LiteEdgeNet: A lightweight edge diagnosis network using pyramid spectral convolution and efficient grouped multihead self-attention

Health condition of rotating machinery is vital for industrial safety, yet traditional methods struggle with computational efficiency and large-scale data, while deep learning models are often too complex for resource-constrained edge devices. A lightweight edge diagnosis network named LiteEdgeNet is proposed in this paper. To minimize the model complexity and enhance the shallow network’s ability of feature representation, Pyramid Spectral Convolution module (PSConv) is designed. PSConv transforms time-domain features into the frequency-domain by Real-valued Fast Fourier Transform (RealFFT), compresses feature length by half, captures multiscale information by pyramid convolution, and employs a feature fusion layer to establish inter-channel relationships. In addition, an Efficient Grouped Self-Attention module (EGSA) is introduced, which groups feature channels, applies independent linear layers to each group, and generates weight matrices for attention heads separately. The design of EGSA, alleviates the information coupling caused by the attention heads over-focusing on the same or similar regions, while reducing the number of parameters and computational cost. Evaluated on two benchmark datasets, LiteEdgeNet exhibits exceptional diagnosis accuracy and robust performance with minimal resource demands. The study also finds that LiteEdgeNet exhibits adaptability to few-shot scenarios. Furthermore, when deployed on edge computing devices, LiteEdgeNet exhibits efficient real-time diagnosis performance in resource-constrained situations.