Personalized exercise load and monitoring driven by multi-source heterogeneous sensor data

Abstract In view of the problems faced by exercise load monitoring in the internet of things environment, such as the heterogeneous coupling of multi-source sensing data and the lack of individual adaptability of general models, this study constructs a personalized exercise load and monitoring model driven by multi-source heterogeneous sensor data. This model combines individual physiological portraits with temporal convolutional networks-bidirectional long short-term memory (BiLSTM) networks. This study first uses a sliding window mechanism to solve the spatiotemporal alignment bottleneck of signals with different frequencies, and then captures instantaneous motion mutations in acceleration through a temporal convolution network module. Moreover, this study combines the BiLSTM network and multi-head attention to deeply mine the cumulative fatigue characteristics in long-term data, effectively solving the problem of key risk signals being submerged in the stationary period. The results showed that the model performed well in convergence efficiency, with a comprehensive F 1-Score of 0.968 on the test set, which was significantly better than the benchmark model. Meanwhile, the predicted load was highly linearly correlated with the real physiological indicators ( R 2 = 0.982), and the accuracy remained at 92.5% even under 5 dB strong noise interference. This study successfully breaks through the generalization bottleneck of traditional models in the face of individual differences by embedding static signs into dynamic networks. This provides solid technical support for the realization of precise proactive intervention in sports risks from professional competition to mass fitness.