Research On Gas Source Localization Based On Neural Networks And Evolutionary Algorithms

Traditional methods for gas source localization are usually designed based on a gas diffusion model combined with probabilistic estimation theory.However,in practical monitoring environments,it is difficult to establish an accurate gas diffusion model because the gas source diffusion is mostly influenced by random turbulent flows.In order to localization of the gas source rapidly,it is essential to identify the leaking gas immediately.Traditional methods for gas identification are mainly based on gas stability characteristics or maximum response,which usually take a long time.Moreover,these traditional methods highly dependent on gas sensor responses,while the difficulty lies identifying very low response from low concentration of the leaking gas.In order to address the above problems,this thesis firstly proposes a fast identification algorithm for low concentration gases,and then proposes a fast gas source localization method in large scale 3D space based on data-driven ideas,combining static sensor networks and neural networks,and finally proposes an accurate gas source localization method in small space based on moving sensors and evolutionary algorithms.A new integrated gas localization method is designed,and the specific research results are:1.The first objective is to investigate a fast identification algorithm for low concentration gases based on convolutional and recurrent neural networks.Firstly,in order to mitigate the performance drift present in mainstream gas-sensitive devices,a generalisation step is used as a data pre-processing technique to extract hidden discrepancy information.Then the convolutional kernel structure and size are optimised to fit the structure of gas sensor data,and attention mechanism and Bayesian optimisation algorithm are added to enable the model to make full use of the initial curve of sensor response,shortening the classification time from 4s to 0.5s,and achieving 91%classification accuracy for gases within 125 ppm.We eventually achieved fast identification of low concentration gases,and the related results was published.2.The second objective is to design a three-dimensional spatial gas source localization method,based on static sensor networks and neural networks.A data preprocessing method based on the structural features of the static sensor network is designed to convert the sensor data at each sampling moment into a matrix similar to a monochrome image,which is similar to the actual position of the sensor.The data in continuous time are converted into a tensor,which retains spatial characteristics of the original data without affecting the temporal continuity of the data.For data labelling,a multi-label classification method is used,which divides the three-dimensional space into numerous small spaces,converting the positions into a multi-label list,and the labels are independent of each other,making them suitable for the localization of single and multiple air sources.In the method,two layers of convolution are used as filters to automatically extract features from the input tensor,and long and short memory neurons are added after the convolution layer to extract sequence information from the instantaneous response sequence.The method uses only the first 75 s of gas diffusion and achieves a localization accuracy of 91%.The result shows that fast localization of single and multiple gas sources in large scale 3D space is achieved.3.The third objective is to design a gas source localization method based on mobile sensors and evolutionary algorithms.Firstly,a genetic algorithm is used for local search,which increases the range and randomness of the search and speeds up the search speed.Then,an ant colony algorithm is used for global search,and a parameter adaptive strategy is designed for the slow convergence of the ant colony algorithm.The result realized an average positioning time of 217 s and an average positioning error of 0.366 m,demonstrating that an accurate positioning of gas sources in small spaces is achieved.