A Generative Model for Zero-Shot Learnning Fault Diagnosis of Rolling Bearings

The current technology for fault diagnosis, based on deep learning, is experiencing rapid development. Its layer-by-layer processing and inherent feature transformation have garnered significant attention. Nevertheless, traditional deep networks utilized for fault diagnosis necessitate a substantial amount of labeled data, and the outcomes are heavily dependent on the quantity and precision of these labels. In practical applications, the challenging working environment often makes it difficult to gather labels, thus limiting the application of these traditional deep networks. In response to these challenges, we have developed a Zero-Shot Learning (ZSL) fault diagnosis model, leveraging a generative model. This model comprises three key modules: a semantic definition module, a feature extraction module, and a generative adversarial module. Firstly, we designed a novel semantic definition module to address the limitations associated with the low dimensionality and limited quantity of manually defined semantic attributes. Secondly, a feature extractor was developed to convert vibration signals into time-frequency maps and extract fault features using a Convolutional Neural Network (CNN). Furthermore, an adversarial module was introduced to train the semantic attributes and fault features of visible class faults in an adversarial manner, enabling the generation of faults belonging to unseen classes. Finally, the model identifies faults by measuring the similarity between the actual unseen class fault and the generated unseen class fault. The effectiveness of our model was validated using the dataset provided by Case Western Reserve University. Notably, the recognition accuracy reached 87.32% without the need for any faulty samples.