Method for predicting the state of a machining operation

Abstract

To provide a method for predicting status of machining operation, in particular chatter occurrence.SOLUTION: The method comprises the following steps: training a neural network having an input layer, at least one hidden layer, an output layer and a plurality of weights in a pre-training phase and a final-training phase, in which in the pre-training phase a pre-training data set is provided to the neural network to obtain a pre-trained neural network and in the final-training phase a final-training data set is fed to the pre-trained neural network to obtain a final-trained neural network, in which the pre-training data set comprises simulated data and the final-training data set comprises experimental data; and performing prediction by utilizing the final-trained neural network to derive prediction data.SELECTED DRAWING: Figure 3

Description

[Technical Field]

[0001] The present invention relates to a method for predicting the state of a machining operation, in particular the occurrence of chatter. [Background technology]

[0002] Today, prediction plays an important role in most technical fields: for example, methods for predicting the occurrence of defects during manufacturing, the quality of manufactured products, and the state of machining operations are widely deployed.

[0003] Various techniques have been developed in the past to make predictions. However, the prediction accuracy is not high enough, and the requirements for specific applications are not met. Therefore, there is an increasing demand for improving the prediction accuracy. Furthermore, when predictions are used in industrial environments, the effort required to make the predictions should be low.

[0004] One approach to making predictions is based on physical models, such as analytical and numerical models. For example, models of the thermal behavior of a system or the stresses and strains induced by external forces acting on the system, modeling the system's stability. To provide accurate predictions, a good understanding of the physical phenomena is required. When the ability to understand and characterize a particular phenomenon is limited, large uncertainties in the model parameters can lead to erroneous physical models or unreliable predictions.

[0005] Another recent and popular approach to making predictions is machine learning. However, to obtain accurate predictions, a large amount of experimental data is required. In industrial fields, collecting a large amount of experimental data can be a difficult task.

[0006] US Patent Application Publication No. 2020 / 0073343 discloses a machine learning method for optimizing the coefficients of a filter provided in a motor control device that controls the rotation of a motor of a machine tool.

[0007] In the machining industry, predicting the state of machining operations is becoming increasingly important in order to improve production quality and productivity. For example, in milling processes, machining instabilities, or chatter, can reduce the quality of manufactured parts, significantly increase the wear on cutting tools, and even damage machine tools. Furthermore, to ensure stable, chatter-free machining, process parameters are often selected very conservatively. Therefore, chatter vibrations are among the most critical phenomena limiting productivity.

[0008] One known approach to predicting chatter is to use physical models. A stability lobe diagram is typically used to represent the predicted results. The stability lobe diagram indicates whether a machining operation is stable or unstable under specific cutting conditions. From the stability lobe diagram, optimal process parameters can be selected for stable machining and maximum productivity. Typically, the stability lobe diagram separates the region between stable and unstable depth of cut as a function of spindle speed.

[0009] EP 2916187 A1 relates to a chatter database system including a central chatter database. The chatter database system is supplied with data corresponding to the machining and chatter conditions of a machining tool, in particular a milling machine, a lathe, a drilling machine or a boring machine. The invention is characterized in that the data supplied to the central chatter database is acquired and collected from at least two individual machining tools included in the chatter database system. The data is thereby sent to the central chatter database via a data connection, preferably via a secure network, and a chatter stability map is created based on the conditions actually encountered.

[0010] However, experimental stability limits often differ from theoretical ones. This is due, on the one hand, to inaccuracies in the stability models that produce the theoretical stability limits, and, on the other hand, to parameters required for these models that may not be precisely known during the cutting operation. Uncertainty in the stability model parameters directly affects the reliability of the stability model's output. Because both model-based and experimental methods for obtaining these model parameters typically require intensive preparation and analysis, recent attempts have been made to use machine learning techniques to predict stability limits. However, this approach requires a large number of samples to learn the shape of the stability lobes. This is one of the reasons why only simulated data was used. Furthermore, this method is limited to one specific toolholder combination with one specified tool length and workpiece material. This means that all training points must be acquired under these specified conditions. Summary of the Invention [Problem to be solved by the invention]

[0011] It is an object of the present invention to provide a method for predicting the state of a machining operation with improved accuracy and reliability, in particular to provide a method for predicting chatter in machine tools that can be deployed in an industrial environment. [Means for solving the problem]

[0012] According to the present invention, the above-mentioned object is achieved by the features set out in the independent claims, with further advantageous embodiments being set out in the dependent claims and the description.

[0013] In the present invention, a method for predicting a state of a machining operation, in particular the occurrence of chatter, includes training a neural network in a pre-training stage and a final training stage. The neural network has an input layer, at least one hidden layer, an output layer, and a plurality of weights. In the pre-training stage, a pre-training data set is provided to the neural network to obtain a pre-trained neural network. In the final training stage, a final training data set is provided to the pre-trained neural network to obtain a final trained neural network. The pre-training data set includes simulated data, and the final training data set includes experimental data. The method further includes performing prediction by utilizing the final trained neural network to derive prediction data.

[0014] A multilayer classification neural network is first trained in a pre-training phase with simulated data created from a physical model. In particular, the physical model represents the state of a machining operation. In the pre-training phase, the goal is for the pre-trained neural network to learn the general dependencies between the inputs and outputs of the physical model.

[0015] Next, in a final training stage, the pre-trained network is fed with a final training data set, which includes experimental data that is relevant to the conditions of the machining operation and obtained during machining conditions, to fine-tune the pre-trained neural network with experimentally measured data, enabling more accurate predictions in the subsequent process.

[0016] In this way, prediction accuracy can be improved and measurement effort can be minimized. These advantages are important for the industrial deployment of this method. The required experimental data for fine tuning can be collected during normal manufacturing.

[0017] After the pre-training stage, the pre-trained neural network has weights determined during pre-training using simulated data. The weights of the pre-trained neural network are used as initial values ​​for the final training stage. During the final training stage, the weights of the pre-trained neural network are adapted by feeding them a final training data set, which is, inter alia, empirical data.

[0018] In an advantageous variant, the amount of data contained in the pre-training data set is significantly greater than the amount of data contained in the final training data set, in particular the amount of data contained in the pre-training data set is at least 10 times greater than the amount of data contained in the final training data set.

[0019] In an advantageous embodiment, the pre-training data set includes only simulated data created from a physical model and / or the final training data set includes only experimental data. The simulated data can be created independently outside of a manufacturing environment. Thus, a large amount of simulated data can be obtained without the need to perform machining. The accuracy of neural network training is a function of the large amount of training data.

[0020] A pre-training dataset is a collection of samples, each of which includes at least one input value and at least one output value, where the output value is determined by providing the input values ​​to a physical model as input data. Logically, the number of nodes in the input layer of a neural network is the same as the number of inputs, and the number of nodes in the output layer is the same as the number of outputs.

[0021] In some applications, other parameters affect the output of the physical model and are known with uncertainties. These parameters are defined as variable parameters and are derived from experiments or simulations. The variable range of the variable parameters can be defined by a reference value and a standard deviation.

[0022] To further improve the prediction accuracy, particularly to reduce the influence of uncertainty in the variable parameters, at least two neural networks are independently trained using at least two different pre-training datasets, resulting in at least two final trained neural networks. Each pre-training dataset is created by varying at least one variable parameter according to its uncertainty range. Each pre-training dataset is created using the same type of physical model.

[0023] The number of neural networks is independent of the number of variable parameters, so the number of neural networks required may be the same as the number of variable parameters, may be less than the total number of variable parameters, or may be greater than the total number of variable parameters.

[0024] For example, the physical model is int input parameters and M output parameters, where N int and M is an integer. A neural network is int The matrix is ​​chosen to have an input layer with M input nodes and an output layer with M output nodes.

[0025] If there are L variable parameters, L is an integer. net There are N neural networks net are trained to obtain N final trained neural networks. net For each of the N neural networks, a pre-training data set is created by varying the value of at least one of the L variable parameters. net pre-training datasets must be created. In the pre-training stage, N net There are N neural networks netEach neural network is individually trained by supplying one of the L pre-training datasets. Each pre-training dataset contains a plurality of samples, including all inputs and corresponding outputs determined from the physical model given specific values ​​of the variable parameters. Different pre-training datasets for different neural networks differ from each other in that the value of at least one of the L variable parameters is changed.

[0026] To keep the amount of required empirical data low, the same final training data set may be provided to all pre-trained neural networks. To train the pre-trained neural networks to further improve accuracy, it is conceivable to provide different final training data sets. For example, multiple empirical data sets may be derived from different machine tools and / or under different machining conditions.

[0027] After the pre-trained neural networks are trained using experimental data, the final trained neural networks are ready for use. Prediction data can be determined from each final trained neural network, thereby generating N net The prediction datasets are N net An optimized prediction dataset with high accuracy can be obtained from each of the N final trained neural networks. net The predicted value can be determined by averaging the predicted data sets determined by the final trained neural networks.

[0028] If multiple prediction data sets are averaged, the inaccuracy of the prediction caused by the uncertainty of the variable parameters can be minimized. Furthermore, from the results of multiple predictions, the propagation of the uncertainty of the variable parameters to the final prediction can be estimated, thereby further increasing the reliability of the prediction. net It is advantageous to determine the trimmed mean of this prediction data set.

[0029] The method of the invention is suitable for many industrial applications where physical models cannot provide sufficient predictive accuracy, for example the prediction of the thermal behavior of machines, the wear of components or the power consumption of machines.

[0030] One particular application is the prediction of chatter occurrence in machine tools. Prediction data defines chatter occurrence in machine tools, particularly for milling, turning, or grinding. Chatter vibrations are self-excited vibrations caused by the interaction of the cutting edge of the tool with the surface of the workpiece being machined. Chatter occurrence in a milling process is related to the cutting coefficient, spindle speed, depth and width of cut, and the dynamics of the cutting tool at the Tool Center Point. One known method for predicting chatter occurrence is the construction of a physical analysis model.

[0031] In this variant, the simulated data is derived from a physical model representing stability. Such a stability model requires four different inputs: the dynamics in the area where the tool and workpiece contact, process information such as engagement conditions, information about the tool geometry, and a cutting coefficient, which relates to the thickness of the uncut chip due to the resulting cutting forces. While the tool geometry and engagement conditions are usually known with sufficient accuracy, the cutting coefficient and tool tip dynamics are subject to high uncertainty.

[0032] In some embodiments, chatter is predicted using a deep neural network (DNN). The deep neural network has an input layer, an output layer, and multiple hidden layers between the input layer and the output layer. The input layer includes one or more input nodes, and the output layer includes one or more output nodes. In particular, a hyperbolic tangent is used as an activation function for the hidden layer, and / or a softmax function is selected as the activation function for the output layer. However, any suitable neural network may be used to implement the method of the present invention.

[0033] The inputs include machining parameters, such as axis positions, axis feed directions, depth of cut, spindle speeds, and workpiece parameters, and the outputs include steady-state data sets representing the occurrence of chatter in the machine tool, including the steady-state data sets.

[0034] Some of the variable parameters are related to the dynamics in the contact area between the tool and the workpiece, such as Young's modulus of the tool, Young's modulus of the holder, density of the tool, loss factor of the tool, loss factor of the holder, outer diameter equivalent to the cylinder of the grooved part, contact stiffness between the translational tool and the holder, contact stiffness between the rotating tool and the holder, and contact damping between the rotating tool and the holder.

[0035] Additional variable parameters are the tangential cutting coefficient and the radial cutting coefficient.

[0036] In the application of chatter prediction, each input in the input layer of the neural network corresponds to one of the precisely known inputs describing the cutting process, such as spindle speed, cutting depth, entry angle, exit angle or clamp length, and each output node in the output layer corresponds to one output of the stability model, i.e., stability or chatter.

[0037] The method of the present invention can reduce the number of required experimental training datasets by approximately an order of magnitude while simultaneously enabling learning from and prediction against multiple dynamic configurations. This is achieved by utilizing transfer learning for deep neural networks (DNNs). A multilayer classification network is trained with artificial stability data created using simple dynamic and stability models of the tool and holder. This is done to teach the neural network the general dependence of the stability boundaries on several influential and easily measurable parameters. The pre-trained network is then fed with experimentally measured stable states and arbitrary cutting process conditions, which allows the network to be fine-tuned with actual data and enable more accurate stability predictions in subsequent processes. One of the key goals of the presented approach is to minimize measurement effort, making it a promising approach for industry. The required experimental data for fine-tuning can be easily collected during normal cutting operations.

[0038] Determination of stability lobe diagrams from the final trained neural network and / or predicted data can improve the visual representation of chatter occurrence associated with various cutting conditions.

[0039] In the present invention, a prediction unit is configured to implement the method disclosed in the present invention.

[0040] In the present invention, the machine tool comprises a controller, a monitoring unit and a prediction unit, the monitoring unit being connected to the controller and the prediction unit, the monitoring unit being configured to detect and characterize the occurrence of chatter during machining based on data provided by the controller and potentially additional sensors, and to prepare empirical data that is fed to the prediction unit.

[0041] In the present invention, the system comprises a plurality of machine tools and a prediction unit. The prediction data generated by the prediction unit can be shared between the machine tools. Experimental data can also be derived from different machine tools included in the system. This has the advantage that the experimental data can be collected efficiently, and the collected experimental data can correspond to various machining conditions.

[0042] To describe the manner in which the advantages and features of the present disclosure can be obtained, a more particular description of the principles briefly described above will now be made by reference to specific embodiments thereof, which are illustrated in the accompanying drawings. These drawings illustrate only examples of the present disclosure and therefore should not be considered as limiting its scope. The principles of the present disclosure will be described and explained with additional specificity and detail through the use of the accompanying drawings. [Brief explanation of the drawings]

[0043] [Figure 1] FIG. 1 is a diagram illustrating a physical model. [Figure 2] FIG. 1 illustrates a neural network. [Figure 3] FIG. 1 illustrates pre-training, final training, and prediction. [Figure 4] FIG. 1 illustrates an example of a physical model requiring uncertain parameters. [Figure 5] FIG. 1 illustrates an example of a physical model requiring uncertain parameters. [Figure 6] FIG. 1 illustrates multiple neural networks. [Figure 7] FIG. 1 illustrates a model for stability prediction. [Figure 8] FIG. 1 illustrates a model for stability prediction. [Figure 9] FIG. 10 is a diagram showing a stability lobe diagram. [Figure 10] FIG. 1 is a diagram illustrating an embodiment for predicting the occurrence of chatter. [Figure 11a]FIG. 1 is a diagram illustrating an embodiment for predicting the occurrence of chatter. [Figure 11b] FIG. 1 is a diagram illustrating an embodiment for predicting the occurrence of chatter. [Figure 12] FIG. 1 is a diagram illustrating an embodiment for predicting the occurrence of chatter. [Figure 13] FIG. 1 is a diagram illustrating an embodiment for predicting the occurrence of chatter. [Figure 14] FIG. 1 is a diagram illustrating an embodiment for predicting the occurrence of chatter. DETAILED DESCRIPTION OF THE INVENTION

[0044] 1, 2, and 3 illustrate an embodiment of the present invention based on transfer learning. Transfer learning is a method in which a model trained on one problem is used as a starting point for a slightly different but related problem. FIG. 1 shows a schematic diagram of a physical model including three inputs x1, x2, and x3 and two outputs y1 and y2. However, the number of inputs and outputs is not limited to those shown in FIG. 1. A pre-training dataset is derived from the physical model. The pre-training dataset contains a large number of samples, e.g., in the range of 1,000 to 10,000. For different samples, different values ​​of the inputs are set and the corresponding outputs are calculated.

[0045] During pre-training, a pre-training dataset is fed into the neural network shown in Figure 2, which has an input layer, several hidden layers, and an output layer. The input layer has the same number of nodes as the inputs, i.e., three in this example. The output layer has the same number of nodes as the outputs, i.e., two in this example.

[0046] Figure 3 shows an overview of the pre-training stage, final training stage, and prediction. In the pre-training stage, the weights of the neural network are determined by using simulated data from a physical model, and at the end of the pre-training stage, a pre-trained neural network with the determined weights is obtained. Therefore, this pre-trained neural network is a mapping of the physical model and understands the main influencing factors and general behavior of the physical model.

[0047] The pre-trained neural network with the determined weights is further trained in a final training stage, in contrast to the pre-training stage, in which an experimental dataset is provided to adjust the weights determined in the pre-training stage and further improve the training accuracy.

[0048] After final training, the final trained neural network is ready for deployment for prediction.

[0049] Figures 4, 5, and 6 illustrate embodiments that can further improve prediction accuracy by independently training multiple neural networks. Figure 4 illustrates an example physical model that includes variable parameters with uncertainty. In addition to inputs x1, x2, and x3, additional variable parameters p1, p2, and p3 affect the output of the physical model. These parameters are not precisely known. The values ​​of these variable parameters can vary within a range, which introduces uncertainty into the output and reduces prediction accuracy. To minimize the impact of substantial uncertainty on prediction accuracy, multiple neural networks are applied. The three neural networks illustrated in Figure 6 merely illustrate that multiple neural networks can be implemented; the number of neural networks is not limited to three.

[0050] Before pre-training begins, a pre-training dataset must be created using a physical model. To calculate the model's output by varying the input, the values ​​of the variable parameters must be determined in advance. Because the variable parameters are uncertain, different values ​​of the variable parameters are selected to be used to determine the pre-training datasets for different neural networks. In this way, the negative impact of the uncertainty of the variable parameters on the final prediction results can be reduced. As shown in Figure 5, three groups of values ​​for the variable parameters p1, p2, and p3 are determined. To create a simulated dataset for NN1, values ​​v11, v21, and v31 are used. Multiple samples are created for NN1 using the same values ​​v11, v21, and v31 and by varying the values ​​of the inputs x1, x2, and x3. To create a simulated dataset for NN2, values ​​v12, v22, and v32 are used. Multiple samples are created for NN2 using the same values ​​v12, v22, and v32 and by varying the values ​​of the inputs x1, x2, and x3. The values ​​v13, v23, and v33 are used to create a simulated data set for NN3. Multiple samples are created for NN3 using the same values ​​v13, v23, and v33, and by varying the values ​​of the inputs x1, x2, and x3.

[0051] At the end of the pre-training stage, three pre-trained neural networks are obtained. In the final training stage, the same experimental data is fed to the pre-trained neural networks. Three different final training data sets may be optional if sufficient experimental data is available. The three final trained neural networks may be used individually to make predictions. The optimized prediction data is determined by calculating the average of the prediction data sets obtained from each of the three final trained neural networks.

[0052] Figures 7 to 14 demonstrate the application of this method to chatter prediction. The physical model for stability prediction used for machine tools, specifically for milling, shown in Figure 7, takes five inputs: clamping length (S cl ), input x1 is the spindle speed (n), input x2 is the cutting depth (a p ) is the input x3, the approach angle (φ st ) and the exit angle (φ ex ) input x5. Additional information provided to the model, namely, variable parameters p1-p13, are listed in Figure 8. These variable parameters have uncertainties. The two output parameters of this physical model represent two states: stable and chatter. This physical model provides a prediction of whether the machining will be stable or not under given inputs. Typically, a stability lobe diagram is used to distinguish between stable and unstable depth of cut as a function of spindle speed.

[0053] Each variable parameter has an estimated reference value. However, this value can vary within a range according to a given probability model that can be assumed for a normal distribution from the value of the standard deviation. The values ​​shown in Figure 8 are for illustrative purposes only.

[0054] In a first step, a pre-training dataset is created by selecting variable parameters, preparing example inputs, and feeding these inputs to an existing stable model so that the stable model output can be derived. For the creation of a simulated dataset, it is not directly clear which values ​​should be assumed in the modeling stage for the variable parameters summarized in Figure 7. Here, an extension to the classical transfer learning idea is applied, which can take modeling uncertainties into account and further improve the accuracy of chatter prediction. It is based on the idea of ​​ensemble learning, where multiple networks are trained and their individual estimates are combined to obtain a single prediction.

[0055] Figure 10 shows the variable parameters selected to prepare different pre-training datasets for different neural networks. All variable parameters are sampled 20 times from their distributions defined in Figure 8 because 20 neural networks are applied. Simultaneously, 1,000 simulated cutting samples are generated, in which the spindle speed, depth of cut, entry angle, and exit angle are uniformly sampled from a defined range. These ranges can be derived from the range of the experimental dataset. For example, if the experimental dataset was obtained with a spindle speed between 6,000 rpm and 15,000 rpm, the same range can be selected for the simulated dataset. The created pre-training datasets, consisting of the inputs spindle speed, depth of cut, entry angle, exit angle, and clamp length, and the output stability / unstability, are used to pre-train one network. Figures 11a and 11b show how to prepare pre-training datasets for NN1 and NN20. Each pre-training dataset contains 1,000 samples. To determine the output, different values ​​of the five inputs are fed into the physical model, and the same variable parameters for NN1 are used for all samples.

[0056] The created pre-trained dataset is then used to train neural networks; in this example, 20 neural networks are applied. However, the number of neural networks is merely for illustrative purposes and is not limited to 20. The pre-trained neural networks then recognize the main influences on stability lobes and also learn the concept of recurring stability pockets in the spindle speed. Nevertheless, these networks may have poor performance when comparing their predictions to actual, empirical stable states.

[0057] While the simulated data aims to train the neural network to recognize the general shape of the stability lobes and their fundamental dependencies, as shown in Figure 9, the goal of fine-tuning is to compensate for four sources of error that may have been introduced in the pre-training stage: inaccuracies in modeling the dynamics in the tool-workpiece contact area, uncertainties in the cutting coefficients, potential operational changes in the dynamics and cutting coefficients (e.g., spindle speed dependence), and uncertainties in the stability model used. This problem is resolved in the fine-tuning stage. Here, a much smaller experimental data set (e.g., 50 entries, as shown in Figure 13) is fed to the pre-trained network. Its initial weights are equal to the optimized network weights from the pre-training stage. Here, the network weights will be slightly adapted to match the neural network predictions with the experimentally observed stable states.

[0058] In the next step, the fine-tuned network can be used for stability prediction of new cutting scenarios, resulting in much more accurate stability prediction. Figure 14 shows the use of multiple neural networks for prediction.

[0059] Once the stability graph for a new process state is predicted, each network makes a prediction. For example, the stability lobes shown in Figure 14 are results from various final trained neural networks. All network predictions are averaged using a trimmed mean approach, which eliminates the very high and very low predictions.

Claims

1. 1. A method for predicting the occurrence of chatter, a condition in a machining operation, comprising: a) training a neural network in a pre-training stage and a final training stage, the neural network having an input layer, at least one hidden layer, an output layer, and a plurality of weights, wherein in the pre-training stage a pre-training data set is provided to the neural network to obtain a pre-trained neural network, and in the final training stage a final training data set is provided to the pre-trained neural network to obtain a final trained neural network, the pre-training data set including simulated data, and the final training data set including experimental data; b) performing predictions by deriving prediction data using the final trained neural network; A method comprising:

2. 2. The method of claim 1, wherein the weights of the pre-trained neural network determined during the pre-training stage are adapted in the final training stage by utilizing the final training data set.

3. The method of claim 1 or 2, wherein the amount of data contained in the pre-training data set is greater than the amount of data contained in the final training data set.

4. 4. The method of claim 1, wherein the pre-training data set contains only simulated data created using a physical model and / or the final training data set contains only experimental data.

5. The method of claim 4 , wherein the pre-training dataset is a collection of multiple samples including at least one input value and at least one output value, and the output value is determined by providing the input value to the physical model as input data.

6. 6. The method of claim 4 or 5, wherein at least two final trained neural networks are obtained by independently training at least two neural networks using at least two different pre-training data sets, each pre-training data set being created by varying at least one variable parameter.

7. The method of claim 5 , wherein the physical model is a stable model defining the occurrence of chatter in a machine tool, the inputs including machining parameters, and the outputs are steady states of the machining operation.

8. 7. The method of claim 6, wherein the variable parameters include one or more of: Young's modulus of the tool, Young's modulus of the holder, density of the tool, loss factor of the tool, loss factor of the holder, cylindrical outer diameter of the grooved portion, translational tool-holder contact stiffness, rotary tool-holder contact stiffness, rotary tool-holder contact damping, tangential cutting coefficient, and radial cutting coefficient.

9. 9. The method of claim 7, further comprising determining the optimized prediction data by averaging the prediction data determined using each final trained neural network.

10. The method according to any one of claims 1 to 9, further comprising determining a stability lobe diagram from the predicted data and / or the optimized predicted data.

11. A prediction unit adapted to implement a method according to any one of claims 1 to 10.

12. A machine tool comprising a controller configured to control the machine tool, a monitoring unit, and a prediction unit as described in claim 11, wherein the monitoring unit is configured to detect and characterize the occurrence of chatter during machining and to prepare empirical data that is supplied to the prediction unit.

13. A system comprising a plurality of machine tools according to claim 12.

Images