Research On The Novelty Detection Methods Based On Equipment's Odor Analysis

With the development of manned spaceflight, deep-sea exploration, aircraft, etc., the reliability and safety of equipments are more and more important in large equipments with non-open space, such as spacecraft, submarine, aircraft and so on. For such systems, one of main reasons that affect the air environment of non-open space cockpit is the pollution caused by long-term running large instruments, especially when it is in abnormal state. On the other hand, the polluted air is also including some important information to reflect the equipments condition.Based on bionic principle, electronic nose technology is an important means of odor detection, which is expected to provide a new approach for achieving health monitoring of intensive system and early novelty detection with no-open space. In this thesis, the present condition of electronic nose technology is summarized firstly. Then, based on the experiments to simulate oil leakage and wire overheating that cause from mechanical equipments running in abnormal state, several methods for odor identification and separation are analyzed.The main contents and conclusions are as follows:(1) Based on the principle and structure of electronic nose system, olfactory detection technology is introduced. The present research condition of involved key technologies is discussed, which includes olfactory sensor array, signal pre-processing, pattern recognition, etc.(2) A series of simulative experiments and some basic analysis of response signals of oil leakage and wire overheating are introduced. The experimental environment, experimental methods and experimental equipment are described. The process of odor concentration change is analyzed with odor extraction and air replenishment in the simulative experiment. The response signals of olfactory sensor array are analyzed.(3) Aimed at the qualitative identification of odors, the experimental data are pre-processed, and the main features are extracted based on principal component analysis. Linear discriminate analysis and feed-forward neural network are used to identify odors qualitatively.(4) Aimed at the quantitative identification of odors, complex frequency-domain analysis method that is used to compensate the change of odor concentration with odor extraction and air replenishment is investigated. A nonlinear correlation model between sensor response and odor concentration as well as a principal component regression model are established and validated.(5) To separate the mixed-odor shed by mechanical equipments at abnormal state, a blind source separation model is proposed based on the independent component analysis, and applied in the separation of mixed odors.