PaperlayerView exact revision on arXiv

Source-linked AI summary

A Continuous High Precision Localization Method for AGVs Based on Interferometric FBG Array Sensing and Hybrid Deep Learning

Laiming Wu, Xin Gui, Hui Cao, Yunchuan Zhang, Haiyang Wang, Jiaqi Wang, Zhengying Li

arXiv:2610.08041v1eess.SP

TL;DR

Continuous AGV localization is challenging in electromagnetically sensitive industrial settings, where existing approaches have limitations. The paper combines interferometric FBG array sensing with a hybrid deep-learning framework and reports measured-signal MAE below 2 cm under varied operating conditions.

  • Problem

    Continuous, high-precision AGV localization remains challenging in industrial settings, where RF or vision methods may be ineffective and optical sensing methods can have limited spatial resolution.

    Source: A. Context and Motivation · p. 1, B. Related Work · p. 2

  • Method

    The method combines a shallowly buried interferometric FBG track, task-oriented spatiotemporal features, and hybrid deep learning to estimate position from multichannel signals.

    Source: C. Main Contributions · p. 2, §V · p. 12

  • Results

    Measured-signal MAE is 1.9203 cm, while simulated-signal MAE is 0.1278 cm; average error remains below 2 cm across repeated experiments and varying speed and load.

    Source: §V · p. 12

  • Takeaways & Limitations

    The experiments report accuracy, repeatability, and robustness under repeated runs and varying speed and load conditions.

    Source: §V · p. 12

Abstract

from arXiv · show

High precision continuous localization is a fundamental challenge for automated guided vehicles (AGVs) operating in electromagnetically sensitive industrial environments. To mitigate the continuity limitations of conventional discrete localization methods and the accuracy limitations of existing continuous localization methods, this paper proposes a continuous AGV localization method that integrates interferometric fiber Bragg grating (FBG) array sensing with a hybrid deep learning framework. We first design a shallowly buried interferometric FBG array sensing track to enable the stable acquisition of high-quality signals. The received sensing signals are subsequently parameterized and preprocessed in a task-oriented manner, in which a multichannel joint representation is constructed using a sliding time window. A hybrid deep learning framework is developed to estimate AGV position from multichannel FBG response signals by jointly exploiting inter-channel spatial correlation and temporal continuity in the measurement sequence. Experimental results show that the proposed method achieves a mean absolute error (MAE) of less than 2 cm under complex industrial conditions, proving its effectiveness and robustness in practical AGV localization scenarios.

I. INTRODUCTION

The paper addresses continuous AGV localization challenges in industrial environments and combines interferometric FBG sensing with deep learning. Experiments report localization errors below 2 cm under complex industrial conditions.

A. Context and Motivation

  • Industrial localization is challenging because RF- and vision-based methods can become ineffective amid high electromagnetic interference and information-security constraints.Localization accuracy also contributes to operational safety, task scheduling efficiency, and system quality of service.Source: A. Context and Motivation · p. 1

B. Related Work

  • Conventional φ-OTDR can detect weak dynamic disturbances, but its typical meter-level spatial resolution is inadequate for continuous high-precision AGV localization.The paper identifies improving spatial resolution while retaining optical-fiber sensitivity as a fundamental challenge.Source: B. Related Work · p. 2
  • Prior FBG-based AGV positioning work established feasibility, but feature-point or discrete-correction methods struggle to localize continuously along the entire path.Earlier traffic-sensing studies largely focused on vehicle state assessment rather than AGV localization.Source: B. Related Work · p. 2
  • Earlier deep-learning optical-fiber studies addressed sensing interpretation and monitoring, but not continuous localization of moving AGVs from multichannel FBG responses in enclosed factories.The paper proposes deep learning to model spatiotemporal heterogeneity in those responses.Source: B. Related Work · p. 2

C. Main Contributions

  • The framework combines a parametric signal model, spatiotemporal feature representation, and hybrid deep learning to model waveform morphology, spatial correlation, and temporal continuity.The shallowly buried FBG track supplies multichannel responses for continuous localization.Source: C. Main Contributions · p. 2
  • Experimental results show the proposed localization method achieves a mean absolute error (MAE) of less than 2 cm under complex industrial conditions, indicating strong robustness and promising engineering potential.Source: C. Main Contributions · p. 2

II. FBG ARRAY-BASED AGV LOCALIZATION SYSTEM

The localization system uses a shallowly buried interferometric FBG array track and an optical acquisition and demodulation chain. It is designed for predefined routes, with the route coordinate as the localization target.

A. Localization Track Design

  • Reflections from adjacent gratings interfere at the demodulator; strain or vibration changes the optical phase, which is demodulated to detect disturbances.The grating spacing and effective refractive index vary slightly as ground strain and vibration are transmitted to the fiber.Source: A. Localization Track Design · p. 3
  • The shallowly buried track combines a sensing cable, epoxy resin, fixing modules, and cement mortar to maintain stable mechanical coupling and strain transfer.The epoxy layer suppresses interfacial slip and microscale decoupling, improving strain-transfer efficiency and spatial consistency.Source: A. Localization Track Design · p. 3, Figure 2 · p. 3
  • The infrastructure is intended for fixed or semi-fixed predefined routes, trading route flexibility for long-term stable sensing.Routes may be straight, curved, or segmented, with arc length serving as the coordinate on curved tracks.Source: A. Localization Track Design · p. 3

B. Signal Acquisition and Demodulation System

  • The acquisition system sends optical pulses through the FBG array and uses an unbalanced Michelson interferometer to demodulate phase variations into time-domain response signals.Photodetector outputs are acquired by a DAQ card and transmitted to a server for analysis.Source: B. Signal Acquisition and Demodulation… · p. 4, Figure 3 · p. 4

C. Modeling of FBG Signals Induced by AGV

The section models FBG responses induced by AGV motion and selects signal components for preprocessing. It fits a standard waveform to measured responses after analyzing axle-loading patterns and quasi-static signals.

C. Modeling of FBG Signals Induced by AGV

  • The study AGV has three axles and six wheels, with 6 cm wheel diameters and 70 cm axle width.The section uses vehicle loading characteristics to analyze signal frequencies; adjacent axle spacing is stated as 40 cm in the frequency analysis.Source: C. Modeling of FBG Signals Induced by… · p. 4
  • Above 50 Hz at 0.5 m/s, wheel-loading excitation overlaps factory noise and is excluded from further analysis.The relevant wheel–ground contact length is typically smaller than 1 cm.Source: C. Modeling of FBG Signals Induced by… · p. 4
  • At 0.5, 1.0, and 1.5 m/s, axle-loading frequencies are 1.25, 2.50, and 3.75 Hz, producing three dominant peaks from adjacent axles.The axle responses superimpose with temporal offsets determined by axle spacing and vehicle speed; see Fig. 5.Source: C. Modeling of FBG Signals Induced by… · p. 4, Figure 5 · p. 5
  • Below 1 Hz, a common frequency band extracts the quasi-static response across vehicle speeds by treating the AGV as an overall moving load.This response does not require floor-surface irregularities or vehicle-speed fluctuations; see Fig. 6.Source: C. Modeling of FBG Signals Induced by… · p. 4, Figure 6 · p. 5
  • Measured responses have a stable, Gaussian-like time-domain profile despite amplitude and phase variation across sensing zones and speeds.The standard waveform is statistically fitted to a large set of measured data and shown in Fig. 7.Source: C. Modeling of FBG Signals Induced by… · pp. 4–5, Figure 7 · p. 5

D. Problem Formulation

A single sensing zone does not suffice for continuous localization because position is encoded in synchronized response evolution and spatial distribution across the FBG array. The task is therefore formulated as mapping multichannel signals to longitudinal position.

D. Problem Formulation

  • A single sensing zone is insufficient for continuous localization because position is jointly encoded in synchronized temporal evolution and spatial response distribution across the array.This motivates formulating localization at the multichannel system level.Source: D. Problem Formulation · p. 5
  • Each input Xk is an M × T multichannel signal segment from a sliding window, where M is sensing zones and T is temporal samples.The model learns a nonlinear mapping from this segment to position.Source: D. Problem Formulation · p. 5
  • The target yk is the AGV's reference longitudinal position along the predefined sensing route.For a curved route, it can be defined as arc length measured from the route origin.Source: D. Problem Formulation · pp. 5–6
  • For curved routes, localization error is measured against the longitudinal route coordinate, defined as arc length from the route origin.The study reports errors with respect to this coordinate.Source: D. Problem Formulation · pp. 5–6
  • The predicted position is the estimated AGV position for the current sample.The formulation maps synchronized multichannel FBG responses to corresponding AGV position.Source: D. Problem Formulation · p. 6
  • The localization task is a nonlinear regression problem from synchronized multichannel FBG responses to AGV position.The deep learning framework is developed to learn this mapping from multichannel sensing data.Source: D. Problem Formulation · p. 6

A. Framework Overview

The framework preprocesses multichannel FBG signals, constructs feature representations, learns deep spatiotemporal features, and regresses AGV position. Its representation combines original responses with temporally derived, energy-weighted features.

A. Framework Overview

  • The four-stage pipeline preprocesses signals, constructs feature representations, learns deep spatiotemporal features, and estimates position by regression.The network learns spatial correlations and temporal dependencies before mapping high-level features to position estimates.Source: A. Framework Overview · p. 6

B. Signal Preprocessing and Feature Construction

  • Disturbances and sensor noise can cause amplitude fluctuations and outliers, so signals are preprocessed before network training.The section introduces filtering to suppress high-frequency noise and random interference.Source: B. Signal Preprocessing and Feature… · p. 6
  • A causal, right-aligned sliding window uses current and historical samples, paired with the reference position at the window endpoint.Online predictions update by sliding stride without future observations; see Fig. 10.Source: B. Signal Preprocessing and Feature… · pp. 6–7, Figure 10 · p. 7
  • First- and second-order temporal differences are energy-weighted and concatenated with clipped response signals to form a multifeature input.The energy coefficient adaptively modulates temporal-difference features to strengthen transient vibration information associated with AGV motion.Source: B. Signal Preprocessing and Feature… · p. 7
  • The representation preserves original responses while adding dynamic temporal differences and energy-adaptive weighting for a more informative, stable input to deep spatiotemporal learning.The enhanced multifeature input is standardized using training-data mean and standard deviation.Source: B. Signal Preprocessing and Feature… · p. 7

C. Hybrid Deep Learning Localization Network

The localization network combines convolutional feature extraction, self-attention, LSTM temporal modeling, and position regression. It uses these modules to model local signal patterns, cross-channel correlations, and temporal dependencies in FBG responses.

C. Hybrid Deep Learning Localization Network

  • The hybrid network combines CNN feature extraction, self-attention enhancement, LSTM temporal modeling, and position regression to estimate AGV location from multichannel signals.It models correlations across sensing zones and temporal dependencies in FBG responses.Source: C. Hybrid Deep Learning Localization… · pp. 7–8
  • The CNN extracts local temporal patterns using 64 filters with kernel size 11, followed by 128 filters with kernel size 5 and max pooling.Max pooling reduces sequence length; the resulting feature representation has 128 channels.Source: C. Hybrid Deep Learning Localization… · p. 8
  • Multi-head self-attention models long-range temporal dependencies and emphasizes informative segments, while residual fusion preserves convolutional features.The network adds the attention output to the original convolutional features.Source: C. Hybrid Deep Learning Localization… · p. 8
  • A two-layer unidirectional LSTM with hidden dimension 128 models each right-aligned window using only current and historical signals.Its output sequence is aggregated into a compact representation for position regression.Source: C. Hybrid Deep Learning Localization… · p. 8

D. Training Strategy and Loss Function

Training combines a composite regression loss with optimization strategies intended to regularize learning and stabilize updates. The loss uses Huber regression and an additional penalty for errors exceeding a threshold.

D. Training Strategy and Loss Function

  • AdamW applies decoupled weight decay for parameter regularization, which the authors say helps mitigate overfitting and improve generalization.Source: D. Training Strategy and Loss Function · p. 9
  • A one-cycle learning-rate policy increases the rate early and reduces it later, while gradient clipping mitigates exploding gradients.The authors describe these strategies as stabilizing training and accelerating convergence.Source: D. Training Strategy and Loss Function · p. 9
  • The composite loss combines Huber regression with a penalty for prediction errors exceeding threshold τ, scaled by λ.The Huber component uses the ground-truth and predicted AGV positions, with δ as its switching threshold.Source: D. Training Strategy and Loss Function · p. 9
  • Huber loss grows linearly beyond its switching threshold, reducing the influence of abnormal samples on training.For smaller errors, it behaves quadratically to support accurate fitting.Source: D. Training Strategy and Loss Function · p. 9

A. Experimental Platform

The experimental platform pairs an FBG sensing track and interrogator with the hybrid localization framework. An offline laser SLAM-aided system supplies reference positions for evaluation, but is not used during online inference.

A. Experimental Platform

  • The evaluation platform consists of an FBG sensing track, an FBG array interrogation system, and the hybrid deep-learning framework.It is used for AGV motion sensing and localization performance evaluation.Source: A. Experimental Platform · p. 9
  • FBG sensors are shallowly embedded and mechanically coupled to the track, while the interrogator acquires and demodulates signals at 1000 Hz.The sensors are fixed periodically and bonded with epoxy resin.Source: A. Experimental Platform · p. 9
  • Laser SLAM-aided positions are aligned and synchronized with FBG signals as offline reference labels, with positioning deviation verified within ±5 mm.Online position estimates use multichannel FBG responses without the SLAM-aided system.Source: A. Experimental Platform · pp. 9–10

B. Dataset Construction and Parameter Settings

The study uses simulated signals for pretraining and experimentally acquired vibration signals to evaluate localization under varied speeds and loads. Its reported inference latency is a small fraction of the prediction update interval on a GPU-assisted platform.

B. Dataset Construction and Parameter Settings

  • The dataset includes simulated and experimentally acquired vibration signals, with simulated signals at 1.00 m/s and 300 kg used for model pretraining.The simulated dataset is used to assess feature extraction under controlled noise conditions.Source: B. Dataset Construction and Parameter… · p. 10
  • Experimental vibration signals cover AGV speeds of 0.75–1.50 m/s and unloaded 300 kg and loaded 500 kg conditions, with samples split 80%, 10%, and 10% into training, validation, and test sets.Source: B. Dataset Construction and Parameter… · p. 10
  • 0.32 ms average single-inference latency is about 3.2% of the 10 ms prediction update interval, indicating feasibility for online localization on a GPU-assisted platform.The configuration uses a 0.5 s causal input window from signals sampled at 1000 Hz.Source: B. Dataset Construction and Parameter… · p. 10

C. Localization Performance Evaluation

The proposed framework localizes AGVs from synthetic and measured signals, with measured-signal error remaining at the centimeter level. Tests report repeatability and stable accuracy across the evaluated speed and vehicle-load conditions.

D. Ablation and Robustness Analysis

The ablations report that the full CNN-Attention-LSTM combines complementary spatial and temporal modeling, while the composite loss best suppresses large errors. The conclusion reports average localization error below 2 cm across repeated, speed-varying, and load-varying experiments, and identifies further conditions for evaluation.

D. Ablation and Robustness Analysis

  • The full CNN-Attention-LSTM performs best among the compared architectures, with self-attention and LSTM providing complementary global-dependency and sequential modeling.The pure CNN is weakest; adding LSTM reduces localization error, while CNN-Attention outperforms the CNN baseline.Source: D. Ablation and Robustness Analysis · pp. 11–12, Table V · p. 12
  • The proposed composite loss achieves the best overall performance, particularly for MaxAE and large-error suppression.The comparison covers MSE, standard Huber, and composite losses; the composite loss adds an excess-error penalty, and its error distribution is more concentrated with a shorter tail.Source: D. Ablation and Robustness Analysis · p. 12, Table VI · p. 12

V. CONCLUSION AND FUTURE WORK

  • Average localization error remains below 2 cm in repeated experiments and under varying speed and load conditions.The conclusion reports good accuracy, repeatability, and robustness.Source: §V · p. 12
  • Future evaluation will cover acceleration, deceleration, stop-and-go, off-track conditions, floor structures, and sensing-track installation conditions.The authors also propose examining calibration or model fine-tuning for new deployments and joint longitudinal and lateral localization with multiline FBG sensing.Source: §V · p. 12
Loading 2610.08041v1…