Positive thinking meditation, as an effective method that can reduce stress and enhance the sense of well-being in life, is becoming a hot research topic in related fields in various countries. In this paper, bio-signal data, such as EEG, of subjects were collected through meditation experiments, and combined with fNIRS signals to form a bio-signal dataset for research. The collected data is processed using independent component analysis to reduce the interference of noise in the data for the study. The stages of the subjects’ positive thinking meditation were classified and identified by deep learning models and long and short-term memory neural networks. To address the problem of misidentification caused by uneven data distribution, this paper uses data enhancement methods to expand the diversity of data and improve the classification accuracy of the model. To address the shortcomings of deep learning models and data enhancement methods in data labelling, this paper integrates the transfer learning method with the MobileNetV2 model, the RestNet50 network model and the Xception model through weighted averaging to form a network integration model for classifying the phases of mindfulness meditation. The model achieves up to 99.4% accuracy on the subject’s biosignal dataset, an improvement of 8.4% compared to the data-enhanced deep learning model. The deep learning machine optimization model can effectively classify the stages of positive thinking meditation, which can be a highly valuable reference for the rational arrangement of positive thinking meditation training strategies.