A Semi-Supervised Fault Diagnosis Method for Gearbox Based on Convolutional Autoencoder

To address the issues of limited labeled data and the tendency of traditional supervised neural networks to overfit, resulting in low fault diagnosis accuracy in real industrial scenarios, this paper innovatively proposes a semi-supervised learning method based on Convolutional Autoencoder (SSCAE) for gearbox fault diagnosis. First, both labeled and a large number of unlabeled samples are used to pre-train the improved convolutional autoencoder(CAE), introducing a classification loss to guide the encoder in learning features beneficial for sample classification. The pre-trained encoder parameters are then loaded into a classifier, and the classifier is trained using the labeled information, with supervised contrastive loss introduced to fully utilize the labeled information. Finally, the method was validated on a gearbox dataset. The experimental results show that the proposed method, in situations with a small number of labeled samples, effectively utilizes the unlabeled information and achieves better fault identification performance than traditional supervised and semi-supervised fault diagnosis methods.