Aircraft paint cleaning monitoring based on laser induced breakdown spectroscopy (LIBS) technology requires a limited peak power density range to ensure the stability of plasma excitation and paint cleaning. However, for the widely used high-frequency (kHz-level) pulsed laser paint removal technology, its peak power density is relatively low, and plasma excitation is limited during the paint removal process; and the strong continuous background spectrum generated by high-frequency laser ablation materials interferes with plasma spectrum acquisition. Based on the controllable cleaning requirements of the skin functional paint layer, this paper designs a LIBS monitoring platform suitable for high-frequency laser paint removal based on the control software writing of the LabVIEW embedded development system and the integration of laser cleaning, spectrum acquisition, control and display modules. The 2024-T3 aluminum alloy double-layer paint sample was selected as the research object, and the paint layer/substrate system spectrum (topcoat layer: Tc; primer layer: Pr; substrate: As) in the wavelength range of 360-700 nm was collected. The original spectrum was preprocessed by smoothing filtering, baseline correction and normalization, and 12 characteristic spectral lines were selected for principal component analysis (PCA). The dimension reduction data was used as the input variable of linear discriminant analysis (LDA) to establish the PCA-LDA discriminant model. Finally, the constructed model was imported into the LIBS monitoring platform, and the classification accuracy of the high-frequency laser paint removal LIBS monitoring platform was verified through experiments. The results show that the principle of selecting the principal components only when the cumulative variance explanation rate is greater than 85% cannot meet the classification needs of LDA in the paint removal process; by optimizing the number of principal components of LDA, the first 9 principal components are finally selected as the input of LDA, which significantly improves the detection accuracy of the LIBS platform. At this time, the classification accuracy of the PCA-LDA model based on the LIBS spectrum reaches 92.5%. It can be seen that the designed high-frequency laser paint removal LIBS monitoring platform can complete the material identification of different structural layers of the paint layer/substrate system, thereby realizing the effective monitoring of high-frequency pulse laser controllable paint removal.

During aircraft inspection and maintenance, the original paint layer needs to be removed in a controllable manner. Common cleaning methods include chemical cleaning, mechanical grinding, water jet, etc., but the existing processes have technical limitations and cannot better meet the high-precision cleaning needs of aircraft paint layers. Laser controllable cleaning is a cleaning method that is controlled and adjusted according to the actual needs of aircraft skin paint removal. Due to its “green, environmentally friendly, and efficient” technical advantages, it has become a research hotspot in the aviation manufacturing and maintenance industry.

At present, the laser cleaning monitoring methods mainly include acoustic monitoring, image recognition, and laser induced breakdown spectroscopy (LIBS) monitoring. Among them, the LIBS technology has the characteristics of small damage, fast, in-situ, and no sample pretreatment, which makes it widely used in the field of laser cleaning monitoring. However, since the spectrum collected during the continuous surface scanning process of laser paint removal is a high-dimensional data formed by the integral accumulation of plasma emission light, the traditional data processing method cannot meet the needs of efficient controllable cleaning monitoring; and the types of elements contained in the paint layer/matrix system are similar, it is not convenient to carry out efficient spectral interpretation and analysis based only on the peak position and peak intensity information of the characteristic peaks of the elements. In order to reveal the inherent laws of LIBS spectral data and realize the identification and classification of unknown samples, domestic and foreign scholars have carried out a lot of research work on machine learning algorithms such as spectral preprocessing and classification modeling. Tong Yanqun et al. used the K-nearest neighbor algorithm to conduct real-time evaluation and automatic classification of the cleanliness of the sample surface. When K=3, a 100% cleanliness level classification accuracy rate can be achieved. Shen et al. evaluated the ability of support vector machine (SVM) and random forest (RF) machine learning methods to predict unknown iron ore samples, proposed a LIBS-RF iron ore sample identification method, and then formed a LIBS ore classification judgment rule. The development of machine learning algorithms provides technical support and guarantee for LIBS technology in the field of laser controllable cleaning monitoring.

In recent years, with the continuous increase in monitoring needs and the innovation of monitoring component technology, domestic and foreign research teams have developed a variety of laser cleaning monitoring platforms based on LIBS technology. Halah A. Jasim’s team developed a high-frequency laser aluminum alloy paint cleaning monitoring system. The study found that the intensity of the characteristic peak spectrum is positively correlated with the depth of paint removal. Sun Lanxiang et al. developed a high-frequency laser carbon fiber composite cleaning monitoring system. Through Pearson correlation coefficient analysis, the automatic control process of laser cleaning was optimized. The excitation threshold of the plasma in the high-frequency laser paint removal process is different from the paint layer cleaning threshold. It is not easy to take into account the paint removal effect and the effective excitation of the spectrum; compared with low-frequency laser excitation, the high-frequency laser excitation spectrum contains a strong continuous background spectrum, and the characteristic spectral line acquisition is disturbed. In order to overcome these problems, a single high-frequency pulsed laser should be used to take into account both paint layer cleaning and spectral excitation. While ensuring the paint removal effect and spectral excitation, the monitoring platform needs to have stable spectral acquisition and efficient classification and recognition capabilities. Therefore, in order to meet the needs of controllable cleaning monitoring of aircraft paint layers, the design of a monitoring platform suitable for high-frequency laser paint removal has significant application value.

Aiming at the demand for controllable cleaning of aircraft skin and the problem of high-frequency laser paint removal monitoring, this paper designs a LIBS monitoring platform suitable for high-frequency laser paint removal based on laser induced breakdown spectroscopy technology. The platform mainly includes laser cleaning module, spectrum acquisition module, control and display module. Based on LabVIEW embedded development system, the high-frequency laser paint removal LIBS system control software is written, and the PCA-LDA discrimination model of different structural layers of the paint layer/substrate system is established to realize the effective monitoring of high-frequency pulse laser controllable paint removal.

1 Design of LIBS monitoring platform for high-frequency laser paint removal

1.1 Platform design principle
The high-frequency laser focuses on the surface of the paint layer to induce the generation of high-temperature plasma. After the laser pulse ends, the excited atoms and ions undergo energy level transitions to produce characteristic spectra. Since the strong continuous background caused by high-frequency laser surface scanning drowns out the spectral lines of atoms and ions, the traditional LIBS platform cannot meet the needs of paint removal monitoring. Therefore, this device suppresses the continuous background intensity and extracts available spectral information from it by optimizing the spectral acquisition path. Then, through the information interaction processing of spectral acquisition, laser cleaning and control and display modules, the materials of different structural layers of the paint layer/substrate system are identified as monitoring indicators to realize the controllable cleaning monitoring of high-frequency laser paint removal. The platform is mainly composed of laser cleaning module, spectral acquisition module, control and display module. Among them, the control module is mainly composed of three functional components: data processing, timing control and closed-loop control, which mainly completes the platform performance regulation and spectral data analysis and interpretation. The laser cleaning module is composed of a high-frequency laser, a galvanometer, and a focusing field mirror, which plays the role of paint layer cleaning and spectral excitation. The spectrum collection system consists of Glan prism, collection lens, filter, optical fiber, and spectrometer. This module converts the spectrum into photoelectricity and transmits the information to the control module. The structure of the LIBS monitoring platform based on high-frequency laser paint removal is shown in Figure 1.

1.2 Platform hardware design
In order to meet the needs of laser cleaning and LIBS monitoring, the platform hardware system is designed in accordance with the principle of “economic applicability” and the idea of ​​”overall planning and step-by-step implementation”.

1.2.1 Laser cleaning module
The selection of laser for the laser cleaning module depends on the application scenario and laser wavelength requirements. The fiber optic pulse laser (model: MFPT-120P, wavelength: 1064nm, frequency range: 1-200kHz, power range: 0-120W, pulse width range: 60-350ns, beam energy follows Gaussian distribution) was selected as the platform light source due to its good optical quality, small pulse energy fluctuation, and high repetition frequency. The cleaning processing head adopts the galvanometer scanning optical path principle. The galvanometer includes an X scanning galvanometer group, a Y scanning galvanometer group, and an electronic drive amplifier. The galvanometer is connected to the control board of the control cabinet system through a signal line. The drive amplifier receives the action command and introduces the laser beam into the F-theta lens along the X and Y directions through the synchronous action of the two servo-controlled rotating mirrors, thereby focusing on the surface of the paint layer. The dust removal equipment includes a vacuum cleaner and a rubber hose. The dust removal equipment is installed on the side of the F-theta lens, which optimizes the spectral excitation environment of the paint removal process and reduces the risk of damage to the F-theta lens.
The optical path of the cleaning module is shown in Figure 2.

1.2.2 Spectral acquisition module
The spectrometer is the core component of the spectrum acquisition module. In order to determine the requirements of the monitoring platform for the overall type and specific parameters of the spectrometer, the spectrum acquisition effects of different spectrometers are compared and analyzed. According to the preliminary experiment, the Ocean Optics MX2500+ fiber optic spectrometer was selected. The spectrometer consists of a collimator, a diffraction grating, a focusing optical system and a detector. The spectrometer uses a planar array CCD technology with high resolution and signal-to-noise ratio, and its measurement range is between 340 and 720 nm. At the same time, the spectrometer is combined with an integrated circuit board to provide USB3.0 and Gigabit Ethernet communication interfaces, so that the spectrometer can be efficiently integrated into the laser paint removal LIBS platform.
In order to solve the problem of plasma spectrum acquisition in the high-frequency laser paint removal monitoring system, this study optimizes the spectrum acquisition path and adds Glan prisms and filters. The schematic diagram of the acquisition path optimization design is shown in Figure 3 (a). Due to the chromatic aberration of the lens, it has different focusing characteristics for various wavelengths of plasma emission light. Therefore, the relative positions of the end faces of each device in the acquisition path must be accurately aligned to achieve efficient coupling of the emitted light. The above devices are all encapsulated in a dust-free box to reduce the interference of environmental factors on spectral line acquisition. Before and after optimization, the spectrum is collected with an integration time of 30ms. The characteristic spectrum of the paint layer in the 360-430nm band is taken as an example, as shown in Figure 3 (b). The shapes of the spectral curves before and after path optimization are roughly similar, but the optimized acquisition path effectively reduces the intensity of the spectrum continuum and improves the quality of spectrum acquisition.

1.2.3 Control and display module
While the control module completes the performance regulation of the platform, it is also necessary to judge the quality of the plasma spectrum associated with the paint removal process and monitor the paint removal effect based on the real-time spectral characteristic information of the paint layer. Therefore, the control module requires faster data processing capabilities and high-performance peripheral interfaces to take into account both paint removal effects and effective spectral excitation. This platform selects ODROID DXU4 as the control unit, adopts ARM architecture, and is equipped with a Samsung Excelynos 5422 eight-core processor with a main frequency of 2 GHz. The motherboard is equipped with an Ethernet interface, 3 universal serial interfaces and an HDMI audio interface to meet the computing needs of data interaction. The display module selects a DELL 15.6-inch touch screen with a resolution of 1920 pixels × 1080 pixels. The actual control unit is shown in Figure 4.

1.3 Platform software design
Combined with the working principle of the high-frequency laser paint removal LIBS platform, the module division principle of high cohesion and low coupling is adopted. Based on the LabVIEW visualization system, the high-frequency laser paint removal LIBS system control software is written in G language. The system software architecture is shown in Figure 5.

This software ensures stable paint layer plasma excitation and spectrum acquisition by regulating the information interaction of the three modules of data processing, timing control and closed-loop control. The control process is shown in Figure 6. Start the LIBS monitoring platform. The system first creates communication between the program and the module. After confirming that the connection is correct, it reads the device information and performs initialization settings. Then the platform completes self-checking and loads process parameters; if the self-checking is abnormal, the system will re-initialize and reconnect with the device until the error disappears. The process parameters can be directly called from the process library; if there is no cleaning parameter of the target material in the process library, it can be manually input through the user editing page. Then click the “Start Acquisition” button to synchronize the laser cleaning module and the spectrum acquisition module through the timing control module to complete the spectrum acquisition of the sample to be tested. Finally, the model described in Section 3.2 is implanted into the data processing module to complete the analysis and interpretation of the spectrum. If the spectrum information is abnormal during the acquisition process (the data analysis module determines that the substrate layer is damaged), the closed-loop control module will immediately shut down the cleaning laser and mark the process parameters to avoid laser damage to the substrate material.

2 Integration and application of high-frequency laser paint removal monitoring platform

2.1 Platform integration
Taking into account the needs of high-frequency laser controllable cleaning monitoring, the relative positions of the laser cleaning module, spectrum acquisition module, control and display module are reasonably set, and the cleaning processing head of the laser cleaning module is integrated at the end of the ＫＵＫA robot arm, which effectively improves the scanning accuracy of the paint removal area. The spectrum acquisition dust-free box is installed on the movable corner bracket to meet the requirements of efficient spectrum acquisition. At the same time, the ODROID XU4 control unit is integrated into the control cabinet, and a high-frequency laser controller is set between the control cabinet and the laser. The power supply of each module is separately equipped with a protection circuit to ensure the smooth operation of the platform. The high-frequency laser paint removal LIBS monitoring platform is shown in Figure 7.

2.2 Sample preparation
The experiment selected high-strength aluminum-magnesium-copper hard aluminum alloy 2024-T3 as the base material with a thickness of 2 mm. According to the manual requirements, the surface was treated and 30 μm white CA8000 polyurethane topcoat and about 30 μm green CA7700 epoxy primer (topcoat: Tc, primer: Pr, 2024-T3 base: As) were evenly sprayed in sequence. The prepared large-size paint plate was cut into 50 mm × 50 mm rectangular samples using an EXAKT cutting machine. The schematic diagram of laser paint removal is shown in Figure 8.

Combined with the test results of inductively coupled plasma emission spectroscopy (ICP-OES), Al, Ti, Ba, and Cr were selected as spectral tracer elements, and their contents are shown in Table 1.

2.3 Platform parameter setting
Run the high-frequency laser LIBS paint removal monitoring platform, edit the paint removal process parameters and spectrometer acquisition parameters. Based on previous experiments, the laser frequency was selected as 100 kHz; the scanning speed was 2800 mm·s’-1; the integration time was 10 ms; the initial laser power was 25 W, and the pulse energy density was adjusted by the gradient method (adjustment step 0.75 J·cm’-2, adjustment range 12.75~20.25 J·cm’-2, a total of 10 groups) to ensure the efficiency of laser paint removal while obtaining stable spectrum acquisition. In order to reduce the influence of normal temperature and pressure environment on spectrum acquisition, the indoor temperature was maintained at 20 ℃ and the humidity was 45%.

3 Results and discussion

3.1 Platform spectrum acquisition
3.1.1 Spectral acquisition and selection principles
The above 10 sets of process parameters were used for the three types of materials in the paint layer/substrate system, and a total of 30 sets of experiments were conducted. For each rectangular sample, 10 spectra were intermittently collected in the 10mm×10mm laser cleaning area. A total of 300 (3×10×10=300) spectra were collected for the three types of materials. The traditional characteristic spectral line selection principle: (1) The spectral line intensity is stable and there is no self-absorption phenomenon; (2) The characteristic spectral lines of each element in the spectrum are obviously different; (3) High signal-to-noise ratio. In practical applications, laser process parameters vary with the controllable cleaning requirements of aircraft paint layers. The systematic errors, random errors and accidental errors generated by them lead to the intensified fluctuations of characteristic spectral lines. Following the existing spectral line selection conditions, it is impossible to meet the needs of laser paint removal LIBS controllable cleaning monitoring. The study shows that the spectral similarity of different tissues increases the difficulty of spectral tracer peak selection and classification identification. By selecting multiple spectral lines, the detection misjudgment can be effectively reduced. In summary, referring to the atomic spectrum data of NIST in the United States, the peaks of 12 characteristic spectral lines in the spectrum are extracted for spectral analysis. The elements and wavelengths corresponding to the selected characteristic spectral lines are shown in Table 2.

3.1.2 Spectral preprocessing
The optimized spectral acquisition path suppresses the continuous spectrum intensity to a certain extent, which can effectively improve the quality of spectral acquisition. However, due to the unstable laser energy of the laser, the external environment and the uneven spraying of the paint layer, the collected spectral data still have measurement errors and irrelevant information. Therefore, in order to eliminate and weaken the interference of non-target information variables on spectral analysis, combined with the processing methods of the literature, the original spectral data are preprocessed by using algorithms such as ruler-sampling method and smoothing filtering. Taking the topcoat spectrum as an example, the peak intensity and peak position of the treated topcoat spectrum are prominent, which is conducive to the extraction and analysis of the paint layer element information, as shown in Figure 9 (a) and (b). The fluctuation of process parameters leads to the change of the ionization degree of the paint layer plasma, and the change of the continuous transition range of electrons makes the spectrum difference significant. In order to further improve the accuracy of model classification and recognition, the full spectrum intensity in the data set is normalized, and the spectral line characteristics after normalization are selected as the input of the model, as shown in Figure 9 (c).

3.2 Platform identification model establishment and verification
3.2.1 Spectral PCA cluster analysis
Principal component analysis (PCA) is a multivariate analysis technology that reveals the internal laws of LIBS spectral data. By converting multiple unrelated variables into comprehensive variables (principal components) that cover more initial variable information, the original data set is reduced in dimension to achieve the attribution and classification of unknown samples.
Each spectrum collected during the laser paint removal process is a high-dimensional data set formed by the integral accumulation of plasma emission light, which contains redundant information and noise information. If the data is directly used as the input of linear discriminant analysis (LDA) to establish a paint layer classification and recognition model, it will lead to a slow algorithm optimization rate and reduced classification accuracy. In order to avoid overfitting in the material identification process of different functional paint layers during the paint removal process, the data needs to be processed by PCA dimensionality reduction. The study selected 12 characteristic spectral line peaks, including AlⅠ396.15nm, AlⅡ622.79nm, BaⅡ485.01nm, 614.14nm, CrⅠ463.76, 520.69, 553.56nm, TiⅠ429.90, 453.21, 517.86, 545.94nm, and TiⅢ494.53nm, as the input variables of PCA, and transformed them into new variables with low correlation through orthogonal transformation. The results of KMO and Bart-lett test showed that the 12 input variables supported principal component analysis (KMO=0.877>0.7; Bartlett: p=0.00001<0.001). The cumulative contribution rates of the first six principal components are shown in Table 3.

It can be seen from Table 3 that when the number of principal components exceeds 3, the contribution rate tends to grow slowly. Combined with the results of the principal component weights, the variance explanation rates of the first three principal components PC1, PC2 and PC3 are 84.063%, 10.517% and 3.931% respectively, and the cumulative variance percentage is 98.511%, that is, the three principal components can basically characterize the LIBS spectral feature information of different structural layers of the paint layer/matrix system. Therefore, the first three principal components are proposed to be used as the input of LDA. The principal component weight analysis is shown in Table 4.

After the three types of spectra in the paint removal process are preprocessed and PCA linear dimension reduction, the distribution of the principal component vector scores in three-dimensional space is shown in Figure 10. As can be seen from the figure, the LIBS spectral data of the same sample are regionally aggregated, and effective regional division can be performed. However, there is a partial overlap in the data, indicating that the spectra of the three types of materials have certain similarities, and are similar in element content and composition, which is consistent with the previous ICP-OES test results of different functional paint materials. Therefore, based on the framework of the principal component analysis algorithm, this paper combines LDA to classify and identify the reduced-dimensional data.

3.2.2 Establishment and verification of the identification model
LDA is a supervised classification algorithm that removes redundant data in a projection manner. The characteristic spectrum of the paint layer in the 360-700nm band is used as the input of the LDA to train the classifier after dimensionality reduction. According to Section 3.1, 12 characteristic spectra of 4 elements were selected to construct a spectral data matrix (300×12=3600), corresponding to the output of 3 types of materials of the paint layer/matrix system. In order to obtain the optimal model, the spectral data were randomly divided into a training set and a test set in advance. The training set consisted of 180 spectra, and the remaining 120 spectra were used as the test set. The 120 spectra of the test set were used as the input of the LDA model fitted by the training set, and the model classification evaluation was performed. The classification accuracy of the training set and the test set under different numbers of principal components is shown in Figure 11. Among them, the blue triangle represents the classification accuracy of the training set, and the red dot represents the classification accuracy of the test set. As can be seen from the figure, if the first three principal components are used to construct the linear discriminant analysis model, the classification accuracy of the training set and the test set are 79.4% and 84.2% respectively; however, as the number of principal components increases, the classification accuracy of the training set and the test set still shows an upward trend. Therefore, even if the first three principal components only lose 1.489% of the original data information, the optimal number of principal components for the paint layer/substrate system PCA-LDA recognition model is 3, which is not appropriate. The reason for the above results may be that the elemental composition of the three types of materials is too similar, and in the high-frequency laser continuous paint removal process, the plasma of the topcoat layer, primer layer, and aluminum alloy substrate is mixed together, which reduces the model’s recognition ability of the functional paint layer. When the number of principal components exceeds 5, the accuracy of the training set and the test set fluctuates alternately, and the model classification accuracy fluctuates between 91.7% and 92.5%, which remains stable overall. When the number of principal components is 9, the model classification accuracy is the highest (92.5%), and the principal component covers 99.983% of the original data information. Therefore, it is inappropriate to use only the cumulative variance explanation rate greater than 85% as the principle of principal component selection in the PCA-LDA recognition model of the high-frequency laser paint removal process. Based on this, this monitoring model selects the first 9 principal components as the input of LDA to construct a discriminant model.

The heat map of the PCA-LDA confusion matrix of the test set is shown in Figure 12. It can be seen that there are a few misjudgments of the spectra of the primer layer and the substrate layer. Five primer layer spectra are misjudged as substrate spectra; three substrate spectra are misjudged as primer spectra, and one is misjudged as topcoat spectrum. The reason for this phenomenon may be that during the coating spraying process, the coating particles diffuse from the substrate pores to the substrate, and the materials at the interface between the primer and the substrate interpenetrate each other; and the increase in the depth of paint removal increases the complexity of the plasma of the three types of materials. Although the above-mentioned interference factors lead to model misjudgment, the classification accuracy of the model for 120 spectra in the test set is still as high as 92.5%, among which the accuracy of topcoat layer recognition reaches 100%. The results show that the high-frequency laser paint removal LIBS platform can complete the material recognition of different functional paint layers and realize the effective monitoring of high-frequency laser paint removal.

4 Conclusions
Based on the laser induced breakdown spectroscopy technology, a LIBS monitoring platform suitable for high-frequency laser paint removal is designed to meet the needs of controlled skin cleaning and the spectrum excitation and acquisition problems of the high-frequency laser paint removal monitoring system. The platform mainly includes laser cleaning module, spectrum acquisition module, control and display module. Among them, the control software is written based on the LabVIEW embedded development system to realize the performance regulation of the platform and the analysis and interpretation of spectral data. Taking the 2024-T3 aluminum alloy double-layer paint sample as the research object, a classification and discrimination model of the paint layer/substrate system in the laser cleaning process is established. The number of principal components of the model input was verified and optimized through experiments to improve the classification accuracy of the LIBS platform. The first 9 principal components were selected as the input of LDA. At this time, the classification effect of the PCA-LDA model based on the LIBS spectrum was optimal, and the classification accuracy reached 92.5%. The designed high-frequency laser paint removal LIBS platform can complete the material identification of different structural layers of the paint layer/substrate system by importing the discrimination model, and can realize the effective monitoring of aircraft skin laser paint removal based on LIBS technology.