Physical information-based artificial intelligence device and prediction model training method therefor

Abstract

The present invention relates to a physical information-based artificial intelligence device and a prediction model training method therefor, the device being capable of reducing training time and increasing prediction accuracy of a prediction model for predicting experimental results of a product, and comprising: a database for storing a dataset including a physical information function; and a processor for generating a prediction model that predicts experimental results of a product on the basis of the dataset, wherein the processor can: acquire, for each pre-selected initial design of experiment (DOE), a first dataset including a physical information function; train a neural network model to predict experimental results of a product by inputting the first dataset to the neural network model; select, on the basis of the training result, DOEs having an accuracy lower than a first reference value from among initial DOEs; extract, for each selected DOE, a design factor having a relative standard deviation (RSD) lower than a second reference value; acquire, for each new DOE including the extracted design factor, a second dataset including a physical information function; retrain the neural network model on the basis of the second dataset; and generate a prediction model on the basis of the retrained neural network model when the retraining result reaches a target error.

Description

Physical Information-Based Artificial Intelligence Device and Method for Learning Its Predictive Model

[0001] The present disclosure relates to a physical information-based artificial intelligence device and a method for training a prediction model therewith, which can reduce the training time of a prediction model for predicting experimental results of a product and increase prediction accuracy.

[0002] Generally, artificial intelligence is a field of computer science and information technology that studies methods to enable computers to perform thinking, learning, and self-development capable of human intelligence, and refers to making computers mimic intelligent human behavior.

[0003] Furthermore, artificial intelligence does not exist in isolation but is closely related, directly and indirectly, to many other fields of computer science. In particular, in the modern era, there are very active attempts to introduce AI elements into various fields of information technology and utilize them to solve problems within those areas.

[0004] Meanwhile, technologies that utilize artificial intelligence to perceive and learn surrounding situations, provide information desired by the user in the desired format, or perform actions or functions desired by the user are being actively researched.

[0005] And, electronic devices that provide these various operations and functions can be referred to as artificial intelligence devices.

[0006] Recently, artificial intelligence models are being applied to test the quality of manufactured products.

[0007] In the case of TV products, a drop impact test is performed for product quality testing, and an artificial intelligence model is used to predict the corresponding maximum stress and strain during a drop impact.

[0008] However, existing artificial intelligence models require large-scale Design of Experiments (DOEs) for training, and generating these DOEs requires significant time and resources, which can lead to major constraints during the actual design and development process.

[0009] Furthermore, since existing AI models do not reflect physical laws or governing equations, they lack physical consistency, which can lead to a problem where prediction accuracy drops sharply outside the range of training data.

[0010] Therefore, in the future, it is necessary to develop an artificial intelligence model training method that can increase prediction accuracy while minimizing the number of DOEs for training.

[0011] The present disclosure aims to solve the aforementioned problems and other problems.

[0012] The present disclosure aims to provide an artificial intelligence device and a method for training a prediction model that can generate a prediction model with high prediction accuracy and minimize training time by reducing the number of DOEs for training, by training a neural network model with a dataset containing physical information functions for each initial DOE to extract design factors with a low relative standard deviation (RSD) from DOEs with low accuracy, and retraining the neural network model to reach a target error with a second dataset containing new physical information functions for each DOE including the extracted design factors.

[0013] An artificial intelligence device according to one embodiment of the present disclosure includes a database that stores a dataset containing a physical information function, and a processor that generates a prediction model for predicting experimental results of a product based on the dataset. The processor acquires a first dataset containing a physical information function for each of the initial Design Of Experiment (DOE) selected in advance, inputs the first dataset into a neural network model to train the neural network model to predict experimental results of the product, selects DOEs among the initial DOEs whose accuracy is less than a first threshold value based on the training results, extracts design factors for each selected DOE whose relative standard deviation (RSD) is less than a second threshold value, acquires a second dataset containing a physical information function for each new DOE containing the extracted design factors, retrains the neural network model based on the second dataset, and when the retraining result reaches a target error, generates a prediction model based on the retrained neural network model.

[0014] A method for training a prediction model of an artificial intelligence device according to one embodiment of the present disclosure may include: acquiring a first dataset containing a physical information function for each of the initial DOEs selected in advance; inputting the first dataset into a neural network model to train the neural network model to predict the experimental results of a product; selecting a DOE among the initial DOEs whose accuracy is less than a first threshold value based on the training results; extracting a design factor for each selected DOE whose relative standard deviation (RSD) is less than a second threshold value; acquiring a second dataset containing a physical information function for each new DOE containing the extracted design factor; retraining the neural network model based on the second dataset; and generating a prediction model based on the retrained neural network model when the retraining result reaches a target error.

[0015] According to one embodiment of the present disclosure, an artificial intelligence device can generate a prediction model with high prediction accuracy and minimize training time by reducing the number of DOEs for training, by training a neural network model with a dataset containing physical information functions for each initial DOE to extract design factors with a low relative standard deviation (RSD) from DOEs with low accuracy, and retraining the neural network model to reach a target error with a second dataset containing physical information functions for each new DOE including the extracted design factors.

[0016] In addition, the present disclosure can increase data efficiency while maintaining physical consistency by training a neural network model with a dataset containing physical information functions.

[0017] FIG. 1 shows an artificial intelligence device according to one embodiment of the present disclosure.

[0018] FIG. 2 shows an artificial intelligence server according to one embodiment of the present disclosure.

[0019] FIG. 3 is a drawing for explaining the operation of an artificial intelligence device according to one embodiment of the present disclosure.

[0020] FIGS. 4 and FIGS. 5 are drawings for explaining the initial DOE selection process of an artificial intelligence device according to one embodiment of the present disclosure.

[0021] FIG. 6 is a diagram illustrating the process of selecting packaging design factor variables by DOE of an artificial intelligence device according to one embodiment of the present disclosure.

[0022] FIGS. 7 to 13 are drawings for explaining the process of acquiring a dataset including physical information functions for each DOE of an artificial intelligence device according to one embodiment of the present disclosure.

[0023] FIG. 14 is a diagram illustrating the neural network model learning process of an artificial intelligence device according to one embodiment of the present disclosure.

[0024] FIGS. 15 and 16 are drawings for explaining the process of selecting a new DOE based on the RSD criteria of an artificial intelligence device according to one embodiment of the present disclosure.

[0025] FIGS. 17 and 18 are drawings for explaining the reduction of the prediction model generation period of an artificial intelligence device according to one embodiment of the present disclosure.

[0026] FIG. 19 is a drawing for explaining the overall operation flow of an artificial intelligence device according to one embodiment of the present disclosure.

[0027] Hereinafter, embodiments disclosed in this specification will be described in detail with reference to the attached drawings. Identical or similar components, regardless of drawing symbols, are assigned the same reference number, and redundant descriptions thereof will be omitted. The suffixes "module" and "part" used for components in the following description are assigned or used interchangeably solely for the ease of drafting the specification and do not inherently possess distinct meanings or roles. Furthermore, in describing embodiments disclosed in this specification, if it is determined that a detailed description of related prior art could obscure the essence of the embodiments disclosed in this specification, such detailed description will be omitted. Additionally, the attached drawings are intended only to facilitate understanding of the embodiments disclosed in this specification; the technical concept disclosed in this specification is not limited by the attached drawings, and it should be understood that they include all modifications, equivalents, and substitutions that fall within the concept and technical scope of this disclosure.

[0028] Terms including ordinal numbers, such as first, second, etc., may be used to describe various components, but said components are not limited by said terms. These terms are used solely for the purpose of distinguishing one component from another.

[0029] When it is stated that one component is "connected" or "connected" to another component, it should be understood that while it may be directly connected or connected to that other component, there may also be other components in between. On the other hand, when it is stated that one component is "directly connected" or "directly connected" to another component, it should be understood that there are no other components in between.

[0030] Additionally, throughout this specification, the terms neural network, neural network, and network function may be used interchangeably. A neural network may be composed of a set of interconnected computational units, which may generally be referred to as “nodes.” These “nodes” may also be referred to as “neurons.” A neural network is composed of at least two nodes. The nodes (or neurons) constituting neural networks may be interconnected by one or more “links.”

[0031] Artificial intelligence refers to the field of researching artificial intelligence or the methodologies capable of creating it, while machine learning refers to the field of researching methodologies to define and solve various problems addressed within the field of artificial intelligence. Machine learning is also defined as an algorithm that improves performance on a task through continuous experience.

[0032] An Artificial Neural Network (ANN) is a model used in machine learning that can refer to any model capable of problem-solving, composed of artificial neurons (nodes) that form a network through the connection of synapses. An artificial neural network can be defined by connection patterns between neurons in different layers, a learning process that updates model parameters, and an activation function that generates output values.

[0033] An artificial neural network may include an input layer, an output layer, and optionally one or more hidden layers. Each layer may include one or more neurons, and the artificial neural network may include synapses connecting the neurons. In an artificial neural network, each neuron may output a function value of an activation function for input signals, weights, and biases input through the synapses.

[0034] Model parameters refer to parameters determined through learning, including synaptic connection weights and neuron biases. Hyperparameters, on the other hand, refer to parameters that must be set before training in a machine learning algorithm, including the learning rate, number of iterations, mini-batch size, and initialization function.

[0035] The objective of training an artificial neural network can be viewed as determining model parameters that minimize the loss function. The loss function can be used as an indicator to determine optimal model parameters during the training process of an artificial neural network.

[0036] Machine learning can be classified into supervised learning, unsupervised learning, and reinforcement learning depending on the learning method.

[0037] Supervised learning refers to a method of training an artificial neural network with labels provided for the training data; a label can refer to the correct answer (or result) that the neural network must infer when training data is input. Unsupervised learning refers to a method of training an artificial neural network without labels provided for the training data. Reinforcement learning refers to a learning method in which an agent defined within an environment is trained to select an action or sequence of actions that maximizes the cumulative reward in each state.

[0038] Among artificial neural networks, machine learning implemented by a deep neural network (DNN) containing multiple hidden layers is also called deep learning, and deep learning is a part of machine learning. Hereinafter, machine learning is used in a sense that includes deep learning.

[0039] FIG. 1 shows an artificial intelligence device (100) according to one embodiment of the present disclosure.

[0040] The artificial intelligence device (100) can be implemented as a stationary device or a mobile device, such as a TV, projector, mobile phone, smartphone, desktop computer, laptop, digital broadcasting terminal, PDA (personal digital assistants), PMP (portable multimedia player), navigation, tablet PC, wearable device, set-top box (STB), DMB receiver, radio, washing machine, refrigerator, desktop computer, digital signage, robot, vehicle, etc.

[0041] Referring to FIG. 1, the artificial intelligence device (100) may include a communication unit (110), an input unit (120), a learning processor (130), a sensing unit (140), an output unit (150), a memory (170), and a processor (180), etc.

[0042] The communication unit (110) can transmit and receive data with external devices, such as other artificial intelligence devices (100a to 100e) or an artificial intelligence server (200), using wired and wireless communication technology. For example, the communication unit (110) can transmit and receive sensor information, user input, learning models, control signals, etc., with external devices.

[0043] At this time, the communication technologies used by the communication unit (110) include GSM (Global System for Mobile communication), CDMA (Code Division Multi Access), LTE (Long Term Evolution), 5G, WLAN (Wireless LAN), Wi-Fi (Wireless-Fidelity), Bluetooth (Bluetooth), RFID (Radio Frequency Identification), Infrared Data Association (IrDA), ZigBee, NFC (Near Field Communication), etc.

[0044] The input unit (120) can acquire various types of data.

[0045] At this time, the input unit (120) may include a camera for inputting a video signal, a microphone for receiving an audio signal, a user input unit for receiving information from a user, etc. Here, the camera or microphone may be treated as a sensor, and the signal obtained from the camera or microphone may be referred to as sensing data or sensor information.

[0046] The input unit (120) can obtain training data for model training and input data to be used when obtaining an output using the training model. The input unit (120) may also obtain unprocessed input data, in which case the processor (180) or the learning processor (130) can extract input feature points as a preprocessing step for the input data.

[0047] The learning processor (130) can train a model composed of an artificial neural network using training data. Here, the trained artificial neural network may be referred to as a learning model. The learning model can be used to infer a result value for new input data other than the training data, and the inferred value can be used as a basis for judgment to perform an action.

[0048] At this time, the learning processor (130) can perform artificial intelligence processing together with the learning processor (240) of the artificial intelligence server (200) of FIG. 2.

[0049] At this time, the learning processor (130) may include memory integrated into or implemented in the artificial intelligence device (100). Alternatively, the learning processor (130) may be implemented using memory (170), external memory directly coupled to the artificial intelligence device (100), or memory maintained in an external device.

[0050] The sensing unit (140) can acquire at least one of internal information of the artificial intelligence device (100), surrounding environment information of the artificial intelligence device (100), and user information using various sensors.

[0051] At this time, the sensors included in the sensing unit (140) include a proximity sensor, an illuminance sensor, an accelerometer, a magnetic sensor, a gyroscope, an inertial sensor, an RGB sensor, an IR sensor, a fingerprint recognition sensor, an ultrasonic sensor, a light sensor, a microphone, a lidar, a radar, etc.

[0052] The output unit (150) can generate output related to sight, hearing, or touch.

[0053] At this time, the output unit (150) may include a display unit that outputs visual information, a speaker that outputs auditory information, a haptic module that outputs tactile information, etc.

[0054] The memory (170) can store data that supports various functions of the artificial intelligence device (100). For example, the memory (170) can store input data, learning data, learning models, learning history, etc. obtained from the input unit (120).

[0055] The processor (180) can determine at least one executable operation of the artificial intelligence device (100) based on information determined or generated using a data analysis algorithm or a machine learning algorithm. The processor (180) can perform the determined operation by controlling the components of the artificial intelligence device (100).

[0056] To this end, the processor (180) can request, search, receive, or utilize data from the learning processor (130) or memory (170), and can control the components of the artificial intelligence device (100) to execute a predicted operation or a preferred operation among the at least one executable operation.

[0057] At this time, the processor (180) can generate a control signal to control the external device when the connection of the external device is required to perform a determined operation, and transmit the generated control signal to the external device.

[0058] The processor (180) can obtain intention information regarding user input and determine the user's requirements based on the obtained intention information.

[0059] At this time, the processor (180) can obtain intent information corresponding to the user input by using at least one of a Speech To Text (STT) engine for converting voice input into a string or a Natural Language Processing (NLP) engine for obtaining intent information of natural language.

[0060] At this time, at least one of the STT engine or NLP engine may be composed of an artificial neural network in which at least a portion is learned according to a machine learning algorithm. Also, at least one of the STT engine or NLP engine may be learned by a learning processor (130), learned by a learning processor (240) of an artificial intelligence server (200), or learned by distributed processing thereof.

[0061] The processor (180) may collect history information, including the operation details of the artificial intelligence device (100) or user feedback regarding the operation, and store it in memory (170) or a learning processor (130), or transmit it to an external device such as an artificial intelligence server (200). The collected history information may be used to update a learning model.

[0062] The processor (180) can control at least some of the components of the artificial intelligence device (100) to run an application stored in memory (170). Furthermore, the processor (180) can operate two or more of the components included in the artificial intelligence device (100) in combination with each other to run the application.

[0063] FIG. 2 shows an artificial intelligence server (200) according to one embodiment of the present disclosure.

[0064] Referring to FIG. 2, the artificial intelligence server (200) may refer to a device that trains an artificial neural network using a machine learning algorithm or uses a trained artificial neural network. Here, the artificial intelligence server (200) may be composed of multiple servers to perform distributed processing and may be defined as a 5G network. At this time, the artificial intelligence server (200) may be included as a part of the artificial intelligence device (100) and may perform at least a part of the artificial intelligence processing together.

[0065] The artificial intelligence server (200) may include a communication unit (210), memory (230), a learning processor (240), and a processor (260), etc.

[0066] The communication unit (210) can transmit and receive data with an external device such as an artificial intelligence device (100).

[0067] The memory (230) may include a model storage unit (231). The model storage unit (231) may store a model (or artificial neural network, 231a) that is being learned or has been learned through a learning processor (240).

[0068] The learning processor (240) can train the artificial neural network (231a) using training data. The training model may be used while mounted on the artificial intelligence server (200) of the artificial neural network, or may be used while mounted on an external device such as an artificial intelligence device (100).

[0069] The learning model may be implemented in hardware, software, or a combination of hardware and software. If part or all of the learning model is implemented in software, one or more instructions constituting the learning model may be stored in memory (230).

[0070] The processor (260) can infer a result value for new input data using a learning model and generate a response or control command based on the inferred result value.

[0071] FIG. 3 is a drawing for explaining the operation of an artificial intelligence device according to one embodiment of the present disclosure.

[0072] As illustrated in FIG. 3, the artificial intelligence device (500) of the present disclosure may include a database (510) that stores a dataset containing physical information functions and a processor (520) that generates a prediction model that predicts experimental results of a product based on the dataset.

[0073] Here, the processor (520) acquires a first dataset containing a physical information function for each of the initial DOEs (Design Of Experiment) selected in advance, inputs the first dataset into the neural network model (530) to train the neural network model (530) to predict the experimental results of the product, selects among the initial DOEs whose accuracy is less than a first threshold value based on the training results, extracts design factors for each selected DOE whose relative standard deviation (RSD) is less than a second threshold value, acquires a second dataset containing a physical information function for each new DOE containing the extracted design factors, retrains the neural network model (530) based on the second dataset, and when the retraining result reaches the target error, can generate a prediction model based on the retrained neural network model (530).

[0074] When the processor (520) receives a user input requesting the selection of an initial DOE when selecting an initial DOE, it can select a preset number of initial DOEs based on a DOE variable with low linear correlation from among a plurality of DOEs stored in advance in the database (510).

[0075] For example, when selecting an initial DOE, the processor (520) may select sampling methods in which DOE variables with low linear correlation have a high standard deviation greater than or equal to a first specific value, and select an initial DOE by extracting duplicate DOEs from sampling methods among the selected sampling methods that have high accuracy greater than or equal to a second specific value.

[0076] And, when an initial DOE is selected, the processor (520) can select packaging design factor variables for each initial DOE and perform packaging design corresponding to the product based on the selected packaging design factor variables.

[0077] For example, when selecting packaging design factor variables, the processor (520) may select packaging design factor variables including design factor variables for the packing configuration of the product, design factor variables for the packing edge of the product, design factor variables for the cushioning thickness of the product, design factor variables for the core thickness of the product, design factor variables for the riv of the product, and design factor variables for the core out of the product.

[0078] Here, design factor variables for the packing configuration of the product may include basic design factor variables for the top, bottom, and side of the packing, and additional design factor variables for the front and rear of the packing.

[0079] In addition, design factor variables for the packing hole of the product may include design factor variables for the top and bottom of the packing hole located at the long side edge of the product.

[0080] In addition, design factor variables for the cushioning thickness of the product may include design factor variables for the front, top, left, and right sides of the cushion located at the edge of one side of the product.

[0081] In addition, design factor variables for the basic wall thickness of the product may include design factor variables for the top and bottom of the basic wall thickness of the product.

[0082] In addition, design factor variables for the ribs of the product may include design factor variables for the width and position of the ribs of the product.

[0083] In addition, design factor variables for weight loss of the product may include design factor variables for the location of weight loss of the product.

[0084] Next, when the processor (520) receives a user input requesting the creation of a prediction model when acquiring the first dataset, it can extract pre-set experimental scenes for each pre-selected initial DOE from the database (510) and acquire the first dataset containing a physical information function corresponding to the extracted experimental scene from the database (510).

[0085] Here, the processor (520) can extract a plurality of preset experiment scenes for each initial DOE and obtain a first dataset corresponding to one experiment scene.

[0086] For example, when the processor (520) obtains the first dataset, the total number of the first dataset is obtained through a formula consisting of: total number of the first dataset = total number of initial DOEs × number of experimental scenes per DOE.

[0087] Additionally, when the processor (520) extracts experimental scenes pre-set for each initial DOE, when the experimental scenes for each initial DOE are extracted, the extracted experimental scenes are grouped according to the experimental type and classified into multiple experimental scene groups, and can obtain a first dataset including a physical information function for the classified experimental scene groups.

[0088] That is, when the processor (520) acquires the first dataset, it can acquire the first dataset including a physical information function for the experiment scene based on the following mathematical formula 1.

[0089]

[0090] Here, is the total loss function, and is the weight of the Mean Squared Error (MSE), and is the mean squared error, and is the weight of the physical information-based error, and is a physical information-based error.

[0091] For example, when the processor (520) acquires the first dataset, if the first dataset is acquired to learn to predict the maximum stress and maximum strain of a product for a collision test scene, the processor (520) can acquire the first dataset to learn to predict the maximum stress and strain of a product for a collision test scene based on the following mathematical formula 2.

[0092]

[0093] Here, is the total loss function, and is the weight of the Mean Squared Error (MSE), and is the weight of the physical information-based error.

[0094] For example, when the processor (520) obtains physical function stress, if the collision test scene is a collision type with the front and rear surfaces of the product, the maximum stress of the product calculated by the bending equation based on external force, Young's modulus, and moment of inertia can be obtained as physical function stress.

[0095] As another example, the processor (520) can obtain the maximum stress of the product as a physical function stress when the collision test scene is a collision type with the side of the product, calculated by a bending equation based on an external force, Young's modulus, and moment of inertia.

[0096] As another example, the processor (520) can obtain the maximum stress of the product as the physical function stress when the collision test scene is in the form of a collision on the upper and lower surfaces of the product, calculated by a bending equation based on external force, Young's modulus, and moment of inertia.

[0097] As another example, when the processor (520) obtains physical function stress, if the collision test scene is a collision type with the corner of the product, the maximum stress of the product calculated through a collision energy absorption formula based on constant, volume, and strain can be obtained as physical function stress.

[0098] In some cases, when a user input requesting the creation of a prediction model is received, the processor (520) may extract pre-set experimental scenes for each pre-selected initial DOE from the database (510), extract design factors and physical information functions corresponding to the extracted experimental scenes, and create a first dataset based on the design factors and physical information functions.

[0099] Here, the processor (520) can extract multiple experimental scenes pre-set for each initial DOE and generate a first dataset for each experimental scene based on design factors and physical information functions corresponding to the experimental scene.

[0100] For example, when generating the first dataset, the processor (520) can generate the total number of the first dataset through a formula consisting of: total number of the first dataset = total number of initial DOEs × number of experimental scenes per DOE.

[0101] And, when the processor (520) extracts experimental scenes pre-set for each initial DOE, when the experimental scenes for each initial DOE are extracted, the extracted experimental scenes are grouped according to the experimental type and classified into multiple experimental scene groups, and can generate a first dataset including a physical information function for the classified experimental scene groups.

[0102] Next, when the processor (520) generates the first dataset, it can generate the first dataset including a physical information function for the experiment scene based on the mathematical formula 1.

[0103] For example, when the processor (520) generates the first dataset, if the first dataset is to be trained to predict the maximum stress and maximum strain of a product for a collision test scene, the processor (520) can generate the first dataset based on the above mathematical formula 2 to predict the maximum stress and strain of a product for a collision test scene.

[0104] Here, the processor (520) can generate the maximum stress of the product as a physical function stress, calculated by a bending equation based on external force, Young's modulus, and moment of inertia, when generating the physical function stress, if the collision test scene is a collision type with the front and rear surfaces of the product.

[0105] Additionally, when generating physical function stress, the processor (520) can generate the maximum stress of the product, calculated by a bending equation based on external force, Young's modulus, and moment of inertia, as physical function stress if the collision test scene is a collision type with the side of the product.

[0106] Additionally, when generating physical function stress, the processor (520) can generate the maximum stress of the product as physical function stress, calculated by a bending equation based on external force, Young's modulus, and moment of inertia, if the collision test scene is a collision type with the upper and lower surfaces of the product.

[0107] Additionally, when generating physical function stress, the processor (520) can generate the maximum stress of the product as physical function stress by using a collision energy absorption formula based on constant, volume, and strain if the collision test scene is a collision type with the corner of the product.

[0108] Next, the processor (520) inputs a first dataset containing a physical information function corresponding to an experiment scene into the neural network model (530) when training the neural network model (530) to train the neural network model (530) to predict the experimental results of the product, and can store the predicted experimental results output through the training of the neural network model (530) for each initial DOE.

[0109] For example, if the first dataset is a dataset containing a physical information function for a collision experiment scene, the processor (520) can store prediction result data for the maximum stress and maximum strain of the product output through the learning of the neural network model (530) for each initial DOE.

[0110] And, when selecting a DOE whose accuracy is less than a first threshold value, the processor (520) calculates the accuracy for each initial DOE by comparing the prediction result data output through the learning of the neural network model (530) with the actual data, and can select a DOE among the initial DOEs whose accuracy is less than a first threshold value based on the calculated accuracy for each initial DOE.

[0111] Next, the processor (520), when extracting design factors whose relative standard deviation (RSD) is less than a second reference value, can select a DOE whose accuracy is less than a first reference value, calculate the relative standard deviation (RSD) of the design factors included in each selected DOE, extract design factors whose relative standard deviation is less than a second reference value among the calculated design factors, and select a new DOE that includes the design factors extracted from among a plurality of DOEs stored in advance in the database (510).

[0112] Here, when selecting a new DOE, the processor (520) may select a new DOE that includes design factors extracted from the remaining DOEs, excluding the initial DOE, among the plurality of DOEs stored in advance in the database (510).

[0113] At this time, the processor (520) can select new DOEs in the same number as the initial number of selected DOEs when selecting new DOEs.

[0114] Next, when a new DOE is selected, the processor (520) can select packaging design factor variables for each new DOE and perform packaging design corresponding to the product based on the selected packaging design factor variables.

[0115] For example, when selecting packaging design factor variables, the processor (520) may select packaging design factor variables including a design factor variable for the packing configuration of the product, a design factor variable for the packing edge of the product, a design factor variable for the cushioning thickness of the product, a design factor variable for the core thickness of the product, a design factor variable for the riv of the product, and a design factor variable for the core out of the product.

[0116] Here, design factor variables for the packing configuration of the product may include basic design factor variables for the top, bottom, and side of the packing, and additional design factor variables for the front and rear of the packing.

[0117] In addition, design factor variables for the packing hole of the product may include design factor variables for the top and bottom of the packing hole located at the long side edge of the product.

[0118] In addition, design factor variables for the cushioning thickness of the product may include design factor variables for the front, top, left, and right sides of the cushion located at the edge of one side of the product.

[0119] In addition, design factor variables for the basic wall thickness of the product may include design factor variables for the top and bottom of the basic wall thickness of the product.

[0120] In addition, design factor variables for the ribs of the product may include design factor variables for the width and position of the ribs of the product.

[0121] In addition, design factor variables for weight loss of the product may include design factor variables for the location of weight loss of the product.

[0122] Next, when acquiring the second dataset, the processor (520) can extract pre-set experimental scenes for each new DOE from the database (510) and acquire the second dataset containing physical information functions corresponding to the extracted experimental scenes from the database (510).

[0123] Here, the processor (520) can extract multiple experimental scenes pre-set for each new DOE and obtain a second dataset corresponding to one experimental scene.

[0124] For example, when the processor (520) obtains the second dataset, the total number of the second dataset is obtained through a formula consisting of: total number of the second dataset = total number of new DOEs × number of experimental scenes per DOE.

[0125] And, when the processor (520) extracts experimental scenes pre-set for each new DOE, when the experimental scenes for each new DOE are extracted, the extracted experimental scenes are grouped according to the experimental type and classified into multiple experimental scene groups, and can obtain a second dataset including a physical information function for the classified experimental scene groups.

[0126] Here, when the processor (520) acquires the second dataset, it can acquire the second dataset including a physical information function for the experiment scene based on the mathematical formula 1.

[0127] For example, when the processor (520) acquires the second dataset, if the second dataset is acquired to learn to predict the maximum stress and maximum strain of a product for a collision test scene, the processor (520) can acquire the second dataset based on the above mathematical formula 2 to learn to predict the maximum stress and strain of a product for a collision test scene.

[0128] Here, when the processor (520) obtains physical function stress, if the collision test scene is a collision type with the front and rear surfaces of the product, the maximum stress of the product calculated by the bending equation based on external force, Young's modulus, and moment of inertia can be obtained as physical function stress.

[0129] Additionally, when the processor (520) obtains physical function stress, if the collision test scene is a collision type with the side of the product, the maximum stress of the product calculated by the bending equation based on external force, Young's modulus, and moment of inertia can be obtained as physical function stress.

[0130] Additionally, when the processor (520) obtains physical function stress, if the collision test scene is a collision type with the upper and lower surfaces of the product, the maximum stress of the product calculated by the bending equation based on external force, Young's modulus, and moment of inertia can be obtained as physical function stress.

[0131] Additionally, when the processor (520) obtains physical function stress, if the collision test scene is a collision type with the corner of the product, the maximum stress of the product calculated through a collision energy absorption formula based on constant, volume, and deformation can be obtained as physical function stress.

[0132] In some cases, when acquiring a second dataset, the processor (520) may extract pre-set experimental scenes for each new DOE from the database (510), extract design factors and physical information functions corresponding to the extracted experimental scenes, and generate a second dataset based on the design factors and physical information functions.

[0133] Here, the processor (520) can extract multiple experimental scenes pre-set for each new DOE and generate a second dataset for each experimental scene based on design factors and physical information functions corresponding to the experimental scene.

[0134] For example, when generating the second dataset, the processor (520) can generate the total number of the second dataset through a formula consisting of: total number of the second dataset = total number of new DOEs × number of experimental scenes per DOE.

[0135] Here, when the processor (520) extracts pre-set experimental scenes for each new DOE, when the experimental scenes for each new DOE are extracted, the extracted experimental scenes are grouped according to the experimental type and classified into multiple experimental scene groups, and a second dataset including a physical information function for the classified experimental scene groups can be generated.

[0136] And, when the processor (520) generates the second dataset, it can generate the second dataset including a physical information function for the experiment scene based on the mathematical formula 1.

[0137] At this time, the processor (520) can generate a second dataset that learns to predict the maximum stress and maximum strain of a product for a collision test scene based on the above mathematical formula 2 when generating the second dataset.

[0138] Here, the processor (520) can generate the maximum stress of the product as a physical function stress, calculated by a bending equation based on external force, Young's modulus, and moment of inertia, when generating the physical function stress, if the collision test scene is a collision type with the front and rear surfaces of the product.

[0139] Additionally, when generating physical function stress, the processor (520) can generate the maximum stress of the product as physical function stress, calculated by a bending equation based on external force, Young's modulus, and moment of inertia, if the collision test scene is a collision type with the side of the product.

[0140] Additionally, when generating physical function stress, the processor (520) can generate the maximum stress of the product as physical function stress, calculated by a bending equation based on external force, Young's modulus, and moment of inertia, if the collision test scene is a collision type with the upper and lower surfaces of the product.

[0141] Additionally, when generating physical function stress, the processor (520) can generate the maximum stress of the product as physical function stress, calculated through a collision energy absorption formula based on constant, volume, and deformation, if the collision test scene is a collision type with the corner of the product.

[0142] Next, the processor (520) can retrain the neural network model (530) by inputting a second dataset containing a physical information function corresponding to an experiment scene into the neural network model (530) to predict the experimental results of the product, and can store the predicted experimental results output through the retraining of the neural network model (530) for each new DOE.

[0143] Here, the processor (520) can store prediction result data for the maximum stress and maximum strain of the product output through retraining of the neural network model (530) for each new DOE if the second dataset is a dataset containing a physical information function for a collision experiment scene.

[0144] Next, when generating a prediction model, the processor (520) calculates the accuracy for each new DOE by comparing the prediction result data output through the retraining of the neural network model (530) with the actual data, and determines whether the retraining result of the neural network model (530) reaches the target error based on the calculated accuracy, and if the retraining result of the neural network model (530) reaches the target error, it can generate a prediction model based on the retrained neural network model (530).

[0145] Here, the processor (520) can, if the retraining result of the neural network model (530) does not reach the target error, select a DOE among the new DOEs whose accuracy is less than a first threshold value based on the accuracy of the new DOEs calculated, extract a design factor for each selected DOE whose relative standard deviation (RSD) is less than a second threshold value, obtain a third dataset containing a physical information function for another new DOE containing the extracted design factor, and retrain the neural network model (530) based on the third dataset.

[0146] At this time, the processor (520), when extracting design factors whose relative standard deviation (RSD) is less than the second reference value, can calculate the relative standard deviation (RSD) of the design factors included in each selected DOE when selecting a DOE whose accuracy is less than the first reference value, extract design factors whose relative standard deviation is less than the second reference value among the calculated design factors, and select another new DOE that includes the design factors extracted from among a plurality of DOEs stored in advance in the database (510).

[0147] For example, when selecting another new DOE, the processor (520) may select another new DOE N times, including design factors extracted from the remaining DOEs, excluding the initial DOE selected first and the new DOE selected second, from among the multiple DOEs stored in the database (510).

[0148] Here, the processor (520) can select another new DOE N times, with the same number of selections as the initial DOE.

[0149] And, when another new DOE is selected N times, the processor (520) can select packaging design factor variables for each new DOE selected N times and perform packaging design corresponding to the product based on the selected packaging design factor variables.

[0150] In this way, the present disclosure enables the generation of a prediction model with high prediction accuracy and minimized training time by reducing the number of DOEs for training, by training a neural network model with a dataset containing physical information functions for each initial DOE to extract design factors with a low relative standard deviation (RSD) from DOEs with low accuracy, and retraining the neural network model to reach a target error with a second dataset containing new physical information functions for each DOE that includes the extracted design factors.

[0151] In addition, the present disclosure can increase data efficiency while maintaining physical consistency by training a neural network model with a dataset containing physical information functions.

[0152] FIGS. 4 and FIGS. 5 are drawings for explaining the initial DOE selection process of an artificial intelligence device according to one embodiment of the present disclosure.

[0153] As illustrated in FIGS. 4 and 5, the present disclosure can select a preset number of initial DOEs based on DOE variables with low linear correlation from among a plurality of DOEs stored in a database when a user input requesting the selection of initial DOEs is received.

[0154] As shown in FIG. 4, the present disclosure can select sampling methods among a plurality of sampling methods in which DOE variables with low linear correlation have a high standard deviation greater than or equal to a first specific value.

[0155] That is, the present disclosure extracts DOE variables with low linear correlation that an artificial intelligence device cannot predict for each sampling method, and can select sampling methods having high standard deviations of DOE variables with low linear correlation.

[0156] Next, as shown in FIG. 5, the present disclosure can select an initial DOE by extracting duplicate DOEs from sampling methods having a high accuracy of at least a second specific value among selected sampling methods.

[0157] That is, the present disclosure can select an initial DOE by extracting DOEs that overlap from the third sampling method and the fourth sampling method, which have high accuracy among the first to fifth sampling methods, as illustrated in FIG. 5.

[0158] FIG. 6 is a diagram illustrating the process of selecting packaging design factor variables by DOE of an artificial intelligence device according to one embodiment of the present disclosure.

[0159] As illustrated in FIG. 6, the present disclosure allows for the selection of packaging design factor variables for each initial DOE once an initial DOE is selected, and enables packaging design corresponding to the product based on the selected packaging design factor variables.

[0160] For example, the present disclosure may select packaging design factor variables including design factor variables for the packing configuration of a product, design factor variables for the packing edge of a product, design factor variables for the cushioning thickness of a product, design factor variables for the core thickness of a product, design factor variables for the riv of a product, and design factor variables for the core out of a product.

[0161] As shown in FIG. 6(a), design factor variables for the packing configuration of the product may include basic design factor variables for the top, bottom, and side of the packing, and additional design factor variables for the front and rear of the packing.

[0162] As shown in FIG. 6(b), design factor variables for the packing height of the product may include design factor variables for the top and bottom of the packing height located at the long side edge of the product.

[0163] As shown in FIG. 6(c), the design factor variables for the cushioning thickness of the product may include design factor variables for the front, top, left, and right sides of the cushion located at the edge of one side of the product.

[0164] As shown in FIG. 6(d), the design factor variables for the basic wall thickness of the product may include design factor variables for the top and bottom of the basic wall thickness of the product.

[0165] As shown in FIG. 6(e), design factor variables for the ribs of the product may include design factor variables for the width and position of the ribs of the product.

[0166] As shown in Fig. 6(f), the design factor variable for weight loss of the product may include the design factor variable for the location of weight loss of the product.

[0167] In addition, the present disclosure allows for the extraction of design factors with a low relative standard deviation (RSD) from DOEs with low accuracy based on prediction experiment results output through the training of a neural network model when a neural network model is trained based on a dataset containing physical information functions for each initial DOE, and when a new DOE containing design factors with a low relative standard deviation is selected, packaging design factor variables are selected for each new DOE, and packaging design corresponding to a product is performed based on the selected packaging design factor variables.

[0168] For example, the present disclosure may reselect packaging design factor variables including, as shown in FIGS. 6(a) to (f), design factor variables for the packing configuration of the product, design factor variables for the packing edge of the product, design factor variables for the cushioning thickness of the product, design factor variables for the core thickness of the product, design factor variables for the riv of the product, and design factor variables for the core out of the product.

[0169] FIGS. 7 to 13 are drawings for explaining the process of acquiring a dataset including physical information functions for each DOE of an artificial intelligence device according to one embodiment of the present disclosure.

[0170] The present disclosure can, when a user input requesting the acquisition of a prediction model is received, extract pre-set experimental scenes for each pre-selected DOE from a database and obtain a dataset containing physical information functions corresponding to the extracted experimental scenes from the database.

[0171] In some cases, the present disclosure may, upon receiving user input requesting the creation of a prediction model, extract pre-set experimental scenes for each pre-selected DOE from a database, extract design factors and physical information functions corresponding to the extracted experimental scenes, and create a dataset based on the design factors and physical information functions.

[0172] Herein, the present disclosure can extract a plurality of pre-set experimental scenes for each DOE and generate a dataset for each experimental scene based on design factors and physical information functions corresponding to the experimental scene.

[0173] For example, the present disclosure can generate the total number of datasets through a formula consisting of: total number of datasets = total number of DOEs × number of experimental scenes per DOE.

[0174] In addition, the present disclosure can generate a dataset including physical information functions for the classified experiment scene groups by grouping the extracted experiment scenes according to the experiment type when experiment scenes are extracted by DOE.

[0175] As illustrated in FIGS. 7 to 13, the present disclosure can generate a dataset for training to predict the maximum stress and maximum strain of a product for a crash test scene based on Equation 1 and Equation 2.

[0176] As shown in FIGS. 7 and 8, the present disclosure can generate a maximum stress of a product as a physical function stress, calculated by a first bending equation based on an external force, Young's modulus, and moment of inertia, when the collision test scene is a collision type with the front and rear surfaces of the product.

[0177] In addition, as shown in FIGS. 9 and 10, the present disclosure can generate the maximum stress of a product as a physical function stress, calculated by a second bending equation based on an external force, Young's modulus, and moment of inertia, when the collision test scene is a collision type with the side of the product.

[0178] In addition, as shown in FIGS. 11 and 12, the present disclosure can generate the maximum stress of a product as a physical function stress, calculated by a third bending equation based on external force, Young's modulus, and moment of inertia, when the collision test scene is a collision type with the upper and lower surfaces of the product.

[0179] In addition, as shown in FIG. 13, the present disclosure can generate the maximum stress of a product as a physical function stress, calculated through a collision energy absorption formula based on a constant, volume, and deformation, when the collision test scene is a collision type with the corner of the product.

[0180] In addition, the present disclosure can generate a new dataset to retrain the product to predict the maximum stress and maximum strain of a product for a crash test scene, as shown in FIGS. 7 to 13, when a new DOE is selected and packaging design factor variables are re-selected for each new DOE.

[0181] FIG. 14 is a diagram illustrating the neural network model learning process of an artificial intelligence device according to one embodiment of the present disclosure.

[0182] As illustrated in FIG. 14, the present disclosure allows a neural network model (530) to be trained to predict experimental results of a product by inputting a dataset containing physical information functions corresponding to an experimental scene into the neural network model (530), and the predicted experimental results output through the training of the neural network model (530) can be stored for each DOE.

[0183] For example, the present disclosure may store prediction result data for the maximum stress and maximum strain of a product output through the learning of a neural network model (530) for each DOE, provided that the dataset includes a physical information function for a collision test scene.

[0184] FIGS. 15 and 16 are drawings for explaining the process of selecting a new DOE based on the RSD criteria of an artificial intelligence device according to one embodiment of the present disclosure.

[0185] As illustrated in FIG. 15, the present disclosure calculates accuracy for each DOE by comparing the prediction result data output through the training of a neural network model with actual data, and can select DOEs among the DOEs whose accuracy is less than a threshold value based on the calculated accuracy for each DOE.

[0186] That is, in FIG. 15, the present disclosure can determine that the accuracy for DOE 7, DOE 46, DOE 52, and DOE 85 is low if, when comparing the prediction result data output through the training of a neural network model with the actual data, the error corresponding to DOE 7, DOE 46, DOE 52, and DOE 85 among the DOEs is relatively higher than that of other DOEs.

[0187] Next, as illustrated in FIG. 16, the present disclosure can calculate the relative standard deviation (RSD) of design factors included in each selected DOE when a DOE with accuracy less than a first reference value is selected, extract design factors with a relative standard deviation less than a second reference value from among the calculated design factors, and select a new DOE including the extracted design factors from among a plurality of DOEs stored in advance in a database.

[0188] That is, in FIG. 16, the present disclosure can calculate the relative standard deviation (RSD) of design factors included in DOE 7, DOE 46, DOE 52, and DOE 85 (A) with low accuracy among DOEs, and extract a third design factor, a fourth design factor, and a seventh design factor having a low relative standard deviation (B) among the calculated design factors.

[0189] Furthermore, the present disclosure may select a new DOE including a third design factor, a fourth design factor, and a seventh design factor having a low relative standard deviation (B) among a plurality of DOEs stored in advance in a database.

[0190] Herein, the present disclosure may select a new DOE comprising design factors extracted from the remaining DOEs, excluding the initial DOE, among a plurality of DOEs stored in advance in a database.

[0191] In this case, the present disclosure may select new DOEs in the same number as the initial number of selected DOEs, but this is merely one embodiment and is not limited thereto.

[0192] FIGS. 17 and 18 are drawings for explaining the reduction of the prediction model generation period of an artificial intelligence device according to one embodiment of the present disclosure.

[0193] As illustrated in FIG. 17, it can be seen that the present disclosure can significantly shorten the prediction model generation period compared to existing methods.

[0194] In other words, the existing method acquires a dataset for all DOEs and then trains a neural network model based on the dataset for each DOE, so the training period for the model is very long, and it may take about 120 days to generate the prediction model.

[0195] In contrast, the present disclosure selects a small number of DOEs and trains a neural network model in the first stage based on a dataset containing physical information functions for each small number of DOEs. Then, based on the results of the first stage of training, design factors with a low relative standard deviation (RSD) are extracted from DOEs with low accuracy. Then, new DOEs containing design factors with a low relative standard deviation (RSD) are selected and the neural network model is trained in the second stage based on a dataset containing physical information functions for each small number of new DOEs. Afterward, the neural network model is trained in the same manner until the target error is reached, thereby shortening the training period of the model and reducing the generation period of the prediction model to approximately 40 days.

[0196] In addition, as shown in FIG. 18, the number of DOE analyses for model training in the conventional method is about 2,700, but in the present disclosure, the number of DOE analyses for model training is about 900, which can be seen to be reduced by about 67%.

[0197] In addition, while the conventional method requires a DOE analysis period of about 120 days for model training, the present disclosure requires a DOE analysis period of about 40 days for model training, which can be seen to be reduced by about 67%.

[0198] FIG. 19 is a drawing for explaining the overall operation flow of an artificial intelligence device according to one embodiment of the present disclosure.

[0199] As illustrated in FIG. 19, the present disclosure can obtain a first dataset containing a physical information function for each pre-selected initial DOE (Design Of Experiment) (S10).

[0200] Herein, the present disclosure can select a preset number of initial DOEs based on DOE variables with low linear correlation from among a plurality of DOEs stored in advance in a database when a user input requesting the selection of initial DOEs is received.

[0201] In addition, the present disclosure allows for the selection of packaging design factor variables for each initial DOE once an initial DOE is selected, and enables the performance of packaging design corresponding to the product based on the selected packaging design factor variables.

[0202] In addition, the present disclosure may extract experimental scenes pre-set for each pre-selected initial DOE from a database when a user input requesting the creation of a prediction model is received, and obtain a first dataset from the database that includes a physical information function corresponding to the extracted experimental scene.

[0203] In some cases, the present disclosure may, upon receiving a user input requesting the creation of a prediction model, extract pre-set experimental scenes from a database for each pre-selected initial DOE, extract design factors and physical information functions corresponding to the extracted experimental scenes, and create a first dataset based on the design factors and physical information functions.

[0204] In addition, the present disclosure can train a neural network model to predict experimental results of a product by inputting a first dataset into the neural network model (S20).

[0205] Herein, the present disclosure inputs a first dataset containing a physical information function corresponding to an experiment scene into a neural network model to train the neural network model to predict the experimental results of a product, and can store the predicted experimental results output through the training of the neural network model for each initial DOE.

[0206] Next, the present disclosure may select a DOE among the initial DOEs whose accuracy is less than a first threshold value based on the learning results (S30).

[0207] Herein, the present disclosure calculates accuracy for each initial DOE by comparing the prediction result data output through the training of a neural network model with actual data, and can select among the initial DOEs whose accuracy is less than a first threshold value based on the calculated accuracy for each initial DOE.

[0208] Next, the present disclosure can extract design factors for each selected DOE in which the relative standard deviation (RSD) is less than a second reference value (S40).

[0209] Herein, the present disclosure can select a DOE whose accuracy is less than a first threshold value, calculate the relative standard deviation (RSD) of the design factors included in each selected DOE, extract design factors whose relative standard deviation is less than a second threshold value among the calculated design factors, and select a new DOE that includes the design factors extracted from a plurality of DOEs stored in advance in a database.

[0210] In addition, the present disclosure allows for the selection of packaging design factor variables for each new DOE when a new DOE is selected, and enables the performance of packaging design corresponding to the product based on the selected packaging design factor variables.

[0211] And, the present disclosure can obtain a second dataset containing physical information functions for each new DOE containing extracted design factors (S50).

[0212] Herein, the present disclosure can extract experimental scenes pre-set for each new DOE from the database and obtain a second dataset from the database that includes a physical information function corresponding to the extracted experimental scene.

[0213] In some cases, the present disclosure may extract pre-set experimental scenes for each new DOE from a database, extract design factors and physical information functions corresponding to the extracted experimental scenes, and generate a second dataset based on the design factors and physical information functions.

[0214] Next, the present disclosure can retrain a neural network model based on a second dataset (S60).

[0215] Herein, the present disclosure inputs a second dataset containing a physical information function corresponding to an experiment scene into a neural network model to retrain the neural network model to predict the experimental results of the product, and can store the predicted experimental results output through the retraining of the neural network model for each new DOE.

[0216] Next, the present disclosure can determine whether the retraining result reaches the target error (S70).

[0217] Herein, the present disclosure calculates accuracy for each new DOE by comparing the prediction result data output through the retraining of a neural network model with actual data, and determines whether the result of retraining the neural network model reaches the target error based on the calculated accuracy.

[0218] And, the present disclosure can generate a prediction model based on the retrained neural network model when the retraining result reaches the target error (S80).

[0219] Herein, the present disclosure may select a DOE among the new DOEs whose accuracy is less than a first threshold value based on the accuracy of the new DOE calculated when the result of retraining a neural network model does not reach a target error, extract a design factor for each selected DOE whose relative standard deviation (RSD) is less than a second threshold value, obtain a third dataset including a physical information function for each other new DOE containing the extracted design factor, and retrain a neural network model based on the third dataset.

[0220] Additionally, the present disclosure may select a DOE whose accuracy is less than a first threshold value, calculate the relative standard deviation (RSD) of the design factors included in each selected DOE, extract design factors whose relative standard deviation is less than a second threshold value among the calculated design factors, and select another new DOE that includes the design factors extracted from a plurality of DOEs stored in advance in a database.

[0221] Here, the present disclosure may select another new DOE including the extracted design factor from among the remaining DOEs, excluding the initial DOE selected first and the new DOE selected second, from among a plurality of DOEs stored in advance in a database.

[0222] In this way, the present disclosure enables the generation of a prediction model with high prediction accuracy and minimized training time by reducing the number of DOEs for training, by training a neural network model with a dataset containing physical information functions for each initial DOE to extract design factors with a low relative standard deviation (RSD) from DOEs with low accuracy, and retraining the neural network model to reach a target error with a second dataset containing new physical information functions for each DOE that includes the extracted design factors.

[0223] In addition, the present disclosure can increase data efficiency while maintaining physical consistency by training a neural network model with a dataset containing physical information functions.

[0224] The above-described disclosure can be implemented as computer-readable code on a medium on which a program is recorded. A computer-readable medium includes all types of recording devices in which data that can be read by a computer system is stored. Examples of computer-readable media include a Hard Disk Drive (HDD), a Solid State Disk (SSD), a Silicon Disk Drive (SSD), ROM, RAM, a CD-ROM, a magnetic tape, a floppy disk, an optical data storage device, etc. Additionally, the computer may include a processor of an artificial intelligence device.

[0225] According to the physical information-based artificial intelligence device of the present disclosure, by training a neural network model with a dataset containing physical information functions for each initial DOE to extract design factors with a low relative standard deviation (RSD) from DOEs with low accuracy, and retraining the neural network model to reach a target error with a second dataset containing new physical information functions for each DOE including the extracted design factors, the number of DOEs for training is reduced to minimize training time and generate a prediction model with high prediction accuracy, thus demonstrating significant industrial applicability.

Claims

A database storing a dataset containing physical information functions; and, It includes a processor that generates a prediction model for predicting experimental results of a product based on the above dataset, and The above processor is, An artificial intelligence device characterized by acquiring a first dataset containing a physical information function for each of the initial Design Of Experiment (DOE) selected in advance, inputting the first dataset into a neural network model to train the neural network model to predict the experimental results of the product, selecting among the initial DOEs whose accuracy is less than a first threshold value based on the training results, extracting design factors for each selected DOE whose Relative Standard Deviation (RSD) is less than a second threshold value, acquiring a second dataset containing a physical information function for each new DOE containing the extracted design factors, retraining the neural network model based on the second dataset, and generating the prediction model based on the retrained neural network model when the retraining result reaches a target error. In Article 1, The above processor is, An artificial intelligence device characterized by selecting a preset number of initial DOEs based on DOE variables with low linear correlation from among a plurality of DOEs stored in advance in the database when a user input requesting the selection of the initial DOEs is received. In Article 1, The above processor is, An artificial intelligence device characterized by selecting packaging design factor variables for each of the initial DOEs when the initial DOEs are selected, and performing packaging design corresponding to the product based on the selected packaging design factor variables. In Article 1, The above processor is, An artificial intelligence device characterized by, when acquiring the first dataset, receiving a user input requesting the creation of the prediction model, extracting pre-set experimental scenes for each of the pre-selected initial DOEs from the database, and acquiring the first dataset including a physical information function corresponding to the extracted experimental scene from the database. In Paragraph 4, The above processor is, When acquiring the above first dataset, (Here, is the total loss function, and is the weight of the Mean Squared Error (MSE), and is the mean squared error, and is the weight of the physical information-based error, and An artificial intelligence device characterized by acquiring a first dataset containing a physical information function for the experimental scene based on a formula consisting of (which is a physical information-based error). In Article 5, The above processor is, When acquiring the first dataset above, if the acquisition of the first dataset is used to train the product to predict the maximum stress and maximum strain for a crash test scene (Here, is the total loss function, and is the weight of the Mean Squared Error (MSE), and An artificial intelligence device characterized by acquiring a first dataset to train to predict the maximum stress and strain of a product for a collision test scene based on a formula consisting of (which is the weight of the error based on physical information). In Article 1, The above processor is, An artificial intelligence device characterized by, when training the neural network model, inputting a first dataset containing a physical information function corresponding to an experiment scene into the neural network model to train the neural network model to predict the experimental results of the product, and storing the predicted experimental results output through the training of the neural network model for each initial DOE. In Article 1, The above processor is, An artificial intelligence device characterized by, when selecting a DOE whose accuracy is less than a first threshold value, calculating the accuracy for each initial DOE by comparing the prediction result data output through the learning of the neural network model with the actual data, and selecting a DOE among the initial DOEs whose accuracy is less than the first threshold value based on the calculated accuracy for each initial DOE. In Article 1, The above processor is, An artificial intelligence device characterized by, when extracting design factors whose relative standard deviation (RSD) is less than a second reference value, selecting a DOE whose accuracy is less than a first reference value, calculating the relative standard deviation (RSD) of the design factors included in each selected DOE, extracting design factors whose relative standard deviation is less than the second reference value among the calculated design factors, and selecting a new DOE including the extracted design factors from among a plurality of DOEs stored in advance in the database. In Article 9, The above processor is, An artificial intelligence device characterized by selecting a new DOE containing the extracted design factor from among the remaining DOEs, excluding the initial DOE, among a plurality of DOEs stored in advance in the database when selecting the new DOE. In Article 9, The above processor is, An artificial intelligence device characterized by selecting packaging design factor variables for each of the new DOEs when the new DOEs are selected, and performing packaging design corresponding to the product based on the selected packaging design factor variables. In Article 1, The above processor is, An artificial intelligence device characterized by, when acquiring the second dataset, extracting experimental scenes pre-set for each of the new DOEs from the database, and acquiring the second dataset including a physical information function corresponding to the extracted experimental scene from the database. In Article 1, The above processor is, An artificial intelligence device characterized by, when generating the above prediction model, comparing the prediction result data output through the retraining of the neural network model with actual data to calculate accuracy for each new DOE, determining whether the result of retraining the neural network model reaches a target error based on the calculated accuracy, and if the result of retraining the neural network model reaches the target error, generating the above prediction model based on the retrained neural network model. In Article 13, The above processor is, An artificial intelligence device characterized by, when the retraining result of the above neural network model does not reach a target error, selecting among the new DOEs whose accuracy is less than a first threshold value based on the accuracy for each new DOE calculated above, extracting design factors for each selected DOE whose relative standard deviation (RSD) is less than a second threshold value, acquiring a third dataset including a physical information function for each other new DOE containing the extracted design factors, and retraining the neural network model based on the third dataset. A step of acquiring a first dataset containing physical information functions for each pre-selected initial DOE (Design Of Experiment); A step of inputting the above-mentioned first dataset into a neural network model to train the neural network model to predict the experimental results of the above-mentioned product; A step of selecting DOEs among the initial DOEs whose accuracy is less than a first threshold value based on the above learning results; A step of extracting design factors for each of the selected DOEs for which the relative standard deviation (RSD) is less than a second reference value; A step of obtaining a second dataset containing physical information functions for each new DOE containing the design factors extracted above; A step of retraining the neural network model based on the second dataset; and A method for learning a prediction model of an artificial intelligence device, characterized by including the step of generating the prediction model based on the retrained neural network model when the retraining result reaches the target error.

Images