TwinMind AI

The Probem: Factories collect huge volumes of motor vibration data, but technicians lack time to analyze it. Late fault detection can cause unplanned downtime and costly repairs. What TwinMind AI Does TwinMind AI is a web dashboard that converts motor sensor recordings into maintenance insights using the Bently Nevada System 1 digital-twin dataset. It covers six conditions: baseline, imbalance, bent shaft, eccentric rotor, faulted bearing, and faulted coupling. Features: - Classifies conditions using 12 vibration features and shows sample agreement. - Explains predictions with alerts and practical inspection steps. - Flags missing values and sensor shifts, recommending human review when evidence is weak. - Provides sensor and anomaly explorers, model explainability, and what-if simulation. - Includes a maintenance assistant that uses saved screening results and inspection notes, identifies tools used, and records technician outcomes. How AMD Powers It: Powered by PyTorch and ROCm on an AMD Radeon Pro W7900D: - Screening: 21,095 rows in 0.55 ms on GPU vs. 2.91 ms on CPU (5.3x faster). - Temporal model: 7.23 ms on GPU vs. 100.08 ms on the best CPU setting (13.8x faster). The AMD GPU tab displays benchmarks, screening results, and flagged samples ranked by score. The public demo runs on CPU; GPU results appear in the demo video. Business Value: Helps teams prioritize inspections, identify potential faults earlier, and document outcomes. Responsible Design: TwinMind AI provides decision support, not confirmed diagnoses. Qualified technicians must verify alerts. The 96.9% accuracy comes from a time-based holdout within one recording and does not prove performance on other motors. It analyzes historical recordings and is not a live safety monitor. Built With: Python, Streamlit, scikit-learn, PyTorch, AMD ROCm, AMD Developer Cloud, Plotly, and pandas