應用分治法於刀具壽命預測模型之研究;Application of Divide and Conquer Approach to Tool Life Prediction Model

NCUIR > Engineering College > Graduate Institute of Mechanical Engineering > Electronic Thesis & Dissertation > Item 987654321/93830

Please use this identifier to cite or link to this item: http://ir.lib.ncu.edu.tw/handle/987654321/93830

Title:	應用分治法於刀具壽命預測模型之研究;Application of Divide and Conquer Approach to Tool Life Prediction Model
Authors:	王治全;Wang, Chih-Chuan
Contributors:	機械工程學系
Keywords:	刀具狀態監測;分治法;深度學習;遷移式學習;聚類演算法;Tool Condition Monitoring;Divide and Conquer Approach;Deep Learning;Transfer Learinging;Clustering Algorthm
Date:	2023-07-27
Issue Date:	2024-09-19 17:40:59 (UTC+8)
Publisher:	國立中央大學
Abstract:	刀具殘餘壽命受到多種因素的影響，因此應用深度學習來分析資料集特徵並推論時常會面臨一定的困難。常見的做法是提高模型的學習能力，以便辨識複雜特徵趨勢。然而，本研究認為降低資料集的特徵複雜度也能夠實現同樣的目標。本研究提出的方法是基於參考分治法，將複雜目標域拆分成數個簡單的局部域，再使用多個神經網路擬合局部域的特性，最後將其統整以預測目標域的表現。首先，應用像是隨機森林回歸之類的弱回歸器找出目標域資料集的重要特徵，用於降低其特徵維度。其次，利用聚類演算法將目標域資料集拆分為數個局部域資料集，藉此讓個別局部域的資料具有相近特徵。第三，利用適當的神經網路學習個別局部域的資料特徵，故個別的局部網路能推論對應之局部域表現。最終，提出一決策網路來統整數個局部網路，達成推論目標域表現之目的。透過既有的資料集進行驗證，本研究提出的模型架構能夠以適當參數量對複雜的資料集快速學習與有效推論。此外，決策網路能夠解釋各局部網路對全域表現推論的貢獻狀況，顯示此方法的合理性。最後，探討局部網路之間遷移式學習。本研究發現以凍結少許推論層或微調等兩種方式能夠有效地降低個別局部網路的訓練時間，進而進一步地降低時間上的訓練成本。 ;Tool residual life is affected by many causes, so it is quite difficult to apply deep learning to analyze dataset features and make inferences. A common solution is to increase the capability of the model in order to identify complicated feature trends. However, this study suggests that the same goal can be achieved by reducing the complexity of the dataset. The method proposed in this study refers divide and conquer appeorach, in which the complex target domain is split into several simple local domains, and then multiple neural networks are used to fit the properties of the local domains and finally integrate them to predict the representation of the target domain. First, a weak regressor such as random forest regression, is applied to find the important features of the target domain dataset, which is used to reduce its dimension of features. Second, a clustering algorithm is used to split the target domain dataset into several local domain datasets, so that the data in individual local domains have similar and close features. Third, we use appropriate neural networks to learn the features of individual local domain, so that the decision local networks can infer the corresponding local domain performance. Finally, a network is proposed to integrate several local networks to achieve the purpose of inferring the performance of the target domain. The model framework with appropriate quantity of parameters proposed in this study can learn quickly and infer effectively from complex datasets.In addition, the decision network is able to explain the contribution of each local network to the inference of the global performance, which shows the rationality of this approach. Finally, we investigate the transfer learning among local networks. It is found that freezing a few inference layers or fine-tuning can effectively reduce the training time of local networks, which further reduces the training cost in terms of time.
Appears in Collections:	[Graduate Institute of Mechanical Engineering] Electronic Thesis & Dissertation

Files in This Item:

File	Description	Size	Format
index.html		0Kb	HTML	23	View/Open

社群 sharing

Loading...