Objective  Knives are common physical evidence at crime scenes, and identifying their brands can provide important investigative leads. Laser-induced breakdown spectroscopy (LIBS) is widely used in material analysis due to its broad elemental detection range, high sensitivity, and spatial resolution. Deep learning has recently shown strong performance in spectral recognition; however, its effectiveness depends on sufficiently large training datasets. And data augmentation is one of the ways to solve this question. A shift-centered and correlation-aligned variational autoencoder (SCCA-VAE) was developed. The model integrates diversity-driven sampling and correlation-based constraints to generate high-quality synthetic LIBS spectra after dimensionality reduction using principal component analysis (PCA), and forms an efficient and accurate identification system when combined with LightGBM.

Methods  LIBS spectra were collected from 138 knife blades, with 22 spectra obtained for each sample (20 used for training and 2 for testing). Because raw spectra contain substantial noise, PCA was applied to extract the principal features for subsequent data synthesis and classification.The proposed SCCA-VAE improves the conventional VAE by optimizing the sampling strategy and loss function. To enhance diversity, synthetic samples are generated around the class-specific latent center rather than from a standard normal distribution. A similarity module is introduced to maintain consistency with real data: a correlation-matrix matching loss is incorporated as a regularization term in the variational lower bound, and a priori feature-mean alignment is adopted to reduce deviations caused by diversity-oriented sampling. The SCCA-VAE data augmentation flowchart is shown (Fig.1). Both comparative and ablation experiments were performed. In the comparison, SCCA-VAE was tested against Gaussian noise and linear random interpolation, with data similarity evaluated by the Wasserstein distance, KS statistic, and correlation-matrix MAE, and diversity measured by the diversity ratio. Classification performance and generalization were assessed using 1D-CNN and LightGBM trained on 4× and 10× augmented datasets. Ablation studies were conducted by separately removing the similarity and diversity modules to analyze their functions. Synthetic data quality and classification results on all augmented datasets were used for quantitative comparison.

Results and Discussion Based on quantitative analysis using evaluation metrics (Tab.3), all three data augmentation methods maintain reasonable similarity between generated data and original samples. However, both Gaussian noise and linear interpolation exhibit diversity ratios below 1, indicating that while they preserve similar features, they fail to effectively expand the variability distribution of the data. In contrast, the SCCA-VAE model successfully establishes a balancing mechanism between similarity and diversity, generating spectral data that combines physical realism with information richness.Results from comparative experiments (Tab.4) show that 4-fold augmentation improves the performance of 1D-CNN for all methods. At 10-fold augmentation, Gaussian noise and linear interpolation suffer from overfitting, while the classification accuracy of SCCA-VAE continues to rise. For the LightGBM classifier, only SCCA-VAE enhances performance, as its synthetic data preserves the discriminative structure of original samples and matches the learning characteristics of LightGBM. This combination achieves the highest accuracy among all augmentation-classifier pairs. The impact of SCCA-VAE data augmentation multiples on LightGBM was further explored, with results indicating that a 4x augmentation achieves optimal performance.Ablation studies verify the effectiveness of each module in SCCA-VAE. Removing the similarity module increases diversity but reduces similarity (Tab.5), and classification performance (Tab.6), though improved, is still inferior to the full model, meaning diversity must be constrained to avoid deviation from real data. Removing the diversity module severely degrades data quality (Tab.5) and classification performance (Tab.6), even below that of the original dataset, demonstrating that diversity is key to performance improvement, while the similarity module ensures reliable and controllable synthesis. The two modules complement each other: the diversity module strengthens data variability and model learning ability, and the similarity module guarantees that generated data faithfully represents the real spectra after dimensionality reduction.

Conclusions  To address the scarcity of training samples in laser-induced breakdown spectroscopy (LIBS) technology, this study developed the SCCA-VAE model. Based on the variational autoencoder (VAE) architecture, this model significantly enhances data augmentation capabilities by integrating diversity-driven sampling strategies with correlation-based similarity constraints. Test results across four key evaluation metrics demonstrate that SCCA-VAE effectively expands the diversity distribution of generated data while preserving its intrinsic similarity. Comparative experiments further validate the model's outstanding generalization performance and robustness. Particularly when applied in conjunction with the LightGBM algorithm, SCCA-VAE achieves higher recognition accuracy at lower data augmentation rates, fully validating the dual advantages of precision and efficiency in the recognition system built by combining the two. In summary, SCCA-VAE not only outperforms conventional data augmentation methods but is also more suitable for the expansion task of high-dimensional spectral datasets; it significantly expands the model's generalization boundaries while forming an efficient, precise recognition closed-loop through deep coupling with LightGBM. This research ultimately establishes a robust data augmentation and recognition framework tailored for LIBS tool attribution, providing strong technical support for precise identification in forensic evidence analysis.