LAFICNN: A Novel Convolutional Adaptive Fusion Framework for Fault Diagnosis of Rotating Machinery.

CNN-based approaches have been widely developed and applied in intelligent diagnosis. However, only some CNN models make sufficient use of various favorable information in images, limiting adequate feature extraction. Moreover, existing models are still dominated by single-channel types, with limited multisensor collaborative diagnostic framework development. To alleviate the above problems, we propose a diagnostic framework called lightweight adaptive fusion improved convolutional neural network (LAFICNN). Specifically, a global spatial attention mechanism is first developed, which considers the nature of images and enhances the model’s attention to the long-range dependencies of signals. Then, a novel Conv Block that can better utilize the spatial information in the image is designed, and a single-channel lightweight improved convolutional neural network (LICNN) is constructed. In addition, an adaptive fusion module (AFM) is proposed, which can adaptively assign learnable weights to features from different branches at a shallow layer of the network without increasing the training parameters. Finally, the LAFICNN is proposed by embedding the AFM into the LICNN to realize multisensor collaborative diagnosis. Experimental results on a rolling bearing dataset show that LAFICNN improves the performance by 0.8% to 4.2% compared to the existing works. The stability of the LAFICNN is verified using a gearbox dataset.