Event pattern recognition of distributed optical fiber sensing system based on FES-RDB-CNN and Voting classifier combination

Event pattern recognition technology has become an important research direction of distributed fiber optic vibration sensors. In this article, an event pattern recognition scheme based on feature-enhanced and simplified residual dense block (FES-RDB)-convolutional neural network (CNN) and voting classifier combination (VCC) is proposed and applied to event pattern recognition for the Sagnac distributed fiber sensing system. The FES-RDB proposed in this article is a new RDB that replaces the convolution block in the RDB with the residual unit in the 34-layer residual nets (ResNet-34) and replaces the ReLU activation function in the ResNet-34 with the Leaky ReLU activation function. By introducing FES-RDB in the feature extraction stage of conventional CNN, the capability of high-dimensional feature extraction, transmission, and reuse of neural networks is greatly improved. The 3-D map obtained by the t-distributed stochastic neighbor embedding (t-SNE) algorithm shows that FES-RDB makes the data points of different types of events have significantly farther distances, more distinct boundaries, and higher aggregation of event data points of the same type. Using the event pattern recognition scheme proposed in this article, the average recognition accuracy of nine types of events reaches 99.46%. Therefore, the event pattern recognition scheme based on FES-RDB-CNN+VCC has excellent performance in practicability and recognition accuracy and has a good application prospect.