Abstract
Thanks to the technical development of target tracking, multi-source data fusion and target tracking have become hotspots in the research of large-scale heterogeneous networks. Based on millimeter-wave (mmWave) heterogeneous networks, this paper constructs a multi-source data fusion and target tracking model centered on a Data Mining Deep Q Network (DM-DQN). Through image filling, the length of the input vector (time window) was extended from 25 to 31 frames using CNN heterogeneous network technology. This alignment of input and output vector lengths preserves the temporal features of eye-tracking data to the greatest extent, thereby expanding the recognition range. Experimental results show that the proposed model achieved a modified mean error of only 1.5m with a tracking time of 160s, indicating an ideal tracking effect. Consequently, the DM-DQN outperformed other algorithms in total user delay and network energy efficiency while ensuring Quality of Service (QoS). While DM-DQN initially underperformed compared to structured data mining in the first 50 iterations, it stabilized and achieved superior performance after 100 iterations, demonstrating fast convergence and robust performance.
Introduction
As telecommunications transition toward 5G and Beyond 5G (B5G) architectures, networks are increasingly characterized by their heterogeneity. These systems integrate traditional ground base stations (G-BS) with mobile units, such as Unmanned Aerial Vehicle (UAV)-mounted base stations (U-BS). Tracking targets—whether users or vehicles—within these complex environments presents significant challenges, including data fragmentation across diverse sources, domain shifts between varying data resolutions, and severe resource constraints. Traditional static algorithms often struggle to balance real-time tracking accuracy with the energy demands of modern mmWave networks.
Methodology
The researchers propose a multi-layered architecture centered on a Data Mining Deep Q Network (DM-DQN) framework. This reinforcement learning approach allows an agent to interact with the network environment, learning optimal strategies for Resource Block (RB) and power allocation.
To address temporal data loss during convolution, a novel "image filling" technique was employed, extending the input time window from 25 to 31 frames. This ensures alignment between input and output vectors, preserving critical temporal features essential for accurate path prediction. The model further integrates Deep Convolutional Neural Networks (DCNN) for spatial feature extraction with Bidirectional Long Short-Term Memory (BLSTM) networks for long-term temporal dependencies.
The system was validated using datasets such as the Sketchy Dataset for cross-domain retrieval and eye-tracking data for temporal preservation, utilizing VGG-19 pre-trained weights as a spatial backbone. A Higher-Order Learnable Energy Function (HOLEF) was utilized as a loss metric to mitigate feature misalignment between network branches.
Results and Discussion
The DM-DQN model demonstrates significant improvements in both accuracy and efficiency. Experimental results indicate a modified mean error of just 1.5 meters over a 160-second tracking duration. A key observation was the "50-iteration pivot": while the DM-DQN initially underperformed compared to traditional structured data mining, it surpassed these methods after 100 iterations as the reinforcement learning agent optimized its strategy.
Furthermore, the model significantly enhances network energy efficiency by dynamically adjusting transmission power based on real-time demand. As shown in the performance analysis, the DM-DQN-based partition module maintains high monitoring values across various network parameters. By jointly optimizing RB and power allocation, the DM-DQN reduces total user delay compared to traditional "two-stage" algorithms, ensuring high Quality of Service (QoS) for all network participants.
Conclusion
The proposed DM-DQN framework effectively addresses the complexities of target tracking in heterogeneous mmWave networks. By preserving temporal integrity and leveraging reinforcement learning for resource management, the system achieves high precision and energy efficiency. Future research will focus on multi-agent distributed management to reduce interference and expanding the model to handle 3D multi-target tracking in cluttered environments.
Summary updated