Product design support system, product design support method, and program

Abstract

To provide a product design support system, a product design support method, and a program which can estimate a performance evaluation value with satisfactory accuracy.SOLUTION: A target specification data acquiring unit 240 acquires target specification data x2 indicating specification of a target tire. A target profile estimation unit 244 inputs the target specification data x2 into a leaned estimation model M, and estimates a target profile f2 based on an output from the leaned estimation model M. A target resonance frequency calculation unit 244, based on the target profile f2, calculates a target resonance frequency y2. The learned estimation model M is generated by a model generating unit 22 including a learning unit 226 for executing learning of an unlearned estimation model m, based on learning data D including learning specification data x3 indicting specification of a learning tire, and a correct-answer profile F.SELECTED DRAWING: Figure 6

Description

[Technical Field]

[0001] The present invention relates to a product design support system, a product design support method, and a program. [Background technology]

[0002] 2. Description of the Related Art Conventionally, a technique is known in which a performance evaluation value indicating the performance of a product is estimated using a machine learning model based on specification data indicating the specifications of the product. [Prior art documents] [Non-patent literature]

[0003] [Non-Patent Document 1] Iwayama M, Wu S, Liu C, Yoshida R. Functional Output Regression for Machine Learning in Materials Science. J Chem Inf Model. 2022 Oct 24;62(20):4837-4851. Summary of the Invention [Problem to be solved by the invention]

[0004] There is a demand for a technology that can estimate a performance evaluation value more accurately than the above-mentioned conventional technology that directly estimates a performance evaluation value. Note that Non-Patent Document 1 describes that the maximum absorption wavelength can be estimated more accurately by estimating the absorption spectrum of a substance and then indirectly calculating the maximum absorption wavelength from the estimated absorption spectrum, rather than directly estimating the maximum absorption wavelength of the substance.

[0005] The present invention has been made in view of the above-mentioned problems, and one of its objects is to provide a product design support system, a product design support method, and a program that can accurately estimate a performance evaluation value. [Means for solving the problem]

[0006] A product design support system according to one embodiment of the present invention comprises: a specification data acquisition means for acquiring specification data indicating product specifications; a product profile estimation means for inputting the specification data into a trained estimation model and estimating a product profile indicating an expected change in a specified physical quantity of the product when specified conditions are changed based on the output from the trained estimation model; and a performance evaluation value calculation means for calculating a performance evaluation value indicating the expected performance of the product based on the product profile, wherein the trained estimation model is generated by a model generation means including a learning means for executing training of an untrained estimation model based on training data including training specification data indicating the specifications of a training product and a ground truth profile indicating an actual change in the specified physical quantity of the training product when the specified conditions are changed. [Effects of the Invention]

[0007] According to the present invention, a performance evaluation value can be estimated with high accuracy. [Brief explanation of the drawings]

[0008] [Figure 1] 1 is a diagram illustrating an overview of a product design support system according to an embodiment of the present invention. [Figure 2] FIG. 10 is a diagram illustrating an outline of a resonance frequency calculation phase. [Figure 3] FIG. 1 is a diagram illustrating an overview of a model generation phase. [Figure 4] FIG. 10 is a diagram illustrating smoothing. [Figure 5] 1 is a diagram illustrating an example of a hardware configuration of a product design support system according to an embodiment of the present invention. [Figure 6] FIG. 1 is a functional block diagram showing an example of functions realized by a product design support system according to an embodiment of the present invention. [Figure 7] FIG. 10 is a flowchart illustrating an example of processing related to a model generation phase. [Figure 8] FIG. 10 is a flowchart illustrating an example of processing related to a resonance frequency calculation phase. DETAILED DESCRIPTION OF THE INVENTION

[0009] Hereinafter, an embodiment of the present invention will be described with reference to the drawings. In this embodiment, a case where a product design support system 1 (see FIG. 5, etc.) is used to estimate the resonance frequency of a tire will be described as an example. That is, in this embodiment, a case where a product is a tire and a performance table value is a resonance frequency will be described as an example.

[0010] [1. Overview of the product design support system according to this embodiment] First, an overview of a product design support system 1 according to this embodiment will be described.

[0011] [1-1. Overview (Part 1)] A technique for estimating a performance evaluation value indicating the performance of a product using a machine learning model based on specification data indicating the specifications of the product has been known. In this embodiment, as described above, as an example, it is considered that the resonance frequency of a tire is estimated using a machine learning model based on specification data indicating the specifications of the tire.

[0012] However, there is a demand for a technology that can estimate a performance evaluation value with higher accuracy than the conventional technology that directly estimates a performance evaluation value. Non-Patent Document 1 describes that the maximum absorption wavelength can be estimated more accurately by estimating the absorption spectrum of a substance and then indirectly calculating the maximum absorption wavelength from the estimated absorption spectrum, rather than directly estimating the maximum absorption wavelength of the substance.

[0013] Therefore, in the product design support system 1 according to this embodiment, the resonance frequency of the tire is not directly estimated, but a frequency response profile f as shown in Fig. 1 is estimated using a trained estimation model M, and the resonance frequency y is calculated based on the estimated frequency response profile f. Fig. 1 is a diagram illustrating an overview of the product design support system 1 according to the embodiment of the present invention.

[0014] According to the product design support system 1, the resonance frequency y is calculated based on the frequency response profile f estimated by the trained estimation model M, so that the resonance frequency of the tire can be estimated with high accuracy.

[0015] Below, an overview of the calculation of the resonance frequency (resonance frequency calculation phase) realized by the product design support system 1 will be described with reference to Fig. 2. Fig. 2 is a diagram showing an overview of the resonance frequency calculation phase. Note that in the following description, Fig. 1 may be referenced in explaining the target profile f2 (corresponding to the frequency response profile f) and the target resonance frequency y2 (corresponding to the resonance frequency y).

[0016] (1) Acquisition of target specification data As shown in Fig. 2, first, the product design support system 1 acquires target specification data x2 indicating the specifications of a target tire (S20). The target tire refers to a tire for which the resonance frequency is to be calculated. The tire specifications include, for example, the tire structure (dimensions, shape, etc.), the tire material, or the tire manufacturing method. In other words, the target specification data x2 is data indicating at least one of, for example, the target tire structure, the target tire material, and the target tire manufacturing method.

[0017] (2) Estimation of target profile Next, the product design support system 1 inputs the target specification data x2 to the trained estimation model M, and estimates the target profile f2 based on the output from the trained estimation model M (S22).

[0018] The target profile f2 is a frequency response profile predicted for the target tire. That is, the target profile f2 indicates the predicted change in the vibration transmissibility (unit: Hz) of the target tire when the frequency (unit: Hz) of the vibration applied to the target tire is changed (see Figure 1). That is, the target profile f2 is data predicted to be obtained when a vibration test is performed on the target tire. The frequency response profile may also be referred to as a frequency response function or a frequency transfer function.

[0019] Here, a description will be given of the internal processing of the trained estimation model M in S22. Specifically, the trained estimation model M executes the following processing in S220 and S222.

[0020] That is, the target specification data x2 is input to the trained neural network NN2, and a plurality of target coefficients β2 are output from the trained neural network NN2 (S220). Note that, although the present embodiment illustrates a case where a neural network is used to estimate a plurality of target coefficients β2 from the target specification data x2, the present invention is not limited to this example, and any known regression model such as a support vector machine or a decision tree may be used to estimate the plurality of target coefficients β2.

[0021] Then, an object profile f2 is calculated based on the plurality of object coefficients β2 and the plurality of kernel basis functions k (S222). Specifically, in S222, a weighted sum of the kernel basis functions k by the object coefficients β2 is calculated as the object profile f2. That is, the object profile f2 estimated by the trained estimation model M is a kernel function. In this embodiment, as an example, the kernel basis function k is a Gaussian function, but the kernel basis function k is not limited to this example, and various basis functions such as a polynomial function or a sigmoid function can be used.

[0022] (3) Calculation of target resonant frequency Finally, the product design support system 1 calculates a target resonance frequency y2, which is an expected resonance frequency of the target tire, based on the target profile f2 (S24). Specifically, the product design support system 1 calculates the frequency corresponding to the peak position of the target profile f2 as the target resonance frequency y2 (see FIG. 1). In this embodiment, the target resonance frequency y2 to be estimated is the first-order resonance frequency (F1). If multiple peaks appear in the target profile f2 (i.e., if multiple resonance frequencies exist), the frequency corresponding to the position of the lowest-frequency peak is calculated as the target resonance frequency y2.

[0023] [1-2. Overview (Part 2)] From here, an overview of the model generation phase, i.e., the generation of a trained estimation model M (training of an untrained estimation model m) used to calculate the resonant frequency, which is realized by the product design support system 1, will be described with reference to FIG. 3. FIG. 3 is a diagram showing an overview of the model generation phase. In this specification, "untrained" means that learning has not been completed. That is, for example, the untrained estimation model m is a machine learning model that has the same configuration (algorithm) as the trained estimation model M, but has unadjusted parameters that are not necessarily the same as the adjusted parameters of the trained estimation model M. In the following description, FIG. 1 may be referenced to explain the learned profile f3 (corresponding to the frequency response profile f) and the learned resonant frequency y3 (corresponding to the resonant frequency y).

[0024] (1) Acquisition of training data 3, first, the product design support system 1 acquires learning data D (S30). The learning data D includes learning specification data x3, a correct profile F, and a correct resonance frequency Y.

[0025] The learning specification data x3 is data indicating the specifications of a learning tire. A learning tire is a tire that is the subject of learning. Details of the learning specification data x3 are the same as those of the target specification data x2, so a description thereof will be omitted.

[0026] The correct profile F is the actual frequency response profile of the learning tire. That is, the correct profile F shows the actual change in the vibration transmissibility of the learning tire when the frequency of the vibration applied to the learning tire is changed (see Figure 1). That is, the correct profile F is data obtained by conducting a vibration test on the learning tire. The correct profile F is used as the correct data for the learning specification data x3.

[0027] The correct resonance frequency Y is the actual resonance frequency of the learning tire calculated based on the correct profile F. Specifically, the frequency corresponding to the peak position of the correct profile F is calculated as the correct resonance frequency Y.

[0028] (2) Smoothing of the correct profile Next, as shown in FIG. 4, the product design support system 1 performs a smoothing process on steep portions S in the supervised profile F that have a gradient equal to or greater than a predetermined magnitude, thereby obtaining a smoothed profile F' in which the steep portions S in the supervised profile F have been smoothed (S31). FIG. 4 is a diagram illustrating smoothing. In this embodiment, as an example, the product design support system 1 obtains the smoothed profile F' by applying a low-pass filter to the supervised profile F as the smoothing process. Note that the smoothing process is not limited to that using a low-pass filter, and various known methods such as spline approximation can be used as the smoothing process.

[0029] If an untrained estimation model m is trained to accurately estimate a frequency response profile including the steep portions, the difficulty of estimation increases, which may result in a decrease in the accuracy of the frequency response profile estimated by the trained estimation model M. In other words, if the ground truth profile F including the steep portions S is used as is as ground truth data to generate the trained estimation model M, the accuracy of estimation by the trained estimation model M may decrease. As a result, the accuracy of the resonant frequency calculated based on the estimated frequency response profile may also decrease.

[0030] Therefore, in this embodiment, as described above, a smoothed profile F' is obtained by smoothing the steep portions S in the ground truth profile F, and this smoothed profile F' is used as ground truth data to generate the estimation model M. This reduces the difficulty of estimation, improving the accuracy of the frequency response profile estimated by the trained estimation model M. As a result, the accuracy of the resonant frequency calculated based on the estimated frequency profile also improves.

[0031] (3) Training an untrained estimation model Then, the product design support system 1 executes learning of the unlearned estimation model m based on the learning specification data x3, the smoothed profile F', and the correct resonance frequency Y (S32).

[0032] However, since a frequency response profile generally has a complex waveform, the accuracy of the frequency response profile estimated by the trained estimation model M may not be sufficient. In this case, the accuracy of the resonant frequency calculated based on the estimated frequency response profile may also decrease.

[0033] In the process of examining this issue, the inventors of the present application noticed that when a portion of the frequency response profile used to calculate the resonant frequency is designated as a region of interest A and the region other than the region of interest A is designated as a region of non-interest NA (see Figure 1), there is no need to accurately estimate the non-region of interest NA of the smoothed profile F' that is not used to calculate the resonant frequency.

[0034] Therefore, in this embodiment, the learning of the unlearned estimation model m is performed so that the learning weight for the target region A of the smoothed profile F' is greater than the learning weight for the non-target region NA of the smoothed profile F'.

[0035] This allows for particularly accurate estimation of the region of interest A used to calculate the resonant frequency. Therefore, according to the product design support system 1, even when a frequency response profile having a generally complex waveform is estimated, the resonant frequency can be calculated with high accuracy based on the estimated frequency response profile.

[0036] Specifically, S32 includes the processes (S320 to S324) described in the following (3-1) to (3-3).

[0037] (3-1) Estimation of learning profile As shown in FIG. 3, the product design support system 1 inputs learning specification data x3 into an unlearned estimation model m, and estimates a learning profile f3 based on the output from the unlearned estimation model m (S320). The learning profile f3 is a frequency response profile predicted for the learning tire. That is, the learning profile f3 indicates the predicted change in the vibration transmissibility (unit: Hz) of the learning tire when the frequency (unit: Hz) of the vibration applied to the learning tire is changed (see FIG. 1). That is, the learning profile f3 is data predicted to be obtained when a vibration test is performed on the learning tire.

[0038] Specifically, the following processes S3200 and S3202 are executed for the unlearned estimation model M.

[0039] That is, target specification data x2 is input to an untrained neural network NN3, and multiple learning coefficients β3 are output from the untrained neural network NN3 (S3200). Then, a learning profile f3 is calculated based on the multiple learning coefficients β3 and multiple kernel basis functions k (S3202). The details of S3200 and S3202 are the same as those of S220 and S222, so a description thereof will be omitted.

[0040] (3-2) Calculation of learning resonance frequency The product design support system 1 calculates a learning resonance frequency y3, which is an expected resonance frequency of the learning tire, based on the learning profile f3 (S322). Details of S322 and the learning resonance frequency y3 are similar to those of S24 and the target resonance frequency y2, so a description thereof will be omitted.

[0041] (3-3) Parameter adjustment 3, the product design support system 1 adjusts the parameters of the unlearned estimation model m so as to reduce the first error L1 and the second error L2 (S324). The first error L1 is an error based on the smoothed profile F' and the learned profile f3. The second error L2 is an error based on the correct resonance frequency Y and the learned resonance frequency y3. As the error, various known loss functions (error functions) such as the mean square error and the mean absolute error can be used.

[0042] That is, the product design support system 1 calculates the first error L1 based on the smoothed profile F' and the learned profile f3 in S324 (S3240).

[0043] Specifically, the product design support system 1 calculates a focus error L based on the focus area F'-A of the smoothed profile F' and the focus area f3-A of the learning profile f3. A Furthermore, the product design support system 1 calculates a non-interest error L based on the non-interest region F'-NA of the smoothed profile F' and the non-interest region f3-NA of the learning profile f3. NA Calculate.

[0044] Then, the product design support system 1 calculates the target error L as shown in the following formula (1): A and the non-focus error L NA In the following formula (1), w is a weight having a value greater than 1. That is, in S324, the target error L A The weight for the non-focus error L NA The parameters of the unlearned estimation model m are adjusted so that the weight of the learning for the region of interest A of the smoothed profile F' is set to be larger than the weight of the learning for the region of non-interest NA of the smoothed profile F'. Note that the formula used to calculate the first error L1 is not limited to formula (1).

[0045]

number

[0046] Furthermore, in S324, the product design support system 1 calculates the second error L2 based on the correct resonance frequency Y and the learning resonance frequency y3 (S3242).

[0047] Then, in S324, the product design support system 1 calculates the sum L of the first error L1 and the second error L2, and executes training of the untrained estimation model m so as to reduce the sum L (S3244). In this manner, in this embodiment, the parameters of the untrained estimation model m are adjusted so as to reduce not only the first error L1 but also the second error L2, and therefore training of the untrained estimation model m is performed so as to enable even more accurate estimation of the resonant frequency.

[0048] The product design support system 1 will be described in detail below.

[0049] 2. Hardware Configuration of the Product Design Support System According to the Present Embodiment 5 is a diagram showing an example of the hardware configuration of a product design support system 1 according to an embodiment of the present invention. The product design support system 1 is a computer system made up of one or more computers. As shown in FIG. 5, the product design support system 1 includes a control unit 10, a storage unit 12, a communication unit 14, a display unit 16, and an operation unit 18.

[0050] The control unit 10 is a program-controlled device such as a CPU that operates according to a program stored in the memory unit 12. The memory unit 12 is, for example, a storage element such as a ROM or RAM, or a hard disk drive. The memory unit 12 stores programs and the like to be executed by the control unit 10. The communication unit 14 is a communication interface such as a network board or a wireless LAN module. The display unit 16 is a display such as a liquid crystal display or an organic EL display. The operation unit 18 is an input device such as a keyboard, a mouse, or a touch panel.

[0051] [3. Functions realized by the product design support system according to this embodiment] 6 is a functional block diagram showing an example of functions realized by the product design support system 1 according to the embodiment of the present invention. As shown in FIG. 6, the product design support system 1 functionally includes a model storage unit 20, a model generation unit 22, and a product performance estimation unit 24.

[0052] [3-1. Model storage section] The model storage unit 20 stores an unlearned estimation model m and a trained estimation model M. The model storage unit 20 is mainly implemented by the storage unit 12. Specifically, the model storage unit 20 stores the configurations (algorithms) and parameters of the unlearned estimation model m and the trained estimation model M. As described above, the configuration of the unlearned estimation model m and the trained estimation model M are common. The model generation unit 22, which will be described later, adjusts the parameters of the unlearned estimation model m.

[0053] [3-2. Model generation section] The model generation unit 22 executes learning of an unlearned estimation model m to generate a trained estimation model M. The model generation unit 22 acquires the unlearned estimation model m (more specifically, the parameters of the unlearned estimation model m) stored in the model storage unit 20, executes learning of the unlearned estimation model m, and generates a trained estimation model M. The model generation unit 22 then stores the generated trained estimation model M (more specifically, the parameters of the trained estimation model m) in the model storage unit 20. Note that the model generation unit 22 may acquire the unlearned estimation model m from an external device or information storage medium. Also, the model generation unit 22 may store the generated trained estimation model M in an external device or information storage medium.

[0054] 6 , the model generation unit 22 includes a training data acquisition unit 220, a training data storage unit 222, a smoothing unit 224, and a training unit 226. The training data acquisition unit 220, the smoothing unit 224, and the training unit 226 are mainly implemented by the control unit 10. The training data storage unit 222 is mainly implemented by the storage unit 12.

[0055] The learning data acquiring unit 220 acquires learning data D including learning specification data x3, a correct profile F, and a correct resonant frequency Y (see FIG. 3). In this embodiment, the learning data acquiring unit 220 acquires the learning data D stored in the learning data storage unit 222, but the learning data acquiring unit 220 may acquire the learning data D stored in an external device or information storage medium.

[0056] The smoothing unit 224 performs a smoothing process on steep portions S in the correct profile F that have a gradient equal to or greater than a predetermined magnitude, and obtains a smoothed profile F' in which the steep portions S in the correct profile F have been smoothed (see Figures 3 and 4).

[0057] The learning unit 226 learns an unlearned estimation model m based on the learning data D. In this embodiment, the learning unit 226 learns the unlearned estimation model m based on the learning specification data x3, a smoothed profile F′ obtained based on the correct profile F, and the correct resonance frequency Y.

[0058] As shown in FIG. 6, the learning unit 226 includes a learning profile estimation unit 2260, a learning resonance frequency calculation unit 2262, and a parameter adjustment unit 2264.

[0059] The learning profile estimation unit 2260 inputs the learning specification data x3 to the unlearned estimation model m, and estimates the learning profile f3 based on the output from the unlearned estimation model m (see FIG. 3).

[0060] The learned resonance frequency calculation unit 2262 calculates a learned resonance frequency y3, which is the predicted resonance frequency of the learning tire, based on the learned profile f3 (see FIG. 3).

[0061] The parameter adjusting unit 2264 adjusts the parameters of the unlearned estimation model m so that the first error L1 and the second error L2 become smaller (see FIG. 3).

[0062] [3-3. Product performance estimation section] The product performance estimation unit 24 calculates (estimates) the resonance frequency of a target tire. As shown in Fig. 6, the product performance estimation unit 24 includes a target specification data acquisition unit 240, a specification data storage unit 242, a target profile acquisition unit 244, and a target resonance frequency calculation unit 246. The target specification data acquisition unit 240, the target profile acquisition unit 244, and the target resonance frequency calculation unit 246 are mainly implemented by the control unit 10. The specification data storage unit 242 is mainly implemented by the storage unit 12.

[0063] The target specification data acquisition unit 240 acquires target specification data x2 indicating the specifications of the target tire (see FIG. 2). In this embodiment, the target specification data acquisition unit 240 acquires the target specification data x2 stored in the specification data storage unit 242, but the specification data storage unit 242 may acquire the target specification data x2 stored in an external device or information storage medium.

[0064] The target profile acquisition unit 244 inputs the target specification data x2 to the trained estimation model M, and estimates the target profile f2 based on the output from the trained estimation model M (see FIG. 2).

[0065] The target resonant frequency calculation unit 246 calculates a target resonant frequency y2, which is an expected resonant frequency of the target tire, based on the target profile f2 (see FIG. 2).

[0066] [3. Processing Executed in the Product Design Support System According to the Present Embodiment] The processing executed by the product design support system 1 will be described in detail below.

[0067] [3-1. Processing related to the model generation phase] 7 is a flow diagram showing an example of processing related to the model generation phase. As shown in FIG. 7, the control unit 10 of the product design support system 1 first acquires learning data D (S700). Next, the control unit 10 performs a smoothing process on the steep portion S of the correct profile F to acquire a smoothed profile F' (S701). The control unit 10 inputs learning specification data x3 to an unlearned estimation model m, and estimates a learning profile f3 based on the output from the unlearned estimation model m (S702). Next, the control unit 10 calculates a learning resonance frequency y3 based on the learning profile f3 (S703).

[0068] The control unit 10 calculates a first error L1 based on the smoothed profile F' and the learning profile f3, and calculates a second error L2 based on the correct resonance frequency Y and the learning resonance frequency y3 (S704). Then, the control unit 10 adjusts the parameters of the unlearned estimation model m so that the first error L1 and the second error L2 become smaller (S705). For parameter adjustment, a known parameter adjustment method (learning algorithm) such as backpropagation or gradient descent may be used. The control unit 10 determines whether to end the parameter adjustment (S706), and if it determines not to end the parameter adjustment (S706; N), it repeats the process of S705. If it determines to end the parameter adjustment (S706; Y), the control unit 10 ends this process.

[0069] In this embodiment, in S706, the control unit 10 determines whether to end the parameter adjustment based on, for example, whether the sum L of the first error L1 and the second error L2 is equal to or less than a predetermined threshold. Note that the method for determining whether to end the parameter adjustment is not limited to this example, and for example, the control unit 10 may determine whether to end the parameter adjustment based on whether the process of S705 has been executed a predetermined number of times.

[0070] In addition, in FIG. 7, the case where the processing is executed in the order of the smoothing processing (S701), the processing related to the estimation of the learning profile f3, and the processing related to the calculation of the learning resonance frequency y3 (S702 to S703) is shown, but the order in which S702 to S703 and S701 are executed may be reversed.

[0071] [3-2. Processing related to the resonance frequency calculation phase] 8 is a flow diagram showing an example of processing related to the resonant frequency calculation phase. As shown in FIG. 8, the control unit 10 first acquires target specification data x2 (S800). Next, the control unit 10 inputs the target specification data x2 to a trained estimation model M, and estimates a target profile f2 based on the output from the trained estimation model M (S801). Finally, the control unit 10 calculates a target resonant frequency y2, which is the predicted resonant frequency of the target tire, based on the target profile f2 (S802), and ends this processing.

[0072] [4. Modifications] It should be noted that the present invention is not limited to the above-described embodiment. The present invention can be modified as appropriate without departing from the spirit of the present invention. Furthermore, the specific character strings and numerical values ​​described above, as well as the specific character strings in the drawings, are merely examples, and the present invention is not limited to these character strings and numerical values.

[0073] For example, the tire performance evaluation value to be estimated is not limited to the resonance frequency, and may be any physical property value indicating tire performance. The performance evaluation value may be, for example, the tire's loss factor (tan δ) or rolling resistance coefficient. When the performance evaluation value is the tire's loss factor, a dynamic viscoelastic profile (e.g., the waveform of a sinusoidal strain applied to the tire and the waveform of a stress generated in response to the sinusoidal strain) may be estimated using an estimation model, and the loss factor may be calculated based on the estimated dynamic viscoelastic profile. Such a dynamic viscoelastic profile can be considered a product profile that indicates the expected change in strain or stress over time.

[0074] Furthermore, the product is not necessarily limited to a tire, but may be any product other than a tire.

[0075] In short, the product design support system comprises a specification data acquisition means for acquiring specification data indicating product specifications, a product profile estimation means for inputting the specification data into a trained estimation model and estimating a product profile indicating an expected change in a specified physical quantity of the product when specified conditions are changed based on the output from the trained estimation model, and a performance evaluation value calculation means for calculating a performance evaluation value indicating the expected performance of the product based on the product profile, and the trained estimation model may be generated by a model generation means including a learning means for executing training of an untrained estimation model based on training data including training specification data indicating the specifications of a training product and a correct answer profile indicating an actual change in the specified physical quantity of the training product when the specified conditions are changed.

[0076] Furthermore, the product design support system does not need to perform learning so that the learning weight for the target region of the correct profile (smoothed profile F' in this embodiment) is greater than the learning weight for the non-target region of the correct profile.

[0077] Furthermore, the product design support system does not need to include a configuration equivalent to the smoothing unit 224. That is, the training data may be used as is to train an untrained estimation model.

[0078] Furthermore, the product design support system does not need to calculate the learning performance evaluation value (the learning resonance frequency y3 in this embodiment) when generating a trained estimation model. That is, the product design support system does not need to adjust the parameters of the untrained estimation model so as to reduce the correct performance evaluation value and the second error based on the learning performance evaluation value (the correct resonance frequency Y in this embodiment).

[0079] In this embodiment, the training data includes a correct performance evaluation value, but the training data does not necessarily need to include a correct performance evaluation value. In this case, the product design support system may calculate the correct performance evaluation value based on the correct profile.

[0080] Furthermore, although the trained estimation model M in this embodiment is generated by the model generation unit 22 of the product design support system 1, the trained estimation model M may also be generated by an external device, etc.

[0081] [5. Summary] According to the present embodiment described above, the performance evaluation value is calculated based on the product profile estimated by the estimation model, so that the resonant frequency of the product can be estimated with high accuracy.

[0082] Furthermore, in this embodiment, the untrained estimation model is trained so that the learning weight for the region of interest in the correct profile is greater than the learning weight for the region of non-interest in the correct profile, so the region of interest used to calculate the performance evaluation value can be estimated with particularly high accuracy. Therefore, according to this embodiment, even when estimating a complex product profile, the performance evaluation value can be calculated with high accuracy based on the estimated product profile.

[0083] Furthermore, in this embodiment, a smoothed profile is obtained by smoothing the steep portions of the ground truth profile, and an untrained estimation model is trained based on this smoothed profile and the training specification data. This reduces the difficulty of estimation using the trained estimation model, thereby improving the accuracy of the product profile estimated by the trained estimation model. As a result, the accuracy of the performance evaluation value calculated based on the estimated product profile also improves.

[0084] In addition, in this embodiment, the unlearned estimation model parameters are adjusted so that not only the first error based on the correct profile and the learned product profile but also the second error based on the correct performance evaluation value and the learned performance evaluation value are reduced, so that the unlearned estimation model is trained so that the performance evaluation value can be estimated with even greater accuracy. [Explanation of symbols]

[0085] 1 Product design support system, 10 Control unit, 12 Memory unit, 14 Communication unit, 16 Display unit, 18 Operation unit, 20 Model memory unit, 22 Model generation unit, 24 Product performance estimation unit, 220 Learning data acquisition unit, 222 Learning data memory unit, 224 Smoothing unit, 226 Learning unit, 240 Target specification data acquisition unit, 242 Specification data storage unit, 244 Target profile acquisition unit, 246 Target resonance frequency calculation unit, 2260 Learning profile estimation unit, 2262 Learning resonance frequency calculation unit, 2264 Parameter adjustment unit, f Frequency response profile, y Resonance frequency, A Region of interest, NA Non-interest region, M Trained estimation model, x2 Target specification data, NN2 Trained neural network, β2 Target coefficient, k Kernel basis function, f2 Target profile, y2 Target resonance frequency, D Training data, m Untrained estimation model, x3 Learning specification data, NN3 untrained neural network, β3 learning coefficient, f3 learning profile, F correct answer profile, F' smoothed profile, L1 first error, L2 second error, L A Focus error, L NA Non-focus error, y3 learning resonance frequency, Y correct resonance frequency, S steepness.

Claims

1. a specification data acquisition means for acquiring specification data indicating product specifications; a product profile estimation means for inputting the specification data into a trained estimation model and estimating a product profile indicating an expected change in a predetermined physical quantity of the product when a predetermined condition is changed, based on an output from the trained estimation model; and a performance evaluation value calculation means for calculating a performance evaluation value indicating an expected performance of the product based on the product profile; and The trained estimation model is The model is generated by a model generation means including a learning means for executing learning of an unlearned estimation model based on learning data including learning specification data indicating the specifications of a learning product and a correct answer profile indicating an actual change in the predetermined physical quantity of the learning product when the predetermined condition is changed. Product design support system.

2. The learning means a learning product profile estimation means for inputting the learning specification data into the unlearned estimation model and estimating a learning product profile indicating an expected change in the predetermined physical quantity of the learning product when the predetermined conditions are changed, based on an output from the unlearned estimation model; a parameter adjusting means for adjusting parameters of the unlearned estimation model so that a first error based on the correct answer profile and the learned product profile becomes smaller; Including, The product design support system according to claim 1 .

3. The learning means When a part of the product profile used in calculating the performance evaluation value is designated as a region of interest and a region other than the region of interest is designated as a region of non-interest, the learning is performed so that a learning weight for the region of interest in the correct answer profile is greater than a learning weight for the region of non-interest in the correct answer profile. The product design support system according to claim 2.

4. The first error is A focus error based on the focus area of ​​the correct answer profile and the focus area of ​​the learning product profile; a non-interest error based on the non-interest region of the correct answer profile and the non-interest region of the learning product profile; contains a weighted sum of The parameter adjustment means setting a weight for the error of interest greater than a weight for the error of non-interest, and adjusting the parameters so that the first error becomes smaller; The product design support system according to claim 3.

5. A learning performance evaluation value calculation means for calculating a learning performance evaluation value indicating the expected performance of the learning product based on the learning product profile. and The learning data further includes a correct performance evaluation value calculated based on the correct profile and indicating an actual performance of the learning product; The parameter adjustment means further adjusting the parameters so that a second error based on the correct performance evaluation value and the learning performance evaluation value becomes small; 4. The product design support system according to claim 2 or 3.

6. The model generation means a smoothing means for performing a smoothing process on steep portions of the correct profile having a gradient equal to or greater than a predetermined magnitude, and obtaining a smoothed profile in which the steep portions of the correct profile have been smoothed; Further comprising: The learning unit learning an unlearned estimation model based on the learning specification data and the smoothing profile; 3. The product design support system according to claim 1.

7. a specification data acquisition step of acquiring specification data indicating product specifications; a product profile estimation step of inputting the specification data into a trained estimation model and estimating a product profile indicating an expected change in a predetermined physical quantity of the product when a predetermined condition is changed, based on an output from the trained estimation model; a performance evaluation value calculation step of calculating a performance evaluation value indicating the performance of the product based on the product profile; and The trained estimation model is The model is generated by a model generation means including a learning means for executing learning of an unlearned estimation model based on learning data including learning specification data indicating the specifications of a learning product and a correct answer profile indicating an actual change in the predetermined physical quantity of the learning product when the predetermined condition is changed. Product design support method.

8. a specification data acquisition means for acquiring specification data indicating the specifications of the product; a product profile estimation means for inputting the specification data into a trained estimation model and estimating a product profile indicating an expected change in a predetermined physical quantity of the product when a predetermined condition is changed, based on an output from the trained estimation model; a performance evaluation value calculation means for calculating a performance evaluation value indicating the performance of the product based on the product profile; A program for causing a computer to function as The trained estimation model is The model is generated by a model generation means including a learning means for executing learning of an unlearned estimation model based on learning data including learning specification data indicating the specifications of a learning product and a correct answer profile indicating an actual change in the predetermined physical quantity of the learning product when the predetermined condition is changed. program.

Images