In-situ Roughness Evaluation of Milling Machined Surface Based on Lightweight Deep Convolutional Neural Network
-
摘要:
传统机器学习类方法对光源类型、设备安装误差等因素较为敏感，需要反复调试与实验，难以实现规模化生产的自动检测. 针对上述问题，提出了一种铣削粗糙度在机监测方法，有效提升了检测效率和准确性. 首先，采用低感度参数设置的方向梯度直方图特征的候选框提取算子实现铣削工件的定位，并基于点匹配算法校正安装误差；然后,通过清晰度评价指标实现工业相机对焦过程优化；最后，构建了一种面向移动端实时计算的轻量级卷积神经网络模型，可对不同粗糙度工件表面纹理进行分类，并在立铣加工纹理数据集上进行了实验验证. 实验结果表明：相比普通卷积神经网络，在模型复杂度相似的情况下，以乘、加运算次数为指标，提出模型推理所需运算量减少55%；代价敏感函数的引入能有效提升粗糙度识别模型对不平衡数据的稳定性；所提方法与传统机器学习方法相比，在检测帧率、图像分辨率相同的实验条件下，精准率、召回率分别提高了8%、21%.
Abstract:Traditional machine learning methods (e.g., hand-coded feature extraction) are sensitive to light sources, equipment installation errors and other factors, which require repeated debugging and experiments and make it difficult to achieve automatic detection in large-scale production. Considering the above-mentioned problems, an in-situ roughness evaluation method is proposed to effectively enhance the efficiency and accuracy of the detection processes. Firstly, an enhanced candidate frame extraction operator for the histogram-of-gradient feature set with low sensitivity parameters is proposed to locate the milling workpiece, and the installation error is corrected using the point matching algorithm. Then, the focusing process of the industrial camera is optimized via the sharpness evaluation metrics. Finally, a lightweight convolutional neural network model for real-time computing at mobile terminals is constructed. The proposed method realizes the classification of surface textures of workpieces with different roughness values, and is experimentally verified on the end milling texture data set. Taking the times of multiplication and addition as the metrics, the performed experiments indicate that the number of floating-point operations (e.g., add and multiply) required for model inference is reduced by 55%, compared with the general convolutional neural network. In addition, the introduced cost-sensitive loss effectively improves the model’s stability to unbalanced data. Compared with the traditional machine learning methods, the accuracy of the proposed model is improved under the same experimental conditions (i.e., detection frame rate and image resolution), where the recall rate is increased by 21%, and the accuracy rate is enhanced by 8% simultaneously.
-
图 3 直方图均衡化前后纹理图像对比
Figure 3. Comparison of texture images before and after histogram equalization
图 6 代价敏感损失中置信度与损失值之间变化关系
Figure 6. Relationships between confidence and loss value in cost-sensitive loss
图 7 不同倍率、粗糙度下铣削纹理图样
Figure 7. Texture images under different microscopic magnifications and roughness levels
图 8 立铣加工监测实验的纹理图像采集与标注
Figure 8. Texture image collection and labeling in end milling monitoring experiment
图 10 不同清晰度指标的采集过程响应曲线
Figure 10. Responding curves of collection process with different clarity indexes
表 1 监测实验工况设置明细
Table 1. Details of monitoring experiment
工况 主轴转速/
（r·min−1）进给速度/
（mm·min−1）切削深度/mm 刀具
类型机床
型号1 3000 200 1 C1 KVC650 2 4000 200 1 C2 KVC650 3 2000 200 0.5 C2 VMC850 
下载: 导出CSV
表 2 可分离卷积结构的消融实验结果
Table 2. Ablation experiment for separable convolution
模型结构 推理速度/ms 准确率/% 普通卷积 33.7 77.6 可分离卷积（α=0.2） 25.2 75.7 可分离卷积 2（α=0.3） 25.1 72.1
下载: 导出CSV
表 3 普通交叉熵损失和代价敏感损失函数测试结果
Table 3. Test results of standard cross-entropy and cost-sensitive loss
% 损失函数 精准率 召回率 工况 1 工况 3 工况 1 工况 3 普通交叉熵 79.1 77.5 80.2 78.3 代价敏感损失 78.7 79.9 80.6 80.3
下载: 导出CSV
表 4 模型性能评估测试
Table 4. Model performance evaluation
方法 P/% R/% 计算时间/（s·帧−1） GLCM + SVM （RBF） 73.7 65.3 0.208 GLCM + SVM （linear） 73.2 65.5 0.208 GLCM + RF 73.4 65.1 0.208 提出方法 81.6 86.2 0.150
下载: 导出CSV
-
[1] RIFAI A P, FUKUDA R, AOYAMA H. Surface roughness estimation and chatter vibration identification using vision-based deep learning[J]. Journal of the Japan Society for Precision Engineering, 2019, 85(7): 658-666. doi: 10.2493/jjspe.85.658 [2] SHI W S, CAO J, ZHANG Q, et al. Edge computing: vision and challenges[J]. IEEE Internet of Things Journal, 2016, 3(5): 637-646. doi: 10.1109/JIOT.2016.2579198 [3] WANG B, LEI Y G, YAN T, et al. Recurrent convolutional neural network: a new framework for remaining useful life prediction of machinery[J]. Neurocomputing, 2020, 379: 117-129. doi: 10.1016/j.neucom.2019.10.064 [4] 高宏力,李登万,许明恒. 基于人工智能的丝杠寿命预测技术[J]. 西南交通大学学报,2010,45(5): 685-691. doi: 10.3969/j.issn.0258-2724.2010.05.006GAO Hongli, LI Dengwan, XU Mingheng. Intelligent monitoring system for screw life evaluation[J]. Journal of Southwest Jiaotong University, 2010, 45(5): 685-691. doi: 10.3969/j.issn.0258-2724.2010.05.006 [5] ZHANG L, GUO L, GAO H L, et al. Instance-based ensemble deep transfer learning network: a new intelligent degradation recognition method and its application on ball screw[J]. Mechanical Systems and Signal Processing, 2020, 140: 106681.1-106681.14. [6] PAPANDREA P J, FRIGIERI E P, MAIA P R, et al. Surface roughness diagnosis in hard turning using acoustic signals and support vector machine: a PCA-based approach[J]. Applied Acoustics, 2020, 159: 107102.1-107102.9. [7] PLAZA E G, NÚÑEZ LÓPEZ P J. Surface roughness monitoring by singular spectrum analysis of vibration signals[J]. Mechanical Systems and Signal Processing, 2017, 84: 516-530. doi: 10.1016/j.ymssp.2016.06.039 [8] TANGJITSITCHAROEN S, THESNIYOM P, RATANAKUAKANGWAN S. Prediction of surface roughness in ball-end milling process by utilizing dynamic cutting force ratio[J]. Journal of Intelligent Manufacturing, 2017, 28(1): 13-21. doi: 10.1007/s10845-014-0958-8 [9] RIFAI A P, AOYAMA H, THO N H, et al. Evaluation of turned and milled surfaces roughness using convolutional neural network[J]. Measurement, 2020, 161: 107860.1-107860.11. [10] KRIZHEVSKY A, SUTSKEVER I, HINTON G E. Imagenet classification with deep convolutional neural networks[J]. Communications of the ACM, 2017, 60(6): 84-90. [11] HE K M, ZHANG X Y, REN S Q, et al. Deep residual learning for image recognition[C]//2016 IEEE Conference on Computer Vision and Pattern Recognition. Las Vegas: IEEE, 2016: 770-778. [12] 夏毅敏,李清友,邓朝辉,等. 基于轻量级模型的隧道岩性快速识别方法[J]. 西南交通大学学报,2021,56(2): 420-427.XIA Yimin, LI Qingyou, DENG Chaohui, et al. Rapid identification method for lithology of tunnel based on lightweight model[J]. Journal of Southwest Jiaotong University, 2021, 56(2): 420-427. [13] BHAT N N, DUTTA S, PAL S K, et al. Tool condition classification in turning process using hidden Markov model based on texture analysis of machined surface images[J]. Measurement, 2016, 90: 500-509. doi: 10.1016/j.measurement.2016.05.022 [14] DATTA A, DUTTA S, PAL S K, et al. Progressive cutting tool wear detection from machined surface images using Voronoi tessellation method[J]. Journal of Materials Processing Technology, 2013, 213(12): 2339-2349. doi: 10.1016/j.jmatprotec.2013.07.008 [15] CHANG S I, RAVATHUR J S. Computer vision based non-contact surface roughness assessment using wavelet transform and response surface methodology[J]. Quality Engineering, 2005, 17(3): 435-451. doi: 10.1081/QEN-200059881 [16] DUTTA S, PAL S K, SEN R. Progressive tool condition monitoring of end milling from machined surface images[J]. Proceedings of the Institution of Mechanical Engineers, Part B: Journal of Engineering Manufacture, 2018, 232(2): 251-266. doi: 10.1177/0954405416640417 [17] RIFAI A P, FUKUDA R, AOYAMA H. Image based identification of cutting tools in turning-milling machines[J]. Journal of the Japan Society for Precision Engineering, 2019, 85(2): 159-166. doi: 10.2493/jjspe.85.159 [18] LIU Y K, GUO L, GAO H L, et al. Machine vision based condition monitoring and fault diagnosis of machine tools using information from machined surface texture: a review[J]. Mechanical Systems and Signal Processing, 2022, 164: 108068.1-108068.30. [19] NATHAN D, THANIGAIYARASU G, VANI K. Study on the relationship between surface roughness of AA6061 alloy end milling and image texture features of milled surface[J]. Procedia Engineering, 2014, 97: 150-157. doi: 10.1016/j.proeng.2014.12.236 -
下载:
点击查看大图
计量
- 文章访问数: 422
- HTML全文浏览量: 121
- PDF下载量: 71
- 被引次数: 0