Ziyang Xiao
Ziyang Xiao has a total of 5 co-authors. Here are the top co-authors by number of shared publications:
- Haomao Wang
State Key Laboratory of Information Photonics and Optical Communications, Beijing University of Posts and Telecommunications, Beijing, China
1 shared publication Luming Li
State Grid JiangXi Information and Telecommunication Branch, Nanchang, China
1 shared publicationRui Zhou
University of Science and Technology Beijing, Beijing, China
1 shared publicationYifan Wang
State Key Laboratory of Information Photonics and Optical Communications, Beijing University of Posts and Telecommunications, Beijing, China
1 shared publication- Zhiguo Zhang
School of Electronic Engineering, Beijing University of Posts and Telecommunications, Beijing, China
1 shared publication
Latest Publications
- 20 February 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Sensors Journal
- Vol. 26 (7) , 10394-10414
OFDR Machine Learning Approach With Principal Component Analysis and AutoGluon AutoML Framework for High-Precision Distributed Strain Sensing
Optical Frequency Domain Reflectometry (OFDR) has extensive application prospects in structural health monitoring, but faces technical challenges in simultaneously achieving high-precision continuous strain monitoring and large strain monitoring capabilities. This study proposes a novel OFDR machine learning approach targeting the requirements of structural health monitoring, aiming to develop a model with high-precision capabilities for both continuous strain and large strain demodulation. Based on the characteristics of OFDR local spectra, the proposed sequence exchange augmentation and coefficient scaling augmentation methods help the model focus on intrinsic spectral variation patterns rather than surface intensity fluctuations, enhancing generalization capability and prediction stability. The PCA dimensionality reduction method provides a novel interpretive perspective for spectral data, abstracting time-correlated spectral data into principal component feature information and extracting novel spectral features and variation patterns. This method achieves a 90% reduction in dimensions while preserving over 99% of the cumulative explained variance, providing machine learning models with a more structured and information-dense feature space. The weighted ensemble capability of the AutoGluon framework can integrate different high-performance models, maintaining excellent demodulation performance in both continuous strain and large strain scenarios by synthesizing the predictive capabilities of various models. Furthermore, the framework’s automated processing and multi-layer stacking architecture efficiently explore model combinations, significantly improving training efficiency and providing researchers with more comprehensive model performance perspectives and data evidence. The trained weighted ensemble model achieved a Mean Absolute Error (MAE) of 2.4 in continuous strain measurement and an MAE of 16.4 under 3000μϵ large strain conditions, simultaneously possessing high-precision demodulation capabilities for both continuous varying strain and large strain scenarios. This study provides a new technical paradigm for the deep integration of OFDR with high-performance machine learning algorithms, offering efficient and effective solutions for the entire OFDR machine learning pipeline from data augmentation to data processing, model training and optimization. The approach meets structural health monitoring application requirements and provides technical support for OFDR system applications in this field.
- 20 February 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Sensors Journal
- Vol. 26 (7) , 10394-10414
OFDR Machine Learning Approach With Principal Component Analysis and AutoGluon AutoML Framework for High-Precision Distributed Strain Sensing
Optical Frequency Domain Reflectometry (OFDR) has extensive application prospects in structural health monitoring, but faces technical challenges in simultaneously achieving high-precision continuous strain monitoring and large strain monitoring capabilities. This study proposes a novel OFDR machine learning approach targeting the requirements of structural health monitoring, aiming to develop a model with high-precision capabilities for both continuous strain and large strain demodulation. Based on the characteristics of OFDR local spectra, the proposed sequence exchange augmentation and coefficient scaling augmentation methods help the model focus on intrinsic spectral variation patterns rather than surface intensity fluctuations, enhancing generalization capability and prediction stability. The PCA dimensionality reduction method provides a novel interpretive perspective for spectral data, abstracting time-correlated spectral data into principal component feature information and extracting novel spectral features and variation patterns. This method achieves a 90% reduction in dimensions while preserving over 99% of the cumulative explained variance, providing machine learning models with a more structured and information-dense feature space. The weighted ensemble capability of the AutoGluon framework can integrate different high-performance models, maintaining excellent demodulation performance in both continuous strain and large strain scenarios by synthesizing the predictive capabilities of various models. Furthermore, the framework’s automated processing and multi-layer stacking architecture efficiently explore model combinations, significantly improving training efficiency and providing researchers with more comprehensive model performance perspectives and data evidence. The trained weighted ensemble model achieved a Mean Absolute Error (MAE) of 2.4 in continuous strain measurement and an MAE of 16.4 under 3000μϵ large strain conditions, simultaneously possessing high-precision demodulation capabilities for both continuous varying strain and large strain scenarios. This study provides a new technical paradigm for the deep integration of OFDR with high-performance machine learning algorithms, offering efficient and effective solutions for the entire OFDR machine learning pipeline from data augmentation to data processing, model training and optimization. The approach meets structural health monitoring application requirements and provides technical support for OFDR system applications in this field.