Fault Diagnosis with Imbalanced Data Based on Auto-encoder

In class imbalance problem, researchers proposed data-level and algorithm-level methods successively. However, we found the imbalance is usually accompanied by the problem of small samples. For example, in the case of few fault samples and many normal samples, we need to deal with both imbalance problem and small sample problem. Based on the influence mechanism of the imbalance on the algorithm, this paper proposes a novel neural network model for solving the class imbalance problem with small samples. By changing the feature extraction process of auto-encoder, we can extract vector features, while the standard auto-encoder can only extract scalar features. By stacking the new auto-encoder model, we can deeply mine the data characteristics and extract the diversified vector features to complete the fault diagnosis task. Vector features effectively improve the feature extraction ability of the model and reduce the sensitivity of the model to class imbalance. The proposed deep model is verified using two types of fault diagnosis datasets and compared with other methods. The results indicate that the proposed method significantly improves the fault diagnosis accuracy in the task with class imbalance and small samples.