Emotion recognition is an important research direction in the field of artificial intelligence, which can help machines understand and adapt to human emotional states. Physiological signals (PS), such as electroencephalogram (EEG), galvanic skin response (GSR), respiration rate, and body temperature, are effective biomarkers reflecting changes in human emotions. However, the collection of PS requires strict and complex operation procedures and professional personnel to label the data, resulting in a shortage of effectively labeled PS data. We propose a based on PS multi-head attention contrastive learning method(PS-MHACL), to address these issues. PS-MHACL learns meaningful feature representations from unlabeled PS and improves recognition performance through feature fusion using a multi-head attention mechanism. We use the genetics-inspired Meiosis method to augment unlabeled signals and train the feature extractor, which is then applied to labeled PS. A contrastive loss function maximizes the similarity of feature representations of augmented groups with the same stimulus, predicting emotions’ Arousal and Valence. Experiments on two public datasets, DEAP and MAHNOB-HCI, showed our method’s superior performance in terms of accuracy and cross-individual generalization ability. On DEAP, accuracies for Valence and Arousal dimensions were 96.39% and 93.44% respectively, with 89.35% for a four-class classification task. On MAHNOB-HCI, these were 94.89%, 96.11%, and 93.72% respectively.