Aiming at the problems of frequent bursty traffic, delayed congestion feedback, and severe queue oscillation in data center networks (DCNs), this study proposes an explicit congestion notification (ECN) network optimization algorithm based on deep Q-learning (DQECO). Firstly, based on the proportional integral controller enhanced (PIE) queue management mechanism, the algorithm introduces the deep Q-learning network (DQN) to realize adaptive and dynamic adjustment of the ECN marking probability. Secondly, by fusing multi-dimensional features including the queue delay, queue occupancy rate, and ECN marking ratio, the algorithm accurately perceives the network congestion state and realizes refined dynamic control of network congestion. Finally, by designing a nonlinear multi-objective reward function, the algorithm guides the agent to maximize the long-term cumulative reward in continuous interaction with the network environment, and realizes continuous iterative optimization of algorithm performance. To verify algorithm performance, this study constructs a simulation environment based on ns3-ai and carries out comparative tests with NewReno and QTCP algorithms in multiple scenarios. Experimental results show that the comprehensive performance of the DQECO algorithm outperforms the comparison algorithms. In high-bandwidth scenarios, its throughput is improved by about 160% compared with NewReno and about 13.4% compared with QTCP. Meanwhile, its average round-trip time (RTT), queue occupancy rate, and packet loss rate are controlled at about 0.08s, below 0.2 and below 0.4, respectively. The algorithm can ensure stable throughput, suppress queue oscillation, and has better stability and adaptability.