MachineMind: Predictive Maintenance Assistant

Problem Statement Unplanned mechanical downtime in rotating machinery costs industrial plants over $50B annually. Plant technicians face alert fatigue from noisy threshold alarms, lack of explainability from black-box AI models, and an absence of structured, safety-compliant physical inspection procedures. The MachineMind Solution MachineMind is an intelligent predictive maintenance assistant designed for plant reliability engineers and maintenance technicians. It monitors continuous multi-sensor telemetry (bearing temperature, vibration RMS, hydraulic pressure, and rotational speed) from critical industrial assets like centrifugal pumps and AC induction motors. Key Architectural Highlights 1. Unsupervised ML Anomaly Detection: An Isolation Forest model trained on nominal operational baselines detects multidimensional degradation without needing labeled failure datasets. 2. Dynamic Machine Health & Hysteresis: Maintains a continuous 0-100% health score and multi-sample temporal hysteresis, eliminating alert flickering and preventing false alarms during transient load shifts. 3. 45-Second Early Warning Lead Time: Empirically validated on progressive bearing degradation to alert crews 45 seconds prior to critical mechanical vibration thresholds. 4. Root-Cause Sensor Attribution: Quantifies relative percentage deviations per sensor to pinpoint mechanical root causes (e.g., bearing raceway spalling vs. coolant hydraulic failure). 5. Prescriptive Action & LOTO Checklists: Translates telemetry deviations into physical, step-by-step Lockout/Tagout (LOTO) compliant inspection tasks. 6. Certified Work Orders: Exports standardized maintenance tickets in JSON, Markdown, and ReportLab PDF formats. 7. AMD Hardware Acceleration: Features an honest hardware probe architecture detecting AMD ROCm and AMD Vitis AI / Ryzen AI NPU execution providers via ONNX Runtime, with verifiable CPU reference fallbacks.