Group optimisation-based adversarial domain generalisation network for rotating machinery cross-domain fault diagnosis using acoustic signals

Highly efficient and accurate fault diagnosis technologies for rotating machinery have significant importance for industrial manufacturing, operational safety and economic development. Acoustic signals offer advantages such as low cost, non-contact operation and simple installation, demonstrating significant application potential in condition monitoring for rotating machinery. However, the feature distribution is susceptible to variations in operating conditions, leading to severe domain shift problems. Particularly when target domain data is unavailable, model diagnostic performance degrades significantly. Meanwhile, existing domain generalisation methods struggle to effectively learn domain-invariant discriminative features when the number of source domains is limited. This leads to constrained generalisation capability and imbalanced generalisation performance across different classes. To address these challenges, this paper proposes a group optimisation-based adversarial domain generalisation (GOADG) network for cross-domain fault diagnosis of rotating machinery acoustic signals. Firstly, a feature extraction method for acoustic signals is designed. The acoustic signal is transformed into a log-mel spectrogram, followed by fault feature extraction using a designed two-dimensional deep convolutional neural network (DCNN). Secondly, the combination of group optimisation and adversarial learning enhances the capability of the model for fault diagnosis under unseen operating conditions. Group optimisation enables the model to learn genuinely domain-invariant discriminative features by focusing on the worst-performing classes within the source domain. Adversarial learning aids in extracting domain-invariant features across multiple source domains, improving the generalisation capability of extracted fault features. Experimental results on the planetary gearbox fault acoustic signal dataset demonstrate that the proposed GOADG model achieves superior diagnostic performance under unseen operating conditions compared to four other domain generalisation methods.