AI-Powered Maintenance Forecasting in Mechanical Systems: A Data-Driven Approach

Human-centered AI has entered the field of queries for PdM prediction to change mechanical maintenance from a reactive-based approach to a more failure-predictive and intervention-oriented method.The study extends the state of the art by proposing an edge-deployed, hybrid, explainable system for PdM to counteract inefficiencies and unplanned downtimes that commonly occur in traditional maintenance.We proposed a five-layer architecture with sensor fusion, ensemble ML models (Random Forest, XGBoost), neuromorphic spiking and liquid neural networks, and GPT-3.5-levelfine-tuned LLMs for diagnostic explanations.Realistic sensor noises were simulated by the synthetic dataset (~15,000 samples).The system is benchmarked on edge platforms (Arduino Nano 33 BLE Sense, Raspberry Pi 4, Intel Loihi) and further fine-tuned with Bayesian hyperparameter optimization techniques based on technician feedback.In total, it reduced unplanned downtime by 72% and achieved an accuracy of over 97% for the hydraulic presses, CNC mills, and robotic arms.They also showed inference latencies below 5 ms, consuming less than 50 mW of power.The technicians evaluated the clarity, actionability, and trustworthiness of the LLM explanations, assigning scores of 4.6/5, 4.4/5, and 4.2/5, respectively.The human-in-loop adjustments reduced false negatives by 4%.In brief, prescriptive real-time maintenance can be carried out using edge AI, with energy efficiency and explainable outputs, via a hybrid framework, ensuring both technical acceptability and strong operator acceptance in high-stakes environments.