Research on Deep Learning-based Anomaly Event Detection Using Fiber Optic Vibration Data

This study aims to develop and evaluate machine learning and deep learning models for classifying electrical signals collected by fiber optic sensors under three conditions: calm, shaking, and pressing. First, time-domain and frequency-domain features were extracted, and key features were selected using the recursive feature elimination method to reduce dimensionality. Subsequently, support vector machine (SVM), linear discriminant analysis (LDA), and long short-term memory (LSTM) models were constructed based on the time-frequency domain features of the signals. Additionally, gated recurrent unit (GRU) and convolutional neural network (CNN) models were constructed and improved based on the complete signal data. The results show that the CNN model performed best in terms of classification accuracy and generalization ability, with a validation accuracy of 96.81%, low risk of overfitting, and good classification balance. Furthermore, by optimizing the GRU model, the recognition ability of minority classes was significantly improved, providing valuable experience for the application of the GRU model.