Network failures in multi-service scenarios in a military broadband-narrowband converged packet switching system can lead to service interruptions, degraded service quality, and low fault-tolerance efficiency. This paper designs a three-stage fault-tolerant architecture for a military broadband-narrowband converged packet switching system, covering the data plane, control plane, and management plane. The data plane implements N + 1 backup and microsecond-level bandwidth adjustment through hardware redundancy modules and dynamic bandwidth allocation units (DBAUs). The control plane uses the extended OpenFlow protocol and multi-service scheduling algorithm to implement intelligent path planning based on service priorities. The management plane integrates an LSTM prediction engine to predict network risks based on historical fault data and generate pre-configured policies. This mechanism achieves a maximum mean-time-to-return (MTTR) of only 82 ms, a service interruption rate of no more than 0.48%, service latency below 45 ms, jitter below 13 ms, and a reduction in redundant resource usage by approximately 32.5%-45.0%, effectively improving the system's service continuity and resource utilization efficiency under fault conditions.