Research On Physiological Signal Emotion Recognition For Multi-scenarios

Emotion recognition based on physiological signals has attracted the attention of many researchers due to its objective and real-time characteristics.With the increasing demand for emotional experience,this technology has broad application prospects in many fields.However,when facing different scenarios,existing methods have the problem of poor recognition performance due to individual user differences and unknown class emotion recognition.Therefore,this thesis proposes corresponding algorithms for different scenarios,and verifies their effectiveness in simulated flight scenarios in civil aviation.It also studies the impact of negative emotions on pilots’ driving and operational abilities based on flight safety requirements.The main work of the thesis is as follows:Firstly,for cross-subject scenarios,aiming at the problem that existing algorithms are affected by feature confusion caused by individual differences between subjects,which leads to reduced recognition accuracy,we propose the Cross-subject Emotion Recognition Based on Multi-source Adversarial Domain Adaptation Strategy(MADAS-ER)algorithm.MADAS-ER effectively alleviates feature confusion caused by individual differences through a three-stage adversarial training paradigm,and improves the stability of the algorithm by utilizing the difference information among source domains.The results of cross-subject emotion recognition experiments on the public dataset SEED show that compared to existing algorithms,MADAS-ER has improved accuracy by 2.71% and6.60% under two experimental paradigms,respectively.Secondly,for the new scenario of discrete emotion open-set recognition,we propose the Open-set Emotion Recognition based on One-vs-Rest Strategy(OVRS-OER)algorithm,which solves the problem of existing emotion recognition algorithms ignoring the recognition of new unknown emotion classes outside the training set in practical applications.The classification network based on the ”one-vs-rest” strategy is used to realize the recognition of unknown emotion classes.At the same time,a hard negative classifier sampling strategy is introduced to avoid artificially setting thresholds,so that the algorithm does not require prior knowledge of unknown emotion classes.The experimental results on the public dataset SEED-V show that compared to the general open-set recognition algorithm in the field of image recognition,OVRS-OER has improved H-score and F1-score by 6.15% and 5.15% on two different experimental settings for classification respectively.Finally,for the practical application scenarios of civil aviation simulation flight,the research team collaborated with the Civil Aviation Flight University of China to collect a pilot discrete emotion dataset based on electrocardiogram signals.On the basis of this self-collected dataset,the effectiveness of the MADAS-ER algorithm for cross-subject scenarios and the OVRS-OER algorithm for discrete emotion open-set recognition scenarios is verified.And combined with flight safety requirements,the correlation between negative emotions represented by doubts,resistance,and panic and the pilot’s driving ability is studied,demonstrating the important practical application value of the proposed algorithms in civil aviation flight scenarios.