Terahertz quantitative evaluation method for thickness of protective coating based on machine learning

Abstract

The invention discloses a protective coating thickness terahertz quantitative evaluation method based on machine learning, and the method comprises the following steps: constructing a data set based on a time domain finite difference method, and extracting the time domain and frequency domain related features of each group of signals; a regression model is established by selecting multiple machine learning algorithms, an optimal prediction model is selected through a root-mean-square error and a decision coefficient, and an optimal hyper-parameter result is selected through a grid search cross validation method for model training. Taking the thickness as an output parameter, taking the important characteristic parameter and the characteristic parameter of the refractive index combination as an input parameter, and respectively training a plurality of regression models; terahertz detection signals of a protective coating sample are obtained through a terahertz time-domain spectroscopy system, and quantitative evaluation of the thickness of the terahertz time-domain spectroscopy protective coating is achieved. The method can be used for quantitatively evaluating the thickness of the protective coating on the surface of the ceramic-based composite material, and the effectiveness and the accuracy of the terahertz time-domain spectrum quantitative evaluation of the thickness of the protective coating on the surface of the ceramic-based composite material are improved.

Description

Technical Field

[0001] The present invention is a terahertz quantitative evaluation method for protective coating thickness based on machine learning, which belongs to the field of non-destructive testing and evaluation. Background Art

[0002] Ceramic matrix composites have a wide range of applications and can be widely used in key components in the aerospace field such as missile air intakes, aircraft flaps, and rocket thrust chambers. As a key material used in high-temperature components, ceramic matrix composites have uneven characteristics. They will inevitably be exposed to harsh environments such as high temperature and gas during service. In order to ensure that they can serve efficiently and for a long time, the materials generally have surface protective coatings. Taking various factors into consideration, coating thinning damage is likely to occur during the production and preparation of materials and during use. In the service process of ceramic matrix composites, there is an urgent problem to be solved in terms of how to improve reliability, that is, an effective detection method is needed for the detection of the thickness of the surface protective coating. In the various stages of the processing and production of ceramic matrix composites and their actual use, including the stage of manufacturing raw materials, the stage of assembling components, and the stage of service process, appropriate non-destructive testing methods are required to obtain information including the thickness of the surface protective coating to achieve effective control of material quality, which is of great significance for product quality assurance and reliability improvement during product use.

[0003] There are many methods available for nondestructive testing of the thickness of protective coatings on the surface of ceramic-based composite materials, including ultrasound, X-ray and infrared thermal imaging. Terahertz time-domain spectroscopy has unique advantages in detecting the thickness of protective coatings on the surface of ceramic-based composite materials. Compared with traditional ultrasonic testing or thermal imaging testing methods, it does not require coupling agents and is not affected by the low thermal conductivity of CMC. It can penetrate non-conductive non-polar materials deeply, with a wider detection range and better spatial resolution. Compared with X-ray detection methods, its energy is lower, only in the order of milli-electron volts, harmless to human tissue, and can avoid the impact of ionization reactions on health, while avoiding the limitations of radiation shielding requirements. Therefore, the existing detection methods face great limitations in practical applications and are difficult to meet the requirements for quantitative evaluation of the thickness of protective coatings on the surface of ceramic-based composite materials. On the other hand, terahertz waves have good penetration in the detection of many non-metallic materials, such as composite materials, ceramics, rubber, etc. Therefore, terahertz time-domain spectroscopy is particularly suitable for comprehensive, rapid, and non-contact nondestructive testing of ceramic-based composite materials. Based on this, it is particularly necessary to develop a terahertz quantitative evaluation method for protective coating thickness based on machine learning.

[0004] The present invention is based on the time-domain finite difference method, and constructs a terahertz simulation model of a protective coating with different refractive indices and thicknesses as a data set, and extracts the time domain and frequency domain related features of each group of signals. After normalization processing is performed using the minimum-maximum normalization method, important characteristic parameters with F scores greater than preset standard values ​​are selected. A variety of machine learning algorithms are selected to establish a regression model, the optimal prediction model is selected by the root mean square error and the coefficient of determination, and the optimal hyperparameter results are selected by the grid search cross-validation method for model training. The thickness is used as the output parameter, and the characteristic parameters of the important characteristic parameters and the refractive index combination are used as input parameters. Multiple regression models are trained separately to select the optimal machine learning model. The terahertz detection signal of the protective coating sample is obtained through the terahertz time-domain spectroscopy system, and the signal characteristics and the refractive index characteristic parameters are selected and input 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 evaluation of the thickness of the terahertz time-domain spectroscopy protective coating. This method can be used to quantitatively evaluate the thickness of protective coatings on the surface of ceramic-based composite materials, solving the problem that it is difficult to effectively obtain thickness information of protective coatings on the surface of ceramic-based composite materials based on traditional non-destructive testing methods, and providing solid technical support for the thickness detection and service status evaluation of protective coatings on the surface of ceramic-based composite materials. This innovative method will open up a new direction for the development of non-destructive testing technology in the field of ceramic-based composite materials. The terahertz quantitative evaluation method of protective coating thickness based on machine learning has not been reported at home and abroad. Summary of the invention

[0005] The purpose of the present invention is to address the deficiencies of the above-mentioned prior art and to develop a terahertz time-domain spectroscopy coating thickness quantitative assessment method based on machine learning algorithm to solve the problem that it is difficult to obtain the thickness information of the protective coating on the surface of ceramic matrix composite materials and to quantitatively assess the coating thickness based on traditional non-destructive testing methods. The proposed terahertz quantitative assessment method for protective coating thickness based on machine learning can improve the quality of quantitative assessment of protective coating thickness and the detection accuracy.

[0006] This method is mainly based on the time-domain finite difference method, and constructs a terahertz simulation model of a protective coating with different refractive indices and thicknesses as a data set to extract the time domain and frequency domain related features of each group of signals. After normalization using the minimum-maximum normalization method, select important characteristic parameters with F scores greater than the preset standard values. Select a variety of machine learning algorithms to establish a regression model, select the optimal prediction model through the root mean square error and the coefficient of determination, and select the optimal hyperparameter results for model training through the grid search cross-validation method. Take the thickness as the output parameter, and the characteristic parameters of the important characteristic parameters and the refractive index combination as the input parameters. Train multiple regression models separately and select the optimal machine learning model. The terahertz detection signal of the protective coating sample is obtained through the terahertz time-domain spectroscopy system, and the signal characteristics and the refractive index characteristic parameters are selected together to input the trained optimal machine learning model. The thickness is used as the output parameter predicted by the optimal machine learning model to achieve the quantitative evaluation of the thickness of the terahertz time-domain spectroscopy protective coating. The implementation flow chart of this method is shown below. Fig.11 shown.

[0007] In order to achieve the above object, 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 is established through simulation software, simulation parameters are set, and simulation results of terahertz detection signals of different medium structural parameters and optical parameters are obtained. Simulation models of protective coatings on the surface of ceramic-based composite materials with multiple groups of different refractive indices are constructed, and each group of data corresponds to multiple groups of different thicknesses. Two types of features of each group of signals are extracted separately, which are divided into multiple parameters of time-domain related features and multiple parameters of frequency-domain related features, and all feature parameters are selected for signal feature extraction. By adopting the minimum-maximum normalization method, the feature values ​​of different dimensions are normalized, and the F scores of all feature parameters are calculated and sorted respectively according to the feature importance sorting results. The important feature parameters with F scores greater than the preset standard values ​​are selected as input for model training. A variety of machine learning algorithms are selected to establish regression models for training, and these multiple regression models are trained and evaluated. The optimal prediction model is selected by two model evaluation indicators, root mean square error and determination coefficient. Through the grid search cross-validation method, all the hyperparameters of the model are arranged and combined, and each combination is evaluated according to the model evaluation index. The best parameter combination result is output and the parameters of multiple machine learning models are optimized. The optimal hyperparameter result is selected for model training. The thickness is used as the output parameter, and the important characteristic parameters with F scores greater than the preset standard value and the characteristic parameters of the refractive index combination are selected as input parameters to train multiple regression models respectively. A scatter plot of model prediction is generated, and the best model evaluation index and predicted scatter plot are selected. Figure 1The model with the highest degree of consistency is recorded as the optimal machine learning model. The terahertz detection signal of the protective coating sample is obtained through the terahertz time-domain spectroscopy system. The feature importance is sorted and selected based on the F score, and the important feature parameters with F scores greater than the preset standard value in the terahertz time-domain spectroscopy detection signal are extracted. After the selected signal features are normalized to the minimum and maximum values, they are input into the trained optimal machine learning model together with the refractive index characteristic parameters. The thickness is used as the output parameter predicted by the optimal machine learning model to achieve quantitative evaluation of the thickness of the protective coating using terahertz time-domain spectroscopy.

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

[0010] The simulation model of the surface protective coating of ceramic-based composite materials with multiple groups of different refractive indices is constructed, and each group of data corresponds to multiple groups of different thicknesses. It is characterized in that all groups of data are used as data sets for model training sets and test sets, 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 characteristic parameters of each group of signals separately is characterized in that two types of characteristics of each group of signals are extracted separately, which are divided into multiple parameters of time domain related characteristics and multiple parameters of frequency domain related characteristics, and all characteristic parameters are selected for signal feature extraction.

[0012] The method of normalizing the eigenvalues ​​and calculating and sorting the F scores of the eigenparameters is characterized in that: the eigenvalues ​​of different dimensions are normalized by adopting the minimum-maximum normalization method, the F scores of all the eigenparameters are calculated and sorted respectively according to the feature importance sorting results, and important eigenparameters with F scores greater than preset standard values ​​are selected as input for model training.

[0013] The method of selecting a machine learning algorithm to establish a regression model for training is characterized in that: multiple machine learning algorithms are selected to establish a regression model for training, including multiple regression models such as extreme gradient boosting regression, support vector regression, random forest regression, and multi-layer perceptron regression. These multiple regression models are trained and evaluated, and the optimal prediction model is selected through two model evaluation indicators, root mean square error and determination coefficient.

[0014] The grid search cross-validation method is used to arrange and combine all hyperparameters of the model, and is characterized in that all parameter combinations are merged to generate a "grid", an exhaustive search method is used to search all hyperparameters for model training, each combination is evaluated according to the model evaluation index, the best parameter combination result is output, and parameter optimization is performed on multiple machine learning models, and the optimal hyperparameter result is selected for model training.

[0015] The method of training multiple regression models separately and generating a scatter plot of model prediction is characterized in that: taking thickness as output parameter, taking important characteristic parameters with F scores greater than preset standard values ​​and characteristic parameters of refractive index combination as input parameters, and training multiple regression models such as extreme gradient boosting regression, support vector regression, random forest regression and multi-layer perceptron regression separately. Generating a scatter plot of model prediction, selecting the best model evaluation index and the predicted scatter plot Figure 1 The model with the highest degree of consistency is recorded as the optimal machine learning model.

[0016] The characteristic parameters are input together into the trained optimal machine learning model to achieve quantitative evaluation of the thickness of the terahertz time-domain spectroscopy protective coating. It is characterized in that: the terahertz detection signal of the protective coating sample is obtained through the terahertz time-domain spectroscopy system, the feature importance is sorted and selected based on the F score, and the important characteristic parameters with F scores greater than the preset standard value in the terahertz time-domain spectroscopy detection signal are extracted. After the selected signal features are normalized to the minimum and maximum values, they are input into the trained optimal machine learning model together with the refractive index characteristic parameters, and 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 detection model of any thickness and refractive index according to the characteristics of the specific detection object;

[0020] 3. Construct multiple sets of simulation models of ceramic-based composite surface protective coatings with different refractive indices, with each set of data corresponding to multiple sets of different thicknesses;

[0021] 4. Extract multiple parameters of time domain related features and multiple parameters of frequency domain related features of each group of signals respectively;

[0022] 5. By adopting the minimum-maximum normalization method, the eigenvalues ​​of different dimensions are normalized;

[0023] 6. Calculate and sort the F scores of all feature parameters according to the feature importance ranking results;

[0024] 7. Select a variety of machine learning algorithms to establish regression models for training, and select the optimal prediction model through two model evaluation indicators: root mean square error and determination coefficient;

[0025] 8. Use the grid search cross-validation method to optimize the parameters of various machine learning models and select the optimal hyperparameter results for model training;

[0026] 9. Select the best model evaluation index and prediction scatter points Figure 1 The model with the highest degree of consistency is recorded as the optimal machine learning model;

[0027] 10. Obtain the terahertz detection signal of the protective coating sample through the 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 the selected terahertz time-domain spectroscopy detection signal characteristics are normalized to the minimum and maximum values, they are input into the trained optimal machine learning model together with the refractive index characteristic parameters. The thickness is used as the output parameter predicted by the optimal machine learning model to achieve quantitative evaluation of the thickness of the terahertz time-domain spectroscopy protective coating. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] Figure 1 It is a simulation modeling flow chart;

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

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

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

[0033] Figure 5 This is the simulation signal diagram of coatings with different thicknesses;

[0034] Figure 6 It is the simulation signal diagram of coatings with different refractive indices;

[0035] Figure 7 It is the result graph of feature importance ranking;

[0036] Figure 8 It is a flow chart of machine learning algorithm;

[0037] Fig. 9 It is a scatter plot comparing the predicted values ​​of different models with the measured values;

[0038] Fig.10 is the prediction result of the optimal model;

[0039] Fig.11It is a flow chart for implementing this method. DETAILED DESCRIPTION

[0040] The present invention will be further described below in conjunction with the accompanying drawings. The following embodiments are only illustrative and not restrictive, and the protection scope of the present invention cannot be limited thereto.

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

[0042] Step 2: Construct multiple sets of simulation models of ceramic matrix composite surface protective coatings with different refractive indices, and each set of data corresponds to multiple sets of different thicknesses. Figure 4 As shown in the figure, the absorption boundary of the PML perfectly matched layer is set around the simulation area, and the fully reflective metal substrate boundary is set on the bottom surface of the model to simulate the coating reflection detection. Some simulation signals of coatings with different thicknesses are shown in the figure. Figure 5 As shown, some simulation signals of coatings with different refractive indices are as follows Figure 6 shown.

[0043] Step 3: Extract signal characteristic parameters, calculate and sort the F scores after normalization. Select material parameters to extract multiple parameters of time domain related characteristics and multiple parameters of frequency domain related characteristics of each group of signals, normalize the characteristic values ​​of different dimensions, and calculate and sort the F scores of all characteristic parameters according to the feature importance sorting results. The feature importance sorting results are shown in the figure below. Figure 7 shown.

[0044] Step 4: Select four machine learning algorithms to build a regression model for training and model evaluation, and select the best model evaluation index and prediction scatter points Figure 1The model with the highest degree of consistency is recorded as the optimal machine learning model. Four machine learning algorithms are selected to establish regression models for training, including extreme gradient boosting regression, support vector regression, random forest regression, and multilayer perceptron regression. These four regression models are trained and evaluated, and the optimal prediction model is selected by two model evaluation indicators: root mean square error and determination coefficient. The four machine learning models are optimized by grid search cross-validation method, and the optimal hyperparameter results are selected for model training. The scatter plots comparing the predicted values ​​of different models with the measured values ​​are shown in the figure below. Fig. 9 As shown, the best model evaluation index and prediction scatter points are selected Figure 1 The model with the highest degree of consistency, namely the multilayer perceptron regression model, is recorded as the optimal machine learning model.

[0045] Step 5: Obtain the quantitative evaluation results of the protective coating thickness. The terahertz detection signal of the protective coating sample is obtained through the terahertz time-domain spectroscopy system. The feature importance is sorted and selected based on the F score, and the important feature parameters with F scores greater than the preset standard value in the terahertz time-domain spectroscopy detection signal are extracted. After the selected signal features are normalized to the minimum and maximum values, they are input into the trained optimal machine learning model together with the refractive index feature parameters, and the thickness is used as the output parameter predicted by the optimal machine learning model. The optimal model prediction results are as follows: Fig.10 As shown in the figure, the absolute error of the predicted value of the multilayer perceptron regression model is 0.01 mm. The small error shows that this method has obvious effectiveness and accuracy in the quantitative evaluation of the thickness of the protective coating on the surface of ceramic matrix composite materials by terahertz time-domain spectroscopy.

[0046] The present invention designs a terahertz quantitative evaluation method for protective coating thickness based on machine learning. Based on the finite difference time domain method, a terahertz simulation model of protective coatings with different refractive indices and thicknesses is constructed as a data set, and the time domain and frequency domain related features of each group of signals are extracted. After normalization processing is performed using the minimum-maximum normalization method, important characteristic parameters with F scores greater than preset standard values ​​are selected. A variety of machine learning algorithms are selected to establish a regression model, the optimal prediction model is selected through the root mean square error and the determination coefficient, and the optimal hyperparameter results are selected through the grid search cross-validation method for model training. The thickness is used as the output parameter, and the characteristic parameters of the important characteristic parameters and the refractive index combination are used as input parameters. Multiple regression models are trained respectively, and the optimal machine learning model is selected. The terahertz detection signal of the protective coating sample is obtained through the terahertz time-domain spectroscopy system, and the signal characteristics and the refractive index characteristic parameters are selected and input into the trained optimal machine learning model. The thickness is used as the output parameter predicted by the optimal machine learning model to achieve terahertz time-domain spectroscopy. Quantitative evaluation of the thickness of the protective coating. This method can be used for quantitative evaluation of the thickness of protective coatings on the surface of ceramic-based composite materials. It can solve the problem that it is difficult to effectively obtain the thickness information of protective coatings on the surface of ceramic-based composite materials based on traditional non-destructive testing methods, and improve the effectiveness and accuracy of terahertz time-domain spectroscopy quantitative evaluation of the thickness of protective coatings on the surface of ceramic-based composite materials. This innovative method will open up a new direction for the development of non-destructive testing technology in the field of ceramic-based composite materials, and provide solid technical support for the thickness detection and service status evaluation of protective coatings on the surface of ceramic-based composite materials. It has great application value in the field 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 is constructed as a data set, and the time domain and frequency domain related features of each group of signals are extracted; after normalization processing using the minimum-maximum normalization method, important characteristic parameters with F scores greater than the preset standard values ​​are selected; a variety of machine learning algorithms are used to establish a regression model, and the optimal prediction model is selected through the root mean square error and the coefficient of determination, and the optimal hyperparameter results are selected through the grid search cross-validation method for model training; the thickness is used as the output parameter, and the characteristic parameters of the important characteristic parameters and the refractive index combination are used as input parameters to train multiple regression models respectively, and select the optimal machine learning model; the terahertz detection signal of the protective coating sample is obtained through the terahertz time-domain spectroscopy system, and the signal characteristics and the refractive index characteristic parameters are selected together to input the trained optimal machine learning model, and the thickness is used as the output parameter predicted by the optimal machine learning model to achieve quantitative evaluation of the thickness of the protective coating using terahertz time-domain spectroscopy.

2. The method for quantitatively evaluating the thickness of protective coatings based on terahertz frequency based 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 is established through simulation software, simulation parameters are set, and simulation results of terahertz detection signals of different medium structure parameters and optical parameters are obtained; When simulating and modeling the protective coating on the surface of ceramic-based composite materials, a protective coating detection model with arbitrary thickness and refractive index is designed according to the specific characteristics of the detection object.

3. The terahertz quantitative evaluation method of protective coating thickness based on machine learning according to claim 2 is characterized in that: Construct multiple groups of simulation models of surface protective coatings of ceramic-based composite materials with different refractive indices, and each group of data corresponds to multiple groups of different thicknesses; all groups of data are used as data sets for model training sets and test sets, in which the refractive index is used as an input parameter and the thickness is used as a predicted output parameter; two types of features are extracted from each group of signals, which are divided into multiple parameters of time-domain related features and multiple parameters of frequency-domain related features, and all feature parameters are selected for signal feature extraction.

4. The method for quantitatively evaluating the thickness of protective coatings based on terahertz frequency based machine learning according to claim 3, characterized in that: By adopting the minimum-maximum normalization method, the feature values ​​of different dimensions are normalized, and the F score is selected as the standard for feature selection. The F scores of all feature parameters are calculated and sorted respectively. The feature parameters with F scores less than the preset standard value are eliminated through the feature importance sorting results, and the important feature parameters with F scores greater than the preset standard value are selected as input for model training.

5. The method for quantitatively evaluating the thickness of protective coatings based on terahertz frequency based machine learning according to claim 4, characterized in that: A variety of machine learning algorithms are used to establish regression models for training, including extreme gradient boosting regression, support vector regression, random forest regression, multilayer perceptron regression and other different regression models; these multiple regression models are trained and evaluated, and the optimal prediction model is selected through two model evaluation indicators: root mean square error and determination coefficient.

6. The method for quantitatively evaluating the thickness of protective coatings based on terahertz frequency based machine learning according to claim 5, characterized in that: Through the grid search cross-validation method, all hyperparameters of the model are arranged and combined, all parameter combinations are merged to generate a "grid", and all hyperparameters are searched for model training using an exhaustive search method. Each combination is evaluated according to the model evaluation index, and the best parameter combination result is output; parameter optimization is performed on multiple machine learning models, and the optimal hyperparameter result is selected for model training.

7. The method for quantitatively evaluating the thickness of protective coatings based on terahertz frequency based machine learning according to claim 6, characterized in that: Taking thickness as the output parameter, and the important characteristic parameters with F scores greater than the preset standard values ​​and the characteristic parameters of the refractive index combination as input parameters, various regression models such as extreme gradient boosting regression, support vector regression, random forest regression and multi-layer perceptron regression are trained respectively; a scatter plot of 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.

8. The method for quantitatively evaluating the thickness of protective coatings based on terahertz frequency based machine learning according to claim 7, characterized in that: The terahertz detection signal of the protective coating sample is obtained through the terahertz time-domain spectroscopy system, and the feature importance is ranked and selected based on the F-score, and the important characteristic parameters with F-score greater than the preset standard value in the terahertz time-domain spectroscopy detection signal are extracted; after the selected signal features are normalized to the minimum and maximum values, they are input into the trained optimal machine learning model together with the refractive index characteristic parameters, and the thickness is used as the output parameter predicted by the optimal machine learning model to realize the quantitative evaluation of the thickness of the terahertz time-domain spectroscopy protective coating.

Images