Abstract: RUL prediction was the core problem of reliability evaluation of high-power semiconductor lasers under various environmental stresses. In practical applications， the existing SVR methods all focused on minimizing the overall errors of the regression curve of the trained model， so as to pursue the generalization， which often resulted in unsatisfactory prediction results at the critical early warning stage， especially before the near-failure， and failed to ensure its reliability. Therefore， the SVR model training method was proposed based on cluster sampling. The observation data in the later period of the test samples were sampled for multiple cluster and then used for the SVR model test. The parameters in the SVR model made the SVR model fit the data in the later period of the training samples better. The effectiveness and robustness of the proposed method were verified by a case study， the results show that the performance and practical value of the proposed method are better than those of several representative small sample analysis methods.

Key words: remaining useful life（RUL）prediction, cluster sampling, support vector regression（SVR）, semiconductor laser

摘要： 剩余使用寿命（RUL）预测是高功率半导体激光器在各种环境应力作用下可靠性评估的核心问题。在实际应用中，现有支持向量回归（SVR）方法均侧重于保证所训练模型的回归曲线的整体误差最小，以追求方法的泛化性，这往往造成关键预警阶段特别是临近故障失效前的预测结果不理想，不能可靠地支持维护决策。提出了一种基于整群抽样的SVR模型训练方法，对测试样本后期观测数据进行多次整群抽样后用于SVR模型测试，SVR模型中的参数使得SVR模型对训练样本的后期数据拟合得更好。实例分析验证了该方法的有效性和稳健性，研究结果表明，所提方法的预测性能和实用价值优于现有几种代表性的小样本分析方法。

关键词: 剩余使用寿命预测, 整群抽样, 支持向量回归, 半导体激光器