Hybrid AI–Polynomial Compensation Framework for Multi-Parameter Error Mitigation in Fiber Bragg Grating Sensors

Abstract

FBG networks fit much into a small space. They are strain gauges, temperature gauges, pressure gauges, and even refractive index gauges, and they do it repeatedly and reliably even in harsh conditions. However, when you play them on the field, it gets dirty. It is contaminated with systematic errors, cross-sensitivities, thermomechanical effects, interrogator hysteresis, nonlinearity of wavelength, and packaging effects. These are not such trifles. They pose a danger to accuracy particularly where measurements count. The real challenge? Paying makes the readings credible once more. In this case, the extra techniques are needed. Learn data using advanced techniques and physics-based polynomial models, and you can learn quickly with interpretable compensation even on hardware with limited resources. Such is the concept of the hybrid frame presented here. Deeper down is an AI regressor constructed using radial-basis expansions that have a ridge regularization. This regressor addresses the multi-parameter corrections which are complex. After that, an elementary poly block intervenes, forcing up the structure you would get as an application of the Bragg condition and known cross-terms. This arrangement specifically addresses such problems as temperature-strain coupling, drift of the interrogator, and sluggish hysteresis of tunable filters. It maintains the calibration even when the ambient conditions vary. In this research, the experiments are conducted with the use of a simulated dataset that is built on the statistic properties of publicly available FBG datasets such as the PHM Society 2015 FBG Data Challenge dataset, the Kaggle FBG Spectrum Dataset, and the IEEE Dataport FBG Time-Series Dataset. So, how does it stack up? Pitting this hybrid approach against an approach that simply uses a plain polynomial compensator you find obvious improvements in the form of lower root-mean-square error (RMSE) and reduced bias in the event of drift and narrower confidence bands around the reconstructed values. Practically, the paper has achieved three items, namely (i) a concise hybrid AI-polynomial design that makes the residual separation explicit; (ii) a production that can produce seven diagnostic plots that can be audited with ease; and (iii) a well- Benchmarked workflow. The bottom line: the hybrid method reduces RMSE, when compared to conventional methods of polynomies, by 25-45 % across all channels, and it also provides transparent coefficients which can be used to recalibrate fields and verify quality assurance.