Data Evaluation for an Electronic Nose

Abstract: An electronic nose is a device that tries to perform the same task as the human olfactory system, to be able to discriminate between different odours. In an electronic nose there are currently a number of individual sensors (typically 5-20). The response from a chemical sensor is usually measured as the change of some physical parameter, e.g. conductivity or current. The responses from all sensors form a response pattern. Response patterns from different odours can be learnt by a computer in order to recognise them when the sensors again are exposed to one of the initially learnt odours. The learning can be made using a number of different techniques, e.g. statistical methods and artificial neural networks. As long as the sensor responses do not vary over time for a given odour, there are many ways to do the learning/recognition. The properties of chemical sensors, however, change over time. Measurements on the same odour will therefore give different sensor responses if they are made with long enough (a few days) intervals. The aim of the work described in this thesis has been to examine different pattern recognition techniques, both for non-drifting (measurements in a shorttime period) and drifting (long term measurements) sensors.The first step in the data evaluation has been to plot individual sensor responses, and to make principal component analysis (PCA) plots in order to get important informationon general trends and problems in the data set. A PCA is a simple way to project data from several sensors to a two dimensional plane. After initial data treatment to reduce the problems encountered, artificial neural networks have been used for pattern recognition for non-drifting sensors. The nature of sensor drift has been analysed. For the case with drifting sensors, a Box-Jenkins model has been used together with a linear classifier for gas identification. The method utilised the fact that similar sensors suffer from similar drift effects. To deal with a measurement situation when also similar sensors drift differently, a recursive updating of the models has been made.Recognition of different paper qualities has been made with an electronic nose consisting of gas sensors of different types. It was not possible to discriminate between all the papers using only one sensor type, which shows the importance of sensor choice for the applications. To get a high recognition rate, all data evaluation had to be made with the responses relative to air. Using responses relative to a reference gas is now customary in all commercial electronic noses.To study the effects of long term drift, a data set with four different alcohols measured over 45 days was used. The first gas recognition method used a reference gas to compensate for the sensor drift. The other methods were based on the Box-Jenkins model with a classifier, as described above. The best results were obtained with a recursively updated Box-Jenkins model, where more than 90% of the samples were correctly classified. This is as good classification as for other published results when the measurements have been made during such short time that no drift effects have occurred.The identification of bacteria from gases evolved during growth has been made. The gas response varied over time during growth, and mathematical fits to the response curves were made. The parameters of the fits were then used in pattern recognition techniques to determine the most important parameters, and also for identification of the bacteria. A relatively good classification rate (76%) was achieved, and this method promises to be useful for medical diagnosis in the future.