Abstract:

To accurately identify physical fatigue in firefighters, lower training injury risks, boost emergency rescue efficiency, 18 firefighters were selected as research subjects. Physical fatigue was induced via routine training. The effectiveness of fatigue induction was verified with subjective questionnaire scores and the mean relative error of time perception. Four types of physiological signals, including heart rate variability(HRV), eye movement, electrodermal activity(EDA) and electromyography (EMG), were synchronously collected before and after fatigue. Fourteen significantly different indicators were screened via power analysis, t-test, rank-sum test and Benjamini-Hochberg (BH) correction. Core features were optimized and selected using the minimum Redundancy Maximum Relevance (mRMR) algorithm. Single-modal, dual-modal, trimodal and four-modal fatigue recognition models were built under the multimodal fusion strategy. Recognition verification was implemented by five algorithms: Logistic Regression (LR), K-Nearest Neighbor (KNN), Support Vector Machine (SVM), Adaptive Boosting (AB) and Random Forest (RF). Result shows that heart rate variability (HRV) serves as the optimal single-modal recognition indicator. Multimodal fusion delivers obviously better recognition performance than the single-modal approach. The trimodal combination of EMG + HRV + eye movement acts as the optimal fusion scheme. The Logistic Regression model achieves a recognition accuracy of 91.4% and an average accuracy of 88.7%. It can judge firefighters' physical fatigue states efficiently and accurately.

Key words: modalities, physiological indicators, firefighters, physical fatigue, machine learning