Machine Learning Algorithms for Predictive Maintenance in Industrial Environments: A Comparative Study

Abstract

In the realm of Industry 4.0, the utilization of artificial intelligence (AI) and machine learning for anomaly detection faces challenges due to significant computational demands and associated environmental consequences. This study aims to tackle the need for high-performance machine learning models while promoting environmental sustainability, contributing to the emerging concept of 'Green AI.' We meticulously assessed a wide range of machine learning algorithms, combined with various Multilayer Perceptron (MLP) configurations. Our evaluation encompassed a comprehensive set of performance metrics, including Accuracy, Area Under the Curve (AUC), Recall, Precision, F1 Score, Kappa Statistic, Matthews Correlation Coefficient (MCC), and F1 Macro. Concurrently, we evaluated the environmental footprint of these models by considering factors such as time duration, CO2 emissions, and energy consumption during training, cross-validation, and inference phases.   While traditional machine learning algorithms like Decision Trees and Random Forests exhibited robust efficiency and performance, optimized MLP configurations yielded superior results, albeit with a proportional increase in resource consumption. To address the trade-offs between model performance and environmental impact, we employed a multi-objective optimization approach based on Pareto optimality principles. The insights gleaned emphasize the importance of striking a balance between model performance, complexity, and environmental considerations, offering valuable guidance for future endeavors in developing environmentally conscious machine learning models for industrial applications