Classification of Distributed Fiber Optic Vibration Signals Using the Diff-ViT Model

A novel method based on ultra-weak fiber grating (UWFBG) arrays and the Differential Vision Transformer (Diff-ViT) model was proposed for the recognition of distributed optical vibration events. Firstly, a distributed optical fiber grating array vibration sensing system based on the phase-sensitive optical time-domain reflectometer (Φ-OTDR) was used to collect waveforms of six types of vibration events, including Digging, Sawing, and Shaking, with a total of 5513 samples. Then, the collected vibration waveforms were converted into two-dimensional images through the Gramian Angular Field (GAF). Finally, the improved Diff-ViT model was used to perform feature recognition on the converted two-dimensional images to complete the classification of the vibration events. The ViT network model has a larger receptive field compared to the traditional CNN model, which can capture the global context information of the entire perturbation event. However, it also allocates certain attention to irrelevant contextual information, which introduces attention noise and results in a decrease in recognition accuracy. Therefore, the Differential Transformer mechanism is introduced to replace ViT’s traditional attention mechanism. It achieves denoising by introducing two groups of attention weight matrices and performing differential amplification, analogous to differential amplifiers in electrical engineering. Experimental results show that the average recognition accuracy of the traditional ViT model is 92.92%, while the Diff-ViT model achieves a recognition accuracy of 96.64%. This indicates that the improved model pays more attention to key features, thereby improving the classification recognition accuracy.