JP7094996B2 - Modeling device, learning device, method and program - Google Patents

Description

The present invention relates to a type allocation model determining device, a learning device, a method and a program.

Conventionally, in the examination of production requirements in the production of parts, if the mold split position used for molding the parts cannot be determined, a problem may occur in a post-process and a design change may occur. Therefore, it is required to automate the mold allocation position determination by eliminating the backtracking due to the design change by determining the mold allocation position at the stage of examining the production requirements.

In determining the mold split position, a mold split model, which is point cloud data of the mold for determining the mold split position, may be used. In addition, when considering a mold split model from the shape of a part, it is necessary to consider various factors such as functional information of the part, formability, and workability. In addition, it is necessary to consider the mold allocation model in consideration of the fact that the factors to be emphasized differ depending on the parts and that the examination process differs for each part. Therefore, there is a problem that it is difficult to make the determination of the type allocation model into logic.

Therefore, Patent Document 1 proposes a means for extracting available mold drawing data from mold drawing data created in the past by using the specification conditions and various parameters of the external structure of the mold to be designed. Has been done.

JP-A-2017-224246

In the above technique, mold drawing data similar to the conditions used is extracted, so if similar mold drawing data exists, it is possible to specify the mold split model based on the extracted mold drawing data. can. However, with the above technology, when manufacturing a new part, the specification conditions and various parameters of the mold of the part are different from the specification conditions and various parameters of the mold of the part manufactured in the past, and it is used for manufacturing the part. It may not be possible to extract suitable mold drawing data.

The technique disclosed in the present disclosure has been made in view of the above situation, and provides a technique capable of easily determining an appropriate mold split model from the initial shape of the mold split model of a part.

One aspect of the technology disclosed in this case is
It is a type allocation model determination device that determines the type allocation model of parts.
An acquisition means for acquiring data representing the shape of the component, data representing the initial shape of the mold allocation model of the component, and functional information of the component.
A determination means that receives data representing the shape of the component, data representing the initial shape of the model of the component, and functional information of the component, and determines the model of the component.
Equipped with
The determination means is
A conversion means for converting data representing the initial shape of the mold division model of the component into point cloud data, and
Machine learning to output the adjustment amount of each point of the point cloud data based on the data representing the shape of the component, the data representing the initial shape of the mold allocation model of the component, and the functional information of the component. With a learning model that has been
A mold split model is obtained by using data representing the initial shape of the mold split model of the component as the first input to the learning model, and the mold split model obtained from the learning model is further used as an input of the learning model. By repeating the process of obtaining the above parts, the mold allocation model of the parts is determined.
It is a model allocation model determination device.

In addition, another aspect of the technique disclosed in this case is
It is a learning model used to determine the mold allocation model of the first component, and is data representing the shape of the first component, data representing the initial shape of the template allocation model of the first component, and the first component. It is a learning device that machine-learns a learning model that outputs the adjustment amount of each point of the point group data representing the initial shape of the mold division model of the first component based on the functional information of the above .
By inputting data representing the shape of the second part, data representing the initial shape of the mold split model of the second part , and functional information of the second part, the workability of the mold split model of the second part and the workability of the mold split model of the second part are input. A learning means for learning the learning model is provided by reinforcement learning using the formability of the molded product molded by using the mold split model of the second component as a reward.
It is a learning device.

Another aspect of the technique disclosed in this case is
A computer-based method for determining a part mold allocation model, which is data representing the shape of the part, data representing the initial shape of the part mold allocation model, and functional information of the part. And get the get steps and
A determination step of accepting data representing the shape of the component, data representing the initial shape of the model of the component, and functional information of the component, and determining the model of the component.
Including
In the determination step, the data representing the initial shape of the mold split model of the component is converted into point group data, and the data representing the shape of the mold split model of the component is used as the first input to the learning model to obtain the mold split model. Then, by repeating the process of obtaining the mold allocation model by further using the mold allocation model obtained from the learning model as the input of the learning model, the mold allocation model of the component is determined.
The learning model outputs the adjustment amount of each point of the point cloud data based on the data representing the shape of the component, the data representing the initial shape of the mold allocation model of the component, and the functional information of the component. Is a learning model machine-learned to do,
This is a method for determining the type allocation model.

Another aspect of the technique disclosed in this case is
It is a learning model used to determine the mold allocation model of the first component, and is data representing the shape of the first component, data representing the initial shape of the template allocation model of the first component, and the first component. It is a learning method performed by a computer that machine-learns a learning model that outputs an adjustment amount of each point of point group data representing the initial shape of the mold division model of the first component based on the functional information of the above.
By inputting data representing the shape of the second part, data representing the initial shape of the mold split model of the second part , and functional information of the second part, the workability of the mold split model of the second part and the workability of the mold split model of the second part are input. A learning step of learning the learning model by reinforcement learning using the formability of the molded product molded using the mold split model of the second component as a reward.
It is a learning method.
Further, another aspect of the technique disclosed in the present disclosure is a computer program for causing a computer to execute the above method.

According to the technique disclosed in the present disclosure, it is possible to determine a mold split model in which an appropriate mold split position is set even when manufacturing a new part, and it is possible to suppress the occurrence of backtracking in the design of the mold split model.

It is a functional block diagram of the learning apparatus which concerns on one Embodiment. It is a flowchart of the learning process of the learning model which concerns on one Embodiment. It is a flowchart of workability evaluation of the mold split model which concerns on one Embodiment. It is a flowchart of the formability evaluation of the part which concerns on one Embodiment. It is a functional block diagram of the type allocation model determination apparatus which concerns on one Embodiment. It is a flowchart of inference processing using the learning model which concerns on one Embodiment. It is a figure which shows the specific example of the learning model which concerns on one Embodiment. It is a figure which shows an example of the part which concerns on one Example. It is a figure which shows an example of the initial type split model which concerns on one Example. It is a figure which shows an example of the type split model which concerns on one Example.

The mold allocation model determination device 100 (see FIG. 5) according to the embodiment of the technique disclosed in the present disclosure determines a mold allocation model to be used for manufacturing parts by using the learning model 2 machine-learned by the learning apparatus 1. The learning model 2 accepts shape data representing the shape of a part (for example, a CAD model), functional information of the part, and initial data representing the initial shape of the type allocation model of the part (hereinafter referred to as an initial type allocation model) as inputs. Then, the learning model 2 outputs a mold allocation model suitable for manufacturing parts based on the input information.

<Learning device>
FIG. 1 is a diagram showing a functional configuration of a learning device 1 for learning a learning model 2 according to the present embodiment. The learning device 1 includes a CAD model acquisition unit 11, a function information acquisition unit 12, an initial type allocation model generation unit 13, a learning unit 20, and a learning model 2. The learning unit 20 includes a data conversion unit 21, a model adjustment amount calculation unit 22, a model update unit 23, a workability evaluation unit 24, a formability evaluation unit 25, and a machine learning unit 26. Further, the learning device 1 is connected to a shape database 30 that stores information on a CAD model and a two-dimensional drawing of pieces constituting a mold split model manufactured in the past, a processing process of each piece, a processing time, and the number of processes. There is.

The learning device 1 is a computer including one or more processors, a main storage device, an auxiliary storage device, an input device, and an output device, and the functions of the above-mentioned parts are realized by the processor executing a computer program. .. The processor included in the learning device 1 includes a CPU (Central Processing Unit) and may further include a GPU (Graphic Processing Unit). Note that some or all of the above functional parts may be realized by a dedicated hardware circuit.

The CAD model acquisition unit 11 acquires a CAD model representing the shape of the component. Here, the CAD model is a three-dimensional model generated by three-dimensional CAD, and is an example of shape data representing the shape of a part. There are various file formats such as IGES and PRT for the CAD model, but any file format may be adopted.

The function information acquisition unit 12 acquires the function information of the parts acquired by the CAD model acquisition unit 11. Here, the functional information includes the functional characteristics of the product in which the parts are used, the functional characteristics of the parts, the product design, the parts design, the appearance characteristics, the appearance step, the appearance burrs, the drawing tolerance, the parts rigidity / deformation, the resin shrinkage rate, and the product size. , Parts wall thickness, assembly information is included.

The initial type allocation model generation unit 13 generates a CAD model (hereinafter referred to as an initial type allocation model) of the type allocation model as an initial value for the CAD model of the parts acquired by the CAD model acquisition unit 11. Here, as a method of generating the initial type split model, for example, a model of the inverted shape of the CAD model of the component may be generated as the initial type split model, and the model thus generated may be generated based on a specific algorithm. The divided model may be generated as an initial type split model. Further, an initial type allocation model may be generated by constructing a database in which the initial type allocation model is stored in advance and selecting an initial type allocation model from the database based on the CAD model of the component. As the file format of the initial type split model, any file format may be adopted as in the CAD model. The CAD model acquisition unit 11, the function information acquisition unit 12, and the initial type allocation model generation unit 13 obtain data representing the shape of the part, data representing the initial shape of the part type allocation model, and functional information of the part. It is an example of the acquisition means to acquire.

The learning unit 20 is a learning model 2 using the CAD model acquired by the CAD model acquisition unit 11, the functional information acquired by the functional information acquisition unit 12, and the initial type allocation model generation unit 13 generated by the initial type allocation model generation unit 13. It is a learning means to learn. The learning model 2 receives the information converted by the following data conversion unit 21 as an input, and obtains the adjustment amount for the initial type allocation model as an output by deep learning. Multilayer neural networks (MLPs) are often used for deep learning. However, in the present embodiment, as the above-mentioned deep learning algorithm, another algorithm such as a convolutional neural network (CNN) may be used, or ensemble learning may be used.

In the present embodiment, the learning unit 20 learns the learning model 2 according to the machine learning algorithm. The learning device 1 aims to output an adjustment amount with respect to the current mold split model in order to obtain a mold split model optimal for molding a part. The learning unit 20 inputs the CAD model, the functional information, and the type allocation model of the parts, outputs the adjustment amount of the point cloud data of the type allocation model, and uses a new type allocation model calculated from the adjustment amount to make a learning model. Learn 2. It can be said that it is difficult to uniquely determine what kind of action (output of the adjustment amount of the point cloud data of the type allocation model) is correct for the CAD model and functional information of the component. Therefore, the machine learning unit 26 adopts a reinforcement learning algorithm for learning only by giving a reward. Reinforcement learning algorithms include Q-Learning, Deep Q-Newwork (DQN), Actor-Critic, and the like, and any algorithm may be adopted in this embodiment.

The learning unit 20 performs machine learning based on the CAD model of the component which is the input data, the functional information, the mold split model, the adjustment amount of the point group data of the mold split model which is the output data, and the workability / formability which is the reward. conduct. As a result, the state is defined by the input data, and the output data calculated for the defined state becomes the action. Furthermore, the value function (evaluation function) is updated with the workability and formability obtained by analyzing the point cloud data of the new mold split model calculated by the behavior as a reward. The above is repeated, and the above learning is repeated for a large number of parts model types. As the learning progresses, a learning model 2 capable of calculating a more suitable type allocation model is generated.

Next, the details of the learning process of the learning model 2 executed by the learning unit 20 will be described with reference to FIG.

In step S100, the machine learning unit 26 initializes the learning model 2. Specifically, the weight between each node in the deep reinforcement learning model is set to a random value.

In step S101, the learning unit 20 uses the CAD model of the parts acquired by the CAD model acquisition unit 11, the functional information of the parts acquired by the function information acquisition unit 12, and the initial type allocation of the parts generated by the initial model allocation model generation unit 13. Get the data for each of the models.

In step S102, the data conversion unit 21 performs data conversion for generating the input data of the learning model 2 with respect to the data acquired in step S101. The data conversion unit 21 uses the CAD model of the parts acquired by the CAD model acquisition unit 11, the functional information of the parts acquired by the function information acquisition unit 12, and the initial type allocation model of the parts generated by the initial type allocation model generation unit 13. It is converted into numerical data which is input data of the learning model 2. Specifically, the data conversion unit 21 converts the CAD model and the initial type split model of the component into point cloud data. Here, the point cloud data is data representing a solid represented by a CAD model as a set of points, and is represented as coordinate data of x, y, and z. The point cloud data is not limited to the coordinate data, and may hold the data of the normal vector of each point. As a method of converting the CAD model to the point cloud data, for example, the CAD model is saved as STL (Stereolithography) data, and the coordinates and normal vectors of the vertices of the triangular patch of the STL data are extracted as the point cloud data. You may. The larger the number of point clouds in the point cloud data, the more faithfully the shape of the component can be expressed, which is preferable. However, since there are restrictions on resources such as memory when performing processing on a computer, the maximum number of point clouds is determined based on the restrictions on the resources responsible for these processes.

In addition, the point cloud data of the mold split model also holds information indicating whether or not each point is a point that breaks the mold. Further, as for the functional information, if the information is the same for all the parts, only one numerical value is held, and if the information is different for each part of the parts, the numerical value for each point of the point cloud data of the CAD model is held. For example, since the rigidity / deformation of the component and the resin shrinkage rate are the same information for the entire component, only one numerical value is retained in the functional information. On the other hand, since the appearance step and the appearance characteristics (presence or absence of grain, etc.) differ depending on the part of the component, the functional information holds the numerical value corresponding to each point of the point cloud data of the CAD model. As a method of expressing numerical values, so-called label expression, One-Hot vector expression, or the like may be adopted, but the method is not limited to these.

In step S103, the model adjustment amount calculation unit 22 calculates the adjustment amount for the mold allocation model. Specifically, the model adjustment amount calculation unit 22 inputs the information converted by the data conversion unit 21 into the learning model 2 and acquires the adjustment amount for the type allocation model output from the learning model 2. The input information includes coordinate data and normal vector data of each point of the point group of the CAD model of the component, numerical data of functional information, coordinate data of each point of the point group of the mold split model, normal vector data and mold. Contains information indicating whether or not it is a point to divide. The adjustment amount for each point of the point cloud data of the mold split model obtained by the model adjustment amount calculation unit 22 by deep learning is whether or not the x, y, z coordinates indicating the changed mold split position and the mold are divided. It is composed of a combination of the information shown.

In step S104, the model update unit 23 updates the coordinates, the normal vector data, and the information indicating whether or not the point is the point to divide the mold for each point of the point cloud in the new type allocation model for the mold allocation model. Here, the adjustment amount for the mold allocation model calculated by the model adjustment amount calculation unit 22 is the adjustment amount at each point of the point cloud data of the mold allocation model. Therefore, the model update unit 23 adds the adjustment amount to the coordinate values of each point of the point cloud data of the type allocation model as an input, so that the point cloud data of the type allocation model as an input is a new type allocation model. Update to the point cloud data of.

Here, the model update unit 23 updates the information on whether or not the mold of each point in the mold split model is to be split with the information of the adjustment amount calculated by the model adjustment amount calculation unit 22. For example, when it is shown that one point in the point cloud data of the mold split model is a point that does not break the mold and the same point is a point that breaks the mold in the calculated adjustment amount. The updating unit 23 updates this point as a point for breaking the mold.

In step S105, the workability evaluation unit 24 converts the point cloud data in the new type allocation model calculated by the model update unit 23 into a CAD model, and evaluates the workability of the converted CAD model. The workability of the CAD model is the ease of making the mold shown by the mold split model, and the evaluation here is the evaluation of the mold shown by the mold split model. The details of the workability evaluation by the workability evaluation unit 24 will be described later.

In step S106, the formability evaluation unit 25 evaluates the formability of the CAD model converted in S105. Here, the formability of the CAD model is the ease of making the parts indicated by the mold split model, and the evaluation here is the evaluation of the parts molded by the mold split model. The details of the moldability evaluation by the moldability evaluation unit 25 will be described later.

In step S107, the machine learning unit 26 determines the reward based on the evaluation results of the workability evaluation unit 24 and the formability evaluation unit 25.

In step S108, the machine learning unit 26 learns the learning model 2. In the present embodiment, the machine learning unit 26 updates the weight between each node in the deep reinforcement learning model by the backpropagation method. As a result, the value function (evaluation function) of the learning model 2 is updated.

In step S109, the machine learning unit 26 determines whether or not the mold split position in the mold split model has been adjusted to a position suitable for molding the part based on the adjustment amount. For example, if the moving distance of the coordinates indicating the mold split position by the adjustment amount in the mold split model is equal to or less than a predetermined threshold value (for example, 0.01 mm or less), the adjustment of the coordinates indicating the mold split position is completed. I can judge.

If it is determined that the adjustment for the mold allocation model is not completed (S109: No), the process returns to step S103. In step S103 when the process returns from step S109 to step S103, the adjustment amount is calculated using the type allocation model obtained in step S104 as an input instead of the initial type allocation model. Then, the learning unit 20 obtains a mold split model by using data representing the initial shape of the mold split model of the component as the first input to the learning model 2, and further uses the mold split model obtained from the learning model 2 as the learning model 2. The process of obtaining the type allocation model as the input of is repeated. Then, in this iterative process, the learning unit 20 uses machine learning based on the evaluation of the workability of the mold split model and the moldability of the molded product molded using the mold split model, and the learning model 2 is performed. To learn. When it is determined that the adjustment to the type allocation model is completed (S109: Yes), the learning unit 20 proceeds to the process to S110.

In step S110, the learning unit 20 determines whether or not the learning of the learning model 2 is completed. The learning end condition may be set as appropriate, but for example, it is possible to set the end condition that the above processing has been performed on a predetermined number of type division models. When the learning of the learning model 2 is completed (S110: Yes), the learning unit 20 ends the learning process. Further, when the learning of the learning model 2 is not completed (S110: No), the learning unit 20 returns the process to S102.

Here, the evaluation of the workability of the CAD model by the workability evaluation unit 24 will be described with reference to FIG. 2. The workability evaluation unit 24 executes the process described below using the shape database 30. FIG. 3 shows a flow of processing executed by the workability evaluation unit 24.

In step S201, the workability evaluation unit 24 searches the shape database 30 for each piece constituting the mold split model converted into the CAD model, and identifies pieces having similar shapes. Since the degree of similarity between the shape data of the pieces constituting the mold allocation model and the shape data of the pieces stored in the shape database 30 can be determined by using a well-known technique, detailed description thereof is omitted here. do.

Next, in step S202, the workability evaluation unit 24 searches the shape database 30, identifies the processing process of each piece specified in step S201, and converts the specified processing process into the above CAD model. It is determined as the processing process of each piece that constitutes the model.

Next, in step S203, the workability evaluation unit 24 estimates the machining time of each step for the machining process of each piece constituting the mold split model determined in S202. As an estimation method, for example, the estimation may be performed based on the processing time of the piece stored in the shape database 30, or the estimation may be performed by executing a processing program for simulating the processing of the piece.

Next, in step S204, the workability evaluation unit 24 evaluates the workability and calculates the reward for the workability. Specifically, the workability evaluation unit 24 searches the shape database 30 for each piece constituting the mold split model converted into the CAD model, and calculates the total number of steps of the machining process determined in step S202. Further, the workability evaluation unit 24 calculates the total processing time of each process of each piece estimated in step S203. Then, the processability evaluation unit 24 adds a reward, that is, a plus when the number of processes is small or the processing time is short, based on the total number of calculated processes and the total processing time. Further, the processability evaluation unit 24 makes a negative value when there are many steps or a long processing time, that is, subtracts a reward. Here, it may be adjusted so that the smaller the number of steps or the shorter the processing time, the larger the reward to be added, and the more the number of steps or the longer the processing time, the larger the reward to be deducted.

Next, the evaluation of the formability of the CAD model by the formability evaluation unit 25 will be described with reference to FIG. 4. The formability evaluation unit 25 executes the process described below using the shape database 30. FIG. 4 shows a flow of processing executed by the moldability evaluation unit 25.

In step S301, the moldability evaluation unit 25 confirms whether or not there is an undercut of the mold split model converted into the CAD model, that is, whether or not there is a shape (part) in which the molded product is caught in the mold and cannot be taken out when the mold is cracked. Perform a cut check. Here, the formability evaluation unit 25 sets the reward in the undercut check as a plus when the undercut does not exist, and sets the reward as a minus when the undercut exists.

Next, in step S302, the formability evaluation unit 25 performs a sharp edge check for confirming whether or not there is a sharp edge in the mold split model. Conditions such as the angle of the edge to be a sharp edge may be appropriately defined, and the formability evaluation unit 25 detects the sharp edge in the mold split model with reference to the condition. Here, the formability evaluation unit 25 sets the reward in the sharp edge check as a plus when the sharp edge is not detected, and sets the reward as a minus when the sharp edge is detected.

Next, in step S303, the formability evaluation unit 25 first generates a two-dimensional drawing for each piece of the mold split model. Here, it is assumed that the two-dimensional drawings are various drawings by trigonometry. The dimensional tolerances in the 2D drawing are determined based on the part data and the shape of the past database. Next, the formability evaluation unit 25 performs a tolerance check to confirm whether or not the pieces are within the determined dimensional tolerance based on the two-dimensional drawing of each piece. Here, the formability evaluation unit 25 sets the reward in the tolerance check as a plus if there is a part that does not fit within the dimensional tolerance, and sets the reward as a minus if there is a part that does not fit within the dimensional tolerance.

Next, in step S304, the formability evaluation unit 25 performs a mold strength check to confirm whether or not there is a thin portion based on the wall thickness of the mold split model. Here, the moldability evaluation unit 25 sets the reward in the mold strength check as positive if there is no thin portion, that is, the portion whose wall thickness is less than the predetermined thickness in the mold split model, and negative if there is a thin portion. do.

Next, in step S305, the moldability evaluation unit 25 performs a sink mark check to confirm whether or not there is a sink mark on the appearance surface due to the influence of the rib thickness in the mold split model. Here, the formability evaluation unit 25 determines whether or not sink marks occur on the appearance surface based on whether or not the rib thickness is less than a predetermined thickness. Then, the moldability evaluation unit 25 sets the reward in the sink mark check to be positive if there is no spot on the appearance surface in the mold split model, and the reward to be negative if there is a spot on the appearance surface in the mold split model. ..

Next, in step S306, the moldability evaluation unit 25 performs a rib feasibility check to confirm whether or not there is a portion where the rib tip has a resin filling defect based on the draft of the mold split model. Here, the formability evaluation unit 25 sets the reward in the rib establishment check to be positive if there is no portion where the rib tip is poorly filled, and sets the reward to be negative if there is a portion where the rib tip is poorly filled.

Next, in step S307, the formability evaluation unit 25 performs an inclined core feasibility check to confirm whether or not there is a portion where the inclined core structure of the mold split model is not established as a mold structure. Here, the formability evaluation unit 25 adds a reward in the inclined core feasibility check if there is no part where the inclined core structure is not established as a mold structure, and minus a reward if there is a part where the inclined core structure is not established as a mold structure. And.

Next, in step S308, the formability evaluation unit 25 establishes a press cut to confirm whether or not there is a portion in the mold split model that cannot secure a hole (push cut) having a size that allows mold matching that constitutes the component claw portion. Perform a sex check. Here, the formability evaluation unit 25 adds a reward in the push-cutting feasibility check if there is no place where a hole that can be molded can be secured, and minus a reward if there is a place where a hole that can be mold-matched cannot be secured. And.

Next, in step S309, the formability evaluation unit 25 confirms whether or not there is a portion in the mold split model where the distance required for mold matching cannot be secured in the protrusion of the component standing wall portion composed of mold matching. Check the feasibility of eating out. Here, the formability evaluation unit 25 sets the reward for the establishment of the cut-off to be positive if there is no part where the distance required for mold matching cannot be secured, and the reward to be negative if there is a part where the distance required for mold matching cannot be secured. do.

Next, in step S310, the moldability evaluation unit 25 performs a mold removal check to confirm whether or not there is a portion on the component cavity side where the mold split model may cause a mold release defect when the mold is opened. Here, the formability evaluation unit 25 adds a reward in the cavity removal check if there is no part that may cause a mold release defect on the component cavity side at the time of mold opening, and a mold release defect on the component cavity side at the time of mold opening. If there is a place that may cause the problem, the reward will be negative.

Next, in step S311, the moldability evaluation unit 25 predicts the amount of warpage of the component by simulating the resin flow analysis by inputting the CAD model of the mold split model and the molding conditions. Then, the formability evaluation unit 25 performs a warp amount check for confirming whether or not the predicted warp amount is within the dimensional accuracy of the part. Here, the formability evaluation unit 25 adds a reward in the warp amount check when the predicted warp amount is within the dimensional accuracy of the part, and when the predicted warp amount is not within the dimensional accuracy of the part, the formability evaluation unit 25 adds a reward. The reward is negative.

Next, in step S312, the formability evaluation unit 25 determines the result of adding up the rewards calculated in each of the above steps S301 to S311 as the formability reward.

The workability reward calculated by the workability evaluation unit 24 and the formability reward calculated by the formability evaluation unit 25 by the above processing are used for determining the reward of the machine learning unit 26 in the above step S107. The reward determined by the machine learning unit 26 becomes larger as the processability reward and the moldability reward are higher. For example, for each of the processability reward and the processability reward, the higher the reward is higher than the preset threshold value, the larger the plus is, and the lower the reward is, the larger the minus is. In addition, the numerical value of the reward may be changed to an appropriate value each time.

<Type allocation model determination device>
FIG. 5 is a diagram showing a functional configuration of a type allocation model determining device 100 that executes inference using the learning model 2 according to the present embodiment. The type allocation model determination device 100 includes a CAD model acquisition unit 111, a function information acquisition unit 112, an initial type allocation model generation unit 113, an inference unit 120, and a learning model 2. The inference unit 120 includes a data conversion unit 121, a model adjustment amount calculation unit 122, and a model update unit 123.

The type allocation model determining device 100 is a computer including one or more processors, a main storage device, an auxiliary storage device, an input device, and an output device, and the functions of the above-mentioned parts can be obtained by the processor executing a computer program. It will be realized. The processor included in the type allocation model determining device 100 includes a CPU and may further include a GPU. Note that some or all of the above functional parts may be realized by a dedicated hardware circuit.

Next, the details of the inference process using the learning model 2 executed by the inference unit 120 will be described with reference to FIG. The inference unit 120 may have a learning model 2 and determines a new type allocation model for parts based on the following inference processing as a determination means.

In step 400, the inference unit 120 uses the CAD model of the parts acquired by the CAD model acquisition unit 111, the functional information of the parts acquired by the function information acquisition unit 112, and the initial type allocation of the parts generated by the initial model allocation model generation unit 113. Get the data for each of the models. The CAD model acquisition unit 111, the function information acquisition unit 112, and the initial type allocation model generation unit 113 provide data representing the shape of the part, data representing the initial shape of the part type allocation model, and functional information of the part. It is an example of the acquisition means to acquire.

In step S401, the data conversion unit 121 converts each data acquired in step S400 into numerical data which is an input of the learning model 2.

In step S402, the model adjustment amount calculation unit 122 inputs the information of the coordinate data and the normal vector data of the point cloud of the CAD model of the component generated by the data conversion unit 121. Further, the model adjustment amount calculation unit 122, the numerical data of the function information, the coordinate data of the point cloud of the initial mold split model, the normal vector data, and the information of whether or not the mold is split are input. Then, the model adjustment amount calculation unit 122 extracts the features of the parts based on these inputs. Deep learning, which has been learned by the learning unit 2 above, is used for extracting the features of the parts. By deep learning, the adjustment amount of the point cloud data of the type split model is obtained as an output. This adjustment amount indicates the adjustment amount at each point of the point cloud data of the initial type split model.

In step S403, the model update unit 123 calculates the coordinates of the point cloud data of the new type allocation model for the input type allocation model, the normal vector data, and the information indicating whether or not the point is the point where the mold is divided. Here, the adjustment amount of the point cloud data of the type allocation model calculated by the model adjustment amount calculation unit 122 indicates the adjustment amount at each point of the point cloud data of the input type allocation model. Therefore, the model update unit 123 adds the adjustment amount to the coordinate values of each point of the point cloud data of the input type allocation model, so that the point cloud data of the initial type allocation model is added to the point cloud of the new type allocation model. Update to data.

In step S404, the inference unit 120 determines whether or not the mold split position in the mold split model is adjusted to a position suitable for molding the part based on the adjustment amount. For example, if the moving distance of the coordinates indicating the mold split position by the adjustment amount in the mold split model is equal to or less than the threshold value (for example, 0.01 mm or less), it can be determined that the adjustment of the coordinates indicating the mold split position is completed. ..

If it is determined that the adjustment for the mold allocation model is not completed (S404: No), the process returns to step S402. In step S402 when the process returns from step S404 to step S402, the adjustment amount is calculated using the type allocation model obtained in step S403 as an input instead of the initial type allocation model. Then, the inference unit 120 obtains a mold split model by using data representing the initial shape of the mold split model of the component as the first input to the learning model 2, and further obtains the mold split model obtained from the learning model 2 as the learning model 2. The process of obtaining the type allocation model as the input of is repeated. When it is determined that the adjustment for the type allocation model is completed (S404: Yes), the inference unit 120 advances the process to S405.

In step S405, the inference unit 120 outputs the point cloud data of the type allocation model for which adjustment has been completed. As described above, according to the present embodiment, the mold allocation model determination device 100 obtains the point cloud data of the mold allocation model more suitable for manufacturing the parts than the initial mold allocation model as an output.

<Example>
First, an embodiment of the learning device 1 will be described. Prepare CAD models and functional information of 20000 kinds of parts. Then, using software that edits the CAD model, the CAD model of one part with PARASOLID format is displayed, and the CAD that becomes the inverted shape of the shape of the part.
A model is generated, and the generated CAD model is used as an initial type split model.

Next, the CAD model of the component and the CAD model of the initial type split model are saved as STL data, and the x, y, z coordinates and the normal vector of the vertices of the triangular patch are extracted and used as point cloud data. The number of points is 2048. Regarding the point cloud data of the initial mold split model, 1 is set for the point where the mold is split at each point, and 0 is set for other points. Regarding the functional information, the resin shrinkage rate adopts a preset numerical value, and the product functional characteristic, the component functional characteristic, the product design, and the component design are defined by the so-called Label expression. In addition, among the functional information, drawing tolerance, component rigidity / deformation, product size, and component wall thickness adopt preset values for each point in the point cloud data, and appearance characteristics, appearance steps, appearance burrs, and assembly. Information is defined by a Label representation for each point in the point cloud data.

FIG. 7 is a diagram showing a network structure of the learning model 2 learned by the learning device 1 in this embodiment. The training model 2 includes point cloud data feature extraction models 50 and 51, and a multi-layer neural network (MLP) 55. The point cloud data feature extraction model 50 extracts features using the point cloud data 52 of the component. Further, the point cloud data feature extraction model 51 extracts features using the point cloud data 53 of the type allocation model. The point cloud data feature extraction models 50 and 51 are known models, respectively, and include a geometric transformation (transform), MLP, Max Pooling, and MLP. The learning model 2 inputs the numerical data 54 of the features and functional information extracted using the point cloud data feature extraction models 50 and 51 into the MLP 55, and obtains the adjustment amount of the point cloud data of the type division model as the output 56.

The learning unit 20 adds the adjustment amount of the point cloud data of the type allocation model obtained as the output 56 to the coordinate values of each point of the point cloud data of the initial type allocation model generated by the initial type allocation model generation unit 13. Get new type split point cloud data with. The learning device 1 converts the acquired type split point cloud data into a CAD model.

Next, in order to determine the reward for processability, the processability evaluation unit 24 of the learning unit 20 searches for a similar shape of the piece from the shape database 30, determines the process and the estimated processing time, and obtains the reward. As for the standard of reward, the process in the past mold processing is defined as the standard process, and if the number of processes is smaller than the standard process, the reward is positive, and if the process is large, the reward is negative. Also, in estimating the machining time, the average machining time of each process in the past die machining is set as the standard machining time, and if it is shorter than the standard machining time, the reward is positive, and if it is long, it is negative.

Next, the formability evaluation unit 25 of the learning unit 20 performs a check for determining the formability reward. Specifically, the formability evaluation unit 25 includes undercut, sharp edge, tolerance, mold strength, sink mark, rib gradient establishment, inclined core establishment, push-cutting establishment, cut-off establishment, mold removal, and warpage. Make each check of quantity. Then, the formability evaluation unit 25 adds up the rewards calculated in each check to obtain the final reward.

The machine learning unit 26 of the learning unit 20 performs reinforcement learning using the above reward. Further, the machine learning unit 26 performs feature extraction by inputting point group data of parts, numerical data of functional information, and point group data of a new type allocation model, and outputs an adjustment amount of point group data of the type allocation model. .. The machine learning unit 26 further generates point cloud data of a new type split model, calculates a reward, and performs reinforcement learning. Then, the machine learning unit 26 repeats reinforcement learning until the adjustment amount of the point cloud data of the mold split model satisfies a predetermined condition. The conditions for ending the reinforcement learning include, for example, that the maximum movement distance of the movement distances of each point in the adjustment amount is 0.01 mm or less, and that the number of repetitions for one type division model reaches a predetermined upper limit. Can be mentioned. In this embodiment, the machine learning unit 26 performs reinforcement learning on CAD models and functional information of 20000 kinds of parts.

The inference unit 120 of the type allocation model determination device 100 outputs a new type allocation model using the learning model 2 in which the learning of the learning device 1 is completed. FIG. 8 shows an example of the CAD model of the component 60 acquired by the CAD model acquisition unit 111. As shown in FIG. 8, the CAD model of the component 60 has a main body portion 66 of the component and a U-shaped portion 67 connected to the main body portion 66. Further, the main body portion 66 is provided with a push-cut shape portion 64, the upper surface 68 of the main body portion 66 is provided with four holes 61, and the bottom surface 62 of the main body portion 66 is lightened. Further, the hole 65 formed by the U-shaped portion 67 is used as a hole for forming the push-cut shape portion 64.

9A and 9B show an example of the CAD models of the pieces 70A and 70B constituting the initial type allocation model generated by the initial type allocation model generation unit 113, respectively. The initial model split model is divided into two pieces, a piece 70A and a piece 70B. As shown in FIGS. 9A and 9B, the CAD model of the piece 70A has a recess 76A in which the shape of a part of the main body portion 66 and the U-shaped portion 67 on the bottom surface 62 side in the CAD model of the component 60 is inverted. Further, the CAD model of the piece 70B has four holes 61 in the component 60, a hole 65 for forming the push-cut shape portion 64, and a convex portion 71 in which a part of the shape of a part of the upper surface 68 side of the main body portion 66 is inverted. , 75, 76B. In the initial mold split model illustrated in FIGS. 9A and 9B, the upper surface 79A of the piece 70A and the upper surface 79B of the piece 70B are the mold split positions.

The data conversion unit 121 converts the CAD models of the parts 60, the pieces 70A, and 70B into point cloud data. Further, the model adjustment amount calculation unit 112 uses the point cloud data converted by the data conversion unit 121, the functional information of the parts acquired by the functional information acquisition unit 112, and the learning model 2, and is an initial type allocation model. Calculate the adjustment amount of the point cloud data of. Then, the model update unit 123 adjusts each point of the point cloud data of the initial type allocation model using the adjustment amount calculated by the model adjustment amount calculation unit 122, and the point of the type allocation model obtained as a result of the adjustment. Output group data. The mold allocation model determination device 100 can convert the point cloud data of the new mold allocation model output by the inference unit 120 to obtain a CAD model of the new mold allocation model for molding the component 60.

10A and 10B show an example of a CAD model of a new mold split model of the component 60 obtained by the mold split model determining device 100. As shown in FIGS. 10A and 10B, the new type split model is divided into two pieces, a piece 80A and a piece 80B. In the CAD model of the piece 80A, the hole 61 of the part 60, the lightened bottom surface 62, the hole 65 for molding the push-cut shape portion 64, and a part of the bottom surface 62 side of the main body portion 66 are inverted. It has recesses 81, 82, 85, 86A. Further, the CAD model of the piece 80B has a concave portion 87 which is an inverted shape of the push-cut shape portion 64 and a convex portion where the shape of the remaining portion of the component 60 formed by the piece 80A is inverted. In the mold split model illustrated in FIGS. 10A and 10B, the upper surface 89A of the piece 80A and the upper surface 89B of the piece 80B are the mold split positions.

As can be seen from FIGS. 10A and 10B, the mold split position in the new mold split model obtained as the output of the inference unit 120 is set so that the upper surface 89A of the piece 80A and the upper surface 89B of the piece 80B form a step 83. Has been done. Further, as can be seen from the comparison between the pieces 70A and 70B and the pieces 80A and 80B, the push-cutting shape portion 64 which cannot be formed by the pieces 70A and 70B is formed in the pieces 80A and 80B. As a result, according to the new mold splitting model, it is possible to manufacture the above-mentioned parts having better workability and moldability.

<Advantageous effect>
According to the above-described embodiment and embodiment, the mold-splitting model (the above-mentioned "new mold") in which a more appropriate mold-splitting position is set with respect to the mold-splitting model of the part (the above-mentioned "initial mold-splitting model"). A split model ") is obtained. As a result, it is possible to suppress the occurrence of design rework due to the use of the mold split model in which the mold split position is not optimized.

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

For example, if the specific model structure of the learning model 2 is such that the behavior (type allocation model in which the type allocation position is adjusted) can be determined from the type allocation model of the parts and the data showing the functional information. Any structure can be adopted. Further, as the learning algorithm, a deep reinforcement learning algorithm other than Deep Q-Network, and a reinforcement learning algorithm such as Q Learning, Actor Critic, and SARSA can be adopted.

Further, in the above description, the learning device 1 for learning the learning model 2 and the type allocation model determination device 100 using the learning model 2 are described as separate devices. However, in the above description, the functions of both the learning device 1 and the type allocation model determination device 100 may be implemented in one device (computer).

The present invention supplies a program that realizes one or more functions of the above-described embodiment to a system or device via a network or 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 the processing to be performed. It can also be realized by a circuit (for example, ASIC) that realizes one or more functions.

2: Learning model 100: Type allocation model determination device 111: CAD model acquisition unit 112: Function information acquisition unit 113: Initial type allocation model acquisition unit

Claims (17)

It is a type allocation model determination device that determines the type allocation model of parts.
An acquisition means for acquiring data representing the shape of the component, data representing the initial shape of the mold allocation model of the component, and functional information of the component.
A determination means that receives data representing the shape of the component, data representing the initial shape of the model of the component, and functional information of the component, and determines the model of the component.
Equipped with
The determination means is
A conversion means for converting data representing the initial shape of the mold division model of the component into point cloud data, and
Machine learning to output the adjustment amount of each point of the point cloud data based on the data representing the shape of the component, the data representing the initial shape of the mold allocation model of the component, and the functional information of the component. With the learning model
Have,
A mold split model is obtained by using data representing the initial shape of the mold split model of the component as the first input to the learning model, and the mold split model obtained from the learning model is further used as an input of the learning model. A mold split model determining device that determines a mold split model of the component by repeating the process of obtaining the above. The mold allocation model determining device according to claim 1, wherein the adjustment amount includes coordinates after the change of the mold allocation position in the mold allocation model and information indicating whether or not the adjustment amount is the mold allocation position. The determination means ends the repetition when the moving distance from the coordinates of the mold split position in the mold split model input to the learning model to the changed coordinates is equal to or less than the threshold value, according to claim 2. The type allocation model determination device described. The acquisition means acquires a CAD model representing the shape of the component as data representing the shape of the component, and a CAD model representing the initial data of the model allocation model of the component as the initial data of the model allocation model of the component. The type allocation model determining apparatus according to any one of claims 1 to 3. The type allocation model determining device according to any one of claims 1 to 4, wherein the functional information includes rigidity / deformation information of the component . It is a learning model used to determine the mold allocation model of the first component, and is data representing the shape of the first component, data representing the initial shape of the template allocation model of the first component, and the first component. It is a learning device that machine-learns a learning model that outputs the adjustment amount of each point of the point group data representing the initial shape of the mold division model of the first component based on the functional information of the above .
By inputting data representing the shape of the second part, data representing the initial shape of the mold split model of the second part , and functional information of the second part, the workability of the mold split model of the second part and the workability of the mold split model of the second part are input. A learning device comprising a learning means for learning the learning model by reinforcement learning using the formability of a molded product molded by using the mold split model of the second component as a reward. The learning means obtains a mold split model by using data representing the initial shape of the mold split model of the second component as the first input to the learning model, and further learns the mold split model obtained from the learning model. Reinforcement learning that repeats the process of obtaining a mold split model as model input and rewards the workability of the mold split model of the second part and the moldability of the molded product molded using the mold split model of the second part. The learning device according to claim 6 , wherein the learning model is learned by the above method. The learning device according to claim 7 , wherein the adjustment amount includes coordinates after the change of the mold split position in the mold split model and information indicating whether or not the adjustment amount is the mold split position. The learning device according to any one of claims 6 to 8 , wherein the reward for processability is determined based on at least one of a processing process and a processing time of each piece constituting the mold split model. The rewards for formability are the presence / absence of undercut, the presence / absence of sharp edges, the dimensional tolerance, the presence / absence of thin-walled parts, the rib thickness, the draft, the establishment of inclined cores, the establishment of push-cutting, and the cutoff in the mold split model. 3 . Learning device. The learning means uses a CAD model representing the shape of the second component as data representing the shape of the second component, and uses a CAD model representing the shape of the second component as initial data of the model allocation model of the second component. The learning device according to any one of claims 6 to 10 , wherein a CAD model representing the initial data of the above is used . The learning device according to any one of claims 6 to 11, wherein the functional information includes rigidity / deformation information of the second component . A computer-based method for determining a part mold allocation model, which is data representing the shape of the part, data representing the initial shape of the part mold allocation model, and functional information of the part. And get the get steps and
A determination step of accepting data representing the shape of the component, data representing the initial shape of the model of the component, and functional information of the component, and determining the model of the component.
Including
In the determination step, the data representing the initial shape of the mold split model of the component is converted into point group data, and the data representing the initial shape of the mold split model of the component is used as the first input to the learning model to generate the mold split model. By repeating the process of obtaining the mold allocation model by further using the mold allocation model obtained from the learning model as the input of the learning model, the mold allocation model of the component is determined.
The learning model outputs the adjustment amount of each point of the point group data based on the data representing the shape of the component, the data representing the initial shape of the mold allocation model of the component, and the functional information of the component. A method of determining a type split model, which is a learning model machine-learned to do so. It is a learning model used to determine the mold allocation model of the first component, and is data representing the shape of the first component, data representing the initial shape of the template allocation model of the first component, and the first component. It is a learning method performed by a computer that machine-learns a learning model that outputs an adjustment amount of each point of point group data representing the initial shape of the mold division model of the first component based on the functional information of the above.
By inputting data representing the shape of the second part, data representing the initial shape of the mold split model of the second part , and functional information of the second part, the workability of the mold split model of the second part and the workability of the mold split model of the second part are input. A learning step of learning the learning model by reinforcement learning using the formability of the molded product molded using the mold split model of the second component as a reward.
A learning method. In the learning step, a mold split model is obtained by using data representing the initial shape of the mold split model of the second component as the first input to the learning model, and the mold split model obtained from the learning model is further learned. Reinforcement learning that repeats the process of obtaining a mold split model as model input and rewards the workability of the mold split model of the second part and the moldability of the molded product molded using the mold split model of the second part. The learning method according to claim 14 , wherein the learning model is learned by the above method. The learning method according to any one of claims 6 to 11, wherein the functional information includes rigidity / deformation information of the second component . A computer program for causing a computer to execute the method according to any one of claims 13 to 16 .

Images