A Cost-Sensitive Dense Network for Fault Diagnosis under Data Imbalance

Intelligent Fault Diagnosis (IFD) is crucial to guarantee the secure and stable functioning of mechanical equipment. The development of deep learning is continuously injecting vitality into IFD. However, in real-world industrial scenarios, obtaining sufficient fault data is difficult, and the fault data is much less than the normal data. Therefore, existing deep learning methods degrade performance when dealing with imbalanced fault diagnosis tasks, which poses a significant challenge to IFD under imbalanced data. To solve the above issue, a cost-sensitive dense network (CSD-Net) based on the improved dense convolutional neural network (DenseNet) and adaptive weighted cross-entropy (AWCE) is proposed, which includes a fault classification module as well as a cost adaptive module. Specifically, the improved DenseNet is used as a feature extractor in the fault classification module to obtain a more efficient feature extraction capability with fewer training parameters. The scaled exponential linear units (SELU) activation function serves to increase the stability of the model. In the cost adaptation module, AWCE adaptively assigns more appropriate misclassification costs to each class to lessen the effects of data imbalance. Eventually, experiments with different levels of class imbalance are designed and confirmed the primacy and efficacy of the proposed method.