Electronic Nose Techniques For Identification And Grade Evaluation Of Many Odors

The research on electronic noses has been very active in the last several decades due to the successful applications in a wide range of areas, such as agriculture, bio-medical, chemistry, environment monitoring, food, manufacture and military. It is more objective and accurate when using an electronic nose, than that of a human sensory panel, a mass spectrometer, a chromatograph, as the method to do quantitative or qualitative analysis on volatile organic compounds. This thesis focuses on the analysis of multiple volatile organic compounds (VOCs) based on an improved electronic nose, including applications to quality-grade identification of petroleum waxes, online monitor and prediction of bio-fermentation processes, hierarchical models on VOCs, improved method on formation of economic training subsets.The main contributions of this thesis are summarized as follows.(1) Two kinds of electronic noses with different structures are designed to meet various purposes. The electronic nose A is for one-time odor measure and analysis of such materials as petroleum waxes, tackiness agents, leathers, glycerol and edible vegetable oils; while the electronic nose B is for continuous online monitor and prediction of bio-fermentation processes. Particularly, improvement is made on gas paths and sampling procedure of the electronic nose B to simultaneously monitor 5 fermentation cylinders. What’s more, a method to choose, replace and validate gas sensors is proposed to solve their poisoning or drifting problem.(2) Hierarchical discrimination and quantification models are studied in order to simultaneously quantify multiple kinds of odors with an improved electronic nose. Such tasks are first regarded as multiple discrimination tasks and then as multiple quantification tasks, and implemented by the hierarchical models with the divide-and-conquer strategy. On one hand, the discrimination models are the common classifiers, including the nearest neighbor classifiers, local Euclidean distance templates, local Mahalanobis distance templates, multi-layer perceptrons (MLPs), support vector machines (SVMs) with Gaussian or polynomial kernels. On the other hand, the quantification models are multivariate linear regressions, partial least square regressions, multivariate quadratic regressions, multi-layer perceptrons, support vector machines. We developed several types of hierarchical models and compared their capabilities for quantifying 12 kinds of volatile organic compounds. The experimental results show that the hierarchical models composed of multiple single-output multi-layer perceptrons followed by multiple single-output multi-layer perceptrons with local decomposition, virtual balance and local generalization techniques, has advantages over the others in the aspects of time complexity, structure complexity and generalization performance.(3) A novel economic training subsets formation technique is proposed according to characteristics of two-dimentional principal component analysis of 12 kinds of VOCs. The method effectively increases the generalization performance of classification model while reduces the average prediction error.(4) A novel method for quantifying odor levels for petroleum waxes based on the improved electronic nose has been proposed to solve problems such as high cost, subjectivity and hard to quantify. Pattern recognition technology has been employed to analyze on the data, resulting into high classification accuracies on more fine-grained odor levels. What’s more, the influence of temperature on sensor response curves and classification accuracies is analyzed. Results show better performance when the experiment is done at 50 ℃.(5) Pattern recognition method and time-series prediction techniques are applied to online monitor and prediction of bio-fermentation processes. Firstly, from the response curves of bio-fermentation processes of polyhydroxyalkanoates(PHA) and Cephalosporin C, we find that the improved electronic nose can predict contamination during fermentation procedure. Secondly, a hybrid time-series prediction model has been presented to real-time prediction during the fermentation processes of tylosin, which combines efficiency and accuracy by applying linear autoregressive method to several specified sensors and nonlinear autoregressive method to the rest sensors.