基于轻量级卷积网络的铣削粗糙度在机监测研究

In-situ Roughness Evaluation of Milling Machined Surface Based on Lightweight Deep Convolutional Neural Network

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%.

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.