Abstract
Fiber optic vibration identification has significant applications in engineering fields, like security surveillance and structural health assessment. However, present methods primarily depend on either temporal-frequency domain or image features simply, challenging the simultaneous consideration of both image attributes and the temporal dependencies of vibration signals. Consequently, the performance of fiber optic vibration recognition remains subject to improvement, and its effectiveness further diminishes under conditions of uneven data distribution. Therefore, this study integrates residual neural networks, long short-term memory networks, and diffusion denoising probabilistic models to propose a fiber optic vibration recognition method DR-LSTM, which incorporates both image and temporal features while ensuring high recognition accuracy across balanced and imbalanced data distributions. Firstly, features of the Mel spectrum image and temporal characteristics of fiber optic vibration events are extracted. Subsequently, specialized neural network models are developed for categories with scarce data to produce similar images for data augmentation. Finally, the retrieved composite characteristics are employed to train recognition models, thereby improving recognition accuracy. Experiments were performed on datasets from natural environment and anthropogenic vibration, including for both balanced and imbalanced data distributions. The results show that on the two balanced datasets, the proposed model achieves improvements in classification accuracy of at least 0.67% and 7.4% compared to conventional methods. In the two imbalanced datasets, the model's accuracy exceeds that of conventional models by a minimum of 18.79% and 2.4%. This validates the effectiveness and feasibility of DR-LSTM in enhancing recognition accuracy and addressing issues with imbalanced data distribution.