[Objective] This paper aims to improve the speed control performance of permanent magnet synchronous motors and ensure that the speed error converges to the preset boundary within a specified time, while avoiding issues such as controller singularities caused by disturbances and estimation peaks in the observer during motor starting. In this paper, a preset performance sliding mode control method based on a time-varying disturbance observer (PPSMC+TDO) is proposed by combining preset performance control with sliding mode control. [Methods] First, a preset performance function capable of setting convergence time was proposed, and an error transformation was performed to convert the inequality-constrained error system into an equality-constrained system. Second, a time-varying disturbance observer was designed to estimate system disturbances. Then, a preset performance sliding mode controller was developed for the transformed error system. [Results] Under the constraint of the preset performance function, PPSMC+TDO ensured that the speed error converged to the minmum preset boundary εT within the preset 0.3 s. When external disturbances were applied, PPSMC+TDO demonstrated the smallest speed fluctuations. Although the instantaneous speed error exceeded the preset boundary, the controller remained stable, without singularities or other unstable behaviors. Furthermore, at the initial moment, the disturbance estimation curves showed no estimation peaks. [Conclusion] Experimental results show that under the influence of the preset performance function, the speed error converges to the preset boundary within the designated time, ensuring good dynamic performance of the speed controller. When disturbances occur, the method addresses the issue of controller singularities caused by system errors exceeding the preset performance function boundary and provides excellent anti-disturbance capability against both constant external load disturbances and periodic disturbances. Additionally, the proposed time-varying disturbance observer effectively eliminates estimation peaks at the initial moment.