將感測器信號和加工參數編碼成圖像用於雷射切割的遷移學習;Images Encoded by Sensor Signals and Processing Parameters for Transfer Learning of Laser Cutting

NCU Institutional Repository > 工學院 > 機械工程研究所 > 博碩士論文 > Item 987654321/87030

請使用永久網址來引用或連結此文件: http://ir.lib.ncu.edu.tw/handle/987654321/87030

題名:	將感測器信號和加工參數編碼成圖像用於雷射切割的遷移學習;Images Encoded by Sensor Signals and Processing Parameters for Transfer Learning of Laser Cutting
作者:	黃鍾易;Huang, Chung-Yi
貢獻者:	機械工程學系
關鍵詞:	雷射切割;切口寬度;深度學習;遷移學習;卷積神經網路;Laser Cutting;Kerf Width;Deep Learning;Transfer Learning;Convolutional Neural Network
日期:	2021-10-22
上傳時間:	2021-12-07 13:49:21 (UTC+8)
出版者:	國立中央大學
摘要:	近年來，雷射切割在工業上有很大的影響力，可透過更改雷射加工參數對不同材質切割來製作成品。但尋找參數與切口關係的過程中會消耗大量材料，尤其在切割貴重材料如藍寶石、鑽石等，會造成很大的損失。為了減少耗材，遷移學習是個可用的選項，利用便宜之耗材建立模型，再遷移至貴重材料模型。基於卷積神經網路(Convolution Neural Network, CNN)在影像辨識上有重大成功與遷移學習的先例，將雷射功率感測器信號與其他加工參數轉換為圖像來建立資料集。本研究使用光纖雷射切割機切割較便宜的不鏽鋼與較貴重的矽鋼片，首先建立不鏽鋼的訓練模型，再利用遷移學習訓練矽鋼片訓練模型，節省大量學習參數的同時還可節省矽鋼片的材料浪費。利用CNN建構出一個切割品質的預測模型，此模型以轉換的圖像資料集當作輸入，由神經網路推導出的雷射切口平均寬度與標準差。在神經網路部分，使用了3種架構來驗證出適合的模型，並在架構之中套用了已訓練好的模型架構(VGG16、ResNet50、GoogleNet)來做比較。本實驗使用了24組不同的切割參數組合，實際切割出120條不鏽鋼切割線與48條矽鋼片切割線，在類神經網路的訓練部分，很好的預測出此24組組合的切口品質，驗證資料的均方誤差約為0.0263，並且在遷移模型上也得到了不錯的效果，驗證資料的均方誤差約為0.0162。經過上述的實驗，由本實驗所建立的雷射切口寬度的預測模型，可以有效地在不同的材質上進行遷移預測，可以減少新材料在實驗中的浪費，降低開銷。 ;Recently, laser has a great influence in the industry. It can produce products through different materials by different parameters. Searching for the best parameters often consumes a lot of materials and increases a lot of cost too much when cutting expensive materials like diamond, sapphire. In order to save the cost, transfer learning is an available option to use cheap material to train a model and then transfer to expensive material model. According to CNN has significant success in images recognition and precedents for transfer learning, we encode the power, repetition rate and cutting speed into images as a dataset. In this study, an optical fiber laser cutting machine was used to cut stainless steel and silicon steel sheets, the stainless steel is cheaper than silicon. First, train a stainless steel training model, and transfer the stainless steel model to the silicon steel training model through transfer learning, which can reduce the training data and the silicon steel sheets. We use the CNN to construct a cutting quality prediction model. This model uses the images dataset as the input, and derives the average width of the cutting kerf and the standard deviation. In the neural network part, three kinds of structures are used to verify suitable models for this case, and the currently pretrained models on the market (VGG16, ResNet50, GoogleNet) are applied to these structure to accelerate the process of calculation. This experiment uses 24 sets of different cutting parameter combinations, and actually cuts 120 stainless steel cutting lines and 48 silicon steel sheets cutting lines. In the training part of the neural network, the cutting quality of the 24 sets of combinations is well predicted, and the mean square error of the verification data is about 0.0263. We also got good results on the transfer learning mode, the mean square error of the verification data is about 0.0162l. After the above experiments, the prediction model of the laser kerf width established by this experiment can effectively predict on the different materials, it also can reduce the waste of new materials in the experiment, and reduce the cost.
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	86	檢視/開啟

在NCUIR中所有的資料項目都受到原著作權保護.

社群 sharing

資料載入中.....