Fault detection and isolation (FDI) systems represent a crucial element for the commercial diffusion of SOFC-based power generation plants. The physical quantities measured in the plant during the functioning feed a statistical classifier, in view of detecting and identifying possible faults. The classifier training is performed off line, by using a large and representative dataset generated through an adequate plant model, which is able to simulate the plant functioning under faulty conditions. However, when the plant operating condition is different from those considered during the classifier training phase, a reduction in the FDI performance is observed. In this study, we show that the adaptation of the statistical classifier to a new operating condition, unexpected and not included in the training phase, can be successfully performed through a machine learning approach called Domain Adaptation (DA), which needs only a few samples of the physical quantities associated with the new operating condition. In this case, the FDI performance approaches the performance achievable if the new operating condition was completely included in the initial classifier training.
展开▼
