An efficient lightweight gearbox fault diagnosis method based on parameter optimization and adaptive attention enhancement across working conditions

Abstract Currently, most mainstream gearbox fault diagnosis methods rely heavily on deep learning techniques. However, such deep learning-based diagnostic models often face challenges such as high computational resource requirements, poor real-time performance, and low diagnostic accuracy when applied to other operating conditions. To address the aforementioned issues, a novel lightweight cross-operating-condition gearbox transfer fault diagnosis framework is proposed, incorporating parameter optimization and adaptive feature attention enhancement. The framework utilizes two-dimensional Gramian angular field images to extract deep gearbox fault information. An improved lightweight mechanism integrating multi-scale channel and spatial adaptive attention enhancement is first adopted to allocate feature weights. This mechanism amplifies critical features while effectively mitigating the influence of redundant or noisy features. Subsequently, a local maximum mean discrepancy loss is employed to achieve cross-domain feature alignment across different operating conditions. Finally, a multi-strategy improved triangular topology optimization algorithm is applied to fine-tune the model’s hyperparameters, boosting its cross-condition diagnostic capabilities and ensuring more accurate fault detection across different operational scenarios. Experimental results indicate that the proposed model achieved accuracy rates of 99.85% and 99.94% on two validation datasets, while maintaining diagnostic accuracies of 74.6% and 79.6% in the presence of significant noise. The model’s diagnostic performance and robustness are superior to those of other models, demonstrating its effectiveness in challenging conditions.