Introduction
In the era of Industry 4.0, real-time tracking of manufacturing objects (MOs) such as tools, materials, and personnel is critical for operational efficiency. Radio Frequency Identification (RFID) has emerged as a preferred technology due to its cost-effectiveness and ease of deployment. However, the Received Signal Strength (RSS) used for localization is highly susceptible to noise, multipath effects from metal machinery, and the non-stationary nature of a dynamic factory floor.
Traditional machine learning models often fail to adapt to these changing environments or require significant computational resources that preclude real-time, "online" learning. This paper presents the evolving Type-2 Quantum Fuzzy Neural Network (eT2QFNN), a novel approach designed to overcome these challenges by learning from data streams in a single pass.
The eT2QFNN Methodology
The core of the proposed solution is the eT2QFNN, which integrates evolving structural learning with quantum fuzzy logic to handle uncertainty and noise.
Evolving Architecture
Unlike static neural networks, the eT2QFNN starts with zero rules. It dynamically creates new fuzzy rules as it encounters novel data points, guided by the modified Generalized Type-2 Datum Significance (mGT2DS) metric. This allows the model to adapt its structure to the specific layout and signal characteristics of a shopfloor without extensive pre-training.
Quantum Fuzzy Sets and Parameter Tuning
The model utilizes Interval Type-2 Quantum Fuzzy Sets. These sets provide a "footprint of uncertainty," which is essential for managing the fluctuating RSS signals common in industrial settings. By using quantum membership functions, the network can better distinguish between overlapping signal profiles at different locations.
To maintain computational efficiency, the authors employ a Decoupled Extended Kalman Filter (DEKF) for parameter updates. Instead of updating the entire network, the DEKF focuses on the "winning" rule most relevant to the current observation, significantly reducing the mathematical overhead and enabling high-speed online performance.
Experimental Validation
The system was tested at the SIMTech laboratory in Singapore using a "smart rack" setup. The dataset comprised over 283,000 observations collected over 20 hours of continuous operation.
The eT2QFNN was benchmarked against several state-of-the-art algorithms, including gClass and eT2ELM. Key findings include:
- High Accuracy: The model achieved a 99% classification rate, outperforming all competitors.
- Compact Model: It successfully represented the environment using only about 6 fuzzy rules, demonstrating high structural efficiency.
- Real-Time Speed: It ranked as the second-fastest algorithm, proving its suitability for low-latency industrial applications.
Conclusion and Future Directions
The eT2QFNN offers a robust solution for real-time RFID localization by combining the adaptability of evolving systems with the noise-handling capabilities of quantum fuzzy logic. While the current model excels at adding rules, future iterations will explore rule pruning and merging mechanisms to ensure long-term stability in environments that undergo drastic, permanent changes over years of deployment.
Summary updated