Contents I. Introduction
High-speed trains operate under challenging conditions, including extreme temperatures, high pressures, rapid rotational speeds, and fluctuating loads. Prognostics and health management (PHM) for high-speed trains are essential to guarantee reliable performance, minimize maintenance expenses, and prevent catastrophic failures once the equipment is in service. [1]. End of Life (EOL), in the context of fault diagnosis, represents the critical juncture at which a system or component can no longer perform its designated functions reliably, necessitating its retirement or comprehensive overhaul. Accurate fault diagnosis is pivotal in determining the EOL by identifying and analyzing failure modes, degradation patterns, and wear-out mechanisms. Advanced diagnostic techniques, including vibration analysis, thermal imaging, and machine learning algorithms, make it possible to pinpoint faults early and understand their progression towards EOL. This proactive approach enhances the precision of EOL predictions and ensures timely maintenance interventions, thereby mitigating risks, avoiding catastrophic failures, and optimizing the lifecycle management of assets. Organizations can seamlessly transition from reactive to predictive maintenance by integrating fault diagnosis with EOL assessments, ensuring sustained operational efficiency and reliability. The automation guided by technical standards and the Open System Architecture for CBM [2], [3] is crucial for developing specialized applications like automobile maintenance [4] and other complex machinery systems [5]. OSA_CBM employs a layered data construct approach, illustrated in Fig. 1(a). Initially, data is collected from sensors and transducers at the base level, then progresses through various stages S1, S2, S3…Sn, being characterized into states, conditions, and diagnoses. This data is then refined into prognostics, ultimately used in final user activities at the highest abstraction level. Mainly, the diagnosis and prognostics, which refer to classification and RUL prediction, are central to OSACBM [6]. Traditionally, OSA-CBM architecture executes diagnosis and prognostics sequentially, necessitating expert knowledge and manual design to transition from diagnosis to prognostics despite originating from the same data source. This classical structure is built on signal processing techniques, including hidden-Markov models [7], different neural networks (NNs) [8] – [10], and support vector machines [11], which typically have shallow network structures.
