AFMCNN: A convolutional fusion diagnostic framework based on global distance matrix for train transmission system

Abstract

Sensor technology and continuously improved convolutional neural networks (CNN) are essential tools for intelligent diagnosis in train transmission systems. Numerous studies have focused on leveraging multi-sensor fusion and two-dimensional CNN to address diagnostic problems. However, research challenges remain to be addressed, such as expertise dependence and inadequate mapping when performing image coding. Additionally, many diagnostic frameworks still rely on conventional convolutional structures, constraining the extraction of features. Furthermore, the existing fusion approaches have seldom considered the issue of unbalanced distribution of diagnostic information among signals from different sources in the transmission system. To fill these research gaps, this paper proposes a global distance matrix (GDM) for image coding and an adaptive fusion multiscale CNN (AFMCNN) for multisensor fusion diagnosis in train transmission systems that can adaptively assign weights to different sensor information. First, the proposed GDM reflects the interrelationships of the time series data while preserving the temporal correlation. Then, a global attention mechanism is designed to improve the network’s attention to the global relationships of the data, considering the characteristics of the GDM. Furthermore, a novel multiscale convolution block is introduced to extract larger spatial information at different scales. Lastly, an adaptive fusion module is proposed to adaptively assign learnable weights for data from different sources at the channel dimension. The weights are visualized to increase the interpretability of the module. The excellent performance and generalization of the proposed methods are verified using bearing and gearbox datasets from the power transmission system.