The invention relates to a ground nuclear magnetic resonance signal noise filtering method based on a model, which is mainly used for filtering the peak noise in a nuclear magnetic resonance signal. The method comprises the steps of firstly, collecting an MRS signal by using a nuclear magnetic resonance water detector, and removing the power frequency harmonics by using a harmonic modeling method;secondly, establishing a peak noise model through two second-order cascade band-pass filters, and meanwhile, positioning the position of the peak noise by adopting an NEO algorithm and intercepting corresponding data segments; and then, performing parameter extraction based on a least square method to obtain peak noise, aand finally, subtracting the MRS data from which the power frequency harmonic waves are removed from the peak noise obtained by modeling to obtain an MRS signal from which the peak noise is removed. According to the method, the problem of removing the peak noise in the MRS signal is solved, and compared with a traditional MRS peak signal denoising method, the method has the advantages of being high in practicability, obviously improving the working efficiency, improving the signal-to-noise ratio and the like.