網宇實體系統設計之土石流預測系統

Landslide Prediction -- A Cyber-Physical System Design

呂俊賢

指導教授 ： 熊博安

國立中正大學/工學院/資訊工程學系/博士(2016年)

摘要

土石流是一種會對於人生安全以及財產帶來巨大損害的天然災害。雖然土石流已經能做到即時偵測，但仍然沒有足夠的時間能夠採取對應的策略來保護生命以及財產的安全。為了最小化土石流所帶來的災害，土石流的預測以及事先預警是必須的。一旦預警事先發送，住在可能受影響的區域內的人們就可以在土石流發生前有更多的反應時間逃離。因此，本論文嘗試提出一個針對土石流預測的網宇實體系統設計。網宇實體系統設計方法提供了一種方式來整合溝通、計算以及控制流程。它能夠良好的利用由無線感測器網路收集來的資料並且用一種智慧方法來解決無線感測器網路上的傳統問題。

基於土石流預測系統的目標上，我們更進一步的討論三個重要的議題，叫做在無線感測器網路中的資料遺失和高耗能以及土石流預測準確度。為了解決遺失的用來做土石流偵測的土壤濕度資料，我們使用有時空關係的異質資料來重建遺失的土壤溼度資料，例如雨量強度。在空間的重建方法中，我們不只考量距離因素同時也考量兩個地點間的雨量強度差距。在時間的重建方法中，如果沒有下雨時，因為蒸發散作用的關係，土壤濕度將會減少。當沒有下雨的時候，我們利用歷史的蒸發散率來重建遺失的土壤溼度資料。如果有雨，我們會感測一段時間內的土壤溼度來重建遺失的資料。我們同時也考量了雨量強度。最後，我們可以用時空重建測量誤差來改進重建的準確率。當有80%隨機遺失的土壤溼度資料時，時空重建的方均根誤差會低於2。為了精準的預測土石流的發生，我們建立了一個人工神經網路的預測模型並且為了預測模型設計一個重新訓練的流程。不同特性的資料會被用來當作訓練資料。為了解決高耗能的問題，我們提出了一個基於模型預測控制的取樣頻率調整方法，根據斜坡的穩定性來動態調整取樣頻率。在高穩定性時，環境資料是不需要的並且監控斜坡的感測器會被設置為低取樣頻率以節省能源。在低穩定性時，感測器被設置為高取樣頻率來感測更多資料做精準的預測。

實驗結果展示了被提出的網宇實體土石流預測系統可以達到83.33%的高預測準確性。對比於準確度高於40%的固定取樣頻率的方法，被提出的基於模型預測控制的取樣頻率調整方法平均可以節省53.33%的耗能。最後，我們提出的系統能夠預先平均38.8分鐘預測土石流的發生。

關鍵字

網宇實體系統 ； 土石流預測 ； 類神經網路 ； 模型預測控制 ； 資料重建 ； 時空關係 ； 異質資料

並列摘要

Landslide is one of the natural disasters that could cause huge damages to properties and great loss of life. Though landslide detection could be done in real-time, there is still not enough time to adopt any reacting policy to keep human life and properties safe. In order to minimize the extent of losses caused by landslides, prediction and pre-alarm of the occurrence of landslides are necessary. Once an alarm is issued in advance, the people living in the possibly affected areas could have more response time to evacuate before the landslide occurs. Hence, this Dissertation tries to propose a Cyber-Physical System (CPS) design for landslide prediction. Cyber-Physical System design approach provides a way to integrates communication, computation and control processes. It could well utilize the data collected from the Wireless Sensor Network (WSN) and solve the traditional problem of WSN in an intelligent way.

Based on the target for landslide prediction system, we further discuss three important issues namely data loss and high energy consumption in the wireless sensor network and landslide prediction accuracy. To solve the missing soil moisture data used in landslide prediction, we use heterogeneous data with spatio-temporal relation to reconstruct missing soil moisture data, such as rainfall intensity. In spatial reconstruction method, we take not only the distance but also the difference of rainfall intensity between two locations into consideration. In temporal reconstruction method, soil moisture will decrease because of evaporation if there is no rainfall. We use historical evaporation rate to reconstruct missing soil moisture data when there is no rainfall. If there is rain, we sense soil moisture data to reconstruct missing data in a period time. We take the rainfall intensity into consideration as well. The last, we can improve the reconstruct accuracy with spatio-temporal reconstruction estimation error. The RMSE of spatio-temporal reconstruction is below 2 when there is 80\% random missing soil moisture data. To accurately predict the occurrence of landslide, we construct an artificial neural networks (ANNs) prediction model and design a retraining flow for the prediction model. The data with different features are used as the retraining data. To solve the high energy consumption problem, we proposed a Model Predictive Control (MPC) based sampling rate tuning method to dynamically adjust the sampling rate due to the stability of the slope. With high stability, the environment data is needless and the sensors monitoring the slope is set to low sampling rate to save the energy. With low stability, the sensors is set to high sampling rate to sense more data for accurate predicting.

Experimental results show the proposed Cyber-Physical landslide prediction system can achieved high prediction accuracy, $83.33\%$. Compare to the fixed sampling rate method with the accuracy more than $40\%$, the proposed MPC based sampling rate tuning method could save $53.33\%$ energy consumption in average. Finally, our proposed system could predict the occurrence of the landslide in average $38.8$ minutes in advance.

並列關鍵字

Cyber-Physical System ； Landslide Prediction ； Artificial Neural Network ； Model Predictive Control (MPC) ； Data Reconstruction ； Spatio-Temporal Relation ； Heterogeneous Data

參考文獻

[1] F. Dai, C. Lee, and Y. Ngai, “Landslide risk assessment and management: an

overview,” Engineering Geology, vol. 64, no. 1, pp. 65–87, April 2002.

[3] A. Musaev, D. Wang, and C. Pu, “LITMUS: A Multi-Service Composition System for Landslide Detection,” IEEE Transactions on Services Computing, vol. 8,

[4] B. Wang, “A Landslide Monitoring Technique Based on Dual-Receiver and

Phase Difference Measurements,” IEEE Geoscience and Remote Sensing Letters,