A feature extraction and recognition algorithm for fiber intrusion signals is proposed based on ensemble empirical-mode decomposition (EEMD) coupled with a random vector-function linked (RVFL) neural network to accurately identify the type of intrusion signal on a distributed optical fiber. The proposed algorithm starts with the preprocessing for the collected fiber intrusion signals,including minimum-maximum normalization processing and the removal of low frequency noise using the db3 wavelet. Then, the intrusion signals are decomposed by the EEMD to obtain five groups of intrinsic mode functions (IMF). Subsequently, the energy ratio of each component of the IMF is calculated, and three feature vectors are filtered using the analysis of variance. Finally, the feature vectors are sent into the RVFL neural network to be trained for the completion of the signal recognition. The experimental results validate that the proposed algorithm can accurately distinguish between different intrusion signals with high recognition rate.