Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Token Entropy Regularization for Multi-modal Antenna Affiliation Identification

Dong Chen, Ruoyu Li, Xinyan Zhang, Jialei Xu, Ruoseng Zhao, Zhikang Zhang, Lingyun Li, Zizhuang Wei

TL;DR

The paper tackles automatic antenna affiliation identification to replace labor-intensive tower inspections by fusing drone-based base-station videos, antenna geometry, and PCI signals within a two-stage training paradigm. It introduces Token Entropy Regularization (TER), comprising Enhanced Token Entropy and Token Entropy Loss, to align heterogeneous visual and signal representations and promote sparse, informative token encodings. Empirical results show TER improves convergence and accuracy across multiple vision backbones, with a Pretrain + SFT regime yielding an overall top-1 antenna–PCI matching accuracy of $87.91\%$ and top-3 of $91.94\%$, and ablations confirming the benefits of token-level sparsity. The work demonstrates a practical, data-efficient path for open-world multimodal calibration in 4G/5G networks, reducing reliance on manual inspections and enabling scalable network optimization and asset management.

Abstract

Accurate antenna affiliation identification is crucial for optimizing and maintaining communication networks. Current practice, however, relies on the cumbersome and error-prone process of manual tower inspections. We propose a novel paradigm shift that fuses video footage of base stations, antenna geometric features, and Physical Cell Identity (PCI) signals, transforming antenna affiliation identification into multi-modal classification and matching tasks. Publicly available pretrained transformers struggle with this unique task due to a lack of analogous data in the communications domain, which hampers cross-modal alignment. To address this, we introduce a dedicated training framework that aligns antenna images with corresponding PCI signals. To tackle the representation alignment challenge, we propose a novel Token Entropy Regularization module in the pretraining stage. Our experiments demonstrate that TER accelerates convergence and yields significant performance gains. Further analysis reveals that the entropy of the first token is modality-dependent. Code will be made available upon publication.

Token Entropy Regularization for Multi-modal Antenna Affiliation Identification

TL;DR

The paper tackles automatic antenna affiliation identification to replace labor-intensive tower inspections by fusing drone-based base-station videos, antenna geometry, and PCI signals within a two-stage training paradigm. It introduces Token Entropy Regularization (TER), comprising Enhanced Token Entropy and Token Entropy Loss, to align heterogeneous visual and signal representations and promote sparse, informative token encodings. Empirical results show TER improves convergence and accuracy across multiple vision backbones, with a Pretrain + SFT regime yielding an overall top-1 antenna–PCI matching accuracy of and top-3 of , and ablations confirming the benefits of token-level sparsity. The work demonstrates a practical, data-efficient path for open-world multimodal calibration in 4G/5G networks, reducing reliance on manual inspections and enabling scalable network optimization and asset management.

Abstract

Accurate antenna affiliation identification is crucial for optimizing and maintaining communication networks. Current practice, however, relies on the cumbersome and error-prone process of manual tower inspections. We propose a novel paradigm shift that fuses video footage of base stations, antenna geometric features, and Physical Cell Identity (PCI) signals, transforming antenna affiliation identification into multi-modal classification and matching tasks. Publicly available pretrained transformers struggle with this unique task due to a lack of analogous data in the communications domain, which hampers cross-modal alignment. To address this, we introduce a dedicated training framework that aligns antenna images with corresponding PCI signals. To tackle the representation alignment challenge, we propose a novel Token Entropy Regularization module in the pretraining stage. Our experiments demonstrate that TER accelerates convergence and yields significant performance gains. Further analysis reveals that the entropy of the first token is modality-dependent. Code will be made available upon publication.
Paper Structure (18 sections, 5 equations, 6 figures, 4 tables)

This paper contains 18 sections, 5 equations, 6 figures, 4 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Multi-modal Encoders
  4. Token Entropy Utilization
  5. Sparse Rate Reduction
  6. Methodology
  7. Task Formulation
  8. Multi-modal Data Acquisition
  9. Token Entropy Regularizer
  10. Experiments
  11. Implementation Detials
  12. Antenna Type Classification
  13. Antenna Affiliation Pretraining
  14. Integrated Antenna Affiliation Evaluation
  15. Visualization
  16. ...and 3 more sections

Figures (6)

  • Figure 1: Schematic diagram of the multi-modal antenna affiliation identification framework. (A) The practical needs of identifying antenna affiliations. (B) A multi-modal approach is proposed with base station videos, antenna geometric features, and PCI signals. (C) The task is achieved through a two-stage training process: classifying antenna types, and matching affiliations.
  • Figure 2: An overview of the proposed multimodal framework for antenna affiliation identification. (A) Antenna type classification via a supervised DINOv3-Transformer. (B, C) The antenna affiliation task comprising a pretraining stage followed by supervised fine-tuning under different data-matching conditions. (D) The Token Entropy Regularization (TER) module, which encourages model convergence by regularizing the distribution of token representations.
  • Figure 3: Visual comparison of antenna type classification. Rows show different antenna instances; columns compare ground truth (left) with predictions from three distinct models. Line style indicates manufacturer (solid: Company A, dashed: Company B); color indicates frequency band (blue: low, red: high).
  • Figure 4: Antenna affiliation matching results. Each row presents a distinct urban scenario with communication infrastructure. Yellow points mark detected PCI signals, with solid circles indicating verified matches and dashed circles indicating unclassified signals. Colors represent different signal classes.
  • Figure 5: Accuracy during pretraining. Our proposed ETE module converges faster and to a higher accuracy than the baseline. Performance is further improved by the full TER method (ETE + TEL).
  • ...and 1 more figures