資料載入中.....
|
請使用永久網址來引用或連結此文件:
https://ir.lib.ncu.edu.tw/handle/987654321/98586
|
| 題名: | 基於多目標域適應深度遷移學習之滾珠螺桿變轉速條件預壓監測研究 |
| 作者: | 柯建瑋;Ke, Jain-Wei |
| 貢獻者: | 機械工程學系 |
| 關鍵詞: | 深度遷移學習;多目標域適應;滾珠螺桿;球通頻率;預壓力 |
| 日期: | 2025-06-18 |
| 上傳時間: | 2025-10-17 12:57:28 (UTC+8) |
| 出版者: | 國立中央大學 |
| 摘要: | 滾珠螺桿作為精密機械及工具機的重要零組件，預壓失效而背隙產生將影響定位精度，故預壓狀態的監測至關重要。本研究提出多目標域適應(Multi-Target Domain Adaption)的深度遷移學習(Deep Transfer Learning)應用在滾珠螺桿預壓狀態監測，以深度遷移學習方法解決滾珠螺桿在不同轉速條件下產生樣本偏差的問題，並憑藉多目標域適應能讓模型識別不同變速的監測資料。論文分為四部分–(1)實驗規劃及數據收集：螺桿載台承載30kg砝碼，分別收集等速1000rpm與變速0-1000rpm運轉的振動訊號，各15000筆資料，研究中將變速資料分割成250-500、500-750及750-1000 rpm等三段，等速與三段變速資料作為源域(Source domain)與三個目標域(Target domain)數據，規劃4%、1%預壓及12um背隙等3種預壓狀態螺桿，從而建立判斷分類基準；(2)分類模型選擇：對卷積神經網路(Convolutional Neural Network, CNN)、多層感知機(Multilayer Perceptron, MLP)與極限梯度提升(eXtreme Gradient Boosting, XGBoost)分別建模以判斷螺桿預壓狀態，比較個別模型的準確率；(3)單源單目標遷移模型建立及探討：針對單一源域遷移單一目標域，提出MMD域適應與域對抗方法來處理數據偏差問題，提出域適應遷移方法解決新舊數據偏差問題，前者比較核函數對模型的穩定度與準確率影響。MMD域適應的平均準確率高達98.8%；域對抗則是89.6%；(4) 單源多目標遷移模型建立及探討：導螺桿需因應製程彈性的需求，需頻繁切換轉速條件以配合加工變化，需跟據不同目標域反覆調參與重新訓練模型，為減少其時間成本及提升模型的泛化性，以多目標域適應方法建模。MMD域適應方法基於單源單目標的模型，將前幾層卷積層分為共享層(Share Layer)和私有層(Private Layer)，使模型同時學習多個目標域特徵；域對抗方法則是域分類器改為成多元分類。MMD域適應平均準確率僅82.0%；域對抗平均準確率99.0%。用域對抗同時學習多個目標域特徵，不僅減少訓練時間也不用對每個目標域反覆調參，然而在訓練過程中，學習曲線跟單目標MMD域適應與域對抗方法相比，較不穩定收斂。;Ball screws are important components of precision machinery and machine tools, so it is extremely important to monitor the preload status. Preload failure and backlash will affect positioning accuracy. The purpose of this study is to propose a method for applying deep transfer learning and multi-target domain adaptation to the preload state monitoring of ball screws, and to solve the sample deviation under different speed conditions of ball screws by using deep transfer learning method. Multi-objective domain adaptation enables the model to predict multiple different speed profiles. This paper is divided into four parts – (1) Experimental planning and data collection: The screw stage is loaded with a 30kg weight, and two signals are collected: a constant speed of 1000rpm and a variable speed of 0-1000rpm. The variable speed data is split into 250-500rpm and 500- 750rpm and 750-1000rpm three-stage speed data, constant speed and three-stage speed data as source domain and three target domain data, the experimental design 4%, 1% preload and 12um backlash a total of 3 (2) Classification model selection: Convolutional Neural Network (CNN), Multilayer Perceptron (MLP) and eXtreme Gradient Boosting (eXtreme Gradient Boosting) are used to select the pre-loaded screws to establish the judgment benchmark; (3) Classification model selection: Convolutional Neural Network (CNN), Multilayer Perceptron (MLP) and eXtreme Gradient Boosting (eXtreme Gradient Boosting) are used to select the pre-loaded screws to establish the judgment benchmark; (4) Classification model selection: Convolutional Neural Network (CNN), Multilayer Perceptron (MLP) and eXtreme Gradient Boosting (eXtreme Gradient Boosting) are used to select the pre-loaded screws to establish the judgment benchmark; (5) Classification model selection: Convolutional Neural Network (CNN), Multilayer Perceptron (MLP) and eXtreme Gradient Boosting , XGBoost) modeling and classification of screw preload status, and compare the accuracy of the model; (3) Single-source single-target migration model establishment and discussion: For the migration of a single source domain to a single target domain, the MMD domain adaptation and domain confrontation method is proposed to solve The problem of data deviation is solved by proposing a domain adaptation migration method to solve the problem of new and old data deviation. The former compares the influence of kernel function on the stability and accuracy of the model; the latter is the lambda value. The average accuracy of MMD domain adaptation is as high as 98.8%; domain confrontation is 89.6%; (4) Establishment and discussion of single-source multi-target migration model: Based on the single-source single-target model, the first few convolutional layers are divided into shared layers and private layers, making the model By learning multiple target domain features simultaneously, the average accuracy of MMD domain adaptation is only 82.0%, while the average accuracy of domain adversarial is 99.0%. Using domain adversarial learning to simultaneously learn features of multiple target domains not only reduces training time but also eliminates the need to repeatedly adjust parameters for each target domain. However, its accuracy is far lower than that of the traditional single target domain MMD domain adaptation transfer model, and there is still room for improvement. |
| 顯示於類別: | [機械工程研究所] 博碩士論文
|
文件中的檔案:
| 檔案 |
描述 |
大小 | 格式 | 瀏覽次數 |
|---|
| index.html | | 0Kb | HTML | 12 | 檢視/開啟 |
|
在NCUIR中所有的資料項目都受到原著作權保護.
|
