A Data Fusion Method in Wireless Sensor Network Based on Belief Structure

Chen‐Feng Long, Xinxing Liu, Yakun Yang, Tao Zhang, Siqiao Tan, Xiaoyong Tang, Gelan YangResearch 2020 · DOI

TL;DR

This paper introduces a novel data fusion method for Wireless Sensor Networks (WSNs) that slims down redundant sensor data by combining multi-granulation rough set theory with belief structures. By identifying and removing unnecessary data attributes, the proposed algorithm significantly reduces communication overhead and energy consumption while maintaining a high degree of information accuracy.

What Problem Does This Paper Address?

Wireless Sensor Networks are often deployed in overlapping patterns to ensure full coverage of an area. While this provides a safety net, it creates a massive headache: data redundancy. Multiple sensors often report the same or highly similar information, leading to high circuit complexity, wasted bandwidth, and rapid battery depletion—a death sentence for remote sensor nodes.

The challenge is "data fusion"—the process of integrating data from multiple sources to produce more accurate or efficient information. Specifically, the authors tackle attribute reduction: how do we determine which pieces of data are essential and which can be safely discarded without losing the "truth" of what the sensors are monitoring?

Methods

The authors bridge two mathematical frameworks to solve this:

  1. Multi-granulation Rough Set Theory: This allows researchers to analyze data at different "scales" or levels of detail (e.g., recording a temperature as an exact integer vs. a category like "Hot" or "Cold").
  2. Belief Structures (Dempster-Shafer Theory): This framework handles uncertainty by using "belief" and "plausibility" functions to quantify how much we trust a particular set of data.

The Innovation: The paper proposes three types of "reductions" (ways to prune data) within an incomplete, ordered information system: positive region reduction, belief reduction, and plausibility reduction.

The core of the paper is Algorithm 1, which calculates the "significance degree" of each attribute. If an attribute doesn't significantly change the "belief" or "positive region" of the data, it is deemed dispensable and removed. Crucially, the authors prove that positive region reduction and belief reduction are equivalent in this context, simplifying the computational requirements.

Datasets

The authors demonstrate their method using a specific case study: Rape pests detected by WSN.

  • The Setup: 8 clusters of sensors monitoring pests like Cabbage butterflies and Aphids.
  • Attributes: 4 specific attributes (pest types/counts) analyzed across 3 levels of granularity (e.g., raw counts vs. grades like "rarely" or "abundance").
  • Characteristics: The dataset is "incomplete," meaning it contains missing values (represented by *), which is a common reality in real-world sensor deployments.

Key Insights and Ideas

  • Mathematical Equivalence: One of the "aha" moments is the proof that "Positive Region Reduction" and "Belief Reduction" are equivalent. This means engineers don't have to calculate both; solving for one automatically satisfies the other, saving precious processing cycles.
  • Transitivity Across Scales: The paper explores how data reduction at a fine-grained level (highly detailed) relates to a coarse-grained level. They found that if the "cover structure" remains the same, the reductions are transitive, allowing for consistent data fusion across different hierarchical levels of a network.
  • Efficiency is King: Traditional rough set methods often have high computational complexity. This method focuses on a linear-style complexity relative to the number of attributes and objects, making it feasible for the limited CPUs found in sensor nodes.

Key Results

  • Data Redundancy Reduction: In the pest-detection example, the method successfully reduced data redundancy by 50% through attribute reduction.
  • Computational Speedup: The proposed algorithm achieves a time complexity of $O(m \times n)$, where $m$ is the number of attributes and $n$ is the number of objects. This is a significant improvement over traditional rough set measures, which often hit $O(m \times n^2)$ or $O(n^2)$.
  • Reliability: The method maintains the "consistency" of the information even after pruning, ensuring that the fused data is still a reliable representation of the environment.

Limitations and Future Work

The authors identify two primary drawbacks:

  1. Discretization Requirement: The algorithm requires continuous data to be discretized (broken into categories/intervals) before processing, which can sometimes lead to information loss if not handled carefully.
  2. Efficacy Threshold: The method loses its edge when the number of reducible attributes is very low; if the data is already lean, the overhead of the algorithm might not be worth the minimal gains.

Future Work: The team plans to focus on the synthesis of evidence within multi-granulation ordered systems to further refine how WSNs handle complex, multi-layered data fusion.

Summary updated

Abstract

Abstract Considering the issue with respect to the high data redundancy and high cost of information collection in wireless sensor nodes, this paper proposes a data fusion method based on belief structure to reduce attribution in multi-granulation rough set. By introducing belief structure, attribute reduction is carried out for multi-granulation rough sets. From the view of granular computing, this paper studies the evidential characteristics of incomplete multi-granulation ordered information systems. On this basis, the positive region reduction, belief reduction and plausibility reduction are put forward in incomplete multi-granulation order information system, and analyze the consistency in the same level and transitivity in different levels. The positive region reduction and belief reduction are equivalent, and the positive region reduction and belief reduction is unnecessary and sufficient conditional plausibility reduction in the same level; if the cover structure order of different levels are the same, the corresponding equivalent positive region reduction. The algorithm proposed in this paper not only performs three reductions, but also reduces the time complexity largely. The above study fuses the node data which reduces the amount of data that needs to be transmitted and effectively improves the information processing efficiency.
BibTeX
@inproceedings{long2020data,
  author = {Chen‐Feng Long and Xinxing Liu and Yakun Yang and Tao Zhang and Siqiao Tan and Xiaoyong Tang and Gelan Yang},
  title = {{A Data Fusion Method in Wireless Sensor Network Based on Belief Structure}},
  booktitle = {Research},
  year = {2020},
  doi = {10.21203/rs.3.rs-96349/v1},
}
Figure