MCAGU-Net: A model for composite fault diagnosis of multi-sensor node networks

Diagnosing sensor node faults into specific types is beneficial for selecting targeted fault handling strategies and recovering fault data. However, current diagnostic models for multi-sensor nodes ignore the potential interactions between sensing units within nodes, resulting in a low diagnosis accuracy and the failure to clearly identify faulty sensing units, hindering effective fault recovery. To address this, this paper proposes, for the first time, a comprehensive solution for sensing unit level composite fault diagnosis in multi-sensor nodes. The approach utilizes graph structures to uniformly model the signal features of multiple sensing units and their correlations, thereby assisting the diagnostic model in effectively capturing composite fault characteristics. Furthermore, a multi-channel attention graph U-Net (MCAGU-Net) diagnostic model is introduced, which reduces channel redundancy and adopts an attention mechanism to enhance multi-channel collaboration, effectively handling the diversity and complexity of composite fault patterns. Empirical results on six datasets show that this approach outperforms the baseline methods on most metrics. Particularly, the diagnostic accuracy reaches 96.9–98.4%, regardless of whether the datasets are class-balanced or imbalanced. Compared to the baseline methods, this approach improves recall rate by over 6.2% on balanced datasets and by over 10.25% on imbalanced datasets. [Display omitted] •Measure heterogeneous correlation using mutual information and symmetric uncertainty.•Graph structure realizes the unified representation of signal and correlation.•MCAGU-Net uses SCConv to reduce feature redundancy to enhance feature representation.•Channel attention mechanism to adaptively learn the importance of different channel.•MSKloss foster composite fault diagnosis on imbalanced class distributions.