Research On Application Of Fire Detection Technology Based On Video In Nuclear Power Plant

Safety is the "lifeline" for nuclear power and our country has demanded that nuclear power "must ensure absolute safety". All workers engaged in nuclear power industry have in their mind that "safety first" and "safety is of the utmost importance". Among many factors threatening nuclear safety, "fire disaster" is destructive with uncertainty. Once the fire bursts, it will exert inestimable effect. However, there are obvious disadvantages of traditional fire detector in irregular large space or environment with dust, humidity and radiation. Thus, it is urgent to find certain non-contact fire detection technology which can adapt to open space with rapid speed of response, strong capacity of resisting disturbance. With the fire detection technology based on video, this problem can be resolved in a good way by adopting such techniques as image processing, feature extraction and fuzzy recognition. In this way, the fire detection system will become timelier and more robust and effective. The main research contents are as follows:(1) It is to analyze the limitations of traditional fire detection system to nuclear power plant in different circumstances such as outdoor areas, high and large space as well as humid and radiant environment. Besides, it is to present areas of potential safety hazards and to make structure topology of network with planning of distribution of video detectors. With original detection system as the main part, there is explanation of ways to combine the new system with the original one. Guided by the principle of centralized processing and decentralized control, the project has helped lay a firm foundation for further image processing and recognition.(2) Preprocessing of images. As the noisy points will exist in videos collected to different degrees due to such factors as electromagnetic interference and sudden change of illumination, there are detailed explanations in this chapter of basic methods such as equalization processing of histogram, mean filtering, median filter and mathematical morphological processing. After comparisons of test of algorithm, it is to select the method which performs best in terms of noise attenuation effect and reservation of original image details.(3) As to features of fire disaster, there is analysis of RGB color and HSV color space model specific to static features of flames and analysis of flames'whole movement and change of areas, height and edge specific to dynamic features of flames. Besides, a model on flames features has been established according to static and dynamic features of flames. As to features of smog, the three features including area change rate, color feature and texture feature have been selected as bases for recognition to effectively prevent from impact of common interference objects. For extraction and segmentation of image objects, the method of Gaussian mixture background modeling is adopted in combination with the method of three-image difference to generate good results.(4) The fire detection and recognition algorithm based on least squares support vector machine (LS-SVM) is proposed. By taking advantage of features of flames and smog extracted and of contextual information of features, it is to first judge the color features of fire disaster and identify the main components (flames or smog) so as to select appropriate SVM fire recognition model. As to elimination of interference, videos of pedestrians, lighting, sunshine as well as wind and rain are selected according to actual situation on site and the reversal training of SVM is conducted. With this algorithm, therefore, it will help eliminate interference information under special circumstances for the nuclear power plant to improve accuracy of fire detection system.As is indicated by the experimental results, the fire recognition based on SVM will help overcome deficiencies of traditional detection methods and improve accuracy of fire recognition so as to ensure both sensitivity and robustness.