運用模糊分群法與車輛偵測器遺失資料插補之研究

Abstract

車輛偵測器(Vehicle Detector, VD)經常作為智慧型運輸系統(Intelligent Transportation System, ITS)的資料來源,然而時常因為故障或其他原因，造成車輛偵測器回傳資料發生資料遺失(Missing Data)的現象。本研究提出一個兩階段的資料插補法來填補所遺失的資料。第一階段利用模糊分群法將歷史資料分群,第二階段利用 k-NN 演算法找尋最接近的 k 筆歷史資料，以其平均值作為差補值。希望藉由此方法可提升插補精準度,並且解決過去運用模糊分群法無法即時插補的缺點。 本研究利用國道一號北上 2012 年 5 月每個周五的實際VD 資料做為測試來源，並將本研究所提出的差補法與 FCM-OC、k-NN 插補法以及平均值插補法等三種常見插補法的結果比較。經由實測結果發現，在精準度方面,本研究所提出的插補法在 20 個測試情境中有 10 個的表現最佳,而在不為最佳解的測試範例中，與表現最佳方法所得結果的MAPE誤差僅有 0.93%，同時本研究所提出的方法，在所需的演算時間方面，也有著大幅的改善。根據以上的測試結果，顯示本研究所提出的插補法能夠即時提供精準的車輛偵測器資料遺失的插補。

Vehicle Detector (VD) is a common data source for Intelligent Transportation System (ITS). However, data are sometimes missing due to mechanical failure or other reasons. We propose a two-stage missing data recovery method to address this issue. In stage 1, we divide historical data into clusters using Fuzzy c-Means (FCM). In stage 2, we use k-nearest neighbor algorithm (k-NN) to obtain k nearest historical data within the most similar cluster and use their average to replace the missing data. We use 5-minute VD data from northbound Freeway No.1 every Friday in May, 2012 as the test data source. We compare the performance of our method with three other commonly used missing data recovery methods (FCM-OCS, k-NN and Mean-Imputation) under different missing-data-percentage scenarios. In 20 different scenarios we’ve tested, our proposed method out-performs the other methods in 10 scenarios. In other 10 scenarios, the largest difference between the most accurate method and our method is only 0.93% in MAPE. Also, our method has a significant improvement in time-consumption comparing to the FCM-based method. These testing results show that our proposed method is practical and sound