Packet-Level Metric-Driven Fault Diagnosis and Knowledge-Enhanced Report Generation for Satellite TCP Accelerators

Satellite TCP accelerators operate over high-latency and error-prone links, where packet-level faults are commonly investigated through manual capture analysis. We present a packet-level fault-diagnosis framework that deterministically extracts 41 scalar statistics from IPv4, TCP, and ICMP headers and packet timestamps. Twenty-six diagnostic rules map these statistics to candidate fault nodes whose signatures can be inferred from single-point pcap traces within the satellite TCP accelerator fault taxonomy. Knowledge-graph, BM25, and vector retrieval provide evidence for a locally executed quantized large language model, which generates structured diagnostic reports. Rule implementation was verified with 26 positive and 26 boundary-negative cases. On 113 strict single-fault traces, the rule engine achieved 86.73% Top-1 accuracy and a 95.58% candidate-set hit rate. On a 104-trace same-generation comparison, Random Forest achieved 89.42% Top-1 accuracy, compared with 87.50% for the rule engine and 71.15% for RBF-SVM; the difference between Random Forest and the rule engine was not significant (exact McNemar p = 0.8318). On 42 independently generated Linux network-stack fault traces covering seven representative fault classes, the rule engine achieved 95.24% accuracy, compared with 57.14% for Random Forest and 0% for RBF-SVM, and rejected five normal controls. In a 30-case report comparison, the complete configuration preserved all upstream candidate sets and introduced no unsupported fault identifier.