Development and validation of an artificial neural networkbased tool for fault detection and diagnosis in vapour compression heat pumps
In this paper, a tool for the detection and diagnosis of soft faults in heat pumping systems, based on Artificial Neural Networks, is presented. The tool is developed following a two-stage approach that consists of a fault detection stage and a fault diagnosis stage. In the former, first a series of experiments is conducted on a fully instrumented, lab-scale water-towater heat pump operating under fault-free conditions. An artificial neural network is then trained to accurately estimate thermodynamic parameters, such as the refrigerant pressures and temperatures in different locations, using the temperatures of the secondary fluids as input variables (predictors) in these conditions. Overall, the accuracy in predicting the pressures or the temperatures lies within ± 5 kPa and ± 0,5 °C respectively. These pressures and temperatures are then combined to define dimensionless predictors, as the ratio between thermodynamic parameters under actual conditions and those computed with the artificial neural network under fault-free conditions. These dimensionless predictors are subsequently compared with a predefined threshold; when this threshold is exceeded, the operating condition is classified as faulty, thereby triggering the fault diagnosis stage. In this, a classification model based on an artificial neural network is developed. This model uses the same set of predictors as input variables and maps them to the corresponding fault category as output. The results demonstrate that the proposed tool achieves an accuracy of approximately 91% in the fault detection, and of around 88% in the fault diagnosis.