Abstract
In modern industrial environments, real-time localization of manufacturing objects (MOs) using Radio Frequency Identification (RFID) is hindered by the high volume, velocity, and non-stationary nature of signal data. This paper introduces Scalable PANFIS, a distributed evolving neuro-fuzzy framework that leverages Apache Spark for parallel processing. By employing a novel rule-merging strategy, the system achieves a 20-fold increase in processing speed compared to traditional methods while maintaining high classification accuracy. Experimental results from a smart rack setup demonstrate the framework's effectiveness in handling massive RFID data streams with minimal accuracy loss.
Introduction
The advent of "smart" factories necessitates precise tracking of tools, personnel, and inventory. While RFID technology is cost-effective and widely deployed, analyzing Received Signal Strength (RSS) data presents significant challenges. RSS streams are characterized by high volume and velocity, non-stationarity due to environmental shifts (e.g., human movement or metallic interference), and high noise levels. Traditional machine learning models often require slow batch retraining, making them unsuitable for real-time applications. While "evolving" algorithms can learn incrementally, they typically lack the scalability required for big data environments.
Methodology
The proposed Scalable PANFIS framework addresses these challenges by adapting the Parsimonious Neo-Fuzzy Inference System (PANFIS) for a distributed architecture.
Distributed Processing with SparkR
To handle large-scale data, the system is integrated with Apache Spark. The data stream is partitioned into chunks, and each worker node in the cluster executes a local instance of PANFIS. This allows for parallelized learning across multiple servers, significantly reducing the computational bottleneck of single-machine systems.
Rule Merging Scheme
A primary challenge in distributed learning is the redundancy of rules generated by independent nodes. Scalable PANFIS incorporates a Rule Merging Scheme to consolidate local models into a single, global model. This process utilizes the Bhattacharyya distance to identify overlapping fuzzy rules and performs homogeneity checks to ensure merging does not compromise precision. This approach ensures the final model remains parsimonious and interpretable without sacrificing accuracy.
Experimental Evaluation
The framework was validated using a real-world dataset from the SIMTech laboratory in Singapore. The setup involved a "smart rack" where passive RFID tags were monitored by a reader. A dataset of 283,100 observations was collected over 20 hours, with the task of classifying objects into one of four shelf locations based on RSS frequency.
Performance Results
Scalable PANFIS demonstrated exceptional efficiency, processing the dataset in 104 seconds—a 20x speedup over the single-machine version (2,130 seconds). Despite the distributed nature of the processing, accuracy remained high at 96.67%, only a marginal 2% decrease from the centralized version (98.71%). Furthermore, the system significantly outperformed standard Spark MLlib algorithms, such as Generalized Linear Models (50.03%) and Gradient Boosted Trees (75.07%).
Conclusion
Scalable PANFIS provides a robust solution for real-time RFID localization in dynamic manufacturing settings. By balancing the trade-off between speed and accuracy through distributed processing and intelligent rule merging, the framework offers a scalable path for big data analytics. Future work will explore the application of this framework to higher-dimensional sensor fusion tasks and broader big data classification challenges.
Summary updated