基于非凸正则化稀疏表示的轮对轴承服役性能评估方法研究

Service performance is an important indicator for the predictive maintenance of wheelset bearing and the operation safety of vehicle. To realize the service performance evaluation of wheelset bearing, the weak feature extraction and accurate amplitude estimation is required. For such a purpose, the non-convex regularized sparse representation is firstly proposed in this project. Then, the performance evaluation method of wheelset bearing in full-life cycle is built by feeding the extracted features into the deep learning models. Specifically, the main research contents consist of: (1) the construction of the parameterized non-convex regularized function; (2) the design of characteristic dictionary, the model establishment of the non-convex regularized sparse representation and the development of algorithms for these established models; (3) the feature extraction of three stages, i.e., the performance degradation stage, initial failure stage and residual life evolution stage of wheelset bearings, using non-convex regularized sparse representation for the three stages in full-life cycle; and (4) the construction of the performance evaluation model of wheelset bearing for the each stage based on a deep contractive auto-encoder and the integration of the three models. To conclude, the non-convex regularized sparse representation and the comprehensive evaluation models of service performance of wheelset bearing are of great significance for the predictive maintenance of wheelset bearing and the safety of vehicle operation.

轮对轴承服役性能是其预知维修的重要参考和车辆运行安全的重要保障。考虑轮对轴承服役性能评估对强噪声背景下弱特征提取及特征成分幅值保真性的要求，提出非凸正则化稀疏表示的轮对轴承服役性能特征提取的方法，并以得到的特征表征指标为深度学习评估模型的输入实现轮对轴承全寿命周期服役性能评估。具体内容：①参数化非凸正则函数构建理论与方法的研究；②轮对轴承服役性能特征成分字典的构建，稀疏表示模型的建立及非凸正则化稀疏表示求解算法的研究；③基于非凸正则化稀疏表示提取特征成分的轮对轴承性能衰退期、初发故障期和剩余寿命演化期服役性能特征指标表征研究；④基于深度学习的压缩自动编码器，分别建立轮对轴承性能衰退评估模型、初发故障检测与诊断模型和剩余寿命评估模型，并研究模型的集成。本项目通过非凸正则化稀疏表示方法的研究，确立轮对轴承服役性能评估的系统方法，对轮对轴承的预知维修和车辆运行安全保障具有重要意义。

轮对轴承服役性能是其预知维修的重要参考和车辆运行安全的重要保障。因此，本项目以轮对轴承为对象开展服役性能研究具有显著的代表性和重要的现实意义。然而，现有轮对轴承服役研究方法普遍存在如下两方面的问题：①在强噪声背景下，轮对轴承服役性能评估特征提取能力弱；②轴承服役状态评估方法对于轮对轴承初发故障的检测能力弱，对于故障类型的精准识别能力差，而导致服役性能评估方法对于轴承性能退化趋势表征能力弱。为解决上述问题，本项目建立了轮对轴承状态特征稀疏表示提取方法与深度学习方法相结合的服役性能评估框架。首先，提出了基于稀疏表示的轴承状态特征提取方法和轴承状态信号特征自适应提取方法，有效地提取轴承状态特征；其次，利用深度神经网络和对抗网络构建故障检测模型，精准地判别轴承故障发生时间；进一步，设计基于深度神经网络的故障诊断方法，准确地识别出轴承故障的类型；最后，在状态特征提取和故障检测的基础上，设计轴承性能退化健康指标，构建轴承性能退化预测模型，实现轮对轴承服役性能退化的监测。