Performance Degradation Estimation of High-speed Train Bogie based on a Novel Multi-Scale Convolution Module

The bogie performance state of high-speed train is one of the most important indexes, which reflects the high-speed train health condition and is closely related to the high-speed train operation safety and comfort. As the high-speed train operates for a long time, the irregular engaging between the wheel and the rail will inevitably cause vibration, which results in physical damage of the train and further causes the bogie performance degradation of the train. Taking the lateral damper on the train bogie as the research object, this paper analyzes the shortcomings of the existing researches about processing the train operation monitoring signal in terms of multi-scale feature learning. This paper proposed a novel multi-scale convolution module named One-Dimensional Feature-Merging-Inception Module (1D-FMI-Module), and based on 1D-FMI-Module, a Multi-Scale Feature-Merging CNN (MSFM-CNN) has been built to estimate the physical damage degree, that is how much the bogie performance has degraded. The 1D-FMI-Module proposed in this paper is a kind of deep attention structure and can determine the response score of different channel features. 1D-FMI-Module can replace ordinary convolution layers or other convolution modules and is flexibly implantable. The proposed MSFM-CNN is an end-to-end structure, which does not need a lot of expert knowledge and avoids the feature loss caused by manual feature selection. Experiments have verified that 1D-FMI-Module and MSFM-CNN proposed in this paper is effective and superior.