Causal discovery-based external attention in neural networks for accurate and reliable fault detection and diagnosis of building energy systems

In the era of big data, data-driven models have become the most promising fault detection and diagnosis solutions to building energy systems, due to their high accuracy and good feasibility. Nevertheless, their high accuracy often benefits from their high model complexity which reduces their interpretability and generalization abilities. To overcome this barrier, this paper proposes a causal attention-based neural network model to enable neural networks to infer like domain experts. It innovatively combines causal discovery with neural networks in the form of external attention. Do-calculus is applied to estimate causal effects of faults on symptoms. The causal effects are adopted to calculate the external attention which constrains the learning of model weights. In this way, this model can learn real causal correlations between faults and symptoms approximately, which improves its interpretability and generalization ability significantly. The experimental data of 12 air handling unit faults from ASHARE RP-1312 are adopted to verify the performance of the proposed model. Seven representative data-driven models are selected as baseline models, including support vector machine, k-nearest neighbors, extreme gradient boosting, classification and regression trees, fully-connected neural networks, convolutional neural networks and backward structural causal model. The proposed model achieves both the highest diagnosis accuracy (100.00%) and the best local generalization ability (100.00%), compared to the seven baseline models. Furthermore, it is discovered that the way it makes decisions is easily interpretable and similar to the way domain experts diagnose faults. • A causal attention-based neural network model is proposed to diagnose faults. • Causality is incorporated into neural networks in the form of external attention. • The model has higher diagnosis accuracy (100%) than seven traditional models. • The model has a better generalization ability (100%) than seven traditional models. • The model has better interpretability than data-driven non-causal models.