Benchmarking Multi-Step Sensor Forecasting for Threshold-Based Predictive Maintenance: An Integrated IoT-to-Enterprise Framework

This work presents an integrated IoT-to-enterprise prototype for threshold-based predictive maintenance and a comparative evaluation of 24-hour multi-step sensor forecasting. The system integrates ESP32-S3-based sensor acquisition, GPS, temperature and tyre-pressure monitoring, the Swift telemetry platform, and Salesforce-based maintenance workflows. The forecasting study compares persistence, seasonal naive, Random Forest, and Bi-LSTM models across three synthetic sensor datasets, with additional robustness, architecture and feature-ablation, threshold-event, and uncertainty analyses. Random Forest achieves the lowest MAE for engine and tyre temperature, while persistence performs best for tyre pressure. An external evaluation on the NASA C-MAPSS FD001 simulated engine-degradation benchmark further assesses the forecasting methods using engine-disjoint data partitions. The study emphasizes evidence-based model selection and distinguishes sensor forecasting performance from direct failure prediction or demonstrated maintenance outcomes.