Fault Detection and Diagnostics for HVAC Systems Using Heat Meter Data, Statistical Features, and Data Mining

• Heating fault diagnostics using energy sub-meters and weather data • Simulation-trained fault model applied to real building data • Correlation and distribution-shape features enable fault separation • Transferable workflow for cross-building diagnostics This study presents an interpretable and scalable fault detection and diagnostics method for space heating and ventilation systems using only system-level heating sub-meter data and outdoor conditions. Although the primary focus is on heating system operation, cooling-related impacts are analyzed where operational faults influence heating energy behavior. The method addresses limited sensor infrastructure in existing buildings by relying on heat meters for space heating and supply-air heating, which are commonly available for billing purposes. Synthetic fault scenarios were generated with a calibrated building energy model, and statistical features were extracted from heating energy-use time series. These included skewness and correlations with outdoor air temperature and solar irradiance. Unsupervised clustering applied to these features distinguished ideal and faulty system behavior. Simulated ventilation-system faults increased annual supply air heating energy use by 22–46 %, while the space heating circuit fault produced only a minor impact. The clearest diagnostic patterns appeared during transitional weather with outdoor air temperatures of 5–20 ∘ C and direct normal irradiance above 35 W / m 2 , when heat recovery capacity should meet most heating demand and control instabilities become more visible. Validation using building management system data from university campus confirmed most ventilation-related faults, including persistent heating coil leakage and unstable supply air temperature control. The workflow narrows the diagnostic space rather than identifying root causes directly, helping building managers determine which systems require inspection. It requires minimal data, preserves interpretability, and can be integrated into existing building management platforms for continuous monitoring. Future work should address partial-fault cases and larger datasets to support probabilistic diagnostics.