A Method for Mental Fatigue State Detection Based on SKM-KNNIC Algorithm

To address the challenges posed by the scarcity of labeled data for mental fatigue state recognition and the inefficiency of processing incremental data, we propose a fatigue state recognition model based on a semi-supervised incremental clustering algorithm. The photoplethysmography (PPG) signals corresponding to various fatigue states are collected using photodetector sensors, establishing a comprehensive fatigue state database. Multidomain feature extraction is performed on the PPG signals, yielding a set of 17 key features from the time domain, frequency domain, and time-frequency domain. Feature dimensions are then optimized using a feature selection algorithm based on mutual information, which constructs an optimal subset of features that maximizes relevance to fatigue states while minimizing redundancy. To process the fatigue state database, we employ the KNNIC algorithm within the Seeded-K-means (SKM) semi-supervised learning framework, training the model using both a small amount of labeled data and a large pool of unlabeled data. The KNNIC algorithm is further extended to handle incremental data, leading to the development of a combined SKM-KNNIC model for fatigue state detection. Experimental results demonstrate that the proposed combined model performs effectively on noisy incremental fatigue state data, achieving a clustering accuracy improvement of approximately 30.67% compared to traditional methods.