Application Of Improved Extreme Learning Machine In Fire Detection

With the rapid development of urbanization,the buildings in city has appeared the characteristics of concentration and high-rise apartment,which should cause the increase of probability of fire.Higher requirements for fire detection methods have been proposed.Due to the dynamic process and non-stationary of fire,the traditional fire detection method is only dealing with the output of a single sensor,and it will easily generate false dismissal and false alarm.At present,the main research of fire detection judge fire.In order to more accurately carry out fire warnings and adopt more effective methods of fire suppression.It is great significance to research the types of combustion during the fire of judgment.The application of machine learning and deep learning has become a hot topic,and it has also driven some devices to develop towards intelligence.These provide theoretical and platform support for solving problems in fire detection.To sum up,in order to overcome the defects of traditional fire detection,this paper compares the principles of traditional and intelligent fire detection methods,and a method of fire detection based on improved extreme learning machine is proposed.To overcome the single sensor susceptible to environmental interference,multi-sensor fusion technology be used to comprehensively consider multiple characteristic signals in the fire combustion process.This avoids the incorrect description of the monitoring environment by a single sensor.Experimental results show that the proposed fire detection algorithm can reduce the false dismissal rate of fire.In addition,this paper improves the extreme learning machine method.The DS-ELM fire combustion identification method was proposed by combining the DS evidence theory.The simulation results show that the proposed fire combustion detection method can achieve a relatively excellent classification performance for different combustion products.The contribution for this paper mainly consists of the following parts:(1)In order to make the experiment closer to the real environment,we constructed a fire burning material collection platform,collecting fire data and different burning data in real environment,and then the data is transmitted to a computer terminal through WIFI;(2)To make the collected combustion data more reliable and efficient,we study some methods to pre-process the original data and reduce noise interference;(3)Since the misclassification cost of fire sample is different from that of non-fire sample in the fire detection,a cost matrix is introduced into the learning process,and a cost-sensitive extreme learning machine fire detection method is proposed.The experimental results show that the false dismissal rate can be decreased near to zero and the false alarm rate of non-fire samples keep at lower level.The method overcome the existing fire detection methods only to pursue high classification accuracy,while ignoring the shortcomings of high false negatives in fire samples.(4)We analyze the possibility of classification between different types of combustibles and compare other recognition algorithms.Finally we select extreme learning machine as a method for fire combustible identification.(5)We analyze the advantages and disadvantages of DS evidence theory and proposes the DS-ELM method to improve the classification accuracy of fire combustion.Compared with the non-optimized ELM method,the proposed method can enhance about the accuracy of 10% and have good adaptability at different combustion points.