GAI內裂瑕疵音頻生成技術

本研究為解決金屬零組件內裂瑕疵品音頻資料取得不易問題，開發內裂瑕疵音頻資料生成技術。 研究之音頻資料取樣頻率128kHz，取樣長度32,768 個資料點。生成式演算法選用SAGAN（Self-Attention Generative Adversarial Networks） 進行稀缺資料生成，藉其自注意力機制的全局資訊關聯生成音頻資料。 資料合理性驗證採用多維KS 檢定法(Kolmogorov-Smirnov Test），生成資料總體相似率達95%。將生成的1,000 筆虛擬資料與真實資料混合訓練，使AI（Artificial Intelligence） 檢測模型之測試準確率從87% 提升至93%。

In order to address the difficulty in obtaining acoustic data of internal crack defects in metal components, this study develops an acoustic data generation technology for internal crack defects. The acoustic data sampling frequency is 128 kHz with a sampling length of 32,768 data points. The generative algorithm employs SAGAN (Self-Attention Generative Adversarial Networks) for scarce data generation, leveraging its selfattention mechanism's global information correlation to generate acoustic data. Data validity verification uses the multivariate KS test (Kolmogorov-Smirnov test), achieving an overall similarity rate of 95% for the generated data. By combining 1,000 generated synthetic data samples with real data for training, the test accuracy of the detection model improved from 87% to 93%.