Modern Enterprise Intrusion Detection Systems (IDS) encounter significant operational hurdles due to the high-dimensional, non-linear, and varied nature of network traffic produced by advanced cyber threats. Conventional kernel classifiers, such as standard Support Vector Machines (SVMs), depend on a single static kernel, which constrains feature transformation to a single homogeneous Reproducing Kernel Hilbert Space (RKHS). As a result, these single-kernel architectures exhibit elevated False Alarm Rates (FAR) and lower detection performance when confronted with complex feature overlap or previously unseen zero-day attacks. To overcome these limitations, this study introduces an Adaptive Multi-Kernel Fusion (AMKF) framework that dynamically refines non-linear mapping for network anomaly detection. The proposed architecture constructs an adaptable composite kernel space combining Gaussian RBF, Polynomial, and Laplacian kernels, optimizing their respective weight coefficients via semi-definite convex optimization based on local manifold properties. Experimental validation on the UNSW-NB15 and CICIDS2017 benchmark datasets demonstrates that the AMKF framework reaches a 99.42% overall accuracy, a 99.15% Detection Rate (DR), and minimizes the False Alarm Rate to 0.48%, outperforming traditional single-kernel models and deep neural networks in both execution efficiency and model interpretability.