A light and shadow scheme optimization method and system based on digital twin and artificial intelligence

Abstract

The present invention discloses a light and shadow scheme optimization method and system based on digital twins and artificial intelligence, belonging to the field of data processing technology, including converting user needs into semantic vectors, mapping user intention vectors to obtain light and shadow control parameters, simulating and calculating the light and shadow control parameters, outputting simulation images and light and shadow physical indicators, extracting key physical quantities under the light and shadow state in the simulation images, and integrating them to form simulation feedback indicators, calculating the difference between the simulation semantic vector and the user intention vector to obtain a semantic error vector, inputting the semantic error vector and light and shadow control parameters, outputting the updated amount of the light and shadow control parameters, establishing a convergence data function, judging the optimization situation, terminating the operation when a preset scenario is obtained, and outputting the final result. The light and shadow scheme optimization method and system based on digital twins and artificial intelligence realizes a closed-loop adaptive optimization process by constructing a linkage mechanism from semantic understanding, parameter reasoning to feedback correction.

Description

Technical Field

[0001] The present invention belongs to the field of data processing technology, and in particular relates to a light and shadow scheme optimization method and system based on digital twins and artificial intelligence. Background Art

[0002] With the rapid development of digital twin technology and artificial intelligence, optimizing physical environments based on virtual simulations has become a key approach across multiple industries, including architectural design, urban lighting, stage sets, and film and television production. The design and optimization of lighting schemes, a crucial component of the spatial visual experience, increasingly relies on sophisticated simulation and intelligent assistance systems. In practice, professional designers typically need to manually set a series of lighting parameters (such as light source type, position, direction, illumination intensity, color temperature, and material reflectivity), then repeatedly test and adjust them through simulation systems to achieve the lighting effects that meet the needs of specific scenarios. However, this process not only relies heavily on professional knowledge and experience, but also suffers from long debugging cycles and low efficiency, making it difficult to adapt to the real-world demands of rapid response and personalized design.

[0003] In order to solve the problems of low efficiency of manual parameter adjustment and unfriendly user interaction, some studies in recent years have begun to try to introduce artificial intelligence to assist in light and shadow optimization. For example, machine learning models are used to predict user preferences, or reference designs with similar light and shadow styles are recommended based on image retrieval. Although these attempts have improved the efficiency of certain links, overall, existing technologies still have obvious defects in dealing with the following three key challenges: (1) Mismatch in the way users express their needs: When describing their light and shadow needs, users often use subjective and vague language (such as "bright and transparent" and "warm and dreamy"), while simulation systems require clear structured parameters. Existing methods are difficult to achieve accurate conversion from subjective intentions to physical parameters; (2) Lack of feedback loop in the optimization process: Most systems regard AI models as static recommenders, ignoring the fact that digital twin simulation systems can provide high-value lighting feedback data, and fail to form a dynamic optimization mechanism of "intention → parameter → simulation → reverse correction"; (3) Lack of explainability and adaptability in the parameter adjustment process: Although traditional black box optimization methods (such as genetic algorithms and particle swarm algorithms) can perform parameter search, the process is not explainable and it is difficult to effectively adapt to complex architectural spatial structures or user preferences. Most existing methods are "static one-way processes": user expression → parameter recommendation → simulation verification. Once the recommended solution deviates from expectations, it needs to be re-entered. The lack of learning and adjustment capabilities seriously limits the intelligent application potential of digital twin systems in light and shadow design.

[0004] To this end, we propose a lighting and shadow solution optimization method and system based on digital twins and artificial intelligence to solve the above problems. Summary of the Invention

[0005] The purpose of this invention is to solve the problems of difficult parameter conversion and lack of optimization feedback in the existing technology, and to propose a light and shadow solution optimization method and system based on digital twins and artificial intelligence.

[0006] In order to achieve the above object, the present invention adopts the following technical solutions:

[0007] A lighting and shadow scheme optimization method based on digital twins and artificial intelligence, including:

[0008] Input user needs, convert them into semantic vectors, and obtain the user intention vector by weighted averaging the semantic vectors generated by different categories of input;

[0009] Mapping the user intention vector in a graph neural network to obtain light and shadow control parameters, wherein the mapping is performed by using the semantic features in the user intention vector as nodes in the graph neural network and the light and shadow control parameters as mapping targets of the nodes, and adjusting the edge weights of the nodes in the graph to ensure the accuracy of the mapping;

[0010] Perform simulation calculations on light and shadow control parameters, output simulation images and light and shadow physical indicators, extract key physical quantities under the light and shadow state from the simulation images, and integrate them to form simulation feedback indicators, wherein the simulation calculations are based on performing lighting simulation in a preset digital twin 3D space, and the simulation images are the rendering results of the scene under given parameters;

[0011] The user intention vector is aligned with the simulation feedback indicator. By constructing a cross-modal mapping relationship, the simulation feedback indicator is mapped to obtain a simulation semantic vector. The difference between the simulation semantic vector and the user intention vector is calculated to obtain a semantic error vector. The semantic error vector is perceptually normalized to obtain a global weighted error. In the cross-modal mapping relationship, nodes represent semantic labels, and edges represent dependencies between semantic labels and physical indicators.

[0012] Establishing an optimization strategy model, inputting a semantic error vector and a light and shadow control parameter, and outputting an update amount of the light and shadow control parameter, wherein the optimization strategy model adopts a graph attention network structure and is configured to map the semantic error vector to a gradient adjustment amount of the light and shadow control parameter;

[0013] Establish a convergence data function, input the semantic error vector, the global weighted error and the adjustment amount of the light and shadow control parameters for judgment. When the preset scenario is obtained, terminate the operation and output the final result.

[0014] Preferably, user demand conversion includes natural language processing and image input processing, wherein the natural language processing converts user demand into a vector representation through a pre-trained language model; and the image input processing extracts visual features of the image through a convolutional neural network to obtain a vector representation.

[0015] Preferably, the features extracted in natural language processing include color temperature, brightness, light source and time.

[0016] Preferably, the features extracted during image input processing include light source position, color temperature and brightness distribution.

[0017] Preferably, the key physical quantities extracted from the simulation image include global average illumination, illumination uniformity, maximum brightness and regional contrast, main light spot area and distribution range, and shadow boundary clarity.

[0018] Preferably, an orthogonality regularization term is introduced into the optimization strategy model, and the difference between the combined gradient and the independent gradient is used as the regularization term to penalize non-orthogonal gradient directions and suppress the directional offset of the adjustment of the same physical parameter by error signals from different semantic directions.

[0019] Preferably, the preset scenarios include:

[0020] If the current weighted semantic error is less than the global threshold, the user's goal is considered to have been achieved and the process stops automatically.

[0021] The error changes in the past w rounds tend to be stable, that is, the error change trend is less than the trend stability threshold, indicating that the optimization falls into the stable fluctuation area and actively converges;

[0022] The parameter update amplitude continues to shrink, indicating that the strategy is approaching ineffective adjustment and active convergence.

[0023] A lighting and shadow optimization system based on digital twins and artificial intelligence, including:

[0024] An input processing module is configured to convert input lighting requirements into structured semantic vectors, wherein the input lighting requirements include natural language input and / or image input; and obtain a user intention vector by weighted averaging the semantic vectors generated by different types of input;

[0025] The parameter mapping module is configured to convert the user semantic vector into specific light and shadow control parameters;

[0026] The simulation calculation module is configured to simulate and calculate the light and shadow control parameters, output a simulation image and light and shadow physical indicators, extract key physical quantities under the light and shadow state from the simulation image, and integrate them to form simulation feedback indicators;

[0027] The error assessment module is configured to align the user intent vector with the simulation feedback indicator. By constructing a cross-modal mapping relationship, the simulation feedback indicator is mapped to a simulation semantic vector. The difference between the simulation semantic vector and the user intent vector is calculated to obtain a semantic error vector. The semantic error vector is perceptually normalized to obtain a global weighted error.

[0028] The strategy optimization module is configured to input the semantic error vector and the light and shadow control parameters and output the updated amount of the light and shadow control parameters;

[0029] The optimization control and output module is set to make judgments based on the convergence judgment function, the input semantic error vector, the global weighted error and the adjustment amount of the light and shadow control parameters. When the preset scenario is obtained, the operation is terminated and the final result is output.

[0030] To sum up, the technical effects and advantages of the present invention are as follows: the light and shadow scheme optimization method and system based on digital twins and artificial intelligence automatically generates structured light and shadow parameters based on the user's subjective expression, performs simulation verification with the help of a digital twin environment, and reversely corrects the system's internal model through the simulation results, gradually optimizes the output scheme, and realizes a closed-loop adaptive optimization process. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] Figure 1 is a flow chart of the method of the present invention;

[0032] Figure 2 Schematic diagram of the system structure of the present invention. DETAILED DESCRIPTION

[0033] The technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments.

[0034] like Figure 1 As shown in FIG, a light and shadow scheme optimization method based on digital twins and artificial intelligence includes:

[0035] Input user needs, convert them into semantic vectors, and obtain the user intention vector by weighted averaging the semantic vectors generated by different categories of input;

[0036] Mapping the user intention vector in a graph neural network to obtain light and shadow control parameters, wherein the mapping is performed by using the semantic features in the user intention vector as nodes in the graph neural network and the light and shadow control parameters as mapping targets of the nodes, and adjusting the edge weights of the nodes in the graph to ensure the accuracy of the mapping;

[0037] Perform simulation calculations on light and shadow control parameters, output simulation images and light and shadow physical indicators, extract key physical quantities under the light and shadow state from the simulation images, and integrate them to form simulation feedback indicators, wherein the simulation calculations are based on performing lighting simulation in a preset digital twin 3D space, and the simulation images are the rendering results of the scene under given parameters;

[0038] The user intention vector is aligned with the simulation feedback indicator. By constructing a cross-modal mapping relationship, the simulation feedback indicator is mapped to obtain a simulation semantic vector. The difference between the simulation semantic vector and the user intention vector is calculated to obtain a semantic error vector. The semantic error vector is perceptually normalized to obtain a global weighted error. In the cross-modal mapping relationship, nodes represent semantic labels, and edges represent dependencies between semantic labels and physical indicators.

[0039] Establishing an optimization strategy model, inputting a semantic error vector and a light and shadow control parameter, and outputting an update amount of the light and shadow control parameter, wherein the optimization strategy model adopts a graph attention network structure and is configured to map the semantic error vector to a gradient adjustment amount of the light and shadow control parameter;

[0040] Establish a convergence data function, input the semantic error vector, the global weighted error and the adjustment amount of the light and shadow control parameters for judgment. When the preset scenario is obtained, terminate the operation and output the final result.

[0041] The specific steps of this embodiment are as follows:

[0042] Step 1: Semantically structured expression of user intent

[0043] The goal of this step is to convert the user's lighting requirements (whether through natural language description or image input) into a structured semantic vector This vector contains the specific characteristics of the user's desired lighting and shadow effects, becoming the basic input for subsequent steps and driving the entire optimization process. Users' lighting and shadow requirements are often vague and subjective, and how to effectively transform these requirements into computable lighting and shadow parameters is the core issue to be addressed in this step.

[0044] First, users can express their needs through natural language descriptions (such as "there is warm sunset light in the room") or upload images (such as reference photos with light and shadow effects). For different input forms, the system needs to adopt different processing methods and finally generate a universal semantic vector. , so that it can be used in subsequent steps.

[0045] Natural Language Processing

[0046] For natural language input, the system converts the input text into a vector representation using a pre-trained language model (such as GPT-4 or BERT). Language models not only understand the meaning of words but also capture the complex semantic associations in the context. For example, when processing the phrase "warm sunset light," the system can extract the following key information from the text:

[0047] Color temperature: A warm hue indicates that the light source has a lower color temperature, close to red or orange.

[0048] Brightness: Soft brightness means the light is relatively even without strong spots or contrast.

[0049] Light source: Natural light at sunset, typically with low-intensity lighting and gradual brightness changes.

[0050] Time: Sunset, usually refers to the twilight period, meaning the direction and intensity of light changes over time.

[0051] By converting these features into a high-dimensional vector, the semantic vector generated by the system Contains the above information.

[0052] Image input processing

[0053] If the user provides an image input, the system will use a convolutional neural network (CNN) to extract the visual features of the image, especially the light and shadow distribution, brightness changes, color tone and other information in the image. The purpose of image processing is to extract information such as light source, brightness, spot distribution, etc., so as to form a structured vector .

[0054] Convolutional neural networks first extract features from an image, identifying the location, intensity, and direction of the light source. For example, in an image of a sunset, a CNN can extract the following features:

[0055] Light source position: the position of the sun and the direction of light;

[0056] Color temperature: The color temperature of the image is warm (low color temperature);

[0057] Brightness distribution: the distribution of bright parts in the image.

[0058] In this way, CNN transforms the image into a feature vector containing the above information This vector is consistent with the semantic vector of the text input Similarly, it represents the user's specific needs for light and shadow.

[0059] Multimodal fusion

[0060] To handle situations involving different input forms (e.g., both natural language and images), the system fuses the language description and image features to generate a unified semantic vector. By using a weighted average approach, the system is able to combine the two into a comprehensive vector based on the contribution of each input:

[0061] ;

[0062] here, is a tuning parameter used to control the relative importance of language input and image input in the final semantic vector. Through this weighted approach, the system can flexibly adjust the model output according to the input type, so that the contribution of each input method to the result can be appropriately reflected. For example, when the user inputs "natural light in the morning", the text information will have a greater impact on the semantic vector, so It may be too large; if the user uploads an image, the weight of the image feature will be relatively large. This approach ensures effective fusion and unified representation of multimodal inputs.

[0063] Generated semantic vector

[0064] Finally, the output of step 1 is a comprehensive semantic vector , which contains all the key information of the user's lighting requirements (such as color temperature, brightness, time, light source type, etc.).

[0065] For example, if the user describes "soft light in the early morning", the system will extract features such as color temperature, brightness, and light source direction from the text, and combine it with the morning scene image that may be uploaded to generate a semantic vector. It also includes characteristics of the early morning, such as low color temperature, soft brightness, and low light intensity. These characteristics will directly affect the generation of light and shadow parameters in subsequent steps.

[0066] Step 2: Inference of the mapping from intent vector to physical parameters

[0067] The goal of this step is to transform the semantic vector generated in step 1 Converted into specific light and shadow control parameters These parameters will be used to drive subsequent lighting simulation and optimization. This process not only involves converting the user's lighting requirements into physical parameters, but also requires using a graph neural network (GNN) to learn how to accurately map abstract lighting characteristics to physical control signals. Through this step, we ensure that the user's subjective lighting requirements are accurately reflected in the physical space and provide a foundation for subsequent lighting simulation.

[0068] The mapping process uses a graph neural network (GNN), in which each semantic feature is represented as a node in the network, and the physical parameters are the mapping targets of these nodes. The GNN learns the relationship between semantic features and physical parameters and adjusts the edge weights between nodes in the graph to ensure that each feature is accurately mapped to the physical parameter. For example, the feature "soft sunset light" in the semantic vector can be mapped to lower light intensity (soft brightness) and higher color temperature (warm light).

[0069] The calculation formula during the mapping process is as follows:

[0070] ;

[0071] in, is a semantic vector The A feature represents a light and shadow feature (such as color temperature, brightness, etc.), is the weight associated with the feature, indicating the degree of influence of the feature on the physical parameter. It is obtained through graph neural network training, which determines the final physical parameters of each feature pair Through this formula, the system can convert the user's light and shadow requirements into specific control parameters.

[0072] In the specific implementation, the semantic vector Each dimension of is mapped to a physical parameter through the graph neural network. Suppose the user inputs "warm sunset light", the system will extract the following features from the semantic vector:

[0073] Color Temperature: Lower color temperature (warmer light source).

[0074] Brightness: Soft brightness (low-intensity light).

[0075] Light source type: Sunset light source (natural light).

[0076] Light source direction: The direction of the light source close to the sunset angle.

[0077] These features will be mapped into specific physical parameters. For example, the system will generate control parameters such as light source position, color temperature, and light intensity for subsequent light and shadow simulation.

[0078] In order to improve the accuracy of the mapping process and avoid overfitting, we introduce the regularization term The regularization term ensures that the generated physical parameters are within a reasonable range and meet the requirements of the actual physical environment by penalizing excessive weights. The calculation formula of the regularization term is:

[0079] ;

[0080] in, The first physical parameters, is the target value of the physical parameter (e.g., set according to lighting design standards or reference simulation data), is a regularization hyperparameter that controls the strength of regularization. By adding this regularization term, the system can ensure the stability of physical parameters, making the output more consistent with the requirements of actual lighting design.

[0081] By combining graph neural networks and regularization terms, the system can automatically optimize the mapping relationship between semantic features and physical parameters, ensuring that each user-entered lighting requirement can be converted into a set of control parameters that conform to physical laws. These control parameters will be used in subsequent lighting simulation and optimization steps to drive the simulation system to generate lighting effects that meet user requirements. By combining graph neural networks (GNNs) and regularization terms, this step not only efficiently converts semantic features into physical control parameters, but also improves the accuracy and stability of the conversion through automatic learning and regularization. This allows the user's lighting requirements to be accurately expressed in physical simulations and provides a solid foundation for subsequent lighting simulation and optimization.

[0082] Step 3: Perform simulation calculations in the digital twin environment

[0083] The task of this step is to control the light and shadow parameters generated in step 2 in the digital twin environment. Perform simulation calculations and output simulation images and light and shadow physical indicators . Input is light and shadow control parameters , where each Represents a structured light and shadow physical parameter, which may include light source position, color temperature, brightness, direction, divergence angle, ambient reflectivity, etc. All parameters are mapped by the previous step of the graph neural network and have structural consistency. Input to simulation module In the pre-set digital twin 3D space, lighting simulation is performed. The twin scene is built with a physics-based rendering engine that supports path tracing or global illumination calculation methods. It can simulate processes such as light propagation, object material response, and spatial reflection to generate realistic and credible simulation images. . Simulation image Is the scene with given parameters The rendering results under , the image size, resolution, and viewing angle are aligned with the twin scene standard to ensure consistency in subsequent analysis.

[0084] In order to provide quantitative feedback capabilities, the system will Input to the lighting index calculation module In the process, the scene geometry and material information are combined to extract the key physical quantities under the light and shadow state to form the index vector These indicators fall into the following categories:

[0085] : global average illumination (unit: Lux);

[0086] : Light uniformity ;

[0087] : Maximum brightness and regional contrast;

[0088] : Main spot area and distribution range;

[0089] : Shadow boundary clarity (measured by edge gradient);

[0090] In order to enhance the controllability of the ability to restore real scenes, we introduce an optional parameter: simulation environment correction item , which is derived from the sampling data of the real physical environment by the on-site sensor and is used to correct the possible parameter deviation between the twin model and the physical world. The authenticity weight in the simulation process can be defined as:

[0091] ;

[0092] in, represents the corrected simulation image, It is a reference image generated by sampling values ​​of on-site sensors (such as environmental HDR map, reflectivity map), This option can be enabled when real environment data is available to enhance the physical consistency of the simulation.

[0093] The final output includes: (1) simulation image (or correct the image ), used for human perception; (2) structured feedback indicators , serving as the basic data for subsequent error modeling and adaptive optimization.

[0094] Step 4: Semantic Error Modeling and Evaluation of Simulation Feedback

[0095] The main task of this step is to establish a semantic error evaluation mechanism in the system: the user intention vector output in step 1 is converted into The simulation feedback index output from step 3 Align by building a cross-modal mapping relationship Physical feedback Map back to the semantic space to form a simulated semantic vector , and finally calculate the difference between the two This is the key step for the entire system to achieve "closed-loop feedback". Without this mechanism, the system cannot determine whether the current simulation is consistent with the user's expectations, and cannot enter the next step of parameter adaptive correction. In the previous step (step 3), we obtained the simulation image and structured feedback indicators These physical indicators are quantitative descriptions of the simulated light and shadow results. The user intention semantic vector generated in the first step is , which includes the user's subjective light and shadow requirements. The two are in completely different data spaces, and it is difficult to compare them in traditional methods. This invention designs a semantic mapping mechanism to convert physical feedback into semantic estimation values. , and then quantify the error.

[0096] First, we construct a physical feedback Nonlinear mapper to semantic space , used to learn the following mapping relationship:

[0097] ;

[0098] Essentially, it is a lightweight graph neural network, where nodes represent semantic labels (such as "soft," "bright," and "spotlight"), and edges represent dependencies between them and physical indicators. Unlike traditional linear regression or MLP, this graph structure can express complex structures where "a semantic concept is determined by multiple physical indicators." For example, "soft" is controlled by the brightness gradient. , illumination uniformity And the main light source diffusion angle and other factors.

[0099] In order to enhance the interpretability and accuracy of the mapping, we introduced a cross-dimensional perception regularization term during network training. , used to penalize the impact of misjudgment of "highly sensitive semantics", so that the error can better reflect the difference in user experience. It is defined as follows:

[0100] ;

[0101] in, Indicates the The perceptual weight of each semantic dimension is obtained based on psychophysical research or system data fitting, reflecting the impact of the dimension on human perception. For example, Usually higher than . This regularization term does not affect the error output , but will be used in subsequent optimization priority judgment and gradient scheduling.

[0102] Next, we calculate the semantic error vector , that is, the difference between user intention and simulation semantics:

[0103] ;

[0104] Each dimension Indicates that the system The degree of deviation in the semantic direction. It is a structured vector in which each dimension has independent interpretability and can be directly used in the next step of "semantic scheduling optimization".

[0105] In order to prevent the error values ​​of different dimensions from affecting the system judgment, it is also necessary to perform perceptual normalization on the error, and finally obtain the global weighted error metric:

[0106] ;

[0107] in, Is the semantic dimension Importance weight (derived from task priority or user preference), is the regularization coefficient, balancing the relationship between error and perceived risk. This will serve as the basis for deciding whether to initiate parameter adjustments and which dimensions to prioritize in subsequent steps.

[0108] In actual system operation, the semantic mapper A small number of user samples can be used for pre-training, and further optimized through online feedback during long-term operation. The system also supports a local fine-tuning mechanism, that is, when the error of a certain semantic dimension deviates for a long time, the mapping edge weight in that direction can be automatically updated locally.

[0109] Step 5: Optimization strategy generation based on error feedback

[0110] The goal of this step is to calculate the semantic error vector output by step 4 Establish an optimization strategy model to guide light and shadow control parameters Adjustment direction and amplitude.

[0111] The input is the semantic error vector generated in step 4 , each Representation system in semantic dimension The auxiliary input is the current light and shadow control parameters generated in step 2 The output is the update amount of the control parameters .

[0112] In this scheme, there is no one-to-one correspondence between semantic errors and physical parameters. The deviation of a semantic dimension may be caused by multiple parameters. Therefore, this step introduces a learnable structured mapping function , build the following optimization strategy model:

[0113] ;

[0114] The model constructs an interpretable gradient graph structure network to convert the error dimension Gradient adjustment amount mapped to control parameters Unlike ordinary backpropagation, this network incorporates a semantic-physical causal matrix into its structure. , making its learned structure more physically relevant.

[0115] In order to avoid conflicting adjustments in multiple semantic directions of the optimization model, an orthogonality regularization term is designed in this step. , suppressing the problem of "directional cancellation" caused by error signals from different semantic directions adjusting the same physical parameter. The regularization term is defined as follows:

[0116] ;

[0117] in:

[0118] Semantic error The propagated parameter gradient vector;

[0119] yes right The impact factor of structure is given);

[0120] The first term represents the total impulse after all error directions are combined, and the second term is the independent cumulative amount of each error direction;

[0121] The larger the difference, the more serious the conflict, and the stronger the penalty during model training, which prompts the optimization strategy to tend towards semantic separation.

[0122] The optimization model This can be achieved using the Graph Attention Network (GAT) structure, which continuously learns the error and parameter adjustment mode through training samples (user evaluation deviation + adjusted convergence effect). The reasoning process is fully automatic and is based only on the error You can output , with high efficiency and real-time performance. It will serve as the input of the parameter adjustment module in the subsequent system. How to perform the adjustment, whether to perform it, and the adjustment range are determined by the system controller or scheduler.

[0123] Step 6: Optimize convergence judgment and result output

[0124] The task of this step is to model the convergence judgment function , used to determine whether the system has achieved the user's light and shadow semantic goals in the current round and whether further parameter optimization and simulation iterations should be terminated. This judgment process does not directly control parameter updates, simulations, or outputs, but only provides a Boolean judgment value at the system level. , as a decision signal for whether to continue the iteration.

[0125] There are three types of input:

[0126] Semantic error vector of the current round (from step 4);

[0127] recent Wheel error history sequence ,in represents the weighted semantic error;

[0128] The control parameter adjustment amount output in step 5 above , used to determine whether the current strategy still has effective adjustment capabilities.

[0129] This scheme proposes a joint convergence criterion function that integrates "error margin", "error trend stability" and "strategy residual capacity" The judgment mechanism is as follows:

[0130] If the current weighted semantic error Less than the global threshold , then it is considered that the user goal has been achieved;

[0131] If the past The wheel error changes tend to be stable, that is , indicating that the optimization falls into a stable fluctuation area;

[0132] If the parameter update amplitude has been continuously reduced, it means that the strategy is approaching "ineffective adjustment" and can also converge.

[0133] The above criteria can be expressed uniformly as follows:

[0134] ;

[0135] in:

[0136] Recently Mean wheel error;

[0137] The variance that indicates the error trend;

[0138] is the average parameter adjustment amplitude of the current round;

[0139] : Error convergence threshold;

[0140] : Trend stability threshold;

[0141] : minimum valid update amplitude threshold;

[0142] : Perceptual sensitivity weight from step 4.

[0143] This joint criterion ensures that the system can automatically stop after the user's intention is achieved ( Small enough), it can also actively converge when the optimization is difficult to advance ( close to zero), while avoiding falling into low-frequency oscillation (through error fluctuation variance Suppression). This allows for a balance between convergence, security, and iteration efficiency. Judgment, this step also constructs an optional learning-based convergence judgment model , used for intelligent judgment in more complex situations. It is based on historical error sequence input and is trained with the label of whether the user terminates the operation as a supervisory signal. The formula is as follows:

[0144] ;

[0145] The model adopts a lightweight recurrent neural network (such as GRU) structure, which can learn the error evolution pattern under different semantic targets, thereby improving the flexibility and adaptability of judgment.

[0146] The purpose of this step is to determine whether the system should terminate. Once the convergence sign is confirmed, , the system controller decides whether to call modules such as result encapsulation and interactive confirmation.

[0147] The technical solution in the above-mentioned embodiment of the present application has at least the following technical effects or advantages: This solution constructs a linkage mechanism from semantic understanding, parameter reasoning to feedback correction, so that users can express their light and shadow requirements only through language or images without professional knowledge, and the system can automatically convert them into high-quality, personalized light and shadow design results.

[0148] The embodiment of the present application also provides a light and shadow solution optimization system based on digital twin and artificial intelligence, such as Figure 2 Shown, including:

[0149] An input processing module is configured to convert input lighting requirements into structured semantic vectors, wherein the input lighting requirements include natural language input and / or image input; and obtain a user intention vector by weighted averaging the semantic vectors generated by different types of input;

[0150] The parameter mapping module is configured to convert the user semantic vector into specific light and shadow control parameters;

[0151] The simulation calculation module is configured to simulate and calculate the light and shadow control parameters, output a simulation image and light and shadow physical indicators, extract key physical quantities under the light and shadow state from the simulation image, and integrate them to form simulation feedback indicators;

[0152] The error assessment module is configured to align the user intent vector with the simulation feedback indicator. By constructing a cross-modal mapping relationship, the simulation feedback indicator is mapped to a simulation semantic vector. The difference between the simulation semantic vector and the user intent vector is calculated to obtain a semantic error vector. The semantic error vector is perceptually normalized to obtain a global weighted error.

[0153] The strategy optimization module is configured to input the semantic error vector and the light and shadow control parameters and output the updated amount of the light and shadow control parameters;

[0154] The optimization control and output module is set to make judgments based on the convergence judgment function, the input semantic error vector, the global weighted error and the adjustment amount of the light and shadow control parameters. When the preset scenario is obtained, the operation is terminated and the final result is output.

[0155] The working principle is as follows: starting from the user's subjective expression, structured light and shadow parameters are automatically generated, simulation verification is carried out with the help of a digital twin environment, and the internal model of the system is reversely corrected through the simulation results, and the output plan is gradually optimized to achieve a closed-loop adaptive optimization process.

[0156] The above description is only a preferred specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any technician familiar with the technical field, within the technical scope disclosed by the present invention, who makes equivalent replacements or changes based on the technical solution and inventive concept of the present invention, should be covered by the scope of protection of the present invention.

Claims

1. A light and shadow scheme optimization method based on digital twins and artificial intelligence, characterized in that: include: Input user needs, convert them into semantic vectors, and obtain the user intention vector by weighted averaging the semantic vectors generated by different categories of input; Mapping the user intention vector in a graph neural network to obtain light and shadow control parameters, wherein the mapping is performed by using the semantic features in the user intention vector as nodes in the graph neural network and the light and shadow control parameters as mapping targets of the nodes, and adjusting the edge weights of the nodes in the graph to ensure the accuracy of the mapping; Perform simulation calculations on light and shadow control parameters, output simulation images and light and shadow physical indicators, extract key physical quantities under the light and shadow state from the simulation images, and integrate them to form simulation feedback indicators, wherein the simulation calculations are based on performing lighting simulation in a preset digital twin 3D space, and the simulation images are the rendering results of the scene under given parameters; The user intention vector is aligned with the simulation feedback indicator. By constructing a cross-modal mapping relationship, the simulation feedback indicator is mapped to obtain a simulation semantic vector. The difference between the simulation semantic vector and the user intention vector is calculated to obtain a semantic error vector. The semantic error vector is perceptually normalized to obtain a global weighted error. In the cross-modal mapping relationship, nodes represent semantic labels, and edges represent dependencies between semantic labels and physical indicators. An optimization strategy model is established, which inputs a semantic error vector and a light and shadow control parameter and outputs an update of the light and shadow control parameter. The optimization strategy model uses a graph attention network structure and is configured to map the semantic error vector to a gradient adjustment of the light and shadow control parameter. An orthogonality regularization term is introduced into the optimization strategy model, which uses the difference between the combined gradient and the independent gradient as a regularization term to penalize non-orthogonal gradient directions and suppress the directional offset of the adjustment of the same physical parameter caused by error signals from different semantic directions. Establish a convergence data function, input the semantic error vector, the global weighted error and the adjustment amount of the light and shadow control parameters for judgment, and terminate the operation and output the final result when the preset scenario is obtained; the preset scenario includes: If the current weighted semantic error is less than the global threshold, the user's goal is considered to have been achieved and the process stops automatically. The error changes in the past w rounds tend to be stable, that is, the error change trend is less than the trend stability threshold, indicating that the optimization falls into the stable fluctuation area and actively converges; The parameter update amplitude continues to shrink, indicating that the strategy is approaching ineffective adjustment and active convergence.

2. The light and shadow scheme optimization method based on digital twin and artificial intelligence according to claim 1 is characterized in that: The user demand conversion includes natural language processing and image input processing. The natural language processing converts user demand into vector representation through a pre-trained language model; the image input processing extracts the visual features of the image through a convolutional neural network to obtain a vector representation.

3. The light and shadow scheme optimization method based on digital twin and artificial intelligence according to claim 2 is characterized in that: The features extracted in the natural language processing include color temperature, brightness, light source and time.

4. The light and shadow scheme optimization method based on digital twin and artificial intelligence according to claim 2 is characterized in that: The features extracted from the image input processing include light source position, color temperature and brightness distribution.

5. The light and shadow scheme optimization method based on digital twin and artificial intelligence according to claim 1 is characterized in that: The key physical quantities extracted from the simulation image include global average illumination, illumination uniformity, maximum brightness and regional contrast, main light spot area and distribution range, and shadow boundary clarity.

6. A light and shadow scheme optimization system based on digital twins and artificial intelligence, characterized by: include: An input processing module is configured to convert input lighting requirements into structured semantic vectors, wherein the input lighting requirements include natural language input and / or image input; and obtain a user intention vector by weighted averaging the semantic vectors generated by different types of input; The parameter mapping module is configured to convert the user semantic vector into specific light and shadow control parameters; The simulation calculation module is configured to simulate and calculate the light and shadow control parameters, output a simulation image and light and shadow physical indicators, extract key physical quantities under the light and shadow state from the simulation image, and integrate them to form simulation feedback indicators; The error assessment module is configured to align the user intent vector with the simulation feedback indicator. By constructing a cross-modal mapping relationship, the simulation feedback indicator is mapped to a simulation semantic vector. The difference between the simulation semantic vector and the user intent vector is calculated to obtain a semantic error vector. The semantic error vector is perceptually normalized to obtain a global weighted error. A strategy optimization module is configured to establish an optimization strategy model, input a semantic error vector and a light and shadow control parameter, and output an update of the light and shadow control parameter. The optimization strategy model adopts a graph attention network structure and is configured to map the semantic error vector to a gradient adjustment of the light and shadow control parameter. An orthogonality regularization term is introduced into the optimization strategy model, and the difference between the combined gradient and the independent gradient is used as a regularization term to penalize non-orthogonal gradient directions, thereby suppressing the directional offset of the adjustment of the same physical parameter caused by error signals from different semantic directions. The optimization control and output module is configured to make judgments based on the convergence judgment function, the input semantic error vector, the global weighted error, and the adjustment amount of the light and shadow control parameters. When the preset scenario is obtained, the operation is terminated and the final result is output; The preset scenarios include: If the current weighted semantic error is less than the global threshold, the user's goal is considered to have been achieved and the process stops automatically. The error changes in the past w rounds tend to be stable, that is, the error change trend is less than the trend stability threshold, indicating that the optimization falls into the stable fluctuation area and actively converges; The parameter update amplitude continues to shrink, indicating that the strategy is approaching ineffective adjustment and active convergence.

Images