基於冰機資料之異常值檢測結果評比與無標註樣本之群集數目預測演算法

NCUIR > College of Electrical Engineering & Computer Science > Graduate Institute of Electrical Engineering > Electronic Thesis & Dissertation > Item 987654321/88381

Please use this identifier to cite or link to this item: http://ir.lib.ncu.edu.tw/handle/987654321/88381

Title:	基於冰機資料之異常值檢測結果評比與無標註樣本之群集數目預測演算法
Authors:	李昱萱;Lee, Yu-Hsuan
Contributors:	電機工程學系
Keywords:	無標註資料;故障檢測與診斷;異常值演算法評估;機器學習;unlabeled data;anomaly detection evaluation;fault detection and diagnosis (FDD);machine learning
Date:	2022-05-09
Issue Date:	2022-07-14 01:08:58 (UTC+8)
Publisher:	國立中央大學
Abstract:	本研究以於冰水主機設備（冰機）的效能分析為契機，發展三個項目：冰機效能分析演算法、異常值檢測結果之評分演算法與無標註資料的群集中心數目預測。冰機是暖通空調設備中耗電的大宗，而暖通空調設備這類大型機器的運轉參數優化依賴於機械專家的知識，隨著老技師的凋零如何調整參數以優化能耗成為企業需解決的問題。本項研究在訓練資料上訓練非監督式的異常值檢測器，檢測來自測試資料的異常徵兆。並透過尋找訓練資料中與測試資料狀態相近的樣本，給予被診斷出狀態不良的測試資料一個建議的調整方向，使它能回到正常的運作狀態上。此效能優化演算法中，挑選好的異常值檢測器以及給予一筆測試資料一個一致的調整方向是演算法的關鍵，然而來自場域的資料屬於無標註資料，既沒辦法透過標註的絕對真相 (ground truth)找出最佳異常值檢測器，在測試資料各筆的調整向量不一致時，在不知曉群集數目的情況下，也無法透過分群手段把不一致的調整向量分離。因此，本研究針對此二問題做進一步的探討。面對無標註資料的困境，先在人工生成的資料上以不依賴絕對真相的前提下，發展出對於不同的異常檢測結果之間在過濾相同比例異常值情況下的結果評分演算法。並觀察到預測錯誤的數量與評分分數之間具有一致性，藉以克服沒有絕對真相難以選擇異常值檢測器的困難點，提供自動化前處理一個輔佐性的指標，以利後續的肇因診斷。對於後者問題，本研究同樣在人工生成的資料上，發展了資料存在雜訊（異常值）的情況下依然可以尋找出資料的群集中心數目的演算法，並且可以透過後續的迭代更新提昇找到的群集中心之準確率，提供 k-means 等需要事先決定群集數量的演算法有關群集數量以及中心初始位置的資訊，解決無標註資料難以下定群集數量的難題。 ;Based on chiller performance analysis, this paper studies three tasks: fault detection and diagnosis (FDD) of chiller, anomaly detection evaluation, and cluster number prediction for unlabeled data. The chiller is one of the main energy consumption machines of HVAC (heating, ventilation, and air conditioning) equipment. The optimization of operating parameters of HVAC equipment relies on the knowledge of mechanical experts. With the aging of machinery, how to adjust parameters to optimize energy consumption has become a problem that enterprises need to solve. The first part is the fault detection and diagnostics (FDD) system. This study detects the bad state of test data by training an unsupervised outlier detector on training data. After a bad state is found, the system will try to make suggested adjustments to the test data to restore the bad state to a normal state. The keys to this FDD system are how to choose a good outlier detector and how to give test data a consistent adjustment. However, for the former point, data from the factory is unlabeled, so we are challenged to get the ground truth to choose the best outlier detector. For the latter point, getting the main adjustment vector by clustering under the unknown number of clusters is difficult. The second and third parts of this study focus to solve the above two problems. Faced with the difficulty of unlabeled data, we developed a scoring algorithm on artificially generated data to evaluate the classification results of different anomaly detectors with the same ratio of outliers. We observed a consistent relationship between the evaluation score and the wrong prediction number. This scoring algorithm can supply the automatic data pre-process in the FDD system with an index and overcome the challenge of choosing a good outlier detector without ground truth. For the problem that the number of clusters on unlabeled data is unknown, we design an algorithm to predict the cluster centers and locations on data with or without noise. The accuracy of the found cluster centers can be improved through iterative updates. It can provide a basis for clustering algorithms such as k-means that need to determine the number of clusters in advance and solve the problem that it is difficult to determine the number of clusters for unlabeled data.
Appears in Collections:	[Graduate Institute of Electrical Engineering] Electronic Thesis & Dissertation

Files in This Item:

File	Description	Size	Format
index.html		0Kb	HTML	66	View/Open

社群 sharing

Loading...