For improving the performance of vibration discrimination in terms of dual Mach–Zehnder interferometric (DMZI) perimeter system, we proposed a novel method based upon the multifractal theory, serial feature fusion as well as improved probabilistic neural network (PNN). By the multifractal theory, the features of original signal are extracted in the form of multifractal spectrum parameters, thereby constructing the feature vector by the method of serial feature fusion (SFF). Then, we employ the simulated annealing algorithm to automatically optimize the smoothing factor of PNN, which can avoid manual selection of empirical thresholds in the process of feature extraction and pattern recognition. Finally, with the simulated annealing-based PNN, the intrusion signal of DMZI system can be discriminated and classified. Compared with traditional methods based upon signal decomposition, our method focuses on the morphological characteristics of original signal, therefore possesses a better ability of detail discrimination. In the DMZI perimeter system, four types of real vibration intrusions are completed to verify the proposed method. The results demonstrate that it is superior to the conventional methods, with an average discrimination rate of over 95%.

为了提高双马赫-曾德干涉（DMZI）周界系统的振动识别性能，我们提出了一种基于多重分形理论、序列特征融合以及改进的概率神经网络（PNN）的新方法。利用多重分形理论，以多重分形谱参数的形式提取原始信号的特征，从而通过串行特征融合（SFF）的方法构建特征向量。然后，我们采用模拟退火算法自动优化 PNN 的平滑因子，避免在特征提取和模式识别过程中手动选择经验阈值。最后，利用基于模拟退火的PNN，可以对DMZI系统的入侵信号进行判别和分类。与基于信号分解的传统方法相比，我们的方法侧重于原始信号的形态特征，因此具有更好的细节辨别能力。在DMZI周界系统中，完成了四种真实的振动侵入，以验证所提出的方法。结果表明，它优于传统方法，平均识别率超过95%。