Chatter Stability Prediction and Process Parameters’ Optimization of Milling Considering Uncertain Tool Information

: Stability is the prerequisite of a milling operation, and it seriously depends on machining parameters and machine tool dynamics. Considering that the tool information, including the tool clamping length, feeding direction, and spatial position, has signiﬁcant effects on machine tool dynamics, this paper presents an efﬁcient method to predict the tool information dependent-milling stability. A generalized regression neural network (GRNN) is established to predict the limiting axial cutting depth, where the machining parameters and tool information are taken as input variables. Moreover, an optimization model is proposed based on the machining parameters and tool information to maximize the material removal rate (MRR), where the GRNN model is taken as the stability constraint. A particle swarm optimization (PSO) algorithm is introduced to solve the optimization model and provide an optimal conﬁguration of the machining parameters and tool information. A case study has been developed to train a GRNN model and establish an optimization model of a real machine tool. Then, effects of the tool information on milling stability were discussed, and an origin-symmetric phenomenon was observed as the feeding direction varied. The accuracy of the solved optimal process parameters corresponding to the maximum MRR was validated through a milling test.


Introduction
Regenerative chatter in the milling process is a kind of self-induced vibration caused by time-delayed displacement feedback in the tool-workpiece system. The negative effects induced by the chatter occurrence include serious noise, poor surface quality, increased tool wear, machine tool failure, and so on, seriously limiting the machining quality and efficiency of the milling process [1,2]. Therefore, milling stability has been deeply studied for decades to avoid and suppress the regenerative chatter vibration [3][4][5]. Generally, the stability lobe diagram (SLD) described by different combinations of cutting depth and spindle speed is utilized to select appropriate chatter-free milling parameters to avoid chatter. And the SLD can be usually obtained by performing chatter tests or solving the analytical chatter stability model [6,7].
Quintana [8] performed a large number of milling tests with varying combinations of spindle speed and depth of cut and utilized the measured sound information to plot the SLD. To decrease the number of milling tests, Grossi et al. [9] proposed a spindle ramp-up test to efficiently identify a high number of chatter frequencies and limit axial cutting depths. For theoretical milling stability analysis, the generally used methods are the zero-order approximation (ZOA) method, semi-discretization method (SDM), and full-discretization method (FDM) [10]. Altintas and Budak [11] represented the milling tool-workpiece system as a two-degree freedom system and proposed the ZOA method to compute the SLD. Since the ZOA method needs no iteration and is efficient in predicting the SLD, it has been widely used in milling stability prediction. Insperger et al. [12] initially proposed the SDM in predicting the milling stability, and some improved SDMs were further presented to increase the prediction efficiency. Ding et al. [13] presented the FDM to analyze the responses of milling dynamics, and other researchers introduced different interpolation algorithms to improve the FDM.
Utilizing the aforementioned analytical methods to predict the milling stability needs to obtain the tool point dynamics in advance. In recent years, more and more researchers focusing on the milling stability prediction with uncertain tool point frequency response functions (FRFs) caused by the change in tool-holder combination or tool spatial position. The response coupling substructure analysis (RCSA) method was first proposed by Schmitz, and the improved RCSA methods have been widely used to efficiently predict the tool point FRFs when the tool-holder assembly changed [14]. Law et al. [15] proposed a reduced model substructural synthesis method to efficiently establish a position-dependent multibody dynamic of a machine tool, benefiting rapid predictions of tool point FRFs and milling stability within the whole machine tool working space. Deng et al. [16] establish a backpropagation neural network model to predict the position and feed direction-dependent tool point FRFs, and then study the effects of tool position and feed direction angle on the milling stability.
The groundbreaking research on milling stability provides multiple means of stable process planning in a real machining environment. However, few researchers have considered the effects of the varying tool clamping length on tool point FRFs and thus on the milling stability, and repeated impact tests are still needed to obtain the tool point FRF and identify the tool-holder joint dynamic parameters as the clamping length changed. Smith et al. [17] established the FEM of the high-speed spindle system and studied the effect of the tool length on achievable stable material removal rate (MRR). Duncan [18] put forward that the tool clamping length was an important parameter to improve cutting system dynamics and MRR. Yan et al. [19] adopted uniform distributed spring and damping elements to simulate the holder-tool joint dynamics and identified the spring and damping coefficients per unit contact area to efficiently compute the joint dynamics and tool point FRFs with varying clamping length. The varying tool clamping length can modify the stable zone of the SLD by affecting the tool dynamics, introducing uncertainties in selecting chatter-free machining parameters. Since machining parameters are the dominant factors affecting the machining quality and efficiency, machining parameters' optimization considering the chatter stability constraint has been researched as the development of the optimization algorithms [20][21][22]. Nevertheless, these optimizations are mainly conducted at a given tool clamping length, spatial position, and feeding direction and ignore the effects of the varying tool information on chatter stability constraint, limiting the selection range of optimal machining parameters for obtaining higher productivity.
Therefore, the purpose of this study is to develop a method to predict the milling stability and obtain optimal process parameters considering the uncertain tool information without repeated impact tests. Initially, typical combinations of tool information including clamping length, feeding direction, and spatial position coordinates are determined by orthogonal experiment method to measure the tool point FRFs through impact testing; then, these tool point FRFs are used to calculate the limiting axial cutting depth (a plim ) under different combinations of machining parameters; on the basis, a generalized regression neural network (GRNN) model whose inputs are the machining parameters and tool information can be trained to predict the a plim ; furthermore, an optimization model for improving the machining efficiency is established by taking the GRNN model to represent the chatter stability constraint, and it can be solved to obtain the optimal combination of machining parameters and tool information. Henceforth, the following contents of this paper are organized as follows. The principles of predicting the tool information dependent-milling stability are given in Section 2. The detailed information and procedures to establish a process parameters optimization model are provided in Section 3. A case study on the milling process of a real machine tool is described in Section 4. Finally, conclusions from the current research are presented in Section 5.

Theoretical Analysis of Milling Chatter Stability
The milling system is generally represented by a two degree of freedom (2-DOF) dynamic model shown in Figure 1a, where the flexible tool is assumed to have two orthogonal degrees of freedom in x and y directions, and the corresponding general equations of motion are expressed in Equation (1).
where m, c, and k mean the mass, damping, and stiffness, respectively, and the cutting force F x and F y can be obtained by Equation (2): a xx a xy a xy a yy where a p is the axial depth of cut, N t is the tool teeth number, K tc is the tangential cutting force coefficient, and [A] is a matrix containing the average directional coefficients, which can be obtained according to the theory proposed by Altintas and Budak [23]: where K rt is the ratio of the radial to tangential cutting force coefficient, and ϕ st and ϕ ex are the start angle and exit angle. Analyzing the dynamics of the milling system in the frequency domain, the dynamic forces can be rewritten in Equation (4): where G(iω c ) is the transfer function matrix which is usually represented by the tool point FRFs in x and y directions as the tool system is more flexible than the workpiece system. Then, the following characteristic equation is derived from Equation (4) to yield the limiting axial cutting depth a plim . where Λ is the eigenvalue and ω c is the chatter frequency. Ultimately, the a plim and corresponding spindle speed n can be obtained: According to Equations (1)-(6), the milling stability depends on the transfer function matrix G(iω) containing the tool point FRFs. Thus, the effects of varying tool information on tool point FRFs can further act on the milling stability. Furthermore, the SLDs describing the relationship between the aplim and n are plotted in Figure 1b, where the red and blue curves represent the SLDs with the tool champing length 60 mm and 50 mm, respectively. Obvious differences between the two SLDs can be observed, indicating the significance of considering the effects of tool clamping length on milling stability prediction and optimization.

The GRNN Model in Predicting the Tool Clamping Depth-Dependent Milling Stability
Since the tool clamping length, feeding direction, and spatial position can vary in continuous intervals, it is time-consuming and impossible to perform the impact testing at the tool point with each different combination of tool information. Furthermore, the dominant mode number is also uncertain when the tool information changed, adding numerous difficulties in predicting the tool information-dependent modal parameters for recognizing the tool point FRFs and further performing the milling stability analysis. As a consequence, a mathematical model needs to be established to predict the tool information dependent-axial limiting cutting depth aplim conveniently.
As the generalized regression neural network (GRNN) has great advantages in learning speed and approximation ability compared with the backpropagation neural network (BPNN) and radial basis function neural network (RBFNN), it is first adopted in this research to establish the mathematical mapping relationship between the tool information and axial limiting cutting depth aplim [24,25]. The typical topologic structure of a GRNN According to Equations (1)-(6), the milling stability depends on the transfer function matrix G(iω) containing the tool point FRFs. Thus, the effects of varying tool information on tool point FRFs can further act on the milling stability. Furthermore, the SLDs describing the relationship between the a plim and n are plotted in Figure 1b, where the red and blue curves represent the SLDs with the tool champing length 60 mm and 50 mm, respectively. Obvious differences between the two SLDs can be observed, indicating the significance of considering the effects of tool clamping length on milling stability prediction and optimization.

The GRNN Model in Predicting the Tool Clamping Depth-Dependent Milling Stability
Since the tool clamping length, feeding direction, and spatial position can vary in continuous intervals, it is time-consuming and impossible to perform the impact testing at the tool point with each different combination of tool information. Furthermore, the dominant mode number is also uncertain when the tool information changed, adding numerous difficulties in predicting the tool information-dependent modal parameters for recognizing the tool point FRFs and further performing the milling stability analysis. As a consequence, a mathematical model needs to be established to predict the tool information dependent-axial limiting cutting depth a plim conveniently.
As the generalized regression neural network (GRNN) has great advantages in learning speed and approximation ability compared with the backpropagation neural network (BPNN) and radial basis function neural network (RBFNN), it is first adopted in this research to establish the mathematical mapping relationship between the tool information and axial limiting cutting depth a plim [24,25]. The typical topologic structure of a GRNN model is described in Figure 2, which was composed of four layers of neurons, including the input layer, pattern layer, summation layer, and output layer [26,27]. The input variable vector X = [X 1 , X 2 , . . . , X p ] lies in the input layer, and its dimension p determined the number of neurons. The input vector was processed by linear functions and then transformed into the pattern layer. The neuron number of the pattern layer equaled the number of training samples, and its transfer function was the radial basis function as expressed in Equation (7).
Symmetry 2021, 13, 1071 where P i is the output of the i th neuron, X i is the training sample corresponding to the i th neuron, and the smoothing factor σ is the width coefficient of the Gaussian function.
Symmetry 2021, 13, x FOR PEER REVIEW 6 of 15 axial cutting depth aplim was taken as the output to predict the tool information dependentmilling stability. On this basis, a GRNN model can be trained according to Equations (7)- (11) to predict the tool information dependent-limiting axial cutting depth.

Milling Process Parameters Optimization
Achieving higher productivity with ideal product quality is an important issue that has been focused on in the manufacturing industry. Generally, the machining parameters are taken as the variables to establish an optimization model which does not consider the effects of tool information, limiting the selection region of an optimal combination of spindle speed and axial cutting depth. Thus, an optimization model of a milling process considering the effects of uncertain tool information is provided in this section.

Variables
The milling efficiency and product quality are highly dependent on the machining parameters and tool-workpiece system dynamics. The machining parameters of a milling process often contain the spindle speed n, axial depth of cut ap, radial cutting width ae, and feed rate per tooth fz. Moreover, the tool-workpiece system dynamics mainly depend on the tool point FRFs which are affected by the tool clamping length lc, feeding direction angle θ, and spatial position coordinates sx, sy, and sz. Therefore, n, ap, ae, fz, lc, θ, sx, sy, and sz are taken as the variables.

Objective Functions
Material removal rate (MRR) stands for the martial volume removed by the tool per unit time in the machining process, and a higher MRR is desired by the industry to improve productivity [28,29]. Then, the MRR is defined as the objective, and it can be described by the machining parameters and tool teeth number Nt shown in Equation (12): Equation (12) indicates that the MRR is linear to the machining parameters. However, increasing the machining parameters tends to scarify the product quality and cause chatter vibration. For instance, a higher axial depth of cut can improve the MRR, but a chatter vibration may occur and affect the machined surface quality. Therefore, the optimization of MRR should be under the following multiple constraints in Section 3.3. The summation layer contained two kinds of neurons marked as S d and S k , respectively. The first kind S d is the sum of the output of each neuron in the pattern layer, and the second kind S k is the weighted sum of the output of each neuron in the pattern layer. S d and S k can be computed through Equation (8).
where S kj means the output of the j th neuron belonging to the second kind S k , and y ij is the expected output of the j th element corresponding to the i th training sample.
The neuron number of output layer equaled the dimension of the output vector Y = [Y 1 , Y 2 , . . . , Y m ], and the output of the j th neuron in the output layer is described in Equation (9): It can be seen from Equations (7)-(9) that the hyper-parameter in the GRNN model was the only smoothing factor σ. When σ had a large value, the result of Equation (9) approximated to the mean value of the training sample outputs, reducing the accuracy of the GRNN model. Conversely, a smaller σ made the output of the trained GRNN model more approximate to the real output of the training sample, but an overfitting phenomenon could be observed. Generally, the value of σ was between 0 and 1, and a linear equation was used to find the optimal σ value as shown in Equation (10).
where t is the iteration number, ∆σ is a constant increment, and σ 0 is an initial value of the smoothing factor, which can be determined by some trial computations. The root mean square error (RMSE) expressed in Equation (11) was used to evaluate the accuracy of the established GRNN model. When the RMSE met the required value, the iteration terminated and the corresponding σ was finally selected to establish the GRNN model. where N o is the number of training samples, and y jr and y j represent the real and predicted values of the output variable, respectively.
The machining parameters and tool information, including the spindle speed n, radial cutting width a e , feed rate per tooth f z , tool clamping depth l c , feeding direction angle θ, and spatial coordinates s x , s y , and s z were taken as the input variables, and the limiting axial cutting depth a plim was taken as the output to predict the tool information dependent-milling stability. On this basis, a GRNN model can be trained according to Equations (7)-(11) to predict the tool information dependent-limiting axial cutting depth.

Milling Process Parameters Optimization
Achieving higher productivity with ideal product quality is an important issue that has been focused on in the manufacturing industry. Generally, the machining parameters are taken as the variables to establish an optimization model which does not consider the effects of tool information, limiting the selection region of an optimal combination of spindle speed and axial cutting depth. Thus, an optimization model of a milling process considering the effects of uncertain tool information is provided in this section.

Variables
The milling efficiency and product quality are highly dependent on the machining parameters and tool-workpiece system dynamics. The machining parameters of a milling process often contain the spindle speed n, axial depth of cut a p , radial cutting width a e , and feed rate per tooth f z . Moreover, the tool-workpiece system dynamics mainly depend on the tool point FRFs which are affected by the tool clamping length l c , feeding direction angle θ, and spatial position coordinates s x , s y , and s z . Therefore, n, a p , a e , f z , l c , θ, s x , s y , and s z are taken as the variables.

Objective Functions
Material removal rate (MRR) stands for the martial volume removed by the tool per unit time in the machining process, and a higher MRR is desired by the industry to improve productivity [28,29]. Then, the MRR is defined as the objective, and it can be described by the machining parameters and tool teeth number N t shown in Equation (12): Equation (12) indicates that the MRR is linear to the machining parameters. However, increasing the machining parameters tends to scarify the product quality and cause chatter vibration. For instance, a higher axial depth of cut can improve the MRR, but a chatter vibration may occur and affect the machined surface quality. Therefore, the optimization of MRR should be under the following multiple constraints in Section 3.3.

Milling Stability Constraint
As the milling process should be developed under a chatter-free condition, the axial limiting cutting depth predicted in Section 2.2 is used to describe the milling stability constraint. a p < a plim = f (n, a e , f z , l c , θ, s x , s y , s z ) (13) where f (·) stands for the GRNN mathematical model established in Section 2.2.

Power Constraint
The power required by the milling process P c should not exceed the power output of the machine tool [30]. where D t is the tool diameter, η is the machine tool efficiency, and K F , C F , x F , y F , z F , u f , v f are the corresponding force coefficients.

Surface Roughness Constraint
The surface roughness R a is an index reflecting the surface quality, and it should meet the technical criteria and requirements of customers. The following empirical formula is commonly used to predict the R a whose unit is µm, and the specific value of f z under the unit of mm/z is substituted in Equation (15) [31,32].
where θ fa and θ ba are the tool rake angle and relief angle, respectively.

Tool Life Constraint
The tool life is an important constraint considering an economical machining process, and it should not be shorter than the required time proposed by the technologists. In practice, the model proposed by Nefedov and Osipov is often introduced to predict the tool life T life [30,33].
where T min is the required tool life defined by users and K v , C v , a, d, e, g, w, and q are constant coefficients.

Optimization Model
On the basis of the defined variables, objectives, and constraints, an optimization model for selecting optimal milling process parameters is formed as follows.
Maxmize : f (V) = max(MRR) V = n a p a e f z l c θ s x s y s z T s.t.
where V min and V max are formed by the minimum and maximum values of each variable, respectively. Particle swarm optimization (PSO) is an algorithm proposed based on the theory of swarm intelligence, which can efficiently search the optimal global particle. Since the PSO has the advantages of simplicity, parallel computing, and fast convergence speed, it has been widely used in the parameter optimization field [34,35]. Therefore, this paper introduced the PSO algorithm to solve the optimization model in Equation (17) to obtain the optimal process parameters. The PSO algorithm is developed through cooperation and competition among the particles. Each individual particle is presented by a D-dimensional position vector (x i1 , x i2 , . . . , x id , . . . , x iD ) which can be updated with the speed vector calculated according to the following algorithm.
where v k id and x k id represent the speed and position of the i th particle, ω is the inertia weight, c 1 and c 2 are the learning factors, r 1 and r 2 are the random numbers between 0 and 1, and P best and G best are the individual and global optimal historical values determined by comparing their fitness values calculated from the fitness function. Generally, ω can be updated using the following Equation (19).
where the ω min and ω max mean the upper and lower limits of the inertial weight.

The Milling Stability Prediction by Establishing a GRNN Model
The axial limiting cutting depth a plim is a significant index to represent the milling stability, and it can be predicted by training a GRNN model under the input variables of machining parameters and tool information. The variation ranges of the input variables n, a e , f z , l c , θ, s x , s y , and s z are listed in Table 1. The sample information of the input and output variables to establish the GRNN model was determined by the following steps.

1.
Step 1: Determine typical combinations of tool information by an orthogonal experiment design method. First, the tool clamping length l c , feeding direction angle θ, and spatial position coordinates s x , s y , and s z are taken as the factors, and eight levels of each factor are determined within its variation interval and listed in Table 2. Then, an orthogonal table shown in Table 3 is used to determine 64 typical combinations of l c , θ, s x , s y , and s z . On this basis, the impact testing has been performed at the tool point using the testing instruments in Figure 3a to obtain the corresponding FRFs under each specific combination of l c , θ, s x , s y , and s z . The tool point FRFs for three different tool clamping lengths are described in Figure 3b, where the dominant mode trends to shit from the higher order to the lower one as the l c increase. This phenomenon may account for the tool point FRF being dominated by the tool mode when l c has a smaller value. On the contrary, the tool point FRF is dominated by the spindle mode when l c has a bigger value.  1  50  0  70  50  15  2  56  25  140  100  50  3  62  50  210  150  100  4  68  75  280  200  150  5  74  100  350  250  200  6  80  125  420  300  250  7  86  150  490  350  300  8  90  175  540 385 335  2.
Step 2: Determine sample information of the GRNN model. For input machining parameters n, a e , and f z , 15 spindle speed values were selected from its interval by an increment of 500 rpm, 8 radial cutting depth values were selected from its interval by an increment of 2 mm, and 10 feed rate per tooth values were selected from its interval By an increment of 0.04 mm/z. Then combining the 64 schemes of l c , θ, s x , s y , and s z shown in Tables 2 and 3, 15 × 8 × 10 × 64 = 76,800 combinations of n, a e , f z , l c , θ, s x , s y , and s z were finally determined. At each combination, the computation for obtaining the related a plim value of a down milling process was developed based on Equations (1)-(6). 3.
Step 3: Obtain basic structural parameters of a GRNN model. Ninety percent of the 76,800 combinations of n, a e , f z , l c , θ, s x , s y , s z , and a plim were randomly selected as the training samples, and 10 percent were taken as the testing samples. Then, six values of the smoothing factor σ, including 0.005, 0.01, 0.05, 0.1, 0.5, and 1 were initially determined to perform some trial computations. A computer with 16 G RAM and a 2.6 GHz Intel i7 processor was used to perform the computation in the MATLAB environment, and one complete computation under a specific σ value needed 20.4 s. A smaller RMSE was observed when the smoothing factor σ = 0.05. Thus, σ 0 and ∆σ in Equation (10) were determined as 0.05 and 0.0001; 4.
Step 4: Train and validate the GRNN model for predicting a plim . Specific values of the 76,800 combinations containing the n, a e , f z , l c , θ, s x , s y , and s z were first normalized. Then, the GRNN model was trained iteratively by modifying σ by an increment of 0.0001. The iteration terminated when the RMSE calculated by Equation (11) was 0.0066, which first met the required 0.01, and the corresponding value of the smoothing factor σ was 0.034. Then, the inputs of the testing samples were used to predict the axial limiting cutting depths by the trained GRNN model. The error percentages between the real and predicted a plim values are shown in Figure 4, and the maximum error percentage lower than 0.6% verifies the accuracy of the established GRNN model. Furthermore, the original and predicted lobe diagrams are also shown in Figure 4, which are plotted under the given tool information a e = 12 mm, f z = 0.08 mm/z, l c = 60 mm, θ = 30 • , s x = 300 mm, s y = 200 mm, and s z = 150 mm. The two lobes have good consistency, which shows that the GRNN model has feasibility in predicting the milling stability.
With the validated GRNN model, the milling stability for different combinations of machining parameters and tool information can be predicted efficiently. Figure 5a,b describes the limiting axial cutting depth varies with the tool clamping depth and feeding direction angle at different tool spatial positions, respectively. For each figure, an originsymmetric phenomenon can be observed as the feeding direction angle changes. Furthermore, the aplim did not decrease monotonously with the decrease in lc, since the tool point FRFs dominated from the spindle mode to tool mode, and the change in natural frequency caused the shift of the lobe.  With the validated GRNN model, the milling stability for different combinations of machining parameters and tool information can be predicted efficiently. Figure 5a,b describes the limiting axial cutting depth varies with the tool clamping depth and feeding direction angle at different tool spatial positions, respectively. For each figure, an origin-symmetric phenomenon can be observed as the feeding direction angle changes. Furthermore, the a plim did not decrease monotonously with the decrease in l c , since the tool point FRFs dominated from the spindle mode to tool mode, and the change in natural frequency caused the shift of the lobe.

The Optimization of Improving the MRR
The optimization model was established based on the trained GRNN model for pre dicting the aplim, and the corresponding coefficients to calculate the power Pc, surfac roughness Ra, and tool life Tlife are listed in Table 1. The PSO algorithm was used to solv the optimization model under the basic information, including the population size 150 the maximum iteration number was 200, the learning factors c1 and c2 were both 1.5, an the initial weights ωmin and ωmax were 0.4 and 0.9, respectively. The position of each part cle had nine dimensions representing the variables n, ap, ae, fz, lc, θ, sx, sy, and sz, and th fitness function was expressed by Equation (12) for computing the optimization objectiv MRR. Then, the positions and velocities of the particles were randomly initialized an further updated based on Equation (18).
After 79 iterations, the value of MRR trended to a constant 1502.6 mm 3 /s shown i Figure 6a. The corresponding global optimal particle position and specific information o the constraints are listed in Table 4. Analyzing each value of the optimal solution, th maximum MRR was not the product of the maximum values of the n, ap, ae and fz, since was optimized under the constraints of stability, power, surface roughness, and tool life

The Optimization of Improving the MRR
The optimization model was established based on the trained GRNN model for predicting the a plim , and the corresponding coefficients to calculate the power P c , surface roughness R a , and tool life T life are listed in Table 1. The PSO algorithm was used to solve the optimization model under the basic information, including the population size 150, the maximum iteration number was 200, the learning factors c 1 and c 2 were both 1.5, and the initial weights ω min and ω max were 0.4 and 0.9, respectively. The position of each particle had nine dimensions representing the variables n, a p , a e , f z , l c , θ, s x , s y , and s z , and the fitness function was expressed by Equation (12) for computing the optimization objective MRR. Then, the positions and velocities of the particles were randomly initialized and further updated based on Equation (18).
After 79 iterations, the value of MRR trended to a constant 1502.6 mm 3 /s shown in Figure 6a. The corresponding global optimal particle position and specific information of the constraints are listed in Table 4. Analyzing each value of the optimal solution, the maximum MRR was not the product of the maximum values of the n, a p , a e and f z , since it was optimized under the constraints of stability, power, surface roughness, and tool life. A milling test was performed on the vertical machining center with the obtained optimal process parameters to validate the accuracy of the optimization. First, the moving components were driven to move according to the specific optimal values of s x , s y , and s z ; then, the tool was inserted into the holder, and the optimal clamping length l c was kept; finally, the ending milling was performed on the workpiece with the optimal machining parameters n, a p , a e , f z along the optimal feeding direction determined by the value of θ. The force signal of the milling process was measured and transformed into a frequency domain. In Figure 6b, the dominant natural frequencies were the tool passing frequency 331.6 Hz and its harmonics, showing that the milling process was under a stable condition. And the measured surface roughness of the workpiece was 4.36 µm, which is close to the predicted value 4.30 µm and meets the required 6.40 µm of the milling process. Accordingly, the obtained optimal configuration of process parameters can correspond to an efficient and stable milling operation under the requirements of stability, power, surface roughness, and tool life.
x FOR PEER REVIEW 13 of 15

Conclusions
Stability is the prerequisite of a milling operation, and it is significantly dependent on the tool point dynamics. Since the tool point FRFs are affected by uncertain tool information, including the tool clamping length, feeding direction, and spatial position, this paper proposed a method to predict the tool information dependent-milling stability and obtain optimal process parameters for an efficient milling operation based on the milling stability theory, GRNN model and PSO algorithm. A case study was developed on a real vertical machining center to validate the feasibility of the proposed method.
1. The tool clamping length lc, feeding direction angle θ, and spatial position coordinates sx, sy, and sz were taken as variables to design an orthogonal table with 64 schemes. Then the impact testing was performed at the tool point of each scheme to obtain the tool point FRFs, which showed obvious differences among the dominant natural frequencies and related amplitudes. In addition, the tool point FRFs for different clamping lengths showed that the dominant modes shifted from the tool mode

Conclusions
Stability is the prerequisite of a milling operation, and it is significantly dependent on the tool point dynamics. Since the tool point FRFs are affected by uncertain tool information, including the tool clamping length, feeding direction, and spatial position, this paper proposed a method to predict the tool information dependent-milling stability and obtain optimal process parameters for an efficient milling operation based on the milling stability theory, GRNN model and PSO algorithm. A case study was developed on a real vertical machining center to validate the feasibility of the proposed method.

1.
The tool clamping length l c , feeding direction angle θ, and spatial position coordinates s x , s y , and s z were taken as variables to design an orthogonal table with 64 schemes. Then the impact testing was performed at the tool point of each scheme to obtain the tool point FRFs, which showed obvious differences among the dominant natural frequencies and related amplitudes. In addition, the tool point FRFs for different clamping lengths showed that the dominant modes shifted from the tool mode to spindle mode as the tool clamping length increased. Furthermore, typical values of each machining parameter, such as the spindle speed n, radial cutting width a e , and feed rate per tooth f z were determined to form different combinations of machining