Zero-shot Fault Diagnosis under Incomplete Information

With the rapid development of industry, fault diagnosis has become the key to ensure the reliability and efficiency of industrial systems. In this work, we proposed a zero-shot fault diagnosis method under incomplete information. Firstly, based on Jensen-Shannon Divergence (JSD) calculation, this method constructs a classifier to divide the test samples into detected known classes and detected unknown classes. Then, the detected unknown classes are clustered by using the K-means algorithm. Also, to alleviate the impact of domain bias in zero-shot fault diagnosis, the classifier is used again to classify the detected known classes into sub-known classes and sub-unknown classes. Afterwards, fault features are extracted via a feature extractor, and they are calculated with the set semantic vectors. According to the calculation results, the sub-known classes and sub-unknown classes are classified. In order to simulate the situation where the fault samples have incomplete information due to missing data, the missing rate is set 10%, 30%, and 50% sequentially. By optimizing the feature extraction network structure, the prediction accuracy increased 10.53% at the missing rate of 30% on the Case Western Reserve University (CWRU) bearing dataset. On the American-Society for Mechanical Failure Prevention Technology (MFPT) dataset, the prediction accuracy increased 10.98% at the missing rate of 50%. These results show an improvement for zero sample fault diagnosis under incomplete information.