Coating layer number automatic classification method based on terahertz time-domain spectroscopy

Abstract

The invention discloses a terahertz time-domain spectroscopy-based coating layer number automatic classification method, and belongs to the field of multi-layer coating thickness detection, and the method comprises the steps: S1, building a detection platform, and collecting the spectral data of different layers of coatings to construct a sample set; s2, processing the data and extracting time domain and frequency domain features; s3, performing kernel principal component analysis dimension reduction to obtain an input feature vector; s4, dividing and normalizing a training set and a test set; s5, optimizing support vector machine model parameters through a differential evolution algorithm; s6, completing model training and verification; and S7, processing to-be-detected data, inputting the data into the model, and automatically outputting the number of coating layers. By the adoption of the method, coating prior information is not needed, the defects of a traditional optimization means are overcome, the recognition problem of a conventional algorithm is solved, the method is suitable for intelligent recognition of the thickness of a complex coating with the unknown layer number in the industrial environment, and the engineering practical value and the automation potential are higher.

Description

Technical Field

[0001] This invention relates to the field of multilayer coating thickness detection, and in particular to an automatic classification method for coating layer number based on terahertz time-domain spectroscopy. Background Technology

[0002] Terahertz waves, electromagnetic waves with frequencies between 0.1 and 10 THz, possess unique physical properties such as high penetration, low photon energy, fingerprint recognition, and transient characteristics, demonstrating great potential in the field of multilayer coating thickness measurement. Compared to traditional thickness measurement methods such as ultrasonic, eddy current, and X-ray thickness measurement, terahertz detection technology offers significant advantages, including non-contact, non-destructive testing, high sensitivity, and the ability to characterize multilayer composite structures. In industrial applications, it not only provides high-precision measurement results but also possesses real-time monitoring capabilities, showcasing unique application value and potential in non-destructive testing scenarios for complex multilayer composite coatings in aerospace equipment, new energy vehicles, and turbine blades.

[0003] However, in special engineering applications, the number of coating layers on the sample being tested is often unknown, posing a significant challenge to existing terahertz thickness measurement algorithms. Currently, high-precision thickness measurement models typically rely heavily on prior information about the number of coating layers. That is, when faced with multi-layered coating samples with unknown layers and micrometer-level thicknesses, the lack of prior information makes it difficult for the algorithm to accurately define the boundaries of each layer, severely impacting the reliability and accuracy of the measurement results, thus limiting the widespread application of terahertz technology in complex multi-layered coatings. Summary of the Invention

[0004] The purpose of this invention is to provide an automatic coating layer number classification method based on terahertz time-domain spectroscopy. This method requires no prior knowledge of the coating's specific layer structure or any other prior information. By combining kernel principal component analysis (KPCA) feature extraction with a support vector machine (SVM) model optimized using differential evolution (DE), it can achieve rapid and accurate identification of unknown multilayer coating structures. This method not only overcomes the limitations of traditional optimization methods such as grid search, which are computationally time-consuming and prone to getting trapped in local optima, but also solves the problems of conventional algorithms struggling with feature extraction and blurred layer number identification boundaries when dealing with thin coatings or complex boundaries, ultimately leading to the failure of subsequent thickness inversion models. It is particularly suitable for the automatic intelligent identification of the thickness of complex coatings with unknown layers in industrial environments, possessing greater engineering practical value and automation potential.

[0005] To achieve the above objectives, this invention provides an automatic classification method for coating layer number based on terahertz time-domain spectroscopy. S1, Build a detection platform that combines a terahertz time-domain spectroscopy system with a robotic arm, calibrate the probe posture so that the terahertz wave is perpendicularly incident on the sample to be tested, and collect time-domain spectral data of coatings with different layer numbers to construct a sample set. S2. Denoise and frequency domain transformation are performed on the time-domain spectral data in S1 to extract time-domain and frequency-domain features; S3. Use kernel principal component analysis to reduce the dimensionality of the features extracted in S2, and select the top N principal components with the highest cumulative contribution rate as the input feature vector. S4. Divide the input feature vectors in S3 into training set and test set and perform mean normalization. S5. Based on the normalized training set in S4, construct a radial basis function kernel support vector machine classification model, and optimize the core parameters of the classification model through differential evolution algorithm to obtain the optimal parameter combination; S6. Train the classification model optimized by S5 using the normalized training set, and then verify the performance of the classification model using the normalized test set. S7. After processing the spectral data of the sample to be tested using the same method as in steps S2-S3, process it according to the normalization standard in S4, input it into the classification model verified in S6, and the classification model will automatically output the number of coating layers.

[0006] Preferably, S1 specifically includes: S11. Platform Construction: An automatic scanning and detection platform is built by combining a terahertz time-domain spectroscopy system with a robotic arm. The probe posture is calibrated by a laser displacement sensor so that the terahertz wave is perpendicularly incident on the surface of the sample to be tested, and time-domain spectral data of different coating layers are collected. S12. Construct a dataset: Summarize the time-domain spectral data collected in S11 into a sample set, with each sample containing complete time-domain waveform points.

[0007] Preferably, S2 specifically includes: S21. Noise reduction processing: Perform ensemble empirical mode decomposition noise reduction processing on the original terahertz signal in S1. S22, Frequency Domain Transformation: Perform a Fast Fourier Transform on the time-domain signal processed in S21 to obtain the corresponding frequency-domain spectral data; S23. Extract the time-domain features of the signal processed by S21 and the frequency-domain features of the frequency-domain spectral data obtained by S22.

[0008] Preferably, the time-domain features in S23 include mean, root mean square, peak-to-peak value, kurtosis, and skewness; Frequency domain characteristics include spectral mean, spectral standard deviation, frequency center, and spectral kurtosis.

[0009] Preferably, S3 specifically includes: S31. Kernel Principal Component Analysis: The kernel principal component analysis method is used to perform nonlinear mapping and principal component extraction on the original feature data extracted in S2 through kernel functions; S32, Nonlinear Mapping: Based on the kernel function, the original feature data extracted in S2 is mapped to a high-dimensional space, and principal component extraction is completed in the high-dimensional space to achieve feature dimensionality reduction; S33. Feature selection: Select the top N principal components with the highest cumulative contribution rate as the final input feature vector.

[0010] Preferably, S4 specifically includes: S41. Dataset partitioning: Randomly divide the dimensionality-reduced dataset from S3 into a training set and a test set; S42. Mean Normalization: Perform mean normalization on the training and test sets after S41, mapping the data to... Interval.

[0011] Preferably, S5 specifically includes: S51. Establish an SVM model: Based on the normalized training set in S4, construct a support vector machine classification model. The penalty coefficient of the classification model is C, and the radial basis function is selected. As the kernel function, the formula is: ; in, and There are two input sample vectors. It is the Euclidean distance between two input sample vectors. These are the kernel parameters of the radial basis functions; Preferably, S52 specifically includes: S521. Initialization: Set the population size, scaling factor, crossover probability, and maximum number of iterations for the differential evolution algorithm. Randomly generate a population containing... The initial population for the parameter combination is given by the formula: ; in It represents a random decimal that is uniformly distributed between 0 and 1. , They are respectively Upper and lower bounds of parameter combinations; S522, Mutation and Crossover: Mutation is performed on individuals in the initial population in S521 to generate difference vectors, and crossover is performed to generate experimental individuals; S523. Selection: Use the cross-validation accuracy of the SVM model on the normalized training set in S4 as the fitness function, compare the fitness of the experimental individuals with that of the target individuals, and retain the better parameter combination for the next generation. S524. Iteration: Repeat steps S522-S523 until the maximum number of iterations is reached or the convergence condition is met, and output the globally optimal parameter combination. .

[0012] Preferably, S6 specifically includes: S61, Model Training: Combining the optimal parameters obtained in S5 The data is assigned to the SVM model, and the classification model is trained using the normalized training set data from S4. S62, Layer Prediction: Input the normalized test set data from S4 into the classification model trained in S61, output the predicted category of each sample, and verify the model performance by evaluating the classification accuracy of the model.

[0013] Therefore, the above-mentioned automatic coating layer number classification method based on terahertz time-domain spectroscopy in this invention has the following beneficial effects: (1) Extremely high degree of automation and system independence: By constructing a layer classification preprocessing mechanism and combining it with a thickness inversion model, this invention can automatically identify unknown multilayer structures without any prior information such as coating refractive index or layer number. This overcomes the drawback of traditional terahertz thickness measurement methods that require preset model parameters, significantly reduces the cost of manual intervention, and enables the terahertz detection system to operate independently on complex and random industrial samples.

[0014] (2) Significant computational efficiency and engineering practical value: This invention utilizes KPCA (Kernel Principal Component Analysis) for nonlinear feature dimensionality reduction, which significantly compresses the data dimension while retaining key physical features. Combined with the efficient prediction capability of DE-SVM, the system can achieve rapid layer number determination after a given spectrum. This efficient algorithm architecture not only meets the timeliness requirements of real-time monitoring in industrial sites, but also provides a reliable prerequisite for subsequent accurate thickness inversion, and has great potential for engineering promotion.

[0015] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0016] Figure 1 This is a flowchart illustrating the automatic classification and identification process for the number of coating layers according to an embodiment of the present invention. Figure 2 This is a diagram of a reflective terahertz time-domain spectroscopy system based on robotic arm drive, according to an embodiment of the present invention. Figure 3 This is a photograph of the three-layer coating sample of Embodiment 1 of the present invention; Figure 4 This is a photograph of the four-layer coating sample of Embodiment 2 of the present invention; Figure 5 This is a flowchart of the ensemble empirical mode decomposition (EEMD) according to an embodiment of the present invention; Figure 6This is a confusion matrix diagram of the 1-3 layer coating identification test set in Embodiment 1 of the present invention; Figure 7 This is the confusion matrix diagram of the 1-4 layer coating identification test set in Embodiment 2 of the present invention; Figure 8 This is a diagram showing the automatic identification results of the number of different types of coating layers in an embodiment of the present invention. Detailed Implementation

[0017] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. In the description of the present invention, it should be noted that the terms "upper," "lower," "inner," and "outer," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings, or the orientation or positional relationship commonly used when the product is in use. They are only for the convenience of describing the present invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on the present invention.

[0018] Example 1 like Figure 1 As shown, this invention discloses an automatic classification method for the number of coating layers based on terahertz time-domain spectroscopy, the steps of which include: S1. Construct a detection platform that combines a terahertz time-domain spectroscopy system with a robotic arm, such as... Figure 2 As shown, the probe orientation is calibrated so that the terahertz wave is incident perpendicularly on the sample under test, and time-domain spectral data of different coating layers are collected to construct a sample set. S1 specifically includes: S11. Platform Construction: An automatic scanning and detection platform is built by combining a terahertz time-domain spectroscopy system with a robotic arm. The probe posture is calibrated by a laser displacement sensor so that the terahertz wave is perpendicularly incident on the surface of the sample to be tested, and time-domain spectral data of different coating layers are collected. S12. Constructing the dataset: Summarize the terahertz time-domain spectral data of coatings with different numbers of layers (1-3 layers) collected in S11 into a sample set. The actual coating samples are as follows: Figure 3 As shown. Each spectral sample contains 6750 complete time-domain waveform points, and the dataset consists of 300 different layers of spectral data.

[0019] S2. The time-domain spectral data in S1 is subjected to noise reduction and frequency domain transformation to extract time-domain and frequency-domain features. S2 specifically includes: S21. Noise Reduction Processing: Perform ensemble empirical mode decomposition (EEMD) noise reduction processing on the original terahertz signal in S1, such as... Figure 5 As shown, noise caused by the environment and mechanical structure is filtered out from the terahertz signal; S22, Frequency Domain Transformation: Perform a Fast Fourier Transform (FFT) on the time-domain signal processed in S21 to obtain the corresponding frequency-domain spectral data; S23. Extract the time-domain features of the signal processed in S21 and the frequency-domain features of the frequency-domain spectral data obtained in S22. The time-domain features include mean, root mean square, peak-to-peak value, kurtosis, and skewness; the frequency-domain features include spectral mean, spectral standard deviation, frequency center, and spectral kurtosis.

[0020] S3. Kernel principal component analysis is used to reduce the dimensionality of the features extracted in S2, and the top N principal components with the highest cumulative contribution rate are selected as the input feature vectors.

[0021] S3 specifically includes: S31. Kernel Principal Component Analysis (KPCA): To address the computational redundancy and "curse of dimensionality" issues caused by high-dimensional feature data, the kernel principal component analysis method is used to perform nonlinear mapping and principal component extraction on the original feature data extracted in S2 through kernel functions. S32, Nonlinear Mapping: Based on the kernel function, the original feature data extracted in S2 is mapped to a high-dimensional space, and principal component extraction is completed in the high-dimensional space to achieve feature dimensionality reduction; S33. Feature Selection: Select the top 5 principal components with the highest cumulative contribution rate as the final input feature vector to improve the computational efficiency of the subsequent classifier.

[0022] S4. Divide the input feature vectors in S3 into training and test sets and perform mean normalization. S4 specifically includes: S41. Dataset partitioning: The dimensionality-reduced dataset in S3 is randomly divided into a training set and a test set, with a ratio of 7:3, i.e., 210 training sets and 90 test sets. S42. Mean Normalization: Perform mean normalization on the training and test sets after S41, mapping the data to... Interval. This step is to eliminate the dimensional differences between different feature metrics, ensuring the convergence speed and accuracy of model training.

[0023] S5. Based on the normalized training set in S4, construct a radial basis function kernel support vector machine (SVM) classification model. Optimize the core parameters of the classification model using the differential evolution (DE) algorithm to obtain the optimal parameter combination. S5 specifically includes: S51. Establish an SVM model: Based on the normalized training set in S4, construct a support vector machine classification model. The penalty coefficient of the classification model is C, and the radial basis function is selected. As the kernel function, the formula is: ; in, and There are two input sample vectors. It is the Euclidean distance between two input sample vectors. These are the kernel parameters of the radial basis functions; S52, Differential Evolution Optimization: The core parameters of the classification model in S51 are optimized using the differential evolution algorithm to obtain the optimal parameter combination. S52 specifically includes: S521. Initialization: Set the population size, scaling factor, crossover probability, and maximum number of iterations for the differential evolution algorithm. Randomly generate a population containing... The initial population for the parameter combination is given by the formula: ; in It represents a random decimal that is uniformly distributed between 0 and 1. , They are respectively Upper and lower bounds of parameter combinations; S522, Mutation and Crossover: Mutation is performed on individuals in the initial population in S521 to generate difference vectors, and crossover is performed to generate experimental individuals; S523. Selection: Use the cross-validation accuracy of the SVM model on the normalized training set in S4 as the fitness function, compare the fitness of the experimental individuals with that of the target individuals, and retain the better parameter combination for the next generation. S524. Iteration: Repeat steps S522-S523 until the maximum number of iterations is reached or the convergence condition is met, and output the globally optimal parameter combination. .

[0024] S6. Train the classification model optimized in S5 using the normalized training set, and then verify the performance of the classification model using the normalized test set. S6 specifically includes: S61, Model Training: Combining the optimal parameters obtained in S5 The data is assigned to the SVM model, and the classification model is trained using the normalized training set data from S4. S62, Layer Prediction: Input the normalized test set data from S4 into the classification model trained in S61, output the predicted class for each sample, and generate the test set confusion matrix as follows. Figure 6As shown, the accuracy rate reached 98.89%, and the model performance was verified by evaluating the classification accuracy of the model.

[0025] S7. After processing the spectral data of the sample to be tested using the same method as in steps S2-S3, process it according to the normalization standard in S4, input it into the classification model verified in S6, and the classification model will automatically output the number of coating layers.

[0026] Example 2 This core solution is adapted and optimized based on Example 1, specifically as follows: A detection platform is built by combining a reflective terahertz time-domain spectroscopy system with a robotic arm (e.g., Figure 2 As shown), terahertz time-domain spectral data of coatings of layers 1-4 were collected to construct a dataset containing 400 samples (the actual coating samples are shown in the image). Figure 4 As shown, each sample contains 6750 time-domain waveform points; noise reduction is performed using EEMD (process as follows). Figure 5 As shown in the figure, after extracting time and frequency domain features by FFT frequency domain transformation, the first 5 principal components are selected by KPCA dimensionality reduction. The training set and test set are divided in a 7:3 ratio and mean normalization is performed. Based on the normalized training set, an RBF kernel SVM classification model is constructed. Its core parameters are optimized by DE algorithm. The KPCA-DE-SVM classification model is trained and used for coating layer number prediction.

[0027] The technical effects of this embodiment are significant: it adapts to the classification requirements of 1-4 coating layers, expanding the classification range and meeting more complex coating detection scenarios compared to Embodiment 1; the dataset of 400 samples further improves the model's generalization ability. Testing has verified that the model has high classification accuracy and good stability, and the generated test set confusion matrix is ​​as follows. Figure 7 As shown, the accuracy rate reaches 97.50%, which can effectively distinguish between 1-4 different coatings, has strong anti-interference ability, and can quickly and accurately complete the automatic classification of coating layers, greatly improving the efficiency and accuracy of coating detection, and has stronger engineering application value.

[0028] This technology also tested the recognition effect of different types of coating layers, and the automatic recognition results are as follows: Figure 8 As shown in the results, the model accurately outputs the predicted number of layers for samples from various industrial applications, including aircraft coatings, automotive coatings, thermal barrier coatings, and oil and gas pipeline coatings. This demonstrates that the technology can adapt to complex and ever-changing practical testing needs, providing universal technical support for non-destructive testing of coatings in multiple fields.

[0029] Therefore, this invention employs the aforementioned automatic coating layer classification method based on terahertz time-domain spectroscopy. Without requiring prior knowledge of the coating's specific layer structure or any other prior information, it achieves rapid and accurate identification of unknown multilayer coating structures through kernel principal component analysis (KPCA) feature extraction combined with a support vector machine (SVM) model optimized by differential evolution (DE). This method not only overcomes the limitations of traditional optimization methods such as grid search, which are computationally time-consuming and prone to getting trapped in local optima, but also solves the problems of conventional algorithms struggling with feature extraction and blurred layer identification boundaries when dealing with thin coatings or complex boundaries, ultimately leading to the failure of subsequent thickness inversion models. It is particularly suitable for the automatic intelligent identification of the thickness of complex coatings with unknown layers in industrial environments, possessing greater engineering practical value and automation potential.

[0030] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. An automatic classification method for the number of coating layers based on terahertz time-domain spectroscopy, characterized in that the steps are as follows: include: S1. Build a detection platform that combines a terahertz time-domain spectroscopy system with a robotic arm, calibrate the probe posture so that the terahertz wave is incident perpendicularly on the sample to be tested, and collect time-domain spectral data of coatings with different numbers of layers to construct a sample set. S2. Denoise and frequency domain transformation are performed on the time-domain spectral data in S1 to extract time-domain and frequency-domain features; S3. Use kernel principal component analysis to reduce the dimensionality of the features extracted in S2, and select the top N principal components with the highest cumulative contribution rate as the input feature vector. S4. Divide the input feature vectors in S3 into training set and test set and perform mean normalization. S5. Based on the normalized training set in S4, construct a radial basis function kernel support vector machine classification model, and optimize the core parameters of the classification model through differential evolution algorithm to obtain the optimal parameter combination; S6. Train the classification model optimized by S5 using the normalized training set, and then verify the performance of the classification model using the normalized test set. S7. After processing the spectral data of the sample to be tested using the same method as in steps S2-S3, process it according to the normalization standard in S4, input it into the classification model verified in S6, and the classification model will automatically output the number of coating layers.

2. The automatic coating layer number classification method based on terahertz time-domain spectroscopy according to claim 1, characterized in that, S1 specifically includes: S11. Platform Construction: An automatic scanning and detection platform is built by combining a terahertz time-domain spectroscopy system with a robotic arm. The probe posture is calibrated by a laser displacement sensor so that the terahertz wave is perpendicularly incident on the surface of the sample to be tested, and time-domain spectral data of different coating layers are collected. S12. Construct a dataset: Summarize the time-domain spectral data collected in S11 into a sample set, with each sample containing complete time-domain waveform points.

3. The automatic coating layer number classification method based on terahertz time-domain spectroscopy according to claim 1, characterized in that, S2 specifically includes: S21. Noise reduction processing: Perform ensemble empirical mode decomposition noise reduction processing on the original terahertz signal in S1. S22, Frequency Domain Transformation: Perform a Fast Fourier Transform on the time-domain signal processed in S21 to obtain the corresponding frequency-domain spectral data; S23. Extract the time-domain features of the signal processed by S21 and the frequency-domain features of the frequency-domain spectral data obtained by S22.

4. The automatic coating layer number classification method based on terahertz time-domain spectroscopy according to claim 3, characterized in that: The time-domain features in S23 include mean, root mean square, peak-to-peak value, kurtosis, and skewness. Frequency domain characteristics include spectral mean, spectral standard deviation, frequency center, and spectral kurtosis.

5. The automatic coating layer number classification method based on terahertz time-domain spectroscopy according to claim 1, characterized in that, S3 specifically includes: S31. Kernel Principal Component Analysis: The kernel principal component analysis method is used to perform nonlinear mapping and principal component extraction on the original feature data extracted in S2 through kernel functions; S32, Nonlinear Mapping: Based on the kernel function, the original feature data extracted in S2 is mapped to a high-dimensional space, and principal component extraction is completed in the high-dimensional space to achieve feature dimensionality reduction; S33. Feature selection: Select the top N principal components with the highest cumulative contribution rate as the final input feature vector.

6. The automatic coating layer number classification method based on terahertz time-domain spectroscopy according to claim 1, characterized in that, S4 specifically includes: S41. Dataset partitioning: Randomly divide the dimensionality-reduced dataset from S3 into a training set and a test set; S42. Mean Normalization: Perform mean normalization on the training and test sets after S41, mapping the data to... Interval.

7. The automatic coating layer number classification method based on terahertz time-domain spectroscopy according to claim 1, characterized in that, S5 specifically includes: S51. Establish an SVM model: Based on the normalized training set in S4, construct a support vector machine classification model. The penalty coefficient of the classification model is C, and the radial basis function is selected. As the kernel function, the formula is: ； in, and There are two input sample vectors. It is the Euclidean distance between two input sample vectors. These are the kernel parameters of the radial basis functions; S52, Differential Evolution Optimization: The core parameters of the classification model in S51 are optimized by using the differential evolution algorithm to obtain the optimal parameter combination.

8. The automatic coating layer number classification method based on terahertz time-domain spectroscopy according to claim 7, characterized in that, S52 specifically includes: S521. Initialization: Set the population size, scaling factor, crossover probability, and maximum number of iterations for the differential evolution algorithm. Randomly generate a population containing... The initial population for the parameter combination is given by the formula: ； in It represents a random decimal that is uniformly distributed between 0 and 1. , They are respectively Upper and lower bounds of parameter combinations; S522, Mutation and Crossover: Mutation is performed on individuals in the initial population in S521 to generate difference vectors, and crossover is performed to generate experimental individuals; S523. Selection: Use the cross-validation accuracy of the SVM model on the normalized training set in S4 as the fitness function, compare the fitness of the experimental individuals with that of the target individuals, and retain the better parameter combination for the next generation. S524. Iteration: Repeat steps S522-S523 until the maximum number of iterations is reached or the convergence condition is met, and output the globally optimal parameter combination. .

9. The automatic coating layer number classification method based on terahertz time-domain spectroscopy according to claim 1, characterized in that, S6 specifically includes: S61, Model Training: Combining the optimal parameters obtained in S5 The data is assigned to the SVM model, and the classification model is trained using the normalized training set data from S4. S62, Layer Prediction: Input the normalized test set data from S4 into the classification model trained in S61, output the predicted category of each sample, and verify the model performance by evaluating the classification accuracy of the model.

Images