Deep modeling and analysis of human big data deepens our understanding of human activities. Periodic time-series signals, e.g., electrocardiographs, collected by health monitoring sensors reflect human health status and assist in disease diagnosis. However, long-term prediction of these signals using deep learning models poses three challenges, namely, sparse features, conservative prediction of extreme peaks, and varying periodic intervals. We address these issues by proposing a prediction framework called EPIC with extreme peak and interval calibrations. EPIC consists of a triple-channel prediction network and a calibration network. The prediction network learns the time-domain, frequency-domain, and deformable features of time-series patterns simultaneously. Amplitude residuals of extreme peaks are emphasized in the designed training loss function. In addition, to alleviate the problem of unaligned predictions resulting from inaccurate periodic intervals, we further design a calibration module to reduce the deviation of periodic intervals. The experimental results and ablation studies indicate that EPIC achieves excellent performance in long-term prediction tasks.

对人类大数据的深度建模和分析加深了我们对人类活动的理解。健康监测传感器采集的周期性时间序列信号（例如心电图）可以反映人体健康状况并辅助疾病诊断。然而，使用深度学习模型对这些信号进行长期预测提出了三个挑战，即稀疏特征、极端峰值的保守预测以及变化的周期间隔。我们通过提出一个名为 EPIC 的预测框架来解决这些问题，该框架具有极端峰值和间隔校准。EPIC由三通道预测网络和校准网络组成。预测网络同时学习时间序列模式的时域、频域和可变形特征。在设计的训练损失函数中强调了极端峰值的幅度残差。此外，为了缓解周期间隔不准确导致的预测不对齐的问题，我们进一步设计了一个校准模块来减少周期间隔的偏差。实验结果和消融研究表明，EPIC 在长期预测任务中取得了优异的性能。