Acta Optica Sinica, Volume. 45, Issue 16, 1606009(2025)

Spectral Drift Correction for FBG-based Shape Sensing using a Hybrid BiLSTM-LSTM Neural Network

Ming Zhang, Qianshi Zhang, Yifei Xuan, Weijuan Cheng, Lixing Shi, Yuanhan Wang, Hailin Wen, and Ying Du*
Author Affiliations
  • College of Information Science and Engineering, Zhejiang University, Hangzhou 310023, Zhejiang , China
  • show less
    Figures & Tables(13)
    FBG sensor configuration. (a) Grating arrangement; (b) sensor cross-section
    Sensor spectral shift error. (a) FBG spectral cumulative variation in multiple measurements; (b) spectral drift of two periodic experiments
    Comprehensive technical framework for FBG-based shape reconstruction. (a) Experimental setup and data acquisition; (b) signal preprocessing; (c) neural network architecture
    MSEs of STFT, WT, and EMD feature extraction methods. (a) MSE comparison; (b) MSE variations
    Schematic diagram of hyperparameter optimization by PSO and comparison with GA and BO methods. (a)‒(d) Hyperparameter optimization by PSO; (e) feature importance analysis; (f) MSE comparison with GA and BO
    Overall methodology framework
    Training-validation loss convergence profiles
    MSE distributions among model configurations in ablation study
    Curvature prediction results analysis. (a) Box plot of comparison for 4D actual curvature and predicted curvature; (b) residual kernel density of predicted curvature
    Comparison of true values and reconstructed values under different models. (a) Constant-curvature circular arc curve; (b) variable-curvature circular arc curve; (c) variable-curvature S-shaped curve; (d) reverse S-shaped curve
    • Table 1. Mean error range at center wavelength for four FBG

      View table

      Table 1. Mean error range at center wavelength for four FBG

      SensorCumulative error in the first five days /nmCumulative error in the last five days /nm
      FBG 1[-1.10, -0.28][-0.40, 0.46]
      FBG 2[-2.32, -0.64][-0.40, 0.44]
      FBG 3[-2.58, -1.10][-1.34, 1.98]
      FBG 4[-1.46, -1.22][-2.40, 2.04]
    • Table 2. Core parameters used in final training of model

      View table

      Table 2. Core parameters used in final training of model

      ParameterSetting
      BiLSTM unit170
      LSTM unit66
      Dropout rate0.1178
      Loss functionMSE
      OptimizerAdam
      Number of epochs200
      Learning rate0.0039
      Batch size64
    • Table 3. Comparison of error performance across different shapes for various models

      View table

      Table 3. Comparison of error performance across different shapes for various models

      ShapeFrenetCNNLSTM-SelfAttentionBiLSTM-LSTM
      MAE/TEMAE/TEMAE/TEMAE/TE
      Shape 115.7609/47.839010.6135/28.96610.3527/1.07980.1391/0.6545
      Shape 214.4787/53.54321.6732/6.44062.4054/11.81702.5222/6.2552
      Shape 320.5649/56.12633.7366/7.35821.5599/5.12971.7354/4.3629
      Shape 418.2307/55.04261.5148/6.21612.7448/3.98291.2951/2.2291
      Best MSE26.43028.38810.8295
      Average MSE33.219412.84701.4202
      R20.74930.93680.9918
    Tools

    Get Citation

    Copy Citation Text

    Ming Zhang, Qianshi Zhang, Yifei Xuan, Weijuan Cheng, Lixing Shi, Yuanhan Wang, Hailin Wen, Ying Du. Spectral Drift Correction for FBG-based Shape Sensing using a Hybrid BiLSTM-LSTM Neural Network[J]. Acta Optica Sinica, 2025, 45(16): 1606009

    Download Citation

    EndNote(RIS)BibTexPlain Text
    Save article for my favorites
    Paper Information

    Category: Fiber Optics and Optical Communications

    Received: Mar. 31, 2025

    Accepted: May. 18, 2025

    Published Online: Aug. 15, 2025

    The Author Email: Ying Du (duying@zjut.edu.cn)

    DOI:10.3788/AOS250818

    CSTR:32393.14.AOS250818

    Topics