Feature-based intelligent models for optimisation of percussive drilling.

Area of Science:

• Engineering
• Machine Learning
• Geology

Background:

• Vibro-impact drilling (VID) systems use high-frequency impacts and rotation for rock penetration.
• Drilling through inhomogeneous rock causes multi-stability and varied impact motions, affecting performance.
• Some impact motions improve rate of penetration (ROP) and bit lifespan, while others hinder them.

Purpose of the Study:

• To develop intelligent models for classifying VID impact motions.
• To enable detection and maintenance of high-performance impacts.
• To avoid low-performance impacts for optimized drilling.

Main Methods:

• Utilized feature-based classification algorithms: multi-layer perceptron, support vector machine, and long short-term memory networks.
• Trained models using a combination of limited experimental and extensive simulated impact data.
• Validated trained networks using cross-validation on separate simulation-only and experimental-only datasets.

Main Results:

• Feature extraction from raw impact data is crucial for model performance.
• 42% of feature-based networks achieved >91% accuracy.
• 67% of networks achieved >77% accuracy on both simulated and experimental data.

Conclusions:

• Machine learning models can effectively classify vibro-impact drilling motions.
• Optimized impact classification can significantly improve drilling efficiency and reduce operational costs.
• Further research into feature engineering can enhance model accuracy for real-world applications.