Knives is one of the most common types of physical evidence found at crime scenes, and establishing an accurate method for classifying them is of paramount importance. In this experiment, a laser-induced breakdown spectroscopy (LIBS) analyzer is used to detect 138 types of knife samples, resulting in 2760 training data sets and 276 prediction data sets. After preprocessing, principal component analysis is employed to reduce the dimensionality of both the training and prediction datasets, retaining the top 400 principal components with a cumulative variance contribution rate of 81%. The dimensionality-reduced data are then fed into one-dimensional convolutional neural network (1D-CNN) model for training and prediction. After 200 iterations, the average prediction accuracy reaches 96.02%. To further improve the prediction accuracy, additional data augmentation is performed based on the spectral features of the data using methods involving normal distribution and linear combination to generate additional 40, 80, and 120 data sets for each sample. Comparative experiments are conducted to investigate the impact of different data augmentation methods and quantities on the performance of the 1D-CNN model. Finally, using the normal distribution method, the model's predictive accuracy gradually increases with the increase in synthetic data volume, reaching a maximum of 97.46%. The study demonstrates that a combination of LIBS spectroscopy, 1D-CNN models, and data augmentation using the normal distribution method can effectively achieve precise identification of knife samples, thereby providing valuable clues for investigation and case resolution.