Transfer learning for process monitoring using reflection-mode ultrasonic sensing

Area of Science:

• Industrial Digital Technologies
• Machine Learning
• Ultrasonic Sensing

Background:

• The Fourth Industrial Revolution integrates digital technologies for manufacturing optimization.
• Ultrasonic sensors offer cost-effective, non-invasive real-time data collection.
• Supervised machine learning requires labeled data, which is scarce in industrial settings.

Purpose of the Study:

• To compare two domain adaptation methods for training machine learning models without labeled data.
• To enable accurate monitoring of industrial processes using ultrasonic sensor data.
• To investigate the transferability of models across different manufacturing processes.

Main Methods:

• Comparison of Single Feature transfer learning and Transfer Component Analysis (TCA) with three features.
• Utilizing a reflection-mode ultrasonic sensing technique.
• Case studies on industrial mixing and pipe cleaning processes.

Main Results:

• The Single Feature method achieved high prediction accuracies: 96.0% for mixing and 98.4% for cleaning completion.
• R-squared values reached up to 0.947 for mixing and 0.999 for cleaning time prediction.
• Demonstrated successful model transfer for monitoring industrial processes without target domain labels.

Conclusions:

• Ultrasonic measurements combined with transfer learning effectively monitor industrial processes.
• The Single Feature transfer learning approach shows significant promise for data-scarce environments.
• Further research is needed to address sensor location variability between process domains.