To address the noise interference during the demodulation of fiber Bragg grating (FBG) by the CCD-based spectral diffraction method, which leads to the reduction of the demodulation accuracy and stability of FBG center wavelength, a fusion noise reduction algorithm based on the variational mode decomposition (VMD-WT-SG) is proposed. By combining the advantages of different noise reduction algorithms and adjusting algorithm parameters, the effective removal of noise components in the spectral signals is realized. The spectral waveforms are smoother and more continuous while retaining most of the features of the original signals, and the noise reduction effectiveness is obvious. Compared with the traditional three methods, such as the VMD noise reduction, the SG noise reduction, and the Kalman filter, the spectral SNR of the spectral signal-to-noise ratio processed by the VMD fusion noise reduction method (VMD-WT-SG) is 20.14 dB, and the root-mean-square error is 0.017. The signal-to-noise ratio is improved by 23.94%, 41.14%, and 94.97%, respectively. The root mean square error is reduced by 39.29%, 45.16%, and 67.92%, respectively. It is the best among the four noise reduction methods. To improve the demodulation rate, an intersection peak-finding algorithm is proposed for the noise-reduced multi-peak spectra based on the commonly used peak-finding algorithm. The center wavelength is obtained by calculating the coordinates of the intersection of the peaks with the intersection of the left and right sag lines of the next largest points. The peak-finding accuracy is compared with the polynomial fitting, the center-of-mass, and the Gaussian fitting algorithms through experiments. The results show that the average peak-finding deviation of the proposed algorithm is 3.3 pm, which is better than that of the center-of-mass and the polynomial fitting. The average running time of the algorithm is 0.261 ms, which is better than those of the polynomial fitting and the Gaussian fitting. In real-time applications, the demodulation rate can reach 4 kHz while ensuring accuracy, Meanwhile, the algorithm has good stability, the average standard deviation of the demodulated wavelengths is 1.8 pm, and it can meet the requirements of practical applications, which has some reference values for the fast multi-peak real-time detection in the FBG sensing networks.