domain spectroscopy system; feature importance is ranked and selected based on F-scores; and important feature parameters with F-scores greater than a preset standard value are extracted from the terahertz time-domain spectral detection signals. After performing min-max normalization on the selected signal features, they are input into the pre-trained optimal machine learning model along with the refractive index feature parameters. The thickness is used as the output parameter predicted by the optimal machine learning model.

[0017] The present invention can achieve the following effects:

[0018] 1. Establish a terahertz simulation model for protective coating detection based on the finite-difference time-domain method;

[0019] 2. Design a protective coating testing model with arbitrary thickness and refractive index based on the characteristics of the specific testing object;

[0020] 3. Construct multiple simulation models of protective coatings on the surface of ceramic matrix composites with different refractive indices, with each set of data corresponding to multiple sets of different thicknesses;

[0021] 4. Extract multiple parameters of time-domain correlation features and multiple parameters of frequency-domain correlation features for each group of signals;

[0022] 5. Normalize the eigenvalues ​​of different dimensions by adopting the min-max normalization method;

[0023] 6. Calculate and sort the F-scores of all feature parameters based on the feature importance ranking results;

[0024] 7. Multiple machine learning algorithms are used to build regression models for training, and the optimal prediction model is selected using two model evaluation metrics: root mean square error and coefficient of determination.

[0025] 8. Optimize the parameters of various machine learning models using the grid search cross-validation method, and select the optimal hyperparameter result for model training;

[0026] 9. Select the optimal model evaluation index and the predicted scatter plot. Figure 1 The model with the highest degree of consistency is denoted as the optimal machine learning model;

[0027] 10. Obtain the terahertz detection signal of the protective coating sample through a terahertz time-domain spectroscopy system, sort and select the feature importance based on the F score, and extract the important feature parameters in the terahertz time-domain spectroscopy detection signal whose F score is greater than the preset standard value.

[0028] 11. After performing minimum-maximum normalization on the selected terahertz time-domain spectral detection signal features, the results are input together with the refractive index feature parameters into the trained optimal machine learning model. The thickness is used as the output parameter predicted by the optimal machine learning model to achieve quantitative assessment of the thickness of the terahertz time-domain spectral protective coating. Attached Figure Description

[0029] Figure 1 It is a simulation modeling flowchart;

[0030] Figure 2 It is a waveform diagram of a simulated light source;

[0031] Figure 3 This is a simulated electric field intensity distribution diagram of the light source;

[0032] Figure 4 It is a coating simulation model;

[0033] Figure 5 These are simulation signal diagrams of coatings with different thicknesses;

[0034] Figure 6 These are simulation signal graphs of coatings with different refractive indices;

[0035] Figure 7 This is a graph showing the results of the feature importance ranking.

[0036] Figure 8 This is a flowchart of a machine learning algorithm;

[0037] Figure 9 This is a scatter plot comparing the predicted values ​​from different models with the measured values.

[0038] Figure 10 This is the prediction result of the optimal model;

[0039] Figure 11This is a flowchart of the implementation of this method. Detailed Implementation

[0040] The present invention will be further described below with reference to the accompanying drawings. The following embodiments are merely descriptive and not limiting, and should not be construed as limiting the scope of protection of the present invention.

[0041] Step 1: Based on the finite-difference time-domain method, establish a terahertz simulation model for the protective coating detection using simulation software. The simulation software used is FDTD Solutions, version 2018a, which performs electromagnetic field calculations based on the finite-difference time-domain method. Establishing the terahertz simulation model requires six steps, and the simulation modeling process is as follows: Figure 1 As shown. The simulated light source waveform is as follows. Figure 2 As shown, the pulse width is set to 577.35 fs, and the time delay is set to 8653.8 fs. The simulated electric field intensity distribution of the light source is as follows. Figure 3 As shown, a Gaussian pulsed light source is set up. The surface and line distribution diagrams of the electric field intensity show that the electric field intensity is high at the center of the light source and gradually decreases at the periphery. The electric field intensity distribution conforms to a Gaussian distribution.

[0042] Step 2: Construct multiple simulation models of protective coatings on the surface of ceramic matrix composites with different refractive indices. Each set of data corresponds to multiple sets of different thicknesses. The coating simulation models are as follows: Figure 4 As shown, an absorbing boundary of the PML perfectly matched layer was set around the simulation area, and a totally reflective metal substrate boundary was set on the bottom surface of the model to simulate the reflective detection of the coating. Simulation signals of some coatings with different thicknesses are shown below. Figure 5 As shown, some simulation signals of coatings with different refractive indices are as follows: Figure 6 As shown.

[0043] Step 3: Extract signal feature parameters, normalize them, and calculate and sort the F-scores. Select material parameters to extract multiple time-domain and frequency-domain correlation features for each signal group. Normalize the feature values ​​of different dimensions. Calculate and sort the F-scores of all feature parameters based on the feature importance ranking results. The feature importance ranking results are as follows: Figure 7 As shown.

[0044] Step 4: Use four machine learning algorithms to build a regression model for training and evaluation, and select the optimal model evaluation metric and predicted scatter plot. Figure 1The model with the highest degree of convergence is denoted as the optimal machine learning model. Four machine learning algorithms were selected to build regression models for training: extreme gradient boosting regression, support vector regression, random forest regression, and multilayer perceptron regression. These four regression models were trained and evaluated, and the optimal prediction model was selected using two model evaluation metrics: root mean square error and coefficient of determination. Grid search cross-validation was used to optimize the parameters of the four machine learning models, and the optimal hyperparameters were selected for model training. A scatter plot comparing the predicted and measured values ​​of different models is shown below. Figure 9 As shown, the optimal model evaluation index and the predicted scatter plot are selected. Figure 1 The model with the highest degree of convergence, namely the multilayer perceptron regression model, is denoted as the optimal machine learning model.

[0045] Step 5: Obtain quantitative assessment results of the protective coating thickness. The terahertz detection signal of the protective coating sample is obtained using a terahertz time-domain spectroscopy system. Based on the F-score, the importance of features is ranked and selected, and important feature parameters with F-scores greater than a preset standard value are extracted from the terahertz time-domain spectral detection signal. After min-max normalization processing of the selected signal features, they are input along with the refractive index feature parameter into the trained optimal machine learning model. The thickness is used as the output parameter predicted by the optimal machine learning model. The optimal model prediction results are as follows: Figure 10 As shown, the absolute error of the predicted value of the multilayer perceptron regression model is 0.01 mm. The small error indicates that the method has significant effectiveness and accuracy in the quantitative assessment of the thickness of the protective coating on the surface of ceramic matrix composites using terahertz time-domain spectroscopy.

[0046] This invention designs a terahertz quantitative assessment method for the thickness of protective coatings based on machine learning. Using the finite-difference time-domain method, a terahertz simulation model of protective coatings with different refractive indices and thicknesses is constructed as a dataset, and time-domain and frequency-domain correlation features of each signal are extracted. After normalization using the min-max normalization method, important feature parameters with F-scores greater than preset standard values ​​are selected. Multiple machine learning algorithms are used to establish regression models, and the optimal prediction model is selected through root mean square error and coefficient of determination. The optimal hyperparameter result is selected for model training using a grid search cross-validation method. Thickness is used as the output parameter, and the feature parameters combining important feature parameters and refractive index are used as input parameters to train multiple regression models, and the optimal machine learning model is selected. The terahertz detection signal of the protective coating sample is obtained through a terahertz time-domain spectroscopy system. The signal features and refractive index feature parameters are selected and input into the trained optimal machine learning model. Thickness is used as the output parameter predicted by the optimal machine learning model, realizing the quantitative assessment of the protective coating thickness using terahertz time-domain spectroscopy. This method can be used for quantitative assessment of the thickness of protective coatings on ceramic matrix composites, solving the problem that traditional non-destructive testing methods are insufficient to effectively obtain information on the thickness of protective coatings on ceramic matrix composite surfaces. It improves the effectiveness and accuracy of quantitative assessment of the thickness of protective coatings on ceramic matrix composite surfaces using terahertz time-domain spectroscopy. This innovative method will open up new directions for the development of non-destructive testing technology in the field of ceramic matrix composites, providing solid technical support for the detection of protective coating thickness and service condition assessment of ceramic matrix composite surfaces. It has significant application value in the fields of non-destructive testing and structural health monitoring.

Claims

1. A terahertz quantitative evaluation method for protective coating thickness based on machine learning, characterized in that: Based on the finite-difference time-domain method, a terahertz simulation model of protective coatings with different refractive indices and thicknesses was constructed as a dataset. Time-domain and frequency-domain correlation features of each signal group were extracted. After normalization using the min-max normalization method, important feature parameters with F-scores greater than preset standard values ​​were selected. Multiple machine learning algorithms were used to establish regression models. The optimal prediction model was selected using root mean square error and coefficient of determination, and the optimal hyperparameters were selected for model training using a grid search cross-validation method. Thickness was used as the output parameter, and the feature parameters of the combination of important feature parameters and refractive index were used as input parameters. Multiple regression models were trained, and the optimal machine learning model was selected. The terahertz detection signal of the protective coating sample was obtained through a terahertz time-domain spectroscopy system. The feature parameters of the combination of important feature parameters and refractive index were input together into the trained optimal machine learning model. Thickness was used as the output parameter predicted by the optimal machine learning model, achieving quantitative assessment of the thickness of the protective coating using terahertz time-domain spectroscopy. Multiple machine learning algorithms were used to build regression models for training, including extreme gradient boosting regression, support vector regression, random forest regression, and multilayer perceptron regression. Multiple regression models were trained and evaluated, and the optimal prediction model was selected using two model evaluation metrics: root mean square error and coefficient of determination. Using thickness as the output parameter, and the selected important feature parameters with F scores greater than the preset standard value and the feature parameters of the refractive index combination as the input parameters, various regression models such as limit gradient boosting regression, support vector regression, random forest regression, and multilayer perceptron regression are trained respectively. A scatter plot of the model prediction is generated, and the model with the best model evaluation index and the highest degree of consistency with the predicted scatter plot is selected as the optimal machine learning model.

2. The terahertz quantitative evaluation method for protective coating thickness based on machine learning according to claim 1, characterized in that: Based on the finite-difference time-domain method, a terahertz simulation model for protective coating detection was established using simulation software. Simulation parameters were set, and simulation results of terahertz detection signals with different medium structural parameters and optical parameters were obtained. When simulating and modeling protective coatings on ceramic matrix composite surfaces, a protective coating detection model with arbitrary thickness and refractive index is designed based on the characteristics of the specific object being tested.

3. The terahertz quantitative evaluation method for protective coating thickness based on machine learning according to claim 2, characterized in that: Multiple simulation models of protective coatings on ceramic matrix composites with different refractive indices were constructed, with each set of data corresponding to multiple sets of different thicknesses. All sets of data were used as datasets for model training and testing, with refractive index as the input parameter and thickness as the predicted output parameter. Two types of features were extracted from each signal, namely multiple parameters of time-domain correlation features and multiple parameters of frequency-domain correlation features. All feature parameters were selected for signal feature extraction.

4. The terahertz quantitative evaluation method for protective coating thickness based on machine learning according to claim 3, characterized in that: By adopting the min-max normalization method, feature values ​​of different dimensions are normalized. The F-score is selected as the standard for feature selection. The F-scores of all feature parameters are calculated and sorted. Feature parameters with F-scores less than the preset standard value are eliminated based on the feature importance ranking results. Important feature parameters with F-scores greater than the preset standard value are selected as inputs for model training.

5. The terahertz quantitative evaluation method for protective coating thickness based on machine learning according to claim 1, characterized in that: The grid search cross-validation method is used to arrange and combine all hyperparameters of the model, merge all parameter combinations to generate a "grid", use an exhaustive search method to search for all hyperparameters for model training, evaluate each combination according to model evaluation metrics, and output the best parameter combination result; optimize parameters for multiple machine learning models, and select the optimal hyperparameter result for model training.

6. The terahertz quantitative evaluation method for protective coating thickness based on machine learning according to claim 1, characterized in that: Terahertz detection signals of protective coating samples are obtained using a terahertz time-domain spectroscopy system. Based on the F-score, the importance of features is ranked and selected, and important feature parameters with F-scores greater than preset standard values ​​are extracted from the terahertz time-domain spectroscopy detection signals. After the selected signal features are subjected to min-max normalization, they are input together with the refractive index feature parameters into a pre-trained optimal machine learning model. The thickness is used as the output parameter predicted by the optimal machine learning model to achieve quantitative assessment of the thickness of the protective coating using terahertz time-domain spectroscopy.

Images