Intelligent Analysis And Research On Gas Sensor Array Data

In recent years, along with the advancement of machine olfaction technology, scilicet gases recognition has been applied to numerous areas(e.g., food safety inspection, medical diagnosis, environmental monitoring). So as to identify special odors, in general, a typical detection system invokes its gas sensor array to collect complex data from a gaseous matter being touched, sends the data gathered to the system’s computer for pattern analysis thereupon. It is worth noting that after the detection process lasts for a period of time, sensor drift(i.e., output signal slowly varies independent of the measured gas) occurs, resulting in data distortion which makes the data more difficult to be analyzed. In order to prevent the accuracy of a measurement from being reduced by the issue, signals are processed for drift compensation.This thesis initially compares the veracity of different recognition methods. As far as characteristic processing is concerned, wavelet transforming and principal component analysis(PCA) are applied. Support vector machines(SVM) and artificial neural networks(ANN) are used for classification. Further, an experiment is devised to evaluate stacked autoencoders in deep learning. Results of the experiment indicate that, on average, deep learning is qualified for all kinds of occasions while traditional methods have both applicable situations and not applicable situations. In other words, deep learning has broad application prospects.The training time of stacked autoencoder network is long for it need a large number of iterative training, this study proposes a deep network for PCA(i.e., it consists of multilayer networks to identify odors, each of which has its own principal component transformation and nonlinear mapping). In this work, mentioned algorithm is utilized to compensate drifts by means of finding characteristic independent of drifts. A conducted experiment announces that using characteristics collected in this way to train classifiers can significantly improve the efficiency of odor detection under the influence of drifts.Last but not least, considering that even a slight misidentification under special circumstances(e.g., poison gas detection in wars, medical diagnosis) could lead to serious consequences, this research puts efforts into enhancing recognition accuracy, at the same time, reducing risks accompanying recognition. For this purpose, this study takes advantage of ideas based on Three-way Decision. First calculate the probability of a gas belonging to one category through nearest neighbor search. Afterwards, compare the probability with the threshold value obtained from a designed loss function. Then choose to accept, postpone or decline. A poison gas together with one kind of common gas is tested to show such approach’s sufficient capability of decreasing risks.