Lightweight and Highly Robustness Network With Multi-Domain Adaptability

Traditional deep learning methods have problems such as low efficiency and poor robustness in satellite gyroscope rolling bearing fault diagnosis. This paper presents an advanced fault diagnosis model that enhances traditional convolutional neural networks (CNN) to address these limitations. The proposed model utilizes an innovative architecture consisting of a hierarchical fusion of multidimensional features, convolutional block attention module (CBAM), residuals and fast fourier transform (FFT) to process raw data into a one-dimensional signal format suitable for model input. This design facilitates efficient feature extraction and data processing, leading to the output of classification results through a specially designed classification header. Compared with a single model, the proposed model has a 4.97% improvement on the gyroscope-bearing fault diagnosis dataset. In addition, compared with existing deep learning methods, our model offers significant improvements in feature extraction and parallel computing capabilities. Notably, it achieves good performance in presenting strong noise interference with only a small amount of computational resources (parameters: 10.005 K) and (MFLOPs: 0.792). The diagnostic model can be practically applied to sensor-equipped rotating machinery for real-time online monitoring and early detection of potential bearing faults. It provides new ideas for subsequent research; Based on the existing methods, more efficient and robust solutions can be developed through the optimization combination of different methods.