Publication: Predicting human stress emotions using machine learning models
Abstract
Bu tez stresin insan üzerindeki etkisini araştırmakta ve otomobil sürücülerinde ortaya çıkan stres seviyelerini tahmin etmeye çalışmaktadır. DRIVE veri seti, Elektrokardiyogram, Elektromiyogram, Galvanik Cilt Tepkisi ve Solunum gibi fizyolojik sinyalleri içeren halka açık bir veri setidir. Sinyal verileri 10 saniyelik bölümlere ayrıldı ve yeni girdileri çıkarmak için Temel Bileşenler Analizi, Doğrusal Ayrımcılık Analizi, Rastgele Ormanlar ve Otokodlayıcılar dahil olmak üzere çeşitli boyut azaltma tekniklerinden yararlanılarak yeni özellikler tasarlamak için kullanıldı. Sonuç olarak, beş farklı veri seti Evrişimli Sinir Ağları (CNN), Tekrarlayan Yapay Sinir Ağları (RNN) ve Uzun Kısa Süreli Bellek ile test edilmiştir. Sonuçlar, LSTM modellerinin, önerilen mimarinin %75 doğruluk elde etmesiyle diğer derin öğrenme algoritmalarından daha iyi performans sergilediğini gösteriyor.
This thesis investigates the effect of stress in human beings and attempts to predict its occurrence in automobile drivers. The DRIVE dataset contains physiological signals such as Electrocardiogram, Electromyogram, Galvanic Skin Response, and Respiration, and is available to the public for further experimentation. The signal data was segmented into intervals of 10 seconds and utilized to engineer new features from which various dimensionality reduction techniques, including Principal Component Analysis, Linear Discriminant Analysis, Random Forests, and Autoencoders, were used to extract new inputs. Consequently, five different sets of data were tested with Convolutional Neural Networks, Recurrent Neural Networks, and Long Short-Term Memory. The results show that the LSTM models outperform the rest of the deep learning algorithms with the proposed architecture achieving an accuracy of 75%.
This thesis investigates the effect of stress in human beings and attempts to predict its occurrence in automobile drivers. The DRIVE dataset contains physiological signals such as Electrocardiogram, Electromyogram, Galvanic Skin Response, and Respiration, and is available to the public for further experimentation. The signal data was segmented into intervals of 10 seconds and utilized to engineer new features from which various dimensionality reduction techniques, including Principal Component Analysis, Linear Discriminant Analysis, Random Forests, and Autoencoders, were used to extract new inputs. Consequently, five different sets of data were tested with Convolutional Neural Networks, Recurrent Neural Networks, and Long Short-Term Memory. The results show that the LSTM models outperform the rest of the deep learning algorithms with the proposed architecture achieving an accuracy of 75%.