• In this paper, a peak detection algorithm for FBG sensor system is proposed. The background noise in the system is reduced by wavelet packet decomposition threshold method. The peak interval can be adaptively divided by Hilbert transform, and the parabola fitting method is used for peak detection. • Through simulation and experiments, the algorithm has good denoising effect, high stability and accuracy, the stability of the algorithm is better than 0.5 pm and the temperature detection accuracy is better than 0.352 ℃. The calculation speed is 81.2% higher than that of Gaussian LM method. It is suitable for FBG multi peak real-time detection system. An accurate multi-peak detection algorithm based on wavelet packet decomposition (WPD) denoising and Hilbert transform (HT) is proposed. WPD and wavelet threshold method are used to denoise the high-frequency part of the spectrum. HT and parabola fitting are used to divide the peak area adaptively and calculate the central wavelength. Simulation results show that compared with the other four denoising methods, this denoising algorithm can solve the problem of background noise more effectively. Compared with the Gaussian LM (Levenberg-Marquardt) algorithm, Centroid method, and polynomial fitting method, this method has higher precision. Experimental results of real-time temperature detection show that the stability is better than 0.5 pm at 0℃ and the temperature monitoring accuracy is better than 0.352℃ in the range of −20℃∼ 40℃, which is the best among the four algorithms. The computational speed is 81.2% higher than the Gauss LM algorithm. In general, this method can be applied to FBG real-time temperature monitoring and demodulation system under the condition of background noise.