Aiming at the problems of severe noise interference of fiber-optic vibration signals， single feature extraction and long recognition time， an improved local characteristic-scale decomposition and ant colony algorithm optimize deep belief network are proposed. Firstly， cubic B-spline function interpolation is used to fit the mean curve to improve the local characteristic-scale decomposition algorithm， and the sum of a series of intrinsic scale components is obtained by decomposing the original signal. Secondly， the fusion index is formed by kurtosis factor and energy spectrum coefficient to screen the effective component. Then， the entropy features of the effective components in the time domain， frequency domain and time-frequency domain are extracted respectively to perform feature fusion and dimensionality reduction. Finally， the integrated feature vectors are fed into ant colony algorithm optimized deep belief network for training and recognition to improve the algorithm efficiency and recognition rate. Experimental verification using measured data shows that the signal-to-noise ratio is increased by 8 dB on average， the average signal recognition rate can reach 95.83%， and the average recognition time is 0.715 s.