Dynamically Optimized SVDD Based Mental State Recognition Method

Intelligent mental state detection methods require training based on a large amount of foundational data. However, in real-world applications, it is often challenging to obtain sufficient data, which limits the applicability of these methods. Therefore, this paper proposes a mental state prediction and recognition method with dynamically updated classification boundaries, capable of achieving precise and timely diagnosis based on a small number of real-time target samples. The paper introduces a dual-boundary SVDD method based on a relaxation threshold. By applying relaxation variables, this method can accurately filter out data points located between the core normal region and the potential abnormal region, thereby determining the range of effective support vectors. The paper also proposes a model dynamic updating method based on spatial/temporal weighting, which improves the training efficiency and shortens the modeling cycle. Additionally, a classification hypersphere's center trajectory offset index is proposed, which uses offset acceleration to identify mental abnormal states, thereby enhancing accuracy and timeliness. The proposed method does not require a large amount of target sample data in advance and can adjust the induction or pre-screening strategy in real-time according to the situation. The model undergoes online training and updating throughout the entire process, making it suitable for precise mental recognition in real-world scenarios. To verify the effectiveness of the proposed method, the SEED dataset from the BCMI laboratory at Shanghai Jiao Tong University and the publicly available DEAP emotional dataset were used for validation. The experimental results confirm the effectiveness and superiority of this method in practical applications.