Fault diagnosis of rotating machinery gearboxes based on low-rank adaptive convolutional network

Traditional fault diagnosis methods face challenges in extracting temporal features from measured signals and achieving noise-resistant robustness in practical measurement scenarios. To address this, this paper proposes a lightweight compound fault diagnosis model (TLCN) based on a low-rank adaptation (LoRA) convolutional network. Specifically, the model captures local transient features and global periodic patterns of vibration signals via multi-scale dilated convolution, enabling fine-grained measurement of critical signal characteristics. It employs LoRA matrix factorization to efficiently fine-tune convolution kernels, reducing parameter redundancy while preserving the integrity of measured feature representation. Additionally, hierarchical sampling and weight decay strategies are integrated to optimize classification boundaries, improving the discriminability of fault categories in noisy measurement environments. Results from three experiments demonstrate that the TLCN model achieves superior generalization and robustness, with quantified improvements in measurement accuracy and noise tolerance.