Project Description #
More than 1.7 million U.S. households rely on mini-split heat pump (MSHP) systems, and the global market is projected to exceed $28 billion by 2035. Yet many of these systems fail silently. Soft faults such as airflow restriction, refrigerant undercharge, and coil fouling reduce efficiency by 15 to 30 percent long before diagnostic codes appear. During these undetected periods, homeowners spend more on energy, equipment experiences unnecessary wear, and harmful refrigerants may leak into the environment. Only after a system eventually shuts down do contracted repair technicians (“contractors”) receive emergency calls, often with little insight into what went wrong. No existing tool converts the operational telemetry already collected by MSHP service infrastructure into early-stage diagnostic insight for contractors.
Airlytics addresses this gap by giving contractors a predictive diagnostic tool that identifies emerging faults before they become severe. The system uses only data available through the Fujitsu UTY-ASSXZ1 service connector to detect outdoor coil airflow restriction before the unit throws an error code or shuts down. Airflow restriction was validated as a high-frequency field failure mode by Ryan O’Grady, a Fujitsu Airstage Services engineer, and can be safely and repeatably induced in a laboratory setting for controlled experimentation. The system employs a two-stage machine learning pipeline. An Isolation Forest model performs unsupervised anomaly detection by learning patterns of healthy system behavior and flagging deviations. A Logistic Regression classifier then estimates the probability that the detected anomaly corresponds to an airflow restriction fault. The pipeline was trained on 2.28 million data windows derived from five weeks of continuous operation and 12 controlled fault induction experiments conducted in the Thayer Fluids Lab. The system achieved an AUC of 0.8516 on cross-validated fault classification, a false alarm rate below 5 percent, and detection lead times under 2 hours from fault onset. A structured usability evaluation with 10 participants, including an HVAC technician and a property manager, produced a mean overall usefulness score of 3.60/5.0, exceeding the 3.0 threshold established in the design requirements. Deliverables include a validated end-to-end data pipeline, a library of 12 fault induction experiments with ground-truth labels, trained ML models with verified performance metrics, a live contractor dashboard deployed on Heroku, user testing results, and complete system documentation for sponsor handoff. Because Airlytics relies entirely on existing OEM telemetry, it requires no additional hardware and can be deployed at minimal cost across large installed fleets, establishing the foundation for proactive HVAC maintenance in the $80 billion U.S. service market.
My Role #
I served as the Data Collection & DevOps/Cloud Lead on the Airlytics team. As one of four engineers working with our industry sponsor Terry McDonnell, I was responsible for the infrastructure that connected the physical HVAC system to the cloud and ultimately to the contractor-facing dashboard.
Specifically, I owned the data ingestion pipeline that captured raw .hst telemetry logs from a Fujitsu mini-split heat pump via a proprietary Android service connector, parsing and cleaning signals sampled every 5 seconds and saving them to a Neon PostgreSQL database at 10-minute intervals. I managed the cloud deployment on Heroku, where a live contractor dashboard surfaced real-time NORMAL, WARNING, and FAULT status alerts driven by our team’s dual-model ML pipeline. My work ensured that data flowed reliably from edge collection through to the deployed application across 5 weeks of total uptime and 12 test periods.
I also handled the DevOps side of the project, managing encrypted data transmission and secure server storage while ensuring no personal occupant data was collected—only system-level telemetry such as compressor RPM, fan speed, EEV position, and heat exchanger temperatures. The biggest roadblock I faced was working with the proprietary data system and the Android platform, which required significant effort to integrate into the broader cloud pipeline.
Code #
Available upon request, as this was a course assignment.