Industrial Transfer Learning for Multivariate Time Series Segmentation: A Case Study on Hydraulic Pump Testing...

Area of Science:

• Machine Learning
• Industrial Manufacturing
• Data Science

Background:

• Industrial data scarcity hinders machine learning adoption in manufacturing.
• Transfer learning is a promising approach to mitigate data limitations.
• Time series segmentation is crucial for industrial process monitoring.

Purpose of the Study:

• Investigate the effectiveness of transfer learning for deep learning-based time series segmentation.
• Analyze the impact of data relatedness (close, distant, unrelated) on transfer learning performance.
• Evaluate transfer learning's benefits for industrial use cases, specifically end-of-line pump testing.

Main Methods:

• Applied deep learning models for time series segmentation.
• Utilized transfer learning by pretraining models with data from different domains.
• Compared model performance across three scenarios: closely related, distantly related, and unrelated source/target data.
• Assessed accuracy and training speed as performance metrics.

Main Results:

• Transfer learning enhanced accuracy and reduced training time for time series segmentation models.
• The most significant benefits were observed when source and target data were closely related, especially with limited target data.
• Negative transfer learning effects were noted when using unrelated datasets.

Conclusions:

• Transfer learning offers substantial potential for improving industrial machine learning applications like time series segmentation.
• Careful consideration of data relatedness is crucial to avoid negative transfer learning outcomes.
• Further research is needed to address the challenges and optimize industrial transfer learning strategies.