[论文解读] A specifically designed machine learning algorithm for GNSS position time series prediction and its applications in outlier and anomaly detection and earthquake prediction

Mostafa Kiani Shahvandi|arXiv (Cornell University)|Jun 16, 2020

Time Series Analysis and Forecasting参考文献 21被引用 7

一句话总结

本文提出了一种专为GNSS设计的新型机器学习算法，通过整合均值去除、趋势减去、基于小波的频率分离以及正弦和余弦函数的三角加权，实现对位置时间序列的预测。该算法在3,000个全球GNSS站的全球研究中，性能优于17种其他算法，展现出卓越的准确性和速度，并实现了早期地震预测（例如在2011年东北地震前约2小时）以及高度精确的异常值检测（比其他方法高出3.22%）。

ABSTRACT

We present a simple yet efficient supervised machine learning algorithm that is designed for the GNSS position time series prediction. This algorithm has four steps. First, the mean value of the time series is subtracted from it. Second, the trends in the time series are removed. Third, wavelets are used to separate the high and low frequencies. And fourth, a number of frequencies are derived and used for finding the weights between the hidden and the output layers, using the product of the identity and sine and cosine functions. The role of the observation precision is taken into account in this algorithm. A large-scale study of three thousand position times series of GNSS stations across the globe is presented. Seventeen different machine learning algorithms are examined. The accuracy levels of these algorithms are checked against the rigorous statistical method of Theta. It is shown that the most accurate machine learning algorithm is the method we present, in addition to being faster. Two applications of the algorithm are presented. In the first application, it is shown that the outliers and anomalies in a time series can be detected and removed by the proposed algorithm. In a large scale study, ten other methods of time series outlier detection are compared with the proposed algorithm. The study reveals that the proposed algorithm is approximately 3.22 percent more accurate in detecting outliers. In the second application, the suitability of the algorithm for earthquake prediction is investigated. A case study is presented for the Tohoku 2011 earthquake. It is shown that this earthquake could have been predicted approximately 2 hours before its happening, solely based on each of the 845 GEONET station time series. Comparison with four different studies show the improvement in prediction of the time of the earthquake.

研究动机与目标

• 开发一种针对GNSS位置时间序列独特特征（如低振幅坐标变化、趋势和混合频率信号）量身定制的机器学习算法。
• 在3,000个全球GNSS时间序列上，将该算法与17种传统机器学习方法及统计基准（Theta）进行性能对比评估。
• 展示该算法在实时GNSS数据中检测异常值和异常现象的实用性。
• 探究利用GNSS数据中的前震形变信号实现早期地震时间预测的潜力。

提出的方法

• 该算法首先从GNSS时间序列中减去均值，以使数据居中。
• 通过专用的趋势检测与去除步骤，去除线性或非线性趋势。
• 应用小波变换将信号分解为高频和低频分量。
• 算法提取关键频率，并利用恒等函数与正弦/余弦函数的乘积，计算隐藏层与输出层之间的权重。
• 在模型训练过程中显式引入观测精度，以反映测量的可靠性。
• 该方法在60%的数据上进行训练，实现对未来位置和异常值的低延迟预测。

实验结果

研究问题

• RQ1 是否一种专为GNSS时间序列设计的机器学习算法，能在预测准确率和速度方面超越通用算法？
• RQ2 与现有异常值检测技术相比，该方法在真实GNSS时间序列中检测模拟异常值的效率如何？
• RQ3 能否在足够早的时间检测到GNSS数据中的前震形变信号，从而实现可靠的地震时间预测？
• RQ4 在GNSS时间序列建模中，引入观测精度在多大程度上能提升预测性能？

主要发现

• 在全球3,000个GNSS时间序列的研究中，该算法在17种机器学习方法和统计Theta方法中实现了最高的预测准确率。
• 该算法的运行速度至少是其他方法的三倍，展现出显著的计算效率。
• 在异常值检测方面，当在2,000条真实时间序列（含10,000个模拟异常值）上测试时，该方法的准确率比10种竞争方法高出3.22%。
• 对于2011年东北地震，该算法仅使用60%的数据，即在破裂前约2小时45秒预测到该事件，地面运动误差约为19%。
• 仅使用10%的数据时，该算法提前12小时4分钟30秒预测到地震，但幅度估计误差达40%。
• 该方法成功检测到振幅小至3厘米的前震信号，表明其具备可靠预测的阈值。