Deep Learning with Emerging New Labels for Fault Diagnosis

Abstract

When recognizing conditions such as mechanical faults, the operating data samples each have a label indicating the sample type, whether normal or a known fault category. However, the class label set of the test data is the same as that in the training set, which means that samples having an unexpected or unknown type are simply incorporated into existing classes rather than being assigned a new label for inclusion in another class. To address this problem, we propose a method called the emerging new labels using a sparse auto-encoder (SAE) method. After extracting features using the traditional SAE, it constructs a distribution from the training sample and identifies those samples that do not fit into known categories. The latter receive new labels and are used for retraining the model, enabling it to identify the new condition. The simulation and gearbox fault diagnosis experiments confirm that our proposed diagnostic model effectively identifies new fault types in samples, improving the diagnostic model's general application.