Open-set domain adaptation fusion method based on weighted adversarial learning for machinery fault diagnosis

Traditional closed-set diagnostic methods assume identical label spaces for the source and target domains. Nevertheless, unlike the source domain, the target domain can include emerging unknown categories, and the target known categories are usually the subset of the source known categories. To deal with this open-set diagnostic problem, the paper presents a new open-set domain adaptation fusion approach using weighted adversarial learning (OSDAF). The OSDAF method requires no threshold judgment, and integrates three sub-models to achieve higher accuracy. An adaptive weighted learning strategy is developed and introduced to two adversarial learning classifiers for the first sub-model. The output difference between the constructed classifiers is maximized to identify target unknown category. Also, the output discrepancy is minimized to align target known-category samples with the same category of the source known-category samples. Then, three binary cross-entropy strategies and an entropy minimization scheme are designed to promote the discrepancy between known and unknown categories, thereby generating discriminant features and establishing the decision boundaries. For the other two sub-models, discriminant features are extracted from the first sub-model and applied to two different label propagation methods, so as to enhance the diversity of the recognition sub-models and facilitate the identification of the target samples. Finally, the efficiency and advantage of the presented approach are assessed using three machinery datasets. The comparison results reveal that the presented OSDAF is an effective method and outperforms typical closed-set and open-set diagnostic methods.