The spectral method is based on analyzing the transient temperature changes in the collector circuit after the pump is started [5; 6]. Temperature signals on a secondly basis are transformed with a Fourier transformation in the spectral range. A failure free training phase results in a characteristic vector and an uncertainty boundary. A measured vector out of this range indicates a failure. Only one extra temperature sensor about a meter after the collector exit in the collector pipe is necessary. Several larger failures could be recognized, especially in high flow systems. These are e. g. a 40 % reduction of collector performance, a 20 % change in pump power and air in the heat exchanger. However, a failure free training phase of at least half a year is necessary and that may be difficult or even impossible.
3.2. Fault Detection with Artificial Neural Networks (ANN)
The development of a neural network-based fault diagnostic system for the solar circuit is still in a research phase. The method consists of three steps. In the prediction module, artificial neural networks are trained with fault-free system operating data obtained from a TRNSYS model. The model is trained so that 4 temperature values (collector in and output and storage in and output) can be predicted for different environmental conditions. The input consists of weather data (global and beam radiation, ambient temperature, incidence angle, wind speed, relative humidity, flow availability and inlet temperature), together with one of the other measured temperature values. In the second step residual values are calculated, which characterize e. g. the actual temperature increase in the collector compared to the predicted one. In the last step a diagnosis module is run. The failure detection was only successfully tested for introduced failures in TRNSYS [7; 8]. Since the network was trained with TRNSYS, and there are no measurement uncertainties it has to be seen how it compares to real system behaviour.