Exploiting Linear Support Vector Machine for Correlation-Based High Dimensional Data Classification in Wireless Sensor Networks

Lawrence Mwenda Muriira, Zhiwei Zhao, Geyong Min2018 · DOI

TL;DR

This paper introduces a data-driven framework using Kernelized Linear Support Vector Machines (KLSVM) to reliably classify links in Wireless Sensor Networks (WSNs). The approach effectively handles high-dimensional data, captures spatiotemporal correlations, and detects network anomalies, leading to improved link quality estimation and fault detection.

What Problem Does This Paper Address?

Wireless Sensor Networks (WSNs) are increasingly deployed for monitoring physical phenomena, but their dynamic environments and limited resources make reliable communication a significant challenge. Data generated by these networks is often high-dimensional and exhibits complex correlations (temporal, spatial, and spatiotemporal). Accurately classifying the quality of links between sensor nodes and detecting anomalies (like faulty links or outliers) is crucial for maintaining network performance and ensuring the integrity of collected data. Existing methods often struggle with the high dimensionality and the need for efficient, accurate classification and anomaly detection in these resource-constrained environments.

Methods

The core of this work is the application and adaptation of a Kernelized Linear Support Vector Machine (KLSVM). The authors propose a data-driven framework that leverages KLSVM for:

  • High-Dimensional Multi-Category Classification: KLSVM is used to classify link quality into multiple categories (e.g., "Very Good Link" to "Faulty Link"). The linear kernel is chosen for its efficiency, especially when dealing with high-dimensional data, and the "kernel trick" allows it to implicitly map data into a higher-dimensional space where it might be linearly separable.
  • Spatiotemporal Data Correlation: The framework analyzes correlations between different link features (like Packet Reception Rate (PRR), retransmission count, and CPDF) over time and space. This helps in understanding how link behavior evolves and how different metrics relate to each other.
  • Anomaly Detection: KLSVM is employed to identify outliers and faulty links within the network. The model's ability to learn decision boundaries helps in distinguishing normal behavior from abnormal patterns.

The authors detail the mathematical formulation of SVM, including the primal and dual forms, the optimization process using Lagrange multipliers, and the handling of linearly non-separable data through slack variables. They specifically focus on the linear kernel due to its computational efficiency and suitability for the observed linearly correlated nature of WSN link data.

Datasets

The study utilizes a dataset gathered from a 20-sensor-node testbed deployed at the New Generation Mobile Internet Research Laboratory (NGMI) at the University of Electronic Science and Technology of China. The experiment involved broadcasting 400,000 packets with an inter-packet interval of 50 ms across 16 channels. The data collected includes packet traces from the 802.15.4 network. The authors also mention using a similar approach to [16] for beacon-based measurements (BCM) to capture received beacon messages.

Key Insights and Ideas

  • Linear Separability in WSN Link Data: The research suggests that WSN link quality data, when analyzed through specific features, tends to exhibit linear correlations, making a linear SVM approach effective. This is a key justification for choosing LSVM over more complex kernel methods.
  • KLSVM for Multi-faceted WSN Analysis: The framework demonstrates that a single KLSVM model can simultaneously address high-dimensional classification, spatiotemporal correlation analysis, and anomaly detection, offering a unified approach to WSN link management.
  • Correlation as a Predictor of Link Quality: The study highlights that strong correlations between link features (like PRR and retransmission count) are indicative of consistent link behavior and can be used for prediction and verification of link quality.
  • Outlier Detection via Spatial Correlation: The technique's attention to spatial correlation of neighboring nodes significantly improves the accuracy of outlier detection, distinguishing them from normal network behavior.

Key Results

  • The KLSVM model achieved good performance on link quality estimation accuracy.
  • The technique demonstrated precise anomaly detection of sensor nodes within the network.
  • The study identified four outliers and one faulty link with detailed descriptions of their behavior and causes (e.g., hardware failure for the faulty link).
  • Analysis of sensor node observations showed that links with PRR > 50% exhibit a high degree of negative correlation, while others show a moderate degree of positive correlation, providing visual intuition of network behavior.

Limitations and Future Work

  • Real-time Detection: The paper notes that the outlier detection technique, while accurate, is not performed in real-time.
  • Generalization of Linear Kernel: While the linear kernel is effective for the observed data, the paper acknowledges that for datasets that are not linearly separable, more complex kernel functions might be necessary.
  • Future Work: The authors suggest that this technique paves the way for a sizable number of possible domains that can benefit from it, implying potential applications beyond link classification and anomaly detection in WSNs. The unified approach to spatiotemporal correlation and high-dimensional classification in a single algorithm is presented as a significant contribution with broad applicability.

Summary updated

BibTeX
@misc{muriira2018exploiting,
  author = {Lawrence Mwenda Muriira and Zhiwei Zhao and Geyong Min},
  title = {{Exploiting Linear Support Vector Machine for Correlation-Based High Dimensional Data Classification in Wireless Sensor Networks}},
  year = {2018},
  doi = {10.3390/s18092840},
}

Loading related content...

Figure