基于机械声场时空全息诊断模型的弱故障特征提取研究

The acoustic-based diagnosis (ABD) technique is an important developing tendency of fault diagnosis technology. Sound field is a three-dimensional space which contains the temporal-spatial attributes. The conventional ABD technique uses the information at one point in the sound field for fault diagnosis, and the new developed acoustic image based fault diagnosis uses the local characteristics on the source surface for fault diagnosis, by which the sound field characteristics are not fully utilized and poor robustness are obtained in some local weak fault conditions. In order to improve the problems, the sound field containing temporal-spatial attributes is modeled and its high dimensional characteristics are applied to diagnostic analysis directly. The array measurement and the near field acoustic holography (NAH) techniques are employed in the sound field. The phase information at the source surface is mapped to time-space domain, and a series of dynamic time-space sequence image models with different temporal-spatial resolutions are obtained. The optimum temporal-spatial resolution is selected by considering the information redundancy index of the sequence image, such as information content, correlation and information entropy. And then the high dimensional sound field diagnosis model containing the temporal-spatial correlation information is constructed. The feature processing techniques in video and spatial fields are adopted to extract the temporal-spatial characteristics of the sound field diagnosis model, such as principal components analysis (PCA), spatial spectrum, temporal-spatial slice manifold and space tensor. Since the changes of sound pressure distribution and temporal-spatial correlation information are considered altogether, the diagnosis robustness is improved. And due to the application of temporal-spatial characteristics, the acoustic diagnosis procedure is replaced by space target recognition, which expands the application scope of acoustic imaging technique and provides a new idea and option for ABD.

噪声诊断由于其非接触测量的优点，成为故障诊断技术重要分支之一。然而，目前基于单点测试的常规声诊断技术和基于阵列测试的声像诊断技术都仅以声场某个点和某个面为研究对象，对于一些机械局部弱故障工况存在信号信噪比和诊断鲁棒性差等问题。针对这些问题，本项目通过构建声场全息诊断模型和提取声场动态时空特征，实现机械微弱故障的有效诊断。首先通过对机械声场采用阵列测量和近场声全息技术，把源像相位信息映射到时空域，获得时空序列声像样本空间，并综合考虑序列声像的信息量、相关性和信息熵等信息冗余指标，选定最优时空分辨率，构建蕴含时空关联信息的声场全息诊断模型；然后基于声场模型的时空属性，引入视频和地质空间领域数据处理分析技术，如主成分、空间谱、时空切片流形和空间张量分析等，提取蕴含时空关联信息的声场弱故障特征进行诊断。本项目直接以高维全息声场信息进行诊断，融合多维度特征提升诊断鲁棒性，为声诊断技术提供新的思路。

噪声诊断由于其非接触测量的优点，成为故障诊断技术重要分支之一。但声信号抗干扰能力较差，高信噪比测点位置的选择严重影响常规声诊断技术的诊断鲁棒性。基于阵列测试的近场声全息声像诊断技术以2维声场为研究对象，以面代点，规避了测点选择，但多源信息的融合也带来了信息冗余度大和声场特征利用不足等问题。本课题以机械声场为研究对象，分别对2维声场特征的冗余特性、2.5维声场和3维声场特征提取方法及相应的诊断机理进行了研究。在2维声场特征提取环节引入了灰色关联聚类技术，通过融合特征序列中关系紧密的特征，构建低冗余性的有效声场特征模型，进一步提升了诊断鲁棒性。基于近场声全息阵列测量理论和信息融合思想，提出了一种基于双面声像模式识别的2.5维声场故障诊断方法。通过融合全息测量面声像、近场重构源像和两者差值声像构建含有部分法向声场变化信息的2.5维声场模型，然后提取Gabor小波纹理特征，并基于随机森林特征选择算法进行特征降维，构建有效声场特征模型进行状态诊断识别。研究结果表明，融合法向变化信息的2.5维声场故障诊断方法比单声像的2维声场诊断方法具有更好的诊断精度。进一步挖掘近场声全息逆向和正向声场重建技术的应用价值，结合2维声场近场声全息诊断技术相位信息利用不充分的情况，提出了一种基于三维声场物理空间特征的故障诊断方法。把二维声场中的相位信息映射到三维空间中，利用近场声全息技术重构机械三维声场，然后在一个波长范围内序列拾取13个辐射声场空间切片，对每个切片面提取Gabor小波纹理特征，并顺序排列各切片的小波特征，构建含有法向空间变化的三维声场空间特征模型进行故障诊断。研究结果表明，3维声场包含了更为完善的机械状态信息，在机械弱故障工况下具有更好的诊断鲁棒性。本课题对基于阵列测试的近场声全息诊断技术进行了补充和完善研究，进一步挖掘了近场声全息声场重构技术的应用价值，也为基于阵列测试的声诊断技术提供了新思路。