A zero-shot learning method for fault diagnosis under unknown working loads

Data-based fault diagnosis is an important technology in modern manufacturing systems. However, most of these diagnosis methods assume that all the data should be identically distributed. In diagnosis tasks, this assumption means that these methods can only handle faults from the same working load. In real-world applications, the working load of the equipment varies for the different productions; if an unknown working load with no prior data available is given, then these traditional methods may be invalid. Zero-shot learning, using known data to diagnose the fault under unknown working loads, provides a transfer approach to solve this problem. In this paper, a zero-shot learning method based on contractive stacked autoencoders is proposed. The proposed method is only trained by the data from the known working load and can diagnose the fault from unknown but related working loads without prior data. The experimental results on the Case Western Reserve University dataset indicate that the proposed method performs better than the traditional methods under unknown working loads and has an accuracy of 97.82%. In addition, the analysis of the singular value and feature space also suggests that the proposed method is more robust and the feature representation is more contractive.