To effectively identify the vibration signals of the fiber optic perimeter system, a method was presented, which combines the multi-fractal spectrum parameters with the improved probabilistic neural network. This method could avoid the shortcomings of experience threshold selecting in extracting features and smoothing factor determining in the process of pattern recognition. First of all, the existence and validity of multi-fractal in optical fiber vibration signals were examined and analyzed. Then, the multi-fractal spectrum parameters of the fiber vibration signals were calculated and extracted to form the feature vectors which could accurately describe the nonlinear and complexity of the signals. Finally, the improved probabilistic neural network algorithm was used for adaptive learning and classification to realize the identification of different optical fiber vibration signals. Four kinds of vibration signals collected from field tests were used to verify the method and the results show that the average recognition rate reaches 96.25 % and the recognition time is 1.63 s. This method is superior to the traditional probabilistic neural network algorithm in terms of correct recognition rate.