A health monitoring and exercise optimization method based on fusion of action recognition and physiological signals—A study on students at the Zhejiang police college

With the growing emphasis on health management and the widespread adoption of smart devices, the demand for motion recognition and health anomaly monitoring has become increasingly pressing. In particular, for students at the Zhejiang Police College, maintaining good physical fitness and health is critical because they often face high-intensity work pressure and unpredictable work environments. Fundamentally, health monitoring is a pattern recognition challenge, wherein the objective is to extract critical features from multimodal data to detect abnormal behaviors or health conditions, thereby enabling real-time surveillance of individual health statuses. Against this backdrop, this paper introduces a novel health monitoring framework (PSCLM), which integrates convolutional neural networks (CNN), long short-term memory networks (LSTM), and modal decomposition techniques, aiming to enhance the precision and resilience of action recognition and health anomaly detection. Initially, the proposed framework analyzes inertial sensor data and constructs a convolutional long short-term memory (CLM) model utilizing CNN and LSTM to facilitate movement type recognition. Subsequently, given the significance of heart rate as a vital indicator during physical activity, heart rate and electrocardiogram (ECG) signals, along with their variational mode decomposition (VMD) decomposition features, are incorporated with movement recognition features to achieve multidimensional and hierarchical fusion of physiological signals during exercise. Finally, leveraging these fused features, the framework achieves comprehensive monitoring of exercise-induced health states. Experimental evaluations reveal that the PSCLM framework attains an average accuracy exceeding 0.9 for action recognition on publicly available datasets, while its accuracy for abnormal state detection reaches 0.95, outperforming traditional approaches and single deep learning models significantly. This research not only provides an innovative technological approach for sports health management and anomaly detection but also establishes a critical foundation for advancing next-generation intelligent health monitoring systems.