In response to the issue of substantial interference cancellation errors arising from insufficient background interference correlation in actual environments, a frequency-domain multi-path electromagnetic interference cancellation approach was proposed. The frequency splitting property of the variational modal decomposition (VMD) was used to decompose the two-channel signals into multiple modal signals with the same center frequency and decomposition bandwidth. Then the decomposed modal signals were used to construct a multi-path interference cancellation system. Finally, a back propagation (BP) neural network was used to nonlinearly fit and cancel the signals on each path so as to obtain the real equipment-under-test (EUT) radiated emission signals. The simulation and experimental test results show that the EUT radiated emission signal obtained by this method has a power error of less than 1 dB and a signal correlation of more than 96% compared with the actual EUT radiated emission signal. Compared with the variable step size LMS algorithm and EMD-LMS algorithm, the power error is reduced by at least 36 dB, and the signal correlation is improved by at least 15%. This method has a reference value for in-situ testing of electromagnetic radiation emissions from large, movable EUTs in open fields.