Multi-label learning for fault diagnosis of pumping units with one positive label

Fault diagnosis using the indicator diagram is a fundamental method to evaluate the working status of pumping units. In applications, human experts typically identify only one fault for each indicator diagram. However, multiple types of faults may occur simultaneously. In this paper, we propose a Single-Positive Multi-label learning for Fault Diagnosis of Pumping Units (SPM-FDPU) algorithm to address this issue. Although trained on single-label data, it is capable of multi-label prediction. First, HU invariant moments and convolutional neural networks are used to extract common and label-specific features, respectively. Second, instance, feature, and label correlations are injected into the training process by feature and label manifolds to enhance supervised information. Third, the manifold is used to augment the latent label matrix to help explore discriminant information. Experiments are conducted on the three real indicator diagram data of an oil field and sixteen multi-label benchmark datasets. The results show that the accuracy of the proposed method has achieved 98% in diagnosing multiple faults on indicator diagram datasets, and the mean rank of the proposed method is optimal in terms of six popular evaluation metrics on multi-label benchmark datasets. The source code is available at github.com/Kqian2020/SPM-FDPU.