Please use this identifier to cite or link to this item:
|Title:||Wi-fi-based indoor positioning by distributed machine-learning data analytics on smart gateways||Authors:||Cai, Yuehua||Keywords:||DRNTU::Engineering::Electrical and electronic engineering||Issue Date:||2016||Source:||Cai, Y. (2016). Wi-fi-based indoor positioning by distributed machine-learning data analytics on smart gateways. Master's thesis, Nanyang Technological University, Singapore.||Abstract:||A smart building management system (SBMS) can achieve building automation and energy optimization through control of an interconnected network of physical components. Indoor positioning system (IPS) is a critical input to optimize the operational performance of a building. However, to perform real-time data analytics for indoor positioning is challenging due to limited computing resources on SBMS Internet-of-things (IoT) platform, i.e., smart gateways. This thesis studies a Wi-Fi-based IPS that is specifically designed for smart gateways and explores the performance improvement with proposed distributed machine learning algorithms. Conventional machine learning approaches usually follow centralized data analytic architectures with all the computation performed on a central server. As indoor environments may change over time, off-line trained predictors may be out-of-date due to different scenarios between training and prediction phases. As such, the initial machine learning predictor for positioning becomes less precise and requires real-time update with parameters fused from all different locations. However, limited computing resources on smart gateways make real-time data analysis challenging and a changing environment where received signal strength (RSSI) varies with time makes it more complex to update position predictors for real-time indoor positioning. There is a need to develop an efficient distributed machine learning on computationally resource limited gateways. In this thesis, an indoor positioning system structure is proposed based on smart gateways, which are designed as IoT platforms for smart buildings. Compared to a positive positioning by the user himself, a passive positioning is applied thus the indoor occupancy profile can be monitored. Besides, to solve the side effects caused by unstable received signal strength indicator (RSSI) values, a positioning algorithm is introduced with an integration of Kalman Filter, Dijkstra’s algorithm and machine learning algorithms, which will smoothen signal variation and improve positioning precision. The IPS is constructed in NTU-Virtus lab environment and with functionality including real-time tracking and indoor occupancy profile monitoring. Experimental results show that the proposed IPS can achieve a positioning accuracy of 2.9m with 88% precision. To overcome the positioning precision decrease cased by out-of-date predictor because of unstable indoor environment and Wi-Fi signal, we need to improve the system’s performance with a fast training process and real-time positioning predictor. In this thesis, distributed machine learning data analytics are researched and distributed machine learning algorithms are developed for the IPS. Each gateway performs its own data storage and processing portion, and then fuses the data with an overall RSSI profile model to report the indoor positioning with the ability to predict and update. The developed distributed machine learning algorithms are with multi-category classification and they are developed on the computationally resource limited gateway network such that the computation can be evenly allocated among the gateways according to the proposed strategies, while the conventional approaches are not suitably adapted for the gateway network. Firstly, based on the distributed collection and analytics of RSSI values in a gateway network, a time-efficient workload-based (WL) distributed support vector machine (WL-DSVM) algorithm is introduced to perform the indoor positioning. The experimental test-bed is at NTU-Virtus lab and experimental results show that with 5 distributed sensor nodes running in parallel, the proposed WL-DSVM can achieve a 3.2x performance improvement compared with centralized SVM on a single node. When comparing to conventional sequential distributed support vector machine (SQ-DSVM), a 27% performance improvement in computational time is achieved with developed WL-DSVM. Indoor environment simulation results show that the proposed DSVM solution is also promising in maintaining stable positioning precision, while traditional solution results in decreasing precision with a changing environment. Secondly, computationally efficient soft-voting distributed extreme learning machine (SV-DELM) based data analytics is also studied. In this part, by adopting incremental Cholesky factorization, a low complex training algorithm on smart gateway networks is developed to reduce the time for training and positioning process. And a DELM architecture with maximum posteriori probability based soft-voting is designed to improve the accuracy of positioning. Comparison between SV-DELM and WL-DSVM shows that both of them have advantages over traditional machine learning and they can be applied separately based on different situations and requirements. SV-DELM is more efficient in computational time while DSVM has advantages in stability, flexibility and prediction precision. Thirdly, a scalable distributed support vector machine (S-DSVM) is developed for the purpose of realizing high scalability for indoor positioning with distributed gateways. Different from WL-DSVM/SV-DELM where one master node and several slave nodes are defined, each smart gateway under S-DSVM constructs local predictors and no master node is required. In WL-DSVM/SV-DELM, if some nodes are broken and no enough dimensions of input data, the training/predicting process will get stuck, which greatly challenges the system’s scalability. S-DSVM only needs each gateway to use local datasets and positioning accuracy can be enhanced with active joined nodes. Experimental results show that scalability of the system can be improve by S-DSVM with 3x from WL-DSVM and 9x from SV-DELM when stability of each gateway is 80%. For S-DSVM, with a weighted scheme applied, computational efficiency can be improved to 1.7x with a sacrifice of 2m positioning precision compared to centralized SVM. In summary, the main contribution of this thesis can be summarized as follows. Firstly, a practical indoor positioning system is developed and integrated with smart building management system. Secondly, a reliable indoor positioning algorithm is developed with an acceptable positioning precision for indoor environment. Thirdly, 3 types of distributed machine learning algorithms are developed and integrated in positioning system, and their performances are analyzed and compared. Results show that they can be implemented according to different requirements and in different situations. As a conclusion, the developed indoor positioning system with distributed machine learning have shown a great potential in maintaining positioning accuracy & precision and proved to be reliable to be implemented in smart buildings in the near future.||URI:||http://hdl.handle.net/10356/68865||Fulltext Permission:||restricted||Fulltext Availability:||With Fulltext|
|Appears in Collections:||EEE Theses|
Updated on May 12, 2021
Updated on May 12, 2021
Items in DR-NTU are protected by copyright, with all rights reserved, unless otherwise indicated.