Future wireless networks will need to improve adaptive resource allocation and decision-making to handle the increasing number of intelligent devices. Unmanned aerial vehicles (UAVs) are being explored for their potential in real-time decision-making. Moreover, cognitive non-orthogonal multiple access (Cognitive-NOMA) is envisioned as a remedy to address spectrum scarcity and enable massive connectivity. This paper investigates the design of joint subchannel and power allocation in an uplink UAV-based cognitive NOMA network. We aim to maximize the cumulative sum rate by jointly optimizing the subchannel and power allocation based on the UAV's mobility at each time step. This is often formulated as an optimization problem with random variables. However, conventional optimization algorithms normally introduce significant complexity, and machine learning methods often rely on large but partially representative datasets to build solution models, assuming stationary testing data. Consequently, inference strategies for non stationary events are often overlooked. In this study, we introduce a novel active inference-based learning approach, rooted in cognitive neuroscience, to solve this complex problem. The framework involves creating a training dataset using random or iterative methods to find suboptimal resource allocations. This dataset trains a mobile UAV offline, enabling it to learn a generative model of discrete subchannels and continuous power allocation. The UAV then uses this model for online inference. The method incrementally derives new generative models from training data by identifying dynamic equilibrium conditions between required actions and variables, represented within a unique dynamic Bayesian network. The proposed approach is validated through numerical simulations, showing efficient performance compared to suboptimal baseline schemes.