A New Multi-Sensor Stream Data Augmentation Method for Imbalanced Learning in Complex Manufacturing Process

Area of Science:

• Industrial Engineering
• Data Science
• Machine Learning

Background:

• Manufacturing processes utilize multiple sensors for failure detection, but data is often imbalanced due to high process reliability.
• Imbalanced datasets lead to biased supervised learning models, hindering accurate failure prediction.
• High-dimensional multi-sensor stream data presents additional challenges for model development.

Purpose of the Study:

• To propose a novel and practical data augmentation method, imbalanced multi-sensor stream data augmentation (IMSDA), for imbalanced learning in failure detection.
• To generate high-quality synthetic failure data that preserves the temporal properties of multivariate time series.
• To enhance the performance of supervised failure detection models using augmented data.

Main Methods:

• Development of the imbalanced multi-sensor stream data augmentation (IMSDA) technique.
• Generation of synthetic failure data across all sensor dimensions while maintaining temporal characteristics.
• Training supervised failure detection models using a combination of real and IMSDA-generated data.
• Evaluation of model performance on a real-world industrial dataset.

Main Results:

• IMSDA effectively generates high-quality failure data, successfully reducing data imbalance.
• The augmented data preserves the temporal properties of the original multivariate time series.
• Supervised failure detection models trained with IMSDA show significantly improved performance.

Conclusions:

• IMSDA is an effective approach for addressing data imbalance in multi-sensor failure detection.
• The method enhances the reliability and accuracy of predictive maintenance in industrial settings.
• IMSDA offers a practical solution for improving supervised learning in complex manufacturing environments.