In the realm of network security, advancements like 5G, IoT, and cloud computing have expanded network environments and real-time traffic complexity, accompanied by a rise in diverse and sophisticated cyber-attacks. This paper introduces a self-adaptive discriminative autoencoder (SADAE) method which is developed through deep metric learning aiming to effectively detect various network intrusions. SADAE integrates K local autoencoders and one global autoencoder: the former captures diverse data distributions, whilst the latter governs network traffic representation scale. Through effective self-adaptive metric learning, the proposed SADAE is able to identify and extract discriminative features to automatically detect various network traffic classes, enhancing data separability and improving detection accuracy. For validation and evaluation, the proposed approach was applied to the binary NSL-KDD datasets and multi-class CSE-CIC-IDS2018 datasets. The experimental results demonstrate SADAE’s effectiveness in detecting and categorising network intrusion anomalies in reference to other popular deep learning and metric learning approaches.