Compound fault identification has always been a challenge in bearing fault diagnosis. Existing learning-based compound fault diagnosis methods require numerous labeled compound fault samples. However, it is impractical to collect numerous samples for each compound fault in industrial scenarios. Based on this, a Zero-shot compound fault diagnosis method based on semantic learning and discriminative features is proposed in this paper. Firstly, to solve the absence of fault semantic vectors, we construct novel single and compound fault semantics, which contain both manually constructed low-dimensional semantics from the original vibration signal and high-dimensional semantics learning from the Convolutional Autoencoder. Then a convolution neural network-based feature extractor is designed, and an adaptive edge center loss is defined on the feature extractor to maximize the inter-category distance and minimize the intra-category distance. Finally, the semantic vectors of the fault are embedded into the feature space, and the Euclidean distance is used to measure the distance between the fault features and the semantic vectors to identify the category of unknown compound fault. The proposed method is validated on a self-built testbed. The results showed that the accuracy of compound fault identification reached 78.74% in the absence of compound fault samples.

Several deep learning methods have emerged for fault diagnosis of industrial equipment in recent years. However, the realistic data set is often much smaller than the benchmark diagnostic data set due to the difficulty of fault data collection and labeling in realistic scenarios. Moreover, the collected fault data may come from various components with different fault categories. Therefore, existing deep learning based models have poor generalization capabilities for cases with only a few data when faced with new components. Herein, a Triplet Relation Network (TRNet) is proposed for cross-component few-shot fault diagnosis by learning from several related meta-tasks iteratively. We construct a dual-channel feature embedding module with shared weights to extract fault features and a relation metric module to adaptively measure the feature similarity of sample pairs. Further, in order to distinguish the most dissimilar samples in the same category (i.e., hard positive samples) and the most similar samples in different categories (i.e., hard negative samples), the hard sample recognition module is designed, combined with a triplet loss, to weaken the hard-to-discriminate feature of hard sample pairs, such that the TRNet are capable for task learning and feature learning to improve classification accuracy on target components. We conduct experiments on two publicly available datasets and one lab-built datasets. We validate the proposed method to classify with one, three or five instances in each category of the target component. The results demonstrate that the fault diagnosis performance of our model is superior to the state-of-the-art methods.