Self-supervised learning (SSL) has emerged as a promising approach to address the challenges of labeled data scarcity in electroencephalogram (EEG) analysis. Traditional supervised methods rely on labeled EEG data, which is difficult and expensive to obtain due to the need for expert annotation and the variability of EEG signals across subjects. SSL, on the other hand, leverages unlabeled data through pretext tasks to learn effective representations without requiring explicit labels. This paper provides a systematic survey of SSL methods applied to EEG signals, discussing their theoretical foundations, methodologies, and applications in various downstream tasks. The paper first introduces the concept and theory of SSL, including typical frameworks such as prediction-based, generation-based, contrastive-based, and hybrid methods. It then provides a comprehensive review of existing SSL approaches for EEG analysis, categorizing them into four major classes based on the type of pretext tasks used. The review discusses the differences between these methods, their adaptation to various downstream tasks, and the potential directions for future research. The paper also explores the application of SSL in EEG analysis, focusing on predictive-based methods that use spatial, temporal, and transformation-based pretext tasks to extract features from EEG signals. These methods include spatial predictive tasks such as EEG jigsaw and channel correlation prediction, temporal predictive tasks such as relative positioning and temporal shuffling, and transformation predictive tasks such as stopped band prediction and multi-transformation recognition. Each method is analyzed in terms of its approach, performance, and effectiveness in capturing relevant features for downstream tasks. In addition, the paper discusses generative-based SSL methods, which use reconstruction and generation tasks to learn effective representations. These methods include temporal reconstruction tasks, multi-domain reconstruction tasks, and generative adversarial tasks. The paper highlights the importance of these methods in capturing contextual and temporal features from EEG signals, which are essential for various EEG-based tasks. Overall, the paper provides a comprehensive overview of SSL methods for EEG analysis, emphasizing their potential to overcome the challenges of labeled data scarcity and improve the performance of EEG-based tasks. The review highlights the strengths and limitations of different SSL approaches and suggests future research directions for advancing SSL in EEG analysis.Self-supervised learning (SSL) has emerged as a promising approach to address the challenges of labeled data scarcity in electroencephalogram (EEG) analysis. Traditional supervised methods rely on labeled EEG data, which is difficult and expensive to obtain due to the need for expert annotation and the variability of EEG signals across subjects. SSL, on the other hand, leverages unlabeled data through pretext tasks to learn effective representations without requiring explicit labels. This paper provides a systematic survey of SSL methods applied to EEG signals, discussing their theoretical foundations, methodologies, and applications in various downstream tasks. The paper first introduces the concept and theory of SSL, including typical frameworks such as prediction-based, generation-based, contrastive-based, and hybrid methods. It then provides a comprehensive review of existing SSL approaches for EEG analysis, categorizing them into four major classes based on the type of pretext tasks used. The review discusses the differences between these methods, their adaptation to various downstream tasks, and the potential directions for future research. The paper also explores the application of SSL in EEG analysis, focusing on predictive-based methods that use spatial, temporal, and transformation-based pretext tasks to extract features from EEG signals. These methods include spatial predictive tasks such as EEG jigsaw and channel correlation prediction, temporal predictive tasks such as relative positioning and temporal shuffling, and transformation predictive tasks such as stopped band prediction and multi-transformation recognition. Each method is analyzed in terms of its approach, performance, and effectiveness in capturing relevant features for downstream tasks. In addition, the paper discusses generative-based SSL methods, which use reconstruction and generation tasks to learn effective representations. These methods include temporal reconstruction tasks, multi-domain reconstruction tasks, and generative adversarial tasks. The paper highlights the importance of these methods in capturing contextual and temporal features from EEG signals, which are essential for various EEG-based tasks. Overall, the paper provides a comprehensive overview of SSL methods for EEG analysis, emphasizing their potential to overcome the challenges of labeled data scarcity and improve the performance of EEG-based tasks. The review highlights the strengths and limitations of different SSL approaches and suggests future research directions for advancing SSL in EEG analysis.