Generalized zero-sample industrial fault diagnosis with domain bias

Abstract

Generalized zero-sample fault diagnosis (GZSFD) is a challenging task involving the diagnosis of all samples from both previously seen and unseen faults. However, the scarcity of unseen samples for training causes that existing methods are hindered by domain bias, where unseen faults are more likely to be misclassified as seen faults. In this article, an efficacious solution is proposed by constructing an unseen fault detector for test samples in GZSFD with domain bias, which utilizes the detected unseen-sample knowledge to enhance the diagnosis performance. Specifically, a ResNet-based one-dimensional convolutional neural network is designed for high-quality feature extraction. Also, a Kullback–Leibler divergence-based distribution threshold detector is constructed for the identification of test samples. Afterwards, test samples are detected and distinguished into seen or unseen classes. In detected unseen classes, a zero-sample fault diagnosis (ZSFD) problem is undertaken, while in detected seen classes, a sub-GZSFD problem is addressed. For ZSFD tasks, to leverage the unseen samples in the test set, a clustering-based scheme without a predefined cluster number is used for the detected unseen fault. For sub-GZSFD tasks, combined with classification results in the ZSFD task, two embedding strategies are proposed to further mitigate the domain bias. They learn a shared weight and the optimal weights of semantic attributes from the feature space to the semantic embedding space, respectively. Using the shared fine-grained semantic attribute descriptions as auxiliary information, the final fault category can be determined. Experimental results showcase that the proposed strategies effectively alleviate the domain bias in GZSFD tasks.