MicroAware

On a space habitat, the life-support fan is one of the most important machines on board, and one of the easiest to ignore until it fails. Crews are busy, sensors are expensive, and when something goes wrong the news has to reach Earth over a radio link that a solar storm can wreck. We wanted a system that does what an experienced engineer does: listens to a machine, notices when it sounds wrong, explains the problem in plain language, and makes sure the message gets through.

What it does?

MicroAware listens to a habitat’s life-support fan, catches faults by ear, explains them out loud, and gets the alert to Earth intact through a solar storm.

Hear it: An autoencoder learns what a healthy fan sounds like. Whatever it can’t reconstruct is the anomaly. We show this live as a “cancelled” spectrogram: the healthy sound is subtracted away, and only the fault is left on screen. Name and explain it: A small classifier identifies known fault types. Gemini turns each alert into a one-sentence explanation plus three things the crew should check. If Gemini isn’t available, a canned fallback takes over. ElevenLabs speaks the alert aloud, with the browser’s built-in voice as a backup. A voice agent answers crew questions using live system data. Send it home: Storm Shield sends each alert from one ESP32 to another over ESP-NOW. Erasure coding and interleaving let the message survive lost or corrupted packets. The level of protection adapts to live NOAA space weather data, or to a manual “solar flare” trigger. A third ESP32 plays the Sun and injects interference so the storm is real, not simulated on a slide. How we built it Audio anomaly detection: We trained an autoencoder on healthy fan audio, using reconstruction error as the anomaly signal. A lightweight classifier labels the known fault types. Explanations and voice: Gemini generates the alert explanations and checklists. ElevenLabs handles text-to-speech and the conversational voice agent. Both have offline fallbacks so the system never goes silent. Hardware link: Three ESP32 boards talk over ESP-NOW on 2.4 GHz: a habitat transmitter, an Earth receiver, and a “Sun” node that creates interference. Storm Shield: We split each alert into packets with erasure coding, interleaved them to spread out burst errors, and scaled redundancy up or down based on NOAA space weather feeds. Dashboard: A web dashboard shows the live spectrogram, fault classifications, alert history, link health, and current space weather. Challenges we ran into Getting the autoencoder to separate real faults from ordinary background noise. Making ESP-NOW transmission reliable under deliberate interference and tuning how much redundancy to add without wasting bandwidth. Keeping the whole pipeline working when cloud services drop, which is why every AI component has a fallback. Syncing three ESP32s, the ML pipeline, and the dashboard into one live demo. Accomplishments that we’re proud of: The “cancelled” spectrogram makes an invisible fault visible at a glance. Alerts survive a real, hardware-generated solar storm, not just a simulated one. The system adapts to live NOAA data, connecting a hackathon demo to actual space weather. Graceful degradation: no API key, no internet, no problem. MicroAware still detects, explains, and speaks. What we learned Anomaly detection with autoencoders is powerful when you can’t collect examples of every possible failure. Erasure coding and interleaving are practical tools, not just textbook theory, and they work on cheap microcontrollers. Designing for failure from the start (fallbacks everywhere) made the project much more robust. What’s next for MicroAware? Run the anomaly model directly on the ESP32 for fully edge-based detection. Expand to more machines: pumps, compressors, and CO₂ scrubbers. Grow the fault library with more labeled data. Test over longer distances and with real radio delay to better simulate Earth–habitat links