| 摘要: | 雷射熱裂（Laser Thermal Cracking, LTC）為一種應用於脆性材料精密加工的非接觸式切割技術。即使在相同加工參數條件下，裂縫的傳播行為仍可能因材料內部應力場與熱擴散之不穩定性而產生顯著差異，導致裂縫停滯、熱損傷，甚至切割失敗等異常情形。傳統方法多著重於加工參數的調控，缺乏即時掌握材料內部應力變化的能力，限制了對異常狀況的早期識別與精準控制。
本研究導入卷積長短期記憶網路（ConvLSTM）進行時間序列影像預測。模型訓練階段分別採用平均平方誤差（Mean Squared Error, MSE）與結構相似度指標（Structural Similarity Index, SSIM）作為損失函數與性能評估指標，以同時評估影像亮度與結構還原的準確性。
實驗結果顯示，當採用 MSE 作為損失函數時，所建構之 ConvLSTM 模型能有效預測裂縫應力場之時序變化，其測試樣本平均 SSIM 為 0.842、MSE 為 0.0281，整體預測結果穩定，且能有效保留原始影像的結構特徵。進一步分析指出，在約 77% 的測試樣本中，模型預測結果通常可較實際影像提前一影格出現應力變化。本研究進一步整合既有之「影像結構特徵指標系統」，包含亮度、輪廓面積、長寬比與影像重心等多項特徵，達到於異常發生前0.8秒成功預測之結果。
;Laser Thermal Cracking (LTC) is a non-contact cutting technique for precision machining of brittle materials. Even under the same processing parameters, the crack propagation behavior may still be significantly different due to the instability of the internal stress field and heat dissipation of the material, resulting in abnormal situations such as crack stagnation, thermal damage, and even cutting failure. Traditional methods focus on the regulation of processing parameters and lack the ability to immediately grasp the changes in the internal stress of the material, which limits the early identification and precise control of abnormal conditions.
In this study, a convolutional short-term and long-term memory network (ConvLSTM) is used to predict time series images. In the model training stage, Mean Squared Error (MSE) and Structural Similarity Index (SSIM) are used as the loss function and performance evaluation index to evaluate the accuracy of image brightness and structural restoration, respectively.
The experimental results show demonstrate that the ConvLSTM model can effectively predict the temporal changes of in the crack stress field when MSE is used employed as the loss function, and the yielding an average SSIM and MSE are of 0.842 and 0.0281, respectively. The overall prediction results are stable, and the structural features of the original images are effectively preserved. It is further analyzed that in about 77% of the test samples, the model predictions usually show the stress changes one frame earlier than the actual images. This study further integrates the existing classification framework, which includes brightness, contour area, aspect ratio, and image center of gravity, to achieve a successful prediction result of 0.8 seconds before the occurrence of the anomaly. |
