JP7534138B2 - Structural design support program, structural design support device, and structural design support method - Google Patents

Description

The present invention relates to a structural design support program, a structural design support device, and a structural design support method.

In recent years, the design of structures, such as buildings, automobiles, bridges, and integrated circuits, has become increasingly sophisticated, and technology to support these designs has attracted attention. One example of such technology is the design device disclosed in Patent Document 1. This design device includes an optimization calculation process in which a computer calculates an optimal solution for at least one design factor based on a genetic algorithm for at least one structural member including at least a column, beam, and load-bearing wall. In addition, because this design device uses a genetic algorithm, it has the advantage of being applicable to a variety of problems.

JP 2014-153994 A

However, because the above-mentioned design device uses a genetic algorithm, it may calculate a structure that falls into a local optimum. In addition, because the above-mentioned design device needs to perform optimization for each individual problem, it may take a long time to calculate a structure.

The present invention has been made in consideration of the above-mentioned circumstances, and aims to provide a structural design support program, a structural design support device, and a structural design support method that can provide more thorough support for the work of designing the structure of a structure.

One aspect of the present invention is a computer that includes a condition data acquisition function for acquiring condition data indicating conditions to be satisfied when designing a structure of a structure and prerequisites to be referenced when designing the structure of the structure, a first model that designs the structure of the structure as a machine learning device under conditions indicated by the condition data, generates generated data indicating the structure of the structure, and is linked to a third model with fixed weights and learns using second error data, a generated data acquisition function for acquiring the generated data generated by the first model, a structural calculation function for outputting structural calculation result data using the generated data and the condition data, and a second model that generates first structural calculation result estimation data from the first structural calculation result data and learns using the first error data; a first error calculation function that generates and outputs the first error data from the structural calculation result data and the first structural calculation result estimation data; a weight data acquisition function that acquires weight data from the second model; a third model that has the same configuration and the same weights as the second model using the weight data and generates and outputs second structural calculation result estimation data from the generated data; and a second error calculation function that calculates the second error data from at least a portion of the condition data and the second structural calculation result estimation data.

In addition, in the structural design support program described above, the first model may be divided into a process of generating the generated data for a first component that is arranged along a first direction defined in the structure, and a process of generating the generated data for a second component that is defined in the structure and arranged along a second direction that intersects with the first direction.

In addition, in the structural design support program described above, at least one of the first model, the second model, and the third model may include at least one of a neural network and an attention mechanism.

One aspect of the present invention includes a condition data acquisition unit that acquires condition data indicating conditions to be satisfied when designing a structure of a structure and prerequisites to be referenced when designing the structure of the structure, a first model that designs the structure of the structure as a machine learning device under the conditions indicated by the condition data, generates generated data indicating the structure of the structure, and is connected to a third model with fixed weights and learns using second error data, a generated data acquisition unit that acquires the generated data generated by the first model, a structural calculation unit that outputs structural calculation result data based on the generated data and the condition data, and a structural calculation unit that calculates structural calculation results from the generated data. The structural design support device includes: a second model that generates first structural calculation result estimation data and learns using first error data; a first error calculation unit that generates and outputs the first error data from the structural calculation result data and the first structural calculation result estimation data; a weight data acquisition unit that acquires weight data from the second model; a third model that has the same configuration and the same weights as the second model using the weight data, and generates and outputs second structural calculation result estimation data from the generated data; and a second error calculation unit that calculates the second error data from at least a portion of the condition data and the second structural calculation result estimation data.

Furthermore, in the above-mentioned structural design support device, the first model whose weighting has been updated based on the second error data may design a structure of the structure using the condition data , and output generated data indicating a result of designing the structure of the structure.

One aspect of the present invention includes a step of acquiring condition data indicating conditions to be satisfied when designing a structure of a structure and prerequisites to be referred to when designing the structure of the structure; a step of a first model, which is a model that generates generation data indicating the structure of the structure under the conditions indicated by the condition data and is a model that learns using second error data, designing the structure of the structure as a machine learning device and acquiring the generated generation data; a step of outputting structural calculation result data using a structural calculation device based on the generation data and the condition data; a step of generating and outputting first error data based on the first structural calculation result estimation data generated by a second model, which is a model that generates first structural calculation result estimation data from the generation data and is a model that learns using first error data, and the structural calculation result data; The structural design support method includes the steps of: learning by updating weights of a model; acquiring updated weight data from the second model ; copying the weight data to a third model, which has the same configuration as the second model and is connected to the first model, and which generates and outputs second structural calculation result estimation data from the generation data, thereby making the third model a model having the same weights as the second model ; calculating second error data from at least a portion of the condition data and the second structural calculation result estimation data; learning a connected model in which the first model and the third model are connected, by updating only the weight of the first model of the connected models from the second error data; and acquiring at least one of the first model and the second model as a learned model.

In addition, the above-mentioned structural design support method may further include a step in which the first model, whose weights have been updated based on the second error data , designs a structure of the structure using the condition data, and outputs generated data indicating a result of designing the structure of the structure.

The present invention provides more thorough support for the work of designing the structure of a structure.

FIG. 2 is a diagram showing an example of functions included in a structural design support program according to an embodiment. 1 is a conceptual diagram illustrating an example of processing executed by a structural design support program according to an embodiment. FIG. 13 is a diagram showing an example of generated data that is generated at a stage when the structural design support program according to the embodiment generates a trained model, and indicates a position where it is determined that it is preferable to place a load-bearing wall in the X direction on the first floor of a building. FIG. 13 is a diagram showing an example of generated data that is generated at the stage when the structural design support program according to the embodiment generates a trained model, and indicates a position where it is determined that it is preferable to place a load-bearing wall in the X direction on the second floor of a building. FIG. 13 is a diagram showing an example of generated data generated at the stage when the structural design support program according to the embodiment generates a trained model, and indicating a position where it is determined that it is preferable to place a load-bearing wall in the Y direction on the first floor of a building. FIG. 13 is a diagram showing an example of generated data generated at the stage when the structural design support program according to the embodiment generates a trained model, showing a position where it is determined that it is preferable to place a load-bearing wall in the Y direction on the second floor of a building. A conceptual diagram showing an example of processing executed by a trained model acquisition function according to an embodiment. FIG. 13 is a diagram showing an example of the relationship between the number of epochs of the machine learning device and the value of the loss function calculated by the first error calculation function at the stage when the structural design assistance program according to the embodiment generates a trained model. FIG. 13 is a diagram showing an example of the relationship between the number of epochs of the machine learning device and the value of the loss function calculated by the second error calculation function at the stage when the structural design support program according to the embodiment generates a trained model. FIG. 13 is a diagram showing an example of the relationship between the number of epochs of the machine learning device and the value of the loss function calculated by the second error calculation function when test data is input at the stage when the structural design support program according to the embodiment generates a trained model. FIG. 11 is a diagram showing an example of a relationship between the eccentricity ratio estimated by a third model in the X direction on the first floor of a building related to learning of a first model and a second model, at a stage when the structural design support program according to the embodiment generates a first model and a second model as trained models, and the number of epochs. FIG. 11 is a diagram showing an example of a relationship between the eccentricity ratio estimated by a third model in the X direction on the second floor of a building related to learning of a first model and a second model, at a stage when the structural design support program according to the embodiment generates a first model and a second model as trained models, and the number of epochs. FIG. 13 is a diagram showing an example of the relationship between the eccentricity estimated by the third model in the Y direction on the first floor of a building related to the learning of the first model and the second model, at a stage when the structural design support program of the embodiment generates the first model and the second model as trained models, and the number of epochs. FIG. 13 is a diagram showing an example of the relationship between the eccentricity estimated by the third model in the Y direction on the second floor of a building related to the learning of the first model and the second model, at a stage when the structural design support program of the embodiment generates the first model and the second model as trained models, and the number of epochs. FIG. 11 is a diagram showing an example of a relationship between the number of epochs and a load-bearing wall test value estimated by a third model for the X-direction on the first floor of a building related to learning of the first model and the second model, at a stage when the structural design support program according to the embodiment generates a first model and a second model as trained models. FIG. 13 is a diagram showing an example of a relationship between the number of epochs and a load-bearing wall test value estimated by a third model for the X-direction on the second floor of a building related to learning of the first model and the second model, at a stage when the structural design support program according to the embodiment generates a first model and a second model as trained models. This figure shows an example of the relationship between the load-bearing wall test value estimated by the third model for the Y direction on the first floor of a building related to the learning of the first model and the second model, at the stage when the structural design support program of the embodiment generates the first model and the second model as trained models, and the number of epochs. This figure shows an example of the relationship between the load-bearing wall test value estimated by the third model for the Y direction on the second floor of a building related to the learning of the first model and the second model, at the stage when the structural design support program of the embodiment generates the first model and the second model as trained models, and the number of epochs. FIG. 1 is a diagram illustrating an example of a two-story building according to an embodiment. FIG. 20 is a diagram showing an example of generated data indicating a position determined by the structural design support program according to the embodiment as being preferable for placing a load-bearing wall in the X direction on the first floor of the building shown in FIG. 19 . FIG. 20 is a diagram showing an example of generated data indicating a position determined by the structural design support program according to the embodiment to be preferable for placing a load-bearing wall in the X direction on the second floor of the building shown in FIG. 19 . 20 is a diagram showing an example of generated data indicating a position determined by the structural design support program according to the embodiment to be preferable for placing a load-bearing wall in the Y direction on the first floor of the building shown in FIG. 19 . FIG. FIG. 20 is a diagram showing an example of generated data indicating a position determined by the structural design support program according to the embodiment to be preferable for placing a load-bearing wall in the Y direction on the second floor of the building shown in FIG. 19 . FIG. 13 is a diagram showing an example of the relationship between the number of epochs of a machine learning device and the value of a loss function at a stage when the structural design support program according to the embodiment re-learns a trained model used by the machine learning device. FIG. 13 is a diagram showing an example of the relationship between the number of epochs and the value of the loss function of the machine learning device at a stage where the structural design assistance program according to the embodiment re-learns a trained model. FIG. 20 is a diagram showing an example of the relationship between the eccentricity ratio estimated by a trained model undergoing re-learning in the X-direction on the first floor of the building shown in FIG. 19 and the number of epochs. FIG. 20 is a diagram showing an example of the relationship between the eccentricity ratio estimated by a trained model undergoing re-learning in the X-direction on the second floor of the building shown in FIG. 19 and the number of epochs. FIG. 20 is a diagram showing an example of the relationship between the eccentricity ratio estimated by a trained model undergoing re-learning in the Y direction on the first floor of the building shown in FIG. 19 and the number of epochs. FIG. 20 is a diagram showing an example of the relationship between the eccentricity ratio estimated by a trained model undergoing re-learning in the Y direction on the second floor of the building shown in FIG. 19 and the number of epochs. FIG. 20 is a diagram showing an example of the relationship between the load-bearing wall test value estimated by a trained model undergoing re-learning for the X-direction of the first floor of the building shown in FIG. 19 and the number of epochs. FIG. 20 is a diagram showing an example of the relationship between the load-bearing wall test value estimated by a trained model undergoing re-learning for the X-direction on the second floor of the building shown in FIG. 19 and the number of epochs. This figure shows an example of the relationship between the load-bearing wall test value estimated by a trained model undergoing re-learning for the Y direction on the first floor of the building shown in Figure 19 and the number of epochs. This figure shows an example of the relationship between the load-bearing wall test value estimated by a trained model for which re-learning is in progress for the Y direction on the second floor of the building shown in Figure 19 and the number of epochs. FIG. 34 is a flowchart showing an example of processing executed by the structural design support program according to the embodiment in the learning phase.

[Embodiment]
An embodiment of the present invention will be described with reference to the drawings. Fig. 1 is a diagram showing an example of functions provided in a structural design support program according to an embodiment. Fig. 2 is a conceptual diagram showing an example of processing executed by a structural design support program according to an embodiment. As shown in Figs. 1 and 2, a structural design support program 10 includes a condition data acquisition function 11, a first model 102, a generated data acquisition function 103, a structural calculation function 104, a second model 105, a first error calculation function 106, a weight data acquisition function 107, a third model 108, a second error calculation function 109, and a trained model acquisition function 110.

At least some of the functions of the structural design support program 10 are realized by, for example, a hardware processor such as a CPU (Central Processing Unit) reading and executing the structural design support program 10 stored in a storage device. The storage device referred to here is, for example, a hard disk drive (HDD: Hard Disk Drive), a solid state drive (SSD: Solid State Drive), a flash memory, or a read only memory (ROM).

The condition data acquisition function 11 acquires condition data indicating the conditions that should be satisfied when designing the structure of a structure and the prerequisites that should be referred to when designing the structure of a structure. The structures referred to here are, for example, buildings, automobiles, bridges, integrated circuits, and molecules. The condition data is, for example, data that specifies the conditions required for at least one of the design of the structure, the framework of the structure, and the layout of the main components of the structure as prerequisites. Furthermore, the condition data is data that specifies, for example, the allowable load, deformation limit, transmitted energy, and absorbed energy for the materials or parts that make up the structure.

When the structure is a building, the preconditions indicated by the condition data are data indicating conditions imposed on at least one of the shape, dimensions, material, and position of at least one of the exterior walls, interior walls, floor slabs, and windows. For example, the condition data is data that sets a condition that a value indicating the degree to which it is recommended to place a bearing wall in a location where it is possible to place a bearing wall because an exterior wall or interior wall exists is multiplied by a coefficient "1". Also, for example, the condition data is data that sets a condition that a value indicating the degree to which it is recommended to place a bearing wall in a location where it is impossible to place a bearing wall because an exterior wall or interior wall does not exist is multiplied by a coefficient "0". Alternatively, the condition data is data that sets a condition that at least one bearing wall is placed for each of the exterior walls and interior walls along a first direction, and at least one bearing wall is placed for each of the exterior walls and interior walls along a second direction intersecting the first direction. Note that the condition data may be data that indicates conditions imposed on at least one of the shape, dimensions, material, and position of the bearing wall, for example.

In addition, the prerequisites indicated by the condition data may be that, if the structure is a building and it is desired to appropriately place the load-bearing walls, "0.9" may be given as the preferred load-bearing wall test value, and "0.0" may be given as the preferred value for the eccentricity ratio of each floor of the building calculated after the load-bearing walls are placed. Note that the load-bearing wall test value referred to here is the ratio of the load-bearing wall's strength to the external force acting on the load-bearing wall.

Among the condition data, the conditions to be satisfied when designing the structure of a structure may be, if the structure is a building and it is desired to properly position load-bearing walls, "1.0" may be given as the upper limit of the load-bearing wall test value, and values such as "0.3" or "0.15" may be given as the upper limit of the eccentricity ratio of each floor of the building calculated after the load-bearing walls are placed.

As shown in FIG. 2, after acquiring the condition data, the condition data acquisition function 101 causes at least a portion of the condition data to be referenced by the first model 102, the second model 105, the structural calculation function 104, the third model 108, the first error calculation function 106, and the second error calculation function 109.

The first model 102 designs the structure of a structure as the machine learning device shown in FIG. 2 under the conditions indicated by the condition data, and generates generated data indicating the structure of the structure. For example, the first model 102 designs the structure of a structure using a neural network to which the condition data is input, and generates the generated data shown in FIGS. 3 to 6. The first model 102 is also linked to a third model 108 with fixed weights, and learns using second error data, which will be described later. The first model 102 may further include at least one of a neural network and an attention mechanism. The first model 102 is a model that can be learned, i.e., a model whose weights can be updated.

3 is a diagram showing an example of generated data generated at the stage of generating a trained model by the structural design support program according to the embodiment, showing a position determined to be preferable for arranging a load-bearing wall in the X direction on the first floor of a building. FIG. 4 is a diagram showing an example of generated data generated at the stage of generating a trained model by the structural design support program according to the embodiment, showing a position determined to be preferable for arranging a load-bearing wall in the X direction on the second floor of a building. FIG. 5 is a diagram showing an example of generated data generated at the stage of generating a trained model by the structural design support program according to the embodiment, showing a position determined to be preferable for arranging a load-bearing wall in the Y direction on the first floor of a building. FIG. 6 is a diagram showing an example of generated data generated at the stage of generating a trained model by the structural design support program according to the embodiment, showing a position determined to be preferable for arranging a load-bearing wall in the Y direction on the second floor of a building. The generated data shown in FIGS. 3 to 6 show a position determined to be preferable for arranging a load-bearing wall by the first model when the building related to the condition data is viewed from above, and the closer the position is to black, the more strongly it is recommended to place the load-bearing wall.

The first model 102 may also separate the process of generating the generated data shown in FIG. 3 from the process of generating the generated data shown in FIG. 4. Similarly, the first model 102 may separate the process of generating the generated data shown in FIG. 5 from the process of generating the generated data shown in FIG. 6. That is, the first model 102 may separate the process of generating the generated data for the bearing walls arranged along the X-direction defined in the building from the process of generating the generated data for the bearing walls arranged along the Y-direction defined in the building and intersecting the X-direction. The X-direction is an example of a first direction defined in the structure. The bearing walls arranged along the X-direction are components of the structure and are an example of a first component arranged along the first direction. Similarly, the Y-direction is an example of a second direction defined in the structure. The bearing walls arranged along the Y-direction are components of the structure and are an example of a second component arranged along the second direction.

As shown in FIG. 2, the generated data acquisition function 103 acquires generated data by inputting condition data into the first model 102, and inputs the generated data into the structural calculation function 104, the second model 105, and the third model 108.

The structural calculation function 104 performs structural calculations using condition data and generated data as input, and outputs structural calculation result data that indicates the results of the structural calculations.

The second model 105 receives condition data and generated data as input and outputs first structure calculation result estimation data. The second model 105 may also include at least one of a neural network and an attention mechanism. The second model 105 is a learnable model, i.e., a model whose weights can be updated.

The first error calculation function 106 receives the structural calculation result data and the first structural calculation result estimation data as input, and generates and outputs first error data.

The second model 105 updates the weights using the backpropagation method based on the first error data. This allows the second model 105 to learn as a function close to that of a structural calculation device having a structural calculation function 104.

The weight data acquisition function 107 acquires weight data indicating the weight of the second model 105 from the second model 105.

The third model 108 copies the weights indicated by the weight data acquired from the second model 105 as the third model 103's own weights. Therefore, the third model 108 has the same configuration and the same weights as the second model due to the weight data. The third model 108 also generates and outputs second structure calculation result prediction data from the generated data. The third model 108 is a model that cannot be learned, i.e., a model whose weights cannot be updated.

The model linking the first model 102 and the third model 108 receives condition data as input and outputs second structure calculation result estimation data. In the following description, the model linking the first model 102 and the third model 108 may be referred to as the linked model.

The second error calculation function 109 calculates and outputs second error data from at least a portion of the condition data and the second structural calculation result estimation data. Here, at least a portion of the condition data is a desired evaluation result from the condition data.

The weights of the linked model are updated using the backpropagation method based on the second error data. Here, since the weights of the third model 108 among the linked models are fixed, the weights of only the first model 102 are updated.

By repeating the above-mentioned series of processes, the structural design support program 10 causes the second model 105 to output data that is closer to that of a structural calculation device having a structural calculation function 104, and causes the first model 102 to output generated data that can provide more preferable structural calculation results.

Here, a system can be considered in which a structural calculation device having a structural calculation function 104 learns the first model 102 using the backpropagation method from error data calculated from the difference between the structural calculation result and the desired evaluation result. However, in order to realize such a system, it is necessary to rewrite the program of the structural calculation device as a function to which the backpropagation method can be applied, and the cost of modifying the program is high. Therefore, by using the above-mentioned method, it is possible to provide thorough design support while maintaining existing program assets. Since it is possible to suppress the introduction cost, it has high industrial value.

A method called a generative adversarial network is known as a similar generative model. A generative adversarial network combines a generative model that generates an image with a judgment model that judges its truth and falsehood, and the generative model and judgment model learn while maintaining a balance. It is known that high-quality generated images can be obtained by appropriately setting hyperparameters, etc. In the present invention, this mechanism is applied, and the system is composed of a generative model that obtains generated data using design conditions as input, a system that evaluates the performance of the generated data, and an evaluation result estimation model that estimates the evaluation of the generated data using the evaluation result as training data. Here, a model with fixed weights for the evaluation result estimation model is linked to the generative model, and the generative model is trained using the desired evaluation result as training data. This allows the generative model to be trained to obtain a more desirable evaluation result. In other words, instead of the truth value of the correct label of the generative adversarial network, a multi-value regression problem is created with a desirable evaluation index value as the goal.

Figure 7 is a conceptual diagram showing an example of a process executed by the trained model acquisition function according to the embodiment. As shown in Figure 7, the trained model acquisition function 110 performs at least one of a process of acquiring the first model 102 as a first trained model and a process of acquiring the second model 105 as a second trained model.

Next, with reference to Figures 8 to 18, we will explain the loss function value, eccentricity ratio, and load-bearing wall test value at the stage when the structural design support program 10 generates a trained model using the above-mentioned method.

Figure 8 is a diagram showing an example of the relationship between the number of epochs of the machine learning device and the value of the loss function by the first error calculation function at the stage when the structural design support program according to the embodiment generates a trained model. The horizontal axis of Figure 8 indicates the number of epochs. The vertical axis of Figure 8 indicates the value of the loss function when a model in which the first model 102 and the third model 108 are linked is trained using training data.

The number of epochs referred to here is the number of times the above-mentioned processes are executed by the condition data acquisition function 11, the first model 102, the generated data acquisition function 103, the structural calculation function 104, the second model 105, the first error calculation function 106, the weight data acquisition function 107, the third model 108 and the second error calculation function 109.

The loss function value referred to here is the value of the loss function calculated by the loss function incorporated in the second error calculation function by comparing the second structural calculation result estimated data with the desired structural calculation result from the structural calculation result data, as described above. The dotted line shown in FIG. 8 represents the value of the loss function for each epoch number. The solid line shown in FIG. 8 represents the moving average of the loss function values represented by each point on the dotted line shown in FIG. 8. As shown in FIG. 8, as the number of epochs increases, the value of the loss function decreases, and it can be seen that the learning of the first model 102 is progressing. Here, the first model 102 and the third model 108 are linked and trained. However, since the weight of the third model 108 is fixed, only the first model 102 is trained.

Figure 9 is a diagram showing an example of the relationship between the number of epochs of the machine learning device at the stage when the structural design support program according to the embodiment generates a trained model and the value of the loss function by the second error calculation function. In Figure 9, the horizontal axis shows the number of epochs, and the vertical axis shows the value of the loss function when the second model 105 is trained.

The number of epochs referred to here is the number of times the above-mentioned processes are executed by the condition data acquisition function 11, the first model 102, the generated data acquisition function 103, the structural calculation function 104, the second model 105, the first error calculation function 106, the weight data acquisition function 107, the third model 108 and the second error calculation function 109.

The loss function value referred to here is the value of the loss function calculated by comparing the structural calculation result data with the first structural calculation result estimated data and using the loss function built into the first error calculation function. The dotted line shown in FIG. 9 represents the value of the loss function for each number of epochs. Furthermore, the solid line shown in FIG. 9 represents the moving average of the loss function values represented by each point on the dotted line shown in FIG. 9. As shown in FIG. 9, as the number of epochs increases, the value of the loss function decreases, indicating that the learning of the second model 105 is progressing appropriately.

Figure 10 is a diagram showing an example of the relationship between the number of epochs of the machine learning device and the value of the loss function by the second error calculation function when test data is input at the stage when the structural design support program according to the embodiment generates a trained model. The horizontal axis of Figure 10 indicates the number of epochs when test data is input. The vertical axis of Figure 10 indicates the value of the loss function when the model in which the first model 102 and the third model 108 are linked generates generation data using test data and outputs second structural calculation result estimation data.

The number of epochs referred to here is the number of times the above-mentioned processes are executed by the condition data acquisition function 11, the first model 102, the generated data acquisition function 103, the structural calculation function 104, the second model 105, the first error calculation function 106, the weight data acquisition function 107, the third model 108 and the second error calculation function 109.

The loss function value referred to here is the value of the loss function calculated by the loss function incorporated in the second error calculation function 109 when test data is input and the second structural calculation result estimated data is compared with a desirable structural calculation result from among the structural calculation result data.

The dotted line in Fig. 10 represents the loss function value for each number of epochs. The solid line in Fig. 10 represents the moving average of the loss function values represented by each point on the dotted line in Fig. 10. As shown in Fig. 10, when the number of epochs exceeds 5k, which represents 5,000, the value of the loss function decreases as the number of epochs increases. The test data is not included in the training data used for learning. In other words, it can be seen that appropriate inferences can be made even for unknown data, and so-called generalization ability has been acquired.

11 is a diagram showing an example of the relationship between the eccentricity estimated by the third model for the X direction of the first floor of a building related to the learning of the first model and the second model at the stage when the structural design support program according to the embodiment generates the first model and the second model as trained models, and the number of epochs. FIG. 12 is a diagram showing an example of the relationship between the eccentricity estimated by the third model for the X direction of the second floor of a building related to the learning of the first model and the second model at the stage when the structural design support program according to the embodiment generates the first model and the second model as trained models, and the number of epochs. FIG. 13 is a diagram showing an example of the relationship between the eccentricity estimated by the third model for the Y direction of the first floor of a building related to the learning of the first model and the second model at the stage when the structural design support program according to the embodiment generates the first model and the second model as trained models, and the number of epochs. FIG. 14 is a diagram showing an example of the relationship between the eccentricity estimated by the third model for the Y direction of the second floor of a building related to the learning of the first model and the second model at the stage when the structural design support program according to the embodiment generates the first model and the second model as trained models, and the number of epochs.

The horizontal axis in Figures 11 to 14 indicates the number of epochs. The number of epochs here refers to the number of times the above-mentioned processes are executed by the condition data acquisition function 11, the first model 102, the generation data acquisition function 103, the structural calculation function 104, the second model 105, the first error calculation function 106, the weight data acquisition function 107, the third model 108, and the second error calculation function 109.

The vertical axis in Figures 11 to 14 shows the eccentricity of the structure calculated by the second model 105 as a trained model. The eccentricity is the ratio of the distance between the center of gravity and the center of rigidity of the structure to the torsional resistance.

The dotted lines shown in Figures 11 to 14 all represent the eccentricity for each number of epochs. Furthermore, the solid lines shown in Figures 11 to 14 all represent the moving average of the eccentricity represented by each point on the dotted lines shown in Figures 11 to 14. As shown in Figures 11 to 14, it can be seen that as the number of epochs increases, the eccentricity decreases slightly, generally within a range of 0.3 or less. In other words, with reference to Figures 11 to 14, it can be seen that as the structural design support program 10 progresses with the learning of the second model 105, it becomes possible to generate a second model 105 capable of outputting structural design data showing a structure with a small eccentricity.

15 is a diagram showing an example of the relationship between the number of epochs and the load-bearing wall test value estimated by the third model for the X direction of the first floor of a building related to the learning of the first model and the second model at the stage when the structural design support program according to the embodiment generates the first model and the second model as trained models. FIG. 16 is a diagram showing an example of the relationship between the number of epochs and the load-bearing wall test value estimated by the third model for the X direction of the second floor of a building related to the learning of the first model and the second model at the stage when the structural design support program according to the embodiment generates the first model and the second model as trained models. FIG. 17 is a diagram showing an example of the relationship between the number of epochs and the load-bearing wall test value estimated by the third model for the Y direction of the first floor of a building related to the learning of the first model and the second model at the stage when the structural design support program according to the embodiment generates the first model and the second model as trained models. FIG. 18 is a diagram showing an example of the relationship between the number of epochs and the load-bearing wall test value estimated by the third model for the Y direction of the second floor of a building related to the learning of the first model and the second model at the stage when the structural design support program according to the embodiment generates the first model and the second model as trained models.

15 to 18, the horizontal axis indicates the number of epochs, and the vertical axis indicates the load-bearing wall test value of the structure of the structure calculated by the second model 105. The dotted lines shown in Figs. 15 to 18 all represent the load-bearing wall test value for each epoch number. In addition, the solid lines shown in Figs. 15 to 18 all represent the moving average of the load-bearing wall test value represented by each point on the dotted lines shown in Figs. 15 to 18. The load-bearing wall test value is the ratio between the magnitude of the load-bearing wall's strength and the magnitude of the external force applied to the load-bearing wall, and when the load-bearing wall is insufficient, the value exceeds 1, and when the load-bearing wall B is too much, the value becomes small, and it is preferable that the value is about 0.9. As shown in Figs. 15 to 18, it can be seen that the load-bearing wall test value approaches a value less than 1 as the number of epochs increases. In other words, referring to Figures 15 to 18, it can be seen that as the structural design support program 10 progresses with the learning of the second model 105, it becomes possible to generate a learned model capable of outputting structural design data that shows a structure in which the load-bearing wall test value shows an appropriate value.

In addition, when the learning of the trained model used by the second model 105 has progressed to a certain extent, the structural design support program 10 may extract condition data from a specific building and re-train the second model 105 using the condition data. The condition data is acquired by the condition data acquisition function 11.

Figure 19 is a diagram showing an example of a two-story building according to an embodiment. The building H shown in Figure 19 is an example of the above-mentioned structure, and includes an exterior wall E, an interior wall G, floor slabs F1, F2, F3, windows W1 and W2. The floor slab F1 is a plate-shaped member that forms the floor structure of the first floor of the building H and supports a load perpendicular to the floor. The floor slab F2 is a plate-shaped member that forms the floor structure of the second floor of the building H and supports a load perpendicular to the floor. The floor slab F3 is a plate-shaped member that forms the roof structure of the building H and supports a load perpendicular to the roof. The window W1 is a window installed in the first floor part of the exterior wall E of the building H. The window W2 is a window installed in the second floor part of the exterior wall E of the building H. In addition, the building H has a load-bearing wall B installed in at least one part of the interior wall G and the part of the exterior wall E where the windows W1 and W2 are not installed.

The first model 102, as a machine learning device, designs the structure of the building H shown in FIG. 19 under the conditions indicated by the condition data, and generates generated data indicating the structure of the structure. For example, the first model 102 generates the structural data shown in FIGS. 20 to 23.

20 is a diagram showing an example of generated data indicating a position where the structural design support program according to the embodiment judges that it is preferable to place a load-bearing wall in the X direction of the first floor of the building shown in FIG. 19. FIG. 21 is a diagram showing an example of generated data indicating a position where the structural design support program according to the embodiment judges that it is preferable to place a load-bearing wall in the X direction of the second floor of the building shown in FIG. 19. FIG. 22 is a diagram showing an example of generated data indicating a position where the structural design support program according to the embodiment judges that it is preferable to place a load-bearing wall in the Y direction of the first floor of the building shown in FIG. 19. FIG. 23 is a diagram showing an example of generated data indicating a position where the structural design support program according to the embodiment judges that it is preferable to place a load-bearing wall in the Y direction of the second floor of the building shown in FIG. 19. The generated data shown in FIG. 20 to FIG. 23 indicate positions where the structural design support program judges that it is preferable to place a load-bearing wall in the X direction of the first floor of the building shown in FIG. 19 when the building H shown in FIG. 19 is viewed from above, and the closer the position is to black, the more strongly it is recommended to place a load-bearing wall.

The first model 102 may also separate the process of generating the structure data shown in FIG. 20 from the process of generating the structure data shown in FIG. 21. Similarly, the structure data generation function 12 may separate the process of generating the structure data shown in FIG. 22 from the process of generating the structure data shown in FIG. 23.

When performing re-learning, the structural design support program 10 executes processing by the condition data acquisition function 11, the first model 102, the generated data acquisition function 103, the structural calculation function 104, the second model 105, the first error calculation function 106, the weight data acquisition function 107, the third model 108, and the second error calculation function 109, as well as processing by the learned model acquisition function 110, in the same manner as when performing the above-mentioned learning.

Next, with reference to Figures 24 to 33, we will explain the loss function value, eccentricity ratio, and load-bearing wall test value at the stage when the structural design support program 10 re-learns the trained model using the above-mentioned method.

24 is a diagram showing an example of the relationship between the number of epochs of a machine learning device and the value of a loss function at the stage when the structural design support program according to the embodiment re-learns a trained model used by the machine learning device. The horizontal axis of FIG. 24 indicates the number of epochs. The vertical axis of FIG. 24 indicates the value of the loss function when the second model 105 is re-learned. The number of epochs and the value of the loss function are the same as those in FIG. 9. The dotted line shown in FIG. 24 represents the value of the loss function for each number of epochs. The solid line shown in FIG. 24 represents the moving average of the values of the loss function represented by each point on the dotted line shown in FIG. 24. As shown in FIG. 24, as the number of epochs increases, the value of the loss function decreases, and it can be seen that the re-learning of the second model 105 is proceeding appropriately.

25 is a diagram showing an example of the relationship between the number of epochs and the value of the loss function of a machine learning device at the stage when the structural design support program according to the embodiment re-learns the trained model. The horizontal axis of FIG. 25 indicates the number of epochs. The vertical axis of FIG. 25 indicates the value of the loss function of the trained model. The number of epochs and the value of the loss function are the same as those in FIG. 8. The dotted line shown in FIG. 25 represents the value of the loss function for each number of epochs. The solid line shown in FIG. 25 represents the moving average of the values of the loss function represented by each point on the dotted line shown in FIG. 25. As shown in FIG. 25, as the number of epochs increases, the value of the loss function decreases, and it can be seen that the re-learning of the first model 102 is proceeding appropriately.

Figure 26 is a diagram showing an example of the relationship between the eccentricity estimated by a trained model undergoing re-learning for the X direction on the first floor of the building shown in Figure 19 and the number of epochs. Figure 27 is a diagram showing an example of the relationship between the eccentricity estimated by a trained model undergoing re-learning for the X direction on the second floor of the building shown in Figure 19 and the number of epochs. Figure 28 is a diagram showing an example of the relationship between the eccentricity estimated by a trained model undergoing re-learning for the Y direction on the first floor of the building shown in Figure 19 and the number of epochs. Figure 29 is a diagram showing an example of the relationship between the eccentricity estimated by a trained model undergoing re-learning for the Y direction on the second floor of the building shown in Figure 19 and the number of epochs.

The horizontal axis in Figs. 26 to 29 indicates the number of epochs. The vertical axis in Figs. 26 to 29 indicates the eccentricity of the structure calculated by the trained model in which re-learning is in progress. The dotted lines in Figs. 26 to 29 all indicate the eccentricity for each number of epochs. The solid lines in Figs. 26 to 29 all indicate the moving average of the eccentricity represented by each point on the dotted lines in Figs. 26 to 29. As shown in Figs. 26 to 29, it can be seen that the eccentricity decreases as the number of epochs increases. In other words, referring to Figs. 26 to 29, it can be seen that as the structural design support program 10 progresses in re-learning the second model 105, it becomes possible to generate a second model 105 capable of outputting structural design data showing a structure with a small eccentricity.

Figure 30 is a diagram showing an example of the relationship between the number of epochs and the load-bearing wall test value estimated by a trained model undergoing re-learning for the X direction of the first floor of the building shown in Figure 19. Figure 31 is a diagram showing an example of the relationship between the number of epochs and the load-bearing wall test value estimated by a trained model undergoing re-learning for the X direction of the second floor of the building shown in Figure 19. Figure 32 is a diagram showing an example of the relationship between the number of epochs and the load-bearing wall test value estimated by a trained model undergoing re-learning for the Y direction of the first floor of the building shown in Figure 19. Figure 33 is a diagram showing an example of the relationship between the number of epochs and the load-bearing wall test value estimated by a trained model undergoing re-learning for the Y direction of the second floor of the building shown in Figure 19.

30 to 33, the horizontal axis indicates the number of epochs, and the vertical axis indicates the load-bearing wall test value of the structure of the structure calculated by the trained model in which re-learning is in progress. The dotted lines shown in Figs. 30 to 33 all represent the load-bearing wall test value for each epoch number. In addition, the solid lines shown in Figs. 30 to 33 all represent the moving average of the load-bearing wall test value represented by each point on the dotted lines shown in Figs. 30 to 33. As shown in Figs. 30 to 33, it can be seen that as the number of epochs increases, the load-bearing wall test value approaches a value less than 1. In other words, referring to Figs. 30 to 33, it can be seen that as the structural design support program 10 progresses in re-learning the second model 105, it becomes possible to generate a second model 105 capable of outputting structural design data showing a structure in which the load-bearing wall test value shows an appropriate value.

Next, an example of processing executed by the structural design support program 10 according to the embodiment will be described with reference to FIG. 34. FIG. 34 is a flowchart showing an example of processing executed by the structural design support program according to the embodiment in the learning phase. Each time the processing shown in FIG. 34 is executed once, the number of epochs described above is incremented by 1.

In step S10, the condition data acquisition function 11 acquires condition data indicating the conditions that should be satisfied when designing the structure of the structure and the prerequisites that should be referred to when designing the structure of the structure.

In step S20, the first model 102 designs the structure of the structure as a machine learning device under the conditions indicated by the condition data, and generates generated data indicating the structure of the structure.

In step S30, the generated data acquisition function 103 acquires the generated data generated by the first model.

In step S40, the structural calculation function 104 outputs structural calculation result data using a structural calculation device based on the generated data and the condition data.

In step S50, the second model 105 generates first structural calculation result estimation data from the generated data.

In step S60, the first error calculation function 106 generates and outputs first error data from the structural calculation result data and the first structural calculation result estimation data.

In step S70, the second model 105 learns by updating the weights of the second model from the first error data.

In step S80, the weight data acquisition function 107 acquires weight data from the second model.

In step S90, the third model 108 copies the weight data to the third model.

In step S100, the second error calculation function 109 calculates second error data from at least a portion of the condition data and the second structural calculation result estimation data.

In step S110, the concatenated model learns about the concatenated model that concatenates the first model and the third model by updating only the weights of the first model in the concatenated model from the second error data.

In step S120, the structural design support program 10 determines whether the termination condition is satisfied. If the structural design support program 10 determines that the termination condition is satisfied (step S120: YES), it terminates the process. On the other hand, if the structural design support program 10 determines that the termination condition is not satisfied (step S120: NO), it returns the process to step S10.

The structural design support program 10 according to the embodiment has been described above. The structural design support program 10 includes a condition data acquisition function 101, a first model 102, a generated data acquisition function 103, a structural calculation function 104, a second model 105, a first error calculation function 106, a weight data acquisition function 107, a third model 108, a second error calculation function 109, and a trained model acquisition function 110.

The condition data acquisition function 101 acquires condition data indicating the conditions to be satisfied when designing the structure of a structure and the prerequisites to be referenced when designing the structure of a structure. The first model 102 designs the structure of a structure as a machine learning device under the conditions indicated by the condition data, generates generated data indicating the structure of the structure, is linked to a third model with fixed weights, and learns using error data. The generated data acquisition function 103 acquires the generated data generated by the first model. The structural calculation function 104 outputs structural calculation result data based on the generated data and the condition data.

The second model 105 generates first structural calculation result estimation data from the generated data and learns using the first error data. The first error calculation function 106 generates and outputs first error data from the structural calculation result data and the first structural calculation result estimation data. The weight data acquisition function 107 acquires weight data from the second model. The third model 108 has the same configuration and the same weights as the second model using the weight data, and generates and outputs second structural calculation result estimation data from the generated data. The second error calculation function 109 calculates second error data from at least a portion of the condition data and the second structural calculation result estimation data. The learned model acquisition function 110 performs at least one of a process of acquiring the first model 102 as a first learned model and a process of acquiring the second model 105 as a second learned model.

Therefore, the structural design support program 10 can provide more thorough support for the work of designing the structure of a structure.

The structural design support program 10 also separates the process of generating generated data for a first component that is arranged along a first direction defined in the structure from the process of generating generated data for a second component that is defined in the structure and arranged along a second direction that intersects with the first direction. This allows the structural design support program 10 to use convolution in at least a portion of the process shown in FIG. 9, for example, to avoid the structure shown by the generated data becoming unclear even if there are locations where the first component arranged along the first direction and the second component arranged along the second direction are in contact or close to each other.

In the above embodiment, the functions shown in FIG. 1 are implemented by a hardware processor that reads and executes the structural design support program 10, but the present invention is not limited to this.

At least some of the functions shown in FIG. 1 may be realized by hardware including circuitry such as a large-scale integration (LSI), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or a graphics processing unit (GPU). Alternatively, at least some of the functions shown in FIG. 1 may be realized by a combination of software and hardware. Furthermore, these pieces of hardware may be integrated into one, or may be divided into multiple pieces.

In addition, when at least a part of the functions shown in FIG. 1 are realized using hardware, at least one of the components shown in FIG. 2 may be incorporated into the hardware. Furthermore, at least one of the components shown in FIG. 2 may be realized as software.

The above-mentioned neural network may also be implemented with an attention mechanism. In general, a convolutional neural network (CNN) has the advantage of being excellent at capturing partial features of data. On the other hand, a convolutional neural network has the disadvantage that, as long as partial features are included, an image may be mistaken for an appropriate image even if the overall configuration combining the partial features is inappropriate. To compensate for this disadvantage, a method is known in which a model that learns which parts of the data should be focused on is added to the learning model as an attention mechanism.

When an attention mechanism is implemented in the neural network included in the first model 102, the neural network can generate structural data while taking into account the balance of the structure of the entire structure. When an attention mechanism is implemented in the neural network included in the second model 105, the neural network can learn the weights indicated by the weight data while taking into account the balance of the structure of the entire structure. Furthermore, when an attention mechanism is implemented in the neural network included in the second model 105, the neural network can generate estimated data while taking into account the balance of the structure of the entire structure. In other words, when an attention mechanism is implemented in at least one of these neural networks, the structural design support program 10 can generate a learned model and structural design data while taking into account the balance of the structure of the entire structure using the attention mechanism.

The above describes the embodiments of the present invention in detail with reference to the drawings. However, the specific configurations of the embodiments of the present invention are not limited to the above-described embodiments, and at least one of various combinations, modifications, substitutions, and design changes may be added to the above-described embodiments without departing from the gist of the present invention.

10...Structural design support program, 101...Condition data acquisition function, 102...First model, 103...Generation data acquisition function, 104...Structural calculation function, 105...Second model, 106...First error calculation function, 107...Weight data acquisition function, 108...Third model, 109...Second error calculation function, 110...Trained model acquisition function

Claims (7)

On the computer,
a condition data acquisition function for acquiring condition data indicating conditions to be satisfied when designing a structure of a structure and prerequisites to be referred to when designing the structure of the structure;
A first model that designs the structure of the structure as a machine learning device under conditions indicated by the condition data, generates generation data indicating the structure of the structure, and is connected to a third model with fixed weights and learns using second error data;
A generated data acquisition function that acquires generated data generated by the first model;
a structural calculation function for outputting structural calculation result data based on the generated data and the condition data;
A second model that generates first structure calculation result estimation data from the generated data and learns using first error data;
a first error calculation function that generates and outputs the first error data from the structural calculation result data and the first structural calculation result estimation data;
a weight data acquisition function for acquiring weight data from the second model;
a third model that has the same configuration and the same weights as the second model using the weight data, and generates and outputs second structure calculation result estimation data from the generated data;
a second error calculation function for calculating the second error data from at least a part of the condition data and the second structure calculation result estimation data;
A structural design support program that makes this possible. 2. The structural design support program according to claim 1, wherein the first model separates a process of generating the generated data for a first component arranged along a first direction defined in the structure from a process of generating the generated data for a second component defined in the structure and arranged along a second direction that intersects with the first direction. 3. The structural design support program according to claim 1, wherein at least one of the first model, the second model and the third model includes at least one of a neural network and an attention mechanism. a condition data acquisition unit that acquires condition data indicating conditions that should be satisfied when designing a structure of a structure and prerequisites that should be referred to when designing the structure of the structure;
A first model that designs the structure of the structure as a machine learning device under conditions indicated by the condition data, generates generation data indicating the structure of the structure, and is connected to a third model with fixed weights and learns using second error data;
A generated data acquisition unit that acquires generated data generated by the first model;
a structural calculation unit that outputs structural calculation result data based on the generated data and the condition data;
A second model that generates first structure calculation result estimation data from the generated data and learns using first error data;
a first error calculation unit that generates and outputs the first error data from the structural calculation result data and the first structural calculation result estimation data;
a weight data acquisition unit that acquires weight data from the second model;
a third model that has the same configuration and the same weights as the second model using the weight data, and generates and outputs second structure calculation result estimation data from the generated data;
a second error calculation unit that calculates the second error data from at least a part of the condition data and the second structure calculation result estimation data;
A structural design support device comprising: The structural design support device according to claim 4, wherein the first model, whose weights have been updated based on the second error data , uses the condition data to design a structure of the structure, and outputs generated data indicating a result of designing the structure of the structure. obtaining condition data indicating conditions to be satisfied when designing a structure of a structure and prerequisites to be referred to when designing the structure of the structure;
A step in which a first model, which is a model that generates generation data indicating a structure of the structure under conditions indicated by the condition data and learns using second error data, designs the structure of the structure as a machine learning device and acquires the generated generation data;
a step of outputting structural calculation result data using a structural calculation device based on the generated data and the condition data ;
a step of generating and outputting first error data based on the first structural calculation result estimation data generated by a second model, which is a model that generates first structural calculation result estimation data from the generated data and learns using first error data, and the structural calculation result data;
learning by updating weights of the second model based on the first error data;
obtaining updated weight data from the second model;
a step of making the third model a model having the same weight as that of the second model by copying the weight data to a third model, the third model being a model having the same configuration as the second model, linked to the first model, and generating and outputting second structure calculation result estimation data from the generated data ;
calculating second error data from at least a part of the condition data and second structure calculation result estimation data;
a step of learning a concatenated model obtained by concatenating the first model and the third model by updating only the weights of the first model in the concatenated model from the second error data;
acquiring at least one of the first model and the second model as a trained model;
A structural design support method comprising: 7. The structural design support method according to claim 6, further comprising a step of: using the first model, whose weights have been updated based on the second error data , to design a structure of the structure using the condition data; and outputting generated data indicative of a result of designing the structure of the structure.

Images