In　this　paper,　a　novel　metric　learning　framework　with　dual　one-class　units　(MLF-DOU)　is　proposed　to　address　the　challenges　of　low　accuracy　and　poor　generalization　ability　associated　with　existing　models　in　network　intrusion　detection.　Specifically,　the　merits　of　one-class　units　are　leveraged,　enabling　compact　feature　representations　of　both　normal　and　attack　traffic　to　be　sufficiently　extracted.　This　extraction　is　beneficial　for　mitigating　the　overfitting　phenomenon.　On　this　basis,　a　metric　learning　method　is　introduced　to　further　enhance　the　recognition　ability　of　the　model　for　traffic　in　different　categories.　The　inter-class　distance　is　increased,　and　the　fine-grained　representations　of　intra-class　similarity　are　strengthened.　By　these　means,　both　the　detection　performance　and　the　generalization　ability　of　the　proposed　MLF-DOU　are　significantly　improved.　Extensive　experimental　results　are　presented　to　demonstrate　the　effectiveness　of　MLF-DOU　across　three　datasets,　showing　its　superiority　over　other　state-of-the-art　methods　in　achieving　accurate　intrusion　detection.　The　effectiveness　of　key　components　within　MLF-DOU　is　validated,　contributing　to　robust　feature　learning　for　each　class.　Moreover,　the　adaptability　of　the　proposed　framework　is　proven,　as　it　can　be　integrated　with　various　network　architectures,　demonstrating　promising　potential　for　real-world　deployments.