关键词: 加速退化试验渊ADT冤曰长短期记忆渊LSTM冤网络曰小样本退化数据曰置信区间曰早期分型 

Abstract:  To address the scarcity of degradation data and failure samples and the difficulty in quantifying lifetime-predic鄄
tion uncertainty for high-reliability devices, this paper proposes a small-sample degradation lifetime prediction
method. The method integrates a long short-term memory (LSTM) network, Bootstrap ensemble learning and
Monte Carlo dropout (MCD). Firstly, the degradation sequences are transformed into sliding-window samples, and
Bootstrap is used to train a model ensemble. Secondly, in the prediction stage, by incorporating MCD sampling,
the data uncertainty and model uncertainty are jointly characterized, and point predictions as well as the 85%,
90%, and 95% prediction intervals are obtained. Early observations are then used for degradation typing, and
sub-models are trained separately to improve the stability of long-term extrapolation and to reduce prediction in鄄
terval width. Finally, the proposed method is validated using the illuminance sequences of 20 tungsten filament
bulbs obtained from an accelerated degradation test (ADT) at 80 益. The results show that the proposed method
can stably identify degradation types, accurately predict degradation trends and provide reliable remaining useful
life intervals. It thus offers a quantitative basis for the reliability assessment and life substantiation of highly reli鄄
able devices.

Key words: accelerated degradation test (ADT), long short-term memory (LSTM) network, small-sample degradation da原 ta, confidence interval, early typing