Xinghu Fu
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Previous Organizations / Institutes
The organizations / institutes from the scholar's other articles in the past ten years
Research Subjects
Xinghu Fu has a total of 190 co-authors. Here are the top co-authors by number of shared publications:
Wa Jin
41 shared publicationsWeihong Bi
29 shared publications- Weihong Bi
School of Information Science and Engineering, The Key Laboratory for Special Fiber and Fiber Sensor of Hebei Province, Yanshan University, Qinhuangdao, China
27 shared publications - Guangwei Fu
School of Information Science and Engineering, The Key Laboratory for Special Fiber and Fiber Sensor of Hebei Province, Yanshan University, Qinhuangdao, China
23 shared publications Guangwei Fu
22 shared publicationsLijiao Zu
12 shared publications
Latest Publications
- 12 October 2026
- Published by Elsevier in Optics & Laser Technology
- Vol. 204
A High-Accuracy human posture prediction based on Quasi-Distributed flexible FBG arrays and random forest algorithm
- 12 October 2026
- Published by Elsevier in Optics & Laser Technology
- Vol. 204
A High-Accuracy human posture prediction based on Quasi-Distributed flexible FBG arrays and random forest algorithm
- Open Access
- 16 September 2026
- Published by Elsevier in SSRN Electronic Journal
High-Sensitivity Vibration Sensor Based on Inter-Modal Interference in a Tapered Few-Mode Fiber Cantilever
In this paper, a tapered all-fiber cantilever vibration sensor based on an inter-modal interference mechanism is proposed and experimentally validated. The sensor features a micro/nano tapered structure fabricated at the tip of a few-mode fiber (FMF) through a combined process of oxy-hydrogen flame thermal tapering and hydrofluoric acid (HF) wet chemical etching. The tapered region breaks the total internal reflection condition of the fiber, enabling the fundamental core mode to leak into the cladding and excite higher-order modes, thereby significantly enhancing the sensitivity to external perturbations. Under dynamic vibration, the cantilever undergoes bending deformation, which synchronously modulates the effective refractive index difference between modes and the effective interference cavity length, resulting in a wavelength drift of the interference spectrum. By locking the laser wavelength at the quadrature operating point (Q-point) of the interference spectrum, a linear phase-to-intensity conversion is achieved, ultimately yielding a high-fidelity voltage output via a photodetector. Experimental results demonstrate that the sensor exhibits an excellent linear frequency response over the range of 70 Hz to 1500 Hz with a coefficient of determination R²=0.999.Its dynamic acceleration sensitivity reaches 129.122mV/(m/s²), representing an approximate 4-fold enhancement compared to an untapered FMF counterpart.Under a weak acceleration excitation of 0.05m/s², a signal-to-noise ratio (SNR) of 7.24dB is achieved, indicating a limit of detection (LOD) better than 0.05 m/s², and the SNR reaches 47.5dB at 3.0m/s². The sensor also demonstrates excellent long-term stability and measurement repeatability. Featuring a compact structure, high sensitivity, and immunity to electromagnetic interference, this device holds great promise for applications in equipment condition monitoring and weak vibration detection.
- 16 September 2026
- Published by Elsevier in SSRN Electronic Journal
Open AccessHigh-Sensitivity Vibration Sensor Based on Inter-Modal Interference in a Tapered Few-Mode Fiber Cantilever
In this paper, a tapered all-fiber cantilever vibration sensor based on an inter-modal interference mechanism is proposed and experimentally validated. The sensor features a micro/nano tapered structure fabricated at the tip of a few-mode fiber (FMF) through a combined process of oxy-hydrogen flame thermal tapering and hydrofluoric acid (HF) wet chemical etching. The tapered region breaks the total internal reflection condition of the fiber, enabling the fundamental core mode to leak into the cladding and excite higher-order modes, thereby significantly enhancing the sensitivity to external perturbations. Under dynamic vibration, the cantilever undergoes bending deformation, which synchronously modulates the effective refractive index difference between modes and the effective interference cavity length, resulting in a wavelength drift of the interference spectrum. By locking the laser wavelength at the quadrature operating point (Q-point) of the interference spectrum, a linear phase-to-intensity conversion is achieved, ultimately yielding a high-fidelity voltage output via a photodetector. Experimental results demonstrate that the sensor exhibits an excellent linear frequency response over the range of 70 Hz to 1500 Hz with a coefficient of determination R²=0.999.Its dynamic acceleration sensitivity reaches 129.122mV/(m/s²), representing an approximate 4-fold enhancement compared to an untapered FMF counterpart.Under a weak acceleration excitation of 0.05m/s², a signal-to-noise ratio (SNR) of 7.24dB is achieved, indicating a limit of detection (LOD) better than 0.05 m/s², and the SNR reaches 47.5dB at 3.0m/s². The sensor also demonstrates excellent long-term stability and measurement repeatability. Featuring a compact structure, high sensitivity, and immunity to electromagnetic interference, this device holds great promise for applications in equipment condition monitoring and weak vibration detection.
- Open Access
- 1 September 2026
- Published by Elsevier in SSRN Electronic Journal
Visible-near-Infrared Hyperspectral Imaging for Non-contact Current Quantification of Surface-mount Resistors using Deep Learning
Conventional automatic optical inspection has limited capability to characterize the operating current of surface-mount resistors on printed circuit boards (PCBs), while complex backgrounds can introduce substantial interference into the target spectra. To address these issues, this study proposes a non-contact current quantification method that integrates hyperspectral imaging, SegNeXt-based semantic segmentation, and ResNet regression. Experiments were conducted under nine current conditions ranging from 0 to 40 mA, and 18 original hyperspectral datasets were acquired and subsequently subjected to region-of-interest (ROI) extraction, black-and-white reference correction, and data augmentation. Spectral analysis showed that the near-infrared region of 830–920 nm was more sensitive to current variations, with pronounced responses observed around 879.19 and 903.35 nm. For SegNeXt segmentation, MSC+GF achieved the best preprocessing performance, yielding PA, mAcc, IoU, mIoU, and Dice values of 98.5%, 97.8%, 96.5%, 95.9%, and 97.2%, respectively. For ResNet-based current quantification, BC-ResNet achieved a RMSEP of 0.050657 and an R²P of 0.820570. After further constructing the MSC+GF+SegNeXt+BC+ResNet cascade, the RMSEP decreased to 0.039364, while R²P increased to 0.900579. These results demonstrate that target segmentation combined with stage-specific spectral preprocessing can effectively suppress interference from the PCB background and improve the accuracy of current prediction for surface-mount resistors, providing a feasible approach for non-contact hyperspectral monitoring of electronic-component operating states.
- 1 September 2026
- Published by Elsevier in SSRN Electronic Journal
Open AccessVisible-near-Infrared Hyperspectral Imaging for Non-contact Current Quantification of Surface-mount Resistors using Deep Learning
Conventional automatic optical inspection has limited capability to characterize the operating current of surface-mount resistors on printed circuit boards (PCBs), while complex backgrounds can introduce substantial interference into the target spectra. To address these issues, this study proposes a non-contact current quantification method that integrates hyperspectral imaging, SegNeXt-based semantic segmentation, and ResNet regression. Experiments were conducted under nine current conditions ranging from 0 to 40 mA, and 18 original hyperspectral datasets were acquired and subsequently subjected to region-of-interest (ROI) extraction, black-and-white reference correction, and data augmentation. Spectral analysis showed that the near-infrared region of 830–920 nm was more sensitive to current variations, with pronounced responses observed around 879.19 and 903.35 nm. For SegNeXt segmentation, MSC+GF achieved the best preprocessing performance, yielding PA, mAcc, IoU, mIoU, and Dice values of 98.5%, 97.8%, 96.5%, 95.9%, and 97.2%, respectively. For ResNet-based current quantification, BC-ResNet achieved a RMSEP of 0.050657 and an R²P of 0.820570. After further constructing the MSC+GF+SegNeXt+BC+ResNet cascade, the RMSEP decreased to 0.039364, while R²P increased to 0.900579. These results demonstrate that target segmentation combined with stage-specific spectral preprocessing can effectively suppress interference from the PCB background and improve the accuracy of current prediction for surface-mount resistors, providing a feasible approach for non-contact hyperspectral monitoring of electronic-component operating states.
- Open Access
- 1 September 2026
- Published by Elsevier in SSRN Electronic Journal
Hyperspectral Geographical Origin Identification of Bupleurum Powder by Integrating Three-Dimensional Residual Convolution and a Transformer
Variations in the geographical origin of Bupleurum may affect its quality, making origin identification essential for traceability and quality control. This study proposes a hyperspectral image classification method that integrates three-dimensional residual convolution and a Transformer for rapid, nondestructive origin identification of powdered Bupleurum. The method takes the complete hyperspectral data cube as input, uses three-dimensional residual convolution to extract local spatial–spectral features, adaptively weights them via a spectral attention module, converts them into Token sequences, and employs a Transformer encoder to model global dependencies. Meanwhile, a masked autoencoder is introduced for self-supervised pretraining to learn spatial–spectral representations from unlabeled data, alleviating the risk of overfitting under limited samples. To evaluate the contribution of each module, three model configurations were established, and five spectral input schemes and two three-dimensional convolutional front-end architectures were compared. On a dataset comprising 280 independent powdered samples from seven geographical origins, the combination of Raw input, a lightweight three-dimensional residual network, and the Full model achieved the best performance, with a five-fold cross-validation average accuracy of 98.57%, a macro-average F1 of 98.57%, and a Kappa of 0.9833. Ablation experiments demonstrated that the Transformer encoder consistently improved performance across different conditions (average accuracy gain of 2.22 percentage points), and MAE pretraining provided stable positive gains under the residual network architecture. This method provides a feasible technical reference for origin identification of powdered Chinese medicinal materials and spatial–spectral feature learning from hyperspectral data under limited samples.
- 1 September 2026
- Published by Elsevier in SSRN Electronic Journal
Open AccessHyperspectral Geographical Origin Identification of Bupleurum Powder by Integrating Three-Dimensional Residual Convolution and a Transformer
Variations in the geographical origin of Bupleurum may affect its quality, making origin identification essential for traceability and quality control. This study proposes a hyperspectral image classification method that integrates three-dimensional residual convolution and a Transformer for rapid, nondestructive origin identification of powdered Bupleurum. The method takes the complete hyperspectral data cube as input, uses three-dimensional residual convolution to extract local spatial–spectral features, adaptively weights them via a spectral attention module, converts them into Token sequences, and employs a Transformer encoder to model global dependencies. Meanwhile, a masked autoencoder is introduced for self-supervised pretraining to learn spatial–spectral representations from unlabeled data, alleviating the risk of overfitting under limited samples. To evaluate the contribution of each module, three model configurations were established, and five spectral input schemes and two three-dimensional convolutional front-end architectures were compared. On a dataset comprising 280 independent powdered samples from seven geographical origins, the combination of Raw input, a lightweight three-dimensional residual network, and the Full model achieved the best performance, with a five-fold cross-validation average accuracy of 98.57%, a macro-average F1 of 98.57%, and a Kappa of 0.9833. Ablation experiments demonstrated that the Transformer encoder consistently improved performance across different conditions (average accuracy gain of 2.22 percentage points), and MAE pretraining provided stable positive gains under the residual network architecture. This method provides a feasible technical reference for origin identification of powdered Chinese medicinal materials and spatial–spectral feature learning from hyperspectral data under limited samples.
- 3 April 2026
- Published by Elsevier in Measurement
- Vol. 275
A fiber-optic SERS probe based on femtosecond laser-processed square microcavity with high-sensitivity detection of antiepileptic drugs
- 3 April 2026
- Published by Elsevier in Measurement
- Vol. 275
A fiber-optic SERS probe based on femtosecond laser-processed square microcavity with high-sensitivity detection of antiepileptic drugs
- 14 March 2026
- Published by Elsevier in Optics & Laser Technology
- Vol. 200
A torsion sensor based on a tapered and cascaded helical four-core fiber
- 14 March 2026
- Published by Elsevier in Optics & Laser Technology
- Vol. 200
A torsion sensor based on a tapered and cascaded helical four-core fiber
- 13 March 2026
- Published by Optica Publishing Group in Optics Express
- Vol. 34  (6) , 10276
Double-layer staggered IWFBGs enable multi-dimensional tactile sensing on scalable flexible skin
Flexible tactile skins with dense, multimodal sensing are important for dexterous robots and wearable human-machine interfaces, yet electrical and magnetic sensors often suffer from electromagnetic interference and complex wiring, while existing optical designs have limited functionality or scalability. We present a flexible tactile skin based on double-layer staggered identical weak fiber Bragg gratings (IWFBGs) interrogated by optical frequency-domain reflectometry (OFDR) for simultaneous sensing of contact location, normal force, and contact orientation. A simulation-driven design framework combining finite element analysis and neural networks is used to compare four IWFBG layouts and identify the double-layer staggered architecture as optimal. To suppress temperature drift and laser instability, we further introduce a MAD-assisted moving-average detrending algorithm that improves the linearity of wavelength–force responses. Subsequently, this study implemented tactile information prediction on a 70 mm × 70 mm sensor using a backpropagation (BP) neural network. In terms of localization, a root mean square error of 3.2–3.5 mm was achieved, while force sensing attained a mean absolute error of 0.27 N. Directional accuracy across four contact directions reached 88%. Furthermore, by further testing the predictive performance of the BP algorithm on unseen locations of the sensor, it is demonstrated that the system possesses spatially continuous resolution capability, highlighting its potential for optical tactile skin applications.
- 13 March 2026
- Published by Optica Publishing Group in Optics Express
- Vol. 34  (6) , 10276
Double-layer staggered IWFBGs enable multi-dimensional tactile sensing on scalable flexible skin
Flexible tactile skins with dense, multimodal sensing are important for dexterous robots and wearable human-machine interfaces, yet electrical and magnetic sensors often suffer from electromagnetic interference and complex wiring, while existing optical designs have limited functionality or scalability. We present a flexible tactile skin based on double-layer staggered identical weak fiber Bragg gratings (IWFBGs) interrogated by optical frequency-domain reflectometry (OFDR) for simultaneous sensing of contact location, normal force, and contact orientation. A simulation-driven design framework combining finite element analysis and neural networks is used to compare four IWFBG layouts and identify the double-layer staggered architecture as optimal. To suppress temperature drift and laser instability, we further introduce a MAD-assisted moving-average detrending algorithm that improves the linearity of wavelength–force responses. Subsequently, this study implemented tactile information prediction on a 70 mm × 70 mm sensor using a backpropagation (BP) neural network. In terms of localization, a root mean square error of 3.2–3.5 mm was achieved, while force sensing attained a mean absolute error of 0.27 N. Directional accuracy across four contact directions reached 88%. Furthermore, by further testing the predictive performance of the BP algorithm on unseen locations of the sensor, it is demonstrated that the system possesses spatially continuous resolution capability, highlighting its potential for optical tactile skin applications.
- Open Access
- 3 February 2026
- Published by Elsevier in SSRN Electronic Journal
Optical Coherent Gas Absorption Spectroscopy based on Space-Frequency Joint Multiplexing for Multipoint Detection
Multipoint gas detection in complex industrial environments faces inherent challenges in balancing the trade-offs among limit of detection (LOD), response speed, and spatial coverage in traditional sensor networks. To address these constraints, this paper proposes an optical coherent gas absorption spectroscopy (OC-GAS) system based on a space-frequency joint multiplexing mechanism. By employing fiber delay lines to map spatially distributed sensing units to distinct beat frequencies, the system enables precise spatial addressing and parallel demodulation, effectively resolving the conflict between multipoint capacity and real-time performance. A theoretical model for light field transmission is established, and a four-channel experimental system is constructed for acetylene C2H2 detection. Experimental results demonstrate a coefficient of determination (R2=0.99978) over a wide dynamic range of 0–10,000 ppm, achieving a limit of detection (LOD) of 45.7 ppm. Furthermore, the system exhibits exceptional crosstalk immunity (> 47 dB), ensuring signal integrity even under extreme concentration surges. Consequently, the proposed method offers a robust parallel star-topology paradigm that simultaneously satisfies the requirements for high sensitivity, fast response, and distributed monitoring in complex industrial environments.
- 3 February 2026
- Published by Elsevier in SSRN Electronic Journal
Open AccessOptical Coherent Gas Absorption Spectroscopy based on Space-Frequency Joint Multiplexing for Multipoint Detection
Multipoint gas detection in complex industrial environments faces inherent challenges in balancing the trade-offs among limit of detection (LOD), response speed, and spatial coverage in traditional sensor networks. To address these constraints, this paper proposes an optical coherent gas absorption spectroscopy (OC-GAS) system based on a space-frequency joint multiplexing mechanism. By employing fiber delay lines to map spatially distributed sensing units to distinct beat frequencies, the system enables precise spatial addressing and parallel demodulation, effectively resolving the conflict between multipoint capacity and real-time performance. A theoretical model for light field transmission is established, and a four-channel experimental system is constructed for acetylene C2H2 detection. Experimental results demonstrate a coefficient of determination (R2=0.99978) over a wide dynamic range of 0–10,000 ppm, achieving a limit of detection (LOD) of 45.7 ppm. Furthermore, the system exhibits exceptional crosstalk immunity (> 47 dB), ensuring signal integrity even under extreme concentration surges. Consequently, the proposed method offers a robust parallel star-topology paradigm that simultaneously satisfies the requirements for high sensitivity, fast response, and distributed monitoring in complex industrial environments.