During the operation of an oil-immersed transformer, aging of the insulating cardboard, mechanical failure of the submersible pump, and action arc of the tap change introduce different particulate impurities into the transformer oil. If the particulate impurities suspend in the oil flow or adhere to the surface of the transformer windings and components, the safety of the transformer will be compromised. In recent years, laser-induced breakdown spectroscopy (LIBS) has been widely used for the detection of nonmetals and metals. There are also reports on the detection of particulate matter in transformer oil, and the filter-paper-assisted LIBS map-scanning method has proven to be effective for the quantitative analysis of particles. In this study, in addition to the target spectral lines, molecular bands and bremsstrahlung and recombination radiation in the plasma generated by laser ablation are present, and the resulting continuous background spectrum cannot be shielded during spectral analysis, which adversely affects the spectral intensity of the characteristic spectral lines of the target element. However, the ablation spots obtained by laser scanning the deposition area of the sample do not contain characteristic spectral lines of the target element, hence necessitating binary classifications to screen out the spectrally effective spots for quantitative analysis. Therefore, there is an urgent need for a set of data-processing algorithms for baseline correction, dimensionality reduction, and classification of the original data to meet the processing requirements of a large number of high-dimensional spectral data.
Based on the sparsity of characteristic spectral lines, this study investigates the application of the baseline estimation and denoising algorithm (BEADS) in LIBS spectral baseline correction and subtraction. The results of the dimensionality reduction using the target analysis line and principal component analysis (PCA) are compared. In addition, the binary classification effects of different machine learning algorithms (e.g., decision tree, support vector machine, K-nearest neighbor classification, and ensemble classifier) on laser ablation spots are studied, in which the ensemble classifier perfects the classifier model through different optimization methods. Finally, based on the above spectrum-processing algorithm, a quantitative analysis and calibration of Fe particle detection is completed.
Spectral analysis reveals that the characteristic spectral lines such as Fe Ⅰ 360.89 nm, Fe Ⅰ 361.88 nm, Fe Ⅰ 363.15 nm, Fe Ⅰ 364.78 nm, and Fe Ⅰ 371.99 nm have excellent sparsity, whereas Fe Ⅰ 373.49 nm, Fe Ⅰ 373.71 nm, Fe Ⅰ 374.83 nm, Fe Ⅰ 374.95 nm, and Fe Ⅰ 376.38 nm exhibit relatively poor sparsity owing to interference with each other or overlap with the background spectrum. The experimental results show that, as shown in Fig. 3, the application of asymmetric penalty functions and convex optimization techniques is beneficial for reducing overfitting. When the parameters are adjusted to fc=0.15, d=1, r=6, and λ=0.8, the accuracy of the baseline estimation is very high, and the residual value obtained by the baseline fitting is small. For the detection of Fe particles in transformer oil, the BEADS algorithm can excellently deduct the continuous background, so that the intensity of the Fe characteristic spectral lines can be accurately corrected. As shown in Fig. 5, in terms of the classification accuracy of ablation spots by the classifier, the spectral data after dimensionality reduction using the target analysis line is better than the original spectral data and the spectral data after PCA is performed. This demonstrates that the dimensionality reduction processing method using the target analysis line is scientific and reasonable. Based on the decision tree algorithm, the Bagged Trees ensemble classifier constructs multiple decision tree models through the extraction of different samples, thereby reducing the variance, optimizing the classification model, and improving classification accuracy. The spectral classification accuracy for data after dimensionality reduction is as high as 98.33%.
Based on the spectral processing method, the linear correlation coefficient between the spectral correction intensity and particle mass ratio is 0.9983, and the relative standard deviation of repeated experiments is small, which proves the scientificity and robustness of the method. The method can realize batch processing of a large number of ablation spot data generated by laser map scanning, which greatly improves data-processing efficiency while reducing errors introduced by manual processing. It provides convenience for LIBS automatic acquisition and data processing and lays a theoretical foundation for LIBS detection of particulate matter in transformer oil.
