Abstract: To overcome the poor generalization and low accuracy of the traditional surface roughness prediction model of CNC milling, a novel surface roughness prediction method of CNC milling was proposed based on multi-source heterogeneous data. Firstly, the static data such as processing parameters, tool diameter and workpiece material and dynamic data such as vibration signals, force signals and power signals were collected in CNC milling with variable technologies. Then, particle swarm optimization(PSO) algorithm was used to optimize the network structure parameters of CNN for obtaining PSO-CNN, which might adaptively extract the features of dynamic data. Features of static data were manually extracted. A shallow neural network was carried out to fuse the features of multi-source heterogeneous data such as dynamic data and static data, which might be used to build surface roughness prediction model of CNC milling with variable technologies. Finally, the superiority of the proposed method was demonstrated according to the performance comparison tests with different surface roughness prediction models. And, the validity of the proposed method was verified by the example of two workpiece machining.

Key words: surface roughness prediction, CNC milling, multi-source heterogeneous data, convolutional neural network（CNN）

摘要： 针对传统数控铣削表面粗糙度预测模型泛化性差、精度较低等问题，提出了一种基于多源异构数据的数控铣削表面粗糙度预测方法。获取变工艺条件下数控铣削的工艺参数、刀具直径及工件材料等静态数据和振动信号、力信号及功率信号等动态数据；采用粒子群优化算法(PSO)优化卷积神经网络(CNN)的网络结构参数得到PSO-CNN；运用PSO-CNN自适应提取动态数据特征并对静态数据特征进行人工提取，再通过浅层神经网络融合动、静态数据等多源异构数据的特征，建立变工艺下的表面粗糙度预测模型；通过不同模型的预测性能对比试验，验证了该方法的优越性，并以两个工件加工过程为例，验证了该方法的有效性。

关键词: 表面粗糙度预测, 数控铣削, 多源异构数据, 卷积神经网络