Enhancing Fault Diagnosis with a Hybrid Attention Mechanism and Spatio-Temporal Feature Mining Model Using Small Sample Data

Fault diagnosis in rotating machinery presents unique challenges due to the limited availability of effective samples, which can lead to a long-tail effect during model training. To overcome these challenges, this paper introduces a novel small-sample fault diagnosis model for rotating machinery named the parallel convolution and bidirectional gated recurrent unit of attention model (PConvBiGA), which focuses on spatial and temporal feature extraction. The model is comprised of two core components: the wide convolutional denoising block (WCDB) and the hybrid attention mechanism spatio-temporal feature mining model (HAM-STM). The WCDB employs a one-dimensional wide convolutional layer coupled with a parallel pooling strategy, which helps preserve essential information from the original vibration data while filtering out weak noise and irrelevant features for diagnosis. To counter the feature loss typically encountered in traditional spatio-temporal feature fusion methods, the HAM-STM is developed. This structure utilizes a hybrid attention mechanism to effectively allocate high-dimensional feature weights across different paths, enhancing the model’s ability to mine features from various channels’ spatio-temporal characteristics. This, in turn, boosts the feature extraction capabilities of the model. Experimental results demonstrate that PConvBiGA successfully extracts more comprehensive features from vibration signals, enabling precise diagnosis with small samples and showing strong generalization and robustness.