Performance Improvement Of A Portable Electronic Nose For Indoor Air Quality Monitoring Based On Neural Networks Ensemble

If there is a modernization process that has successfully changed human livingconditions since its inception, it is undoubtedly the industrialization process. However,like any process, industrialization has many negative aspects. Environmental pollution,more specifically the outdoor as well as the indoor air pollution is one of these negativeaspects. Consequently, there has been a resurgence of interest in the development ofsmart sensing systems for real-time air quality monitoring, especially in indoorenvironments where we usually spend most of our time and from which serioushealth-related problems may emerge. Electronic nose (E-nose) systems are found asgood alternative to existing techniques such as the use of human panel, analyticalmethods based on gas chromatography or mass spectrometry, to name a few. Moreover,these systems can be constructed using cost-effective off-the-shelf metal oxidesemiconductor gas sensors. However, these sensors are prone to drift and interferencefrom non-target gas analytes, which may jeopardize the performance of E-nose systemsif drift counteraction and interference removal are not accounted by the patternrecognition algorithm. Therefore, robust signal processing algorithms that considerthese factors are of paramount importance in E-nose systems. Designing such signalprocessing algorithms is the main objective of this thesis work.Although E-nose systems have been around for decades, up to now, many potentialusers do not know about these systems. Therefore, the history, key concepts andarchitecture (this includes sampling and delivery systems, sensor array, signalpreprocessing, feature extraction, and pattern recognition) of E-nose systems are firstdiscussed, and then some of their applications are described. Furthermore, someavailable commercial E-nose systems are enumerated. Finally, current developmentsand problems associated with E-nose systems are pointed out.Calibration of E-nose systems requires some initial data sets that are mostlygenerated through several experiments under controlled atmospheric conditions. Aself-made E-nose system is first introduced, and then the experimental setup andprocedure to generate initial data sets are described. These data sets constitute a basisfor data preprocessing and pattern recognition.Orthogonal signal processing (OSC) is a preprocessing technique that has beensuccessfully applied in electronic nose systems. To investigate the effectiveness of OSC, an empirical study using two different multivariate regression methods, multilayerperceptron (MLP) and partial least squares (PLS), was carried out. Experimental resultsusing data sets of six indoor air pollutants show that, combination of OSC and MLP iseffective only in the presence of very strong background noise, whereas thecombination of OSC and PLS is very effective regardless of the level of backgroundnoise. However, the performance of MLP was better than that of PLS, which implies theneed for nonlinear pattern recognition methods.Artificial neural networks (ANNs) and support vector machines (SVMs) arepattern recognition methods that are widely used in E-nose systems. ANNs are based onempirical risk minimization; while SVMs are grounded in the framework of statisticallearning theory which is based on structural risk minimization. The basic principle ofANNs (with emphasis on MLP) and SVMs was thoroughly discussed. Owing to itsglobal search capability, genetic algorithm was used to optimize the initial weights andthe hyper-parameters of MLP and SVM, respectively. Experimental results using datasets of five indoor air pollutants show that, although both MLP and SVM modelsprovide satisfactory results, the latter have better generalization performance, which isin line with the theoretical assumption. However, for embedded applications, MLPmodels involve less computational complexity than SVM models. This is the rationalebehind stressing on MLP models, at the cost of improving their generalizationperformance.There are many methods to improve the generalization performance of MLP neuralnetworks. These include regularization, cross-validation, training with jitter (noise), andensemble method. The latter is the focus of this thesis work. The success of ensemblemethod can be explained based on the bias-variance decomposition of error, whichshows that ensemble method can reduce variance and also bias. Ensemble learningrefers to techniques which generate multiple base models using traditional machinelearning algorithms and combine them into an ensemble model. In the generation stagethe objective is to create base models that are sufficiently accurate and diverse in theirpredictions. This can be done through three categories of methods: methods based onthe modification of the learning set (e.g. bagging, boosting), methods based on themodification of the training algorithm (e.g. Negative Correlation Learning), andmethods based on selection (e.g. ambiguity based method, GASEN). As in thecombination stage, linear methods (e.g. simple averaging, weighted sum), and nonlinearmethods (median rule,“stacked generalization”) are commonly used. Another important alternative for creating an ensemble of base models is the mixture of experts, which isbeyond the scope of this thesis.Owing to their good empirical results and theoretical support, bagging andboosting are the most widely used ensemble learning algorithms. Moreover, bagging hasbeen found to be more effective on unstable estimators (predictors) such as supportvector machines, artificial neural networks, to name a few. A new selection basedensemble method is proposed and discussed. The method combines variance inflationfactor (VIF) as diversity metric, performance measure (either the mean squared error, orthe mean absolute relative error of prediction) and genetic algorithm to select optimumnumber of base networks (models) from a pool of bagged neural networks. Results fromtwo empirical studies show that the proposed method compares unfavorably with othersimilar methods in only few cases. Moreover, this method performs better than the bestbase network and the standard bagging method. More research on the rules regardingVIF will significantly improve the performance of this method.Long or short term drift is one of the most serious problems associated with gassensors. It can drastically affect the performance of electronic nose systems if nocounteraction is performed. In the last empirical study, an ensemble method that cancope with drift problem is proposed for online gas concentration estimation.Experimental results show that the method is not only effective but also attractive whencompared with other ensemble methods.