基于深度学习的工业气体识别研究

发布时间：2018-06-24 02:28

本文选题：机器嗅觉 + 气体识别　；参考：《上海应用技术大学》2017年硕士论文

【摘要】：由于信息技术的迅猛发展,在安全检查、工业生产等很多的领域当中都出现了机器学习的影子。许多相关的技术都随之产生,为人类的生活带来了方便。电子鼻可用来检测鉴别和分析各种各样气味的气体,其作为机器嗅觉的典型应用,已经渗入到各行各业之中,在安全检查、工业生产、环境保护、医疗诊断等领域中,气体识别系统应用尤其广泛。气体识别系统采用一种较为复杂的时间序列信号,这种信号由敏感膜材料因被气体分子吸附而震动所出现。敏感膜材料、外界的环境因素、包括气体的种类和密度都会影响到这种信号的产生。因此,人们通常通过手工设计特征,并在原有的基础上结合支持向量机等方法来解决气体数据波形复杂、分析困难等问题。在识别的过中,传感器的响应值会逐渐下降,数据整体的分布也会随之产生变化,这就导致对气体数据的分析变得难上加难,准确率也有所降低。这种使得气体数据分析变难,准确率下降的现象叫做传感器漂移。人们为了消去这种现象采用了很多信号处理领域的方法。本文引入了深度学习的方法来解决传统气体识别过程中特征提取方法的不足,利用深度学习网络的特点以及深度学习过程中无监督学习的特征来自动地获取更能反映原始气体数据本质的特征,从而解决传统方法中特征提取的不足,以及对传统方法中不能有效解决的无标注气体识别问题。其次本论文验证了在气体数据明显漂移的情况下,利用深度学习技术可以提高气体识别的准确率。这是因为利用此方法可以得到比起初输入的数据更加完善、准确的数据。这对漂移补偿问题提供了更具可行性的解决方案,也对在工业生产中检测有毒气体、未知气体识别提供了参考。
[Abstract]:With the rapid development of information technology, machine learning appears in many fields, such as safety inspection, industrial production and so on. Many related technologies have followed, bringing convenience to human life. Electronic nose can be used to detect and analyze various odour gases. As a typical application of machine smell, electronic nose has penetrated into all walks of life, in the fields of safety inspection, industrial production, environmental protection, medical diagnosis, etc. Gas identification systems are particularly widely used. The gas recognition system uses a complex time series signal, which is vibrated by the sensitive membrane material adsorbed by gas molecules. Sensitive membrane materials, external environmental factors, including the type and density of gases, will affect the production of this signal. Therefore, people usually design features by hand, and combine the support vector machine and other methods to solve the problem of complex gas data waveform and difficult analysis. In the process of recognition, the response value of the sensor will decrease gradually, and the distribution of the whole data will change, which makes the analysis of gas data more difficult, and the accuracy is also reduced. This makes it difficult to analyze gas data and reduces accuracy, called sensor drift. In order to eliminate this phenomenon, many methods in the field of signal processing have been adopted. In this paper, the method of depth learning is introduced to solve the shortcoming of the traditional method of feature extraction in the process of gas recognition. The features of the deep learning network and the unsupervised learning in the process of the deep learning are used to automatically acquire the features which can reflect the essence of the original gas data, thus solving the shortcoming of feature extraction in the traditional methods. And the problem of unlabeled gas identification which can not be effectively solved by traditional methods. Secondly, this paper verifies that the accuracy of gas recognition can be improved by using depth learning technology when the gas data drift obviously. This is because using this method can get more perfect and accurate data than the original input data. This provides a more feasible solution to the drift compensation problem, and also provides a reference for the detection of toxic gases and the identification of unknown gases in industrial production.
【学位授予单位】：上海应用技术大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP181

【相似文献】