无线传感器网络中的声目标识别问题研究
文献类型：学位论文
| 作者 | 管鲁阳 |
| 学位类别 | 博士 |
| 答辩日期 | 2008-09-06 |
| 授予单位 | 中国科学院声学研究所 |
| 授予地点 | 声学研究所 |
| 关键词 | 声目标识别 双耳模型 特征提取 数据不均衡 流形学习 |
| 其他题名 | Research on acoustic target recognition in wireless sensor networks |
| 学位专业 | 信号与信息处理 |
| 中文摘要 | 随着MEMS、数字信号处理、无线通信等相关技术的发展和高性能DSP芯片的广泛应用，由多个传感器节点组成的无线传感器网络能够实现对布设区域内特定目标的实时监控、区域态势分析及预测，在工业、环境、交通、军事等众多领域得到了越来越广泛的应用。 无线传感器网络中，目标识别是传感器节点所需具备的重要功能之一。基于声信号的目标识别由于具有隐蔽性好、抗电磁干扰能力强、功耗低等特点，是传感器节点感知周边环境的重要手段。本文在无线传感器网络的应用背景下，重点研究了声目标识别中的特征提取和数据不均衡问题，论文主要工作和创新点如下： （1）研究了流形学习在声信号处理中的应用及特点，提出一种基于流形学习的单类分类算法，从分类器的角度解决数据不均衡问题。仿真试验表明流形学习能够对一段连续目标声信号进行有效分析，发现其短时谱中所包含的低维流形，并且保持各类目标声信号的流形的相对独立性。在此基础上，论文提出了基于流形学习的单类分类算法，利用单类分类算法识别样本较少的正类目标，流形则是判决其是否属于正类目标的依据。车辆识别试验结果表明此算法能够有效地解决数据不均衡条件下正类目标的识别问题。 （2）提出基于人耳听觉模型的特征提取方法。针对无线传感器网络识别的主要目标（履带车和轮式车）声信号的特点，以人耳生理模型为信号分析手段，设计了适用于单通道信号基于耳蜗模型的倒谱系数（Monaural Model based Cepstrum Cofficient, MoMCC）和适用于双通道信号基于双耳模型的倒谱系数特征（Binaural Model based Cepstrum Cofficient, BiMCC），试验证明上述特征具有良好的分类性能和对噪声干扰的鲁棒性。 （3）讨论了基于双耳模型的目标方向估计。在基于双耳模型的特征提取算法中，利用双耳模型计算结果，根据特征频率峰值相对于对角线的偏移估计目标所处方向。仿真试验和对真实车辆信号的计算结果显示利用双耳模型可在特征提取的同时估计目标方向。 （4）综述模式识别中数据不均衡问题的研究现状并提出一种基于代价敏感学习的数据处理方法。由于实际应用中数据采集条件的限制，不同类别目标的数据可能存在显著差异而导致数据不均衡问题。论文概述了数据不均衡问题的本质、分类器评价标准及主要解决方法等方面的研究现状并从数据处理的角度提出了一种数据重采样方法，以代价敏感学习为准则确定正负类样本比例，通过对样本集重采样实现不同类样本之间的均衡。在车辆识别及UCI数据库多种类型目标识别的测试中有效避免了数据不均衡对正类样本识别正确率的影响，并且对不同领域数据具有很好的适应性。 |
| 英文摘要 | Along with the development of MEMS, DSP, wireless communication and other related techniques, wireless sensor networks (WSN) have emerged, which is consisted of many multifunctional sensor nodes, can detect, recognize and monitor the special target, and predict its future status by collaboratively real time processing. Today, WSN has extensive applications in many fields such as industrial and environmental monitoring, traffic control, and battlefield surveillance. In WSN system, target recognition is an important function of the sensor node. Since the microphone works passively with low cost and is robust to the electromagnetic interference, acoustic target recognition is especially meaningful for various applications. Furthermore, if the sensor node can estimate the direction of the target, it would facilitate target location or tracking of WSN system. In this background, the dissertation aims at the development of feature extraction and the solution of imbalanced dataset problem for robust acoustic target recognition. The main contributions of the research works are as follows: (1) Manifold learning was explored as a method of acoustic signal processing and a new one-class learning algorithm based on it was proposed to address the imbalanced data problem. Simulation experiments showed that manifold learning could be used to analyze the shot-time spectrum of the acoustic signal, find the manifold embedding in it effectively and keep the manifolds of the different targets’ signals independently. A new one-class learning algorithm based on manifold learning was proposed to recognize the target class with few data in which the manifold was the criteria to decide if the sample belongs to this class. The performance of vehicle recognition showed that this classification algorithm eliminated the influence of imbalanced dataset successfully. (2) New feature extraction algorithms were proposed based on the human physiological hearing model. According to the characteristics of the targets, tracked and wheeled vehicles, MoMCC (Monaural Model based Cepstrum Coefficient) was proposed based on the cochlear model for the mono acoustic signal and BiMCC (Binaural Model based Cepstrum Coefficient) was designed based on the binaural hearing model for the stereo signals. Experimental results of vehicle target recognition showed that these two features obtained better performance and were more robust to the noise. (3) Direction of Arrival (DOA) estimation based on the binaural model was explored primarily. In the 2-D output pattern of the binaural model, the peak of the characteristic frequency would deviates from the main diagonal obviously because of the ITD and ILD between the two input signals. And the more interaural time delay, the larger the disparity from the main diagonal. Experimental results showed that it estimated the direction of the vehicle successfully based on the disparity. (4) A survey of research on the imbalanced data problem was presented and a new cost sensitive re-sampling method for imbalanced data was proposed. Learning from imbalanced data is an important topic, arising very often in practice in classification problem. The nature of this problem, its influence on the performance of classifier and a variety of methods addressing this problem were presented in this survey. To eliminate the influence of the imbalanced data, a re-sampling method was proposed to balance the data among the different classes according to the rule of cost sensitive learning. In the test of UCI database and vehicle recognition, this method achieved better performance, comparing with other data operation methods. |
| 语种 | 中文 |
| 公开日期 | 2011-05-07 |
| 页码 | 121 |
| 源URL | [http://159.226.59.140/handle/311008/194]  |
| 专题 | 声学研究所_声学所博硕士学位论文_1981-2009博硕士学位论文 |
| 推荐引用方式 GB/T 7714 | 管鲁阳. 无线传感器网络中的声目标识别问题研究[D]. 声学研究所. 中国科学院声学研究所. 2008. |
入库方式： OAI收割
来源：声学研究所
浏览0
下载0
收藏0
其他版本
除非特别说明，本系统中所有内容都受版权保护，并保留所有权利。