Journal Article

Robustification of the Random Forest: A Multitude of Decision Trees for Fault Diagnosis of Face Milling Cutter Through Measurement of Spindle Vibrations

Sonar, Pradnya R., Atnurkar, Atharva M., Jogdeo, Atharva A., Patange, Abhishek D.

2024

Overview

Purpose Recognition of tool failure is an everlasting problem for the manufacturing industry, which leads to diminishing productivity and quality of the product. Much research has been conducted on intelligent tool condition monitoring (TCM) techniques. Machine-learning-based techniques comprise pattern recognition of signals to obtain intelligent decision-making. However, dealing with diversified data distributions’ present and future moments is complex and needs robust models. This paper aims to develop a generalized statistical model of spindle vibrations trained using random forest. Methods The real-time vibration evolved during machining was collected by performing experimentation considering different tool wear conditions of the face milling cutter. A thorough statistical analysis was undertaken considering hyperparameter tuning of the random forest—a multitude of decision trees, thereby attempting to provide generalization and robustness to the model. Results Compared to various hyperparameters, the maximum depth of the tree showed a more significant influence on the performance of the vanilla model as it regulates its growth on a macro-level. While tuning the tree concerning minimum sample split, no notable splits were observed because the minimum splitting requirement of the node attained a peak and led to the underfitting of random forests. Thus, the allocation of the improvised value of minimum sample split eliminates the anomaly posed by underfitting. Conclusion The robustness analysis of the random forest tree consisting of hyperparameter tuning has successfully eliminated the underfitting and overfitting of the model by showcasing an improvement and downfall in the performance, respectively. Here, the downfall in the accuracy indicates the ability of the tuned model to access unknown instances, thereby showing the generalization of the model concerning blind data. The results obtained using the overall framework indicate that the tuned random forest tree algorithm is appropriate and proficient for fault classification. This ML-based classification approach shall offer effective and optimum cutting tool usage and save considerable tooling costs.

• Publisher: Springer Nature Singapore,Springer Nature B.V
• Subject: Acoustics / Classification / Control / Cutting tools / Decision trees / Dynamical Systems / Engineering / Engineering Acoustics / Face milling / Face milling cutters / Fault diagnosis / Machine learning / Machinery condition monitoring / Original Paper / Pattern recognition / Real time / Robustness / Spindles / Statistical analysis / Statistical models / Tool wear / Tooling / Tuning / Vibration