基于改进神经网络的声发射信号识别算法研究

发布时间：2019-01-04 10:33

【摘要】：声发射检测技术是一种通过瞬态弹性波记录材料内部振动变化的无损检测技术,具有实时性、适用性、广泛性等特点,适合材料形变、管道泄漏检测、容器压力检验等多方面的研究。声发射信号的模式识别是检测中的重要环节,只有对故障模式进行及时准确的判别,才能有效排查故障,避免监测失误带来的材料损耗和安全隐患。目前,基于神经网络的声发射识别技术存在特征信息含量少、网络易陷入局部最优等缺点。本文在现有技术的基础上,在以下几个方面开展了研究工作：第一,对声发射检测技术进行了详细的研究,对声发射(Acoustic Emission, AE)信号的处理方法展开讨论,并通过实验比较了不同碰摩程度下AE信号波形的区别。第二,对声发射信号的特征提取进行了研究,提出将Hurst指数和近似熵加入AE特征表示中,从统计相关性角度和不确定性角度进行了分析,通过实验证明了新型AE特征的有效性。第三,对神经网络算法结构进行了研究,将高斯混合模型(Gaussian Mixture Model, GMM)与反向传播(BP, Back Propagating)的优势相结合交替训练模型参数,提出将混合模型GMM/ANN(Artificial Neural Network, ANN)网络应用于AE识别中,达到优化网络性能的效果。第四,提出了用前向混沌神经网络识别声发射的算法研究,针对AE系统的不确定性、非线性性质,利用Logistic映射单元增强网络的混沌特性,改善了系统的识别性能。第五,提出了用深信度网络识别声发射的算法研究,在受限玻尔兹曼机(Restricted Boltzmann Machines, RBM)模型基础上构建深信度网络(Deep Belief Network, DBN),设计和改进模型参数,减弱局部最优效应,通过与BP进行比较证明了DBN在AE识别中的优越性。
[Abstract]:Acoustic emission (AE) is a kind of nondestructive testing technology which records the internal vibration of material by transient elastic wave. It has the characteristics of real time, applicability and universality, and is suitable for material deformation and pipeline leakage detection. Vessel pressure inspection and other aspects of research. Pattern recognition of acoustic emission signals is an important part of detection. Only by timely and accurate identification of fault modes can we effectively detect faults and avoid material losses and hidden dangers caused by monitoring errors. At present, the acoustic emission recognition technology based on neural network has some disadvantages, such as less feature information content, and the network is prone to fall into local optimum. Based on the existing technology, this paper has carried out research work in the following aspects: first, the acoustic emission detection technology is studied in detail, and the processing method of acoustic emission (Acoustic Emission, AE) signal is discussed. The differences of AE signal waveforms with different rubbing degrees are compared by experiments. Secondly, the feature extraction of acoustic emission signals is studied. It is proposed that Hurst exponent and approximate entropy be added to the AE feature representation, and analyzed from the angle of statistical correlation and uncertainty. The experimental results show that the new AE feature is effective. Thirdly, the structure of neural network algorithm is studied. The advantages of Gao Si hybrid model (Gaussian Mixture Model, GMM) and backpropagation (BP, Back Propagating) are combined to train the model parameters alternately, and the hybrid model GMM/ANN (Artificial Neural Network, is proposed. ANN) is applied to AE recognition to optimize network performance. Fourthly, a forward chaotic neural network algorithm for acoustic emission recognition is proposed. Aiming at the uncertainty and nonlinearity of AE system, the chaotic characteristics of the network are enhanced by Logistic mapping unit, and the recognition performance of the system is improved. Fifthly, an algorithm for acoustic emission recognition using belief network is proposed. Based on the (Restricted Boltzmann Machines, RBM) model of constrained Boltzmann machine, the (Deep Belief Network, DBN), design of belief network is constructed and the parameters of the model are improved to reduce the local optimal effect. The advantages of DBN in AE recognition are proved by comparison with BP.
【学位授予单位】：东南大学
【学位级别】：硕士
【学位授予年份】：2015
【分类号】：TP183;TB302.5

【参考文献】