A Scalable and Interpretable Machine Learning Framework for Predictive Maintenance of Soil Sensors in Precision Farming Using Edge–Cloud Architecture

Abstract

Precision agriculture and efficient environmental management depend on the dependability of sensor-based monitoring systems in soil and climatic applications. The upkeep and long-term dependability of the sensors producing this data have received little attention, despite the widespread use of machine learning (ML) for agronomic parameter estimation. A scalable and interpretable machine learning system for environmental sensor monitoring, diagnostics, and predictive maintenance is presented in this paper. The system can identify both sudden sensor failures and slow performance deterioration by combining time-series forecasting models with unsupervised anomaly detection methods. Through the use of SHAP and LIME, interpretability is integrated, allowing for clear and understandable diagnostics. With lightweight models appropriate for edge devices in low-resource contexts, the suggested architecture facilitates deployment in dispersed environments via edge-cloud integration.  The system's efficacy in problem detection, data quality preservation, and actionable insights is demonstrated by experimental findings using real-world sensor datasets and generated degradation scenarios.  An important step toward independent and reliable environmental monitoring systems is represented by this research.