Space habitats are unforgiving environments where a failing pump or a drifting CO2 scrubber can escalate from a warning to a catastrophe. LHADA - Lunar Habitat Anomaly Diagnosis Agent - is an AI-powered diagnostic system built to close that gap. LHADA ingests continuous telemetry from habitat life-support systems and runs it through a multi-method anomaly detection pipeline, combining threshold-based monitors, matrix-profile analysis (DAMP/MDI via STUMPY), and statistical drift detectors operating in parallel. When an anomaly is flagged, the system doesn't just raise an alert: it hands the signal to a reasoning agent powered by Qwen 2.5 32B, served on AMD MI300X hardware via vLLM, which cross-references a structured knowledge base to generate a human-readable diagnosis, likely root cause, and recommended action. The system was validated against the EDEN ISS 2020 dataset - a closed-loop Antarctic greenhouse and the closest publicly available analog to a space habitat life-support system. The implemented scenario is a thermal-loop coolant leak, inspired by ISS operational history, where gradual temperature deviations across the coolant circuit must be caught and diagnosed before they compromise habitat integrity. The full stack can run on-premises with no cloud dependency, making it viable for the high-latency, air-gapped reality of lunar operations. A real-time Gradio frontend streams live sensor data, anomaly annotations, and LLM diagnostic reasoning via WebSocket. A FastAPI backend exposes REST and streaming endpoints for integration with broader mission-control systems. LHADA demonstrates that AMD GPU infrastructure combined with open-weight LLMs can deliver mission-critical, explainable AI at the edge, where astronaut safety depends on fast, trustworthy answers.
Category tags: