Method for predicting chatter of machine tools

Through the combination of artificial neural network and stability model, and the training of evolution algorithms, the problem of limited accuracy of flutter prediction in the existing technology is solved, and high-precision flutter prediction in real production environment is achieved, which improves the quality of the workpiece and Tool lifespan.

CN111985148BActive Publication Date: 2024-11-01GF MACHINING SOLUTIONS AG +1
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202010435329.6
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Priority Date
2019-05-21
Filing Date
2020-05-21
Publication Date
2024-11-01
Estimated Expiration
2040-05-21

AI Technical Summary

Technical Problem

When predicting machine tool flutter, the accuracy of the prior art is affected by uncertain parameters and differences between the experimental environment and the production environment, making it difficult to obtain sufficient measurement data in the real production environment, affecting the quality of the workpiece and the life of the tool.

Method used

Using artificial neural network combined with stability model, train through the first input data and the second input data, adjust the weight to generate prediction data, and compare it with experimental data, adjust the neural network weight to improve the accuracy of flutter prediction, and utilize evolution Algorithms such as genetic algorithms are trained to ensure the accuracy of obtaining data in a real production environment.

Benefits of technology

The flutter prediction of a wide range of machine processing parameters is realized, the accuracy of prediction and stability of the prediction are improved, the accuracy of the leaf gallery diagram is reduced, the measurement cost is reduced, and the high-precision workpiece quality and tool integrity are ensured.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN111985148B_ABST
    Figure CN111985148B_ABST
Patent Text Reader

Abstract

The present invention relates to a method for predicting chatter of a machine tool. The method comprises the following steps: feeding first input data into an artificial neural network, the artificial neural network comprising a plurality of weights; determining first output data at the output of the artificial neural network based on the first input data and the plurality of weights; providing the first output data into a stability model to generate prediction data; comparing the prediction data with measured stability data, and adjusting the plurality of weights of the artificial neural network.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The invention relates to a method for predicting chatter of a machine tool and to a chatter prediction unit for carrying out the method. Background Art

[0002] Chatter is a self-excited vibration and has a crucial influence on the quality of the machined workpiece. If chatter occurs during the milling process, the quality of the workpiece can be significantly degraded, for example the surface quality of the workpiece is poor. In addition to this drawback, the life time of the cutting tool and / or machine components, such as the spindle, can be shortened by chatter. Chatter usually occurs due to a non-optimal combination between machine dynamics and machine processing parameters. Therefore, if chatter can be accurately predicted for different machine processing parameters, tools and workpieces, the optimal machine processing parameters can be pre-selected to reduce the risk of chatter and thereby improve the quality of the machined workpiece and ensure the integrity of the machine including the cutting tool.

[0003] Methods for predicting the occurrence of chatter are well established. However, in known methods, the accuracy of the prediction depends on several parameters; in particular, certain parameters cannot be measured or calculated accurately because they vary under different machine processing conditions. Furthermore, certain parameters can only be obtained in an experimental environment, which is different from a real production environment. Therefore, a system and method for acquiring measurement data in a real production environment to improve the accuracy of the prediction is proposed.

[0004] EP 2916187 discloses a chatter database system, in which data corresponding to machining and chatter conditions are obtained from a plurality of machine tools, in particular, data can be obtained while the machine tools are machining a workpiece. These data can be used to generate experimentally based stability diagrams, which illustrate the occurrence of chatter with respect to machining parameters. In this system, a large amount of data can be collected from machine tools in a real production environment. However, this system still has the disadvantage that not all data required for generating stability diagrams can be obtained in this way, such as the dynamics of the tool center point. Summary of the invention

[0005] It is an object of the present invention to provide a method and system for achieving improved accuracy of chatter prediction for a wide range of machine processing parameters.

[0006] According to the invention, these objects are achieved by the features of the independent claims.Further advantageous embodiments result from the dependent claims and the description.

[0007] In the present invention, a method for predicting chatter of a machine tool comprises the following steps: feeding first input data into an artificial neural network, the artificial neural network comprising a plurality of weights; determining first output data at the output of the artificial neural network based on the first input data and the plurality of weights; providing the first output data and second input data into a stability model to generate prediction data; and comparing the prediction data with measured stability data and adjusting the plurality of weights of the artificial neural network. The stability model input comprises second input data and first output data from the output of the artificial neural network.

[0008] Different models can be used to predict stability conditions. Examples are semi-discretization methods based on analytical zero-order solutions or in the time domain as well as full discretization methods. References related to different models and flutter are:

[0009] - Budak E, Altintas Y (1998) Analytical Prediction of ChatterStability in Milling—Part I: General Formulation . Journal of Dynamic Systems, Measurement, and Control , 120(1): p. 22-30.

[0010] - Insperger T, Stépán G (2002) Semi-discretization method for delayedsystems. International Journal for Nu-merical Methods in Engineering, 55(5):p. 503-518.

[0011] - Ye Ding, LiMin Zhu, XiaoJian Zhang, Han Ding, A full-discretization method for prediction of milling stability, International Journal of MachineTools and Manufacture, Volume 50, Issue 5, 2010, Pages 502-509,

[0012] - Chatter suppression techniques in metal cutting by J. Munoa, X. Beudaert, Z. Dombovari et al., CIRP Ann. Manuf. Technol. 65 (2) (2016) 785-808.

[0013] For computational reasons it is advantageous to use a zero-order solution model as stability model for calculating the stability limits. However the method of the invention is not limited to this stability model and any stability model may be used.

[0014] The stability lobe diagram is a well-known visual representation for distinguishing stable from unstable machine parameters in a milling operation. In the stability lobe diagram, the stability limit is depicted, for example in terms of the critical depth of cut as a function of the spindle speed.

[0015] One method for creating a stability lobe diagram is to establish a stability model and use the model to obtain the stability lobe diagram. This model-based approach requires information about structural parameters, such as the relative dynamics in the tool-workpiece contact area, and about process parameters, such as the cutting force coefficient that relates the chip thickness to the resulting force. In order to determine the dynamics of the machine tool and the cutting force coefficient, extensive measurements are mandatory. Such measurements incur additional costs. In addition, measurements cannot be made during production and therefore the measurement conditions do not correspond to the actual machine processing conditions. Since the machine dynamics can change significantly under actual machine processing conditions, the accuracy of the measurements is not satisfactory. Therefore, stability lobe diagrams established based on inaccurate data do not provide sufficient accuracy, particularly in the application of high-precision milling machine tools. The method of the present invention enables the data required for the stability lobe diagram to be obtained with higher accuracy.

[0016] For certain given machining conditions, stability models can be applied to predict the occurrence of chatter and the corresponding chatter frequency, which defines the frequency of vibration between the tool and the workpiece. Figure 3 The prediction data depicted in . The stability model requires several input parameters to make a flutter prediction. Some of these are reliably known, and some others may be uncertain. In the following, the uncertain parameters are referred to as first output data, and the known parameters are referred to as second input data. At the beginning, it is necessary to define which parameters are uncertain and whether such parameters depend on other known parameters. This enables the so-called first input data to be defined. Both the first input data and the second input data are known parameters, but they represent different types of parameters. The first output data describes the parameters estimated based on the first input data.

[0017] The first input data represent the known parameters of the first output data, which are likely to affect the uncertain input parameters of the stability model. The first input data may be machining parameters set by a machine-tool operator, such as spindle speed, feed rate, cutting depth, lubrication conditions, workpiece material, or machining parameters measured during machining, such as spindle bearing temperature, shaft load.

[0018] Second input data are parameters known with sufficiently high accuracy, such as the number of flutes of the tool, or the cutter engagement, such as the entry and exit angles. If known reliably, it is also possible to specify as second input data parameters related to the machine dynamics or the cutting force interaction, such as, for example, the modal parameters of the tool center point or the cutting force coefficients.

[0019] The parameters required as input to the stability model are defined by the type of stability model used. If a parameter is known exactly, it is fed directly into the stability model as second input data. If a parameter is unknown, or its dependence on other known parameters is uncertain, the parameter is derived from a neural network. For example, the entry and exit angles or the number of flutes are typically second input data. However, modal parameters or cutting force coefficients defining the machine dynamics can be second input data or first output data depending on how accurately and reliably their values ​​are known.

[0020] In order to determine the functional relationship between the first input data and the first output data, an artificial neural network is implemented upstream of the stability model. The artificial neural network is trained by comparing the prediction data generated by the stability model with experimental data, which are generated from a real cutting process under corresponding machine processing conditions. The actual stability of the cutting process is monitored by using a vibration detection system, which identifies whether a vibration phenomenon occurs during machine processing and outputs the vibration frequency if vibration is detected. This collected information constitutes experimental data and must be related to the corresponding machine processing conditions. Therefore, the learning process of the artificial neural network does not explicitly require dedicated tests, but can be carried out during production conditions. Moreover, it is possible to identify any interactions between parameters, such as modal parameters that depend on spindle speed and load conditions.

[0021] Artificial neural networks generally consist of An input layer with nodes, and a hidden layer with The input and hidden layers use weights are connected, the weights are connected to the corresponding input Multiply it and pass it to the node At each node, the sum of all inputs multiplied by the corresponding weights is used to activate the function The results and weights are multiplied and summed to obtain the desired output The hidden and output layers also have a bias term and , which is a constant added to the corresponding layer. In the simplest case where constant parameters are assumed, the network reduces to the output bias value .

[0022] In the present invention, the first output data in the output layer becomes the input to the stability model. One challenge is to choose a number of nodes that is suitable for a particular application. In one variant, this number is therefore chosen as a hyperparameter that is looped over during training.

[0023] For a given first output data, the stability model generates prediction data, which includes stability constraints and the associated flutter frequency The predicted data is then compared to the experimental data to adjust the weights of the neural network. The goal is to tune the weights in the artificial neural network so that any combination of network inputs produces a stability model input such that the predicted process stability and chatter frequency are consistent with the experimental data.

[0024] Depending on the application, an appropriate training algorithm must be selected. The most common learning algorithm for a simple feedforward neural network is back propagation. Back propagation describes a technique in which the error between the output and the target value is calculated and back propagated through the network to adjust the weights in the network. However, this technique requires that all operations in the network structure are differentiable. In this work, the zero-order solution (ZOS) is used as a stability model. Since the ZOS includes non-differentiable operations and the error term is only calculated downstream of the stability model, it is impossible to back propagate the error. Therefore, different training algorithms need to be applied. In the present invention, an evolutionary algorithm is used to train the neural network. This means that the weights and bias terms of the neural network are adjusted by using an evolutionary algorithm. In a preferred embodiment, a genetic algorithm is used to train the neural network. The algorithm is initialized by creating a population of individuals. Each individual includes several genes, which represent the variables of the problem. All genes therefore together constitute a possible solution to the problem under consideration. The fitness of each individual is evaluated using a defined objective function. Individuals with high fitness values ​​are then selected for performing gene operations of crossover and mutation.

[0025] In one embodiment, each individual in the population represents a possible solution for the weights and biases of the network. For each measured sample, first input data is fed into the network and first output data is calculated using the weights encoded in the corresponding individual. The network output is fed into the stability model along with the second input data and predicted data including stability limits and chatter frequencies are calculated for each sample. After comparing the predicted and measured stability data, i.e., the predicted data and the experimental data, the weights and biases of the neural network can be tuned until the deviation of the predicted data and the measured stability data is below a defined value.

[0026] In some embodiments, the obtained relationship between the first input data and the first output data can also be used in other models besides the stability model. For example, the identified relationship between the modal parameters defining the machine dynamics and the spindle speed can be used as an input to a spindle speed dependent tool displacement prediction model. The modal parameters include, for example, natural frequency and damping ratio.

[0027] As outlined above, it is necessary to ensure the accuracy of the prediction using data obtained in a real production environment. Therefore, the first input data and / or the second input data and / or the experimental data are at least a part of the collected data, which can be acquired when the machine tool is machining the workpiece. The use of experimental data measured during machining improves the accuracy of the prediction. The collected data includes not only experimental data but also machining parameters set by the operator, such as axis positioning, axis feed direction, cutting depth, spindle speed and workpiece parameters, or measured machining parameters, such as spindle temperature and axis load.

[0028] In the training process of artificial neural networks, overfitting to the presented data set may be a concern. Overfitting describes the state when the network performs very well on the data used for training but performs poorly on the data it has never seen. In other words, the resulting model will not generalize well to new data, which is one of the key ideas of artificial neural networks. In order to reduce the risk of overfitting, the method can be further optimized by dividing the collected data into a training set and a validation set for verifying the trained data. In one variant, the collected data is split into training and validation data sets, such as approximately 70% and 30% of the total data, respectively. The idea is to perform weight tuning until the validation error begins to increase and the training error continues to decrease. This behavior may be a hint that overfitting is occurring.

[0029] In a preferred variant, at least a second artificial neural network is connected to the stability model to feed the output data of the second neural network into the stability model. In particular, if necessary, further neural networks can be connected to the stability model. The number of neural networks to be applied is application-dependent, but is not limited to 2. The two independent neural networks connected to the stability model each have their own structure with a specific number of nodes and network input and output layers. If no interaction between the input parameters is expected, it can be advantageous to create multiple neural networks instead of combining all input and output parameters into one neural network. This leads to an increased identification speed and avoids the inclusion of physically absurd solutions. For example, there can be one network for identifying feed-related cutting force coefficients and another network for identifying TCP dynamics as a function of spindle speed and positioning in the workpiece.

[0030] In the present invention, the processing unit is configured to execute the method of the present invention and includes a neural network module, a stability model module and a comparison module.

[0031] Preferably, the chatter prediction unit is furthermore configured to establish a stability map. If the chatter prediction unit is located in a central system, which can be accessed by different machine tools, the generated stability map can be centralized and easily made accessible to the machine tools. The collected data are advantageously obtained during machining of the workpiece. For this purpose, the machine tool comprises a sensing unit for obtaining experimental data. Furthermore, the machine tool comprises a communication unit, which is configured to send the experimental data and the machine machining parameters set by the operator to the chatter prediction unit, and to use the stability map generated by the chatter prediction unit to select the machine machining parameters for machining the workpiece.

[0032] In the present invention, a system includes a plurality of machine tools and a chatter prediction unit. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] In order to describe the manner in which the advantages and features of the present disclosure can be obtained, hereinafter, a more particular description of the principles briefly described above will be presented by reference to specific embodiments thereof, which are illustrated in the accompanying drawings. These drawings depict only exemplary embodiments of the present disclosure and, therefore, are not to be considered limiting of its scope. The principles of the present disclosure are described and explained with additional specificity and detail through the use of the accompanying drawings, in which:

[0034] Figure 1 The method of the present invention is illustrated;

[0035] Figure 2 An artificial neural network is illustrated;

[0036] Figure 3 A first embodiment of the present invention is illustrated;

[0037] Figure 4 A second embodiment of the present invention is illustrated;

[0038] Figure 5 The simulated results of the second embodiment are shown;

[0039] Figure 6 A third embodiment of the present invention is illustrated;

[0040] Figure 7 The measurement results of the third embodiment are shown;

[0041] Figure 8 The system of the present invention is illustrated. DETAILED DESCRIPTION

[0042] The present invention presents a method for chatter prediction. In particular, the method aims to identify the unknown stability model input parameters required for calculating the stability boundary in a milling operation. The main steps of the method are Figure 1 In step 1, during machining of a workpiece, the collected data, including current machining parameters and experimental data of the corresponding machining, are continuously recorded. The collected data can be stored in Figure 8 In the central database of the flutter prediction unit 1 shown in . In steps 2 and 3, part of the collected data is fed into an artificial neural network for identifying unknown parameters required for generating a stability diagram by using a stability model. One condition for identifying the unknown parameters is to obtain the best possible agreement between the model stability prediction and the measurement results. In order to achieve this, a comparison module is provided for comparing the predicted data and the experimental data. The deviation of these data is used to train the neural network so that the deviation is optimized to a defined threshold. In step 4, after the unknown parameters have been identified, a stability diagram can be generated by using the identified parameters. In addition, with the identified parameters and relationships, stability prediction for new parameter combinations is possible.

[0043] Figure 2 shows the basic structure of a neural network with: An input layer with nodes, and a hidden layer with The input and hidden layers use weights are connected, the weight With the corresponding input Multiply it and pass it to the node At each node, the sum of all inputs multiplied by the corresponding weights is used to activate the function The results and weights are multiplied and summed to obtain the desired output The hidden and output layers also have a bias term and , which is a constant added to the corresponding layer.

[0044] Figure 3 A first embodiment of the present invention is illustrated. First input data is fed into a neural network. First output data can be calculated based on the weights and bias terms of the neural network. The calculated first output data is then fed into a stability model to obtain prediction data. A comparison module is provided, and the comparison module is configured to compare the prediction data and experimental data for the same cutting spindle speed and cutting depth. The output of the comparison module is fed back to the neural network to train the neural network. Training means adjusting the weights and bias terms of the neural network to achieve a minimum value of the deviation between the prediction data and the experimental data. In this embodiment, in addition to the first output data set, a second input data is fed into the stability model. Stability is calculated based on a zero-order solution. The stability model generates prediction data, which includes stability limits and associated chatter frequencies. The first input data includes machine processing parameters, such as spindle speed, spindle bearing temperature. The second input data includes parameters that are known due to high precision and can be directly fed into the stability model, such as inlet and outlet angles. The prediction data is then compared with the actual stability state of the experiment, i.e., the experimental data.

[0045] Figure 4 A second embodiment of the invention is illustrated. In this embodiment, two neural networks are applied and accordingly the first input data are divided into two groups and fed into the two neural networks. The first neural network is applied to identify the feed-related cutting force coefficients, while the second neural network aims to identify the TCP dynamics, the modal parameters. The inputs to the first neural network are therefore the feed rate and the workpiece material. A neural network with two output nodes is selected, which defines the cutting force coefficients. The inputs to the second neural network are the spindle speed and the machine processing temperature. The second neural network also has two output nodes defining the modal parameters, which are the natural frequency and the damping. In this embodiment, the first output data comprises the outputs of the first neural network and the second neural network.

[0046] Figure 5 Shows Figure 4Simulation results for an embodiment of a process in which four parameters are affected by four other parameters. The underlying relationships are identified from the results of 50 simulated cuts without prior knowledge about the behavior of the stability model inputs. It is assumed that the dynamics of a given machine-spindle-tool combination shows a strong dependence on the spindle speed and the spindle bearing temperature. Three slot cutters are used for cutting in two different workpiece materials (Workpiece 1 and Workpiece 2). These cuts are performed in both the X- and Y-directions of the machine tool at different feed rates, spindle speeds, and depths of cut. The following assumptions are made about the unknown parameters:

[0047] (1) Natural frequency and damping ratio Depends on spindle speed and spindle bearing temperature ;

[0048] (2) TCP dynamics are equal in the X- and Y-directions; and

[0049] (3) Cutting force coefficient and Depends on nominal feed rate and workpiece material WP .

[0050] Mathematically, these assumptions can be stated as follows:

[0051]

[0052]

[0053] in and are the tangential and radial cutting force coefficients for workpieces 1 and 2, respectively, and and The natural frequency and damping ratio are corresponding. The stability of 50 different cuts was evaluated in the simulation. For each simulation, the spindle speed, bearing temperature, entry and exit angles, feed per tooth, and cutting direction and workpiece type were randomly sampled in the defined range. The cutting depth for each simulation sample was selected as follows with respect to the theoretical stability limit: ,in is a random number drawn uniformly from the range -0.3 to 0.3. This approach corresponds to the idea of ​​performing a cut with a deviation of up to 30% from the theoretical stability limit, with the same chance of performing a cut above and below the theoretical boundary, respectively. The possibility of including such a cut away from the stability limit in the opposite sign is an important advantage of the invention.

[0054] To identify the first output data of the neural network, the first input data is split into training and validation data sets, which contain 70% and 30% of the data, respectively. The optimization is run for a maximum of 250 generations, each with a population size, i.e., number of individuals, of 250. Number of nodes and weight and bias constraints is chosen as a hyperparameter. The number of nodes is and The range between is looped over, and = 0.1 and = 10, evaluating the maximum weight in 25 steps with logarithmic spacing.

[0055] The goal is now to identify the difference between the input n and T and the output and , and input and workpiece material WP and output K tc and K rc The network output is used to calculate the speed-dependent stability diagram. The predicted stability limits are compared with Figure 5 For cases 1 and 2, very good agreement of the predicted and actual stability diagrams can be observed. Case 3 yields a less good prediction. In this case, the temperature is T = 50 degrees Celsius, which is exactly the upper limit of the distribution from which the random samples were drawn.

[0056] Figure 6 A third embodiment is illustrated, in which two neural networks are applied. However, the first input data is fed into only one neural network. The first neural network is applied to estimate the TCP dynamics, while the second network is used to identify the cutting force coefficient depending on the feed and spindle speed. The first output data includes the output of the first neural network and the second neural network. Since there is no input data in the first neural network, the neural network is defined only by the bias term and the output, such as natural frequency, damping ratio.

[0057] For this example, an experimental study was conducted to verify this approach using real-world cutting data. Two slotted 12 mm diameter end mills were fitted to a shrink fit holder with a 64 mm protrusion length. The holder was clamped to a Mikron high performance five-axis milling machine and slotting experiments were performed in Aluminum 7075 at feed rates of 0.03 mm / tooth, 0.05 mm / tooth, and 0.12 mm / tooth.

[0058] The measurement data were acquired by performing 40 cuts (21 stable and 19 unstable) at arbitrary spindle speeds between 7000 rpm and 13000 rpm and cutting depths between 1.3 mm and 2.5 mm. Since the dominant mode is expected to be that of the tool-holder combination and a shrink-fit holder is used, no dependence of the dynamics on the spindle speed is assumed. In addition to the convenience of this approach in identifying the dynamic parameters, it is shown that this method can produce more accurate results than if conventional impact tests of thin tools are considered. On the other hand, it is well known that the cutting force coefficients obtained from mechanical calibration can differ significantly at different nominal feed rates and spindle speeds. Therefore, K is set as a function of spindle speed and feed rate. tc and K rc .

[0059] Two separate networks are created. One has the purpose of identifying the modal parameters and the other is used to identify the cutting force coefficients as a function of feed rate and spindle speed. Since there are no input parameters in the first network, it is reduced to the bias term b ol (1) ,l=1…4. For the second network, one hidden layer with two nodes is used for the approximation of the relationship between the spindle speed and feed rate and the cutting force coefficient. The data set is randomly split into training and validation sets, which contain 28 and 12 samples respectively. A population of 100 individuals is used and the optimization is stopped after 150 generations. For |w in the range between 0.3 and 1.1 max The optimization process is run for different values ​​of |. max The lowest validation error is obtained with |=0.7.

[0060] In the next step, the identified network parameters are used to construct the speed and feed related stability lobes. Figure 7 The results are shown in . Overall, good predictions are obtained for the three feed rates of interest. The trend of increasing stability limits with increasing feed rate can be well predicted and the pocket position is well matched. However, for the highest feed rate, the unconditional limit cutting depth is overestimated by approximately 15%.

[0061] Figure 8A system 1 is shown, comprising a chatter prediction unit 20, a central database 21 and three machine tools 10, 11 and 12. The number of machine tools is merely an illustration of the inclusion of a plurality of machine tools and is therefore not limited to three. Experimental data are obtained by the machine tools and sent to the central database to be stored. The chatter prediction unit interrogates the required collected data from the central database to build a stability map. The built stability map is stored in the central database and is accessible to the machine tools.

[0062] Reference Mark List

[0063] 1 System

[0064] 10 Machine Tools

[0065] 11 Machine Tools

[0066] 12 Machine Tools

[0067] 20 Flutter Prediction Unit

[0068] 21 Central database.

Claims

1. A method for predicting chatter of a machine tool occurring during machining of a workpiece by the machine tool, include: - feeding first input data into an artificial neural network, the artificial neural network comprising a plurality of weights, wherein the first input data comprises known machining parameters of a first type applied to machining a workpiece by a machine tool, and the known machining parameters are machining parameters set by an operator or measured machining parameters; - determining said first output data at the output of the artificial neural network based on said first input data representing known parameters influencing first output data and said plurality of weights, wherein said first output data represents uncertain parameters required as input data for a stability model; - providing said first output data and second input data to said stability model for predicting the occurrence of chatter and a corresponding chatter frequency defining a vibration frequency between said tool and said workpiece to generate prediction data comprising the occurrence of chatter and said corresponding chatter frequency, wherein said second input data represents a second type of known machine processing parameters required as input data to the stability model; - Comparing the predicted data with the experimental data and adjusting the plurality of weights by a comparison module, wherein an evolutionary algorithm is applied to adjust the weights of the artificial neural network. 2 . The method of claim 1 , wherein the method further comprises obtaining the collected data from at least one machine tool while the machine tool is machining a workpiece. 3 . The method of claim 2 , wherein the collected data includes experimental data, machining parameters set by an operator, and machining parameters measured during machining. The method according to claim 2 , wherein a portion of the first input data and / or the second input data is obtained from collected data.

5. The method according to claim 2, wherein the collected data is divided into a training set and a validation set, and the evolutionary algorithm uses the training set to train the neural network and uses the validation set to check the accuracy of the trained neural network.

6. A chatter prediction unit, configured to execute the method according to any one of claims 1 to 5, comprising a neural network module, a stability model module and a comparison module.

7. The chatter prediction unit according to claim 6, further being configured to create a stability map.

8. A system comprising a plurality of machine tools and a chatter prediction unit according to claim 6, wherein the machine tools include: a. a sensing unit configured to obtain experimental data; b. A communication unit configured to send the collected data, including the experimental data, to a central database connected to the chatter prediction unit.