【作者】 李晶；

【导师】 赵力；

【作者基本信息】 东南大学 ， 信号与信息处理， 2017， 博士

【摘要】 声发射(Acoustic Emission,AE)信号处理是声发射技术的重点研究内容,也是无损检测评估中的关键环节。AE信号对结构内部损伤程度、类型和位置的识别是AE检测技术研究的核心。AE信号按照其产生的物理本质分为典型和二次型,传统的信号处理方法对此并没有区分。分别探讨这两种类型AE信号在产生机理、传播特性和波形特征等方面的共性和特性,研究相应的AE信号处理和分析方法,也是急需解决的问题。本文从工程应用的角度,分别研究了两种AE信号的产生机理和处理方法,一方面,针对二次型AE信号,以旋转机械碰摩AE信号为研究对象,深入研究了 AE信号的预处理、定位和识别的新方法;另一方面,针对典型AE信号,以煤岩样冲击破裂的AE信号为研究对象,提出了对AE波形的特征提取方法和对煤岩体冲击破裂状态预测的新方法。本文主要内容包括:针对金属自由板内AE信号的传播特性进行了深入的理论分析和实验验证,提出了一种对AE传感器信号的多模抑制和频散补偿的预处理方法。实验结果表明,该方法能有效分离出AE信号中的主要成份A0和S0,滤除高阶模态波的干扰,减弱AE监测信号中波包扩展和变形程度,以及合成与频散补偿出单一模态波,进而为AE源类型分析提供依据。针对旋转机械碰摩AE信号的宽带、多模态和频散特性,引入了基于频率聚焦理论的近场多重AE源定位算法,并推导了频率聚焦矩阵的计算公式。针对算法对聚焦频率和声源位置初值的依赖性,进一步对算法进行完善,提出了基于频率自动聚焦的近场AE源定位方法。研究结果表明,两种方法均具有较好的分辨相干信号的能力,后者对AE源的定位精度较高、计算量小,可为碰摩初期故障的检测提供有效依据。结合AE信号能量在空间分布上具有稀疏性的特点,引入稀疏分解理论,提出了时域多快拍近场AE源定位算法,构建了特征子带阵列信号在空间上稀疏分解的凸优化模型,得到AE信号在整个空域中能量分布的稀疏系数。针对近场双AE源定位中较远AE源的定位精度不理想且计算复杂度高等问题,从子带分解和粗-细网格的优化搜索策略角度出发,提出了基于频率多快拍的近场AE源稀疏分解定位算法。实验结果表明,改进的方法定位精度高,计算量小,实用性强,具有很好的解相干能力,可以有效地应用于碰摩AE信号的定位检测中。基于深度学习框架下的卷积神经网络,提出了一种转子碰摩故障AE信号识别新方法。对频散补偿后的AE信号提取语谱图特征,从时间、频率和能量角度构建不同工况下AE信号的状态参数,利用卷积神经网络对碰摩故障进行识别。该方法直接对AE信号的语谱图特征学习和识别,避免了人为选择某些局部特征而造成的信息丢失,能更全面的描述AE信号的碰摩特征。实验结果表明,该方法具有较好的碰摩故障识别性能。在AE信号预测煤岩冲击破裂研究中,首先,分析了煤岩内部断裂发出的AE信号的产生机理,从无标度区域的优化搜索方面改进了 GP关联维数特征。实验结果表明,改进的GP关联维数对岩样从稳定期到破裂阶段的演变更加敏感,在噪声环境下具有较强的鲁棒性,是用于岩石破裂状态识别的有效特征。在此基础上,针对AE信号的高频和低频特征对岩石破裂过程的不同反应,提出了多分辨率特征融合的岩石失稳破坏预报方法。从识别结果可以看出,该方法对煤岩所受应力状态的识别性能较好,且对岩样危险状态预测具有较好的灵敏度,为煤岩体危险状态的识别和分析提供了一条新的途径,类似识别模型在AE预测煤岩体冲击破裂领域尚无先例。更多还原

【Abstract】 Acoustic emission(AE)signal processing is the key research area in acoustic emission technology,which is also the important step in nondestructive testing and evaluation application.Using AE signal to detect structural damage degree,damage type and damage location is the core topics in AE detection technology.According to the physical nature of different sources,AE signal should be divided into two types:typical and quadratic.It is of great theoretical and practical value to study the generation mechanism,propagation characteristics,waveform characteristics and processing and analysis methods in different kinds of AE sources.In the thesis,we discuss the generation mechanism and processing methods involving these two kinds AE sources from the perspective of engineering application.One is the quadratic AE source in rub-impact fault of rotating machinery to be processed for recovery,localization and recognition.Another is the typical AE source in coal and rock mass fracture to be studied for feature parameters extraction and bursting condition prediction.The main contents of this paper include as follows:Based on theoretical analysis and experimental tests of AE propagation characteristics in metal free plate,a new preprocessing method to filter high order modal waves and compensate frequency dispersion is put forward.Simulation results demonstrate that it is an effective method to recover all frequency information of AE source by extracting dominant components A0 and S0 in the modal cluster.Besides,every modal wave is also be extracted by propagation forward calculation.Then,the processed AE signal served as the important evidence is further uesed for fault analysis.For the wideband,multimodal and dispersion characteristics of AE signals,two new near-field multiple AE source localization algorithms using frequency focusing theory are proposed for the first time.The first near-field coherent Signal-Subspace Method(N-CSM)is proposed,which is demonstrated that results are extremely depended on the focusing frequency and the initial position.According to this point,the second algorithm(N-AFCSM)improves frequency focusing method in auto way.Experimental results show that these two methods both have well ability to distinguish coherent signals.The latter method has higher precision and lower computational complexity for AE sources localization,which provides an effective evidence for the early fault detection.For AE signal energy distributed sparsely in the space,the sparse decomposition theory is creatively introduced to locate near-field AE sources.Considering multiply snapshots of AE signal in time domain,a convex optimization model using received AE sub-band array signal is constructed for AE energy sparse coefficient in the entire space.This method is demonstrated that the localization accuracy of the ulterior AE source is still unsatisfactory with low computational efficiency.Then,the improved algorithm using multiply AE snapshots in frequency domain is proposed based on frequency sub-band decomposition method and the coarse-fine grid optimal search strategy.The localization results present that the improved method has higher accuracy,lower computational complexity,preferable practicability and better decorrelation ability,which is an effective way to detect rub-impact AE sources for practical engineering application.Based on the deep learning frame,a new recognition algorithm is proposed for rotor rub-impact fault recognition by using frequency compensated AE signal spectrum features and deep Convolutional Neural Network(CNN).The proposed AE signal spectrum features describe AE waveform variation from several three perspectives including:time,frequency and energy.Compared with traditional AE waveform features,it provides complete AE information source in rub-impact fault and avoids information lossing by artificial selection.Combined with CNN model,the experimental results show that the proposed method achieves an approving performance in rotor rub-impact fault recognition.For the other kind AE source the thesis applies AE signal combined with fracture mechanics to predict coal rock mass bursting condition.Firstly,the designed algorithm improves the search strategy of GP correlation dimension for the optimal scaleless region localization in auto way.Results achieve better sensibility to the evolution of rock mass from stable to burst and preferable reliability in noise environment.On this basis,a new coal rock burst prediction method using multi-resolution AE waveform feature fusion is proposed The recognition results present that it has excellent performance to recognize the coal rock current conditions and is sensitive to coal rock crisis condition.Its good performance provides a new way to identify and analyze the stress conditions of coal rock mass.There is no precedent in similar recognition models for coal rock burst prediction using AE technology.更多还原