In the traditional Raman spectrum detection process, the effective signal of Raman spectrum is sometimes submerged by the fluorescent background and is difficult to be recognized. In order to separate differential signals and baseline deviations accurately and effectively, combining differential Raman technique with error back propagation algorithm neural network, a differential Raman demodulation and denoising algorithm was proposed. The theoretical and experimental validation was carried out. The fluorescent substances such as paraquat, kidney-tonifying pills, engine oil and heroin were detected and analyzed. The results show that Raman characteristic spectra can be obtained effectively. The study solves the difficult problem of detection in industry application.