@Article{sdhm.2026.087157,
AUTHOR = {Qingmin Hou, Guanghua Xiao, Tongtong Dai},
TITLE = {Reflective Fiber Optic Angle Sensor for Monitoring the Rotation State of Monopolar Photovoltaic Tracking Mounts},
JOURNAL = {Structural Durability \& Health Monitoring},
VOLUME = {20},
YEAR = {2026},
NUMBER = {5},
PAGES = {0--0},
URL = {http://www.techscience.com/sdhm/v20n5/68534},
ISSN = {1930-2991},
ABSTRACT = {To address the issue of angular deviation in inclined single-axis photovoltaic tracking mounts during long-term operation-caused by wind disturbances, mechanical transmission gaps, installation inaccuracies, and environmental factors-a angle sensor based on reflective fiber-optic ranging principles has been developed for monitoring the rotation angle of the main shaft. This sensor employs a structural conversion approach of “using straight lines to replace curves”, transforming the shaft’s rotational angle into linear displacement of a mirror via a gear-rack mechanism, and generating corresponding output voltage signals through variations in reflected light intensity to measure rotation angles. Due to factors such as the optical response characteristics of reflective fiber probes, mechanical transmission gaps, assembly errors, and external environmental disturbances, a significant nonlinear relationship exists between the sensor’s output voltage and rotation angle, making traditional linear fitting methods inadequate for accurately describing this complex input-output mapping relationship. To enhance angle inversion accuracy, this paper proposes a nonlinear modeling and error compensation method utilizing a two-stage neural network. First, a neural network model is established with sensor output voltage as input and standard angle values as output to capture the nonlinear relationship between voltage and angle. Subsequently, an error compensation model is constructed using output voltage and preliminary predicted angles as inputs, with prediction residuals as outputs, to perform secondary refinement of angle inversion results. Experimental results demonstrate that the average nonlinear error of the sensor was 31.96% without compensation; after applying the basic neural network angle inversion model, the nonlinear error decreased to 1.00%; further with two-stage neural network error compensation, it further decreased to 0.31%, representing a reduction of 0.69 percentage points compared to the basic model. On the 15 representative calibration points, the proposed method achieved an MAE of 0.401°, an RMSE of 0.425°, and a maximum error of 0.640°. For the independent measurement dataset used for final evaluation, the corresponding MAE, RMSE, and maximum error were 0.81°, 0.89°, and 1.40°, respectively.},
DOI = {10.32604/sdhm.2026.087157}
}