Problem

The objective of the case study was to develop accurate predictive models that could identify potential commercial HVAC equipment failures. The goal was to detect pre-failure conditions within a specific time frame to allow for proactive maintenance services to be deployed. In addition, the project involved analyzing extensive and diverse data sources, including legacy systems, third-party data, time-series data, and location data.

The Before State

• Develop accurate predictive models to identify commercial HVAC equipment that is trending toward failure
• Models must detect pre-failure conditions within a timeframe that allows proactive maintenance services to be deployed
• Large volumes of disparate data sources (legacy systems, third-party, time-series, location data)

What KPI Delivered

• A series of XGBoost classifiers that could predict individual unit pre-failure conditions with over 90% accuracy

• Harmonized data sets that included terabytes of sub-second time series IoT streams, third-party API weather forecasts, and building/mechanical features

• Integrated MLOps prediction API into existing proprietary dashboards

The After State

• Remote building engineers were now able to be alerted to likely HVAC failures up to 5 days in advance

•  XX% Reduction in unnecessary truck rolls (service tech dispatched)

• Highly scalable and repeatable processes as additional sites and units come online