摘要
Advancements in artificial intelligence have facilitated the widespread application of machine learning techniques in industrial, manufacturing, and scientific domains. However, the scarcity of labeled data continues to pose a major challenge for effective analysis and modeling. To address this limitation, we propose an interpolation-based data augmentation method that generates pseudo-labeled data points between sparsely labeled sequences. Time-series spectral data acquired during atmospheric pressure plasma jet thin film deposition are used to interpolate labels for intermediate data points based on feature similarity and temporal order, thereby generating additional training data with estimated labels. We evaluated data augmentation across multiple cycles on three representative regression models, i.e. Lasso, multilayer perceptron (MLP), and extreme gradient boosting, and found that the implementation of data augmentation effectively reduced the mean absolute error (MAE) and mean absolute percentage error (MAPE) in film thickness prediction. In particular, with four augmentation cycles, the MLP achieved the greatest improvement, reducing MAE by 5.5% and MAPE by 16.5% compared with the baseline without augmentation. In addition to improving prediction accuracy, we further examined model robustness, which is quantitatively defined as the relative percentage change in prediction error when outliers are introduced. Under this definition, the proposed augmentation method enhances robustness by mitigating the adverse effects of outlier datasets. These findings demonstrate the effectiveness of the method for regression on sequential process data, particularly under conditions of limited labeled data and the presence of outliers.