A Weighted Combination Method for Conflicting Evidence in Multi-Sensor Data Fusion

Fuyuan Xiao, Bowen QinSensors 2018 · DOI

TL;DR

This paper introduces a novel weighted combination method for multi-sensor data fusion that effectively handles conflicting evidence. By incorporating a modified cosine similarity measure and belief entropy, the method assigns more appropriate weights to individual sensor readings, leading to improved accuracy in classification and fault diagnosis tasks.

What Problem Does This Paper Address?

Multi-sensor data fusion aims to combine information from various sensors to achieve a more robust and accurate understanding of a system or environment. However, a significant challenge arises when sensors provide conflicting or uncertain information. Traditional methods, like Dempster-Shafer evidence theory, can produce counter-intuitive results in such scenarios. Existing approaches to mitigate this often focus on modifying the combination rule itself or using simple averaging, which may not adequately account for the interplay between different pieces of evidence or the inherent information volume of each sensor's contribution. This paper addresses the need for a more sophisticated method that can intelligently weigh conflicting evidence to improve fusion accuracy.

Methods

The proposed method introduces a two-stage weighting process before applying Dempster's combination rule:

  1. Credibility Degree Calculation:
    • Modified Cosine Similarity Measure: This measure, adapted from Jiang et al., assesses the similarity between different pieces of evidence (represented as Basic Probability Assignments or BPAs). It considers factors like angle, distance, and vector norm to quantify how well one piece of evidence supports another. A higher similarity suggests less conflict.
  2. Support Degree: Based on the similarity matrix, the support degree for each piece of evidence is calculated by summing its similarities with all other pieces of evidence.
  3. Credibility Degree: The support degrees are then normalized to obtain the credibility degree for each piece of evidence. This reflects how much each piece of evidence is supported by the collective evidence.
  4. Information Volume Adjustment:
    • Belief Entropy (Deng Entropy): This measure quantifies the information volume or uncertainty within each piece of evidence. A higher belief entropy indicates more information.
  5. Modified Credibility Degree: The initial credibility degree is adjusted by multiplying it with a factor derived from the normalized information volume. This step ensures that evidence with higher information content receives more appropriate weighting.
  6. Final Weights: The modified credibility degrees are normalized to serve as the final weights for each piece of evidence.

These weighted pieces of evidence are then combined using a weighted average approach before being fused using Dempster's combination rule.

Datasets

The paper evaluates the proposed method using:

  • A numerical example: A synthetic dataset with five BPAs is used to illustrate the method's feasibility and effectiveness.
  • Iris dataset: A standard dataset for classification, used to demonstrate the method's application in data classification.
  • Motor rotor fault diagnosis data: Experimental data from sensors measuring vibration signals at different frequencies (1X, 2X, 3X) is used to validate the method's performance in fault diagnosis.

Key Insights and Ideas

  • Beyond Simple Averaging: The core idea is that not all sensor data is created equal, especially when conflicting. Simply averaging or using basic Dempster-Shafer can lead to errors.
  • Interplay and Self-Impact: The method uniquely considers both how pieces of evidence relate to each other (similarity) and the inherent information content of each piece (belief entropy). This dual consideration is key to handling conflict.
  • Belief Entropy as a Weighting Factor: Using belief entropy to measure the "information volume" of evidence and adjust its credibility is a novel way to dynamically weight sensor inputs. Evidence with more information (higher entropy) gets a more significant say.
  • Improved Handling of Conflict: The weighted approach demonstrably outperforms standard methods like Dempster's rule and even some other advanced techniques when dealing with highly conflicting evidence.

Key Results

The proposed method demonstrated superior performance in several applications:

  • Numerical Example: The method successfully identified the correct target with a high BPA value (97.13%) when dealing with conflicting evidence, outperforming other methods.
  • Iris Data Classification: The proposed method achieved higher accuracy in classifying Iris flower species, correctly identifying the "setosa" class with a belief value of 86.93%, surpassing other methods.
  • Motor Rotor Fault Diagnosis: The method accurately diagnosed fault types (F2) at different frequencies (1X, 2X, 3X) with higher belief values (e.g., 90.55% for F2 at 1X) compared to existing methods like Jiang et al.'s. It also showed reduced uncertainty in its predictions.

Limitations and Future Work

  • Computational Complexity: While not explicitly detailed as a limitation, the multi-step weighting process likely adds computational overhead compared to simpler methods.
  • Parameter Tuning: The modified cosine similarity measure involves a parameter ($\alpha$) that influences the distance impact. The paper doesn't extensively discuss optimal tuning strategies for this parameter across different applications.
  • Future Work: The paper suggests that the proposed method is effective for handling conflicting evidence and can be applied to various domains. Future work could explore its application in more complex, real-world scenarios and potentially optimize the computational efficiency.

Summary updated

BibTeX
@article{xiao2018weighted,
  author = {Fuyuan Xiao and Bowen Qin},
  title = {{A Weighted Combination Method for Conflicting Evidence in Multi-Sensor Data Fusion}},
  journal = {Sensors},
  year = {2018},
  doi = {10.3390/s18051487},
}
Figure