BTA深孔钻削钻头磨损状态卷积神经网络识别技术研究

发布时间：2018-03-30 01:08

本文选题：深孔钻削　切入点：主轴电机电流　出处：《西安理工大学》2017年硕士论文

【摘要】：在制造业智能化、信息化的背景下,通过在生产制造过程中采集与制造相关的信息数据,利用这些信息数据达到对生产过程进行监测,并在这些数据中得到生产过程出现的问题时对生产过程做出调整有着重要的作用。机械切削加工中,刀具作为直接与加工表面接触的部分,刀具也是整个加工系统中最薄弱的环节,要使自动化的加工过程高效稳定地进行,研究和开发加工过程中刀具状态监测技术就显得尤为重要。而现有的刀具状态监测技术大多针对车削、铣削加工,很少有针对深孔钻削的刀具状态监测方法。本文针对BTA深孔加工的特点,整个加工过程在封闭的环境中进行,刀具的磨损状态无法直接观察,由主轴电机电流与钻头磨损之间内在关系,直接从电机驱动器中采集了钻削过程中主轴电机电流信号作为监测信号,并介绍了一种将信号进行连续小波变换的信号分析方法与卷积神经网络的模式识别方法相融合的刀具状态监测方法。建立了基于深孔钻削数控机床的数控系统通讯模块的主轴电机电流信号采集系统,通过钻削实验获取了主轴电机电流信号,对采集得到的主轴电流信号进行初步分析,记录了钻削过程钻头磨损量,获取了钻头的磨损规律信息。结合钻头在不同磨损阶段的主轴电流信号特征以及信号时域分析和频域分析的不足,为了描述信号频率随时间的变化规律,获取钻头的磨损规律信息及变化特征,采用连续小波变换,得到不同磨损阶段的小波尺度谱。并在连续小波变换中,通过不同小波基函数的特征,确定了最佳小波基函数,利用小波信号熵的方法确定了最优小波分解层数。分析结果发现,钻头在不同磨损阶段的小波尺度谱表现出了很明显的不同,很好的反映了信号频率随时间变化规律。在信号小波尺度谱中高频成分随时间逐渐减小,而中频成分则在逐渐增加,很好的映射了钻头磨损规律。针对小波尺度谱在不同磨损阶段的明显不同,直接将小波尺度谱作为状态特征,省去了在模式识别前的前处理过程,结合卷积神经网络很好地识别图像的特性,特征提取和模式识别过程都在网络结构中完成。将采集到的信号的小波尺度谱,一部分作为训练集,一部分作为训练集,通过训练和测试结果确定了网络结构,包括:网络卷积层层数、网络卷积核大小以及网络卷积核个数。对训练完成的各层输出特征图可视化分析,发现图像经过各层卷积和采样之后,图像像素点减少,但是不同的卷积提取不同的图像特征,多个卷积核保证了图像信息的完整性,卷积神经网络在经过训练之后可以很好的提取图像特征。最后利用训练好的网络对一个全新钻头的完整钻削寿命周期中不同磨损状态进行识别,达到了很好的识别效果。
[Abstract]:Under the background of manufacturing intelligence and information, through collecting the information data related to manufacturing in the manufacturing process, the information data can be used to monitor the production process. It is very important to adjust the production process when we get the problems in the production process in these data. In machining, the tool is the part that is in direct contact with the machined surface. The tool is also the weakest link in the whole machining system, so the automatic machining process should be carried out efficiently and stably. The research and development of tool condition monitoring technology is particularly important in the process of machining, and most of the existing tool condition monitoring technology is aimed at turning and milling. There are few tool condition monitoring methods for deep hole drilling. According to the characteristics of BTA deep hole machining, the whole machining process is carried out in a closed environment, and the tool wear state can not be observed directly. According to the inherent relationship between spindle motor current and bit wear, the signal of spindle motor current in drilling process is collected directly from the motor driver as the monitoring signal. This paper also introduces a tool condition monitoring method which combines the signal analysis method of continuous wavelet transform and the pattern recognition method of convolution neural network. The NC system based on deep hole drilling NC machine tool is established. The main shaft motor current signal acquisition system based on signal module, The spindle motor current signal was obtained by drilling experiment, and the spindle current signal was preliminarily analyzed, and the bit wear amount during drilling process was recorded. According to the characteristics of spindle current signal in different wear stages of bit and the deficiency of signal analysis in time domain and frequency domain, in order to describe the variation law of signal frequency with time, the information of wear law of bit is obtained. The wavelet scale spectrum of different wear stages is obtained by means of continuous wavelet transform, and the best wavelet basis function is determined by the characteristics of different wavelet basis functions in continuous wavelet transform. Wavelet signal entropy is used to determine the optimal number of wavelet decomposition layers. The results show that the wavelet scale spectrum of bit is very different in different wear stages. In the wavelet scale spectrum of signal, the high frequency component decreases gradually with time, while the intermediate frequency component increases gradually. According to the obvious difference of wavelet scale spectrum in different wear stages, wavelet scale spectrum is taken as the state feature directly, and the pre-processing process before pattern recognition is eliminated. The feature extraction and pattern recognition process are completed in the network structure. The wavelet scale spectrum of the collected signal is regarded as the training set and the other part as the training set, and the wavelet scale spectrum of the acquired signal is regarded as the training set, and the wavelet scale spectrum of the acquired signal is regarded as the training set. The network structure is determined by training and testing results, including: network convolution layer number, network convolution kernel size and network convolution core number. It is found that the pixel points of the image decrease after each layer of convolution and sampling, but different convolution extracts different image features, and the integrity of image information is guaranteed by multiple convolution cores. After training, the convolution neural network can extract image features well. Finally, using the trained neural network to identify the different wear states in the whole drilling life cycle of a new bit, the recognition effect is very good.
【学位授予单位】：西安理工大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TG523;TP183

【参考文献】