DEVELOPMENT OF OLFACTORY SIMULATOR BY USlNG A NEURAL NETWORK

我试图开发一种使用神经网络的人类嗅觉模拟器，并使用大气压质谱仪作为嗅觉传感器。对现有的大气压质谱仪进行改进，目的是通过改进离子导入部分的透镜组和采用转换双节点的离子检测系统来提高大气压质谱仪的灵敏度。通过这些改进，高质量部分的灵敏度提高了一倍。开发了质谱仪与计算机接口的硬件和软件，建立了对光谱进行连续测量和积累的数据采集系统，提高了测量精度。由于光谱仪的电源等部分存在不稳定部分，导致质量数的测量不稳定。由于这一故障不能预期系统在当年完成，实验继续通过离线操作光谱仪进行。嗅觉模拟器软件的开发及操作结果根据传统的研究，嗅觉可分为七种基本气味。制备了一些化合物，对每种基本气味进行了分类，并测定了其质谱图。将每个质谱图输入到输入层200、中间层50和输出层7的神经网络中，进行学习，直到稳定到相应气味的分类为止，然后将新样品的质谱图输入模拟器，并确定是否可以将样品分类为相应的气味，从而证实了使用神经网络能够从质谱图中对气味进行分类。

I try to develop a human olfactory simulator using a neural network with an atmospheric pressure mass spectrometer as a sensor of the smell.1. Improvement of existing atmospheric pressure mass spectrometerI aimed at rising the sensitivity of the atomospheric pressure mass spectrometer with the improvement of lens group of the ion introduction part and the ion detection system by using the conversion di-node. Sensitivity in a high mass part has doubled by these improvements.2. Developing the hardware and the software to interface between the mass spectrometer and the computer.I try to make the data acquisition system which is measuring the spectrum continuously and accumulating them to improve accuracy of measurements. Because an unstable part existed in the power-supply unit etc. of the spectrometer, the measurement of the number of mass became unstable. As the completion of the system in the current year was not able to be expected by this trouble, the experiment is continued by off-line operation of the spectrometer.3. Development of software for simulator of olfactory and operation resultThe smell is classified into seven kinds of basic smells according to a traditional research. Some compounds which are classified each basic smell were prepared and the mass spectra were measured. Each mass spectrum pattern was put into the neural network in input layer 200, middle layer 50, and output layer 7, which was made to study until settling to the classification of the corresponding smell After that, a mass spectrum pattern of a new sample was put into the simulator, and whether it was possible to classify the sample into the corresponding smell.As a result, it was confirmed to be able to classify the smell from the mass spectrum by using the neural network.