基于深度生成对抗网络的直升机动力传动系统智能健康预示研究

There still exist some challenges in the intelligent health prognosis research of helicopter power transmission system, such as insufficient and unbalanced fault samples, multi-source and redundant condition parameters, subjective and unsatisfactory analyzed results, and so on. In this project, the optimal combined deep generative adversarial network model is constructed, and the generated data with high quality of multi-source monitored parameters of helicopter power transmission system fault samples can be acquired, which is able to overcome the limitation of unbalanced and insufficient fault data. The selected multi-source generated data fusion strategy is designed based on sensitivity and contribution, and the fused feature can be defined to comprehensively characterize performance degradation of helicopter power transmission system, which is able to overcome the limitation of incomplete information contained in a single or a few monitoring parameters. The intelligent health prognosis model designed with deep wavelet network is constructed based on the multi-source generated data and its fused feature, and it can be used to automatically evaluate the current operation conditions and predict the remaining useful life of helicopter power transmission system, which is able to overcome the limitation that the useful information hidden in the highly nonlinear data is very hard to be accurately captured. The simulated verification, experimental verification and engineering application will be performed for the key parts of helicopter power transmission system, such as rotor, bearing, planetary gear, blade, oil pump, disk and so on. This project research is expected to make innovations in fault data generation and multi-source parameter fusion, which has potential to improve intelligent operation and maintenance technology of helicopter power transmission system.

直升机动力传动系统智能健康预示研究仍存在故障样本失衡不足、状态参数多源冗余以及预测效果主观欠佳等挑战。本研究项目拟建立最优组合的深度生成对抗网络模型，获取高质量的直升机动力传动系统故障样本多源传感参数的生成数据，克服故障数据严重失衡不足的局限。拟设计基于敏感性及贡献度的选择性多源生成数据融合策略，定义综合表征直升机动力传动系统性能退化的融合特征量，克服单一或少数监测参数所含信息不完整的局限。拟构造基于多源生成数据及融合特征量的深度小波网络智能健康预示模型，克服高度非线性数据的深层特征信息难以精确捕获的局限，实现直升机动力传动系统运行状态与剩余寿命的自动化推理。针对轴、轴承、行星齿轮、叶片、油泵和轮盘等故障频发部件开展仿真验证、试验验证及工程应用等研究工作。通过本项目的研究，可望在故障数据生成和多源融合等方面有所创新，提高直升机动力传动系统智能运维技术。

本项目旨在针对直升机动力传动系统智能健康预示难题，从高质量的故障数据生成、选择性的多源参数融合和早期异常检测及寿命预测三方面开展了深入研究。三年资助期内取得了如下成果：（1）提出了基于改进辅助分类生成对抗网络的多模式数据生成与评估策略，改善了数据生成中分类与判别性能的兼容性，并在数据-特征层面及个体-整体层面评估生成质量。（2）构建了极少样本下改进局部融合生成对抗网络驱动的智能故障诊断模型，通过融合特征和丰富信息选择性聚焦生成样本的局部细节，在极少样本下实现相似性和多样性的数据增强。（3）开发了多传感器驱动集成堆叠小波自动编码器的协同故障诊断框架，基于并行端对端的多传感器信息特征学习和可调整决策融合权重提升协同诊断准确率和稳定性。（4）提出了多源参数选择性融合驱动深度神经网络的跨域寿命预测技术，通过综合规范化处理、相关分析和流形学习定义融合健康指标，嵌入模型迁移提升寿命预测适用性。重点以转子、轴承、行星齿轮和锥齿轮等关键部件为研究对象，对所提方法开展了仿真和实验验证。基于以上相关研究成果，以第一/通讯作者发表期刊论文34篇（标注基金号），其中非开源SCI论文30篇（ESI热点3篇、ESI高被引11篇），《机械工程学报》3篇（1篇年度Top 10高被引），会议论文6篇（3篇Best Paper Award），申请专利3项。研究成果受到了中国工程院邵新宇院士、张来斌院士、王自力院士；美国工程院院士Ali Mosleh；三院院士Miguel A. F. Sanjuan；三院院士Yan Zhang；两院院士Moncef Gabbouj；两院院士Lihui Wang；加拿大工程院院士Ming J. Zuo；IEEE/ASME/SME/CIRP Fellow、凯斯西储大学Robert X. Gao教授；IEEE/ASME/SAE/IMAP Fellow、马里兰大学Michael Pecht教授；故障诊断国际知名专家Andrew Ball教授；《IEEE COMST》（IF=33.840）主编Dusit Niyato教授；褚福磊教授、陈雪峰教授、林京教授、雷亚国教授、高亮教授和黄洪钟教授等10余位杰青团队的引用和正面评价；上述成果强力支撑了项目负责人晋升长聘副教授，入选科睿唯安2022全球高被引科学家、爱思唯尔2021中国高被引学者、湖南省优青，获批国家自然科学基金面上项目1项。