PaperlayerView exact revision on arXiv

Source-linked AI summary

A Transfer Learning-Driven Methodology for Efficient and Sustainable Indoor Localization Using VLC

Masood Jan, Wafa Njima, Xun Zhang, Alexander Artemenko

arXiv:2610.29594v1cs.LGeess.SP

TL;DR

Industrial indoor localization must contend with changing environments and the data and energy demands of conventional approaches. This paper applies transfer learning to VLC localization and combines accuracy, energy, data use, and transfer gain in a Model Efficiency metric. On a BOSCH factory dataset, it reports up to 47% lower localization error than conventional models and up to 67.5% Model Efficiency.

  • Problem

    Conventional localization approaches face accuracy, energy, computational, data-efficiency, and adaptability challenges in industrial environments.

    Source: §I · p. 1

  • Method

    The framework transfers pretrained DNN knowledge to a target VLC localization model and introduces Model Efficiency to combine accuracy, energy use, data needs, and transfer gain.

    Source: §I · p. 2, A. Modeling Environmental Variability… · p. 4

  • Results

    The framework achieves up to 67.5% Model Efficiency and reduces localization error by up to 47% compared to conventional models.

    Source: §VII · p. 15

  • Takeaways & Limitations

    The authors conclude that the framework has potential for industrial automation and smart logistics, while maintaining high accuracy with 30% of the original dataset.

    Source: §VII · p. 15

    Field deployments with separate, unsynchronized transmitters and receivers may have timing or frequency mismatches that increase measurement variance and bias features; these effects remain unquantified experimentally.

    Source: F. Limitations’ Discussion and Future… · p. 14

Abstract

from arXiv · show

The rapid expansion of the Internet of Things enables seamless collaboration among connected devices, making indoor localization a critical component of industrial automation. However, conventional localization techniques struggle in dynamic environments. These methods depend on extensive, frequent data collection and environment-specific calibration, limiting scalability, interoperability, and effective use of prior research while imposing significant energy penalties that hinder practical deployment. To address these challenges, we propose an energy-efficient transfer learning (TL) based approach for visible light communication based indoor localization. Our method leverages TL to tackle environmental variability, including lighting fluctuations and physical obstacles that typically degrade localization performance, while simultaneously reducing computational overhead and energy consumption compared to conventional approaches. We also introduce a novel model efficiency (ME) metric, designed to integrate localization accuracy, energy efficiency, data efficiency, and transfer gain into a single evaluative measure for comprehensive optimization. Evaluations on a real-world dataset collected from a BOSCH factory demonstrate that our model achieves a 47% improvement in localization accuracy compared to a conventional model that does not utilize TL, and achieves up to 67.5% ME with TL models, compared to only 10.4% in conventional models under high-noise conditions. These results underscore the potential of our approach to deliver a highly efficient and scalable indoor localization system suitable for energy-constrained industrial applications.

I. INTRODUCTION

The paper targets VLC indoor localization that can adapt to changing industrial conditions without heavy data collection or retraining. It positions transfer learning as a response to gaps in VLC adaptation and evaluates it alongside a unified efficiency perspective.

I. INTRODUCTION

  • Dynamic lighting and layouts can degrade offline-trained models, while industrial deployment also faces accuracy, energy, complexity, data, and adaptability constraints.VLC offers advantages including immunity to electromagnetic interference, but ambient light and line-of-sight conditions remain challenges.Source: §I · pp. 1–2

A. Federated Learning for Indoor Localization

  • Federated learning supports privacy-preserving collaborative training, but client-specific models may generalize poorly to unseen conditions and costly retraining can erode energy savings.FeMLoc combines collaborative meta-training with rapid adaptation to new environments; it reports up to 80.95% accuracy improvement and 82.21% fewer training iterations than conventional neural networks.Source: A. Federated Learning for Indoor… · pp. 2–3

B. Continuous Learning for Indoor Localization

  • Continuous learning adapts incrementally, but catastrophic forgetting and the memory, latency, update, or computation costs of mitigation can constrain resource-limited industrial use.The review contrasts these demands with transfer learning’s use of prior knowledge and minimal fine-tuning.Source: B. Continuous Learning for Indoor… · p. 3

II. RELATED WORKS

  • The review frames federated, continuous, and transfer learning around adaptability, energy efficiency, and transferability, while VLC transferability and noise robustness remain open challenges.It notes that most existing VLC studies focus on deterministic modeling or hardware-based noise mitigation rather than learning-based domain adaptation.Source: §II · p. 2, C. Transfer Learning for Indoor… · p. 4

C. Transfer Learning for Indoor Localization

  • Transfer learning reuses pretrained knowledge to reduce retraining and labeled-data needs, but remains underused for VLC-specific ambient-light, photodiode-noise, and LED-flicker challenges.The paper’s stated contribution combines a noise-adaptive architecture, a unified Model Efficiency metric, and industrial-scale validation.Source: C. Transfer Learning for Indoor… · pp. 3–4

III. SYSTEM MODEL AND PROBLEM FORMULATION

The framework compares base, environmental-variation, and transfer-learning DNN configurations, then uses a composite efficiency metric to assess deployment trade-offs. Its transfer process initializes a target model from a source model and jointly accounts for source- and target-domain losses.

III. SYSTEM MODEL AND PROBLEM FORMULATION

  • The framework compares Base, EV, and TL configurations across accuracy, energy use, data utilization, and transfer gains using the Model Efficiency metric.The framework is intended to support comparison under varying environmental conditions and deployment decisions.Source: §III · p. 4

A. Modeling Environmental Variability and TL Framework

  • The proposed process transfers knowledge from pretrained source model fs to target model ft to improve generalization under new environmental conditions.The evaluation considers three model configurations.Source: A. Modeling Environmental Variability… · p. 4, Figure 2 · p. 5
  • The Base model trains on raw VLC data, the EV model adds a Noise Factor to simulate lighting variability, and the TL model fine-tunes source weights on target-environment data.The EV model represents environmental uncertainty in signal strength; the TL model leverages previously learned representations.Source: A. Modeling Environmental Variability… · p. 4

1) Problem Formulation:

  • The source domain maps signal-intensity features to longitude and latitude, while the target task seeks a mapping that generalizes to its domain.The source data contain ns samples and the input features occupy Rd.Source: 1) Problem Formulation · pp. 4–5
  • The target model starts from source-model parameters and is fine-tuned on target-domain data to adapt the learned mapping.The source DNN is parameterized by learned weights θs; the updated target weights are θt.Source: 1) Problem Formulation · p. 5

2) Loss Design:

  • The total objective combines source and target losses with weights λs and λt, which are optimized through cross-validation for target-task performance.The domain-specific losses are computed independently using mean squared error, and jointly guide preservation of source knowledge alongside target adaptation.Source: 2) Loss Design · p. 5

3) Fine-Tuning:

Fine-tuning initializes the target model from source parameters and makes small target-data updates. The section also defines Model Efficiency as a weighted combination of four dimensions, including a validation-error-based accuracy component.

3) Fine-Tuning:

  • Fine-tuning initializes target parameters from the source model and applies small target-data updates to adapt while retaining source knowledge.The gradual updates reduce the risk of performance degradation from source–target domain differences.Source: 3) Fine-Tuning · pp. 5–6

B. Unified Performance Metric Formulation

  • Model Efficiency combines Performance Accuracy, Energy Efficiency, Data Efficiency, and Transfer Gain for balanced model comparison.The study assigns each component a weight of 0.25.Source: B. Unified Performance Metric… · p. 6

1) Performance Accuracy (PA):

2) Energy Efficiency (EE):

The evaluation defines energy efficiency around training energy use and defines data efficiency around achieving strong performance with less training data.

2) Energy Efficiency (EE):

  • Energy Efficiency evaluates training cost as energy consumption per unit of useful work, with higher values indicating more energy-efficient training.The method averages energy consumed per epoch across training epochs.Source: 2) Energy Efficiency (EE) · p. 6
  • Training energy estimation uses epoch duration and sampled GPU and CPU power consumption.The passages describe regular power sampling during training.Source: 2) Energy Efficiency (EE) · p. 6

3) Data Efficiency (DE):

  • Data Efficiency rewards models that achieve strong performance using a reduced share of the available training data.The setup defines Dmax as 100% of available training data.Source: 3) Data Efficiency (DE) · p. 6

4) Transfer Gain (TG):

The section compares transfer gain against base and environmental-variation models, while the pre-transfer workflow creates noisy datasets to represent environmental drift. The reported noise experiments show increased localization error for environmental-variation models relative to the base model.

4) Transfer Gain (TG):

  • Transfer Gain measures relative performance improvement for a TL model against both the environmental-variation and base models.A positive Transfer Gain indicates improvement relative to both comparison models.Source: 4) Transfer Gain (TG) · pp. 6–7
  • Model Efficiency integrates four components to compare models without allowing improvements in one area to compromise another.Source: 4) Transfer Gain (TG) · p. 7

A. Pre-TL Model Development

  • The pre-transfer workflow trains a base model on the original dataset, then injects Gaussian noise to create environmental-variation datasets representing drift over time.Noise intensity is controlled by standard deviation scaled by a noise factor.Source: A. Pre-TL Model Development · p. 7
  • Localization error increases substantially for environmental-variation models trained on noise-injected data compared with the base model on the original dataset.The tested noise factors are 2, 4, and 8.Source: A. Pre-TL Model Development · p. 7

B. TL Algorithm Development and Implementation

The TL algorithm transfers base-model weights, freezes shared layers, and fine-tunes the remaining layers on noisy target data. A separate evaluation algorithm collects metrics for Base, EV, and TL models and returns their Model Efficiency values.

B. TL Algorithm Development and Implementation

C. ME Evaluation Algorithm

  • The ME evaluation algorithm extracts accuracy and energy metrics for Base, EV, and TL models, assigning zero Transfer Gain to Base and EV models.For TL models, Transfer Gain is retrieved according to data size and noise factor.Source: C. ME Evaluation Algorithm · p. 8
  • After collecting component values, the evaluation algorithm computes Data Efficiency and Model Efficiency and returns the resulting ME values for comparison.Source: C. ME Evaluation Algorithm · p. 8

V. EXPERIMENTAL SETUP AND DATA COLLECTION

The experimental setup was implemented in a real-world industrial environment to collect signal-strength data across transmitter-receiver configurations under factory conditions.

V. EXPERIMENTAL SETUP AND DATA COLLECTION

  • The setup was designed to capture signal-strength data across transmitter-receiver configurations for model training and validation under realistic factory conditions.The section covers the VLC transceiver, BOSCH factory measurement setup, and data-collection methodology.Source: §V · p. 8

1) Transceiver Architecture:

The VLC transceiver converts an electrical input into an optical signal at the transmitter and processes the received signal through amplification, equalization, and filtering. Its transmitter and receiver have distinct optical and sensing specifications.

1) Transceiver Architecture:

  • The TX pre-equalizes and amplifies its input before the light source converts it into an optical signal; the RX processes the received signal through amplification, post-equalization, and filtering.An optional lens can adjust beam directivity to suit measurement requirements.Source: 1) Transceiver Architecture · pp. 8–9, Figure 3 · p. 9

2) Transceiver Specifications:

  • The transmitter uses a 450 nm LED with 250–750 mW optical output and a 60° full-angle half-power beamwidth.The LED is driven at 200 mA ± 100 mA and operates across a 450–650 nm spectral range.Source: 2) Transceiver Specifications · p. 9
  • The receiver’s parallel photodetectors provide a 600 mm² sensitive area, with a 60° full-angle half-sensitivity field of view across 380–780 nm.The transceiver units use USB-C power connections, with a converter board supplying stable 5 V DC to the optical front-end and signal-processing circuits.Source: 2) Transceiver Specifications · p. 9

B. VLC Measurement System

Measurements from the BOSCH factory were used to evaluate base, noise-injected, and transfer-learning models under varying noise conditions. Transfer learning maintained near-base localization error under noise, while the noise-injected model degraded substantially.

B. VLC Measurement System

  • The BOSCH measurement setup uses two VLC transceivers, converter boards, DC supplies, and a VNA to capture received signals at different RX positions.The converter boards combine the VNA’s RF signal with DC power for simultaneous transmission and device powering.Source: B. VLC Measurement System · p. 9

C. Data Collection

  • Measurements covered a 46 m × 42 m × 11 m portion of the 74 m × 171 m × 11 m factory hall, with TX and RX heights fixed at 2.9 m and 1.1 m.The receiving area was divided into 0.2 m × 0.2 m grids whose intersections served as measurement points.Source: C. Data Collection · p. 9, Figure 5 · p. 10, Figure 6 · p. 10
  • Ten transmitter areas were sampled at about 160 receiver positions each; five-second signal recordings were averaged into one RSSI value per position.The on-site Base dataset reflects actual factory noise and serves as the ground-truth reference for EV and TL comparisons.Source: C. Data Collection · p. 10

A. TL Model Evaluation under different lighting conditions

  • The TL model is pretrained on the Base dataset, then fine-tuned with 30%, 50%, 70%, or 100% of target-domain training data at NF = 2, 4, or 8.Across Base, EV, and TL conditions, 80% of samples were used for training and 20% held out for validation.Source: A. TL Model Evaluation under different… · pp. 10–11
  • 41.37 cm at NF = 2 and 42.59 cm at NF = 8 were the TL model’s localization errors, versus 70.17 cm and 80.41 cm for EV, respectively.The base model’s error was 41.21 cm; the EV error at NF = 8 represented a 95.14% degradation relative to the base model.Source: A. TL Model Evaluation under different… · p. 11
  • The authors note that localization error may not surpass some state-of-the-art methods; larger architectures could reduce error at higher computational and power costs.The stated aim is to balance accuracy, noise immunity, and minimal data requirements efficiently.Source: A. TL Model Evaluation under different… · p. 11
  • Using noise-trained base models reduced TL localization error by 0.29% at NF = 2 and 0.67% at NF = 4, with only minimal sensitivity to base-model selection.The reported findings also state that precision degrades as noise intensifies, while TL mitigates the effect and enhances stability.Source: A. TL Model Evaluation under different… · p. 11

B. TL model evaluation with limited data

Transfer learning preserves localization performance with limited data and environmental noise, while the unified Model Efficiency metric shows higher efficiency than environmental-variation models across tested conditions.

B. TL model evaluation with limited data

C. Key Metric Evaluation: TG, Accuracy and EE

D. Model Efficiency Evaluation

  • 61.5% ME for TL Model 30% (NF2) exceeds the base model’s 53.7%, while TL Model 100% (NF8) reaches 53.0% versus 10.4% for EV Model 100% (NF8).The paper reports that TL models consistently outperform EV counterparts and maintain efficiency as noise increases.Source: D. Model Efficiency Evaluation · pp. 13–14, Figure 10 · p. 13

E. Generalization capacity of the best ME model

The highest-ME TL model stays below its 80.1 cm error threshold across tested noise conditions. The paper also reports overall accuracy and efficiency gains, while identifying environmental, deployment, sensing, scaling, and energy-accounting boundaries for future work.

E. Generalization capacity of the best ME model

F. Limitations’ Discussion and Future Work

  • Experiments do not fully cover direct sunlight, complete darkness, rapid illumination transients, or very dense multi-obstacle layouts.These conditions may reduce LOS reliability and increase positioning error.Source: F. Limitations’ Discussion and Future… · p. 14
  • Using a VNA as a shared TX/RX reference reduces synchronization-related variance, so unsynchronized field hardware may introduce timing or frequency mismatches and biased features.The authors leave systematic experimental quantification of these deployment effects to future work.Source: F. Limitations’ Discussion and Future… · p. 14
  • The framework currently uses VLC alone; scaling to larger infrastructures or complex networks and fine-tuning on edge devices remain open challenges.The paper identifies multi-sensor fusion and practical edge-platform limits as future directions.Source: F. Limitations’ Discussion and Future… · p. 14
  • The reported energy term covers training only, excluding inference and network communication energy that may matter in deployed IoT scenarios.The authors identify full life-cycle energy accounting as future work.Source: F. Limitations’ Discussion and Future… · pp. 14–15

VII. CONCLUSION

  • Up to 67.5% ME and a localization-error reduction of up to 47% versus conventional models are reported in the conclusion.The conclusion also reports high accuracy using only 30% of the original dataset.Source: §VII · p. 15
Loading 2610.29594v1…