基于多传感器时序信息深度建模的飞航发动机故障预测方法研究

Characterization and analysis of implicit information from multi-sensor time series are one of the core challenges which limit complex aero-engine fault prediction. Through the combination of the time-series theoretical feature measurement ability and the complex conversion ability of the long-term learning data of the memory network, the implied feature extraction can be performed better to significantly improve fault prediction accuracy and expand the prediction range. Therefore, taking the sensor data of aero-engine test as the main research object and aiming at the typical feature prediction problem of Markoff sequence group with rich connotation, this project innovatively proposes a deep memory network prediction method for time series analysis and traction. Through the multi-parameter fusion analysis of the potential correlations among time-frequency, granularity, and scale of sensor data, combined with the process of dissimilar sensor data representation unification, a deep memory network model with hierarchical attention and storage mechanisms based on sequence features was constructed for unknown sequence prediction research. At the same time, the error criteria and training boundaries based on a measurement for the similarity of the sequence are used to achieve efficient model training. Finally, a set of time series depth prediction methods that satisfies the feature of the field, with clear relationship and easy to implement will be established to solve the problems of the fault prediction of the aero-engine from a new perspective and conduct application verification in typical rocket engine flight tests and aeroengine trial tests.

多传感器时间序列隐含信息表征与分析是制约飞航发动机复杂故障预测的核心挑战之一，而通过时间序列理论特征度量能力与记忆网络长时间范围学习数据复杂转换能力的结合可以更好地完成隐含特征提取从而显著提高故障预测精度、扩展预测范围。因此，本课题以飞航发动机试验传感数据为主要研究对象，针对丰富内涵变序马尔科夫序列组上典型特征预测问题，创新性地提出了时间序列分析牵引的深度记忆网络预测方法，通过多参融合量化分析传感数据时频、粒度、尺度间的潜在关联关系，结合相异传感器数据表征一致化过程，构造基于序列特征引导具有分层注意力和存储机制的深度记忆网络模型，开展未知序列预测研究。同时以序列相似性度量为依据的误差评判准则与训练边界，实现高效的模型训练，最终建立一套满足领域特征、关系清晰、便于实施的时间序列深度预测方法，从全新角度解决飞航发动机故障预测难题，并在典型火箭发动机飞行试验和航空发动机试车试验中开展应用验证。

项目以飞航发动机试验传感数据为主要研究对象，针对丰富内涵变序马尔科夫序列组上典型特征预测问题，创新性地提出了时间序列分析牵引的深度记忆网络预测方法，通过多参融合量化分析传感数据时频、粒度、尺度间的潜在关联关系，采用时间弯曲等手段研究数据时间粒度的对齐方法，以及针对不同的传感数据的不同尺度，结合相异传感器数据表征一致化过程；开展面向时序信息的深度记忆网络建模研究，重点研究网络的输入输出和预测机制；参考不同序列特征对预测贡献度的不同，研究如何定义注意力权重；并从序列整体到局部建立多层次的注意力体制；此外，针对保护长程物理量带来的变序马尔科夫问题，研究深度记忆网络模型中的存储机制，进而开展未知序列预测研究。基于序列相似度度量方法，研究深度记忆网络模型训练的误差准则；研究该准则下的训练边界条件设置问题；进而利用时序信息整体特征、分段聚合特征，通过相关分析方法，研究深度记忆网络参数的初始化手段；并结合网络自身特点，研究模型训练时参数的选择性更新和层间连接简化策略，进而实现高效的模型训练。最终建立一套满足领域特征、关系清晰、便于实施的时间序列深度预测方法，从全新角度解决飞航发动机故障预测难题，并在典型火箭发动机飞行试验和飞机飞行姿态试验中开展应用验证。