Adaptive WSN Clustering with Decision Transformer Model and Dynamic Cluster Head Selection
- 주제(키워드) Wireless Sensor Network (WSN) , Clustering , Reinforcement Learning (RL) , Decision Transformer (DT) , Lifetime
- 주제(DDC) 006.31
- 발행기관 아주대학교 일반대학원
- 지도교수 Young-June Choi
- 발행년도 2026
- 학위수여년월 2026. 2
- 학위명 석사
- 학과 및 전공 일반대학원 인공지능학과
- 실제URI http://www.dcollection.net/handler/ajou/000000035524
- 본문언어 영어
- 저작권 아주대학교 논문은 저작권에 의해 보호받습니다.
초록/요약
A wireless sensor network (WSN) is a distributed system in which multiple sensor nodes collect and transmit environmental data to a Base Station (BS) under limited energy resources. However, the energy constraints of sensor nodes limit network lifetime and lead to performance degradation. Inefficient Cluster Head (CH) selection causes energy imbalance and reduces energy efficiency. Existing probabilistic or fixed clustering schemes fail to balance energy consumption in dynamic environments, while reinforcement learning (RL)-based methods still suffer from excessive energy waste due to random exploration during early training. To address these issues, this paper proposes a WSN clustering optimization method that integrates a RL-based Decision Transformer (DT) with a dynamic cluster head selection mechanism. The proposed approach dynamically selects CH by considering network-wide information—such as the number of alive nodes, node locations, distances to the BS, and residual energy—to maintain energy balance. By leveraging past experiences through a Transformer-based sequence model, the DT mitigates the instability and random exploration issues of traditional RL methods and enables stable learning of energy-efficient policies. Furthermore, the combination of a Replay Buffer and Target Network enhances learning convergence and overall training stability. Experimental results demonstrate that the proposed model significantly outperforms conventional WSN clustering and RL-based methods in terms of First Node Death (FND) and average node lifetime. Specifically, the FND improved by up to 893.46%, and the average node lifetime by 174.41%, confirming that the proposed model achieves both stability and energy efficiency. In addition, the model exhibits strong adaptability, maintaining performance across various network settings without retraining. Therefore, this study presents a next-generation Transformer-based RL clustering framework that effectively addresses early exploration and energy inefficiency problems while ensuring robust scalability in large-scale WSNs.
more목차
1. Introduction 1
1.1 Introduction to WSN and Clustering 1
1.2 Limitations of Existing Models and Proposed Solution 1
2. Related Works 5
2.1 Classical clustering techniques 5
2.2 Reinforcement Learning-based clustering technique 6
2.3 Decision Transformer-based clustering 7
3. System Model 9
3.1 Network Model 9
3.2 Energy Model 10
4. Proposed Method 12
4.1 Framework 13
4.2 State, Action and Reward 13
4.3 Decision Transformer structure 14
4.4 Dynamic Cluster Head Selection Mechanism 14
4.5 Concept of reward function model 15
4.6 Definition and importance of performance metrics 17
5. Results 18
5.1 WSN environment and clustering process 18
5.2 Limitations of Comparative Models 20
5.3 . Performance Comparison and Evaluation 21
5.4 Other performance comparisons 26
6. Conclusion 29
Bibliography 30
