Research On Physiological Signal Emotion Recognition Method Based On Deep Learning

Emotion recognition based on physiological signals is a technique for recognizing human emotional states from physiological signals using computer analysis algorithms,which has the advantages of objectivity and real-time processing and has promising applications in a variety of fields.However,the following issues exist in the field of physiological signal-based emotion recognition that prevent it from being used in real life:(1)Subject variability,that is,differences in physiological signal distribution among different individuals affect cross-subject emotion recognition applications,resulting in a decrease in model accuracy;(2)Signal non-stationarity,which occurs when physiological signals from the same subject have different distributions affected by physical changes in different sessions,makes it difficult for the model to identify emotions stably across time;(3)The different levels of pilot panic in the flight scenario are more similar in terms of physiological signals,leading the model to fail to distinguish accurately.To address these issues,the thesis' main work is as follows:1.To address the issue of subject variability,existing studies treat all features as having equal transferability when performing knowledge transfer,leading to the negative tranfer.Base on this,we propose the Cross-subject EEG Emotion Recognition Based on Joint Domain Alignment Strategy(JDAS-ER)method.JDAS-ER combines the attention mechanism to focus on key feature regions to mitigate the impact of negative transfer in local domain alignment.The overall transferring is then performed in the global domain alignment,while the adaptive factor is proposed to dynamically control the weights of local domain alignment and global domain alignment.JDAS-ER achieves a 3.29% accuracy improvement on the SEED dateset compared to existing methods.2.Aiming at the issue of signal non-stationarity,considering that only the overall alignment of the two domains and ignoring the decision boundary leads to the introduction of ambiguous features,we propose the method called Maximizing Domain Discrepancy for Cross-session EEG Emotion Recognition(MDD-ER).MDD-ER constrains the domain discrepancy based on feature characteristics.For shallower features,the mean difference between the source domain and the target domain is maximized to constrain their marginal distribution differences.For deeper features,the prediction discrepancy between the two classifiers is introduced to conduct adversarial learning,constraining its conditional distribution differences.MDD-ER conducts two cross-session experimental paradigms on the SEED dataset with an accuracy of 91.90% and 84.81%,respectively,which are superior to existing studies.3.The above algorithm can accurately classify the emotional categories with a high degree of discrimination,but it is difficult to effectively distinguish the different degrees of panic of the pilots in the actual flight scene.Aiming at this issue,we conducts a research on emotion recognition in flight scenes.We collect a simulated flight emotional ECG dataset with research value using flight training programs,special circumstances,and crew atmosphere to induce emotions.For panic emotions that endanger flight safety,we combine center loss and cross-entropy loss to overcome the problem of confusing features in the recognition of different degrees of panic.The proposed method yielded a more objective and detailed evaluation of pilot emotion recognition,which helps to improve flight safety to some extent.