多源信息融合的核电循泵轴承健康状态增量评估

Incremental Health State Assessment of Guide Bearing in Nuclear Circulating Water Pump with Multi⁃signal Fusion Network

针对基于深度学习的导轴承健康状态评估模型在增量学习新任务后,几乎彻底遗忘之前学习的内容,导致模型对旧任务评估准确率下降的问题,提出一种多源信息融合的核电循环水泵导轴承健康状态增量评估方法。首先,设计多源信息融合网络,提取反映导轴承健康状态的高维融合特征;其次,通过减小多源信息融合网络的分类损失及蒸馏损失保留新旧任务知识,设计范例管理策略及最邻近均值分类器实现导轴承健康状态增量评估;最后,在核电循环水泵实验台架采集导轴承劣化数据,并对所提方法进行实验验证及对比分析。实验结果表明,所提出方法的准确率可达94%以上。

As the traditional deep-learning-based health assessment model trained on previous data is retrained on the newly coming data,catastrophic forgetting usually happens and results in an accuracy drop.To deal with this limitation,this paper proposes an incremental health assessment of guide bearing in circulating water pumps with a multi-signal fusion network.First,the multi-signal fusion network is designed to generate fusion highlevel features.Then,the knowledge from both old task and new task is distilled and learned by minimizing distillation loss and cross-entropy loss during network training under the help of old exemplars,respectively,and the incremental health state assessment is achieved through a nearest-mean-of-exemplars classification strategy.Finally,the effectiveness of the proposed method is verified by the degradation dataset collected under the circulating water pump test bench.Experiment results show that the health state of bearing is assessed with an accuracy of over 94%.