Estimating the remaining useful life of retrieved electrical components using machine learning in remanufacturing applications
Loading...
Date
2026-09-10
Authors
Νόβας, Αναστάσιος
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
Electrical and electronic assemblies recovered from returned domestic appliances are today sorted by visual inspection and fixed rules, so units with years of service left are scrapped and units close to failure are returned to use. A quantitative estimate of how much life remains would replace that judgement with evidence, but the setting denies prognostic methods what they normally require. A recovered component arrives with no monitored operating history, offers a single opportunity for measurement at inspection, and belongs to a class for which no run-to-failure dataset exists.This thesis develops an end-to-end framework that estimates remaining useful life under exactly those constraints, and carries the estimate through to an operational decision. Since no run-tofailure data exists for components of this kind, the training data is constructed. Established physicsof-failure relationships for the transistor, capacitors and inductor of a commercial motor-drive board are combined into a single system-level observable, the total power the assembly draws, and propagated across a Monte Carlo population sampled over manufacturing tolerances and operating conditions. A Neural Stochastic Differential Equation, whose drift and diffusion are both learned rather than assumed, is trained on the resulting run-to-failure trajectories. At inference the model is started from a single inspection snapshot and integrated forward as an ensemble, and remaining life is obtained as the first-passage time of the degradation state to its failure threshold, yielding a distribution rather than a point estimate. That distribution is converted into a disposition between reuse, refurbishment and retirement by selecting the action of lowest expected cost, so the probability threshold separating the outcomes follows from the cost structure of the operation rather than being chosen by the practitioner.The framework is instantiated on that board and evaluated on a population of units held out from training, each assessed from a single inspection snapshot, as a returned unit would be at the inspection station of an automated remanufacturing cell. It proves substantially more accurate than both an age-only reliability baseline and a constant-rate extrapolation from the unit's own signal, and the intervals it reports contain the true remaining life at least as often as they claim. Applied as a sorting policy, every unit it returned to service survived the period it was committed for, at a lower total cost than refurbishing every recoverable unit. The range of cost structures and commitment periods over which the assessment repays its own cost is characterised explicitly.
Description
Keywords
Remaining useful life, Remanufacturing, Neural networks, Predictive maintenance, Synthetic data