Research On Indoor Smoke Source Localization Algorithms Based On Wheeled Mobile Robots

Since the early 1990 s,many researchers have dedicated attention to the design of a robotic system capable of finding a source of chemical or smoke particle release,that is,olfactory robot.Olfactory robots use various kinds of gas concentration sensors or smoke particle distribution density sensors as its olfactory input to search and track gas or smoke source by climbing concentration gradient or imitating biological behaviors,etc.In existing research work,olfactory robots are commonly used to search for toxic volatile gas sources and combustible gas leakage sources.They use volatile organic compound(VOC)sensors to measure the concentration of gaseous chemicals(such as gaseous formaldehyde,benzo,ethanol or acetone)in order to locate gaseous chemical volatile sources.However,in many situations,such as burning some materials,or even more severe-in fire disasters,the burning smoke source does not release gaseous chemical substance that can be detected by VOC sensors,but rather solid airborne particles that cannot be detected by common VOC sensors.Solving the smoke particle source localization problems requires substantially changing the sensing technology and in turn adapting the related localization algorithms.Smoke source localization tasks often occur inside buildings.The indoor airflow field,smoke plume distribution and the movement of the robot are all affected by the indoor structure and the obstacles.Due to the above difficulties,research on smoke particle source localization is still in its early stage.To solve the above issues,research on smoke source location algorithms implemented by wheeled mobile robots in a variety of indoor environments is conducted in this thesis.The main achievements can be concluded as follows:To conduct robotic experiment with related algorithms,two smoke source localization robots are built,mounting a particulate matter sensor,Plantower PMS 7003,selected from numerous similar products.The sensing modalities of the particulate matter sensor are studied comprehensively.Two environment sites,including a wind tunnel and a laboratory,are deployed for systematic robotic experiments.To estimate the location of the smoke source in an empty indoor environment with known arguments,we proposed an Infotaxis-based smoke source localization algorithm and some variants with different probability map integration strategies,based on the dispersion character of smoke particles and the multiple sensing modalities of the particulate matter sensor.Simulations showed that the unweighted multi-modality map strategy is of the best performance.The experimental campaign in the wind tunnel under various environment conditions showed that the median localization error can be lower than 1 m.To tracking smoke plumes in an empty indoor environment with unknown arguments,a local perception window particle filter-based smoke plume path tracking algorithm is proposed,based on a general smoke plume path model.The algorithm can guide a mobile robot to track the smoke plume path in a windy environment by searching for the most information gain at every step.Through a modified firefly algorithm,the bio-inspired anemotaxis behavior is introduced in the proposed algorithm.This adaptation enhances the success rate of the algorithm and make it less computationally expensive.In the simulated wind tunnel,the average source localization error is0.78 m.With the proposed algorithm,the robot can also track the smoke plume in the laboratory,in which the wind field is not uniformly distributed.The results show that the success rate can be up to 90%.To search for the smoke source in a complex indoor environment with obstacles inside or in a multiple-room environment,a Deep Q-Network based algorithm is proposed,utilizing a deep convolutional neural network.A simulated environment dedicated to the smoke source localization task was built,and a smoke hits distribution model is proposed to train the Deep Q-Network to guide the robot take the proper action under various states,maximizing the cumulative future reward.Simulations in various complex indoor environments in the high-fidelity simulator,GADEN demonstrate that the proposed algorithm can guide the robot to the smoke source,while avoiding obstacles.The average distance overhead is 1.502.At last,a conclusion of this thesis is presented,and the future research work is also prospected.Achievements in this thesis can be used to locate burning smoke sources in indoor environments with various levels of complexity,such as houses,workshops,laboratories,warehouses.The Infotaxis-based algorithm is suitable for empty indoor environment with known arguments.The local perception window particle filter-based algorithm is suitable for empty indoor environment with unknown arguments.The Deep Q-Network based algorithm is suitable for complex indoor environment with obstacles inside or multiple-room environment.