【作者】 田华；

【导师】 尚志远；

【作者基本信息】 陕西师范大学 ， 声学， 2004， 硕士

【摘要】 本课题得到国家自然科学基金项目(项目批准号：10274047)资助。动物鸣声是动物的一种很重要的生物行为，动物的声信号包含着很重要的、丰富的生物含义，其中最为重要的一点就是动物的声行为具有种的特异性，利用动物鸣声来进行分类已是动物系统分类学研究的热点。 利用动物鸣声来进行分类在国内仅出现在鸟类和昆虫，哺乳动物利用鸣声分类基本未见。传统的鸣声分类时一般依据不同种动物鸣声信号在时域波形上的波形结构和脉冲率等的不同，功率谱上相对幅度最大所对应的频率即主频的不同来进行分类。传统鼠类的分类首先从外观上区分，如区分不开就需要进行解剖，根据其头盖骨来分类，这样的分类非常复杂而且需要专业人员。人工神经网络网络作为一门新兴的学科，从产生至今已经取得了长足的发展，被广泛深入地应用于模式识别、优化问题、图形图像处理、信号处理、自动控制、预测等等众多的领域，为这些领域相关问题的解决提供了新的有效的手段。本文我们采用人工神经网络的方法根据鸣声对不同种属的鼠类鸣声进行分析研究，有助于鼠类的分类简单化、自动化，且该分类方法分类速度快、准确率高。在国外，人工神经网络被广泛的应用到依据动物鸣声对同种动物不同个体的和同类不同种群的动物的分类的研究；而在国内，这个领域的研究基本处于空白。 我们首先对多种动物的鸣声研究现状作了较为详细的总结分析，探讨了动物声通讯的生物学意义。鸣声作为依据在动物分类中、在害虫、鼠害的监测和防治工作中广泛应用以及应用动物的鸣声进行。 针对鼠类叫声的分析提出了专业的录音话筒+专业声卡+计算机的数字录音方法。通过该方法尽可能的减少信号采集和存储中的中间环节，从而减少了噪声的引入和信号的畸变，避免了人的手工操作，同时也充分发挥了计算机自动程序执行、自动控制的能力，实现全自动不间断录音、存储。存储鸣声信号以WAV文件格式存在计算机的硬盘上，这样信号不易受到电、磁的干扰，使信号安全、准确和可靠。 对录制好的信号，我们利用数字信号处理及神经网络等方面的理论知识，在Matlab的环境下自己编写了程序实现了对信号时域波形和频域功率谱的方便的手工分析，以认识信号的特征;在对信号的各种特征有了较深入的认识之后，用程序实现了信号自动分段分析、叫声信号自动截取以及叫声信号特征自动选择、提取;然后我们对所研究的大白鼠、甘肃盼鼠、黄胸鼠、小白鼠、棕色田鼠的各种状态下的叫声信号分别分成了验证样本集和训练样本集，每个样本集中都含有45个叫声信号文件;利用训练样本集对设计的人工神经网络(BP网络方案二)进行训练，训练成功的网络对训练样本集的样本识别率为100%，对验证样本集中样本的正确识别率可达到90%以上(大白鼠、甘肃黔鼠、黄胸鼠、小白鼠、棕色田鼠正确识别率分别为93.33%、97.78%、97.78%、95.56%、93.33%)，并对实验结果作了进一步分析。实验结果表明，本文采用的人工神经网络的分类分析方法是更容易实现的、可行的和有效的，它较其他方法的突出的优点是极强的对叫声信号特征的自动提取能力和推广能力，是一种很有前途的分析方法。更多还原

【Abstract】 This thesis is supported by National Science Fund (No. 10244047). Call is a main behavior of animals. Calls of animals contain important and abundant biological meanings, among which animal characteristics revealing is the most important. Calls based animal classification has become the hotspot of animal taxonomy.Calls based classification is used in avian and insects only recently and still not in mammals in China. Common methods classify animals by their difference in sound pulse rate, wave form structure in time domain and main peak frequency in frequency domain et cetera. Common methods classify mice first by appearance, then by dissection, which need manual operation of professionals and are sophisticated. As an emerging subject, artificial neural network (ANN) has been seen vast development ever since start, and been widely used in the fields of pattern recognition, optimization, image process, signal process, prediction etc, and provided novel and effective method for resolving relevant problem in these fields. Artificial neural network has been widly used in call based classification of individual from the same group and different group from the same species. Little has been done in this field in our country. In this paper ANN is used for classification of mice of different species by the analysis of mice calls. It makes mice classification simple, automatic, exact and fast.In this paper, Summarization of the status of call research and biological meaning of sound communication comes first, with disscussion of classification based on call-pattern recognition and inspection, prevention and cure of pests, mice using their calls. Followed by the introduction of ANN, its basic theory, basic algorithm, merit and defect of back propagation network and ANN application in pattern recognition field. Call based mice classification comes last.Compared with traditional sound recording, we put forward a method for gathering call signals using professional microphone, professional sound card and acomputer. It reduces the processes needed in singal gathering as far as possible and without manual operation. It also makes full use of the automatic program execution ability of computer. So it can automatically record call signals and save gathered data in waveform format to hardisk without stop.Recorded data are processed by program developed in Matlab6.5 using theories of digital signal process, neural network, et cetera. It includes manual analyzation of signal, automatic sound finding and cutting out, automatic feature extraction, neural network training and validation. Experiments are carried on the recorded data of call of five species of mice. For both each species and for both training and validation 45 calls are used. The trained network obtains 100% recognition and more than 90% validation success. The result shows that ANN as a recognition method in calls of mice is very promising, effective and viable.更多还原