In previous research, most related works on inferring the structures of diffusion networks are designed for recovering the process of information propagation. The learning data adopted by these works is distinct in terms of both format and features from the available surveillance data of epidemics. Therefore, the existing methods are not competent when dealing with epidemic surveillance data with some intractable properties such as coarse granularity, spatial and temporal multi-scale, and incompleteness. To address this issue, an AOC (autonomy oriented computing) based method is proposed to model epidemic networks, as well as to infer their structures from epidemic surveillance data. In this method, the structure of an epidemic network and the process of disease spread are modeled by an autonomous multi-agent system named D-AOC, and the parameters of the system are automatically estimated by a self-discovery process. During this process, the parameters are adjusted and thereafter, the behaviors of agents are updated by a feedback mechanism which combines the Monte Carlo simulation and swarm intelligence. The objective is to reduce the difference between emergent behavior of the D-AOC and observed surveillance data. Regulated by the feedback mechanism, it is expected that the D-AOC will keep evolving toward the real system to be simulated. In this way, the structure of epidemic network and main biological features related to the epidemic will finally be recovered. The effectiveness and applicability of the proposed method have been validated and discussed by analyzing the real surveillance data of the H1N1 swine-flu in Hong Kong during 2009. Moreover, one scenario of applying epidemic network inference is also demonstrated by a case study of epidemic risk assessment in Hong Kong.