This study develops an integrated knee joint monitoring system combining wearable sensors with stationary cycling equipment, designed to detect proper cycling postures and identify three common error patterns: external rotation of the knee joint, internal knee buckle and unstable centre of gravity. The system aims to prevent exercise-induced injuries through real-time biomechanical feedback during cycling training sessions.
The experimental protocol involved collecting kinematic posture data from multiple participants, followed by preprocessing using sliding window segmentation and cycle-split algorithms. Principal component analysis was employed for dimensionality reduction, extracting seven principal features characterizing knee joint kinematics. While establishing baseline performance using conventional machine learning models (support vector machine, decision tree), we developed a novel hybrid deep learning architecture integrating long short-term memory (LSTM) networks with convolutional neural networks (CNN). This configuration enables simultaneous temporal pattern recognition through LSTM layers and spatial feature extraction via convolutional neural networks modules, achieving robust multi-class exercise posture recognition.
Following comparative evaluation, the LSTM-CNN hybrid deep learning architecture emerged as the optimal solution, demonstrating superior performance with 98.4% recognition accuracy. This hybrid model significantly outperformed conventional machine learning benchmarks.
This study realizes a wearable motorcycle riding posture monitoring system at the knee joint. After the functional test of the system, the recognition accuracy of riding posture at the knee joints of four types of dynamic bicycles reaches 94.3%, which can reduce the injuries brought by incorrect postures of the wearer in the process of riding, and it has a certain degree of practicality.
This pioneering work introduces the first successful implementation of hybrid deep learning architecture LSTM-CNN in cycling motion monitoring effectively overcoming the inherent temporal feature extraction limitations of traditional machine learning approaches through its dual-path spatiotemporal processing mechanism.
