Enhancing machinery reliability in lunar bases: Optimized machine learning for bearing fault classification in DC power distribution networks

In space missions and extraterrestrial habitats, ensuring the reliability of power systems is critical, particularly for DC distribution networks supporting lunar bases and space stations. These systems rely on rotating machinery such as motors and pumps, making the integrity of rolling bearings essential. There is a significant gap in robust fault detection and classification for such machinery under harsh, variable conditions similar to those in space. Existing machine learning (ML) methods often struggle to capture complex multi-channel patterns in sensor data due to overfitting, hyperparameter sensitivity, and high computational demands. This study proposes an ML-driven framework for fault classification in rolling bearings under extreme conditions, taking into account varying dataset sizes. Using three datasets, the proposed approach employs multi-variate variational mode decomposition (MVMD) and Hilbert-Huang Transform (HHT) to capture fault signatures and extract relevant features. To address overfitting and account for monotonic fault progression, this framework fuses four feature selection methods —Laplacian Score (LS), Minimum Redundancy Maximum Relevance (mRMR), ReliefF, and Mutual Information (mutInf)—with Spearman's rank correlation. The performance of ML classifiers (Neural Networks, Support Vector Machines, Naïve Bayes, K-Nearest Neighbors, Decision Trees, and Ensemble Methods) is optimized by adjusting hyperparameters using Bayesian Optimization (BO), Asynchronous Successive Halving (ASHA), and Random Search (RS), all in parallel settings to improve computational efficiency. These optimizers also help ML architectures to adapt according to available datasets of diverse types. Key quantitative results show that the ASHA-optimized ML model performs well with larger datasets, providing an overall accuracy of 99.94% with the reduced computational load. Meanwhile, BO and RS attained accuracies of 99.90% and 98.0%, which proved effective for scarce datasets. This innovative framework integrates signal decomposition, feature selection, and optimization techniques, creating an efficient predictive maintenance tool. It improves fault classification, boosting the reliability of machinery in extraterrestrial environments and enhancing the safety and sustainability of long-term space missions.