磨削訊號特徵擷取與加工品質之關聯研究

Abstract

本研究旨在針對磨削加工中砂輪狀態（銳利與鈍化）進行有效分類，以提升製程穩定性與工件表面品質。由於傳統診斷方式多倚賴人工經驗與後製檢測，無法即時掌握砂輪狀態變化，故本研究提出一套結合人工特徵、自動圖像特徵與混合式特徵的機械學習架構，以增進砂輪失效預測的準確性與穩健性。人工特徵模型方面，擷取頻域、離散小波轉換、短時傅立葉轉換及小波包轉換等共 84 個統計特徵，輸入至 XGBoost 分類器建模，達成準確率 97.45%、精確率 97.55%、召回率 97.45%、F1 分數 0.975 的良好表現。自動特徵模型方面，先將原始訊號轉換為六種圖像（STIM、GASF、GADF、FS、STFT*、CWT），輸入 Vision Transformer（ViT-B/16）模型進行訓練，再透過 L8 田口直交表設計與 Soft Voting 策略選出最佳圖像組合。結果顯示，FS、STFT*、CWT 三圖融合輸入至 ViT 可達成分類最佳表現（準確率 98.9%、F1 分數 0.989），展現優異的深度學習辨識能力。最後，混合式特徵模型整合 84 個人工統計特徵與 FS、STFT*、CWT 三圖所對應之 2304 個圖像特徵，並輸入至 XGBoost 進行混合學習。實驗結果顯示其準確率為 97.6%、精確率 97.7%、召回率 97.55%、F1 分數為 0.976，整體表現穩定，惟未超越單純 ViT 圖像模型之最優效能。綜合而言，ViT 圖像模型透過多圖融合，有效捕捉砂輪磨削訊號中的時頻紋理與能量分佈特徵，展現最佳分類效能；而圖像配合 XGBoost 雖略遜一籌，仍具彈性與實用性。混合式模型則展現融合潛力，未來可透過特徵選擇進一步優化。研究結果驗證深度圖像特徵在磨削狀態監測的應用價值，並說明田口方法在特徵組合篩選上的實用性，為智慧製造中的機械狀態診斷系統建立良好基礎。

This study aims to achieve effective classification of grinding wheel conditions (sharp vs. dull) during grinding processes, thereby enhancing process stability and workpiece surface quality. Traditional diagnostic methods often rely on manual experience and post-process inspection, which fail to capture real-time changes in wheel condition. To address this, a machine learning framework combining handcrafted features, automated image features, and hybrid feature fusion is proposed to improve the accuracy and robustness of wheel failure prediction. For the handcrafted feature model, a total of 84 statistical features were extracted from the frequency domain, discrete wavelet transform (DWT), short-time Fourier transform (STFT), and wavelet packet transform (WPT). These were used to train an XGBoost classifier, achieving excellent performance with an accuracy of 97.45%, precision of 97.55%, recall of 97.45%, and F1-score of 0.975. In the automated feature model, the raw signals were first transformed into six types of images (STIM, GASF, GADF, FS, STFT*, and CWT), which were then input into a Vision Transformer (ViT-B/16) for training. Using the Taguchi L8 orthogonal array and a soft voting strategy, the optimal image combination was selected. The best classification performance was achieved by fusing FS, STFT*, and CWT images, with an accuracy of 98.9% and an F1-score of 0.989, demonstrating the powerful recognition capabilities of deep learning. Finally, in the hybrid feature model, the 84 handcrafted features were combined with 2,304 deep image features extracted from FS, STFT*, and CWT images and fed into XGBoost for hybrid learning. Experimental results showed an accuracy of 97.6%, precision of 97.7%, recall of 97.55%, and F1-score of 0.976. While performance was stable, it did not surpass the best results achieved by the ViT image-only model. In summary, the ViT image model, through multi-image fusion, effectively captures the time–frequency textures and energy distributions within grinding signals, achieving the best classification performance. Although the image–XGBoost model is slightly less accurate, it offers greater flexibility and practicality. The hybrid model shows potential for future enhancement through feature selection. Overall, the findings confirm the value of deep image features in grinding condition monitoring and highlight the practical utility of the Taguchi method in feature combination optimization, laying a solid foundation for intelligent diagnostics in smart manufacturing.