基于载荷反演的齿轮箱工况不敏感状态退化评估理论与技术

Condition-based maintenance of gearboxes depends on the performance degradation assessment which is able to provide a condition indicator to quantify the health of gearboxes. Most of the condition indicators available work well under stationary operating regime. Unfortunately, varying operating conditions are very common in the field. Varying load and speed will result in amplitude modulation and frequency modulation which are similar to that of gear faults. Hence, varying conditions will confuse gear fault detection and deteriorate performance degradation assessment. The objective of the present project is to address the performance degradation assessment of general gearboxes working under varying conditions. The relationship between vibrations and varying load and/or speed will be investigated by developing dynamic models of gearboxes with parallel shafts. By an intensive research on dynamic simulation and data processing methods, algorithms to recover load profile from vibration signals are to be developed and validated on simulated and experimental data. Once the load influence has been removed from vibrations, features insensitive to load variations are possible to be extracted. There are numerous feature extraction paradigms for gear diagnosis and health evaluation. It seems necessary to select the most informative and suitable features by some criterions as a subject of the project. A fusion model will be studied to include multiple selected features to give a consistent, monotonous and robust indicator for gearbox health evaluation. All those proposed methods will be validated by experimental and field data. The outcomes of the project are expected to improve the condition-based maintenance of gearboxes especially those working under varying conditions.

齿轮箱性能状态退化评估是实现其智能维护的基础，但目前研究大多集中在平稳工况下进行。工业现场广泛存在的变速变载工况将导致齿轮振动信号出现与故障特征类似的调频和调幅现象，此时若忽略工况变化带来的影响必将引起误诊断并降低状态退化评估的准确性。为此，本项目以通用齿轮箱为主要研究对象，以工况变化、振动响应和载荷反演为主线，以提高变工况下齿轮箱性能状态退化评估的准确性为目标，以变工况齿轮箱动力学仿真和振动/工况响应机制研究为切入点，深入开展载荷反演算法研究，力求从振动信号中剔除工况变化的影响。在此基础上研究面向齿轮箱状态退化评估的特征提取和特征评价方法并建立候选特征库，进而建立基于多退化特征融合的性能退化评估模型，并通过试验台和工程实际信号予以验证，最终提出一套完整的对工况变化不敏感的齿轮箱特征提取和状态退化评估原理和算法，为齿轮箱健康管理和维护水平的提高提供理论基础和技术支撑。

本项目针对齿轮箱性能退化评估进行研究。研究了变工况下齿轮箱振动响应，搭建了齿轮系统动力学仿真模型，利用此模型对齿轮接触瞬态动力学及刚柔耦合动力学进行仿真，对齿轮疲劳寿命进行分析，同时以动力学分析结果为参考对传动齿轮结构进行优化，实现了齿轮副轻量化设计。. 研究了信号采样期间载荷保持稳定但不同采样期间的载荷可能变化的变载荷非平稳问题。提出了基于AR-ARX-AANN的齿轮箱载荷反演技术，消除载荷变化对性能退化评估的影响。为解决齿轮箱早期故障阶段冲击特征微弱且存在信号传输路径、强背景噪声以及高幅值偶然性冲击的干扰，导致故障特征成分难以提取以及大多齿轮箱故障复合诊断方法各处理步骤采用的诊断优化指标不一致等问题，本项目提出多种齿轮箱故障特征自适应增强的联合降噪方法，如相关峭度联合降噪方法、最大相关峭度解卷积联合降噪方法等。. 本项目提出了两种面向状态退化评估的新指标：自适应频带冲击强度和自适应频带熵能比。二者都具有与故障程度一致性好、随故障变化敏感性高等优点，且自适应频带熵能比指标具有较好的工况鲁棒性，全寿命疲劳实验数据验证了二者的有效性和优越性。提出了一种基于相关性、单调性和鲁棒性的特征优选方法，并以实验数据验证了该特征优选方法的有效性和实用性。. 状态退化评估模型可系统性地分为空间距离型、概率相似型和残差重构型三大类模型。本项目提出一种基于一体优化深度自编码高斯混合模型的（概率相似型）状态退化评估新模型，以改进传统基于概率估计的性能退化评估模型易受到解耦模型学习和不一致优化目标的影响，且不能在低维空间中保留必要的信息，从而导致模型次优性能的缺点。提出一种将自编码器和字典学习相融合的AEDL双重构型神经网络（残差重构型），相较于传统单重构模型，其能够更好地识别齿轮箱状态退化时所发生的早期故障。