Hierarchical deep Q-network for anomaly detection and interpretation in multidimensional time-series physical fitness data.

This paper addresses the critical challenge of achieving both accuracy and interpretability in anomaly detection within complex multidimensional time-series physical fitness data. We design and implement an innovative framework based on a Hierarchical Deep Q-Network (HDQN) that effectively processes and learns the inherent multi-scale and multi-granularity characteristics of physical fitness data. The introduction of a hierarchical structure enables the model to capture subtle patterns at different levels of abstraction, from low-level sensor fluctuations to high-level behavioral patterns. The Deep Q-Network component trains an agent that learns an optimal policy for identifying nuanced and complex anomalous data points or sequences, moving beyond traditional threshold-based methods. Evaluated on a real-world dataset comprising 300 participants monitored over 60 days with 8-dimensional physiological metrics, our HDQN framework achieved superior performance with an F1-score of 0.811, outperforming traditional machine learning methods (Isolation Forest: 0.611, One-Class SVM: 0.575), deep learning approaches (LSTM Autoencoder: 0.703), and non-hierarchical Deep Q-Networks (Flat-DQN: 0.743). The hierarchical architecture inherently supports interpretation of detected anomalies by localizing contributory factors within specific data dimensions or temporal segments. Robustness analysis demonstrated that HDQN maintained stable performance with only 3.82% F1-score degradation under 10% Gaussian noise, significantly better than baseline methods. This amalgamation of hierarchical representation learning and deep reinforcement learning significantly enhances the precision of anomaly detection, provides meaningful explanations for identified deviations, and offers a more comprehensive understanding of an individual's dynamic health status, paving the way for proactive and personalized health interventions in physical education and sports science contexts.