Objective To explore the predictive effect of heart rate variability （HRV） on noise annoyance and develop a model for identifying and assessing noise annoyance. Methods A group of employed subway drivers participated in a simulated train driving experiment under different noise conditions. The Weinstein Noise Sensitivity Scale， subjective noise annoyance questionnaire， and electrocardiogram data were collected. HRV features were extracted and transformed into a standard normal distribution using Z-Score normalization. Random Forest （RF） was used for feature selection and important features were inputted to establish various driver noise annoyance identification models based on HRV features. The impact of individual noise sensitivity on accuracy was also discussed. Results Multiple HRV features were found to be related to noise annoyance. Feature selection revealed that individual noise sensitivity significantly influenced the identification and detection of noise annoyance. Among various classification models， the Convolutional Neural Network （CNN） model achieved the best performance in identifying annoyance levels， with an accuracy of 90.03%. Conclusion The deep learning model based on HRV demonstrated excellent performance in identifying noise annoyance， providing a method and theoretical support for real-time recognition of occupational noise annoyance.