Gas source localization is one of the most common applications for gas-sensitive mobile robots. When gas propagates through a medium, it forms plume-like substances. A large body of research indicates that mobile robots can localize gas sources based on these gas plumes. Currently, most research on plume localization is limited to two-dimensional planes, meaning that mobile robots can only detect plumes in a single plane. Although significant progress has been made, this overlooks the objective laws of gas distribution in real three-dimensional environments. Drones, as mobile autonomous platforms with certain payload capabilities and high flexibility, can carry various sensors to perform detection tasks in three-dimensional environments. This paper proposes an olfactory drone, utilizing the Improved PPO algorithm for gas source localization in three-dimensional environments. We constructed a training scenario using a Gaussian puff model to simulate the diffusion of gas in a three-dimensional environment. After training, the trained neural network is deployed on the simulation platform. Using Fluent software, we exported the actual gas diffusion situation to the simulation platform to validate the algorithm. The results indicate that the olfactory drone can locate gas sources in unknown environments.