Research Of Key Technologies In Wireless Indoor Localization Based On Deep Learning

In the recent years,WiFi fingerprint based indoor localization technology becomes a major research area with personal mobile devices.The studies on wireless indoor localization systems focus on two main ways.The one is to derive the specific position coordinates by finding the relations of equations based on some strict assumptions such as line-of-sight propagation path.The way is too ideal to be applied to many other tasks.The other one is to solve the existing problems.By modeling the localization with deep learning algorithm,the localization system can mine the potential features,and thus can weaken the harsh conditions of indoor localization to improve the performance of the system.Above all,wireless indoor localization technology has gained more attention from all walks of life.It is very important for us to develop a practical and universal localization system.In this paper,from the perspective of deep learning algorithm,the main studies on solving the lack of samples and the low quantity of samples are conducted.Main topic of the paper is a key wireless indoor localization technology based on deep learning algorithm.At first,the indoor localization algorithms are introduced based on traditional machine learning and deep learning.A localization algorithm based on neural networks is studied.It can train a model with many parameters in the neural networks and can predict the specific position.However,the disadvantage of the algorithm is that it cannot promise a good result.So a hierarchical indoor wireless indoor localization algorithm is proposed in order to solve the above problems.Its goal is to predict the building and floor,and then to predict the position.The advantage of it is that the variance of localization can be reduced in the practical application.In the second,the localization framework based on data augmentation is proposed.It can also be used to predict the building and floor.The goal of the framework is to increase the samples.There are two main modules of the framework.The one is augmentor and the other is trainer.The augmentor,consisting of a generator and a discriminator,can produce much position data based on the existing data.The discriminator is first trained by deep learning and then fine-tuned by deep neural networks in which data is derived by the generator.For the learning of the generator,we train it with the help of discriminator which gives reward.The trainer,built as a deep model with the original data and the generated data,can estimate the accurate localization.At last,deep neural network based on transfer learning is proposed to solve the problems such as the low quantity of training data and the long time of training.The algorithm can train the model with source domain training data,and then fix the components of the model.Finally,the knowledge of the fixed components can be transferred to other new tasks.According to the testing data,the algorithm can be used to extract the shared features of data.Moreover,performance of the overall model is improved because of the generalization.