RESEARCH INITIATION AWARD: The Application of Neural Networks to Gas Sensors

9309012 SEGEE传感器技术即使在非常低的浓度下也能检测到潜在的危险气体。然而，这些传感器的选择性不是很强，而且响应往往是气体浓度的非线性函数。通过使用具有不同灵敏度的传感器阵列，提高了选择性。校准这种阵列的数学过程非常复杂，如果单个传感器的响应过于相似，可能会导致性能低下，甚至不稳定。如果通常需要忽略来自可用传感器子集的有效传感器数据。为了解决上述问题，人们提出了人工神经网络模型作为一种选择。它们能够学习多维非线性函数，能够利用冗余信息，并且具有自适应能力。这项拟议的工作将使用从受控环境中的传感器收集的数据来训练三种不同类型的人工神经网络，以识别气体类型和浓度。这些神经网络的性能将首先通过与训练数据同时收集的测试集来测量，但不用于训练。最终，将对现场绩效进行测量，并使用在线培训来微调绩效。***

9309012 Segee Sensor technology exist that is capable of detecting potentially dangerous gases even in very low concentrations. However, these sensors are not very selective and often have responses that are nonlinear function of gas concentration. Selectivity has been improved by using an array of sensors having differing sensitivities. The mathematics of calibrating such an array is very complex and poor performance and even instability can result if the responses of individual sensors are too similar. If its generally necessary to disregard valid sensor data from a subset of the available sensor. To address the above problem artificial neural network models have been suggested as an alternative. They are capable of learning multidimensional nonlinear functions, can exploit redundant information and are adaptable. The proposed work will use data gathered form sensors in a controlled environment to train three different types of artificial neural networks to identify gas type and concentration. Performance of these neural networks will be measured first by means of a test set gathered at the same time as the training data, but not used for training. Ultimately, performance in situ will be measured and on-line training used to fine tune performance. ***