Because of the cross sensitivity and other uncontrollable factors, the information data appear abnormal and result in the large deviation of information analysis and low recognition accuracy when using traditional monochannel optical sensing fiber to achieve the measurement. A feature extraction and recognition method based on bicoherence spectrum, sample entropy and singular value decomposition (BSS) and back propagation neural network (BPNN)-Dempster Shafer(DS) is proposed. Assuming the intrusion detection system contains three optic sensing fibers based on the Brillouin optical time domain reflection (BOTDR), the method utilizes the BSS algorithm to extract the different intrusion features of multiplex signal, respectively. The classification of the feature vectors for different intrusion vibrations is realized by using the BPNN algorithm and the spatio-temporal information fusion of multi sensing fibers is acquired by Dempster Shafer (DS) evidence theory. Numerical analysis and simulation results show that the novel method can effectively extract the information characteristics of multi-channel sensor networks and have higher accuracy and credibility based on BPNN-DS evidence theory compared with the monochannel optical sensing fiber. This multi-channel information fusion algorithm can also identify signal types of multiintrusion sensor networks accurately.