A Neural Additive Domain Adaptation Network with Interpretability for Fault Detection in Various Operating Conditions

Domain adaptation (DA) can be used for fault detection under multiple operating conditions. However, the lack of interpretability of DA models prevents potential users from understanding the decisions made by models. To give a solution, a neural additive domain adaptation network (NADAN) is proposed, which is constructed based on a neural additive model (NAM). Benefiting from the structural advantages of NAM, the predictions of NADAN are interpretable. Specifically, NADAN is able to explain how each feature affects the decision of the model. In addition, NADAN consists of two modules, feature learning (FL) and domain adaptation (DA), and achieves domain alignment through the combination of FL and DA. The FL module is pretrained to learn high-order features. The DA module takes learned features as input and achieves domain alignment based on multi-kernel maximum mean discrepancy (MK-MMD). Compared to the raw data, the learned features are more task-relevant, which helps to improve the performance of NADAN on DA. Furthermore, NADAN augments the feature dimensions during DA to measure the inter-domain differences in detail, which in turn achieves better domain alignment. Experiments on datasets generated from wastewater treatment simulation model demonstrate the effectiveness of the proposed NADAN.