An Adaptable Cost-sensitive Imbalanced Fault Diagnosis Method Considering Sample-significance Information

Imbalanced datasets acquired from machinery are common in real cases and some cost-sensitive learning-based methods have been presented to tackle the imbalanced fault diagnosis in recent years. However, these methods manually assign misclassification costs that remain unchanged during training for each health condition and therefore they cannot be applied generally in different scenarios. Moreover, the samplesignificance information of each sample of the same health condition is not considered, causing low diagnostic accuracies. To deal with the weaknesses, this paper proposes a new adaptable cost-sensitive imbalanced fault diagnosis method. The architecture of the proposed method includes a feature extraction module and a cost-sensitive module. Specifically, the adaptable cost-sensitive parameter and the sample-significance information loss function are introduced into the training process to pay more attention to the minority classes and the hard samples. The imbalanced dataset of gears is employed to validate the effectiveness of the proposed method. The experimental results indicate that the proposed method can achieve high diagnosis accuracy and is superior to some existing imbalanced fault diagnosis methods.