A comparative analysis of edge and cloud computing architectures for rolling stock monitoring systems

This paper presents a quantitative comparison of Edge- and Cloud-oriented processing of technical-condition monitoring data for high-speed railway (HSR) rolling stock, based on a verified discrete-time digital twin of a train. Unlike approaches that treat Edge and Cloud as separate, independent configurations, the model simultaneously computes both processing pipelines within a single run, including a stochastic radio channel (correlated fading, Doppler-induced degradation, HARQ with residual block error), priority-based scheduling of two traffic classes (safety-critical and monitoring), and a thermal model of the on-board computer with throttling. A hybrid architecture mode, in which Edge pre-filters monitoring traffic before forwarding it to the Cloud, is additionally implemented and empirically evaluated. The full factorial design shows that the two architectures are bottlenecked by fundamentally different mechanisms: the Cloud degrades due to the physical unavailability of the radio channel, whereas Edge degrades due to thermal throttling under insufficient cooling. The simultaneous action of both factors drives the share of the critical Both Down state to a substantial fraction of the total observation time, whereas increasing the thermal headroom of the hardware platform alone is sufficient to reduce this share several-fold. Hybrid filtering delivers a measurable reduction in network traffic both under nominal conditions and under active anomalous load, but the accompanying ML-inference compute load measurably erodes the Edge node’s thermal headroom. Local computation proved substantially more energy-costly than radio transmission of the same volume of data. The results quantitatively substantiate the need for a redundant hybrid monitoring architecture in which no single subsystem constitutes a single point of failure.