Generative Oversampling and Deep Forest based Minority-class Sensitive Fault Diagnosis Approach

In the actual industrial production processes, various faults occur at different frequencies and the resulting fault data may be class imbalanced. This means machine learning-driven fault diagnosis methods have to learn from imbalanced data, and accordingly lead to lower diagnostic accuracy or even directly errors in identifying minority class. To solve this problem, we present a novel Minority-class Sensitive Fault Diagnosis approach (MSFD), which can reduce the imbalance of data and enhance the sensitivity of our diagnostic model to minority-class samples. Specifically, we first design a new generative oversampling method by combining Wasserstein Generative Adversarial Network (WGAN) with Synthetic Minority Oversampling Technique (SMOTE) to balance the whole dataset and improve the distribution of the minority-class samples. WGAN is adopted to learn the distribution of minority-class samples and generate some minority-class samples as a supplement to the original dataset, while SMOTE is applied to the resulting dataset to further enhance the diversity of synthetic samples for weakening the influence from WGAN's mode collapse. In addition, a deep forest or multi-Grained Cascade Forest (GcForest) based minority-class aware fault classification model is developed. First, during multi-grained scanning processes, we score the forests and select the corresponding forests with higher scores to generate feature representations for accelerating model convergence. Second, weights are introduced for different forests in cascade levels to further improve the overall performance of our fault diagnostic model. A series of experiments are conducted to testify the effectiveness of our proposed method, and the experimental results show that our approach can synthesize new minority-class samples with higher qualities and improve the diagnosis performance for minority-class samples as well as its overall classification accuracy. Meanwhile, in case of extremely imbalanced datasets, the proposed approach still maintains a relatively high recognition rate for minority-class samples.