Siamese DNFCVAE: A semi-supervised model for early fault detection of rolling bearings under noisy background

Abstract

Early fault detection in rolling bearings is essential for monitoring the condition of industrial equipment. However, detection accuracy is often limited by noise interference and weak fault signals, a semi-supervised method based on a Siamese Denoising Normalizing Flow Convolutional Variational Autoencoder (Siamese DNFCVAE) is introduced. Firstly, a novel Denoising Normalizing Flow (DNF) module is designed and embedded into the CVAE encoder, which dynamically adjusts denoising intensity at each transformation step through adaptive noise estimation networks and scaling factors, progressively filtering noise while enhancing posterior distribution flexibility. Secondly, a Siamese architecture is integrated with DNFCVAE to introduce contrastive learning mechanisms, explicitly modeling inter-sample similarities to amplify sensitivity to early weak fault signals. Finally, an adaptive loss weight optimization strategy is developed, which automatically balances reconstruction and contrastive objectives, and realizes more stable and efficient model training. Experimental results on two rolling bearing datasets demonstrate that the Siamese DNFCVAE model achieves superior fault detection performance under complex noise conditions, significantly surpassing existing methods and highlighting its effectiveness and practicality for early fault detection.