The optical fiber early warning system has been widely used in the intrusion detection and early warning of oil and gas pipelines. The current technical difficulty is still how to improve the accuracy of multi-class recognition of optical fiber intrusion signals. In this paper, gradient boosting decision tree (GBDT) algorithm is used to train the multi-classification model of fiber intrusion signal, and a feature extraction and recognition algorithm based on Fourier decomposition method (FDM) and GBDT algorithm is proposed. The algorithm uses FDM to preprocess the fiber intrusion signal, extracts the approximate entropy, energy and spectral entropy characteristics of the signal, and then uses the GBDT algorithm to train the model to identify and classify the fiber intrusion signal. In order to test the performance of the algorithm, use support vector machine and AdaBoost algorithms to train the models and conduct comparative experiments. The results show that the algorithm can effectively identify four types of optical fiber intrusion signals, namely, knocking, trotting, passing and picking, with an average accuracy of 92.5%.