基于MF-DFA特征和LS-SVM算法的刀具磨损状态识别

鉴于多重分形理论在精细刻画系统非线性现象和过程方面具有的独特优势,该文提出了基于多重分形去趋势波动分析和最小二乘支持向量机的刀具磨损状态识别方法。首先,用MF-DFA(multifractal detrended fluctuations analysis)方法处理去噪后的刀具磨损声发射信号,讨论其长程相关性和分形特性;然后,分析对比了不同磨损阶段下多重分形谱参数的变化,筛选出能灵敏表征刀具磨损状态的多重分形谱参数:分形维数最大值点对应的奇异指数α_0,多重分形谱谱宽△α和广义Hurst指数波动均值△h(q)作为特征量;最后,利用LS-SVM(least square support vector machine)算法并对比支持向量机和神经网络算法实现刀具磨损状态识别,结果表明LS-SVM算法识别率最高,平均识别准确率达97.78%,验证了本文所提方法的有效性。试验结果表明,提取的特征对刀具磨损状态的变化非常敏感,可以分离相近的磨损状态,为刀具状态监测提供一种参考方法。

Cutting tool wear recognition based on MF-DFA feature and LS-SVM algorithm

Abstract: Cutting is an important process in machining. In order to improve the automatic and intelligent level of machining and improve the production efficiency and quality, it is urgent to monitor the tool wear state. The feature extraction of wear state is the key to the tool wear monitoring. In view of the unique advantages of multifractal theory in accurately depicting the nonlinear phenomena and processes of the system, a tool wear state recognition method based on multifractal detrended fluctuation analysis (MF-DFA) and least squares support vector machine (LS-SVM) is proposed. The acoustic emission (AE) signal is denoised with wavelet packet analysis, and the best tree of wavelet packet decomposition is determined and reconstruction is performed based on the minimum Shannon criterion so as to achieve the purpose of signal initial denoising. Firstly, the MF-DFA method is used to deal with the noise emission signals of the tool wear after denoising, and the long range correlation and fractal characteristics are discussed. It shows that the tool wear time sequence is an orderly process with long range correlation, and the internal fluctuation is not random, and it has the ability to maintain the trend. Then, the multifractal spectrum parameters of different wear stages were analyzed and compared. The parameters of singular exponent corresponding to the point of extreme value and multifractal spectrum width are increasing with the progression of the wear stage, which indicates that the greater the wear amount, the greater the fluctuation of the AE signal, the more uneven the probability measurement of the whole fractal structure, the more random the fluctuation. The values of the AE signal multifractal dimension under different wear states are less than zero, and the multifractal spectrum is left hook like, indicating the number of the maximum subset in the probability measure is relatively large. The absolute value of the normal wear stage is the smallest, which indicates that the volatility is the smallest in this stage; the value of the parameter increases with the increase of the wear amount, indicating that the greater the fluctuation degree of generalized Hurst exponent, the stronger the multifractal characteristics. The singular exponent corresponding to the point of extreme value, the multifractal spectrum width and the mean of the generalized Hurst exponent, which can sensitively characterize the tool wear state, were selected as the characteristic quantities, and the three-dimensional feature vectors were constructed to characterize the tool wear stage. The clustering effect of the extracted tool wear state characteristics was obvious. The LS-SVM algorithm, SVM algorithm and BP (back propagation) neural network are applied to recognize the tool wear state. Simplex iterative algorithm is used to optimize the parameters, the optimal model is constructed to determine the performance of each group of parameters, and the parameters of regularization and kernel function are determined. The average recognition accuracy is 97.78%. The results show that the tool wear AE signal has long range correlation and obvious multifractal characteristics, the multifractal parameters, i.e. singular exponent corresponding to the point of extreme value, multifractal spectrum width and mean of the generalized Hurst exponent can be used as sensitive characterization for the feature of tool wear stage, and the tool wear stages can be clearly distinguished. The multifractal spectrum features extracted with the method based on MF-DFA and LS-SVM can identify the different wear stages of the tool well, verify the effectiveness of the recognition method, improve the accuracy of recognition, and lay a foundation for the realization of the wear prediction.