【作者】 张婷；

【导师】 苏义鑫；

【作者基本信息】 武汉理工大学 ， 控制科学与工程， 2010， 硕士

【摘要】 温度是科学研究、工业生产及日常生活中经常遇到的一种物理量,随着自动化程度的提高和对产品质量要求的提高,对准确、快速地测温和控温的要求也越来越高,温度的测量已成为科学研究的重要课题之一。尤其是高温物体的温度测量,实现起来难度很大,引起了国内外的重视。在以往的高温物体的温度测量中,我们一般采用的是接触式传感器,像热电偶、热电阻、光学高温计等。但是这些传统的传感器难以得到我们所需的高温物体的实时温度并且精度不高,另外,工业现场存在很多不良环境因素,都会限制一些对环境要求比较苛刻的精度较高的测量仪器和仪表的使用,例如温度过高、过于潮湿、粉尘多、强干扰等都会影响温度测量的精确度。因此,本文研究基于辐射图像颜色和神经网络的高温测量方法。由于神经网络具有对非线性特性的高度拟合能力,而高温物体的颜色与其温度之间存在某种非线性关系,因而采用摄像头摄取高温物体在不同温度时的颜色图像,然后再利用神经网络的非线性逼近能力拟合颜色与温度之间的非线性关系。本文首先通过对颜色模型的分析比较,确立适合本文测温研究的RGB模型,并对采集到的彩色图像进行预处理,得到所需的颜色特征值即R、G、B值。其中,图像的预处理是在MATLAB平台实现的,包括图像的平滑处理和图像的分割。然后将颜色特征值作为神经网络的输入,分别采用BP网络、RBF网络和小波神经网络进行拟合,同时,针对BP网络中隐含层节点数的选择、BP网络学习速率的选择、以及RBF网络中径向基函数扩展系数的选择进行仿真训练,得出仿真结果。最后,根据仿真结果分析比较三种神经网络可知：神经网络应用于基于图像颜色的高温物体温度测量是可行的,并且RBF网络相对于其它两种网络具有训练速度快,误差小,网络稳定等优点,适合于在基于图像颜色的测温系统中使用。更多还原

【Abstract】 Temperature is a physical quantity which we often use in scientific research, industrial production and daily life. With the improvement of automation degree and product quality requirements, the demand for accurately and rapidly temperature measurement and conformity that also increase, so temperature measurement has became one of the most important topics in scientific research. Especially the measurement of high-temperature objects, for it is very difficult to achieve, which has attracted the attention at home and abroad. In the past, we generally used the contact sensor for the measurement of high temperature objects, such as thermocouple, thermal resestance, optical pyrometer and so on. But these traditional sensors are difficult to abtain the real-time temperature that we need and have low accuracy. In addition, the harsh environment of industrial site will also limit the use of some measuring instruments and meters with high accuracy. Therefore, we put into the research of temperature measurement that based on neural network and image color.As the neural network can fit the nonlinear well, while it is existing in a relationship between the color of objects with high temperature and it’s temperature, so that we can use cameras to collect color images of objects with high temperature at different temperatures, after that we use nonlinear approximation capability of neural networks to fit the nonlinear relationship between color and temperature.We first compare the color models to establish a suitable model, that is the RGB model, and preprocess the color images collected to get the color feature values we need. The image preprocessing is implemented in the MATLAB platform, including image smoothing and image segmentation. Then train BP, RBF and wavelet networks with the color feature values as input of neural network. In the meanwhile, we select the hidden nodes, learning rate for BP network and expansion coefficient for RBF network. The analysis of experimental data showes that the temperature measurement based on neural network technology and image color is very feasible, and also analyzed and compared three kinds of neural networks from the convergent speed、the training error and limitations. RBF network can be found for the other two networks has higher training speed、smaller error and more stable. It can be used in the temperature measurement system based on image colors.更多还原