Texture-Aware Ridgelet Transform and Machine Learning for Surface Roughness Prediction
Texture-Aware Ridgelet Transform and Machine Learning for Surface Roughness Prediction
复制标题
DOI：
10.1109/tim.2022.3214630
复制
发表时间：
2022
影响因子：
5.6
通讯作者：
Clayton Cooper;Jianjing Zhang;Liwen Hu;Yuebin Guo;R. X. Gao
中科院分区：
文献类型：
--
作者：
Clayton Cooper;Jianjing Zhang;Liwen Hu;Yuebin Guo;R. X. Gao
加工表面粗糙度的量化对于估计零件性能（如摩擦学和疲劳）至关重要。作为传统接触式轮廓测量的非接触式替代方法，由于图像处理和ML技术的进步，摄影方法已被广泛应用，这些技术允许分析嵌入在光学图像中的表面特征并将这些特征与表面粗糙度相关联。现代摄影方法广泛采用二维小波变换进行图像处理。然而，2-D小波往往是有限的，在捕捉线图案，是普遍的加工表面，由于其径向对称的性质，导致次优的表面表征。此外，表面粗糙度预测主要使用ML方法作为点预测进行，该方法不考虑模型和数据中的不确定性。为了解决这些局限性，本研究提出了一种脊波变换（RT）为基础的加工表面表征方法。RT自动检测主导线模式，即，纹理，并提取拓扑特征，例如沿沿着与引起表面粗糙度最相关的方向嵌入表面轮廓中的组成空间频率。提取的纹理感知功能，然后用作输入随机森林（RF）和核密度估计的表面粗糙度预测和不确定性量化。使用实验数据的评估表明，所开发的方法预测表面粗糙度的误差为0.5%，优于现有的技术，并展示了潜在的RT作为一个可行的技术加工表面分析。
Quantification of machined surface roughness is critical to enabling estimation of part performance such as tribology and fatigue. As a contactless alternative to the traditional contact profilometry, photographic methods have been widely applied due to the advancement of image processing and ML techniques that allow the analysis of surface characteristics embedded in optical images and association of these characteristics with surface roughness. The state-of-the-art of photographic methods make extensive use of 2-D wavelet transform (WT) for image processing. However, a 2-D wavelet is often limited in capturing line patterns that are prevalent in the machined surface due to its radially symmetric nature, leading to suboptimal surface characterization. In addition, surface roughness prediction is primarily carried out as point prediction using ML methods which do not account for uncertainty in the models and data. To address these limitations, this study presents a ridgelet transform (RT)-based method for machined surface characterization. RT automatically detects the dominant line patterns, i.e., texture, in surface images and extracts topological features, such as the constituent spatial frequencies embedded in the surface profile along the direction that is most relevant for inducing surface roughness. The extracted texture-aware features are then used as inputs to random forest (RF) and kernel density estimation for surface roughness prediction and uncertainty quantification. Evaluation using experimental data shows that the developed method predicts surface roughness with an error of 0.5%, outperforming existing techniques and demonstrating the potential of RT as a viable technique for machined surface analysis.