This paper addresses the challenge of assessing and forecasting the technical condition of complex cyber-physical systems (CPS). Traditionally, this problem is formulated as a regression task, where multi-dimensional time series of system features serve as inputs, and the outputs correspond to a health index or time-to-failure metrics. A critical factor influencing prediction accuracy is the composition of the training dataset, which should consist of data from similar or homogeneous systems. To improve model reliability, we propose a novel method for identifying similar objects by extracting latent representation vectors from hidden layers that characterize system states and comparing these vectors. Experimental validation on the TURBOFAN dataset demonstrates the effectiveness of this approach, achieving a 34% reduction in the error of remaining useful life prediction. These results highlight the potential of latent vector analysis for enhancing prognostic models in CPS health management.