Attention-Guided Multifeature Fusion Convolutional Network for Machinery Intelligent Fault Diagnosis

Intelligent fault diagnosis has become a key approach for monitoring equipment conditions in modern industrial systems. For rotating machinery (RM) that operates under varying and noisy conditions, it is crucial to ensure not only diagnostic accuracy but also the trustworthiness of model decisions. However, most existing deep learning methods lack transparency and struggle to effectively utilize low-level information. To address these challenges, we propose an attention-guided multifeature fusion convolutional network (AGMFCN) that integrates low- and high-level features to fully exploit fine-grained information, enabling more comprehensive and trustworthy feature encoding for fault diagnosis. Furthermore, variational self-attention is introduced with sparse constraints on attention weights, allowing the method to focus on highly discriminative fault-related features and enhancing the trustworthiness of diagnostic decisions. Additionally, by fusing multiscale features obtained from convolutional layers with varying receptive fields, the method effectively aggregates crucial information across different scales, further enhancing its performance in various operating conditions. Experiments on two datasets validate the effectiveness of the proposed method, demonstrating high accuracy under noisy conditions and less dependence on the quantity of training sample sizes. Additionally, post hoc visualization of attention heatmaps reveals that the model consistently focuses on impulsive segments associated with fault events, providing interpretive insights and enhancing diagnostic trustworthiness.