Tool State Recognition Based on POGNN-GRU under Unbalanced Data.

Area of Science:

• Engineering
• Data Science
• Machine Learning

Background:

• Accurate tool state recognition is crucial for extending tool life.
• Real-world tool sensor data often exhibits imbalanced characteristics.
• Graph Neural Networks (GNNs) excel at spatial feature extraction but struggle with temporal data.

Purpose of the Study:

• To propose an effective tool state recognition method for imbalanced data.
• To address the limitations of GNNs in temporal feature extraction.
• To improve the accuracy and efficiency of tool state identification.

Main Methods:

• Developed an Improved-Majority Weighted Minority Oversampling Technique (IMWMOTE) to handle data imbalance.
• Proposed a Pruned Optimized Graph (POG) data construction method using multi-scale, multi-metric basis and Gaussian kernel weights.
• Constructed a POGNN-GRU model to integrate GNNs and GRUs for deep spatial-temporal feature mining.

Main Results:

• The proposed IMWMOTE effectively alleviates data imbalance issues.
• The POG graph construction method provides a comprehensive data description.
• The POGNN-GRU model demonstrated superior performance in tool state recognition compared to baseline models.
• Achieved accuracy improvements of 1.62% and 1.86% on the PHM 2010 and HMoTP datasets, respectively.

Conclusions:

• The POGNN-GRU method offers a robust solution for tool state recognition with imbalanced data.
• The integrated approach effectively captures both spatial and temporal dependencies in sensor data.
• This work contributes to enhanced tool life prediction and predictive maintenance strategies.