Fig. 1. Schematic diagram of the design and principle of the DHF sensor. (a) The 3D schematic diagram of the proposed sensor. (b) The schematic illustration of the sensing principle. (c) Cross section of the fabricated DHF.

Fig. 2. Numerical and experimental results of the sensor. (a) The side beam propagation diagram of the numerical simulation of the device. (b) The light intensity distribution at different spaces between DHF and CCD. (c) The experimental transmission results when the distance between the end of the DHF and the receiving CCD is 3 cm.

Fig. 3. The influence of temperature changes on the light spots.

Fig. 4. Bending experimental principle and results. (a) The light of the customized laser (980 nm) goes through the SMF to the DHF, and the final signal was recorded by the CCD (Ncam). The bending radius can be controlled by regulating the distance of ds. As for the different bending directions, it can be achieved by rotating the DHF. (b) The light spot patterns corresponding to bending at different directions with the same radius. d₁, d₂, d₃, and d₄ represent the four directions of “up,” “left,” “down,” “right”, respectively.

Fig. 5. Deep learning network based on 2D-CNN for the multi-direction sensor prediction and the demodulation model. 1280 groups of spot data are input into the model according to the different feature areas of each group of data.

Fig. 6. Deep learning network based on 2D-CNN for the multi-direction bending sensor demodulation results. (a) The root mean square error (RMSE) and loss of model training data and validation data. (b) Using the trained 2D-CNN model to verify the accuracy of the bending direction and bending angle for 128 sets of test data simultaneously. One-hot coding has been carried out in different directions. The 0.01–0.09 mm corresponding to ds is also encoded as 1–9, which facilitates model convergence at the same scale of magnitude. (c) Accuracy verification after sampling different regions of the same input image.