Efficient On-Demand Design of Fiber Optic Vibration Sensor With a Symmetric Bidirectional Neural Network

Fiber optic sensors, prized for their lightweight, compact size, high-temperature resilience, and resistance to electromagnetic interference, find extensive utility in various measurement applications. The performance of these sensors is primarily contingent on their sensitive units, with distinct structures of these units yielding varied performance outcomes. Traditional design methods primarily rely on finite element simulation and optimization, which are subjective and inefficient. Thus, the efficient on-demand design of sensitive structures is essential for different application scenarios. Here, we present a novel approach for both forward performance prediction and inverse structure design employing deep learning techniques based on symmetric bidirectional neural networks (SBNNs), with fiber optic vibration sensors serving as a design example. The proposed method can address the nonunique solution in traditional deep learning techniques for inverse design of 3-D complex structures. By learning the underlying relationships between complex nonintuitive sensitive structures and their performances, the approach can eliminate the need for numerous costly calculations that heavily depend on human experience or intuition. Furthermore, compared to the response surface optimization (RSO) method, this approach saves
$21.1\times $
computation time and has an accuracy improvement of 34.6% when dealing with six samples. The results show that the efficiency of fiber optic sensor design can be significantly improved by employing this novel deep learning technique, which offers new insights for the rapid advancement of the fiber optic sensing field.