This paper develops a convolutional neural networks (CNN) method for inverse heat transfer problems (IHTP) to estimate sources in the whole-body level of insect species for the first time. For the purpose of formulating a transferable and interpretable method, we approximate the inversion operator of the heat transfer equation for different types of boundary conditions by multi-layer convolution and deconvolution operators. The hierarchical training of end-to-end CNN consists of two levels of training. In the first-level training, Gaussian pulse heat flow is used as the preset biological internal heat source to construct a forward first-level training set and obtain a first-level neural network estimation system, which is used to obtain preliminary knowledge. To further improve the results of the first-level network's estimation and with the help of prior physical and biological knowledge, we tune the training set to carry out the secondlevel training and obtain the final results of inverse identification. In order to verify the effectiveness of the deep-learning solver, we take selected biological species as the research objects, retrieve the internal heat source changes of related species in vivo, and explore the biological temperature regulation mechanism. Compared with traditional inversion strategies, the solver proposed in this paper has a solution efficiency of sub-seconds, which is suitable for solving nonlinear transient IHTP in complex biological systems. Furthermore, the proposed deep-learning solver is expected to be applied as an effective real-time non-invasive measurement technique in biomedical fields.