Invariant Feature Exploration Generalization Network for High-Speed Train Brake Pad State Recognition under Variable Speeds

The brake pads are a critical component for the friction braking system of high-speed trains, which are prone to damage during long-term service. Friction-induced vibration signals include rich friction interface features, and reasonable analysis of vibration signals can predict brake damage in advance. However, the characteristics of braking interface signal are different under variable speeds, making it difficult to precisely recognize the health state of brake pads at unknown speeds. This article proposes an invariant feature exploration generalization network (IFEGN). The proposed method focuses on the feature alignment among domains and the differences within the domains. Thus, the invariant features of brake pads under variable rotational speeds are fully explored. This enables the model to generalize well to unseen data, even if there are huge differences among the source domains. The network utilizes knowledge distillation to learn the Fourier phase information to obtain internally-invariant features. It employs the local maximum mean discrepancy to align sub-domains to obtain mutually-invariant features. Various experiments show that the IFEGN model has better generalization ability than other models.