As machine learning for emotion recognition always needs a large number of samples, the method of mining the inherent emotional feature of life with a small number of samples is explored in this study. Neonatal demand for the outside world comes from the instinct without interferences such as intentions, and cry is the main medium of communication between neonates and the outside world. Thus, Neonatal cry is selected as the object of this study. The inherent emotional features of neonatal cry are excavated based on the nonlinear method. The minimum embedding dimension of neonatal cry is taken as the feature representing nervous system activity and emotion. It is found that the minimum embedding dimension of neonatal cry in the state of pain is higher and that in the state of sadness is lower. This result is consistent with related research of brain nerve activity under different emotions. The minimum embedding dimensions of neonatal cry at multiple scales are analyzed. It is also found that the minimum embedding dimension of neonatal cry in the state of pain has a certain change rule in different frequency bands. And this result is also consistent with crying characteristics in the state of pain. The extracted emotion-related parameters, which reflect the inherent physiological feature of the human body, can be used to identify and classify emotions by sounds.