Undercomplete digital twin system, construction method thereof and computer device

Abstract

The invention belongs to the technical field of digital twinning, and particularly relates to an undercomplete digital twinning system, a construction method thereof and a computer device. The method comprises the following steps: S1, inputting running state data of the under-complete digital twin system into a pre-trained complete small model to obtain complete information of the under-complete digital twin system; s2, evaluating whether the complete information meets a preset evaluation standard or not; s3, if the complete information does not meet the evaluation standard, correcting the complete information through a large language model according to success experience and failure experience to obtain corrected complete information; and S2 is executed again according to the corrected complete information until the complete information meets the evaluation standard. According to the invention, the problem of low generalization ability of undercomplete digital twinning in the prior art is solved.

Description

Technical Field

[0001] The present invention belongs to the field of digital twin technology, and specifically relates to an incomplete digital twin system, a construction method thereof, and a computer device. Background Art

[0002] The concept of digital twins can be traced back to NASA's space programs in the 1960s, when they were used to simulate the operating status of spacecraft. With the development of technology, digital twins have evolved from simple simulations to complex simulations. Michael Grieves first formally coined the term "digital twin" in 2002, initially applying it primarily to product design and manufacturing. With the advancement of the Internet of Things, big data, and artificial intelligence, the application of digital twins has continued to expand, from its initial aerospace applications to industrial manufacturing, smart cities, healthcare, and other fields. Digital twins have become a technology that "dynamically presents the past and present behaviors or processes of a physical entity in digital form," used to optimize business processes and decision-making.

[0003] Traditional digital twin technology generally assumes perfect perception of the physical environment and mapping of data to the twin. However, in practice, issues such as communication frequency, latency, and obstruction make perfect data mapping impossible. Consequently, building effective digital twin systems in environments with incomplete data—or "incomplete digital twins"—becomes a pressing challenge.

[0004] In traditional digital twin systems, missing data is typically addressed by completing the missing data. Traditional statistical data processing and completion methods primarily include statistical interpolation, linear fitting, k-nearest neighbor methods, and expectation maximization techniques. However, these methods often rely on strict assumptions, which often compromise their generalization capabilities. While satisfying these assumptions for the current task results in high accuracy, migrating to scenarios where these assumptions are no longer met results in an inability to obtain accurate completed data, impacting subsequent simulations.

[0005] Deep learning-based data interpolation methods primarily focus on short-term prediction and interpolation tasks, making them unsuitable for scenarios like digital twins that require long-term, real-time simulation. Furthermore, these deep learning methods often utilize small neural networks trained in specific scenarios, resulting in low generalization capabilities and poor transferability to new scenarios.

[0006] In summary, the incomplete digital twins in existing technologies suffer from low generalization ability. Summary of the Invention

[0007] The purpose of the present invention is to provide an incomplete digital twin system, a construction method thereof, and a computer device to solve the problem of low generalization ability of incomplete digital twins in the prior art.

[0008] To solve the above technical problems, the present invention provides a technical solution for a method for constructing an incomplete digital twin system, which includes:

[0009] S1. Input the operating status data of the incomplete digital twin system into the pre-trained complete small model to obtain the complete information of the incomplete digital twin system;

[0010] S2. evaluating whether the complete information meets pre-set evaluation criteria;

[0011] S3. If the complete information does not meet the evaluation criteria, the complete information is corrected using a large language model based on the success experience and the failure experience to obtain corrected complete information;

[0012] Re-execute S2 based on the corrected complete information until the complete information meets the evaluation criteria;

[0013] The successful experience is: complete information that meets the evaluation criteria and the corresponding operating status data of the incomplete digital twin system;

[0014] The failure experience is: complete information and corresponding operating status data that do not meet the evaluation criteria.

[0015] The beneficial effect of the above technical solution is that the technical solution of the method for constructing an incomplete digital twin system of the present invention belongs to an improved invention. The incomplete digital twin of the present invention introduces a large language model to correct the complete information output by the complete small model that does not meet the evaluation criteria until the complete information that meets the evaluation criteria is obtained. The present invention introduces a large language model and makes corrections through the powerful generalization ability of the large language model, thereby greatly improving the generalization ability of the incomplete digital twin. The present invention solves the problem of low generalization ability of incomplete digital twins in the prior art.

[0016] Furthermore, the corrected complete information is obtained in the following manner: the successful experience corresponding to the operating status data having a high degree of similarity with the operating status data corresponding to the failed experience is used as the recommended successful experience; a large model prompt word is generated based on the operating status data corresponding to the recommended successful experience and the failed experience, and the large model prompt word is input into the large language model to obtain the corrected complete information.

[0017] Furthermore, the large model prompt words include task description prompts, impact factor prompts, successful experience prompts and failed experience prompts;

[0018] The task description prompt includes an overview of the simulation task, the input and output format definition of the complete small model, and the response format definition of the large model; the influencing factor prompt includes the operating status data corresponding to the failure experience; the successful experience prompt includes the recommended successful experience; the failure experience prompt includes the complete information corresponding to the failure experience and the failure cause analysis.

[0019] Furthermore, the large model prompt words are input into the large language model using CoT prompt technology.

[0020] Furthermore, the successful experience is stored in a memory pool; S3 also includes:

[0021] If the complete information meets the evaluation criteria, the successful experience corresponding to the complete information is stored in the memory pool.

[0022] Furthermore, the incomplete digital twin system includes an intelligent agent, and the operating status data includes historical trajectory information for representing the historical motion status of the intelligent agent, target position guidance information for representing the relationship between the actual position of the intelligent agent and the reference motion trajectory of the intelligent agent, and spatial environment state information for representing the state of the surrounding environment during the historical motion of the intelligent agent; the complete information is decision information including the position information of the intelligent agent.

[0023] Furthermore, the target position guidance information includes the target position guidance vector of each intelligent agent. The target position guidance vector at time t is obtained by: first, projecting the actual position of the intelligent agent at time t onto the reference trajectory of the intelligent agent, and then moving the projection point along the reference trajectory by a set distance to obtain a guidance point. The direction of the target position guidance vector is the direction of the vector from the actual position of the intelligent agent at time t to the guidance point, and the size of the target position guidance vector is the size of the vector from the actual position of the intelligent agent at time t-1 to the actual position of the intelligent agent at time t.

[0024] Furthermore, the complete small model is a spatiotemporal encoder-decoder structure; the input of the time encoder in the spatiotemporal encoder includes the historical trajectory information and target position guidance information; the input of the spatial encoder in the spatiotemporal encoder includes the spatial environment state information.

[0025] Furthermore, the historical trajectory information used when training the complete small model is historical trajectory information with missing data at some moments.

[0026] Furthermore, the historical trajectory information with missing data at some moments is obtained by processing the originally collected historical trajectory information with Gaussian random masking.

[0027] Furthermore, the evaluation criteria include a speed continuity constraint and a collision avoidance constraint; the speed continuity constraint is used to ensure that the speed of the agent is continuous within a unit time interval, and the collision avoidance constraint is used to ensure that the agent will not collide with other agents and obstacles within a set time period in the future;

[0028] When the agent's decision information satisfies both the velocity continuity constraint and the collision avoidance constraint, the decision information is considered to meet the evaluation criteria; otherwise, the decision information is considered to not meet the evaluation criteria.

[0029] Furthermore, if the change in the speed direction within a unit time interval exceeds a direction threshold or the change in the speed magnitude exceeds a magnitude threshold, it is considered that the decision information does not meet the speed continuity constraint.

[0030] Furthermore, when there exists t′∈{0,1,2,...,N} such that the following equation holds, the decision information is considered to not satisfy the collision avoidance constraint:

[0031]

[0032] Where N is the set time period; v i (t) is the velocity vector of agent i at time t; v j (t) is the velocity of agent j at time t; r is the radius of the agent; The position coordinates of agent j at time t output by the complete small model; The position coordinates of agent j at time t output by the complete small model.

[0033] Furthermore, the successful experiences are stored in a memory pool in the form of groups, and the successful experiences are clustered to obtain the groups;

[0034] The recommended successful experience is obtained in the following manner: comparing the similarity between the running status data corresponding to the cluster center of each group and the running status data corresponding to the failed experience, and selecting the successful experience from the group with higher similarity as the recommended successful experience.

[0035] The present invention also provides a technical solution for a computer device: a computer device comprising a processor, wherein the processor is used to execute a computer program to implement the steps of the method for constructing an incomplete digital twin system as described above.

[0036] The present invention also provides a technical solution for an incomplete digital twin system: an incomplete digital twin system, comprising a processor, which is used to execute a computer program to implement the steps of the method for constructing an incomplete digital twin system as described above. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] Figure 1 This is an overall schematic diagram of motion decision-making in an embodiment of a method for constructing an incomplete digital twin system of the present invention;

[0038] Figure 2 A schematic diagram illustrating a task description of an embodiment of a method for constructing an incomplete digital twin system of the present invention;

[0039] Figure 3 A schematic diagram showing the impact factors of an embodiment of a method for constructing an incomplete digital twin system according to the present invention;

[0040] Figure 4 A schematic diagram showing successful experience of an embodiment of a method for constructing an incomplete digital twin system of the present invention;

[0041] Figure 5 A schematic diagram of a failure experience prompt of an embodiment of a method for constructing an incomplete digital twin system of the present invention;

[0042] Figure 6 This is a schematic diagram of a large language model response according to an embodiment of a method for constructing an incomplete digital twin system of the present invention;

[0043] Figure 7 This is a schematic diagram of target position guidance parameters for an embodiment of a method for constructing an incomplete digital twin system of the present invention. DETAILED DESCRIPTION

[0044] In the prior art, the processing of missing data often adopts a data completion algorithm based on statistics or deep learning to fill or interpolate data. However, both the processing method based on statistics and the processing method based on deep learning have the problem of poor generalization ability. When the incomplete digital twin of the present invention generates a state, a large language model is introduced to correct the complete information output by the complete small model that does not meet the evaluation criteria, until the complete information that meets the evaluation criteria is obtained. The present invention introduces a large language model to participate in the state generation process, and makes corrections through the powerful generalization ability of the large language model, thereby greatly improving the generalization ability of the incomplete digital twin. The present invention solves the problem of low generalization ability of incomplete digital twins in the prior art.

[0045] Example of a method for constructing an incomplete digital twin system:

[0046] The method for constructing an incomplete digital twin system in this embodiment combines a small model (SM) with a large model (LM). By leveraging the extensive knowledge of the LM, erroneous information output by the SM is corrected, thereby generating high-quality state data. This enhances cross-scenario generalization and improves real-world applicability.

[0047] To enable the LM to learn effective knowledge from incomplete data, we employ a reconstruction approach to train the model, enabling it to obtain complete information based on existing incomplete data. This approach effectively mitigates the impact of missing data. We also propose a memory-based LM correction method, in which the memory stores past successful experiences and recommends similar experiences to the LM based on the current state. This approach enables the LM to correct complete information in complex situations.

[0048] The construction method referred to in the present invention refers to: constructing corresponding virtual data (i.e., complete information) from incomplete real data, so as to construct a virtual digital twin corresponding to the physical part in the digital twin based on the virtual data.

[0049] This embodiment takes group motion simulation as an example to introduce the construction method of the incomplete digital twin system of the present invention. Figure 1 As shown:

[0050] S1. Input the operating status data of the incomplete digital twin system into the pre-trained complete small model to obtain the complete information of the incomplete digital twin system.

[0051] Specifically for group motion simulation: the operating status data of each intelligent agent in the incomplete digital twin system is input into the pre-trained complete small model (i.e. Figure 1 In the Small Model in the above-mentioned complete small model), the decision information of each agent at the future moment (that is, the above-mentioned complete information) is obtained.

[0052] In other embodiments, the construction method of the incomplete digital twin system of the present invention is applied to other tasks, such as the reconstruction task of missing data, in which case the complete information represents the complete data after reconstruction; the prediction task of the incomplete digital twin system, in which case the complete information represents the predicted state information; the data generation task, in which case the complete information represents the generated data information; other types of control decision tasks, in which case the complete information represents the control decision instruction information.

[0053] The group motion simulation in this embodiment refers to the generation of reasonable motion decisions (such as position, speed, etc.) for each intelligent agent in the incomplete digital twin system based on their historical motion data for the future moment (which can be a certain moment or multiple consecutive moments, that is, a period of time).

[0054] Specifically, in this embodiment, the running state data includes historical trajectory information (i.e. Figure 1Historical Trajectories in the , target position guidance information used to represent the relationship between the current position of the agent and the target position, and spatial environment state information used to represent the state of the surrounding environment during the historical movement of the agent; decision information (i.e. Figure 1 The Decision in includes the location information of the agent, i.e., the location coordinates.

[0055] This embodiment uses the idea of ​​group behavior reconstruction to train the small decision model. The trajectory data used to train the small decision model is a trajectory sequence with some moments missing. The trajectory data between time t-T and time t-1 is used as historical data, and the data at time step t (i.e., the future moment) is used as predicted data.

[0056] The data used to train the small decision model includes the trajectory data set from time t-T to time t Among them, the trajectory data between time t-T and time t-1 is used as the input when training the small decision model, and the trajectory data at time t is the output when training the small decision model:

[0057]

[0058] Where K is the number of agents; is the trajectory sequence of agent i from time t-T to time t; is the position coordinate of agent i at time j. The missing data in is represented by a mask matrix

[0059]

[0060] Among them, when M i When (j) = 0, it means that the position coordinates of agent i at time j are missing; otherwise, M i (j)=1.

[0061] To improve the model's ability to reconstruct missing data, we first manually masked the training data, including random masking of historical data and full masking of predicted data. Random masking of historical data refers to masking a portion of each agent's historical trajectory data, which serves as input to the small decision model. In this example, Gaussian random masking was used to mask the training data. Full masking of predicted data refers to masking all data after time t.

[0062] After artificial mask processing, an artificial mask matrix is ​​generated Element I in i (j) The meaning of M i (j) Same, combined Get the final mask matrix When I i (j) and M i (j) are both 1, corresponding to RM in i (j) = 1; otherwise, RM i (j)=0.

[0063] The decision-making model of this embodiment adopts a spatiotemporal encoder-decoder structure to encode the temporal information and spatial information of the historical data respectively.

[0064] Time encoder: The input data of the time encoder includes historical trajectory information and target position guidance information, that is, the historical trajectory data of each agent from time t-T to time t-1 and target location boot parameters

[0065] in, Represents the historical trajectory data of agent i from time t-T to time t-1; for After transformation, we get: the position of agent i at time t-1 is taken as the origin, the velocity direction is the positive direction of x-axis as the local coordinate system, and the coordinates of all the coordinate systems are Transform to the local coordinate system to get

[0066] It represents the target position guidance parameter of agent i from time t-T to time t-1, as shown in the following formula:

[0067]

[0068] Among them, Pd i (j) represents the optimal state vector of agent i facing the destination at time j (i.e., the target position guidance vector). Figure 7 Here we will show how to get Pd i (j): We will point A i (j) Projected to agent i (i.e. Figure 7 The global path of Trajectory of agent i) (i.e. Figure 7 On the Global path in Figure 7 From this point, we move along the path by a distance of H (this is a hyperparameter, determined empirically, i.e., a set distance) to get point G i (j). Vector Pd i The direction of (j) is equal to the vector G i (j)-Ai (j) direction. Vector Pd i (j) is equal to the size of vector A i (j)-A i The size of (j-1).

[0069] It should be noted that the above global path is represented as the reference path of agent i during movement. i (j) The specific process of projecting the global path to agent i is as follows:

[0070] The global path is represented by a sequence of position coordinates in a global coordinate system. The calculation point A i (j) The distance to the line segment between two adjacent points in the position coordinate sequence, with any one of the two endpoints of the line segment corresponding to the shortest distance as point A on the global path i (j) The corresponding projection point (i.e. Figure 7 red dot in the image).

[0071] The temporal encoder learns temporal correlations via a multi-head self-attention mechanism.

[0072] Spatial encoder: The input data of the spatial encoder includes the spatial environment state information, that is, the spatial environment state perceived by agent i from time t-T to time t-1 Among them, C i (j) represents the grid state perceived by agent i at time j. i (j) is represented by a set of grid states of size s×s. Each grid state contains the relative position and velocity information between the grid and agent i at time j. If an agent exists at the grid, the velocity information is the velocity information of the agent; otherwise, the velocity information is empty.

[0073] The spatial encoder learns spatial correlations through sequentially connected convolutional neural networks and a multi-head self-attention mechanism.

[0074] Decoder: The outputs of the temporal decoder and the spatial decoder are fused (superimposed in this embodiment) and fed into a fully connected neural network, which outputs the state of each agent at time t (i.e., decision information), including the direction and magnitude of the speed.

[0075] In this embodiment, the loss function L of the training decision model is:

[0076]

[0077] in, represents the Hadamard product; represents the reconstructed trajectory;

[0078] The reconstructed trajectory consists of two parts: The position coordinates at time t are the predicted position coordinates output by the small model, and the second is The missing data in (i.e. the corresponding mask matrix The data are reconstructed based on the existing non-missing data by using interpolation or data completion.

[0079] S2. Evaluate whether the complete information (i.e., the decision information in the group simulation motion) meets the pre-set evaluation criteria.

[0080] The evaluation criteria of this embodiment (i.e. Figure 1 Evaluation) includes velocity continuity constraints (i.e. Figure 1 continuity in) and collision avoidance constraints (i.e. Figure 1 The speed continuity constraint ensures the agent's speed is continuous within a unit time interval, and the collision avoidance constraint ensures that the agent does not collide with other agents or obstacles within a set time period. If the agent's decision information satisfies both the speed continuity constraint and the collision avoidance constraint, it is considered to meet the evaluation criteria; otherwise, it is considered to fail the evaluation criteria.

[0081] Speed ​​continuity constraint: In real scenarios, an individual usually does not change its motion pattern frequently. Specifically, the magnitude and direction of the speed remain continuous within a short time interval Δt (i.e., unit time interval).

[0082] For agent i, the change in velocity direction Expressed as:

[0083]

[0084] Among them, v i (t) represents the velocity vector of agent i at time t, ||v i (t)|| represents v i (t) is the modulus of the

[0085] For agent i, the change in velocity direction Expressed as:

[0086]

[0087] Here, ||·||2 represents the two-norm.

[0088] like Exceeds the directional threshold or If the size threshold is exceeded, the decision information (i.e., v i (t)) does not satisfy the velocity continuity constraint.

[0089] Specifically, in this embodiment, the direction threshold is 30°, and the magnitude threshold is 0.7 m / s.

[0090] Collision avoidance constraints: Collision is a key challenge in swarm simulation. To avoid collisions, an agent must predict the motion of surrounding agents and obstacles based on its current velocity and choose a velocity that will prevent collisions within a set time period N. The condition for determining whether agent i will collide with agent j within N steps is as follows, where r is the radius of the agent.

[0091] Make the following equation true

[0092]

[0093] Where N is the set time period; v i (t) is the velocity vector of agent i at time t output by the decision-making model; v j (t) is the speed of agent j at time t output by the decision model; r is the radius of the agent; The position coordinates of agent j at time t output by the small model; The position coordinates of agent j at time t output by the small model.

[0094] S3. If the complete information does not meet the evaluation criteria, the complete information is corrected through the large language model based on the successful experience and the failed experience to obtain the corrected complete information; S2 is re-executed based on the corrected complete information until the complete information meets the evaluation criteria; the successful experience is: the complete information that meets the evaluation criteria and its corresponding operating status data of the incomplete digital twin system; the failed experience is: the complete information that does not meet the evaluation criteria and its corresponding operating status data.

[0095] Specifically for group motion simulation: if the decision information does not meet the evaluation criteria, then the large language model (i.e. Figure 1 The decision information is modified using the Large Language Model (LLM) in the decision tree to obtain modified decision information; S2 is re-executed according to the modified decision information until the decision information meets the evaluation criteria.

[0096] Among them, successful experience (i.e. Figure 1 The success) is the decision information that meets the evaluation criteria and the corresponding operating status data of each intelligent agent; the failure experience (i.e. Figure 1 Where fail) refers to the decision information that does not meet the evaluation criteria and the corresponding operating status data of each intelligent agent.

[0097] The revised decision information is obtained in the following way: Figure 1 The successful experience corresponding to the running status data with higher similarity to the running status data corresponding to the failed experience in the memory is recommended as the successful experience (i.e. Figure 1 Top K in the recommendation); Generate large model prompt words (i.e. Figure 1 Prompts in the large model), input the large model prompt words into the large language model to obtain the revised decision information.

[0098] To enhance the large language model's ability to reflect on failed decisions, we introduced a memory pool to store past successful experiences. This pool then recommends similar experiences based on the agent's current influencing factors (i.e., the input data for the small decision-making model, which in this example includes historical trajectories, spatial environment state information, and goal-directed parameters). These influencing factors generate prompts that are fed into the large model to generate more optimal movement decisions.

[0099] It mainly includes two parts: memory storage and recommendation and prompt generation.

[0100] (1) Memory storage and recommendation: the kth successful experience E in the memory pool k It contains four parts, namely historical trajectory Space environment status information Goal-oriented parameters and decision-making Among them, j represents the intelligent agent that provides the successful experience.

[0101] When the memory capacity is small, similarity can be calculated across all experiences in the memory pool to find matching items. However, when the memory size reaches tens of thousands, brute-force comparison becomes prohibitively expensive. To improve efficiency, we introduce a grouped storage method based on K-means clustering. During the recommendation phase, only the cluster center is compared, and the top k experiences are retrieved from the most similar group. Similarity calculation is based on the proximity between influencing factors, considering three aspects: historical trajectory similarity, spatial environment state information similarity, and goal-oriented parameter similarity. The following uses the example of agent i and the kth experience at time t-1 to explain the three similarity calculation methods in detail.

[0102] Historical trajectory similarity: Based on the dynamic time warping (DTW) algorithm, calculation and The similarity of . Define a size of T R ×T E The similarity matrix S, where T R With T E They are and The element S(j,l) in the matrix S represents the trajectory point RA i (j) with The cumulative similarity is calculated as follows:

[0103]

[0104] in, Represents trajectory point RA i (j) with The Euclidean distance between .

[0105] Similarity of spatial environment state information: The impact of different grids on the agent varies depending on the relative distance and direction angle of the grid. Objects that are adjacent and located in front of the agent have a greater impact.

[0106] Distance influence function I in local coordinate system dis (d j )for:

[0107]

[0108] Among them, d j represents the distance from the jth grid to the origin (i.e. the location of the agent), It means that the pre-set distance affects the variance in the Gaussian function.

[0109] The direction angle influence function in the local coordinate system is:

[0110]

[0111] Among them, a j represents the angle between the jth grid and the x-axis (i.e., the location of the agent), It means that the pre-set angle affects the variance in the Gaussian function.

[0112] but with C j The similarity S between (t-1) per for:

[0113]

[0114] in, express and The cosine similarity between ; the numerator represents the grid Hezhong There are the number of elements, that is, the grid Hezhong The number of non-empty (NULL) values.

[0115] Goal-oriented parameter similarity: Calculate the current goal-oriented parameters and The cosine similarity S betweendes ,Right now:

[0116]

[0117] The final similarity score is the weighted sum of the three: S i,k =ω traj ·S traj +ω per ·S per +ω des ·S des , where ω traj 、ω per and ω des They represent the weights of the three similarity components, ranging from [0, 1], and are all set to 1 in this paper.

[0118] (2) Prompt generation: The large model prompt words include task description prompts (i.e. Figure 1 Task description in the task description), impact factor prompts, successful experience prompts and failed experience prompts.

[0119] like Figure 2 As shown, the task description prompts include an overview of the simulation task, the input and output format definitions of the small decision model, and the response format definition of the large model. To simplify the decision-making, we restructured the motion decision into a classification task: velocity magnitude is divided into 6 categories and velocity direction is divided into 5 categories.

[0120] Figure 2 The first paragraph mainly includes the simulation task overview and the definition of the input and output format of the decision model. The second paragraph defines the speed of decision making (i.e. Figure 2 velocity magnitude) and velocity direction (i.e. Figure 2 The velocitydirection in is classified according to the corresponding values. Figure 2 The "Final Result" section in the Response Format section defines the large model response format. This section defines the format in which the language large model outputs the revised decision information.

[0121] The task description prompts that when the simulation task remains unchanged, each iteration remains fixed.

[0122] like Figure 3 As shown, the impact factor prompt includes three key impact factors: historical trajectory Space environment status information C i (t-1) and goal-oriented parameter Pd i(t-1). Historical trajectories are guided by "Historical Trajectory" and list discrete trajectory points; the perception grid is guided by "External Environment" and traverses the grid states containing objects; the goal-oriented parameters are guided by "Destination Direction" and list corresponding values.

[0123] like Figure 4 As shown, the successful experience tips include recommended successful experiences. Use the above method to obtain the most similar k successful experiences. Each experience contains the impact factor and decision-making decision making Recorded as "Final Result". Each experience uses the same format and constitutes the context learning input of the large model.

[0124] like Figure 5 As shown, the failure experience prompt includes the decision information corresponding to the failure experience input into the decision correction unit, as well as the failure cause analysis. A "Failure Reason" entry is introduced, allowing the simulation evaluation module to identify the failure cause. For example, if the failure is due to a large change in velocity direction, it can be expressed as: "The current velocity direction change is too large, violating velocity direction continuity." If the failure is due to a collision, the position and velocity information of the collision object is added.

[0125] Due to the complexity of the decision-making process in crowd simulation tasks, directly generating decisions from a language model can lead to hallucinations. Instead, we use a Chain of Thought (CoT) prompting technique to guide the large model to generate a series of reasoning sentences.

[0126] The above prompts are input into the large language model. Since the task description prompts define the large model response format, the output of the large language model is as follows: Figure 6 As shown, after decoding, the corrected decision information can be obtained (i.e. Figure 6 fan chart below). Figure 6 The black arrow in the fan diagram indicates the direction of velocity (i.e. direction), and the blue arc indicates the magnitude of velocity (i.e. speed).

[0127] S4. Control each intelligent agent based on the decision information that meets the evaluation criteria.

[0128] Computer device embodiment:

[0129] A computer device includes a processor configured to execute a computer program to implement the steps of the method for constructing an incomplete digital twin system as described above. The specific method for constructing an incomplete digital twin system has been described in sufficient detail in the embodiment of the method for constructing an incomplete digital twin system and will not be repeated here.

[0130] Specifically, the processor may be a CPU, or other general-purpose processor, a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor may be a microprocessor or any conventional processor, etc. The processor may also be a processor that supports the Advanced RISC Machine (ARM) architecture.

[0131] Examples of incomplete digital twin systems:

[0132] An incomplete digital twin system includes a processor configured to execute a computer program to implement the steps of the method for constructing an incomplete digital twin system as described above. The specific method for constructing an incomplete digital twin system has been described in sufficient detail in the embodiment of the method for constructing an incomplete digital twin system and will not be repeated here.

[0133] Specifically, the processor may be a CPU, or other general-purpose processor, a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor may be a microprocessor or any conventional processor, etc. The processor may also be a processor that supports the Advanced RISC Machine (ARM) architecture.

[0134] The present invention has the following characteristics:

[0135] When making decisions, the incomplete digital twin of the present invention uses a large language model to correct decisions output by the small decision-making model that do not meet the evaluation criteria, until decision information that meets the evaluation criteria is obtained. By incorporating a large language model into the decision-making process and leveraging its powerful generalization capabilities to perform decision corrections, the present invention significantly improves the generalization capabilities of the incomplete digital twin's decision-making process. This solves the problem of low generalization capabilities in the decision-making of incomplete digital twins in the prior art.

[0136] Finally, it should be noted that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art can still modify the technical solutions described in the aforementioned embodiments without inventive effort, or replace some of the technical features therein with equivalents. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of the present invention shall be included within the scope of protection of the present invention.

Claims

1. A method for constructing an incomplete digital twin system, characterized in that: The method includes: S1. Input the operating status data of the incomplete digital twin system into the pre-trained complete small model to obtain the complete information of the incomplete digital twin system; S2. evaluating whether the complete information meets pre-set evaluation criteria; S3. If the complete information does not meet the evaluation criteria, the complete information is corrected using a large language model based on the success experience and the failure experience to obtain corrected complete information; Re-execute S2 based on the corrected complete information until the complete information meets the evaluation criteria; The successful experience is: complete information that meets the evaluation criteria and the corresponding operating status data of the incomplete digital twin system; The failure experience is: complete information and corresponding operating status data that do not meet the evaluation criteria.

2. The method for constructing an incomplete digital twin system according to claim 1, characterized in that: The corrected complete information is obtained in the following manner: a successful experience corresponding to running status data having a high similarity to running status data corresponding to a failed experience is used as a recommended successful experience; A large model prompt word is generated according to the operation status data corresponding to the recommendation success experience and the failure experience, and the large model prompt word is input into the large language model to obtain corrected complete information.

3. The method for constructing an incomplete digital twin system according to claim 2, characterized in that: The large model prompt words include task description prompts, impact factor prompts, successful experience prompts and failed experience prompts; The task description prompt includes an overview of the simulation task, the input and output format definition of the complete small model, and the response format definition of the large model; the influencing factor prompt includes the operating status data corresponding to the failure experience; the successful experience prompt includes the recommended successful experience; the failure experience prompt includes the complete information corresponding to the failure experience and the failure cause analysis.

4. The method for constructing an incomplete digital twin system according to claim 2 or 3, characterized in that: The large model prompt words are input into the large language model using CoT prompt technology.

5. The method for constructing an incomplete digital twin system according to claim 1, wherein: The successful experience is stored in the memory pool; S3 also includes: If the complete information meets the evaluation criteria, the successful experience corresponding to the complete information is stored in the memory pool.

6. The method for constructing an incomplete digital twin system according to claim 1, wherein: The incomplete digital twin system includes an intelligent agent, and the operating status data includes historical trajectory information for representing the historical motion status of the intelligent agent, target position guidance information for representing the relationship between the actual position of the intelligent agent and the reference motion trajectory of the intelligent agent, and spatial environment state information for representing the state of the surrounding environment during the historical motion of the intelligent agent; the complete information is decision information including the position information of the intelligent agent.

7. The method for constructing an incomplete digital twin system according to claim 6, characterized in that: The target position guidance information includes the target position guidance vector of each intelligent agent. The target position guidance vector at time t is obtained in the following manner: first, the actual position of the intelligent agent at time t is projected onto the reference trajectory of the intelligent agent, and then the projection point is moved a set distance along the reference trajectory to obtain a guidance point. The direction of the target position guidance vector is the direction of the vector from the actual position of the intelligent agent at time t to the guidance point, and the size of the target position guidance vector is the size of the vector from the actual position of the intelligent agent at time t-1 to the actual position of the intelligent agent at time t.

8. The method for constructing an incomplete digital twin system according to claim 6, wherein: The complete small model is a spatiotemporal encoder-decoder structure; the input of the time encoder in the spatiotemporal encoder includes the historical trajectory information and target position guidance information; the input of the space encoder in the spatiotemporal encoder includes the spatial environment state information.

9. The method for constructing an incomplete digital twin system according to claim 6, wherein: The historical trajectory information used when training the complete small model is the historical trajectory information with missing data at some moments.

10. The method for constructing an incomplete digital twin system according to claim 9, characterized in that: The historical trajectory information with missing data at some moments is obtained by processing the original collected historical trajectory information with Gaussian random masking.

11. The method for constructing an incomplete digital twin system according to claim 6, wherein: The evaluation criteria include a speed continuity constraint and a collision avoidance constraint; the speed continuity constraint is used to ensure that the agent's speed is continuous within a unit time interval, and the collision avoidance constraint is used to ensure that the agent will not collide with other agents or obstacles within a set time period in the future; When the agent's decision information satisfies both the velocity continuity constraint and the collision avoidance constraint, the decision information is considered to meet the evaluation criteria; otherwise, the decision information is considered to not meet the evaluation criteria.

12. The method for constructing an incomplete digital twin system according to claim 11, characterized in that: If the change in speed direction within a unit time interval exceeds a direction threshold or the change in speed magnitude exceeds a magnitude threshold, it is considered that the decision information does not meet the speed continuity constraint.

13. The method for constructing an incomplete digital twin system according to claim 11, wherein: When there exists t′∈{0,1,2,...,N} such that the following equation holds, the decision information is considered to not satisfy the collision avoidance constraint: Where N is the set time period; v i (t) is the velocity vector of agent i at time t; v j (t) is the velocity of agent j at time t; r is the radius of the agent; The position coordinates of agent j at time t output by the complete small model; The position coordinates of agent j at time t output by the complete small model.

14. The method for constructing an incomplete digital twin system according to claim 2, wherein: The successful experiences are stored in a memory pool in the form of groups, and the successful experiences are clustered to obtain the groups; The recommended successful experience is obtained in the following manner: comparing the similarity between the running status data corresponding to the cluster center of each group and the running status data corresponding to the failed experience, and selecting the successful experience from the group with higher similarity as the recommended successful experience.

15. A computer device comprising a processor, characterized in that: The processor is used to execute a computer program to implement the steps of the method for constructing an incomplete digital twin system as described in any one of claims 1 to 14.

16. An incomplete digital twin system, comprising a processor, characterized in that: The processor is used to execute a computer program to implement the steps of the method for constructing an incomplete digital twin system as described in any one of claims 1 to 14.

Images