超声波检测中缺陷的智能识别

为了实现超声检测对缺陷的智能识别,引入小波包分析与人工神经网络技术。该方法利用超声信号进行三层小波包分解,提取各频率成分能量为特征值。建立并训练了一种BP缺陷识别的神经网络,该网络使用Levenberg—Marquardt算法。实验分析表明,小波包分析和人工神经网络的引用能为缺陷类型提供有效的智能识别。     

Study on flaw identification of ultrasonic signal for large shafts based on optimal support vector machine

Automatic identification of flaws is very important for ultrasonic nondestructive testing and evaluation of large shaft. A novel automatic defect identification system is presented. Wavelet packet analysis (WPA) was applied to feature extraction of ultrasonic signal, and optimal Support vector machine (SVM) was used to perform the identification task. Meanwhile, comparative study on convergent velocity and classified effect was done among SVM and several improved BP network models. To validate the method, some experiments were performed and the results show that the proposed system has very high identification performance for large shafts and the optimal SVM processes better classification performance and spreading potential than BP manual neural network under small study sample condition.     

Real-Time Diagnostics of Metal-Cutting Machines by Means of Recurrent LSTM Neural Networks

Real-time diagnostics of modules in metal-cutting machines may be based on neural-network algorithms for simulation of the standard process, identification of defects, and the introduction of corrections in the cutting machine’s control system. The machining conditions in normal operation of the machine are recorded by means of a trained neural network with long short-term memory (LSTM network). In real-time operation, the difference between the standard neural-network model and the actual process characteristics is used to determine the type of defect and the module of the machine where it occurs on the basis of a second neural network, the classification unit.     

Identification of bearing faults using time domain zero-crossings

In this paper, zero-crossing characteristic features are employed for early detection and identification of single point bearing defects in rotating machinery. As a result of bearing defects, characteristic defect frequencies appear in the machine vibration signal, normally requiring spectral analysis or envelope analysis to identify the defect type. Zero-crossing features are extracted directly from the time domain vibration signal using only the duration between successive zero-crossing intervals and do not require estimation of the rotational frequency. The features are a time domain representation of the composite vibration signature in the spectral domain. Features are normalized by the length of the observation window and classification is performed using a multilayer feedforward neural network. The model was evaluated on vibration data recorded using an accelerometer mounted on an induction motor housing subjected to a number of single point defects with different severity levels.     

模糊神经网络法及其在缺陷模式识别中的应用

针对冷轧带钢表面缺陷图像模式识别中出现的问题,引入模糊模式识别和反向传播神经网络识别方法.在研究比较两种识别方法的基础上,利用模糊模式识别在剔除噪音数据和反向传播神经网络在模型拟合和非线性识别上的优势,提出一种新的模糊神经网络方法,并详细讨论了算法的结构特点及其实现方法.对五种出现频率较高的典型缺陷图像进行计算机实验研究,结果表明,该方法能对缺陷图像进行有效的识别,具有良好的性能.     

挠性印制电路板焊盘表面缺陷的检测

针对挠性印制电路板(FPC)上的焊盘表面缺陷,提出一种基于图像处理技术的智能检测方法。首先,根据缺陷的表现形式对焊盘缺陷进行归纳与分类,采用最大熵值法量化提取焊盘的颜色特征和面积特征;然后,通过评估灰度共生矩阵(GLCM)对纹理颜色变化特征与纹理结构特征量化的有效性,将其应用于焊盘纹理特征的量化与提取。实验分析显示,缺陷焊盘与非缺陷焊盘在某个或多个特征上存在着明显的差异。基于该特点,建立了BP神经网络,以焊盘的颜色、面积、纹理结构、纹理颜色变化特征值作为神经网络的输入,通过学习大量样本,获取最佳权值参数,最终实现对FPC焊盘表面缺陷的检测,检测准确率高达94.6%,50个焊盘的检测时间为300ms。实验结果表明:本文提出的检测方法不仅能够有效地对焊盘表面缺陷进行识别,而且能够满足在线检测对速度的要求。     

面向石质文物裂隙的超声定位系统及算法研究

面向石质文物多年风化、干裂，急需修复的场景，针对深度学习在文物裂隙检测应用中的不足，提出一种基于缺陷特征增强和卷积神经网络(CNN)的石质文物缺陷定位算法。算法通过小波包分解进行特征提取并选择包含丰富特征信息的有效频段，作为CNN网络的输入，通过模型训练和波形分类识别，缩小定位范围，提高缺陷定位算法的泛化能力和识别率。本文搭建了以现场可编程门阵列(FPGA)为核心的超声检测平台，并在边长为40 cm的立方体试件上进行了实验验证，实验结果表明，波形识别准确率相较传统算法提高了11.3%,平均定位误差小于10%,为石质文物裂隙检测提供了可靠依据，有助于文物保护和修复。     

基于轴箱振动与动力学模型驱动的高速列车车轮失圆状态识别方法

针对高速列车车轮失圆识别难以兼顾效率与精度问题,提出一种基于轴箱振动与动力学模型的高速列车车轮失圆状态智能识别方法。首先,利用静态检测设备采集车轮非圆原始数据,提出一种数据增强技术构建车轮非圆增强数据集。其次,将增强数据集输入至高速列车车辆—轨道耦合动力学模型,获取车轮不同失圆状态下轴箱振动样本集。最后,通过构建恰当结构与配置参数的一维卷积神经网络(1-dimensional convolutional neural network,1-DCNN),可对轴箱振动信号进行自适应特征提取,实现对车轮失圆状态的智能识别分类。结果表明：提出的车轮失圆状态智能识别方法能实现正常车轮、多边形车轮、擦伤车轮、随机非圆化车轮与局部缺陷车轮5类车轮失圆状态的智能分类,准确率达99.2%(标准差为0.05),且单个样本平均识别耗时为0.4 ms。结合现场试验,所提方法对实测轴箱振动具有较好识别能力,测试精度为95%。与经典的SVM和BP神经网络相比,1-DCNN模型具有更高的识别准确度。     

An effective defect identification scheme in pipeline ultrasonic testing

The identification of a defect from its pulse-echo ultrasonic signal is a difficult task but greatly significant for defect evaluation in pipeline ultrasonic testing. In this work, an effective defect identification scheme, which uses the deconvolution technique, is developed to identify the ultrasonic signals from simulated defects with two different shapes (a notch and a transverse cylindrical cavity) on a pipeline specimen. A DSF (Defect Similarity Function) defined as a deconvolution of a target signal by a reference signal is used to classify the defects. Experimental results indicate that the DSFs for the same type of defect-reference signals are narrow impulselike patterns with larger amplitudes, while those for the different types of defect-reference signals are broad impulselike patterns with relatively smaller amplitudes. Furthermore, the DSF is sensitive to defect shapes but non-sensitive to the measuring distance and defect size. Therefore, the defects can be identified from the pattern of the DSF. The proposed technique shows great potential for the identification of defects via ultrasonic signals in the pipeline inspection. This text was submitted by the authors in English.     

Automated defect recognition of C-SAM images in IC packaging using Support Vector Machines

Ultrasonic C-mode scanning acoustic microscopy (C-SAM) is widely used in the semiconductor industry for reliability testing and product inspection due to its ability to nondestructively detect defects in IC packaging. However, image interpretation and defect identification depend largely on the experience of operators, and there is no defect recognition system; this is partly due to current recognition systems, which are based on computer vision algorithms and are not robust for C-SAM images. A new robust defect recognition system and its application to C-SAM images are described in this paper. The iconic domain of two-dimensional C-SAM grey-level image analysis based on the non-linear Mumford-Shah model is used, and defect recognition is achieved through the use of Support Vector Machines (SVMs). The system is verified through experiments on a sequence of C-SAM images corrupted by synthetically generated noise, bias and different shape. The remarkable defect recognition rates achieved indicate that Support Vector Machines (SVMs) are suitable for IC package defect identification.     

集箱管接头内焊缝表面缺陷识别方法研究

针对集箱管接头内焊缝表面缺陷自动检测需要,进行了基于计算机视觉的集箱管接头内焊缝表面缺陷自动识别方法研究。分析了不同焊缝图像的纹理特征,从焊缝图像的灰度共生矩阵中提取出15种可用于焊缝表面缺陷状态表征的特征参数。在此基础上,研究将BP神经网络应用于焊缝表面缺陷自动识别中。分析了灰度共生矩阵的步长、灰度级、神经网络的结构参数及输入特征参数的数量和种类对焊缝图像识别效果的影响,优化出最佳的识别参数。在以上研究基础上,利用优化后的神经网络对内窥镜获得的不同焊接质量的焊缝图像进行了训练和识别。结果表明,提出的基于图像纹理的神经网络识别系统可以很好实现集箱管接头内焊缝焊接状态的自动评价,整体识别率达91%。研究工作为集箱管接头内焊缝焊接质量自动检测做了有益的探索。     

小波神经网络在基桩缺陷诊断分析中的应用

将小波分析作为神经网络的前置处理手段，从基桩动测信号小波变换的分量中提取特征，然后将这些特征输入人工神经网络进行训练和分类，进而实现基桩缺陷位置和程度的诊断。仿真试验的结果表明，该方法对桩身完整性的评价是快速有效的，特别是对于多个缺陷的判别较其他方法具有优越性，在此基础上进行了模型桩的现场试验研究。     

基于改进CV模型和PCNN的NSST域焊接缺陷提取

为了精确地提取焊接缺陷,进一步提高缺陷检测的准确性,提出了一种基于改进ChanVese(CV)模型和脉冲耦合神经网络(pulse coupled neural network,PCNN)的非下采样Shearlet变换(non-subsampled Shearlet transform,NSST)域焊接缺陷提取方法。首先,对焊接缺陷图像进行NSST分解,对得到的低频分量采用PCNN提取出缺陷的主要区域;然后,利用背景抑制后的低频分量和高频分量构造出高频特征图像,并对其进行粗分割,再利用改进的CV模型寻找最优轮廓,提取出缺陷精细轮廓;最后,融合缺陷的主要区域和精细轮廓信息得到最终的结果。实验结果表明,与其他缺陷提取法相比,所用方法提取的缺陷结构更为完整,缺陷轮廓更为精细。     

基于三级级联架构的接触网定位管开口销缺陷检测

针对高铁接触网定位管开口销在列车长期运行振动中容易松脱并且松脱样本数量匮乏的问题,本文提出一种基于深度卷积生成对抗网络(DCGAN),扩充缺陷样本集后,再训练卷积神经网络(CNN)检测开口销缺陷的三级级联架构。该架构首先采用中心点法提取训练需要的相同规格开口销图像。然后通过改进的DCGAN生成模拟缺陷样本,并搭建轻量级CNN网络对生成的模拟缺陷样本进行筛选。最后将添加了模拟缺陷样本的扩充缺陷样本集与正样本集输入优化后的VGG16卷积神经网络中,以训练分类模型,检测开口销缺陷。实验结果表明,本文所提方法检测接触网定位管开口销缺陷的准确率高达99%。     

基于径向基函数神经网络的发动机磨损预测分析

针对BP神经网络算法的不足,利用径向基函数(RBF)神经网络建立设备的磨损预测模型,对光谱分析数据进行实例仿真,并与BP网络模型进行对比研究.仿真结果表明,该模型预测精度高,训练时间短,大大优于BP神经网络模型.     

稀疏重构SAM芯片焊点检测方法研究

提出了基于稀疏表示的声扫描显微镜(Scanningacousticmicroscope,SAM)图像超分辨率重构方法,以解决其空间检测分辨率受超声波频率和穿透深度的限制,原始SAM图像分辨率较低,不利于封装缺陷辨识等问题。通过字典设计训练和稀疏系数α求解获得了重构的高分辨率SAM图像,利用Levenberg-Marquardt算法改进BP神经网络(LM-BP),并用于倒装芯片焊点缺陷识别。与原始图像及双三次插值图像相比,稀疏重构图像的峰值信噪比明显增大,提高了SAM图像质量,减小了芯片焊点的错误识别数目,错误率降至2.76%。试验结果表明稀疏表示的SAM重构算法和LM-BP神经网络训练速度快、识别精度高,可用于高密度半导体封装缺陷的检测及可靠性评估。     

一种基于神经网络的双足机器人混合动力学系统辨识方法

针对双足机器人的混合动力学系统辨识问题,从系统渐进稳定性角度分析,推导出连续与离散混合系统的可辨识条件,提出了一种基于混沌粒子群优化的径向基函数神经网络与动态模糊神经网络的联合辨识方法。利用混沌粒子群优化的径向基函数神经网络辨识双腿的连续摆动阶段,利用动态模糊神经网络辨识离散的足地碰撞阶段;依据两阶段同一变量的耦合、转换关系,实现了对双足机器人整体混合系统的准确辨识。仿真实验结果表明,该方法辨识和预测结果具有较高的准确度。     

基于ECT投影数据的两相流流型辨识

提出一种基于电容层析成像（ECT）系统的两相流流型辨识新方法。ECT投影数据经特征参数提取前处理环节后馈人事先训练好的神经网络实现流型辩识。文中介绍了特征参数的提取、神经网络的构成以及训练方法。研究结果表明，该方法辨识精度高、辨识速度快，是两相流流型在线辨识的一种有效手段。     

压电工作台的神经网络建模与控制

建立了压电工作台的神经网络在线辨识模型并设计了相应的自适应控制器以抑制压电工作台迟滞特性、蠕变特性及动态特性对其微定位精度的影响.采用双Sigmoid激活函数对神经网络激活函数进行了改进,同时分析了改进激活函数的神经网络模型与PI迟滞模型在迟滞建模上的异同.设计了基于改进激活函数的3层BP神经网络作为压电工作台的在线辨识模型,推导了网络权值、阈值及激活函数阈值修正公式.最后,基于神经网络模型设计了压电工作台的自适应控制方案,该控制方案利用另外一个神经网络来完成对PID控制器参数的自适应调整.实验结果表明:提出的神经网络在线辨识模型平均误差为0.095 μm,最大误差为0.32 μm;自适应控制方案跟踪三角波的平均误差为0.070 μm,最大误差为0.100 μm;跟踪复频波的平均误差为0.80 μm,最大误差为0.105 μm.实验数据显示压电工作台的定位精度得到了有效提高.     

基于改进卷积神经网络的燃气调压器故障识别研究

针对传统故障识别方法不仅过分依赖专家经验对故障特征进行提取且识别准确率不高的问题,在深度学习理论基础上,提出了一种将一维卷积神经网络与SVM分类器相结合的改进深度卷积神经网络,实现调压器“端到端”的故障识别。首先,介绍了传统卷积神经网络结构;其次,将改进后的一维卷积神经网络与SVM相结合,提出了基于1-MsCNN-SVM算法的调压器故障识别模型,并对模型的组成部分进行了介绍;然后,通过对比实验确定了模型的卷积核长度和卷积层组数;最后,为验证模型的有效性,基于燃气调压器故障数据集,开展了燃气调压器故障识别研究。研究结果表明,改进后的1-MsCNN-SVM算法故障识别准确率高达99.20%,模型具有较好的分类准确率。