Online vibration suppression is critical for enhancing the performance of active magnetic bearing (AMB) systems, especially in vibration-sensitive applications. This paper proposes an adaptive gradient linear neuron (AdaGrad-LN)-based online vibration extraction strategy for AMB motors. The approach integrates a second-order generalized integrator frequency-locked loop (SOGI-FLL) to extract rotational frequency from AMB winding currents, eliminating the need for angular sensors. AdaGrad-LN then processes time-domain signals from vibration sensors, extracting vibration amplitudes at the rotational frequency and its harmonics. Additionally, this method allows accurate extraction across multiple frequency points. Experimental results on a 5-DOF AMB motor demonstrate that the proposed method significantly improves convergence speed and accuracy, reducing convergence time by up to 60% and steady-state fluctuations by 62.5% compared to conventional Adaline algorithms. This approach offers a more efficient solution for real-time vibration control in AMB systems.