Pre-trained deep learning models for EEG-based cognitive state recognition for construction workers through transfer learning

In pursuit of proactive safety management in construction, accurate and real-time recognition of workers' cognitive states is essential for crafting targeted interventions to minimize safety risks. Unlike conventional manual and qualitative methods, electroencephalogram (EEG) offers a reliable and objective solution for cognitive state recognition. Notably, efforts that integrate EEG with deep learning have exhibited significant advancements. Nevertheless, certain constraints, such as experimental costs and limited participants, have caused a scarcity of high-quality data, which has impeded performance in recognition tasks. Although transfer learning demonstrates its capability in addressing this challenge, the lack of relevant explorations into EEG-based cognitive state recognition remains a significant research gap. Therefore, customizing pre-trained deep learning models for these tasks would be beneficial. This study aims to develop pre-trained models to advance future studies in this domain through transfer learning, encompassing: 1) extracting extensive and accurately labeled EEG data from the DEAP dataset, 2) selecting appropriate network architectures and implementing pre-trained model development, and 3) collecting EEG data for mental fatigue recognition and evaluating model effectiveness. Rigorous evaluation suggests a substantial improvement in model performance, with test accuracy increasing from 63.19% to 92.29% by leveraging the pre-trained convolutional neural networks (CNN)—long short-term memory (LSTM) model. In conclusion, this study significantly contributes to enhancing safety management for construction workers through EEG by providing validated pre-trained models to address challenges of data scarcity. Future research may advance by exploring additional network architectures, increasing sample size, and considering more specific recognition tasks for model evaluation.