Abstract:  In order to predict the stiffness degradation degree of spindle/tool holder interfaces, a method was proposed based on excitation and response signal fusion. Firstly, side milling experiments of rectangular titanium alloy workpiece were carried out, instantaneous milling force signals and response vibration signals near the spindle-tool holder interfaces were collected, and a database reflecting the stiffness degradation of the spindle-tool holder interfaces was constructed. Then, according to the time-domain, frequency-domain and time-frequency domain features of the instantaneous milling forces and vibration signals in each direction in the database, three features, namely the time-domain mean value, frequency-domain center frequency and time-frequency first-order wavelet packet energy of the instantaneous milling force signals in the X direction and the vibration signals in the Z direction, were optimized based on correlation analysis. The low frequency filter convolution kernel and the high frequency filter convolution check after the preferred eigenmatrix were used for the dual channel convolution pooling processing respectively. The eigenvector of stiffness degradation degree of the deeply fused spindle-tool holder interfaces was obtained. Finally, the probabilistic mode of support vector machine(SVM) model was transformed into the conditional probability of naive Bayes model（NBC）, and the mixed classifier model NBC-SVM was constructed to improve the classification performance of the classifier. On the basis of the stiffness degradation database of the spindle-tool holder interfaces, the two-channel convolution pooled feature fusion method (CP-FF) and NBC-SVM model were used to predict the stiffness degradation degree of the spindle-tool holder interfaces, and the prediction accuracy is as 96%.

Key words: spindle-tool holder interface, stiffness degradation, feature fusion, naive Bayes classifier-support vector machine(NBC-SVM) model

摘要： 为了预测主轴/刀柄结合面刚度退化程度，提出了一种基于激励和响应信号融合的主轴/刀柄结合面刚度退化程度预测方法。首先进行钛合金矩形工件侧铣实验，采集瞬时铣削力信号和主轴/刀柄结合面附近的响应振动信号，构建反映主轴/刀柄结合面刚度退化的数据库。然后根据数据库中瞬时铣削力和振动信号各方向的时域、频域和时频域特征，基于相关性分析优选出瞬时铣削力信号和振动信号的时域均值、频域中心频率、时频域一阶小波包能量3个特征，分别使用低频滤波卷积核和高频滤波卷积核对优选后的特征矩阵进行双通道卷积池化处理，获取深度融合的主轴/刀柄结合面刚度退化程度特征向量。最后以支持向量机模型（SVM）的概率模式转化为朴素贝叶斯分类器（NBC）的条件概率，构建混合分类器模型（NBC-SVM），提高了分类器的分类性能。在主轴/刀柄结合面刚度退化数据库的基础上，基于双通道卷积池化的特征融合方法（CP-FF）和NBC-SVM模型实现了主轴/刀柄结合面刚度退化程度的预测，预测精度达96%。

关键词: 主轴/刀柄结合面, 刚度退化, 特征融合, 朴素贝叶斯分类器支持向量机模型