CHILLER DIAGNOSTICS Machine learning approach Carrier

Sammanfattning: Chillers are large and complex machines that are used for temperature regulation in large buildings and plants. An undetected fault in the machine can lead to extended downtime and cause both great financial losses and increased environmental impact. Therefore Carrier, as manufacturer of chillers, is interested in developing fault detection algorithms capable of detecting the faults early so that they can be addressed before complications can occur. The faults, studied in this thesis, are refrigerant leakage, fouling (accumulation of unwanted material on a surface) in the evaporator and fouling in condenser, some of the main components in the machine. The thesis aims to implement a fault detection and diagnostic algorithm (FDD), using supervised machine learning methods, and provide accurate and reliable identification of chosen faults. FDD is done by creating three binary classifiers, one for each fault, and by training the classifier on a labeled dataset, simulated from a highfidelity chiller model. Inputs for data generation in the chiller model come from the field data from 35 days in summer. Machine learning models - decision tree, logistic regression, and support vector machine, were used to develop the FDD algorithm and their performance was compared with 4 key indicators: accuracy, false alarm rate, latency, and isolation. The results have shown that the problem occurred to be linearly separable, with respect to this particular dataset. Support vector machine model has achieved the highest scores in the detection of all three faults with the lowest fault severity level. Refrigerant leakage has proven to be the most difficult fault to detect, partly because the fault severity was almost close to normal operating conditions. Moreover, it was estimated that three signals were needed to detect leakage, while fouling could be separated by only the two most important signals. Finally, the transient data was removed and the training dataset was reduced to 1 day, which proved to be enough to detect and isolate the selected faults.