Tool Wear Estimation Using Clustering for Predictive Maintenance

The patent transitions from time-based maintenance (fixed schedule) to condition-based maintenance (real-time wear state monitoring) by changing the maintenance parameter from time to actual wear condition. This is achieved through continuous monitoring of machining parameters and using machine learning models to estimate tool wear state, allowing tools to be used until actual wear limits are reached rather than discarding them prematurely based on time schedules.

The patent replaces the mechanical/time-based maintenance system with an intelligent monitoring and prediction system. Instead of relying on fixed time schedules, the system uses sensors to collect machining data and machine learning algorithms (including clustering analysis and neural networks) to predict tool wear and remaining useful life, substituting mechanical judgment with computational intelligence.

The patent introduces machining parameters (current, voltage, power, acoustic emission, vibration) as intermediary indicators of tool wear. Instead of directly measuring tool geometry offline, the system uses these readily available process parameters as proxies for wear state, which can be monitored continuously without removing the tool. Machine learning models correlate these intermediary measurements with actual wear conditions.

The patent replaces direct mechanical measurement methods (optical microscopy requiring tool removal) with indirect sensing and computational analysis. Sensors monitor machining parameters continuously, and machine learning algorithms process this data to estimate wear, substituting physical measurement with intelligent inference from process data.

The patent enables continuous monitoring of tool wear during machining operations without interrupting the cutting process. Tools remain in position and continue machining while sensors continuously collect data, eliminating the stop-start nature of offline measurement and maintaining consistent cutting conditions throughout the monitoring period.

The patent uses machining parameters as intermediary indicators that can be measured without moving the tool. These parameters serve as proxies for wear state, allowing wear assessment while the tool remains in its machining position, thus avoiding the need to change tool position for measurement purposes.

The patent employs dynamic clustering methods (including Dirichlet process mixture models and adaptive neural networks) that can automatically adjust the number and characteristics of wear state clusters based on the data. Instead of using fixed, pre-defined clusters, the system dynamically creates clusters that adapt to the actual wear patterns observed, capturing unpredictable wear behaviors while maintaining model interpretability.

The patent uses unsupervised learning and clustering analysis to preliminarily identify wear state patterns and cluster structures from historical data before applying supervised prediction models. This preliminary exploration of data structure enables the system to adapt to specific wear patterns without requiring pre-defined cluster assumptions, improving both accuracy and adaptability.