Currently, it is a great challenge to accurately diagnose global properties of dusty plasmas from limited data. Based on machine learning, a novel diagnostic method for various global properties in dusty plasma experiments is developed from single particle dynamics. It is found that for both two-dimensional (2D) dusty plasma simulations and experiments, the global properties such as the screening parameters κ and the coupling parameter Γ can be accurately determined purely from the position fluctuations of individual particles. Hundreds of independent Langevin dynamical simulations are performed with various specified κ and Γ values, resulting in a great number of individual particle position fluctuation data, which can be used for training, validating, and testing various convolutional neural network (CNN) models. To confirm the feasibility of this diagnostic method, three different CNN models are designed to determin the κ value. For the simulation data, all these CNN models perform excellently in determining the κ value, with the averaged determined κ value almost equal to the specified κ value. For the experiment data, the distribution of the determined κ values always exhibits one prominent peak, which is very consistent with the κ value obtained from the widely accepted phonon spectra fitting method. Furthermore, this diagnostic method is extended to simulatneously determining both the κ and Γ values, achieving satisfactory results by using 2D dusty plasma data from both simulations and experiments. The excellent performance of the CNN models developed here clearly indicates that through machine learning, the global properties of 2D dusty plasmas can be fully characterized purely from single particle dynamics.