To address the limitations of traditional network gateways in performance and reliability, this paper proposes a multi-gateway load balancing and active-standby switchover strategy based on software-defined networking (SDN). A link quality scoring model is first developed, which normalizes throughput, delay, and packet loss rate into three separate 0–100 scores. To address the drawbacks of fixed weighting strategies, we adopt Q-learning to obtain optimal weight coefficients for the three metrics across nine discrete link states categorized by delay and packet loss. Based on the learned adaptive scoring model, four operation modes were implemented, i.e., normal, load balancing (LB), active standby (AS), and hybrid. Among them, the LB and hybrid modes adopted the Q-learning-based adaptive weighting mechanism, while the normal and AS modes either used fixed weights or no link quality scoring. Experiments are conducted on the Mininet platform under three fixed degradation levels (low, medium, high). Ablation experiments with fixed weight combinations [0.3, 0.4, 0.3] were further carried out to verify the necessity of the adaptive weighting module, and a quasi-dynamic gradual link degradation scenario was supplemented to test the dynamic adaptation capability. TCP single-flow and multi-flow throughput, UDP multi-flow packet loss rate, and self-healing time were measured. Results show that, compared with the normal mode, the LB and hybrid modes increased TCP throughput by 15–42% and reduced UDP packet loss by 7–13%. The hybrid mode achieved a self-healing time of 1.66 s, outperforming the active-standby mode (2.38 s), demonstrating its superior trade-off between performance and availability.