基于卷积球化的齿轮箱早期非平稳信号时变盲反褶积

Gearbox which is liable to break down plays an important role in machine. It is very important to identify gearbox faults. Blind deconvolution has enormous potential for fault diagnosis of gearbox. In this project, two main challenging problems: strong convolution mixture and time-variant channel are considered according to the linear convolution mixture model. In traditional blind deconvolution, there are two main algorithms: a single time-domain algorithm and a single frequency-domain algorithms, which are focussed on the time-domain model and frequency-domain model respectively. Firstly, In order to overcome the shortages of the traditional blind deconvolution algorithms, we establish a simplified and unified block model by using convolutive sphering. Then, fruit fly optimization algorithm (FOA) is applied in estimating the block model structure parameters, and manifold learning method is used in reducing the dimension of the model algorithm. Thirdly, dynamic clustering algorithms are used in reconstructing the separated signals in the solution space in order to obtain results of high precision. On the other hand, reference signals and FOA are applied in extracting a small number of critical signals in convolution mixture. In this way, the semi-blind extraction approach can overcome the disadvantages of traditional simultaneous deconvolution algorithms, which are lack of processing the strong convolution and time-variant mixture. Finally, we can design a simple and dynamical solution and unified blind deconvolution algorithms which are adjustable and suitable for characteristics of initial non-stationary signals of gearbox and engineering application, and then evaluate the performance of all the algorithms.

齿轮箱是机器的重要部位，也是易发生故障的部位，对齿轮箱开展故障诊断有重要意义，应用盲反褶积技术研究齿轮箱故障诊断蕴藏着巨大潜力。本课题在线性卷积混合基础上，针对强卷积混合、传递通道时变两个难点，对齿轮箱早期非平稳信号识别问题展开。针对单一时域模型算法或单一频域模型算法的缺点，拟通过卷积球化统一时域和频域模型，并利用果蝇优化算法准确估计"块"结构参数，建立适合于处理强卷积和时变的块模型；针对块模型数据大和耦合问题，应用流形学习约简模型数据，采用动态聚类方法研究振源贡献率，重构解空间，以提高非平稳信号盲反褶积精度；另一方面，利用参考信号和果蝇优化算法提取卷积混合中的少量关键非平稳信号，以半盲提取和果蝇优化算法的优势弥补传统处理大量源的不足。最后，设计一种可动态调节且参数设置简单的，适合齿轮箱早期非平稳信号识别技术特点和工程应用背景的盲反褶积机制及其实现算法，并对算法的性能进行分析与评价。

齿轮箱是机器的重要部位，也是易发生故障的部位，对齿轮箱开展故障诊断有重要意义。在线性卷积混合基础上，针对强卷积混合和传递通道时变两个难点，对齿轮箱早期非平稳信号识别问题展开，提出了一个基于盲反褶积的齿轮箱故障诊断统一框架和方法。针对单独的时域或频域算法的缺点，通过卷积球化方法统一了时频域模型，并利用果蝇优化算法准确估计时延结构参数，建立适合于处理强卷积和时变的统一模型；针对块模型数据大和耦合问题，应用流形学习约简模型数据，采用动态聚类方法研究振源贡献率，重构解空间，提高了非平稳信号盲反褶积的计算精度；另外，设计了精简的参考信号，并利用参考信号和基于克拉美罗界的稳健独立分量分析提取卷积混合中的少量关键非平稳信号，以半盲提取的优势弥补传统处理大量信号源的不足。最后，设计了一种可动态调节且参数设置简单的，适合齿轮箱早期非平稳信号识别技术特点和工程应用背景的盲反褶积机制及其实现算法，并对算法的性能进行分析与评价。