Spectrum denoising is an important part of spectrum detection. As we know, spectral signal is susceptible to thermal noise, mechanical vibration on site and random noise, etc. However, online monitoring systems require to reduce the impact of parameter selection caused by human operation on denoising, so a method based on singular value decomposition is proposed to denoise spectrum signal. An improved effective singular value selection method is also proposed. First, the author specify the maximum peak of the difference spectrum of singular value for the lower bound which named θ1, using the integrated information of singular value and its difference spectrum to select the upper bound, which is called θ2. The interval θ1～θ2 is defined as a fuzzy area. Then, the membership is obtained with Fuzzy C-means clusting and corresponding weight coefficients to the singular values in the fuzzy area are given. Finally, the proposed method is used to denoise UV spectrum signal with different signal to noise ratio. The signal to noise ratio, root mean square error, normalied correlation coefficient and smoothness radio are used to evaluate the result of denoising. The result shows that: based on data-driven, the proposed method has a good denoising effect, which can effectively restore the original signal.