Multi-Dimensional Joint Attention Fault Diagnosis for Rolling Bearings Using EEMD Signal Decomposition

Fault diagnosis of rolling bearings is challenging due to the difficulty in feature extraction. The integration of deep learning models with signal processing techniques enhances signal quality and enables automatic extraction of intricate information from data, thereby offering a more comprehensive solution for fault diagnosis. To explore this field, this paper proposes a novel fault diagnosis model that leverages dual-scale convolution (DSConv) and a multi-dimensional joint attention mechanism (Mul-Dja) on signals decomposed by EEMD. Initially, the one-dimensional vibration signal is decomposed via EEMD, and the resulting components are filtered based on kurtosis criteria. These selected components are then stacked in the channel dimension to form multi-dimensional data. Subsequently, the DSConv module is employed to extract multi-scale fusion features. To enhance feature representation, two distinct attention mechanisms operating in different dimensional spaces are implemented to allocate attention weights to each feature dimension, both preceding and following the DSConv operation. Finally, a hybrid classifier architecture is constructed by sequentially combining LSTM with a pooling layer and a fully connected layer, which outputs the fault type classification. The evaluation results reveal that the proposed diagnostic method not only provides exceptional fault identification capabilities but also demonstrates remarkable robustness.