JP7171538B2 - Tool Condition Determination Apparatus, Learning Apparatus, and Method - Google Patents

Description

The present invention relates to a tool condition determination device, learning device, and method thereof.

2. Description of the Related Art Conventionally, when a mold or metal part is machined into a target shape, an operator determines the tool to be used. The operator selects tool conditions, including the type of tool, tool axial direction, and machining range, while looking at the target machining shape. Based on the selected tool conditions, the operator generates a tool locus using CAM (Computer Aided Manufacturing), performs a simulation, determines whether the workpiece can be machined into a desired shape, and determines the tool conditions. .

Since the machining time varies greatly depending on the selected tool conditions, experience is required to select the tool conditions to shorten the machining time. rice field. Moreover, the more complicated the desired shape is, the more difficult it is to select tool conditions, and even an experienced operator cannot find appropriate conditions.

On the other hand, in Patent Document 1, a plurality of tool selection patterns are set by selecting a plurality of tools from all tools, the cutting time of all the tool selection patterns is estimated by simulation, and the pattern with the shortest cutting time is selected. We propose a method of selection.

JP 2007-260849 A

However, in the method of Patent Document 1, if the number of possible tool selection patterns is small, the optimum pattern can be selected in a short time. takes time. Moreover, since the direction of the tool axis is not taken into account, the selected pattern does not necessarily have the shortest cutting time.

Accordingly, the present invention has been made in view of such circumstances, and it is an object of the present invention to provide a technique capable of easily determining appropriate tool conditions used for machining to obtain a target shape from an initial shape of a workpiece. do.

One aspect of the present invention is
A tool condition determination device for determining tool conditions for a machine tool using a plurality of types of tools,
Acquisition means for acquiring data representing the initial shape and the target shape of the workpiece;
determining means for receiving data representing an initial shape and a target shape of a workpiece and determining a set of tool conditions used to obtain said target shape from said initial shape;
with
The determining means is
a learning model machine-learned to output tool conditions used in machining for obtaining the target shape from the current shape of the workpiece;
a next shape obtaining means for obtaining a next shape after machining according to the tool conditions output by the learning model;
has
Tool conditions are obtained by using the initial shape of the workpiece as the first input to the learning model, and the next tool condition is obtained by using the next shape after machining by the tool conditions obtained from the learning model as input to the learning model. determining the set of tool conditions by repeating the process;
It is a tool condition determination device.

One aspect of the present invention is
A learning device for machine learning a learning model used to determine tool conditions for a machine tool using a plurality of types of tools,
Acquisition means for acquiring data representing the initial shape and the target shape of the workpiece;
learning means for learning the learning model by reinforcement learning with a reward for the removal volume and machining time from the initial shape of the workpiece to the target shape;
A learning device comprising:

One aspect of the present invention is
A computer-implemented tool condition determination method for determining tool conditions for a machine tool using a plurality of types of tools, comprising: an acquisition step of acquiring data representing an initial shape and a target shape of a workpiece;
a determining step of receiving data representing an initial shape and a target shape of a workpiece and determining a set of tooling conditions to be used to obtain said target shape from said initial shape;
including
In the determining step, the initial shape of the work is transferred to the learning model using a learning model machine-learned so as to output tool conditions to be used in the next machining for obtaining the target shape from the current shape of the work. Obtaining the tool condition as the first input of the above-mentioned series of tools by repeating the process of obtaining the next tool condition by further inputting the next shape after machining according to the tool condition obtained from the learning model as an input to the learning model determine the conditions,
It is a tool condition determination method.

One aspect of the present invention is
A computer-implemented learning method for machine learning a learning model used to determine tool conditions for a machine tool using multiple types of tools, comprising:
an acquisition step of acquiring data representing an initial shape and a target shape of the workpiece;
a learning step of learning the learning model by reinforcement learning with a reward of removal volume and processing time from the initial shape of the work to obtaining the target shape;
A learning method comprising:

According to the present invention, it is possible to easily determine appropriate tool conditions for machining to obtain a target shape from the initial shape of the workpiece.

1 is a functional block diagram of a learning device according to an embodiment; FIG. It is a figure which shows the point-group data of the three-dimensional model concerning embodiment. It is a figure which shows the tool axial direction and machining efficiency which concern on embodiment. It is a figure which shows the image acquisition direction concerning embodiment. 4 is a flowchart of learning processing according to the embodiment; 1 is a functional block diagram of a tool condition determining device according to an embodiment; FIG. It is a flow chart of tool condition decision processing concerning an embodiment. It is a figure which shows the specific example of the learning model concerning embodiment.

A tool condition determination device 3 (see FIG. 6) according to an embodiment of the present invention uses a machine-learned learning model 2 to determine a series of tool conditions used to obtain a target shape from the initial shape of the workpiece. . The learning model 2 receives data representing the current shape and target shape of the workpiece (for example, a CAD model) as input, and outputs tool conditions for the next machining used to obtain the target shape from the current shape. The tool condition determination device 3 obtains the tool conditions by using the initial shape of the workpiece as the first input to the learning model 2, and further uses the obtained next shape after machining according to the obtained tool conditions as an input to the learning model 2 to obtain the next tool conditions. An appropriate series of tool conditions are determined by repeating the process of obtaining .

An embodiment of the present invention includes a learning device 1 for learning a learning model 2 and a tool condition determination device 1 for determining tool conditions using the learning model. Each will be described below.

<Learning device>
FIG. 1 is a diagram showing the functional configuration of a learning device 1 according to this embodiment. The learning device 1 includes a CAD model acquisition unit 11 , a learning unit 12 and a learning model 2 . The learning unit 12 includes a data conversion unit 13 , a tool condition acquisition unit 14 , a next workpiece shape acquisition unit 15 , a machining time/removal volume calculation unit 16 and a machine learning unit 17 .

The learning device 1 is a computer including one or more processors, a main memory device, an auxiliary memory device, an input device, and an output device, and the above functions are realized by the processor executing a computer program. The processor of the learning device 1 includes a CPU (Central Processing Unit) and may further include a GPU (Graphic Processing Unit). Some or all of the above functional units may be realized by dedicated hardware circuits.

The CAD model acquisition unit 11 acquires CAD models of an initial shape and a target shape (also called a machined shape) of a workpiece. Here, the CAD model is a three-dimensional model generated by three-dimensional CAD, and is an example of data representing the shape of the workpiece. There are various file formats for CAD models such as IGES and PRT, but any file format may be adopted.

The data conversion unit 13 converts the CAD model representing the shape of the workpiece into input data for the learning model 2 . The CAD models to be converted by the data conversion unit 13 are the CAD models of the initial shape and the target shape of the workpiece acquired by the CAD model acquisition unit 11 and the CAD model of the shape of the next workpiece acquired by the next workpiece shape acquisition unit 15. be. The data after conversion by the data conversion unit 13 is the input data of the learning model 2, and must contain the features necessary for obtaining the processing conditions. Therefore, in this embodiment, the learning model 2 receives point cloud data and two-dimensional image data of the workpiece shape. In another embodiment, the learning model may receive only one of the point cloud data and the two-dimensional image data.

The data conversion unit 13 converts the CAD model into point cloud data including coordinate data and normal vector data. FIG. 2A shows a three-dimensional shape 20 represented by a CAD model, and FIG. 2B shows point cloud data 21 obtained by converting this CAD model. The point cloud data 21 is data expressing the three-dimensional shape 20 with a collection of points. Also, each point is given coordinate data and normal vector data. For example, the point 22 is given coordinate data (8, 0, 0) and normal vector data (0, -1, 0) with a magnitude of 1. In FIGS. 2A and 2B, the z-axis is perpendicular to the surface on which the workpiece is installed on the machine tool, and the x-axis and y-axis are horizontal to the installation surface.

Here, the coordinate information and normal vector information of the point cloud data are useful parameters for determining the tool diameter, tool length, tool axial direction, and machining range.

This is because the shape is specified by the coordinate information of the point cloud data, it is possible to determine whether cutting is possible geometrically, and to determine the tool diameter, tool length, and machining range. Also, the normal vector information of the point cloud data contributes to the determination of the tool axis direction. 3A and 3B show the difference in efficiency due to the difference in the tool axis when machining the surface 31 forming the shape 30 with the radius end mill 32. FIG. FIG. 3A shows the case where the tool axis 33 is oriented in the (0,0,1) direction, that is, the z-axis direction, and FIG. That is, it shows a case where it is oriented at an angle of 45° with respect to the z-axis. When the normal vector of the surface 31 is directed to (0, 0.71, 0.71), in the case of FIG.
Since the processing is performed at the corner 34 of the scanning line, the processing path of the scanning line satisfies the desired cusp height. On the other hand, in the case of FIG. 3B, since machining is performed on the bottom surface 37 of the radius end mill 32, the machining path is a circular machining path. Therefore, when comparing FIGS. 3A and 3B, the machining path is shorter in FIG. 3B, and the machining time is also shorter in FIG. 3B, so the normal direction vector information is important.

The number of point clouds included in the point cloud data 21 is preferably as large as possible because the shape can be represented more faithfully. However, the maximum number of point clouds is determined by the limitation of memory for computer processing.

Depending on the tool, it may interfere with the workpiece, and it is necessary to determine whether or not there will be interference. Therefore, in this embodiment, the data conversion unit 13 also converts the CAD model into a two-dimensional image of the workpiece viewed from a predetermined direction. FIG. 4A shows the image acquisition directions. When the tool axis direction is the z direction, the presence or absence of interference can be confirmed by looking at the work and the tool from a direction horizontal to the xy plane. Therefore, in this embodiment, as shown in FIG. 4A, the direction perpendicular to the bottom surface 41 (the surface to be installed on the machine tool) in the setup of machining of the workpiece 40 is defined as the z direction 42, and the direction 42 is horizontal to the xy plane, that is, perpendicular to the z direction 42. Acquire 2D images from different directions. In the example of FIG. 4A, the data conversion unit 13 generates a total of 12 two-dimensional image data viewed from directions shifted by 30 degrees. FIG. 4B is a two-dimensional image 43 viewed from direction A, and acquisition of such an image is performed in directions BL as well. Since the more the number of images, the more faithfully the shape can be represented, it is preferable to obtain images from more directions than just the directions A to L. In practice, the memory limit of the computer is the upper limit. Note that the data conversion unit 13 may acquire a two-dimensional image viewed from a direction other than the direction perpendicular to the z-direction 42 .

Here, learning model 2 will be described. The learning model 2 receives point cloud data including coordinate data and normal vector data generated by the data conversion unit 13 and a plurality of two-dimensional image data, and determines the tool type (tool diameter and tool length), tool axis direction, and Output the machining conditions including the machining range. The output of the learning model 2 is, more precisely, action values of possible actions (processing conditions). In order to obtain the target shape from the initial shape of the model, multiple steps of machining are performed. The learning model 2 accepts data representing the current shape and target shape of the workpiece as input and uses it to obtain the target shape. output the tool conditions for the next one-step machining.

In this embodiment, the learning model 2 is generated by unsupervised learning. An example of unsupervised learning is reinforcement learning, especially deep reinforcement learning. In deep reinforcement learning, an action-value function is approximated by a multilayer neural network. A convolutional neural network (CNN) is often used for image data, and a multilayer neural network (MLP) is often used for point cloud data, but any algorithm may be used. Also, ensemble learning may be used.

The tool condition acquisition unit 14 determines machining conditions for the next machining from the output data of the learning model 2 . The machining conditions include tool type including tool diameter and tool length, tool axial direction, and machining range. The tool condition acquisition unit 14 basically adopts an action (machining condition) that is judged to be optimal from the output data of the learning model 2, but adopts an action that is not optimal with a predetermined probability to facilitate utilization and search. Strive for balance.

The tool type (tool diameter and tool length) is selected from predetermined tool group information. This selection can be realized by the classification in the learning model 2.

The tool axis direction is a numerical value obtained as an inclination angle on the x, y, and z axes, and is calculated by regression to the learning model 2. All directions obtained by rotating each of the x, y, and z axes by an arbitrary angle can be adopted, but the tool axis direction should not be selected beyond the movable range of the rotation axis of the machine tool. You can specify the calculation range in .

The processing range is represented by a rectangular area surrounded by two points indicated by arbitrary x, y, and z coordinate values, and is calculated by regression in the learning model 2 . However, if it is a format in which the range can be seen in three dimensions, the above is not the case. Since the maximum range is determined from the workpiece and machining shape, it is calculated within that range.

A next workpiece shape acquisition unit (next shape acquisition unit) 15 acquires a shape obtained by applying the tool conditions acquired by the tool condition acquisition unit 14 to the current shape of the workpiece given as an input to the learning model 2. Calculated by simulation. Any known algorithm can be adopted for the cutting simulation.

The machining time/removed volume calculator 16 calculates the difference between the shape obtained by the next workpiece shape acquisition unit 15 and the workpiece as the removed volume. Also, the volume per unit time for each tool is obtained in advance, and the machining time is calculated by dividing the removed volume by the volume per unit time for each tool. The machining time/removed volume calculator 16 stores the removed volume and machining time in each machining in a memory.

The machine learning unit 17 learns the learning model 2 according to a machine learning algorithm. The purpose of the learning device 1 is to learn the order of tools to be used when using a plurality of tools for machining a workpiece into a target machining shape, and the tool conditions including the tool diameter, tool axial direction, and machining range. Considering that it is difficult to uniquely determine what kind of action (selection of tool conditions) is correct for the workpiece to be machined into the target shape, the machine learning unit 17 only provides rewards. Employs a learning reinforcement learning algorithm. Reinforcement learning algorithms include Q-learning, Deep Q-Network (DQN), Actor-Critic, etc., and any algorithm may be employed in the present invention.

Further, the reward for the machine learning unit 17 is determined based on the machining time and the removal volume calculated by the machining time/removal volume calculation unit 16 . The reward is determined as a large value as the processing time is short, and is determined as a large value as the removal volume is large. For example, if the machining time is shorter than a preset threshold value, the reward is positive; if the machining time is longer than the preset threshold value, the reward is negative; if the removal volume is larger than the preset threshold value, the reward is positive; Note that the numerical value of the reward may be changed to an appropriate value each time. For example, the closer the actual removed volume is to the volume that can be removed using only the smallest tool, the larger the reward based on the removed volume is determined. Also, the shorter the machining time, the larger the reward based on the machining time, and the minimum value is set for the reward based on the machining time when the machining time is equal to the machining time when only the smallest tool is used.

The learning process of the learning model 2 performed by the learning device 1 will be described with reference to FIG.

In step S<b>101 , the machine learning unit 17 initializes the learning model 2 . Specifically, weights between nodes in the deep reinforcement learning model are set to random values.

In step S102, the CAD model acquisition unit 11 acquires CAD data representing the initial shape and target shape of the workpiece.

In step S103, the data conversion unit 13 converts the CAD model representing the initial shape and the target shape of the workpiece acquired by the CAD model acquisition unit 11 into point cloud data and two-dimensional image data from a plurality of directions. As described above, point cloud data is a set of point data each having coordinate data and normal vector data. The two-dimensional image data includes two-dimensional image data of the work viewed from a plurality of directions perpendicular to the z-axis, with the direction perpendicular to the installation surface of the work as the z-axis.

In step S<b>104 , the tool condition acquisition unit 14 inputs the data converted by the data conversion unit 13 to the learning model 2 and obtains tool conditions for the next machining based on the output of the learning model 2 . The tool condition acquisition unit 14 employs random behavior (machining conditions) with a predetermined probability to balance utilization and search.

In step S105, the next workpiece shape acquiring unit 15 obtains a post-machining shape (next shape) by simulation when the current workpiece shape is processed under the machining conditions obtained in step S104.

In step S106, the machining time/removal volume calculator 16 obtains the machining time and the removed volume based on the next shape and the target shape, and the machine learning unit 17 determines a reward based on the machining conditions and the removed volume.

In step S<b>107 , the machine learning unit 17 learns the learning model 2 . In this embodiment, the machine learning unit 17 updates the weights between each node in the deep reinforcement learning model by the back propagation method. As a result, the value function (evaluation function) of learning model 2 is updated.

In step S108, the machine learning unit 17 determines whether the machining for obtaining the target shape has been completed. For example, when the next shape obtained in step S105 and the target shape match within a predetermined allowable error range, it can be determined that the machining has been completed. Alternatively, it may be determined that the machining is completed when the removed volume in the current machining is zero or smaller than the threshold value.

If it is determined that the processing has not ended (S108-NO), the process returns to step S103. In step S103, instead of the initial shape of the workpiece, the CAD model of the next shape obtained in step S105 is converted into point cloud data and two-dimensional image data. As described above, the learning unit 12 obtains the tool conditions by using the initial shape of the workpiece as the first input to the learning model 2, and further calculates the next shape after machining according to the tool conditions obtained from the learning model 2. is repeated until the target shape is obtained. In this repetitive process, the learning unit 12 learns the learning model 2 by machine learning using the removed volume and the machining time in the series of machining as a reward. If the learning unit 12 determines in step S108 that the processing has been completed (S108-YES), the learning unit 12 proceeds to step S1
03 to S107 are completed, and the process proceeds to step S109.

In step S109, the learning unit 12 determines whether or not the learning has ended, and if the learning has not ended (S109-NO), the process returns to step S102. In step S102, the initial shape and target shape of the workpiece are newly obtained, and the above-described steps S103 to S108 are repeated. If the learning has ended (S109-YES), the process ends. The termination condition of learning may be appropriately set. For example, the termination condition is that the above processing is performed for a predetermined number of sets of initial shapes and target shapes.

<Tool Condition Determining Device>
FIG. 6 is a diagram showing the functional configuration of the tool condition determining device 3 according to this embodiment. A tool condition determination device 3 uses the learning model 2 learned by the learning device 1 to determine tool conditions for a machine tool using a plurality of types of tools.

The tool condition determination device 3 includes a CAD model acquisition section 61 , a tool condition determination section 62 and a CAM 66 . The tool condition determination unit 62 includes a data conversion unit 63 , a tool condition acquisition unit 64 , a next workpiece shape acquisition unit 65 and a learning model 2 . The tool condition determination device 3 is a computer including one or more processors, a main memory device, an auxiliary memory device, an input device, and an output device, and the above functions are realized by the processor executing a computer program. . Some or all of the above functional units may be realized by dedicated hardware circuits.

The CAD model acquisition unit 61 acquires CAD models of the initial shape and target shape of the workpiece. Since the CAD model acquisition unit 61 is substantially the same as the CAD model acquisition unit 11, repeated description will be omitted.

The data conversion unit 63 converts the CAD model representing the shape of the workpiece into input data for the learning model 2 . The CAD models to be converted by the data conversion unit 63 are the CAD models of the initial shape and target shape of the workpiece acquired by the CAD model acquisition unit 61 and the CAD model of the shape of the next workpiece acquired by the next workpiece shape acquisition unit 65. be.

The data conversion unit 63, like the data conversion unit 13, converts the CAD model into point cloud data and two-dimensional image data. Since the point cloud data and the two-dimensional image data have been described above, a repeated description will be omitted.

The learning model 2 is a model that has been learned by the learning device 1 . The learning model 2 receives point cloud data and two-dimensional image data of the current shape and target shape of the workpiece as inputs, and outputs data representing tool conditions in the next machining used to obtain the target shape from the current shape.

The tool condition acquisition unit 64 provides the learning model 2 with the point cloud data and the two-dimensional image data, and selects tool conditions based on the obtained output data. The tool conditions include tool type (tool diameter and tool length), tool axial direction, and working range.

The next workpiece shape acquisition unit 65 calculates the shape obtained by applying the tool conditions acquired by the tool condition acquisition unit 64 to the current shape of the workpiece given as an input to the learning model 2 by simulation. Any known algorithm can be adopted for the cutting simulation.

The tool condition determining unit 62 obtains tool conditions by first inputting to the learning model 2 point cloud data and two-dimensional image data converted from a CAD model representing the initial shape and target shape of the workpiece to be acquired. Then, the tool condition determining unit 62 performs the process of obtaining the next tool condition by further inputting the learning model 2 with the point cloud data and the two-dimensional image data converted from the next shape after machining under the obtained machining conditions. Repeat until you get The tool condition determination unit 62 outputs a series of tool conditions thus obtained as a series of tool conditions used to obtain the target shape from the initial shape of the workpiece.

The tool conditions output by the tool condition determination unit 62 may be one set or a plurality of sets. When outputting one set of tool conditions, the tool condition acquisition unit 64 should always select one tool condition that is determined to be optimum based on the output data of the learning model 2 . When acquiring a plurality of sets of tool conditions, in one of the steps, based on the output data of the learning model 2, a plurality of tool conditions judged to be optimal are selected, and the next shape obtained from the plurality of tool conditions is selected. Processing conditions after that may be obtained for each of them.

When the tool condition determination unit 62 outputs a series of tool conditions in a plurality of sets, it is preferable to output the details of each tool condition to an output device so that the user can select which series of tool conditions is appropriate. .

The CAM 66 calculates a machining path based on a CAD model representing the initial shape of the workpiece obtained from the CAM model acquisition unit 61 and a series of machining conditions obtained from the tool condition determination unit 62, and controls the machine tool 67. Generates and outputs control data (NC data). The machine tool 67 can process the work into the target shape by performing processing based on this control data.

A tool condition determination process performed by the tool condition determination device 3 will be described with reference to FIG. 7 .

In step S201, the CAD model acquisition unit 61 acquires CAD data representing the initial shape and target shape of the workpiece.

In step S202, the data conversion unit 63 converts the CAD model representing the initial shape and the target shape of the workpiece acquired by the CAD model acquisition unit 61 into point cloud data and two-dimensional image data from a plurality of directions. As described above, point cloud data is a set of point data each having coordinate data and normal vector data. The two-dimensional image data includes two-dimensional image data of the work viewed from a plurality of directions perpendicular to the z-axis, with the direction perpendicular to the installation surface of the work as the z-axis.

In step S<b>203 , the tool condition acquisition unit 64 inputs the data converted by the data conversion unit 63 to the learning model 2 and obtains tool conditions for the next machining based on the output of the learning model 2 . Note that the tool condition acquisition unit 14 may select one machining condition determined to be optimum from the output of the learning model 2, or may select a plurality of higher machining conditions.

In step S204, the next workpiece shape acquiring unit 65 obtains a post-machining shape (next shape) by simulation when the current workpiece shape is processed under the machining conditions obtained in step S203.

In step S205, the tool condition determination unit 62 determines whether machining for obtaining the target shape has been completed. For example, when the difference between the next shape obtained in step S204 and the target shape is equal to or less than a predetermined threshold value, it can be determined that the machining is completed. More specifically, it can be determined that the machining is completed when the volume difference between the next shape and the target shape obtained in step S204 matches within a predetermined allowable error range. Alternatively, it may be determined that the machining is completed when the removed volume in the current machining is zero or smaller than the threshold value.

If it is determined that the processing has not ended (S205-NO), the process returns to step S202. In step S202, instead of the initial shape of the workpiece, the CAD model of the next shape obtained in step S204 is converted into point cloud data and two-dimensional image data. As described above, the tool condition determining unit 62 obtains tool conditions by first inputting the initial shape of the workpiece to the learning model 2, and further inputs the next shape after machining based on the obtained tool conditions to the learning model 2. , the process for obtaining the next tool condition is repeated until the target shape is obtained.

If it is determined in step S205 that the machining has been completed (S205-YES), the process proceeds to step S206, and the tool condition determination unit 62 outputs a series of machining conditions obtained by the above iterative process. When a plurality of series of machining conditions are obtained, the tool condition determination unit 62 inquires of the user which series of machining conditions should be used, and outputs the series of tool conditions specified by the user. .

In step S207, the CAM 66 generates and outputs NC data for controlling the machine tool 67 based on a series of machining conditions.

<Example>
Specific examples are described below. The following example is merely an example and does not limit the present invention to its content.

First, an embodiment of the learning device 1 will be described.

The CAD model acquisition unit 11 acquires PRASOLID format data as a CAD model representing the initial shape and target shape of the workpiece.

The data conversion unit 13 displays the CAD model using software that displays the CAD model. , two-dimensional images were acquired from a total of 36 directions of 10° around the z-axis. The data conversion unit 13 also converted the initial shape and target shape of the workpiece into point group data of 2048 points, and obtained x, y, z coordinate information and normal vector information of each point.

FIG. 8 is a diagram showing the network structure of the learning model 2 that the learning device 1 learns. The learning model 2 includes an image data feature extraction model 80 , a point cloud data feature extraction model 81 , and a multi-layer neural network (MLP; Multi-Layer Perceptron) 85 . The image data feature extraction model 80 is a well-known model in which features are extracted from each of the 36 images by CNN 81, and after Max Pooling 72, CNN 73 further extracts features. The point cloud data feature extraction model 81 is also a known model, including geometric transform, MLP, MaxPooling, and MLP. The outputs of the image data feature extraction model 80 and the point cloud data feature extraction model 81 are further input to the MLP 85, and the learning model 2 outputs 109-dimensional data. Here, 100 dimensions out of 109 dimensions are outputs for determining tool types by classification, and 100 combinations of tool diameters and tool lengths are assigned. For the remaining nine dimensions, the three dimensions were output obtained by calculating the values indicating the inclinations of x, y, and z in the tool axis direction by regression. Furthermore, the remaining six dimensions were outputs obtained by calculating the values of the x, y, and z coordinates of two points in the rectangular area indicating the machining range by regression.

The next workpiece shape acquisition unit 13 performed a cutting simulation and calculated the workpiece shape after machining when the calculated tool conditions were applied to the workpiece. More specifically, as a method of simulating cutting without NC data, a method of generating a tool path by inversely offsetting the machining shape based on the tool information, and performing a cutting simulation on the workpiece based on the tool path. It was adopted. Then, the final shape of the uncut workpiece was output in the STL format.

The machining time/removed volume calculation unit 16 obtains the difference between the volumes of the uncut workpiece and the workpiece before the cutting simulation, and uses the difference as the removed volume. Furthermore, the machining time was calculated by dividing the removed volume by the previously calculated value of the removed volume per unit time for each tool.

Rewards for removed volume and processing time were given by the formulas shown below.

Reward for volume removed = 2 x volume removed/maximum volume removed - 1
Reward for processing time = 1 - 2 x processing time / maximum processing time

Here, the maximum removal volume is the value obtained by subtracting the machining shape from the workpiece, and the maximum machining time is the value obtained by dividing the maximum removal volume by the removal volume of Φ1 per unit time. As a result, the reward ranges for both the removed volume and the processing time are -1 to 1, and the greater the removed volume and the shorter the processing time, the greater the reward. Separately from the above formula, the reward was set to -2 when the removed volume was 0, and -2 when the processing time was 0.

The machine learning unit 17 performed reinforcement learning for the learning model 2 using the above reward. Furthermore, the uncut work shape and machining shape were converted into data, the next tool conditions were calculated, cutting simulation was performed, the machining time and removal volume were calculated, converted into rewards, and learning was performed. The above was repeated until there was no uncut portion left. Then, a series of operations were performed on the other 20,000 models, and learning proceeded.

Next, an embodiment of the tool condition determining device 3 will be described. Note that the CAD model acquisition unit 61, the data conversion unit 63, and the next workpiece shape acquisition unit 65 are the same as the corresponding functional units of the learning device 1, and therefore description thereof is omitted. The learning model 2 is a model for which learning has been completed by the learning device 1 .

The tool condition determination device 3 receives as input a CAD model representing the initial shape and target shape of the workpiece for determining the tool conditions, and provides the learning model with point cloud data and two-dimensional image data converted from the CAD model to determine the tool conditions. selection. Furthermore, the process of obtaining the tool conditions by providing the learning model with point cloud data and two-dimensional image data corresponding to the work shape after machining and the target shape is repeated until the target shape is obtained. Through the above iterative processing, the tool condition determination device 3 has obtained a series of tool conditions with good machining efficiency for obtaining the target shape from the initial shape of the workpiece.

<advantageous effect>
According to the above-described embodiments and examples, a series of machining conditions including the order of tools to be used for obtaining the target shape, the tool type, tool axial direction, and machining range for each machining can be obtained with high accuracy and in a short time. Since the point cloud data is used as the input data for the learning model 2, the accuracy of machining direction selection can be improved. Moreover, since a plurality of image data viewed from the direction perpendicular to the z-axis are used as input data for the learning model 2, interference between the tool and the workpiece can be taken into consideration. Moreover, since the shape after machining under the selected machining conditions is obtained by cutting simulation, the processing time can be shortened.

<Other Examples>
The above embodiments and examples are specific examples of the invention, and the invention can be implemented in various ways.

For example, as the specific model structure of the learning model 2, any structure can be adopted as long as the action (tool condition) can be determined from the data representing the current shape and target shape of the workpiece. Also, the learning algorithm can employ deep reinforcement learning algorithms other than Deep Q-Network, and reinforcement learning algorithms such as Q Learning, Actor Critic, and SRASA.

In addition, both point cloud data and two-dimensional image data are used as input data for the learning model 2, but only the point cloud data may be used as input, or only two-dimensional image data may be used as input. good. Also, the two-dimensional image data is not limited to an image of the model viewed from a direction perpendicular to the z-axis (perpendicular to the installation surface), and may include images viewed from other directions.

Further, in the above description, the learning device 1 for learning the learning model 2 and the tool condition determination device 3 using the learning model 2 are explained as separate devices. 1 and the tool condition determination device 3 may be implemented.

The present invention supplies a program that implements one or more functions of the above-described embodiments to a system or device via a network or a storage medium, and one or more processors in the computer of the system or device reads and executes the program. It can also be realized by processing to It can also be implemented by a circuit (for example, ASIC) that implements one or more functions.

2: learning model 3: tool condition determination device 61: CAD model acquisition unit 62: tool condition determination unit 63: data conversion unit 65: next workpiece shape acquisition unit

Claims (20)

A tool condition determination device for determining tool conditions for a machine tool using a plurality of types of tools,
Image data of two-dimensional images obtained by viewing the initial shape and target shape of the work from a plurality of directions as data representing the initial shape and target shape of the work, and a series of images used to obtain the target shape from the initial shape a determining means for determining the tool condition of
with
The determining means is
a learning model machine-learned to output tool conditions used in machining for obtaining the target shape from the current shape of the workpiece;
a next shape obtaining means for obtaining a next shape after machining according to the tool conditions output by the learning model;
has
Tool conditions are obtained by using the initial shape of the workpiece as the first input to the learning model, and the next tool condition is obtained by using the next shape after machining by the tool conditions obtained from the learning model as input to the learning model. determining the set of tool conditions by repeating the process;
Tool condition determination device. conversion means for converting the CAD models of the initial shape and the target shape of the workpiece into the image data;
The tool condition determination device according to claim 1. A tool condition determination device for determining tool conditions for a machine tool using a plurality of types of tools,
determining means for receiving data representing an initial shape and a target shape of a workpiece and determining a set of tool conditions to be used to obtain said target shape from said initial shape;
with
The determining means is
a learning model machine-learned to output tool conditions used in machining for obtaining the target shape from the current shape of the workpiece;
a next shape obtaining means for obtaining a next shape after machining according to the tool conditions output by the learning model;
has
The next shape acquisition means generates a tool path by inversely offsetting the machining shape based on the tool information included in the tool condition, and obtains the next shape by simulating the workpiece based on the tool path.
Tool condition determination device. The determining means receives point cloud data of the initial shape and target shape of the workpiece as data representing the initial shape and target shape of the workpiece.
A tool condition determination device according to any one of claims 1 to 3. conversion means for converting the CAD models of the initial shape and the target shape of the workpiece into the point cloud data so that the point cloud data includes coordinate data and normal vector data;
the determining means receives as input coordinate data and normal vector data of the initial shape and target shape of the workpiece, which are included in the converted point cloud data;
The tool condition determination device according to claim 4. A tool condition determination device for determining tool conditions for a machine tool using a plurality of types of tools,
Acquisition means for acquiring data representing the initial shape and the target shape of the workpiece;
conversion means for converting the CAD models of the initial shape and target shape of the workpiece into point cloud data including coordinate data and normal vector data;
determining means for receiving data representing an initial shape and a target shape of a workpiece and determining a set of tool conditions used to obtain said target shape from said initial shape;
with
The acquisition means acquires a CAD model of the initial shape and target shape of the work as data representing the initial shape and target shape of the work,
the determination means receives as input coordinate data and normal vector data of the transformed initial shape and target shape of the workpiece;
The conversion means converts the CAD models of the initial shape and the target shape of the work into the CAD models of the initial shape and the target shape of the work with the axis perpendicular to the plane on which the work is installed on the machine tool as the z axis. Convert to a two-dimensional image viewed from multiple directions perpendicular to the axis,
The determining means receives as input coordinate data of the transformed initial shape and target shape of the workpiece, normal vector data, and a two-dimensional image,
The determining means is
a learning model machine-learned to output tool conditions used in machining for obtaining the target shape from the current shape of the workpiece;
a next shape acquiring means for obtaining a next shape after machining according to the tool conditions output by the learning model;
Tool conditions are obtained by using the initial shape of the workpiece as the first input to the learning model, and the next tool condition is obtained by using the next shape after machining by the tool conditions obtained from the learning model as input to the learning model. determining the set of tool conditions by repeating the process;
Tool condition determination device. The determination means terminates the repetition when a difference between the next shape obtained by the next shape acquisition means and the target shape is equal to or less than a threshold.
7. A tool condition determination device according to any one of claims 1 , 2 and 6. further comprising output means for outputting control data for controlling the machine tool based on the series of tool conditions output by the determination means;
The tool conditions include tool type, axial direction, and machining range.
The tool condition determination device according to any one of claims 1 to 7. A learning device for machine learning a learning model used to determine tool conditions for a machine tool using a plurality of types of tools,
As data representing the initial shape and target shape of the work, at least point cloud data of the initial shape and target shape of the work and image data of two-dimensional images of the initial shape and target shape of the work viewed from a plurality of directions. learning means for receiving one and learning the learning model by reinforcement learning using the removed volume and machining time from the initial shape of the workpiece to the target shape as a reward;
A learning device comprising: conversion means for converting the CAD models of the initial shape and the target shape of the workpiece into the image data;
10. A learning device according to claim 9. conversion means for converting the CAD models of the initial shape and the target shape of the workpiece into the point cloud data;
A learning device according to claim 9 or 10. A learning device for machine learning a learning model used to determine tool conditions for a machine tool using a plurality of types of tools,
a learning means for receiving data representing an initial shape and a target shape of a workpiece, and learning the learning model by reinforcement learning with a reward for the removal volume and machining time from the initial shape of the workpiece to the target shape;
A tool path is generated by inversely offsetting the next shape after machining according to the tool conditions output by the learning model based on the tool information included in the tool conditions, and a workpiece is simulated based on the tool paths. a next shape obtaining means for obtaining by performing
comprising
learning device. The learning model is a model that outputs tool conditions used in machining for obtaining the target shape from the current shape of the workpiece,
The learning means obtains a tool condition by using the initial shape of the workpiece as an initial input to the learning model, and further uses the next shape after machining according to the tool condition obtained from the learning model as an input to the learning model. Repeat the process of obtaining the tool conditions until the target shape is obtained, and learn the learning model by reinforcement learning with the removal volume and machining time as rewards.
A learning device according to any one of claims 9 to 12. The reward is a larger value as the removal volume is larger, and a larger value as the processing time is shorter.
A learning device according to any one of claims 9 to 13. A learning device for machine learning a learning model used to determine tool conditions for a machine tool using a plurality of types of tools,
Acquisition means for acquiring a CAD model of the initial shape and target shape of the work as data representing the initial shape and target shape of the work;
conversion means for converting the CAD models of the initial shape and target shape of the workpiece into coordinate data and normal vector data;
learning means for learning the learning model by reinforcement learning with a reward for the removal volume and machining time from the initial shape of the workpiece to the target shape;
has
the learning means receives as input coordinate data and normal vector data of the transformed initial shape and target shape of the workpiece;
learning device. The conversion means converts the CAD models of the initial shape and the target shape of the work into the CAD models of the initial shape and the target shape of the work with the axis perpendicular to the plane on which the work is installed on the machine tool as the z axis. Convert to a two-dimensional image viewed from multiple directions perpendicular to the axis,
The learning means receives as input coordinate data of the transformed initial shape and target shape of the workpiece, normal vector data, and a two-dimensional image.
16. A learning device according to claim 15. A computer-implemented tool condition determination method for determining tool conditions for a machine tool using multiple types of tools, wherein a plurality of initial shapes and target shapes of the workpiece are used as data representing the initial shape and target shape of the workpiece. a determination step of receiving image data of a two-dimensional image viewed from the direction of and determining a set of tool conditions to be used to obtain the target shape from the initial shape;
including
In the determining step, using a learning model machine-learned so as to output tool conditions used in machining for obtaining the target shape from the current shape of the work, the initial shape of the work is transferred to the learning model. obtains the tool conditions as an input, and repeats the process of obtaining the next tool conditions by using the next shape after machining according to the tool conditions obtained from the learning model as an input to the learning model, thereby obtaining the series of tool conditions decide,
Tool condition determination method. A computer-implemented learning method for machine learning a learning model used to determine tool conditions for a machine tool using multiple types of tools, comprising:
The learning model includes point cloud data of the initial shape and target shape of the workpiece as data representing the initial shape and target shape of the workpiece, and two-dimensional images of the initial shape and target shape of the workpiece viewed from a plurality of directions. A learning step of learning the learning model by reinforcement learning in which at least one of the image data is received and the removal volume and processing time from the initial shape of the workpiece to the target shape are obtained as rewards,
A learning method comprising: The learning model is a model that outputs tool conditions in the next machining used to obtain the target shape from the current shape,
In the learning step, the tool conditions are obtained by using the initial shape of the workpiece as the first input to the learning model, and the next shape after machining according to the tool conditions obtained from the learning model is further input to the learning model. Repeat the process of obtaining the tool conditions until the target shape is obtained, and learn the learning model by reinforcement learning with the removal volume and machining time as rewards.
19. Learning method according to claim 18. A computer program for causing a computer to perform the method according to any one of claims 17 to 19.

Images