Abstract: Under starting and braking conditions or non-uniform load conditions of large belt conveyors, the double-rotor permanent magnet drive system is susceptible to large vibrations caused by periodic excitations and perturbations. In particular, the non-sinusoidal distribution of the magnetic field and the presence of eddy current harmonics make the vibration control of the system particularly important. The double-rotor permanent magnet drive system has the advantages of high efficiency in speed regulation, low operating cost, strong adaptability and so on, but its nonlinear characteristics are obvious. The neural network iterative learning control algorithm makes the network able to fit the nonlinear features of the system by introducing nonlinear activation function and depth structure. During the iterative learning process, the network continuously adjusts the parameters according to the real-time vibration data of the system to achieve the goal of vibration suppression. Therefore, using this algorithm to study the vibration suppression of the rotor system is of great significance to improve the overall stability of the system and ensure its safe and efficient operation. An modified neural network iterative learning control algorithm (MNN−ILC) is proposed for the transmission unbalance and misalignment coupled vibration problems induced by the non-sinusoidal distribution of the magnetic field and the eddy current harmonics in the double-rotor permanent magnet drive system. In order to effectively suppress vibration, an adaptive factor σ based on the error value and a regularized weight attenuation factor λ are innovatively introduced to quickly respond to system changes and reduce the error. Multi-physics field finite element simulation is used to model the vibration characteristics of the system under unaligned unbalanced condition. In order to verify the effectiveness of the proposed algorithm, a 55 kW double-rotor permanent magnet drive train is used as a research object to test its real-time damping capability at different speeds. The results show that the maximum amplitude of the measurement point without applying the control is about 18.7 μm. After applying the MNN−ILC algorithm, the maximum value of the amplitude is reduced to 3.1 μm, which is a reduction of about 83.4%. By comparing the MNN−ILC algorithm with the traditional ILC algorithm and the existing improved ILC algorithm. It is found that the traditional ILC algorithm starts to diverge at about 10 s after applying the control, while the MNN−ILC algorithm still maintains a good and stable control performance. The amplitude is controlled within ±3.5 μm. MNN−ILC algorithm improved vibration suppression by approximately 51.9%, 58.1%, and 61.4% compared to ILC algorithm. The reduction in response time is about 33.4%, 32.0%, and 32.5%, respectively. This study provides an important theoretical reference for the vibration suppression of double-rotor permanent magnet drive systems.