Abstract: Remaining useful life prediction models implement prediction according to monitoring data samples of equipment.When the pattern of the samples changes,models trained by original pattern samples will get poor performance on new pattern samples.To make models retain the ability to handle original pattern samples and get the ability to handle new pattern samples,three incremental learning methods were proposed.The loss function was constructed using labels of new pattern and original pattern samples in the first method;in the second method,a process of initializing models’ parameters was added based on the first method;in the third method a regularization item that represented the difference of distribution between current model and original model’s parameters was appended to the loss function used in the first method.Contrast experiments on turbofan engine dataset were carried out among different learning methods.The results confirmed that the proposed methods realized the target of incremental learning and the third method in proposed methods obtained the best comprehensive performance in the respect of the prediction accuracy and training time of the models.

Key words: remaining useful life, turbofan engine, deep learning, incremental learning

摘要： 机械剩余使用寿命预测模型依据设备状态监测数据样本进行寿命预测,当样本模式发生变化时,基于原模式样本训练好的模型在新模式样本上的预测表现往往较差。为使模型保留对原模式样本的处理能力,同时拓展出针对新模式样本的处理能力,提出了3种增量学习方法：第一种方法使用新模式样本与原模式样本的标签值构建损失函数;第二种方法在第一种方法基础上增加了模型参数初始化步骤;第三种方法在第一种方法的损失函数中增加了表示当前模型与原模型参数分布差异的正则化项。在涡扇发动机数据集上进行了不同学习方法的对比实验,结果表明所提方法实现了增量学习的目标,其中第三种方法在模型预测准确度和模型训练时长两方面取得了最优的综合表现。

关键词: 剩余使用寿命, 涡扇发动机, 深度学习, 增量学习