Online Tool Condition Monitoring Based on Parsimonious Ensemble

Area of Science:

• Manufacturing Engineering
• Machine Learning
• Data Science

Background:

• Existing tool condition monitoring (TCM) methods use batch processing, which is inadequate for high-speed manufacturing.
• Current TCM approaches require complete retraining for new machining parameters, increasing operational inefficiency.

Purpose of the Study:

• To present an advanced online TCM approach using Parsimonious Ensemble+ (pENsemble+).
• To develop a flexible ensemble learning algorithm capable of adapting to data stream characteristics.
• To reduce operator labeling effort through online active learning and feature selection.

Main Methods:

• Implementation of Parsimonious Ensemble+ (pENsemble+), an adaptive ensemble learning algorithm.
• Integration of online feature selection for dynamic attribute sampling.
• Incorporation of online active learning to minimize manual data labeling.
• Development of an ensemble merging strategy to reduce complexity while preserving diversity.

Main Results:

• pENsemble+ demonstrates adaptability with on-the-fly growth and shrinkage of ensemble and base-classifier structures.
• Experimental validation on metal turning and 3-D printing data streams confirmed efficacy.
• Performance was further assessed using benchmark concept drift data streams.
• The method achieved low structural complexity and significantly reduced operator labeling requirements.

Conclusions:

• pENsemble+ offers an effective solution for online tool condition monitoring in dynamic manufacturing environments.
• The adaptive nature of pENsemble+ addresses the limitations of traditional batch-based TCM methods.
• The integrated active learning and feature selection significantly decrease the burden on human operators.