View PDF Version
 
DOI: 10.1039/D5AY00906E (Paper) Anal. Methods, 2025, Advance Article

Rapid detection method for the hardness of louver contacts based on laser-induced breakdown spectroscopy (LIBS)

Yun Xu, Zefeng Yang*, Ziyi Li, Langyu Xia, Kai Liu and Wenfu Wei
School of Electrical Engineering, Southwest Jiaotong University, Chengdu 611756, China. E-mail: yangzefeng@foxmail.com

Received 29th May 2025 , Accepted 23rd July 2025

First published on 18th August 2025

Abstract

The louver contact, a critical component in ultra-high-voltage (UHV) converter transformers, significantly impacts the operational safety of electrical equipment due to its performance stability. The degradation mechanism arises from prolonged exposure to high current, elevated temperatures, and mechanical stress, leading to reduced material hardness and subsequent deterioration of contact performance. In this study, Laser-Induced Breakdown Spectroscopy (LIBS) was employed to characterize aged louver contact specimens. An outlier elimination strategy based on the cosine-distance Local Outlier Factor (LOF) algorithm was implemented. A hybrid dimensionality reduction framework integrating Principal Component Analysis (PCA) with neural encoder architectures was applied to address the dual challenges of high-dimensional data processing and preservation of critical spectral features. This research establishes a LIBS-based surface hardness detection method for louver contacts. Through comparative evaluation of multiple models, the Adam-optimized Gradient Boosting Decision Tree (Adam-GBDT) demonstrated superior performance, achieving a coefficient of determination (R2) of 0.977, and exhibited significant potential for micro-damage surface evaluation. These findings provide technical support for operational monitoring and maintenance strategies for UHV equipment.

1 Introduction

The louver contact, functioning as a crucial electrical connection component between the converter transformer and the converter valve, plays an irreplaceable and significant role in power systems.1,2 The louver contact is typically secured within the groove of the lead plug on the valve-side winding, serving as a core connection component. It operates under harsh conditions such as high current, elevated temperatures, mechanical compression, and wear for extended periods. This continually complex working environment leads to various degrees of deformation, collapse, and wear on the contact body itself, all of which are closely related to the hardness characteristics of the louver contact. It has been shown that there is a significant correlation between the hardness of a material and its mechanical property parameters such as elasticity,3,4 fatigue resistance5–7 and creep resistance.8–10 Increased wear and louver contact collapse caused by changes in hardness will inevitably have a significant impact on the electrical contact performance of the louver contact. Therefore, the hardness parameter of the louver contact is one of the key factors determining the contact performance of an electrical connection.

Traditional hardness testing methods (e.g. Brinell, Rockwell and Vickers hardness tests) are mainly based on the principle of hardness measurement by indentation of the specimen surface by an indenter under static load.11–13 However, these methods have the limitations of being destructive, slightly complicated to operate, needing a long testing time and having high requirements for sample surface cleanliness. Therefore, there is a need to find an efficient, micro-damage hardness testing technique for the louver contact.

The LIBS technique, a rapid atomic spectrometry technique, provides in situ, minimally destructive multi-element analysis capabilities. This technique has demonstrated extensive applications across environmental monitoring,14,15 biomedical diagnostics,16,17 agricultural science,18,19 and industrial quality control.20,21 The LIBS technique works by focusing a high-energy pulsed laser on the surface of a sample, causing ablation of materials on the order of micrograms and the formation of a high-temperature plasma (temperatures up to 10[thin space (1/6-em)]000–20[thin space (1/6-em)]000 K).22 During the plasma cooling process, atoms and ions in the excited state emit spectra with characteristic wavelengths, and this spectral information can be used to analyse the elemental composition of the sample. Many studies have now explored the principle that there is a correlation between LIBS and material hardness.23,27 It was found that there is a positive correlation between hardness and plasma temperature and a negative correlation with ablation mass,23,24 which can be interpreted to mean that at the same laser energy, materials with high hardness have less mass to be ablated and produce higher plasma temperatures.25–27 Further, Wang28 et al. investigated the hardness of aged insulators by employing LIBS and found that there was a linear relationship between the hardness and the ratio of ions to the atomic spectral lines and the plasma temperature. Sattar29 et al. applied the LIBS technique to monitor the surface hardness of tungsten alloys, and found that the measurement based on the linear relationship between plasma electron temperature and hardness resulted in good reproducibility of the results. Xia30 et al. conducted a comparison and analysis of three different methods, namely spectral line intensity ratio, plasma excitation temperature, and machine learning. They utilized machine learning techniques to establish and optimize a multivariate model for hardness analysis of U71Mn railway tracks. However, current research is more focused on verifying the correlation between LIBS and material hardness, lacking studies on data anomalies caused by interferences such as oxides and deposited particles during the operation of electrical contact elements in UHV systems and the insufficiency of traditional dimensionality reduction methods.

This study takes the louver contact as the research object and applies LIBS technology to explore the correlation between hardness and spectral signals. The aim is to develop a rapid, micro-damage, and high-precision method for hardness detection of the louver contact, providing important support for enhancing the operational reliability of critical components in UHV transmission systems.

2 Samples and methods

2.1 Sample preparation

In this study, the LA-CUT type louver contact was selected as the experimental subject. Fig. 1(a) presents a schematic diagram of the louver contact structure in the bushing of an UHV converter transformer. The louver contact, primarily made of copper, is installed in the dedicated grooves of the plugs at both ends through mechanical fixation, forming a reliable connection with the double-ended socket, as shown in the lower right corner of Fig. 1(a). With material degradation, the hardness of the louver contact will decrease, potentially leading to deformation and loss of reliable electrical contact performance in severe cases.
image file: d5ay00906e-f1.tif
Fig. 1 (a) Schematic diagram of the louver contact structure in the bushing of an UHV converter transformer. (b) Testing equipment for measuring the pressure and compression of louver contacts, obtaining stress–strain relationship curves for samples with different hardness, revealing the impact of hardness variations on contact force magnitude.

In high-temperature compression experiments simulating the actual operating conditions of the louver contact, three sets of experiments were conducted at temperatures of 200 °C, 250 °C, 300 °C, and 350 °C. With consistent pressure and duration conditions, four groups of louver contact samples with hardness values of 90.10 HV, 80.57 HV, 71.46 HV, and 65.68 HV (designated as #1, #2, #3, and #4, respectively) were obtained. Subsequently, using the device shown in Fig. 1(b) for testing pressure and compression, multiple tests were conducted on these four samples to plot pressure–compression relationship curves. It was observed that as the hardness decreased, there was a significant change in the slope of the curves. By fitting the data, it was found that the slope of the curves varied by 30.4–60.2%, which means that the contact force decreased by the same amount. When the compression of the contact plate changed by the same amount, the contact force decreased significantly, which could impact the electrical contact performance of louver contact pieces.

2.2 Experimental data acquisition platform

The LIBS platform used in the experiment is shown in Fig. 2. The platform employs an Nd[thin space (1/6-em)]:[thin space (1/6-em)]YAG laser as the laser source, generating a laser wavelength of 1064 nm with a pulse width of 6–8 ns. In order to balance the signal quality and damage to the sample, the energy of a single laser pulse is 200 mJ. The laser beam emitted by the laser is focused onto the sample surface through a 45-degree placed mirror and a horizontally placed convex lens, forming a high-power density ablation area (micron-scale spot). Subsequently, another convex lens placed at a 45-degree angle collects the high-temperature plasma radiation spectrum excited by the laser ablation and directs it to the optical fiber probe. The collected light is coupled through an optical fiber to a mid-range grating spectrograph (Andor ME5000), with a wavelength range of 200–1000 nm, capable of covering the elemental emission lines from ultraviolet to near-infrared. The platform uses a digital delay generator (DG535) to synchronously control the laser and the intensified charge-coupled device (ICCD). In order to avoid the influence of early high-intensity continuous background radiation interference and signal attenuation caused by plasma cooling, the delay time of ICCD (ANDOR istar DH344) is set to 1 μs, and the acquisition time window is set to 3 μs.
image file: d5ay00906e-f2.tif
Fig. 2 LIBS experimental platform for acquiring spectral data.

Through the established LIBS experimental platform, 100 sets of spectral data were collected from each of four samples with different hardness. The same position was continuously stimulated six times, and the data of the last three times were taken for analysis. From the spectral comparison shown in Fig. 3, it can be observed that there are significant differences in the intensity of spectral line signals at wavelengths such as Cu II 521.82 nm and Hα 656.28 nm among the four hardness samples.


image file: d5ay00906e-f3.tif
Fig. 3 Spectral lines of samples of different hardness; the spectral line signal intensities of the four hardness samples at Cu II 521.82 nm, Hα 656.28 nm and N I 746.83 nm are significantly different. It is possible to distinguish samples of different hardness by spectral lines.

3 Data processing and analysis

3.1 Abnormal sample rejection

Addressing the issue of outliers in spectral data due to sampling interference, this study employed the Local Outlier Factor (LOF) algorithm for anomaly detection and removal. Four distance metrics, namely Manhattan distance, cosine distance, Chebyshev distance, and Euclidean distance, were compared in detail. Through quantitative analysis of computational efficiency, intra-class distance, and inter-class distance (Fig. 4), it was found that cosine distance offers dual advantages: it has the shortest computation time, and the dataset formed after removing outliers exhibits the smallest intra-class distance and the largest inter-class distance, significantly enhancing inter-class separability. This may be attributed to the fact that the unique vector angle measurement mechanism of cosine distance aligns well with the high-dimensional characteristics of LIBS spectral data. By capturing the structural similarities and differences in high-dimensional space, cosine distance can more accurately identify the distribution differences between normal and abnormal samples. Experimental results demonstrate that the LOF algorithm based on cosine distance is the optimal choice for anomaly detection tasks in spectral data.
image file: d5ay00906e-f4.tif
Fig. 4 A comparison of the effect of the four different methods of calculating the distance matrix.

Statistical visualization analysis of the LOF scores for all samples (Fig. 5) reveals the following results: The histogram of LOF scores exhibits a strongly clustered distribution centered around an LOF value of 1, which aligns with the density distribution characteristics of normal samples. After setting a 5% anomaly threshold, due to the high-dimensional nature of the data and the unique vector angle measurement mechanism of cosine distance, the two-dimensional scatter plot distribution of high-LOF-value anomaly samples (marked in red) in the Principal Component Analysis (PCA) reduced-dimensional space generally displays outlier characteristics. Collaborative verification of statistical distribution features and spatial topological features demonstrates that the anomaly detection method based on the LOF algorithm with cosine distance effectively achieves precise identification and removal of outlier samples in spectral data.


image file: d5ay00906e-f5.tif
Fig. 5 (a) Histogram showing LOF score statistics for all samples; the closer the LOF score is to 1, the more the data point is in the normal distribution and the less it looks like an outlier. (b) Scatter plot showing samples excluding outliers, set a 5% threshold for abnormal samples, marked in red as abnormal samples.

3.2 Noise reduction and normalization

Noise reduction in spectral data is crucial for effectively eliminating the interference of noise on the signals, thereby improving signal quality and enhancing feature correlation. This study employs a wavelet denoising method based on multi-indicator evaluation and optimization, aiming to address the balance between feature preservation and noise suppression in spectral signals. During the denoising process, a comprehensive scoring mechanism using multiple evaluation indicators is adopted, including Smoothness Ratio (SMR), Signal-to-Noise Ratio (SNR), Pearson Correlation Coefficient (PCC), and Mean Squared Error (MSE). A weighted scoring mechanism is utilized to dynamically optimize the combination of wavelet basis functions and decomposition levels. Among these indicators, SNR is used to quantify the degree of noise reduction, MSE evaluates the overall fidelity of the signal, PCC measures the similarity of waveform structure, and SMR characterizes the improvement in local smoothness. The denoising effect on the spectral lines is shown in Fig. 6. The original signal contains a large amount of high-frequency noise. From the magnified view, the denoised signal is smoother, with a smoothness that is 2.2479 times that of the original signal. The SNR of the denoised signal is 21.6986 dB, and the PCC reaches 0.9961, close to 1. Meanwhile, the MSE is 180.0914, which is relatively low compared to the signal amplitude (on the order of thousands). These indicators indicate that the denoised spectral lines maintain good integrity in their overall trend, and the amplitude of peaks and detailed features are not significantly lost.
image file: d5ay00906e-f6.tif
Fig. 6 Demonstration of the before and after effects of noise reduction on spectral lines. The red chromatographic lines show obvious fluctuation and clutter, which represents the original spectral lines with noise interference. The blue chromatographic line is relatively smooth and stable, and it is the spectrum line after denoising.

To eliminate the significant dimensional differences among features and enable the model to learn the importance of each feature evenly, normalization was performed. The min–max normalization method effectively eliminates the dimensional differences among features while accommodating the characteristics of high-dimensional sparse data. This ensures that each feature has an equal impact on the model, providing a solid foundation for subsequent model training.

3.3 Reduced dimensionality

In this study, a hybrid dimensionality reduction strategy combining Principal Component Analysis (PCA) with deep encoders is adopted. In a LIBS spectral data matrix, each row represents an individual spectral signal, with all signal wavelengths comprising the feature columns, totaling 25[thin space (1/6-em)]777 dimensions. Firstly, PCA linear dimensionality reduction is performed on the standardized and preprocessed data, and it is determined that retaining 173 principal components achieves a cumulative variance contribution rate of 95% (with the first 10 principal components contributing 85.4%). To address the limitations of linear dimensionality reduction in expressing complex features, an encoder network with nonlinear mapping capabilities is added. The 173-dimensional PCA output vector is compressed into an 18-dimensional latent space through a multilayer perceptron. The increase in potential dimensions will significantly affect the computational efficiency of the model. To balance the model's efficiency and performance, 18 dimensions are selected. The specific architecture and data flow are shown in Fig. 7(a). Specifically, the data is first reduced in PCA, and subsequently mapped into a lower-dimensional space through an encoder for further representation. From the scatter plot of PCA output in Fig. 7(b), in the three-principal-component space, samples of the same hardness exhibit obvious clustering effects, while samples of different hardness are hierarchically distributed.
image file: d5ay00906e-f7.tif
Fig. 7 (a) Specific architectures and data flow for PCA and hybrid dimensionality reduction strategies, (b) scatter plot of samples in the space of the first three principal components from PCA, (c) comparison of random forest model results for PCA and hybrid dimensionality reduction strategies.

Fig. 7(c) demonstrates the impact of data processed by two dimensionality reduction methods, PCA and hybrid dimensionality reduction, on model performance. The experiment utilizes a random forest model consisting of 50 decision trees and evaluates model performance through Mean Squared Error (MSE) and the coefficient of determination (R2). The results indicate that compared to using PCA alone, the combined strategy of introducing an encoder significantly enhances model performance: MSE decreases from 12.737 to 8.583 (a reduction of 32.6%), and R2 increases from 0.874 to 0.915 (an increase of 4.7%). From the perspective of model stability, the prediction results of individual decision trees using the hybrid dimensionality reduction strategy exhibit a more concentrated scatter plot distribution, indicating that this strategy effectively reduces the discreteness of model outputs. Combined with the analysis of the fitting effect, the linear principal components extracted by PCA in the hybrid dimension reduction strategy complement the non-local features captured by the autoencoder, enhancing the coverage ability of the features on the data distribution.

3.4 Model prediction

To evaluate the representational performance of the low-dimensional features, a comparative experimental framework is established using Convolutional Neural Networks (CNN), Gradient Boosting Decision Trees (GBDT) in an ensemble form, and Support Vector Regression (SVR). Experimental data indicate the following: during the training process of the CNN model (depicted in Fig. 8(a)), its Mean Squared Error (MSE) exhibits a decreasing trend with increasing iterations, ultimately converging to 8.445. Concurrently, R2 increases to 0.922, validating the CNN's effective extraction capability of low-dimensional features. The response surface for hyperparameter optimization of the SVR model (depicted in Fig. 8(b)) reveals a strong coupling effect between the regularization parameter and the kernel function scale parameter. By searching and locating the global optimal solution within the parameter space, the SVR achieves a prediction accuracy of MSE = 2.628 and R2 = 0.954 with the optimal parameter combination. Experimental results demonstrate that Gradient Boosting Decision Trees (GBDT) exhibit significant predictive advantages in regression tasks. Fig. 8(c) illustrates the variation in GBDT model performance with respect to the number of decision trees. By examining the zoomed-in plot of the last 10 trees, it is evident that performance stabilizes when the number of decision trees reaches 17. The final MSE is 1.230, and R2 reaches 0.977. The GBDT model not only possesses efficient training speed but also achieves excellent predictive performance with a relatively small number of trees. Therefore, GBDT is the optimal choice for the current task.
image file: d5ay00906e-f8.tif
Fig. 8 (a) Predictive effectiveness of CNN during iterative optimization, R2 = 0.922, MSE = 8.445, (b) different predictive effectiveness of SVR in parameter space, R2 = 0.954, MSE = 2.628, (c) predictive effectiveness of GBDT during increasing the number of decision trees, R2 = 0.977, MSE = 1.230, (d) systematic comparison of the five-dimensional evaluation system (time, MSE, RMSE, MAE and R2) for systematic comparison of the four intelligent optimization algorithms' hyper-parameter optimization search effect.

In this study, four intelligent optimization algorithms—Variable Neighborhood Search (VNS), Brain Storm Optimization (BSO), Ant Colony Optimization (ACO), and Adam—were employed for hyperparameter tuning of GBDT. In Fig. 8(d), through a systematic comparison using a five-dimensional evaluation system (including computation time, MSE, Root Mean Squared Error (RMSE), Mean Absolute Error (MAE), and R2), experimental results show that the Adam algorithm has a significant advantage in computational efficiency with a learning and prediction speed of 31.95 seconds, reducing time costs by 91.6%, 95.0%, and 78.9% compared to VNS, BSO, and ACO, respectively. In terms of prediction accuracy, the Adam-optimized GBDT(Adam-GBDT) model achieved stable performance with MSE = 1.23, RMSE = 1.109, and MAE = 0.069, while Adam's R2 reached 0.977. From the perspectives of time efficiency and prediction accuracy, the Adam algorithm is the optimal optimization method in this study.

4 Conclusion

Through experimental and algorithmic optimization efforts, this study has established a novel rapid detection method for the louver contact status based on LIBS and machine learning. Experimental findings reveal that a 10.6–27.1% decrease in hardness results in a 30.4–60.2% change in contact force, underscoring the importance of hardness measurement in the louver contact and motivating the application of LIBS for this purpose. Addressing the challenges posed by sampling interference and high dimensionality in LIBS spectral feature data, this study employed the LOF algorithm based on cosine distance to efficiently eliminate abnormal samples. Compared to a single PCA dimensionality reduction strategy, the combined dimensionality reduction strategy involving PCA and an encoder reduced the MSE of the model by 32.6% and increased R2 by 4.7%. From the combination models constructed using three machine learning algorithms and four parameter optimization algorithms, the optimal Adam-GBDT model was selected. Its learning and prediction process took only 31.95 seconds, significantly reducing time costs by 78.9–95.0% compared to VNS, BSO, and ACO. Furthermore, the Adam-GBDT model exhibited the lowest prediction errors, with the R2 value reaching 0.977, successfully validating the feasibility of applying LIBS technology for rapid detection of the hardness of the louver contact. This research provides a new method for the online monitoring of electrical equipment quality.

Data availability

The data will be made available by the authors upon request through email.

Author contributions

Yun Xu: writing – original draft, data curation, conceptualization. Zefeng Yang: resource, validation, supervision. Ziyi Li: methodology, software. Langyu Xia: investigation, formal analysis. Kai Liu: project administration, conceptualization. Wenfu Wei: writing – review & editing, resources, funding acquisition.

Conflicts of interest

There are no conflicts to declare.

Acknowledgements

This work was supported by a grant from the National MCF Energy R&D Program (2024YFE03250200).

References

  1. Z. Huang, F. Liu, G. Wu, C. Yuan, H. Tang, Y. Tu, H. Zhou, J. Chen, M. Lin and K. Liu, IEEE Trans. Power Delivery, 2025, 40(1), 100–112 Search PubMed .
  2. S. Hu, Z. Huang, J. Chen, C. Yuan, B. Feng, T. Zhao and Y. Liu, IEEE Access, 2025, 13, 14444–14452 Search PubMed .
  3. A. Nagakubo, H. Ogi, H. Sumiya and M. Hirao, Appl. Phys. Lett., 2014, 105(8), 081906 CrossRef .
  4. X. Liu, Y.-Y. Chang, S. N. Tkachev, C. R. Bina and S. D. Jacobsen, Sci. Rep., 2017, 7(1), 42921 CrossRef PubMed .
  5. C. Zhang, J. Tian, S. Yang and S. Mu, J. Mater. Res. Technol., 2024, 28, 2060–2070 CrossRef .
  6. J.-W. Seo, H.-M. Hur and S.-J. Kwon, Metals, 2022, 12(4), 630 CrossRef .
  7. H. Liu, D.-G. Shang, Z.-K. Guo, Y.-G. Zhao and J.-Z. Liu, Fatigue Fract. Eng. Mater. Struct., 2014, 37(8), 937–944 CrossRef .
  8. Y. F. Li, T. Nagasaka, T. Muroga, Q. Y. Huang and Y. C. Wu, J. Nucl. Mater., 2009, 386–388, 495–498 CrossRef .
  9. M. Négyesi, M. Kraus, V. Mareš, D. Kwon and B. Strnadel, Int. J. Pressure Vessels Piping, 2023, 206, 105085 CrossRef .
  10. Y. S. Lee and J. Yu, Metall. Mater. Trans. A, 1999, 30(9), 2331–2339 CrossRef .
  11. G. Leyi, Z. Wei, Z. Jing and H. Songling, Measurement, 2011, 44(10), 2129–2137 CrossRef .
  12. L. Ma, S. Low and J. Fink, J. Test. Eval., 2006, 34(3), 187–199 CrossRef CAS .
  13. S.-K. Kang, J.-Y. Kim, P.-P. Chan, H.-U. Kim and D. Kwon, J. Mater. Res., 2010, 25(2), 337–343 CrossRef CAS .
  14. R. Junjuri, C. Zhang, I. Barman and M. K. Gundawar, Polym. Test., 2019, 76, 101–108 CrossRef CAS .
  15. H. Wang, X. Yan, Y. Xin, P. Fang, Y. Wang, S. Liu, J. Jia, L. Zhang and X. Wan, Chemosensors, 2023, 11(7), 377 CrossRef CAS .
  16. S. Shi, L. Pi, L. Peng, D. Zhang, H. Ma, Y. Liu, N. Deng, X. Wang and L. Guo, J. Anal. At. Spectrom., 2025, 40(2), 478–486 RSC .
  17. P. Winnand, M. Ooms, M. Heitzer, M. Lammert, F. Hölzle and A. Modabber, Oral Oncol., 2023, 138, 106308 CrossRef CAS PubMed .
  18. K. Liu, P. Fan, G. Hou, X. Li, S. Qi and H. Qiu, Results Chem., 2022, 4, 100477 CrossRef CAS .
  19. V. N. Lednev, P. A. Sdvizhenskii, M. Y. Grishin, S. V. Gudkov, A. S. Dorokhov, A. F. Bunkin and S. M. Pershin, J. Anal. At. Spectrom., 2022, 37(12), 2563–2572 RSC .
  20. L. Zhang, W. Yin, L. Dong, W. Ma, L. Xiao and S. Jia, IEEE Photonics J., 2017, 9(5), 1–10 Search PubMed .
  21. G. S. Senesi, Chemosensors, 2025, 13(2), 41 CrossRef CAS .
  22. B. Tang, Z. Yang, Z. Li, W. Wei, L. Xia, Z. Li, P. Li and G. Wu, Talanta, 2025, 292, 127927 CrossRef CAS PubMed .
  23. T. A. Labutin, A. M. Popov, V. N. Lednev and N. B. Zorov, Spectrochim. Acta, Part B, 2009, 64(10), 938–949 CrossRef .
  24. S. Messaoud Aberkane, A. Bendib, K. Yahiaoui, S. Boudjemai, S. Abdelli-Messaci, T. Kerdja, S. E. Amara and M. A. Harith, Appl. Surf. Sci., 2014, 301, 225–229 CrossRef CAS .
  25. K. Lahna, M. Pardede, K. H. Kurniawan, K. Kagawa and M. O. Tjia, Appl. Opt., 2017, 56(36), 9876–9881 CrossRef .
  26. D. A. Rusak, B. C. Castle, B. W. Smith and J. D. Winefordner, Crit. Rev. Anal. Chem., 1997, 27(4), 257–290 CrossRef .
  27. A. M. Marpaung, R. Hedwig, M. Pardede, T. J. Lie, M. O. Tjia, K. Kagawa and H. Kurniawan, Spectrochim. Acta, Part B, 2000, 55(10), 1591–1599 CrossRef .
  28. X. Wang, X. Hong, P. Chen, C. Zhao, Z. Jia, L. Wang and L. Zou, IEEE Trans. Plasma Sci., 2019, 47(1), 387–394 Search PubMed .
  29. H. Sattar, H. Ran, W. Ding, M. Imran, M. Amir and H. Ding, Appl. Phys. B, 2019, 126(1), 5 Search PubMed .
  30. L. Xia, Z. Yang, W. Wei and G. Wu, Spectrochim. Acta, Part B, 2024, 215, 106908 CrossRef .
This journal is © The Royal Society of Chemistry 2025
Click here to see how this site uses Cookies. View our privacy policy here.