[Objective] As an emerging intelligent control technology, deep reinforcement learning (DRL) has demonstrated remarkable potential in the field of motor drive system control. In this regard, this paper researches and designs an advanced DRL-based drive control architecture for permanent magnet synchronous motor (PMSM), aiming to achieve high-precision, model-free robust control without relying on the accurate identification of motor physical parameters. [Methods] This paper combined the deep Q-network with finite control set torque control, directly outputting the switching states of the inverter through online learning, thus enabling the agent to determine the optimal switching states of the inverter directly via continuous online learning and interaction with the motor environment. Firstly, a comprehensive multi-level reward function was designed to reflect the complex characteristics of the PMSM, simultaneously accommodating multiple optimization objectives including high-fidelity torque tracking, stator current amplitude minimization, and overall energy efficiency maximization. Secondly, to bridge the gap between theoretical exploration and practical safety requirements, a novel safety protection and evaluation mechanism based on current constraints was established. This mechanism ensured that the inherent random exploration process of DRL did not lead to system overcurrent or hardware damage. Finally, the convergence and control performance of the algorithm were effectively improved by introducing the Q-learning structure and an automated hyperparameter optimization method. [Results] The simulation results showed that the average reward value stabilized at approximately 1 after 400 training episodes, which verified the excellent convergence of the proposed algorithm. The algorithm accurately tracked the torque commands and maintained fast response speeds with minimal steady-state errors under various speed and load step conditions. With the valid weight coefficients, the system successfully balanced torque precision and operational efficiency. Furthermore, the safety protection mechanism effectively truncated the expected future returns of high-risk states via the done signal, ensuring that the stator current was strictly confined within the safety threshold, which validated the robustness of the model even in small-sample scenarios. [Conclusion] The proposed scheme achieves high-performance model-free torque control, and its integrated safety assessment mechanism provides a scientific foundation and novel insights for preventive operation and maintenance alongside the application of reinforcement learning in the power electronics field, as well as a new research direction for intelligent motor control.