TL;DR
This paper introduces a novel "weighted belief entropy" to better quantify uncertainty in multi-sensor data fusion, especially within the Dempster-Shafer evidence theory framework. By considering not just the mass function but also the scale of the frame of discernment, this new measure reduces information loss and leads to a more accurate data fusion approach, demonstrated in both artificial data experiments and a real-world motor fault diagnosis application.
What Problem Does This Paper Address?
In many real-world applications, data comes from multiple sensors, each with its own level of uncertainty and reliability. Effectively combining this diverse information, known as multi-sensor data fusion, is crucial. The Dempster-Shafer (DS) evidence theory is a powerful tool for handling such uncertain information. However, a significant challenge within DS theory is accurately measuring the degree of uncertainty in sensor reports before fusion. Existing uncertainty measures often overlook important aspects of the information, leading to potential information loss and less reliable fusion results, especially when dealing with conflicting data. This paper aims to address this gap by proposing a more comprehensive uncertainty measure and a subsequent data fusion method that leverages this improved measure.
Methods
The core contribution of this paper is the proposal of a weighted belief entropy. This new measure builds upon Deng entropy, which itself is a generalization of Shannon entropy for the DS framework. The key innovation is the incorporation of two additional factors:
- The scale of the Frame of Discernment (FOD): This refers to the total number of possible events in the system.
- The relative scale of a proposition: This is the size of a specific event (proposition) compared to the entire FOD.
Mathematically, the weighted belief entropy $E_{wd}(m)$ for a mass function $m$ is defined as:
$E_{wd}(m) = \sum_{A \in \Omega} \frac{|A|}{|\mathcal{X}|} m(A) \log_2 \frac{m(A)}{|\mathcal{X}| - |A|}$
where $|A|$ is the cardinality of proposition $A$, $|\mathcal{X}|$ is the cardinality of the FOD, and $m(A)$ is the mass assigned to proposition $A$.
This weighted belief entropy is then integrated into a five-step multi-sensor data fusion approach:
- Information Modeling: Representing uncertain information using Basic Probability Assignments (BPAs) within the DS framework.
- Uncertainty Measure: Calculating the uncertainty of each BPA using the proposed weighted belief entropy.
- Weight Calculation: Assigning a weight to each Body of Evidence (BOE) based on its calculated uncertainty. Higher uncertainty leads to a lower weight, reflecting reduced reliability.
- Weighted Mass Functions: Adjusting the original mass functions by their calculated weights.
- Data Fusion: Combining the weighted mass functions using Dempster's rule of combination.
Datasets
The paper uses two main datasets for evaluation:
- Artificial Data: A synthetic dataset for a target recognition problem, adapted from a previous study [58], involving five sensors reporting on three potential targets (A, B, C).
- Real-World Data: Experimental data for fault diagnosis of a motor rotor, specifically vibration acceleration sensor data collected under different fault conditions (F1, F2, F3) and frequencies (Freq1, Freq2, Freq3), adopted from a previous study [16].
Key Insights and Ideas
- Beyond Mass Functions: Traditional uncertainty measures in DS theory often focus solely on the mass function. This paper highlights that the structure of the problem (the FOD's scale) also carries crucial information about uncertainty.
- Reduced Information Loss: By considering the FOD's scale, the weighted belief entropy captures more nuances of uncertainty, leading to less information loss during processing compared to measures like Deng entropy.
- Improved Conflict Handling: The proposed weighted belief entropy provides a more robust measure of uncertainty, which is essential for effectively managing and fusing conflicting sensor data.
- Weighting for Reliability: The method explicitly weights each sensor's contribution based on its calculated uncertainty, giving more influence to more reliable (less uncertain) sources.
Key Results
- Numerical Example (Artificial Data): The proposed method achieved the highest belief (98.95%) for the recognized target A in the target recognition experiment, outperforming other methods in terms of belief strength.
- Fault Diagnosis Application: In the motor rotor fault diagnosis task, the proposed method successfully identified the fault type (F2) and demonstrated a higher support degree for the diagnosis compared to a previous method [16]. The results were consistent with existing methods but showed improved performance in handling conflicting sensor reports.
- Uncertainty Measure Comparison: A graphical comparison in Figure 1 shows that the weighted belief entropy consistently increases with the complexity of the proposition, unlike some other measures that decrease or remain constant, indicating its ability to capture evolving uncertainty.
Limitations and Future Work
- Limitations: The paper acknowledges that the generation of BPAs from raw sensor data is a complex issue and not the primary focus of this work. The current approach assumes BPAs are already available.
- Future Work: The authors suggest that further study will focus on extending the proposed measure and data fusion approach to solve more problems in industrial applications. This implies a desire to explore its applicability in a wider range of scenarios and potentially integrate it with more sophisticated BPA generation techniques.
Summary updated