A Data-Driven Dynamic Data Fusion Method Based on Visibility Graph and Evidence Theory

Gang Liu, Fuyuan XiaoIEEE Access 2019 · DOI

Introduction

In complex environments such as robotics, sensor networks, and defense systems, identifying a target often requires fusing a continuous stream of uncertain and potentially conflicting data. Dempster-Shafer (D-S) evidence theory has emerged as a powerful mathematical framework for this task due to its ability to handle "unknowns" and conflicting reports. However, traditional approaches to dynamic data fusion often struggle with temporal relevance.

Standard Credibility Decay Models (CDM) typically apply a mathematical "discount" to older data, assuming that information "rots" over time. These models are often highly subjective, requiring researchers to manually tune decay parameters. Furthermore, they frequently over-prioritize the most recent data point, potentially discarding valuable structural patterns embedded in the historical time series. To address these issues, this research proposes a data-driven approach that leverages the structural properties of data over time.

Methodology: The SRCDM Pipeline

The authors introduce the Structure Revision Credibility Decay Model (SRCDM), a multi-step pipeline designed to automatically determine the trustworthiness of data based on its underlying structure rather than just its age.

Visibility Graph (VG) Transformation

The core innovation of SRCDM is the use of Visibility Graphs to capture the geometry of a time series. A time series is converted into a graph where each data point serves as a node. Two nodes are connected if they can "see" each other—meaning no intermediate data point's value is high enough to block the line of sight between them. This transformation allows the system to analyze the data's structure using network theory.

Visibility Graph Transformation: Mapping a time series to a network structure.

Weighting and Aggregation

Following the transformation, the system employs a VGA Operator to calculate weights for past data based on "node degree" (the number of connections a point has). This provides an objective measure of a data point's importance within the historical context. These weights are then used to form a Structure Revision Basic Belief Assignment (SRBBA), which aggregates past beliefs into a single "memory" reflecting the historical structure.

Data-Driven Discounting and Fusion

Instead of a fixed, subjective decay rate, the model calculates the Jousselme distance between the incoming evidence and the SRBBA. If the new data aligns with the historical structure, it is assigned higher credibility; if it appears to be an outlier or interference, its influence is automatically discounted. Finally, this discounted evidence is fused using standard D-S combination rules to produce a final target identification.

Experimental Evaluation

The SRCDM was evaluated across three distinct scenarios:

  1. Numerical Simulation: Testing robustness against artificial interference at specific time steps.
  2. Target Change Analysis: Assessing the model's sensitivity and adaptation speed when the tracked target switches (e.g., from Target A to Target B).
  3. Iris Dataset Application: Using the UCI Iris flower dataset to simulate a real-world classification task by transforming attributes into evidence streams.

Key Results

The experimental results demonstrate significant improvements over traditional models:

  • Robustness: In simulations involving high-interference data, SRCDM remained accurate while traditional CDM and Ordered Weighted Averaging (OWA) models were led astray.
  • Success Rates: In identification tasks with a high confidence threshold ($\kappa = 0.8$), SRCDM achieved a 50% success rate, compared to just 20% for the traditional CDM.
  • Accuracy: On the Iris dataset, SRCDM achieved an average Pignistic Probability (certainty measure) of 0.8016, outperforming the CDM's 0.6895.
  • Stability: The model proved more stable during long or irregular time intervals, as it relies on structural relationships rather than simple exponential time-decay.

Discussion and Limitations

By prioritizing structural relationships over simple recency, the SRCDM provides a parameter-free fusion method that removes the need for subjective tuning. However, the authors note some limitations. The model currently requires a predefined "Frame of Discernment" (a known list of possible targets). Additionally, while it simplifies the fusion process, it still requires storing past belief assignments to construct the SRBBA, posing potential memory challenges for extremely long-duration tasks.

Future research will focus on enabling the model to identify the number of targets automatically without prior knowledge and optimizing performance for complex, multi-target environments.

Summary updated

BibTeX
@article{liu2019data,
  author = {Gang Liu and Fuyuan Xiao},
  title = {{A Data-Driven Dynamic Data Fusion Method Based on Visibility Graph and Evidence Theory}},
  journal = {IEEE Access},
  year = {2019},
  doi = {10.1109/access.2019.2931951},
}
Figure