Open Access
Add to BibSonomy
Abstract
Online monitoring technology plays a pivotal role in advancing the utilization of laser paint removal in aircraft maintenance and automation. Through the utilization of a high-frequency infrared pulse laser paint removal laser-induced breakdown spectroscopy (LIBS) online monitoring platform, this research conducted data collection encompassing 60 sets of LIBS spectra during the paint removal process. Classification and identification models were established employing principal component analysis (PCA), partial least squares discriminant analysis (PLS-DA), and orthogonal partial least squares discriminant analysis (OPLS-DA). These models served as the foundation for creating criteria and rules for the online LIBS monitoring of the controlled paint removal process for aircraft skin. In this research, 12 selected characteristic spectral lines were used to construct the OPLS-DA model, with a predictive root mean square error (RMSEP) of 0.2873. Both full spectrum and feature spectral line data achieved a predictive accuracy of 94.4%. The selection of feature spectral lines maintains predictive performance while significantly reducing the amount of input data. Consequently, this research offers a methodological reference for further advancements in online monitoring technology for laser paint removal in aircraft skin.
© 2024 Optica Publishing Group under the terms of the Optica Open Access Publishing Agreement
1. Introduction
Traditional methods for aircraft skin paint removal involve manual mechanical polishing [1], chemical agents [2], and shot blasting techniques [3,4]. These conventional paint removal methods are commonly associated with significant drawbacks, including severe environmental pollution, potential harm to workers’ health, high costs, and high labor intensity. In contrast, laser paint removal, a novel cleaning technology [5], presents several advantages, including its eco-friendliness, precise targeting, high paint removal accuracy, and efficiency. In the future, laser paint removal has the potential to supplant traditional methods for comprehensive aircraft paint removal. Given the intricate complexity of the paint layer system on the aircraft skin surface and the uneven thickness of these paint layers, this research is crucial to ensure the efficacy of cleaning the aircraft skin's paint layers and achieving controlled removal of multiple paint layers in a stratified manner.
Currently, the primary online monitoring technologies encompass image recognition [6] and acoustic monitoring [7]. However, these technologies frequently encounter challenges related to insufficient real-time performance and accuracy. Laser-induced breakdown spectroscopy (LIBS) is a laser-based spectral technique [8] that employs laser pulses to create a plasma on a material's surface and captures the emission spectrum from the resulting plasma. This process facilitates precise measurement and analysis of material composition. LIBS technology is distinguished by its sensitivity, non-contact approach, and rapid analysis speed, rendering it well-suited for online monitoring of the laser paint removal process in multi-layer aircraft skin.
In the realm of online monitoring for laser cleaning using LIBS, Wang [9] integrated LIBS technology and devised a real-time laser cleaning monitoring system to assess the quality of laser cleaning. They conducted an analysis of the elemental composition and LIBS spectra of epoxy resin matrices and carbon fiber reinforcements in carbon fiber reinforced plastics (CFRP) materials, establishing that LIBS technology is suitable for monitoring the cleaning quality of CFRP. Bian [10] employed an online monitoring approach that integrates LIBS with image binarization to optimize laser cleaning parameters for paint on white marble surfaces. This approach not only enhanced the efficiency of cleaning gold and silver paint layers but also ensured the integrity of the marble substrate. In our previous studies [11] performed an analysis of the variation in peak intensities of characteristic elements within LIBS spectra for various paint layers on the multi-layer structure of aircraft aluminum alloy skin. This study validated the applicability of LIBS technology for monitoring the outcomes of laser-selective paint removal and the removal of paint from aircraft skin. Nonetheless, aircraft skin paint systems are intricate, characterized by notable variations in the types of elements found in different paint layers. The paint layers encompass a wide array of elements, and the LIBS spectral data generated during the laser paint removal process can be extensive. Therefore, during the online monitoring process, it is necessary to consider the use of multivariate statistical analysis methods to reduce the volume of analysis data from LIBS spectra and extract essential information.
Principal component analysis (PCA) is a widely employed statistical analysis method for dimensionality reduction in various applications. PCA is highly effective in reducing the dimensionality of LIBS data, mitigating noise, and preserving the majority of the original data information [12]. Chatterjee [13,14] employed PCA to discern LIBS spectra of soil samples, illustrating the effectiveness of combining LIBS and PCA for identifying and classifying soil samples from geothermal areas. In a prior study, the author also classified thermal waters through correlation analysis and PCA. The results indicated that in PCA, thermal waters are grouped into two distinct clusters. Sarkar [15] conducted research on Laser ablation molecular isotopic spectroscopy (LAMIS), a method used for analysis of the boron isotopic composition in ambient air at atmospheric pressure. The study also explored multiple optimization procedures, including the spectral region of analysis, effect of flicker noise from matrix, spectral normalization and pre-treatment procedures. Nevertheless, PCA, being an unsupervised multivariate statistical analysis method, may yield classification bias when identifying approximate spectral data. Partial least squares discriminant analysis (PLS-DA) is a supervised multivariate statistical analysis method tailored to maximize inter-group distinctions based on predetermined classifications, rendering it more suitable for sample identification and classification compared to unsupervised methods like PCA [16]. Sezer [17] employed LIBS to examine durum wheat and common wheat samples in flour and pasta to detect potential adulteration. In PLS-DA studies, the determination coefficient and detection limit for adulteration in durum wheat flour were 0.999 and 0.52%, respectively. The integration of LIBS and PLS-DA enables swift and dependable identification and detection of wheat flour adulteration in pasta. Orthogonal partial least squares discriminant analysis (OPLS-DA) is a non-parametric linear classification technique [18]. Its primary application is constructing relationship models and model predictions among multiple variables and response variables. In contrast to traditional discriminant analysis models, OPLS-DA not only reduces dimensionality but also minimizes noise impact while preserving the strongest correlations. Su [19] explored the potential of combining Raman spectra with the OPLS-DA multivariate modeling method to identify stamp ink types. The outcomes demonstrated that this method effectively identified six distinct brands of stamp ink on two types of paper, achieving a 100% accuracy rate. PCA, PLS-DA, and OPLS-DA are widely employed multivariate statistical analysis methods in the analysis of LIBS data. These methods place varying emphasis on data processing, sample classification, and identification, making them valuable tools in the realm of spectral analysis.
This paper is centered on the paint layers of aircraft aluminum alloy skin and makes use of a high-frequency laser paint removal LIBS online monitoring platform. Laser single-scan cleaning is employed for the removal of topcoat and primer, simultaneously capturing signatures in real-time from the topcoat, primer, and aluminum alloy substrate throughout the removal process. By reducing the dimensionality of the LIBS spectral data and building PCA, PLS-DA, and OPLS-DA models, this research categorizes and distinguishes blended spectral data originating from the topcoat, primer, and aluminum alloy substrate during the laser cleaning process. This research establishes monitoring criteria that serve as a reference for the controlled laser paint removal of aircraft skin.
2. Materials and methods
2.1 Experimental materials
The utilized substrate is the widely adopted 2024-T3 Al-Cu-Mg series hard aluminum alloy, possessing a thickness of about 2 mm. Following anodization treatment in accordance with the manual's paint specifications, the samples received coatings of approximately 30 µm of PPG CA7700 yellow-green epoxy primer, succeeded by about 30 µm of PPG CA8000 white polyurethane topcoat. Subsequently, the samples were sectioned into specimens measuring 30 mm × 30 mm × 2 mm and denoted as TO (Topcoat). In control experiments, samples lacking the white topcoat were designated as PR (Primer), while the aluminum alloy substrate without paint layers was identified as AL (Aluminum alloy substrate). In order to augment the corrosion resistance, wear resistance, and adhesion of the aluminum alloy substrate, a sulfuric acid anodizing treatment was applied before the aluminum alloy coating. This process resulted in the formation of an approximately 5 µm thick oxide film. Figure 1 displays a schematic diagram of the paint layers on the aircraft's aluminum alloy skin.
During aircraft maintenance, repair, overhaul, or passenger-cargo aircraft modification, it becomes essential to assess the condition of the aircraft substrate material, including potential corrosion, damage, and wear. This frequently necessitates partial or complete removal of the aircraft's paint layers on the skin, followed by repainting in adherence to specific spray standards. As outlined in the A320 SRM 51-23-11 manual, the process of removing paint from the aircraft skin necessitates the elimination of the white topcoat and yellow-green primer while ensuring the preservation of the aluminum alloy substrate without damage. Consequently, this paper predominantly examines the LIBS spectra acquired during the laser paint removal process for the intact paint system sample TO, the sample lacking the white topcoat PR, and the sample devoid of paint layers AL.
The three-dimensional micromorphology of samples TO, PR, and AL was characterized using the 3D optical surface profiler S neox (SENSOFAR, Spain). As depicted in Fig. 2(b), the surface roughness (Sa) of the white topcoat on sample TO measures 0.25 µm, indicating a relatively low surface roughness and excellent smoothness. In Fig. 2(d), the Sa of the yellow-green primer on sample PR is 1.58 µm, signifying a larger surface roughness that contributes to strong adhesion. As illustrated in Fig. 2(f), the Sa of sample AL is 0.48 µm, indicating a relatively low surface roughness. In accordance with the standardized spray process for A320 aircraft, the roughness levels comply with the manual requirements.
Fig. 2. Topcoat (a) macro appearance (b) three-dimensional morphology; primer (c) macro appearance (d) three-dimensional morphology; aluminum alloy substrate (e) macro appearance (f) three-dimensional morphology.
2.2 Experimental setup
The laser paint removal LIBS online monitoring platform comprises two primary components: the laser paint removal system and the LIBS online monitoring system. The laser paint removal system includes components like a pulsed laser, field lens, scanning mirror, and laser processing head. The LIBS online monitoring system consists of a fiber optic spectrometer, data acquisition probe, optical fibers, and a central control computer. Figure 3 depicts a schematic diagram of the laser paint removal LIBS online monitoring platform.
The laser utilized is the MFPT-120P (MAXPHOTONICS, China), a 1064 nm tunable Q-switched fiber laser, with a Gaussian distribution of beam energy, the focal length of the field lens is 160 mm. Table 1 displays the primary parameters of the pulsed fiber laser. The spectrometer in use is the AvaSpec-ULS2048CL-4-EVO fiber optic spectrometer, and Table 2 provides an overview of its primary parameters.
Table 1. Primary parameters of the pulsed fiber laser
Table 2. The main technical parameters of fiber spectrometer
Following the emission of the laser cleaning signal, the laser beam is generated by the fiber laser and directed into the scanning mirror. Subsequently, a focusing lens concentrates the laser beam onto the sample surface. The spectrometer captures signals synchronized with the laser cleaning signal, facilitating real-time online acquisition of plasma spectra.
A planar scan of the laser paint removal region was executed, encompassing a 10 mm × 10 mm scanning area. In order to mitigate the influence of pulse energy fluctuations on spectral stability, the spectrometer integration time was configured to 10 ms, enabling the real-time collection of LIBS spectra. The experiments were carried out with laser average power as the sole variable, while maintaining consistent laser parameters, including a scanning speed of 2000 mm/s, line spacing of 0.02 mm, repetition frequency of 100 kHz, pulse width of 350 ns, and a single laser scan. Laser average power was varied to gather 20 sets of LIBS spectra for each of the topcoat, primer, and aluminum alloy substrate removal processes, yielding a total of 60 sets for model training and testing.
2.3 Spectral preprocessing
Spectral signals are obtained during the high-frequency laser planar scanning process. The obtained spectral lines are relatively wide and are susceptible to influences like stray light, noise, and a pronounced continuous background. Such interference can affect the qualitative analysis of LIBS spectral signals. Hence, preprocessing of the raw spectra is essential. The commonly employed high-precision baseline correction method for eliminating background noise from spectra is adaptive iteratively reweighted penalized least squares (airPLS) [20]. We employed the airPLS algorithm to estimate and correct baselines based on raw spectral data. This algorithm fits the baseline by locally modeling the data and incorporates a penalty term to balance the fitted spectral baseline and the characteristic signals in the samples. This approach effectively prevents overfitting of the baseline. The savitzky-golay (S-G) filter is a time domain filtering method grounded in local polynomial least squares fitting [21,22]. Its primary benefit lies in its capability to eliminate random noise within the spectral signal while retaining the signal's shape and width, thus enhancing the signal-to-noise ratio of the spectral lines. In the S-G smoothing algorithm, two primary parameters are involved: the window size and the polynomial order. We chose a window size of 5 and a polynomial order of 2 to fit the spectral data, aiming to achieve the desired smoothing effect. Min-max normalization standardizes each feature's range to [0,1], ensuring the spectra are uniformly scaled for data analysis. Figure 4 displays representative raw spectrum of unpainted aluminum alloy, spectrum subsequent to baseline correction and S-G smoothing, and spectrum following min-max normalization. We utilized the airPLS, S-G smoothing, and min-max normalization algorithms to preprocess the spectral data in the Matlab R2022a software.
Fig. 4. LIBS spectral preprocessing for unpainted aluminum alloy (a) raw spectrum (b) baseline correction and S-G smoothing (c) min-max normalization.
3. Results and discussion
The adoption of planar scanning during laser paint removal results in relatively broad spectral lines that are susceptible to stray light and noise interference. These spectra frequently feature prominent continuous backgrounds, containing substantial interference information. Moreover, the topcoat and primer compositions share certain common chemical elements, leading to similarities in the morphology of topcoat and primer LIBS spectra, as depicted in Fig. 5. Consequently, PCA, PLS-DA, and OPLS-DA models have been selected to define the criteria for online monitoring of laser paint removal via LIBS, facilitating the identification and categorization of LIBS spectra throughout the laser paint removal process.
3.1 Laser paint removal LIBS spectral clustering analysis based on PCA model
Given the substantial volume of laser paint removal LIBS spectral data, coupled with considerable noise, dimensionality reduction becomes essential for effective clustering analysis. Principal component analysis (PCA) is an unsupervised dimensionality reduction algorithm that reconfigures multiple indicators into a handful of principal components. These principal components are linear combinations of the original variables, exhibit no correlation with one another, and have the capacity to capture the majority of the information from the original data. Following PCA processing, Fig. 6 illustrates the interpretation rate and cumulative interpretation rate of the initial 6 principal components of laser paint removal LIBS spectral data.
PC1, PC2 and PC3 together accumulate an interpretation rate of 90.2%, essentially encapsulating the primary information within the LIBS spectral data. To demonstrate the PCA model in 3D, scores for the top 3 principal components of the LIBS spectral data are computed, as depicted in Fig. 7.
Upon scrutinizing the 3D score scatter plot generated by the PCA model, it becomes apparent that the LIBS spectra of the topcoat, primer, and aluminum alloy substrate during the laser cleaning process predominantly fall within two distinct regions in the scores of the top 2 principal components. This facilitates a distinct demarcation between the paint layers and the aluminum alloy substrate, illustrating discernible clustering patterns. Nevertheless, the LIBS spectra of the topcoat and primer manifest similarities that pose challenges in their differentiation. Given the requisite control of paint layer thickness for stratified and precise paint removal in laser paint removal on aircraft skin, it is imperative to construct an additional model to attain accurate classification of spectral samples.
3.2 Laser paint removal LIBS spectral clustering analysis based on PLS-DA model
PLS-DA has the capacity to mitigate the impact of multicollinearity among variables, thus simplifying the identification of inter-group disparities and yielding improved clustering outcomes. An established PLS-DA model is employed to distinguish the LIBS spectra collected during the laser paint removal process for the topcoat, primer, and aluminum alloy substrate. The 60 sets of LIBS spectral data from the laser paint removal samples undergo preprocessing, and they are subsequently partitioned into training and testing sets at a 7:3 ratio. The PLS-DA model is constructed utilizing SIMCA-P 14.1 software, which automatically fitted with three principal components, resulting in a cumulative interpretation rate of 90.3%. Notably, sample inter-group disparities are more prominent when compared to the PCA model, culminating in the generation of a PLS-DA 3D score scatter plot, as depicted in Fig. 8.
Nevertheless, PLS-DA, being a supervised discriminant analysis method, explicitly defines and groups samples during model construction, enhancing its ability to distinguish inter-group disparities. However, it is susceptible to potential issues of model overfitting. Hence, a permutation test is performed on the PLS-DA model to evaluate the presence of overfitting. This test entails the random permutation of classification variables 200 times, as illustrated in Fig. 9.
The intercept of the R2 fitting line on the left vertical axis is 0.013, while the Q2 fitting line has an intercept of -0.265. A reliable PLS-DA model should exhibit an intercept on the Y-axis for R2 less than 0.3, and for Q2 less than 0.05 [23]. Both the R2 and Q2 values obtained from random permutations are lower than the initial values, indicating that the predictive capability of this PLS-DA model surpasses that of any individual random permutation of the y variable. This demonstrates that the model is not afflicted by overfitting. As depicted in Fig. 8, the PLS-DA model demonstrates superior clustering in contrast to the PCA model. Nevertheless, there are instances of similarities in the LIBS spectra of laser-cleaned topcoat and primer. This could be attributed to the high similarity of spectral data acquired at the boundary between laser-cleaned topcoat and primer, rendering differentiation challenging. Consequently, a more suitable approach is required to distinguish between samples exhibiting significant category similarities.
3.3 Laser paint removal LIBS spectral clustering analysis based on OPLS-DA model
OPLS-DA models differ from PLS-DA models primarily in the introduction of orthogonality. This change results in a partial change in the weights. In comparison to PLS-DA models, OPLS-DA models introduce orthogonal weights, incorporating both predictive and orthogonal weights into the total weight. In the OPLS-DA model, predictive weights are mainly utilized to maximize the differences between categories, while orthogonal weights are employed to maximize the differences within categories. This design effectively simplifies the weight set by removing variations in directions unrelated to the Y variable.
The basic establishment process of OPLS-DA is as follows: 1) Based on the established high-frequency infrared pulse laser paint removal LIBS online monitoring platform, the LIBS spectra during the laser paint removal process are collected in real-time. 2) The raw spectral data undergo baseline correction, S-G smoothing, and min-max normalization preprocessing. 3) Utilizing the preprocessed spectral data, training set samples and testing set samples are created, with the assignment of variable values for three categories—the response variables. 4) Conducting regression analysis on the LIBS spectral data of the training set samples and categorical variables, an OPLS-DA model is established. The model's fit is then evaluated and analyzed. 5) Using the established OPLS-DA model, calculate the response variable values for the testing set samples. Subsequently, categorize the testing set samples based on these values to assess the reliability and predictive capability of the identification model through validation analysis.
OPLS-DA excels in eliminating unrelated data variations in the independent variables, thereby concentrating the primary classification information in a single principal component, leading to a more distinctive discriminative outcome. The 60 sets of spectral data obtained from laser-cleaned samples underwent preprocessing. Subsequently, these spectral data were utilized as input variables to build an OPLS-DA model for laser paint removal process LIBS spectroscopy. Figure 10 illustrates the 3D score scatter plot for OPLS-DA.
The OPLS-DA model demonstrates outstanding performance with cumulative interpretive ability parameters, achieving R2X of 0.961 for the independent variables and R2Y of 0.883 for the dependent variables. The parameter R2X denotes the extent to which the model elucidates the input data (X matrix), whereas R2Y signifies the model's capacity to explain the response variable (Y matrix). Both R2X and R2Y values fall within the range of 0 to 1. A heightened R2X value typically indicates the model's effectiveness in elucidating the data, reflecting its success in providing a robust depiction of the input data. Similarly, an elevated R2Y value often indicates the model's substantial explanatory power for the response variable, signifying a well-fitted model to the Y matrix. The predictive ability parameter, Q2, reaches 0.777. The root mean square error of cross-validation (RMSECV) serves as a metric for evaluating the predictive performance of a model on new data. A diminished RMSECV value signifies enhanced predictive performance of the model on new data. The RMSECV value for the OPLS-DA model is 0.0862. Collectively, these values attest to the model's robust interpretive and predictive capabilities, as well as its high stability. For a more in-depth evaluation of the OPLS-DA model's reliability, a permutation test was conducted. This involved random permutations of the classification variable 200 times, as depicted in Fig. 11.
The results depicted in Fig. 11 show that both fitted lines exhibit relatively steep slopes. The R2 intercept is 0.025, less than the threshold of 0.3, while the Q2 intercept stands at -0.255, also less than 0. These values indicate that the OPLS-DA model is free from overfitting concerns. Cross-validation analysis of variance (CV-ANOVA) is a statistical method employed to evaluate the significance of a model and assess its ability to effectively discriminate between different categories. This approach is commonly utilized to appraise the performance of OPLS-DA models. In CV-ANOVA, the F-value and P-value serve as statistical indicators to determine the existence of significant differences between categories. The F-value gauges the ratio of between-category variance to within-category variance, while the P-value is employed to ascertain the significance of the F-value. A higher F-value signifies a more pronounced between-category variance relative to within-category variance, indicating superior performance of the model in distinguishing between different categories. If the P-value is smaller than the selected significance level (typically 0.05), there exists ample evidence to reject the null hypothesis, signifying significant differences between categories in the model. CV-ANOVA was conducted on the model, revealing an F-value of 8.152 with a P-value of 3.36 × 10−12, which is less than 0.05. These results confirm the model's statistical significance. The 60 sets of LIBS spectral data obtained during the laser paint removal process were split into training and testing sets with a 7:3 ratio. The root mean square error of estimation (RMSEE) serves as a metric for evaluating the fit of a model to training data. A reduced RMSEE value signifies an improved fit of the model to the training data. RMSEE for the fitted model is 0.2472, and the RMSECV is 0.2912. These values indicate that the model demonstrates a good fit and strong stability. Figure 12 provides a visual representation of the training and prediction outcomes of the OPLS-DA model.
To evaluate the stability of the model and ascertain the presence of overfitting, a permutation test and cross-validation variance analysis were conducted on the OPLS-DA regression model. The permutation test results revealed an R2 value of 0.183 and a Q2 value of -0.551, both associated with relatively steep slopes, indicating the absence of overfitting in the model. The cross-validation analysis yielded an F-value of 21.271 and a P-value of 3.44 × 10−11, affirming the model's statistical significance and strong stability. Model testing, as depicted in Fig. 12(b), entailed the random selection of LIBS spectral data from the three sample categories for prediction. The model obtained a root mean square error of prediction (RMSEP) of 0.2238, and the predicted values closely matched the actual values. We conducted sequential training and prediction of OPLS-DA model using datasets collected in different batches. The results indicate that the OPLS-DA model can effectively classify and identify mixed data containing spectra from topcoat, primer, and aluminum alloy. The predictive accuracy of the model reached 94.4%. This demonstrates that the model provides excellent predictive accuracy, making it suitable for cluster analysis and prediction of LIBS full spectrum data during the laser paint removal process. Although utilizing full spectrum data effectively retains spectral information, the substantial volume of data can hinder data analysis efficiency. Achieving prompt identification and classification of LIBS spectra during laser paint removal requires a reduction in the input of spectral data.
3.4 Laser ablation LIBS feature spectrum clustering analysis based on OPLS-DA
The model constructed using full spectrum data helps mitigate the loss of spectral information to a certain degree. Nonetheless, full spectrum data is vulnerable to interference from stray light and noise during collection, resulting in prominent continuous backgrounds. This elevates analytical complexity and diminishes efficiency. Selecting characteristic spectral lines not only enables the differentiation of various paint layers and substrates but also substantially reduces the volume of data input. In previous study [24], we proposed a method for selecting characteristic peaks by identifying feature elements in different materials to represent characteristic spectra. This approach proves particularly crucial for the analysis of spectral data. Consequently, this research centered on the examination of the OPLS-DA model employing characteristic spectral lines. In the case of the frequently employed 2024-T3 Al-Cu-Mg series hard aluminum alloy found in aircraft skin, the predominant elemental component of the substrate is Al. However, the chemical composition of the topcoat and primer varies. The topcoat exhibits a higher concentration of Ti, while the primer shows elevated levels of Cr and Sr, as depicted in the element concentration comparison in Fig. 13.
Leveraging disparities in chemical elements between the aluminum alloy substrate and paint layers and referencing the National Institute of Standards and Technology (NIST) Atomic Spectra Database, 12 characteristic spectral lines were selected from the LIBS spectra obtained during the laser paint removal process, including Al, Ti, Cr, Sr, and Ba elements (as shown in Table 3). This selection of characteristic spectral lines resulted in a data input volume reduction by three orders of magnitude compared to using the entire spectrum data.
Table 3. Selection of characteristic spectral lines
Employing the same training and testing set split ratio as the full LIBS spectral model, constructed an OPLS-DA model based on the first three principal components for analysis, using the 12 selected spectral lines as input data. A permutation test was performed to validate the reliability of the OPLS-DA model for analyzing laser paint removal LIBS spectral features. The R2 intercept was determined to be 0.010, which is less than 0.3, and the Q2 intercept was -0.239, also less than 0.
Partitioned the 60 sets of spectral data chosen based on characteristic spectral lines into training and testing sets, maintaining a 7:3 ratio. The regression model produced an RMSEE of 0.2654 and an RMSECV of 0.2646. These findings indicate that the regression model is reasonably reliable and free from overfitting. Figure 14 presents the training and prediction outcomes of the OPLS-DA regression model utilizing characteristic spectral lines.
Fig. 14. OPLS regression model for characteristic spectral lines (a) training results (b) prediction results.
To evaluate the reliability of the OPLS-DA regression model based on characteristic spectral lines and ascertain the absence of overfitting, conducted permutation tests and cross-validation analysis. The permutation tests disclosed R2 values of -0.052 and Q2 values of -0.175, along with a pronounced slope in the fitted line, confirming the absence of overfitting in the model. The cross-validation analysis demonstrated an F-value of 166.169 and a P-value of 8.23 × 10−20, signifying a high level of model reliability. Figure 14(b) displays the predictions for the testing set. The regression model achieved an RMSEP of 0.2873, indicating a precise alignment between predicted and actual values, resulting in a prediction accuracy of 94.4%. These findings align with those derived from the OPLS-DA model utilizing full spectrum data, laying a theoretical foundation for the differentiation of topcoat, primer, and aluminum alloy substrate spectra in the laser paint removal process.
4. Conclusion
In this work, a high-frequency (kHz-level) nanosecond pulse laser served as the laser source for online monitoring of laser paint removal with controlled layering on aircraft skin. The study involved the analysis of LIBS spectra obtained from the topcoat, primer, and aluminum alloy substrate during the paint removal process. Established PCA, PLS-DA, and OPLS-DA models. The cumulative interpretation rate of the first three principal components in the PCA model is 90.2%. With the first three principal components, the PLS-DA model attains a cumulative interpretation rate of 90.3%. The OPLS-DA model demonstrates a cumulative interpretation rate of 96.1% for the independent variables. Through cross-validation variance analysis, an F-value of 8.152 was obtained with a P-value of 3.36 × 10−12, which is less than 0.05. The RMSEP of the regression model is 0.2238, signifying a high degree of stability, reliability, and accuracy, with a predictive accuracy of 94.4%. Furthermore, this work conducted a comparison of OPLS-DA models using full spectrum and characteristic spectral line data for recognizing different materials within LIBS spectra. The results demonstrated that the OPLS-DA model based on characteristic spectral line data also achieved a prediction accuracy of 94.4%. By selecting characteristic spectral lines to construct the model, data input was substantially reduced, leading to improved data analysis efficiency and reduced complexity. This approach facilitated the swift classification, recognition, and prediction of LIBS spectra data from the topcoat, primer, and aluminum alloy substrate, development of criteria for LIBS online monitoring in laser-controlled layered paint removal from aircraft skin. The integration of LIBS technology with the OPLS-DA model enhanced the applicability of characteristic spectral line data, serving as a methodological reference for advancing online monitoring technology in aircraft skin laser paint removal with controlled layering.
Funding
National Natural Science Foundation of China (52205239); Sichuan Science and Technology Program (2022NSFSC1903); Fundamental Research Funds for the Central Universities (ZJ2022-006); Deyang Science and Technology Plan Project (2022GZ011).
Disclosures
The authors declare no conflicts of interest.
Data availability
Data underlying the results presented in this paper are not publicly available at this time but may be obtained from the authors upon reasonable request.
References
1. E. Irwin and K. Streilein, “Use of Field-based Motion Capture to Augment Observational data in ergonomic Assessment of Aircraft Maintenance,” Procedia Manuf. 3, 4501–4508 (2015). [CrossRef]
2. S. Uang, T. Shih, C. Chang, et al., “Exposure assessment of organic solvents for aircraft paint stripping and spraying workers,” Sci. Total Environ. 356(1-3), 38–44 (2006). [CrossRef]
3. Y. K. Madhukar, S. Mullick, and A. K. Nath, “Development of a water-jet assisted laser paint removal process,” Appl. Surf. Sci. 286, 192–205 (2013). [CrossRef]
4. A. Moridi, S. M. Hassani-Gangaraj, S. Vezzú, et al., “Fatigue behavior of cold spray coatings: The effect of conventional and severe shot peening as pre-/post-treatment,” Surf. Coat. Technol. 283, 247–254 (2015). [CrossRef]
5. X. Li, H. Wang, W. Yu, et al., “Laser paint stripping strategy in engineering application: A systematic review,” Optik 241, 167036 (2021). [CrossRef]
6. Á Borsos, B. Szilágyi, PŞ Agachi, et al., “Real-Time Image Processing Based Online Feedback Control System for Cooling Batch Crystallization,” Org. Process Res. Dev. 21(4), 511–519 (2017). [CrossRef]
7. W. Zou, F. Song, and Y. Luo, “Characteristics of audible acoustic signal in the process of laser cleaning of paint on metal surface,” Opt. Laser Technol. 144, 107388 (2021). [CrossRef]
8. L. J. Radziemski, “From LASER to LIBS, the path of technology development,” Spectrochim. Acta. Part B At. Spectrosc. 57(7), 1109–1113 (2002). [CrossRef]
9. W. Wang, L. Sun, Y. Lu, et al., “Laser induced breakdown spectroscopy online monitoring of laser cleaning quality on carbon fiber reinforced plastic,” Opt. Laser Technol. 145, 107481 (2022). [CrossRef]
10. T. Bian, Y. Bai, L. Yu, et al., “Laser cleaning of paint layers on white marble surface based on cooperative use of laser-induced breakdown spectroscopy and image binarization,” J. Cult. Herit. 62, 124–133 (2023). [CrossRef]
11. W. Yang, Z. Qian, Y. Cao, et al., “LIBS Monitoring and Analysis of Laser-Based Layered Controlled Paint Removal from Aircraft Skin,” J. Spectrosc. 2021, 1–12 (2021). [CrossRef]
12. D. Diaz, A. Molina, and D. W. Hahn, “Laser-Induced Breakdown Spectroscopy and Principal Component Analysis for the Classification of Spectra from Gold-Bearing Ores,” Appl. Spectrosc. 74(1), 42–54 (2020). [CrossRef]
13. S. Chatterjee, M. Singh, B. P. Biswal, et al., “Application of laser-induced breakdown spectroscopy (LIBS) coupled with PCA for rapid classification of soil samples in geothermal areas,” Anal. Bioanal. Chem. 411(13), 2855–2866 (2019). [CrossRef]
14. S. Chatterjee, A. Sarkar, A. S. Deodhar, et al., “Geochemical and isotope hydrological characterisation of geothermal resources at Godavari valley, India,” Environ. Earth Sci. 76(2), 97 (2017). [CrossRef]
15. A. Sarkar, X. Mao, and R. E. Russo, “Advancing the analytical capabilities of laser ablation molecular isotopic spectrometry for boron isotopic analysis,” Spectrochim. Acta Part B At. Spectrosc. 92, 42–50 (2014). [CrossRef]
16. O. Özdikicierler, “Chemometric discrimination of Turkish olive oils by variety and region using PCA and comparison of classification viability of SIMCA and PLS-DA,” Eur. Food Res. Technol. 247(1), 157–168 (2021). [CrossRef]
17. B. Sezer, A. Unuvar, I. H. Boyaci, et al., “Rapid discrimination of authenticity in wheat flour and pasta samples using LIBS,” J. Cereal Sci. 104, 103435 (2022). [CrossRef]
18. Q. Ma, X. Chen, K. Zhang, et al., “Chemical Fingerprint Analysis for Discovering Markers and Identifying Saussurea involucrata by HPLC Coupled with OPLS-DA,” J. Anal. Methods Chem. 2020, 1–8 (2020). [CrossRef]
19. D. Su and K. Li, “Micro Raman spectroscopy combined with chromatographic effect and OPLS-DA for time-resolved detection of stamp inks,” Microchem. J. 190, 108727 (2023). [CrossRef]
20. Z. Zhang, S. Chen, and Y. Liang, “Baseline correction using adaptive iteratively reweighted penalized least squares,” Analyst 135(5), 1138–1146 (2010). [CrossRef]
21. X. Liu, S. Qiao, and Y. Ma, “Highly sensitive methane detection based on light-induced thermoelastic spectroscopy with a 2.33 µm diode laser and adaptive Savitzky-Golay filtering,” Opt. Express 30(2), 1304–1313 (2022). [CrossRef]
22. Y. Shao, Y. He, and J. Mao, “Quantitative analysis of bayberry juice acidity based on visible and near-infrared spectroscopy,” Appl. Opt. 46(25), 6391–6396 (2007). [CrossRef]
23. S. Xu, C. Bai, Y. Chen, et al., “Comparing univariate filtration preceding and succeeding PLS-DA analysis on the differential variables/metabolites identified from untargeted LC-MS metabolomics data,” Anal. Chim. Acta 1287, 342103 (2024). [CrossRef]
24. D. Lin, W. Yang, Z. Qian, et al., “Real-time LIBS monitoring of laser-based layered controlled paint removal from aircraft skin based on random forest,” Appl. Opt. 62(10), 2569–2576 (2023). [CrossRef]
