Homologous association learning network for gearbox fault diagnosis under time-varying speed conditions

Abstract

Under time-varying speed conditions, vibration signals measured from rotating machinery, especially planetary gearboxes, are non-stationary, and fault features of the same category may drift across different speed ranges, thereby causing distribution mismatch among variable-speed samples and weakening the generalization capability of diagnostic models. To address this issue, a homologous association learning network (HALNet) is proposed. The main innovation of HALNet is to effectively associate time-domain and frequency-domain features under variable-speed conditions. First, a homologous time-frequency association learning strategy with a momentum reference mechanism is developed to encode paired time-domain and frequency-domain features from the same vibration segment into a shared latent space, where batch-wise cross-view comparison is performed and the homologous pair is driven by the association loss toward the highest matching response. Second, a similarity-guided heterologous-sample discrimination scheme is developed to select hard negative samples according to the response distribution, where discriminative constraints are imposed to separate heterologous samples with high similarity. In addition, a self-supervised learning scheme is introduced to transfer speed-invariant representations to downstream diagnosis. Experimental results show that HALNet achieves diagnostic accuracies of 100.00% and 95.96% in Cases 1 and 2, respectively, and outperforms five advanced comparison methods across multiple evaluation metrics.