領域自適應於銑削加工中刀具磨耗偵測

Abstract

近年來，製造業對於預測工具機故障和刀具磨耗的模型開發越發重視。傳統的刀具磨耗監測系統研究通常基於已知的切削條件（如轉速和進給）建立模型，但這些模型只能對特定切削條件下的測量數據具有良好的辨識性能。對於變切削條件的刀具磨耗監測，如在材料硬度不同的情況下，切削條件的差異會對磨耗產生影響，傳統方法往往難以有效應對。因此，本研究採用領域自適應的概念，以神經網絡模型為基礎，用於不同硬度材料下的刀具磨耗監測。 研究中利用振動訊號和主軸電流訊號作為預測刀具磨耗的訊號源。採用安裝方便、成本低廉且信噪比高的加速規進行振動訊號擷取，並使用從工具機收集的主軸電流訊號，因其與切削力高度相關。透過對抗性訓練方法，建立的領域自適應模型能夠有效預測刀具磨耗，驗證了該模型的可行性。進一步，通過結合與目標領域相似的切削數據進行訓練，模型的性能得到提升。此外，雖理論上多樣數據可提升泛化，但當源領域中的部分數據與目標領域數據差異大時，效果可能不佳。源領域選擇的相似性對模型性能是關鍵。適當微調和組合源領域數據能增強預測準確性。在減少特徵工程不進行數據標準化下，領域自適應依然能縮小源領域與目標領域特徵分布的差距。本研究間接證明領域自適應與特徵工程對於模型泛化性有相同之功效，對於已進行有效的資料前處理與特徵工程之數據，用於不同硬度材料下的刀具磨耗監測，採用領域自適應，準確度僅能小幅提高。

In recent years, the manufacturing industry has increasingly focused on the development of predictive models for machine failure and tool wear. Traditional research on tool wear monitoring systems typically relies on established cutting conditions (such as spindle speed and feed rate) to build models that only exhibit good recognition performance for specific conditions. However, for tool wear monitoring under variable cutting conditions, especially when dealing with different material hardness, traditional methods often struggle to be effective. Therefore, in this study, we utilize the concept of domain adaptation to construct a neural network model for tool wear monitoring in different material hardness scenarios. The research uses vibration signals and spindle current signals as the signal sources for predicting tool wear. Vibration signals are captured using accelerometers, known for their easy installation, cost-effectiveness, and high signal-to-noise ratio. Additionally, spindle current signals from the machine tool are utilized, given their strong positive correlation with cutting forces. By employing adversarial training methods, the domain adaptation model effectively predicts tool wear, validating the feasibility of this model. Furthermore, by adding cutting data similar to the target domain for training, the performance of the model is enhanced. In theory, diverse data can improve model generalization. However, the effectiveness may be reduced when there is a significant discrepancy between some data in the source domain and the target domain data. The similarity in the choice of source domain is crucial to the performance of the model. Proper fine-tuning and combining of source domain data can enhance the prediction accuracy. Even when reducing feature engineering and without standardization, domain adaptation can still work to minimize the discrepancy in feature distribution between the source and target domains. This research indirectly shows that domain adaptation and feature engineering have similar effects on model generalization. When used for tool wear monitoring under different material hardness, domain adaptation results in only a slight improvement in accuracy for data that has undergone effective preprocessing and feature engineering.