System of Detecting Abnormal Cutting using Neural Network

本研究的目的是开发一个利用频谱分析和神经网络的车削加工异常检测系统。该系统利用神经网络对刀具振动频谱进行模式学习和识别。1993年，进行了以下工作：（1）设计并建立了检测系统的设备;（2）提出并验证了利用谱分析检测再生颤振的方法;（3）提出并验证了利用谱分析和神经网络检测颤振的方法。切削试验以10 ~ 140 m/min的切削速度和1.0 ~ 4.0 mm的切削宽度进行。在切削速度不连续改变切削宽度的情况下，测量了工件的振动、动态切削力和刀具的振动。采用仅用频谱分析的方法，可以提高振动信号的检测成功率， ...更多信息 噪声小的信号检出率为100%，而噪声大的动态推力信号检出率为71%。另一方面，采用频谱分析和神经网络相结合的方法，对工件振动信号的检测率为88%，对动态推力信号的检测率为83%。实验结果表明，采用谱分析和神经网络相结合的方法，检测成功率可达80%以上，且检测成功率几乎不受噪声分量强度的影响。1994年，提出了一种利用光谱分析和神经网络相结合的方法来识别刀具后刀面磨损宽度的方法，并得到了验证。在实验中，使用四种人工磨损刀具，工件正交切削具有不同的后刀面磨损的宽度。结果表明，在神经网络学习的切削条件下，对刀具振动的识别成功率约为80%，而在未学习的切削条件下，识别成功率约为30%。因此，我们得出结论，目前的方法识别刀具后刀面磨损必须加以改进。少

The object of this study is to develop a system of detecting abnormal cutting for turning using spectral analysis and neural network. This system learns and recognizes patterns of the spectrum of cutting tool vibration using neural network. In 1993, the following things were carried out : (1) equipment of the detection system was designed and set up, (2) a method of detecting regenerative chatter using spectral analysis was presented and verified, (3) a method of detecting chatter vibrations using spectral analysis and neural network was presented and verified. The cutting tests were carried out at cutting speed of 10 to 140 m/min and 1.0 to 4.0 mm in width of cutting. And vibration of workpiece, dynamic cutting forces and vibration of cutting tool were measured at a cutting speed as changing the width of cutting discontinuously where self-excited chatter vibrations occurred or not. By the method using only spectral analysis, the rate of successful detection from vibration signal of wo … More rkpiece with little noise was 100%, but that from signal of dynamic thrust force with much noise was 71%. On the other hand, by the method using both spectral analysis and neural network, the rate of detection from vibration signal of workpiece was 88%, and that from dynamic thrust force was 83%. As a result, we concluded that the rate of successful detection by the method using both spectral analysis and neural network is about more than 80%, and this rate is hardly affected by the intensity of noise component. In 1994, a method of identifying width of tool flank wear using both spectral analysis and neural network was proposed and verified. On experiments using four kinds of artificial worn tools, workpieces were orthogonally cut by each tool having different flank wear in width. As a result, the rate of successful identification from vibration of cutting tool was about 80% under cutting conditions learned by the neural network, but that was about 30% under unlearned cutting conditions. Therefore we concluded that present method of identifying tool flank wear has to be improved. Less