Displacement-Sensing Method Based on Residual Scaling for One-Shot MMF Specklegram Prediction

A high-resolution and wide measurement range displacement sensing method based on multimode fiber (MMF) is proposed. To achieve a high-resolution displacement detection model, a one-shot dataset was constructed by collecting MMF specklegram images for 1801 displacements with resolution of 0.01 mm. This work modifies the fully connected layer of a residual network (ResNet) to achieve displacement prediction and applies residual scaling to reduce prediction errors in the one-shot learning task. Under stable environmental conditions, experimental results show that this method achieves an average error as low as 0.0083 mm in displacement prediction with resolution of 0.01 mm; meanwhile, the measurement range reaches 18 mm. Additionally, the model trained on a 0.01 mm resolution dataset was evaluated on a specklegram dataset with a resolution of 0.005 mm for its generalization ability, yielding an average error of 0.0138 mm. Regression evaluation metrics demonstrate that the proposed model has a significant improvement over other displacement-sensing methods based on MMF specklegrams, with prediction errors approximately three times lower than ResNet. Additionally, temperature immunity was studied within an 18 mm measurement range under a temperature range from 21.25 °C to 22.35 °C; the MMF displacement sensor demonstrates a dispersion of 5.08%, an average nonlinearity of 7.71% and a hysteresis of 6.13%. These findings demonstrate the potential of this method for high-performance displacement-sensing in practical applications.