Selective Mixup: Exploring Data Augmentation for Long Time Series in Confectionery Manufacturing

Shuhei J. Yamazaki, Takuya Maekawa

Abstract: In the era of Industry 4.0, accurate prediction based on time series data is increasingly important for improving productivity in manufacturing plants. However, in actual manufacturing, challenges such as noise in the data due to operator intervention and data imbalance are major issues in prediction. Data augmentation is a widely recognized method for addressing these issues and has been extensively studied. However, for relatively long time series data, existing data augmentation techniques often fall short due to the sequential nature of the data and other influencing factors. This paper proposes a novel approach that integrates data augmentation techniques for long time series data from manufacturing machines. We aim to enhance prediction accuracy by employing a novel selective mixup technique alongside classical methods such as scaling and jittering on time series data from manufacturing machines. Our method effectively mitigated the impact of imbalanced time series data and achieved robust predictions in the real-world scenario of predicting product quality in confectionery production. This study suggests that data augmentation for long time series data can contribute to more robust and efficient manufacturing systems.

External IDs:doi:10.1109/icps65515.2025.11087824