A feature extraction framework with entropy on graphs for cross-dataset building fault detection

Simulation is commonly adopted in developing building automated fault detection and diagnosis (AFDD) strategies. However, simulations often fall short in accurately representing real-world scenarios, which hinders the efficacy of models trained on such data for identifying faults in actual buildings. To tackle this challenge, we present a new approach for feature extraction that leverages entropy obtained from graph structures. These structures are constructed based on features that can distinguish between normal and faulty conditions. This method includes acquiring graph structures from simulated data, extracting their entropies as features to train AFDD models. Then, the process of obtaining entropies from graphs is replicated for real building data, and the trained AFDD model is applied to conduct tests on them. Empirical findings illustrate that our suggested approach enables fault detection in real-world scenarios, even when the model is trained with simulated data. The features extracted by our proposed approach surpass the baseline, which consists of GNN embedded features, in terms of fault detection performance. Therefore, we infer that our method has the potential to take advantage of simulation for real building fault detection.