It is very important to secure the purity of the sound source to improve the degree of development of the noise problem, which is one of the important factors in vehicle development. So far, to acquire only the noise of the component, which is a problem element in vehicle driving noise, the component is removed and driven to acquire the noise, or the method of denoising the noise of other parts has been used. However, the method of removing part takes a lot of time to remove the part, and when the noise of the removed part is acquired, it has a disadvantage in that it differs from the characteristics of the noise measured in the mounting state of the vehicle. In addition, the method of denoising may cause data loss due to the deformation of the sound source of the noise. To maintain the label purity of the fuel pump and noise, the method of measuring the noise data of the fuel pump and the method of acquiring pure noise data only for vehicles excluding the noise of the fuel pump are presented. First of all, we propose a method of creating a mixing sound that can be used as training data using the acquired noise. Also, we used a transfer learning technique using an AI speaker-to-speaker separation model of vehicle noise excluding the fuel pump to separate vehicle noise and fuel pump noise, and we intend to demonstrate its performance through accurate indicators