Validation of the self-diagnosis efficiency system (D6.3)

This deliverable describes the experimental validation of a FDD method within the TRI-HP project. A self-learning FDD for variable speed heat pump was developed previously and validated with simulated data. In the present work a series of faults were emulated in a real 10 kW variable-speed water-to-water heat pump developed within the project using propane as a refrigerant. The considered emulated faults were: evaporator fouling, liquid-line restriction, compressor valve leakage, refrigerant overcharge and refrigerant undercharge. A comparison between those faults and the no fault situation was made to obtain useful parameters for fault identification. The variables compared, also called features, were the COP, the electrical power consumption, heating power, subcooling, superheating, evaporation and condensation temperatures, discharge line and liquid line refrigerant temperatures, refrigerant mass flow and compressor speed. Those experimental test update the results obtained using simulated data regarding the fault impact on different heat pump variables. In general, every fault decreases the COP and increases the electrical power consumption. The evaporator fouling and liquid line restriction were the faults that have a lower and higher degradation effect on the COP, respectively. The FDD consist of different modules: the steady-state detector, the input space, the regression model and the diagnosis module. The steady-state detector eliminates non-steady data to increase the prediction efficiency and has worked correctly during the validation. The input space module classifies the data in groups defined by the driving variables. For the validation with simulated data, those driving variables where the compressor speed, the condenser outlet temperature and evaporator inlet temperature. During the experimental tests, the compressor speed changed causing a failure on the algorithm, since it was not able to differentiate if the speed changed due to a fault or due to a change in the heat demand. Therefore, the compressor speed was changed for the condenser inlet water temperature as a driving variable.The regression model represents the heat pump behavior in normal conditions. The detector was trained with 50 hours of data corresponding to 42 hours of dynamic conditions, 6 and 2 hours of static tests at 10.2 kW and 12.3 kW respectively. For dynamic conditions, a virtual model developed in TRNSYS emulated the demand and the heat source of a residential building in Zurich climate. Once trained, the algorithm predictions were compared with the heat pump measurements. If the residual between both was larger than the training error margin this was an indication of a possible fault. To avoid the false alarms, the possible fault indication must be present 80 % of the time for a 10 minutes period. When a fault is detected a warning flag is triggered. After that, a comparison of the trends of the features is made to diagnose which fault is occurring. The diagnose chart developed previously with simulation data does not work for the current heat pump because of the differences between the simulation and experimental trends. Because of that, the diagnose chart was updated with the obtained experimental trends. The algorithm detects COP drifts of 7 % in 10 minutes. The detector has high accuracy as shows no false alarms during the tests. The diagnosis module, after the update, could diagnose each fault, except evaporator fouling which was detected as a fault, but without a diagnose. The FDD run in parallel during those tests giving real-time information about the heat pump soundness. The algorithm run from an industrial computer located in the laboratory. At the same time, it was running in a pocket-size computer called Raspberry Pi. The low computational need of the algorithm allows us to implement it in portable platforms