Interpretable Pruned SpinalNet Network Regressor for Directional Bend Sensing Behavior Prediction of Dumbbell-Core Fiber Sensor

Refractive index-insensitive direction bend sensors made of specialty microstructured optical fibers (MOFs) play important roles in structural health monitoring robotics, medical sensing, IoT, aeronautics, and underwater technologies. However sensor performance optimization is computationally expensive to optimize through Finite Element Method (FEM) based COMSOL simulations. Hence this research focuses on contriving a significant feature pruned Interpretable SpinalNet Neural network regression model to predict the optical transmission spectrum of a fabricated dumbbell-shaped dual-core MOF bend sensor. Initially the baseline Dense neural network regressor with 47,809 trainable parameters trained on MOF fabricated sensor experimental dataset without augmentation achieved an R2score of 0.9218. After Autoencoder based data augmentation the model’s R2score was increased to 0.9480 with a corresponding 95% confidence interval ranging from 0.9443 to 0.9515. A baseline SpinalNet regressor model trained on augmented dataset by gradually feeding the input feature representations at successive stages was able to achieve an R2score of 0.9729 and a 95% confidence interval of 0.9703 to 0.9752, though it required 88,065 trainable parameters. A more efficient Tuned SpinalNet variant regressor built with hyperparameter tuning presented a superior predictive performance with an R2score of 0.9748 and a 95% confidence interval of 0.9723–0.9772 with only 50,689 parameters. In addition, interpreting and identifying the significant feature was performed with Shapley Additive Explanations (SHAP) approach, paved a way to contrive a feature pruned regressor variant namely Interpretable Pruned SpinalNet model with competitively high R2score of about 0.9736 and a 95% confidence interval of 0.9721 to 0.9770 by comprising only 44,545 trainable parameters. Surprisingly, the proposed model demonstrated strong generalization when evaluated on an unseen data set for the bend radius of 3.79 mm. Thus, the proposed Interpretable pruned SpinalNet model offers an efficient, fast and precise method to characterize the sensing behavior of real-time MOF bend sensor.