研磨參數最佳化與以其在線數據預測工件表面粗糙度

Abstract

本研究使用機械手臂進行自動化研磨加工，而為了提升工件的研磨品質，針對不同的研磨參數進行最佳化，選用的研磨參數包括砂紙號數、研磨力、研磨輪轉速、進給速度以及研磨角度，本研究進行研磨參數最佳化主要可分為兩大部分：模型訓練與最佳化參數，首先利用機器學習方法建立表面粗糙度模型，再利用預測模型進行最佳化參數。 由於研磨過程與測量表面粗糙度需耗費大量時間，本研究訓練模型所使用的資料是利用中央合成設計（central composite design）進行蒐集，相較於全因子實驗能夠使用較少的實驗組數蒐集到訓練資料，而研究中選用了線性模型、支持向量機（support vector machine）、神經網路（neural network）分別進行訓練，利用驗證資料集作為訓練超參數的挑選，再使用測試資料集檢驗模型表現，最後再將模型組合成堆疊（Stacking）模型，使預測結果再進一步提升。 訓練後的模型會套用於粒子群最佳化（Particle Swarm Optimization, PSO）演算法進行研磨參數最佳化，分別使用了支持向量機、神經網路與堆疊模型生成了不同砂紙號數下的最佳化研磨參數，接著再透過實驗驗證其效果，驗證實驗除了使用與訓練資料中相同的黃銅片工件，另外也使用了曲面的黃銅棒進行研磨，最終使用最佳化參數所得到的研磨結果皆比未使用的更好，其中240號砂紙使用支持向量機模型的最佳化參數所得到的表面粗糙度最小，400號砂紙則是神經網路模型，800號砂紙則是堆疊模型，將這些參數應用於黃銅棒上同樣也得到比未使用最佳化參數更小的表面粗糙度，以此驗證了預測模型的準確度以及最佳化演算法的效果。

In this research, an automated grinding process using a robotic arm was employed to enhance the grinding quality of workpieces. The optimization of grinding parameters, including grit size, grinding force, wheel speed, feed rate, and grinding angle, was conducted to achieve optimal results. The optimization process can be divided into two main parts: model training and parameter optimization. Initially, a surface roughness model was developed using machine learning techniques, followed by the parameter optimization with the prediction models. Due to the time-consuming nature of grinding and surface roughness measurements, the central composite design was utilized to collect training data with fewer experimental runs compared to the full factorial design. Linear model, support vector machine (SVM), and neural network were trained using the collected data. The selection of training hyperparameters was based on the validation dataset, and the performance of the models was evaluated using the testing dataset. Finally, the models were combined into a stacking model to further improve the prediction accuracy. The trained models were then applied to the particle swarm optimization (PSO) algorithm for grinding parameter optimization. The support vector machine (SVM), neural network, and stacking model are utilized to generate the optimized grinding parameters for different grit sizes. The optimized parameters were experimentally validated using brass specimens, including flat plates and curved rods. The results demonstrated that the surface roughness obtained with the optimized parameters was superior to the non-optimized parameters. The SVM model yielded the lowest surface roughness for the 240-grit size, the neural network model for the 400-grit size, and the stacking model for the 800-grit size. The application of the optimized parameters to the curved brass rods also resulted in improved surface roughness compared to non-optimized parameters, validating the accuracy of the prediction models and the effectiveness of the optimization algorithm.