Abstract: To improve accuracy of emotional recognition for dimensional model on the physiological signals, based on the DEAP dimension emotional physiological data set, this paper uses the AR model power spectrum estimation method to extract the power spectrum density of the EEG θ, α, β and γ rhythms, uses wavelet packet decomposition to extract the EEG wavelet packet coefficients and energy ratio time-frequency characteristics, and uses nonlinear analysis method to extract EEG sample entropy and wavelet packet entropy. A stacked auto-encoder neural network algorithm is designed to perform machine emotion recognition on the EEG combination features in two emotional dimensions: valence and arousal. The effects on emotion recognition results, including EEG features, sample balances and emotion tags are analyzed. The simulation tests show that the stacked auto-encoder neural network is effective for the dimensional emotion recognition of EEG. In the emotional valence dimension, the average recognition accuracy rate can reach 80.3%; and the average recognition accuracy rate reaches 81.5% in the arousal dimension. The research can provide a practical reference for the automatic analysis and machine recognition on continuous dimension emotions.