Fault prediction in HVAC chillers by analysis of internal system dynamics

The Chiller of a Heating, Ventilation, and Air-Conditioning (HVAC) system is a complex and expensive multicomponent appliance that is not impervious to failure. Predicting, or even identifying, a fault at its inception, can reduce the scale of the damage and mitigate the potential losses to be incurred, both financial and operational. This paper presents a systematic approach for the analysis of multiple streams of data from chillers to identify potential failures as soon as they become detectable from the data. The data streams are received from sensors in the IoT ecosystem of chillers to monitor the multitude of processes and parameters that are vital to their operation. Chillers have built-in mechanisms to generate alarms when key sensor values go beyond designated limits. A certain combination of these alarms is responsible for chiller failure, therefore, our proposed method needs to first predict these alarms using multi-sensor data fusion. Thus, in this IoT ecosystem there are two levels of sensor fusion for our predictive models: at the sensor level and at the derived alarms level. The final objective is to determine 'time-time-to-next-alarm' TA). The model for TTA is built using time-shifted sensor values. Since chiller failure is a function of sensor alarms, and both are binary in nature, a special technique of logistic circuits is used to mimic the combination logical circuit to predict the failure of the chiller.