Safety of electronic and computer systems largely depend on their failure-free operation time (“lifetime” (LT)). The problem of predicting the degradation rate of technical system characteristics (“service life”) which determines LT, is often solved within the accelerated testing (AT) paradigm. However, it has its own specifics for prototypes of devices produced at the early stages of development work, when, having only a small number of device copies, it is necessary to estimate their potential service life (LT), if this characteristic is decisive for assessing the feasibility of continuing the development. This may concern both technical devices (and elements of “Cyber-physical systems”, in particular) and software products, for example, machine learning (ML) systems under development. This paper analyzes the extent to which state-of-the-art of machine learning (ML), as well as such specialized LT assessment models as selective censoring, Survival Analysis (SA), Extreme Value theory (EVT), allow obtaining service life forecasts for designed devices under real operating conditions at the early stages of development/design. The difficulties of solving the maintenance forecasting problem using traditional machine learning software tools are analyzed, and a heuristic method for solving the maintenance forecasting problem under given conditions is considered and proposed. It is noted that the problems under consideration have a similar fundamental nature for a wide variety of applications of electronic and computer technology, for which it is necessary to estimate the rate of degradation of efficiency. Examples will be given both from the modern practice of computer networks and from the practice of developing modern solar electric cells (SEC) of autonomous power supply, the lifetime of which is one of the factors of safe operation of various systems. At the same time, a common property of time series representing the efficiency of these systems in certain terms is, in many cases, self-similarity, which, under certain conditions, can allow forecasting based on individual prototypes of the designed devices. Proposed heuristic is demonstrated by the example of predicting the degradation of the performance of new designs of solar electric cells (SEC). The Self-similarity property of corresponding time series is demonstrated. The proposed heuristic refers to the extraction of the Hondrick-Prescott trend from a non-stationary time series representing the degradation of a quality characteristic. Examples of similar behavior of the degradation process for other systems are given.