稀疏深度学习与直升机传动系统故障诊断研究

Transmission system is a life-limited component of the helicopter, and fault diagnosis to transmission system is critical to safety of the helicopter. In this proposal, a sparse deep learning diagnosis method is proposed based on sparse representation and deep learning, which are the frontier of machine learning.. First, sparse separation method to multi-source signals is researched to develop a signal separation model in underdetermined condition. Then, sparse deep learning theory is explored and sparse deep learning model for fault diagnosis is developed based on the sparse characteristic of mechanical signal, by which classification accuracy of the fault would be improved. After that, multi-mode deep learning model is researched and matching mechanism between deep learning and sensor data is illustrated. By using the multi-mode deep learning, a condition index fusing multi-sensors data is constructed to indicate degree of the fault. By the above research aspects, the sparse deep learning model is finally proposed for fault diagnosis of transmission system of helicopter. This model is also validated by test data and real engineering application.. The innovation in sparse deep learning, fault classification and quantitative diagnosis of helicopter transmission system would be realized by the research conducted in this proposal. The research work in this proposal will also provide new theory and method for fault diagnosis of helicopter which has attracted more and more attention in aviation.

传动系统是直升机的限寿关键件，其故障诊断对保障直升机飞行安全具有重要意义。本项目以稀疏表示和深度学习两项机器学习前沿理论为基础，探索稀疏深度学习理论进而开展直升机传动系统故障诊断研究。. 研究传动系统多源耦合信号稀疏分离方法，建立欠定条件下信号分离模型，从根源上克服信号混叠对诊断精度的限制；探索稀疏深度学习理论，构造稀疏深度学习模型，根据物理对象实现具有针对性的深度网络构造，提高故障分类精度；研究多模式深度网络融合方法，阐明状态信息与深度网络的匹配机制，建立融合多模式信息的故障定量诊断指标。在以上研究基础上，提出基于稀疏深度学习的直升机传动系统故障诊断方法，并开展实验验证与工程应用研究。. 本项目研究有望在稀疏深度学习理论、直升机传动系统故障分类与定量诊断方面有所创新和突破，为航空领域中受到广泛关注的直升机运行安全保障提供新理论和新方法。

针对直升机传动系统故障诊断问题，本项目提出多成份混叠信号提取与故障模式稀疏聚类方法，提升了直升机传动系统故障特征的聚集性与故障模式的分离度。提出稀疏深度堆叠网络构造方法，发展了基于故障特征的深度网络构造方法；构建了基于稀疏特性迁移的深度网络模型，实现了直升机传动系统故障模式诊断。提出基于多模式信号特性的深度网络构造方法，构建了协同深度置信网络模型和多感受野图卷积网络模型，实现直升机传动系统多源监测信息的融合，有效提升了诊断精度和鲁棒性。开展了实验研究和工程数据验证，为直升机传动系统传感监测位置优化、工作状态识别和性能退化评估提供了有效方法。. 基于本项目相关研究成果，发表学术论文18篇；申请发明专利 3项。研究成果发表在IEEE Transactions on Industrial Electronics等高水平期刊，相关成果被IEEE神经网络与学习系统汇刊主编等著名专家引用评价。核稀疏局部线性嵌入研究成果被国际著名工程新闻网站Advances in Engineering遴选为有显著贡献的科学论文，并予以高亮专题报道。混合膨胀卷积齿轮故障诊断会议论文获得获2018年ASME 柔性自动化会议的最佳论文提名，相关研究获2019年CIRP 全生命周期工程会议最佳论文提名。直升机传动系统工作状态识别与性能监测相关算法已集成到某型直升机HUMS系统并装机应用；以本项目智能诊断相关算法为支撑，获批陕西省技术转移与科技成果推广引导计划项目一项。