Method and system for constructing simulation environment model based on AIGC and storage medium

Abstract

The invention discloses a method and system for constructing a simulation environment model based on AIGC and a storage medium. The method comprises the following steps that S1, geographic information and obstacle information in information input by a user are extracted through a BiLSTM-CRF model; converting the text data into a word vector sequence through a word embedding layer, and inputting the word vector sequence into BiLSTM; an output vector of the BiLSTM is used as an input vector of a CRF decoding layer, and the CRF outputs an optimal tag sequence corresponding to a word vector sequence; and geographic information and obstacle information in the information input by the user are extracted according to the optimal label sequence. S2, respectively inputting the geographic information and the obstacle information into a trained generative adversarial network for generating a geographic model and an obstacle model to generate corresponding models; s3, combining the geographic model generated in the step S2 with the obstacle model to construct a simulation environment model; the device can quickly and accurately construct a simulation environment model according to information input by a user.

Description

Technical Field

[0001] The present invention relates to a simulation environment model, and in particular to a method, system and storage medium for constructing a simulation environment model based on AIGC. Background Art

[0002] With the continuous development of computer technology, computer simulation has become an indispensable technical tool in many fields. Based on its high efficiency, low cost and safety, simulation technology has been widely used in aviation, aerospace, automobile, medical and other fields.

[0003] In the past, the construction of three-dimensional simulation environment models often adopted a customized and targeted scene construction method, that is, all parameters of the simulation environment needed to be configured in the code algorithm in advance. When the actual operating environment changes, it is impossible to quickly and accurately modify the simulation environment model parameters according to the new environment information input by the user to build a new simulation environment model; that is, it is impossible to quickly and accurately build a simulation environment model based on the user input information. Summary of the invention

[0004] Purpose of the invention: The purpose of the present invention is to provide a method, system and storage medium for constructing a simulation environment model based on AIGC, which can quickly and accurately construct a simulation environment model according to user input information.

[0005] Technical solution: The method for constructing a simulation environment model based on AIGC of the present invention comprises the following steps:

[0006] S1. Extract geographic information and obstacle information from the user input information through the BiLSTM-CRF model;

[0007] S2, inputting the geographic information and obstacle information into the trained generative adversarial network for generating geographic models and obstacle models to generate corresponding models;

[0008] S3, combining the geographical model and obstacle model generated in step S2 to construct a simulation environment model.

[0009] Based on the above technical solution, the BiLSTM-CRF model can quickly and accurately extract the geographic information and obstacle information from the user input information, making the expression of the user input information more free and convenient, without deliberately entering the corresponding information according to a fixed template, and the extracted information is highly accurate, which also ensures the accuracy of the subsequently constructed model; and the extracted information is input into the generative adversarial network used to generate the geographic model and the obstacle model, and the corresponding model can be automatically and quickly generated. The independent generative adversarial networks are trained for the geographic model and the obstacle model respectively, considering the differences between the two models, and separate independent training can improve the accuracy of the corresponding model generation; the accuracy of the final simulation environment model constructed by combining the two generated models with higher accuracy is also higher, and the whole process is completed by the BiLSTM-CRF model and the generative adversarial network in AIGC. The user only needs to input the corresponding information of the new environment to automatically generate it, and there is no need to manually modify the corresponding parameters, which improves the generation speed of the simulation environment model.

[0010] Preferably, the geographic information extracted in step S1 includes longitude and latitude, terrain and landform; the obstacle information includes coordinates, elevation, area and collision volume of the obstacle.

[0011] Preferably, the information input by the user in step S1 includes text information and voice information, and the information input by the user is converted into text data after preprocessing, and then the BiLSTM-CRF model is used to extract information.

[0012] Through preprocessing, all user input information is converted into text data to facilitate information extraction of the BiLSTM-CRF model.

[0013] Preferably, the preprocessing includes converting voice information into text data by voice conversion technology.

[0014] Preferably, in step S1, the text data is converted into a word vector sequence through a word embedding layer, and the word vector sequence is input into BiLSTM; the output vector of BiLSTM is used as the input vector of a CRF decoding layer, and CRF outputs the optimal label sequence corresponding to the word vector sequence; and the geographic information and obstacle information in the information input by the user are extracted according to the optimal label sequence.

[0015] Preferably, the calculation formula for the optimal tag sequence is:

[0016] y * = argmax j∈N score(X,y j )

[0017] Where X is the word vector sequence, y jis the jth predicted label sequence of X, N is the total number of all possible predicted label sequences of X, score(X,y j ) is calculated as

[0018]

[0019] Among them, it means that the i-th word in X is y j The probability score of the i-th label in , Represents y j The transition probability score between the i-1th and i-th labels in .

[0020] Preferably, the target loss function calculation formula of the generated adversarial network in step S2 is:

[0021]

[0022] Among them, x is the real sample vector, P data is the real data distribution, z is the noise vector that follows the Gaussian distribution P Z A random vector of , G(z) is the generator, D(G(z)) is the discriminator, and E represents the expected value of the specified distribution.

[0023] Preferably, in step S3, after the geographic model and the obstacle model are combined, additional environmental information input by the user or the system default environmental information is added to construct a simulation environment model; the environmental information includes light, temperature, humidity, precipitation, wind speed, area type, height or depth limit and geographic coordinates.

[0024] Adding environmental information after combining the geographic model and obstacle model can further improve the accuracy and authenticity of the final constructed simulation environment model.

[0025] The system for constructing a simulation environment model based on AIGC of the present invention comprises:

[0026] Information extraction module: used to extract geographic information and obstacle information from the information input by the user using the BiLSTM-CRF model;

[0027] Preliminary model generation module: used to input geographic information and obstacle information into the trained generative adversarial network for generating geographic model and obstacle model to generate corresponding models;

[0028] Final model generation module: used to combine the generated geographic model and obstacle model to construct a simulation environment model.

[0029] A computer-readable storage medium storing one or more programs according to the present invention includes one or more programs including instructions, and when the instructions are executed by a computing device, the computing device executes any one of the above methods.

[0030] Beneficial effects: Compared with the prior art, the present invention has the following significant effects: the corresponding information in the user input information can be quickly and accurately extracted through the BiLSTM-CRF model, and the extracted information can be input into the trained generative adversarial network to generate a geographic model and an obstacle model, and then the two can be combined to construct a simulation environment model. The whole process is completed based on AIGC, which is fast and simple, and the accuracy of the final simulation environment model is also improved by generating two independent models first. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] Figure 1 It is a schematic diagram of the process of the present invention;

[0032] Figure 2 This is a schematic diagram of the BiLSTM-CRF algorithm model;

[0033] Figure 3 Schematic diagram of the structure of the generated adversarial network;

[0034] Figure 4 This is the effect diagram of the simulation environment model;

[0035] Figure 5 This is a flow chart of the simulation process. DETAILED DESCRIPTION

[0036] As shown in the figure, the method for constructing a simulation environment model based on AIGC according to the present invention comprises the following steps:

[0037] S1. Extract geographic information and obstacle information from the user input information through the BiLSTM-CRF model;

[0038] Construct an input module to provide an interactive interface, collect conditions or descriptions entered by users through text or voice, and provide a data source for subsequent extraction of key information; key information includes but is not limited to the geographical location of the experimental area, meteorological information, building distribution, regional conditions, etc.;

[0039] Preprocess the user input information. First, determine the source data type. If it is text input, save the data directly in text form. If it is voice input, convert the audio into text form through voice conversion technology and save it.

[0040] A BiLSTM-CRF model is built to perform named entity recognition tasks, with geographic information, obstacle information, and environmental information as recognition targets.

[0041] Through the word embedding layer, the saved text data T = (t1, t2, ..., t n ), converted to a word vector sequence W = {w1,w2,...,w n}.

[0042] After the word vector sequence W is input into BiLSTM, vectors in two directions are obtained. These vectors have rich semantic information. The final vector is obtained by concatenating the forward and backward vectors to integrate this information, and the output is a vector containing the forward and backward information corresponding to each word in the word vector sequence, so as to better capture the semantic features in the word vector sequence.

[0043] The output of the BiLSTM network layer is recorded as P = {P1, P2, ..., P n}, where P i Indicates w i The probability of each label, at this time P is the input vector of the CRF decoding layer.

[0044] Finally, the CRF decoding layer adds certain constraints between labels, and uses the transfer matrix to constrain and correct the output vector P of the BiLSTM layer. The values ​​in the matrix are the transfer scores between corresponding labels to avoid invalid label prediction. The BiLSTM-CRF calculation formula is as follows:

[0045]

[0046] Among them, W is the word vector sequence, y j is the jth predicted label sequence of W, N is the total number of all possible predicted label sequences of W, Indicates that the i-th word in W is y j The probability score of the i-th label in , Represents y j The transition probability score between the i-1th and i-th labels in .

[0047] After normalization, the word vector sequence W is labeled as the jth predicted label sequence y j The probability of the possibility is as follows:

[0048]

[0049] The probability is solved by maximizing the likelihood function, and the calculation formula is as follows:

[0050]

[0051] Finally, the Viterbi algorithm is used to solve the optimal label sequence corresponding to the input sequence. The calculation formula is as follows:

[0052] y* = argmax j∈N score(W,y j ) (4)

[0053] Through the above steps, key information extraction is achieved, and the key information is classified and processed according to the set geographical information, obstacle information, environmental information and other goals.

[0054] S2, inputting the geographic information and obstacle information into the trained generative adversarial network for generating geographic models and obstacle models to generate corresponding models;

[0055] Generative adversarial networks are used to automatically generate scene environments. Figure 3 The network structure of the generated adversarial network is constructed, where the noise vector z follows the Gaussian distribution P Z The real data distribution in this structure is P data , G represents the parameter θ g The generator G receives a Gaussian distribution P. Z The noise vector z is passed through the parameter θ g Calculate the noise space P Z To the real data distribution P data The mapping G(z;θ g ); for the discriminator D(x;θ d ), its input is the real sample x and the generated sample X fake =G(z), outputs a single scalar value [0,1]. If the input of the discriminator is a real sample x, the expected output of D(x) is close to 1; if the input is a generated sample G(z), the expected output of D(G(z)) is as close to 0 as possible, that is, maximizing (1-D(G(z))).

[0056] That is, the goal of the entire network is to allow the generator G to gradually distribute the random noise P through adversarial learning. Z Mapping to the real data distribution P data The ability to generate data X fake =G(z) is input to the trained discriminator D, the output can be as close to 1 as possible, that is, D(G(z)) is maximized, so that the expected objective function is minimized. The objective loss function of adversarial training is as follows:

[0057]

[0058] The noise vector z follows the Gaussian distribution P Z A random vector, P data is the real data distribution, G(z) is the generator, D(G(z)) is the discriminator, and E represents the expected value of the specified distribution. Expanding formula 2 yields:

[0059] V(D,G)=∫ x P data (x)log(D(x))dx+∫ x P z (z)log(1-D(G(z)))dx

[0060] =∫ x P data (x)log(D(x))+p g (x)log(1-D(x))dx (6)

[0061] Among them, p g (x) is the model distribution implicitly defined by G(z). The above formula converts the problem of finding the optimal network into the problem of finding the maximum integral value. data (x) and p g (x) are all scalars, so let P data (x) = a, p g (x) = b, D(x) = y, the function can be rewritten as:

[0062] f(y)=alogy+blog(1-y) (7)

[0063] In order to find the most extreme point, we can take the derivative to get the result. If a+b≠0, then

[0064]

[0065] Where a, b∈(0,1). Because the first-order derivative of the integrand is equal to zero and the second-order derivative is less than zero, the discriminator has a maximum value, and the optimal discriminator is:

[0066]

[0067] When p g (x) = P data (x), D * (x) = 0.5, indicating that the discriminator is completely confused and cannot distinguish between real samples and generated samples. This shows that the generator perfectly replicates the data distribution of real samples and can obtain the required generated data.

[0068] The process of generating a corresponding model through a generative adversarial network is as follows:

[0069] Geographic environment model generation

[0070] The geographic information extracted in step S1, including but not limited to longitude and latitude, terrain, landform and other descriptions, is input into a generative adversarial network that has been trained to generate a geographic model, and DEM (digital elevation model) data per unit area is output. Material assignment and texture mapping are performed according to the type of geographic environment, and all unit area models are spliced ​​and smoothed, so as to automatically generate a geographic environment model.

[0071] The generative adversarial network training used to generate geographic models is trained by combining information in the geographic information system and confidence in training samples. The information in the geographic information system is real sample information of the geographic environment obtained from the geospatial database.

[0072] Obstacle model generation

[0073] The obstacle information extracted in step S1, including but not limited to coordinates, elevation, floor area, collision volume and other descriptions, is input into a generative adversarial network that has been trained to generate an obstacle model, and a three-dimensional obstacle model including relevant attributes such as model size, features, type, collision area, etc. is automatically generated; the training of the generative adversarial network can be combined with the three-dimensional model and semantic information and training samples in the model library for training.

[0074] S3, combining the geographical model and obstacle model generated in step S2 to construct a simulation environment model.

[0075] The geographic model generated in step S2 is combined with the obstacle model, and the user input environmental information including but not limited to light, temperature, humidity, precipitation, wind speed, area type, height / depth limit, geographic coordinates, etc. is added to the simulated three-dimensional scene to construct the final simulation environment model. Of course, environmental information can also be added; if the user does not enter relevant data, the system default value is uniformly adopted, and finally the automatic generation of the simulation ring model based on the user input information is realized.

[0076] The generated simulation environment model can be used in simulation experiments of simulation objects to obtain relevant experimental data and reduce experimental costs.

[0077] Specifically, using the simulation environment model generated in step S3 as a carrier, input simulation object information including but not limited to starting position, speed, posture, type, etc., as well as execution task information including but not limited to task duration, waypoints, routes, height / depth, execution mode, task type, etc., and utilize the constructed simulation environment model to realize real-time analysis and visual display of the running status of the simulation object in the generated simulation environment. After a single simulation process is completed, a comprehensive evaluation result report is output based on the status analysis of the entire simulation process to provide decision support for operation staff.

[0078] The system for constructing a simulation environment model based on AIGC of the present invention comprises:

[0079] Information extraction module: used to extract geographic information and obstacle information from the information input by the user using the BiLSTM-CRF model;

[0080] Preliminary model generation module: used to input geographic information and obstacle information into the trained generative adversarial network for generating geographic model and obstacle model to generate corresponding models;

[0081] Final model generation module: used to combine the generated geographic model and obstacle model to construct a simulation environment model.

[0082] A computer-readable storage medium storing one or more programs according to the present invention includes one or more programs including instructions, and when the instructions are executed by a computing device, the computing device executes any one of the above methods.

Claims

1. A method for constructing a simulation environment model based on AIGC, characterized in that: The following steps are involved: S1. Extract geographic information and obstacle information from the information input by the user through the BiLSTM-CRF model; S2, inputting the geographic information and obstacle information into the trained generative adversarial network for generating geographic models and obstacle models to generate corresponding models; S3, combining the geographical model and obstacle model generated in step S2 to construct a simulation environment model.

2. The method for constructing a simulation environment model based on AIGC according to claim 1, characterized in that: The geographic information extracted in step S1 includes longitude and latitude, terrain and landform; the obstacle information includes the coordinates, elevation, area and collision volume of the obstacle.

3. The method for constructing a simulation environment model based on AIGC according to claim 1, characterized in that: The information input by the user in step S1 includes text information and voice information. The information input by the user is converted into text data after preprocessing and then the BiLSTM-CRF model is used to extract information.

4. The method for constructing a simulation environment model based on AIGC according to claim 3, characterized in that: The preprocessing includes converting the voice information into text data through voice conversion technology.

5. The method for constructing a simulation environment model based on AIGC according to claim 3, characterized in that: In the step S1, the text data is converted into a word vector sequence through a word embedding layer, and the word vector sequence is input into BiLSTM; the output vector of BiLSTM is used as the input vector of the CRF decoding layer, and CRF outputs the optimal label sequence corresponding to the word vector sequence; and the geographic information and obstacle information in the information input by the user are extracted according to the optimal label sequence.

6. The method for constructing a simulation environment model based on AIGC according to claim 5, characterized in that: The calculation formula for the optimal tag sequence is: y * =argmax j∈N score(W,y j ) Where W is the word vector sequence, y j is the jth predicted label sequence of W, N is the total number of all possible predicted label sequences of W, score(W,y j ) is calculated as in, Indicates that the i-th word in W is y j The probability score of the i-th label in , Represents y j The transition probability score between the i-1th and i-th labels in .

7. The method for constructing a simulation environment model based on AIGC according to claim 1, characterized in that: The target loss function calculation formula of the generated adversarial network in step S2 is: Among them, x is the real sample vector, P data is the real data distribution, z is the noise vector that follows the Gaussian distribution P Z A random vector of , G(z) is the generator, D(G(z)) is the discriminator, and E represents the expected value of the specified distribution.

8. The method for constructing a simulation environment model based on AIGC according to claim 1, characterized in that: In step S3, after the geographic model and the obstacle model are combined, additional environmental information input by the user or the system default environmental information is added to construct a simulation environment model; the environmental information includes light, temperature, humidity, precipitation, wind speed, area type, height or depth limit and geographic coordinates.

9. A system for constructing a simulation environment model based on AIGC, characterized in that: The system comprises: Information extraction module: used to extract geographic information and obstacle information from the information input by the user using the BiLSTM-CRF model; Preliminary model generation module: used to input geographic information and obstacle information into the trained generative adversarial network for generating geographic model and obstacle model to generate corresponding models; Final model generation module: used to combine the generated geographic model and obstacle model to construct a simulation environment model.

10. A computer-readable storage medium storing one or more programs, characterized in that: The one or more programs include instructions which, when executed by a computing device, cause the computing device to perform any one of the methods according to claims 1 to 8.

Images