Detection of Chiller Energy Efficiency Faults Using Expectation Maximization

To detect degradation in energy efficiency of a chiller in a chiller plant, a multivariate Gaussian mixture model is applied. This classification technique was selected to take advantage of an expected correlation between measurable state variables and equipment and operation specifications and system control targets. The hidden variable is the faultiness of the chiller and can take on one of three possible states. The five observed variables correspond to sensor measurements that are typically available and monitored in commercially available chiller plants. The fault detection algorithm is trained on simulated data for the Molecular Foundry at the Lawrence Berkeley National Laboratory and tested on measured sensor data. The results show that detection of severe faults and no faults are relatively accurate, while detection of moderate faults is sometimes mistaken for severe faults. The computation needs are moderate enough for deployment and continuous energy monitoring. Future research outlines the next steps in regards to sensitivity analyses with alternate probability