Abstract: iming at the problem that the existing milling vibration state identification model has low accuracy and long training time,a milling vibration state identification method based on improved MobileNetV2 was proposed.The MobileNetV2 backbone structure was used as the backbone feature extraction network,and the Multiscale Attention Fusion Layer (MAFL) and Layered Classifier (LC) were combined to reconstruct the top-level structure of MobileNetV2,so as to achieve the purpose of model improvement.Based on variational mode decomposition and Hilbert transform,the data preprocessing of milling vibration state was carried out,and the improved model was trained by combining Transfer Learning (TL) with Fine-tune.Furthermore,the improved MobileNetV2 model and a variety of classical classification models were used to identify and compare the milling vibration state with the variable cutting depth side milling process at different speeds.The results showed that the improved MobileNetV2 had advantages in accuracy and time consumption.The proposed identification method was more suitable for the application requirements of real-time cognition and chatter warning of cutting state in the field of manufacturing engineering,which had a broad engineering application prospect.

Key words: milling vibration, improvement, MobileNetV2, state identification

摘要： 针对现有铣削振动状态辨识模型准确率不高,训练耗时较长的问题,提出基于改进MobileNetV2的铣削振动状态辨识方法。以MobileNetV2骨干结构为主干特征提取网络,联合多尺度注意力聚融层(MAFL)与层递式分类器(LC)对MobileNetV2顶层结构进行重建,从而达到模型改进目的;其次,以变分模态分解与希尔伯特变换为基础开展铣削振动状态数据预处理,并以迁移学习(TL)与Fine-tune相结合对改进模型进行训练;进而,以不同转速下变切深侧铣工艺为对象,利用改进MobileNetV2模型及多种经典分类模型对铣削振动状态进行辨识与对比分析。结果表明,改进MobileNetV2在准确率和耗时方面均具有优势,所提辨识方法更适应制造工程领域对切削状态实时认知与颤振预警的应用需求,具有较广阔的工程应用前景。

关键词: 铣削振动, 改进, MobileNetV2, 状态辨识