Semi-Supervised Learning for Forklift Activity Recognition from Controller Area Network (CAN) Signals.

Area of Science:

• Industrial Engineering
• Machine Learning
• Manufacturing Systems

Background:

• Machine Activity Recognition (MAR) is crucial for optimizing manufacturing processes.
• Existing MAR research has primarily focused on construction equipment, neglecting industrial forklifts.
• Forklift operations are vital in industrial settings but lack dedicated MAR solutions.

Purpose of the Study:

• To develop a data-driven method for forklift activity recognition.
• To leverage Controller Area Network (CAN) signals and semi-supervised learning (SSL) for enhanced MAR.
• To address the gap in MAR research concerning forklift operations.

Main Methods:

• Utilized Controller Area Network (CAN) signals from forklifts.
• Implemented semi-supervised learning (SSL) to train classifiers on large unlabeled datasets.
• Applied a two-step post-processing technique to refine recognition results.

Main Results:

• Achieved 88% balanced accuracy for driving activities and 95% for load-handling activities.
• Obtained a Matthews correlation coefficient of 0.82 for five activity classes, matching human expert performance.
• Successfully detected the transport of small loads, a task beyond the forklift's built-in weight sensor.

Conclusions:

• The proposed data-driven MAR method effectively recognizes forklift activities using CAN signals and SSL.
• This approach offers a valuable tool for monitoring and improving industrial operations, particularly in logistics and warehousing.
• The method demonstrates superior performance, even identifying subtle activities missed by onboard sensors.