Fault detection and diagnostics of air handling units using machine learning and expert rule-sets

To reduce HVAC energy inefficiencies, fault detection and diagnostics (FDD) has become a growing field of interest. In particular, air handling units (AHU), devices that circulate air and regulate room temperature and humidity, are the primary focus of most HVAC FDD systems. A data-driven FDD for AHUs on a university campus would fill a role in the reduction HVAC energy consumption, which remains one of the main drivers in total building energy use and consequently impacts the total global CO2 emissions. Fault detection and diagnostics for HVAC systems can potentially reduce 10-40% of total building energy consumption.

The aim of this project was to:

Determine the optimal fault detection methodology for AHU steam and chilled water leakage in order to prioritize maintenance and rehabilitation of valves, and to monitor and maintain improved AHU operation and energy efficiency.
Provide general fault detection for an AHU system using an expert rule-set which combines the AHU Performance Assessment Rules with rule expressions developed for the 107 buildings (776 AHUs) in the UT Austin main campus dataset.

Web application view of the 15-minute resolution datasets collected between July 2017 and December 2018 for each of the 776 AHUs and 107 UT Austin main campus buildings.

Heatmap of faults detected through the APAR rule-set. AHU faults are aggregated for each campus building; individual building results broken down by their AHUs can be found through the web application along with the results from the additional rule-set algorithms.

Ten different supervised learning classification algorithms were analyzed as fits for leakage detection. The overall goal of detecting AHU leakage in a practical setting is to find a model that is accurate, and additionally meets the criteria of desired fault detection characteristics.

Criteria used in supervised machine learning model selection and evaluation.

Methods to improve the accuracy of classification models included preprocessing data with feature transformations, examination of feature collinearity and skewness, cross validation for resampling, parameter tuning by iterating over a range of input values, and validation of final models for heating and cooling data.

Violin plots showing the predictors versus the outcome classes of normal and leakage and chilled water and steam. The width of the plots represents the number of data observations at that feature correlation.

Results indicating the AHU valves with leakage were field validated. UT Austin facilities personnel can continue to monitor the fault detection and diagnostics results through a dashboard web application developed to automatically update with new data scraped from the building automation systems which manage the campus.