Elsevier

Neurocomputing

Volume 166, 20 October 2015, Pages 282-293
Neurocomputing

Semi-supervised deep extreme learning machine for Wi-Fi based localization

https://doi.org/10.1016/j.neucom.2015.04.011Get rights and content

Abstract

Along with the proliferation of mobile devices and wireless signal coverage, indoor localization based on Wi-Fi gets great popularity. Fingerprint based method is the mainstream approach for Wi-Fi indoor localization, for it can achieve high localization performance as long as labeled data are sufficient. However, the number of labeled data is always limited due to the high cost of data acquisition. Nowadays, crowd sourcing becomes an effective approach to gather large number of data; meanwhile, most of them are unlabeled. Therefore, it is worth studying the use of unlabeled data to improve localization performance. To achieve this goal, a novel algorithm Semi-supervised Deep Extreme Learning Machine (SDELM) is proposed, which takes the advantages of semi-supervised learning, Deep Leaning (DL), and Extreme Learning Machine (ELM), so that the localization performance can be improved both in the feature extraction procedure and in the classifier. The experimental results in real indoor environments show that the proposed SDELM not only outperforms other compared methods but also reduces the calibration effort with the help of unlabeled data.

Introduction

Wireless localization based on Wi-Fi is quite popular, especially in indoor environment [1], [2], [3], for it does not need deploying any extra infrastructure. What is more, the widespread Access Points (APs) and smart mobile devices facilitate the development of Wi-Fi based indoor localization. To implement indoor localization, traditionally, labeled data are required. In Wi-Fi localization field, labeled data means both data and their corresponding locations (coordinates/classes) are known; unlabeled data means only data are available, and whose corresponding locations are unknown. Generally, there are mainly two types of Wi-Fi localization methods: propagation based method and fingerprint based method. The propagation based method takes the advantage of the nonlinear fading characteristics of wireless signal to set up a propagation model [1]. This kind of method can be easily implemented; however, the localization performance is not good enough for it is difficult to set up an accurate propagation model in a complex and dynamic indoor environment. Compared with propagation based method, fingerprint based localization method is widely adopted. “Fingerprints” are features obtained by feature extraction methods, and used to represent the corresponding locations; “localization” means adopting pattern recognition method to estimate the location. Theoretically, the more the labeled data are, the better the localization performance will be. However, the calibration procedure (which is also known as the acquisition of labeled data) is always at the cost of time, man-hour and money, which leads to the limited number of labeled data. According to [4], “calibration procedures are applied in a great number of proposed location techniques and are considered to be not practical or a considerable barrier to wider adoption of such methods”. As reported [5], Ekahau [6], a commercial real-time localization system, spent $10,000 just to collect labeled data in a large office building, which clearly verifies the high cost of obtaining labeled data.

In contrast to the acquisition of labeled data, the collection of unlabeled data can be carried out easily; especially when crowd sourcing is adopted. Crowd Sourcing [7], [8] is a distributed model to solve problems through an open way with different participants. For indoor localization, crowd sourcing (data collection) means gathering data from heterogeneous devices [9], and most of the data are unlabeled. In order to use unlabeled data, works based on manifold learning and semi-supervised learning were proposed. These works indeed improved localization performance with unlabeled data; however, they paid less attention to the problem caused by crowd sourcing.

After data have been collected for localization system, feature extraction becomes an important step. Though Received Signal Strength (RSS) of Wi-Fi signal has real physical meaning and can be directly used as feature; the highly dynamic indoor environment and heterogeneous devices lead to severe fluctuation of wireless signal, which makes this kind of direct feature less representative. Traditional indoor localization methods mainly selected features manually or extracted features by shallow networks, which cannot reflect Wi-Fi characteristics in complex environment well. However, a machine learning method, Deep Learning (DL) has been famous worldwide since 2006 [10]. DL learns high level features and distributed data structure, which can represent original data better than shallow feature. Though DL outperforms traditional neural networks, its training time consumption is high. While, Extreme Learning Machine (ELM) proposed in the same period as DL, is popular for fast learning speed, which can make up the time-consuming shortage of DL.

Considering the analysis above, we put forward a novel localization method, Semi-supervised Deep Extreme Learning Machine (SDELM) to improve the localization performance with unlabeled data. The contribution of this work is:

  • 1)

    Utilize unlabeled data to get discriminative features and better classification ability;

  • 2)

    Propose semi-supervised embedding for deep leaning network;

  • 3)

    Adopt modified ELM to improve the learning speed of SDELM.

The rest of this paper is organized as follows: Section II shows the related works in Wi-Fi localization field. Section III introduces the proposed SDELM in detail. Section IV evaluates the performance of SDELM in real wireless indoor environments. And Section V concludes the work.

Section snippets

Related works

The mainstream approach of Wi-Fi based indoor localization, fingerprint based method [2], [3], consists of two phases: off-line training phase and online locating phase. At training phase, features are extracted according to a certain rule, so that the relationship between features and their corresponding locations can be established, which forms a “radio map”; at online locating phase, features of unknown location are extracted with the same rule, and pattern recognition methods are used to

Methodology--Semi-supervised deep extreme learning machine

Graph based semi-supervised learning is embedded to deep learning structure to get discriminative feature and better classification boundary. To guarantee the learning speed, semi-supervised learning is evolved from ELM. Hence, in this section, we firstly briefly introduce graph based semi-supervised learning, deep leaning, and ELM; then elaborate the proposed SDELM in detail.

Performance Evaluation

In this section, performance of proposed SDELM will be validated from four aspects: feature representation, localization accuracy, time consumption and man-hour cost in real indoor environments.

Conclusion

Wireless localization system based on Wi-Fi faces a practical problem that the number of labeled data is limited. In order to fully utilize a large number of unlabeled data, SDELM method is proposed, which uses unlabeled data to improve localization performance from feature extraction to pattern recognition (classifier). Semi-supervised learning embedded deep learning can not only extract features with high discriminability but also improve the classification ability with the help of unlabeled

Acknowledgement

This work is supported by Natural Science Foundation of China (No.61173066, No.61070110), and Beijing Natural Science Foundation (No.4144085).

Yang Gu received the B.A.Sc. from Beijing University of Posts and Telecommunications, China in 2010. Currently, she is a Ph.D student in the Institute of Computing Technology, Chinese Academy of Sciences (ICT, CAS). Her current research interests include machine learning, indoor localization and pervasive computing.

References (42)

  • G.B. Huang et al.

    Extreme learning machine: Theory and applications

    in Neurocomputing

    (2006)
  • P. Bahl et al.

    RADAR: An in-building RF-based user location and tracking system

    in Proc. IEEE INFOCOM

    (2000)
  • V. Zheng et al.

    Transferring localization models over time

    in Proc. AAAI Conference on Artificial Intelligence

    (2008)
  • Q. Yang et al.

    Estimating location using Wi-Fi

    IEEE Intelligent Systems

    (2008)
  • L. Moraes et al.

    Calibration-free WLAN location system based on dynamic mapping of signal strength

    in MobiWAC

    (2006)
  • J. Ledlie, J. Park, D. Curtis and et al. “Molé: A scalable, user-generated WiFi positioning engine,” in 2011...
  • Ekahau positioning engine. Available from:...
  • G. Chatzimilioudis et al.

    Crowdsourcing with smartphones

    in IEEE Internet Computing

    (2012)
  • D. Zeinalipour-Yazti et al.

    Crowdsourced trace similarity with smartphones

    in IEEE Trans. on Knowledge and Data Engineering

    (2013)
  • C. Laoudias, D. Zeinalipour-Yazti, C. Panayiotou, “Crowdsourced Indoor Localization for Diverse Devices through...
  • G.E. Hinton et al.

    Reducing the Dimensionality of Data with Neural Networks

    in Science

    (2006)
  • A. Haeberlen et al.

    Practical robust localization over large-scale 802.11 wireless networks

    in ACM MOBICOM’04

    (2004)
  • K. Kaemarungsi, “Distribution of WLAN received signal strength indication for indoor location determination,” in 1st...
  • K. Kaemarungsi and P. Krishnamurthy, “Properties of indoor received signal strength for WLAN location fingerprinting,”...
  • Y. Chen et al.

    Sensor-sssisted Wi-Fi indoor location system for adapting to environmental dynamics

    in MSWiM ‘05

    (2005)
  • P. Bahl,V.N. Padmanabhan, “Enhancements of the RADAR user location and tracking system,” Microsoft Corp., Redmond, WA,...
  • J. Zhu et al.

    Equilateral triangle localization algorithm based on average RSSI

    in Journal of Northeast University (Natural Science)

    (2007)
  • S. Fang et al.

    Principal component localization in Indoor WLAN environments

    in IEEE Transaction on Mobile Computing

    (2012)
  • M. Youssef, A. Agrawala, and A.U. Shankar, “WLAN location determination via clustering and probability distributions,”...
  • Y. Chen et al.

    Power-efficient access-point selection for indoor location estimation

    in IEEE Trans. Knowledge and Data Eng.

    (2006)
  • M. Kjaergaard, C. Munk, “Hyperbolic location fingerprinting: a calibration-free solution for handling differences in...
  • Cited by (90)

    • An accuracy-maximization learning framework for supervised and semi-supervised imbalanced data

      2022, Knowledge-Based Systems
      Citation Excerpt :

      To utilize a large amount of unlabeled data as well as a relatively limited amount of labeled data for better classification, semi-supervised learning methods [19] was proposed and demonstrated promising performances in various tasks. Semi-supervised ELM variants have also been proposed [4–9] to specifically tackle the labeled data scarcity problem in ELM learning. However, most of the existing semi-supervised ELM variants can only support learning on balanced datasets but not imbalanced ones, which relates to another real-world challenge facing ELM — class imbalance issues.

    • A fingerprint technique for indoor localization using autoencoder based semi-supervised deep extreme learning machine

      2021, Signal Processing
      Citation Excerpt :

      Semi-supervised Deep Extreme Learning Machine (SDELM) is the other ELM based method, which has the advantages of semi-supervised learning and deep learning [26]. The authors of [26] show that SDELM has 25% success rate by using 500 labeled data and 3000 unlabeled data. However, because of the randomness of input weight matrix and bias of ELM, it cannot extract high level features from the input data.

    • Deep Learning for IoT

      2024, Internet of Things
    • Channel state information-based wireless localization by signal reconstruction

      2023, Eurasip Journal on Wireless Communications and Networking
    • Mechanisms, algorithms, implementation and perspectives of brain⁃ inspired navigation

      2023, Hangkong Xuebao/Acta Aeronautica et Astronautica Sinica
    View all citing articles on Scopus

    Yang Gu received the B.A.Sc. from Beijing University of Posts and Telecommunications, China in 2010. Currently, she is a Ph.D student in the Institute of Computing Technology, Chinese Academy of Sciences (ICT, CAS). Her current research interests include machine learning, indoor localization and pervasive computing.

    Yiqiang Chen received Ph.D. degree from the Institute of Computing Technology, Chinese Academy of Sciences (ICT, CAS), Beijing, China, in 2002. In 2004, he was a Visiting Scholar Researcher at the Department of Computer Science, Hong Kong University of Science and Technology (HKUST), Hong Kong. Currently, he is a Professor and Director of the pervasive computing research center at ICT, CAS. His research interests include artificial intelligence, pervasive computing, and human computer interface. He is a member of IEEE.

    Junfa Liu received the Ph.D. degree in the Institute of Computing Technology, Chinese Academy of Sciences (ICT, CAS) in 2009. Currently, he is a vice Professor in ICT, CAS. His current research interests include machine learning, computational intelligence, human computer interaction and pervasive computing. He is a member of IEEE.

    Xinlong Jiang, received the B.A.Sc. from Beijing University of Posts and Telecommunications, China in 2011. Currently, he is a Ph.D student in the Institute of Computing Technology, Chinese Academy of Sciences (ICT, CAS). His current research interests include indoor localization and pervasive computing.

    View full text