On the Stability and Homogeneous Ensemble of Feature Selection for Predictive Maintenance: A Classification...

Area of Science:

• Machine Learning
• Industrial Engineering
• Data Science

Background:

• Feature selection (FS) is vital for machine learning-based predictive maintenance (PdM), impacting model performance and system interpretability.
• Robustness and reproducibility of selected features are critical, especially for real-world datasets with low sample-to-dimension ratios (SDR).
• The stability of FS methods under data variations has not been extensively studied in the context of PdM.

Purpose of the Study:

• To evaluate the stability and performance of popular filter-based FS methods in PdM applications.
• To investigate the impact of FS ensembles on stability and performance indicators.
• To analyze FS method behavior across diverse milling datasets with varying characteristics.

Main Methods:

• Applied five-fold cross-validation to assess Fisher score, minimum redundancy maximum relevance (mRMR), and ReliefF.
• Evaluated three FS methods using macro-F1 score and feature selection stability metrics.
• Investigated the effect of homogeneous FS ensembles on performance and stability.
• Utilized four milling datasets with different operating conditions, sensors, SDR, and class numbers.

Main Results:

• Different FS methods exhibited comparable macro-F1 scores but significantly varied stability.
• Fisher score, both single and ensemble, demonstrated superior performance and stability in most scenarios.
• mRMR showed the lowest stability, highest variability across settings, and greatest benefit from ensembling.

Conclusions:

• FS method choice significantly impacts stability, even with similar predictive performance.
• Fisher score is a reliable FS method for tool condition monitoring in PdM.
• Ensembling can enhance the stability of FS methods, particularly for mRMR.