深度强化学习模式下空间滚动轴承变工况寿命状态表征和预测方法研究

A novel lifetime identification and prediction method for space rolling bearing is proposed based on deep reinforcement learning theoretical framework, in order to overcome the limitations that the existing prediction method does not suit to space rolling bearing life prediction, and solve the key issues existed in the vibration analysis based identification and screening of bearing, such as weak feature decoupling and extraction, lifetime characterization, identification and prediction. The research stratagem is as follows. A novel integrated framework of perception and decision making, which can deal with weak feature decoupling and extraction, lifetime characterization, identification and prediction, will be proposed to form the new principles and methods for life prediction. A novel method of weak feature extraction based on code fuzzy association and anti coding weighted fusion will be proposed to achieve weak feature decoupling and extraction. A novel method of dynamic pattern recognition based on cross layer fusion of structural feature and feature vector mapping for different conditions will be proposed to achieve same scale and tendency characterization of bearing lifetime state under different conditions. A novel method of evolution trend prediction based on deep network extension and sample knowledge reasoning will be proposed to achieve lifetime prediction of bearing under small samples. Finally, a system prototype of lifetime identification and prediction method for space rolling bearing based on deep reinforcement learning will be developed, and the system performance will be evaluated. This research will have important theoretical and practical significance in ensuring the performance, lifetime and reliability of space rolling bearing.

提出深度强化学习模式下空间滚动轴承寿命状态表征和预测新方法，目的是克服现有预测方法的局限性，在深度强化学习框架下解决基于振动分析的轴承鉴定筛选过程中微弱特征解耦分离、状态表征、识别预测等关键问题。研究思路为：提出集微弱特征解耦分离、状态表征、识别预测等功能于一体的感知与决策集成式滚动轴承寿命状态表征和预测理论架构，形成寿命预测的新理论和新方法；提出信号编码模糊关联和反编码加权融合的微弱特征提取新方法，实现微弱特征的解耦分离；提出结构性特征跨层融合与不同工况信号特征映射新方法，实现变工况下轴承寿命状态的同尺度趋势性量化表征；提出深层网络模型拓展和轴承样本知识推理的预测新方法，解决小样本非类同条件下轴承寿命预测问题；研发基于深度强化学习的空间滚动轴承寿命状态表征与预测系统原型，测评系统性能。该方法对保证空间滚动轴承的性能、寿命和可靠性具有重要理论和现实意义。

空间滚动轴承是卫星、空间站等空间飞行器的关键零部件之一，在地面模拟环境下开展空间轴承试验测试、寿命状态评估与预测、以及轴承鉴定筛选等是保障空间轴承达到预期性能、寿命和可靠性的重要措施。项目提出在深度强化学习框架下基于振动分析的空间滚动轴承寿命状态表征和预测新方法，主要围绕空间轴承寿命状态表征、性能演变趋势特征提取、状态识别预测模型构建与优化等理论与技术开展研究。主要研究成果和创新点包括：. 研究了基于振动特征提取的滚动轴承状态表征方法，针对变工况下滚动轴承振动特征迥异问题，提出了振动信号标准化的变工况下轴承状态同尺度表征方法、以及基于深度自编码网络的滚动轴承状态特征自动提取方法，实现了空间滚动轴承寿命状态表征；. 提出基于相似性度量的轴承性能退化指标构建方法，研究多指标融合的性能退化特征综合评价方法，提出基于深度网络跨层特征融合的性能退化特征提取方法，构建了表征空间滚动轴承性能演变的趋势性健康指标；. 研究了基于深度神经网络的空间滚动轴承自动特征提取与智能识别模型构建方法，针对不同工况下训练样本与测试样本分布差异导致滚动轴承寿命状态难以有效识别的问题，提出改进均衡分布适配的滚动轴承寿命状态识别方法，研究强化学习机制下的深度网络模型自适应构建与优化方法，针对深度模型轻量化部署以及低成本测试系统构建等应用需求，提出基于强化学习的寿命状态识别与预测深度网络模型压缩方法，实现了轴承状态识别与预测模型压缩和优化。. 利用空间滚动轴承振动测试数据对所研究提出的理论和方法进行了验证，项目相关研究工作总结形成了系列成果，可为指导空间滚动轴承设计与制造、开展轴承鉴定筛选、保证空间轴承使用性能提供支撑。