A park digital twin modeling method based on generative AI technology

Abstract

The application discloses a park digital twin modeling method based on generative AI technology and relates to the technical field of digital twin modeling, which comprises the following steps: preprocessing is completed by aligning point clouds, images and state data, topological adaptive filtering and multi-resolution voxelization, frequency domain harmonic fusion, super-surface texture excitation and time delay coupling fuzzy clustering, a holographic entropy film driven cross-modal spine network is constructed, and hierarchical tokens are generated through a graph attention generation layer, a 3D segment is obtained by inputting a surge tuning diffusion transformer model for iterative denoising and focus decoding; the segments are fused in the voxel space through cross-attention kernels, and the parameters are adaptively updated through an entropy pulse closed loop until the error converges, thereby generating a high-precision digital twin model. The spectrum mapping and holographic entropy film are used to construct a cross-modal semantic network, and the 3D segments are efficiently generated through focus fusion in the adaptive surge tuning diffusion transformer framework, so that the cross-modal semantic coupling efficiency, the generative reasoning convergence speed and the multi-segment voxel fusion coherence are improved.

Description

Technical Field

[0001] This invention relates to the field of digital twin modeling technology, specifically to a method for digital twin modeling of a campus based on generative AI technology. Background Technology

[0002] With the deepening development of smart park construction, digital twin technology is becoming a core support for achieving deep integration of physical and virtual spaces. However, existing digital twin solutions for parks typically rely on linear interpolation or simple timestamp mapping to align LiDAR point clouds with high-definition images, making it difficult to accommodate data streams with different frequencies and sampling rates. This results in inaccurate spatiotemporal registration, which in turn affects the accuracy and completeness of subsequent scene reconstruction.

[0003] In point cloud preprocessing, most methods only use fixed window filtering or global voxel lattice downsampling, which cannot preserve the topological information of complex site structures, nor can they take into account multi-resolution details. Image enhancement is often limited to single spatial domain or frequency domain filtering, which cannot simultaneously take into account low-frequency smoothing and high-frequency details. The processing of equipment status data mostly relies on statistical thresholds or empirical rules, lacking a fine characterization of the multi-channel time delay coupling relationship.

[0004] Insufficient cross-modal semantic fusion capability is also a significant bottleneck restricting fine-grained twins. Existing methods often process geometry, texture, and runtime state in stages or in parallel, lacking efficient coupling mechanisms and making it difficult to maintain semantic consistency of multi-source elements within the same network structure. Even when using graph neural networks, most works only stay at single-head convolution or basic attention frameworks, failing to deeply explore the adaptive correlation and residual fusion characteristics between nodes, thus making it difficult to construct efficient propagation models with spine-like topological hierarchies.

[0005] In generative AI modeling, current technologies are mainly based on single diffusion models or pure Transformer structures, lacking dynamic noise tuning and convergence determination mechanisms. This often sacrifices iterative efficiency while ensuring generation diversity. The decoding stage also lacks fusion strategies targeting local semantic focal points, making it difficult to achieve a balance between detail and overall coherence in generated fragments. The fusion of 3D scene fragments usually relies on simple weight stacking or averaging strategies, which cannot take into account the seamless connection of multi-scale elements, and there is no adaptive entropy-driven closed-loop optimization method for real-time self-correction. Summary of the Invention

[0006] In view of the shortcomings of the prior art described above, the purpose of this invention is to provide a method for digital twin modeling of a campus based on generative AI technology to solve the above-mentioned technical problems.

[0007] To achieve the above objectives, the present invention provides the following technical solution: a method for digital twin modeling of a campus based on generative AI technology, comprising:

[0008] S1: Perform spatiotemporal mapping and unified sampling on LiDAR point cloud, panoramic image and device status data to generate a synchronous frame set;

[0009] S2: Apply topological adaptive filtering and multi-resolution resampling to the synchronized point cloud to obtain a voxel set; perform frequency domain harmonic fusion and texture enhancement on the synchronized image to obtain an enhanced image; construct a time delay coupling matrix from the device status data and perform fuzzy clustering to obtain the state membership degree.

[0010] S3: Based on voxel sets, enhanced images, and state membership, a cross-modal semantic spine mapping network is constructed using a holographic entropy membrane;

[0011] S4: Using the semantic spine mapping network as input, a graph attention network is used to perform multiple rounds of information propagation and then linear mapping and normalization to generate hierarchical semantic tokens;

[0012] S5: Input the semantic token into the diffusion transformer hybrid network, and use the adaptive surge tuner to complete the latent variable iterative denoising and decoding to obtain a set of 3D scene fragments;

[0013] S6: Weighted fusion of 3D scene fragments in voxel space based on cross-attention kernels;

[0014] S7: Based on the coupling deviation between the fusion result and the real-time sensing data, the digital twin model is obtained by updating the mapping and generating parameters through adaptive entropy vascularization until the error converges.

[0015] The present invention is further configured such that step S1 includes:

[0016] The raw timestamps of LiDAR point cloud, panoramic image and device status data are linearly normalized to a unified frame domain.

[0017] Construct point cloud image alignment mapping function and state alignment mapping function, and solve for the optimal parameters of the mapping function by minimizing the residual between spatial point cloud and projected image and the residual between state features and frame index;

[0018] According to the preset sampling interval, the mapping function is called in sequence to output the point cloud frame, image frame and status frame at the corresponding time, forming a candidate synchronization frame triplet;

[0019] Threshold comparison is performed on the residuals of point clouds and images on the projection plane in the candidate synchronization frame triplet. For residuals exceeding the threshold, the mapping parameters are corrected by interpolation to meet the residual requirements.

[0020] Output all corrected synchronization frame triples to generate a synchronization frame set.

[0021] The present invention is further configured such that step S2 includes:

[0022] For the synchronous point cloud, an adjacency weight matrix based on distance and normal similarity is constructed. Topological noise is eliminated with adaptive intensity using the graph Laplacian inverse filter operator. Multi-layer voxel grids are divided with a reference voxel scale and a geometric series multiplication factor. Representative points are selected in each layer of the grid according to the principle of local point density priority to form a multi-resolution voxel set.

[0023] The low-frequency and high-frequency components of the synchronous image are extracted sequentially in the frequency domain according to the multi-order harmonic window and superimposed according to the exponential fusion weight to generate full-spectrum features. The full-spectrum features are excited by the second-order Laplacian operator and multiplied by the hypersurface function of the gradient magnitude to obtain the enhanced image.

[0024] The device state sequence is mapped to the power of the delay difference of each channel to generate a delay coupling matrix. The Euclidean distance penalty factor is adjusted by the delay coupling matrix. The coupling penalty is introduced into the membership update by using the fuzzy clustering algorithm. The iteration continues until the membership degree and the cluster center converge to obtain the state membership degree.

[0025] The present invention is further configured such that step S3 includes:

[0026] Feature mapping is performed on the multi-resolution voxel set, corresponding enhanced image region, and state membership of each spatial unit to generate configurational spine spectrum, holographic texture spectrum, and pulse entropy state spectrum.

[0027] For any two spatial units, attenuation calculation is performed based on the higher-order norm distance of the configuration spine spectrum and the holographic texture spectrum. The sum of the pulse entropy state spectrum is set as the amplification factor after entropy increase modulation, and the holographic entropy membrane weight is obtained through exponential mapping.

[0028] The holographic entropy membrane weights are multiplied element-wise with the power-law decay factor based on the spatial distance between unit centers to form the cross-modal node connection weight matrix.

[0029] The edge weight matrix is ​​normalized, and the normalization result, along with the configuration spine spectrum, holographic texture spectrum, and pulse entropy state spectrum, is set as structured input to construct a cross-modal semantic spine mapping network.

[0030] The present invention is further configured to construct a sparse topological Laplacian operator for the multi-resolution voxel set of each spatial unit, and apply high-order filtering to different frequency bands in a cyclic manner using the Chebyshev polynomial spectral filtering basis, and generate a configurational spine spectrum through power regularization of spectral coefficients and skeleton weight allocation.

[0031] Complex rotatable wavelet decomposition is performed on the enhanced image region corresponding to each spatial unit. Rotation scale invariance mapping is achieved through Fourier-Mellin transform. Holographic texture spectrum is constructed through superelliptical phase reconstruction and amplitude saturation operation.

[0032] The fuzzy membership vector and time-delay coupling sequence of each spatial unit are input into the Tsallis non-additive entropy mapping, and combined with time-delay jitter exponential filtering and normalization mapping, the pulse entropy state spectrum is output.

[0033] The present invention is further configured such that step S4 includes:

[0034] For the mapping features of spatial units in the semantic spine mapping network, mapping vector groups are generated through query mapping, key mapping and value mapping; based on the high-order norm distance between mapping vectors and the learnable gating signal, adaptive correlation coefficients are calculated in multi-head adaptive correlation coding;

[0035] The head-level aggregation of corresponding value mapping vectors is driven by adaptive correlation coefficients, and separable convolution is applied to the aggregation results of all heads to form node embeddings after multiple rounds of information propagation.

[0036] In the output of each propagation layer, the embedding of the previous layer is fused by the residual ratio to achieve feature order superposition;

[0037] The final layer embedding obtained through multiple rounds of information propagation is linearly mapped and combined with channel recalibration and dynamic normalization of temperature coefficient to amplify the normalized embedding amplitude and obtain the hierarchical semantic token.

[0038] The present invention is further configured such that step S5 includes:

[0039] Latent variables are generated by initializing and mapping all levels of semantic tokens, and the tokens are concatenated and appended with scene drift vectors to form a conditional context.

[0040] In each iteration, the multimodal aligned loss gradient is calculated based on the self-attention mapping of latent variables and conditional context. Gradient descent is performed on the latent variables with an adaptive gradient step size. In addition, the surge tuner completes the injection of high-order noise based on the dynamic noise ratio generated by the current residual mode and tuning parameters.

[0041] After each iteration, the normalized power ratio of the differences in the latent variables before and after is evaluated, and the iteration is stopped when it is lower than the preset convergence threshold.

[0042] By setting convergence latent variables and semantic tokens at each level as inputs, the 3D geometry and texture information is recovered through decoding mapping. The key information of the tokens is added by the focus fusion operator, and a set of 3D scene fragments is output.

[0043] The present invention is further configured such that step S6 includes:

[0044] For a 3D scene fragment at the target voxel point, a query vector and a key vector are generated through query mapping and key mapping;

[0045] Distance decay is performed using an exponential function based on the difference in the higher-order norms of the query vector and the key vector, and unnormalized attention weights are generated by combining the global coupling coefficient.

[0046] Power regularization and exponential mapping are applied to the unnormalized weights to obtain normalized attention coefficients.

[0047] The absolute values ​​of the voxel attributes of the 3D scene fragments are taken and amplified by powers of the attention coefficient. Then, Minkowski pooling is performed on all fragment values. Weighted fusion is achieved through power accumulation and corresponding root operations, and the fused voxel field is output.

[0048] The present invention is further configured such that step S7 includes:

[0049] A global deviation scalar is constructed by combining the power of the fused voxel field with the real-time point cloud projection error and the image projection error, and the integral along the spatial boundaries of the park.

[0050] The adaptive entropy pulse strength is calculated by applying an adaptive hyperbolic mapping to the global deviation scalar and combining it with an entropy-increasing modulation term.

[0051] The exponential decay update rate of the mapping function and generator parameters is adjusted by the adaptive entropy pulse intensity, and the parameters are updated element by element by driving the bias gradient.

[0052] By comparing the power-normalized difference of the global bias in two consecutive rounds with the preset convergence threshold, the iteration stops after the convergence condition is met, and the final digital twin model is output.

[0053] The invention is further configured to include: exporting the digital twin voxel field and material data that have reached the convergence condition as rendering vertex and texture resources, and mapping them to the scene graph structure of the visualization interaction platform.

[0054] This invention provides a method for digital twin modeling of a park based on generative AI technology. It generates a synchronized frame set by performing spatiotemporal mapping and unified sampling on LiDAR point clouds, panoramic images, and equipment status data. Topological adaptive filtering and multi-resolution resampling are applied to the synchronized point clouds to obtain a voxel set. Frequency domain harmonic fusion and texture enhancement are performed on the synchronized images to obtain enhanced images. A time-delay coupling matrix is ​​constructed from the equipment status data and fuzzy clustering is performed to obtain state membership degrees. Based on the voxel set, enhanced images, and state membership degrees, a cross-modal semantic spine mapping network is constructed using a holographic entropy membrane. Using the semantic spine mapping network as input, a graph attention network is used for multi-round information propagation, followed by linear mapping and normalization to generate hierarchical semantic tokens. The semantic tokens are input into a diffusion transformer hybrid network, and an adaptive surge tuner is used to complete iterative denoising and decoding of latent variables to obtain a set of 3D scene fragments. The 3D scene fragments are weighted and fused in voxel space according to cross-attention kernels. Based on the coupling deviation between the fusion result and real-time perceived data, the mapping and generation parameters are updated using an adaptive entropy vasoconstrictor until the error converges, resulting in a digital twin model. The beneficial effects include:

[0055] 1. Dual enhancement of point cloud and image preprocessing: Topological adaptive filtering and multi-resolution resampling are used in the point cloud to remove noise and retain details, while frequency domain harmonic fusion and hypersurface texture excitation are combined in the image to highlight high-frequency features.

[0056] 2. Efficient construction of cross-modal semantic spine mapping network: Through innovative mapping of configurational spine spectrum, holographic texture spectrum and pulse entropy state spectrum and holographic entropy membrane weighting, a structured backbone network is generated to ensure efficient coupling of multi-source elements;

[0057] 3. Efficient convergence of hybrid generative AI inference and focus decoding: By utilizing an adaptive surge tuner, conditional self-attention mapping and high-order convergence determination, latent variables are quickly denoised and high-fidelity 3D fragments are recovered through focus fusion.

[0058] 4. Seamless fusion of multiple segments: Based on the complex fusion logic of cross-attention kernel and Minkowski pooling, seamless splicing of 3D segments in voxel space is achieved, enhancing the overall coherence and realism of the model.

[0059] The above description is only an overview of the technical solution of this application. In order to better understand the technical means of this application and to implement it in accordance with the contents of the specification, and to make the above and other objects, features and advantages of this application more obvious and understandable, the following are specific embodiments of this application. Attached Figure Description

[0060] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. In the drawings:

[0061] Figure 1 The flowchart illustrates an exemplary embodiment of the present invention of a digital twin modeling method for a campus based on generative AI technology. Detailed Implementation

[0062] The embodiments of the present invention will be described below with reference to the accompanying drawings and preferred embodiments. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and various details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be understood that the preferred embodiments are only for illustrating the present invention and not for limiting the scope of protection of the present invention.

[0063] It should be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of the present invention. Therefore, the drawings only show the components related to the present invention and are not drawn according to the actual number, shape and size of the components in the actual implementation. In the actual implementation, the form, quantity and proportion of each component can be arbitrarily changed, and the layout of the components may also be more complex.

[0064] In the following description, numerous details are explored to provide a more thorough explanation of embodiments of the invention. However, it will be apparent to those skilled in the art that embodiments of the invention may be practiced without these specific details. In other embodiments, well-known structures and devices are shown in block diagram form rather than in detail to avoid obscuring embodiments of the invention.

[0065] A method for digital twin modeling of industrial parks based on generative AI technology, such as Figure 1 As shown, it includes:

[0066] S1: Perform spatiotemporal mapping and unified sampling on LiDAR point cloud, panoramic image and device status data to generate a synchronous frame set;

[0067] S2: Apply topological adaptive filtering and multi-resolution resampling to the synchronized point cloud to obtain a voxel set; perform frequency domain harmonic fusion and texture enhancement on the synchronized image to obtain an enhanced image; construct a time delay coupling matrix from the device status data and perform fuzzy clustering to obtain the state membership degree.

[0068] S3: Based on voxel sets, enhanced images, and state membership, a cross-modal semantic spine mapping network is constructed using a holographic entropy membrane;

[0069] S4: Using the semantic spine mapping network as input, a graph attention network is used to perform multiple rounds of information propagation and then linear mapping and normalization to generate hierarchical semantic tokens;

[0070] S5: Input the semantic token into the diffusion transformer hybrid network, and use the adaptive surge tuner to complete the latent variable iterative denoising and decoding to obtain a set of 3D scene fragments;

[0071] S6: Weighted fusion of 3D scene fragments in voxel space based on cross-attention kernels;

[0072] S7: Based on the coupling deviation between the fusion result and the real-time sensing data, the digital twin model is obtained by updating the mapping and generating parameters through adaptive entropy vascularization until the error converges.

[0073] The present invention is further configured such that step S1 includes:

[0074] The raw timestamps of LiDAR point clouds, panoramic images, and device status data are linearly normalized to a unified frame domain; specifically, the LiDAR point clouds... ,image and device status Original timestamp Power-law mapping is applied to compress and stretch the respective time domains to a unified frame domain [0, N−1] to address the differences in start and end times and sampling rates of heterogeneous data. , , For data dimensions, including LiDAR point clouds ,image and device status ;parameter By imbuing the normalization process with nonlinear controllability, we obtain the output three-way normalized time parameter sequence { , , }, The value range is [1.2, 3];

[0075] A point cloud image alignment mapping function and a state alignment mapping function are constructed. The optimal parameters of the mapping functions are solved by minimizing the residuals between the spatial point cloud and the projected image, as well as the residuals between state features and the frame index. Specifically, two mapping functions are constructed: one maps the point cloud temporal domain to the image temporal domain by minimizing the residuals from the projection of 3D points onto a 2D plane; the other maps the device state sequence to the frame index by minimizing the difference between the state features and the frame index. Both residuals participate in the optimization to ensure geometry and state synchronization. ,in, This is a LiDAR point cloud temporal mapping function used to map normalized frame parameters. The temporal index mapped to the original point cloud sequence; This is a panoramic image time mapping function used to map normalized frame parameters. Time index mapped to the original image sequence; This is a device state time mapping function used to map normalized frame parameters. Time index mapped to the original state sequence; The optimal solution set for the above three mapping functions is obtained by minimizing the overall residual. This is a function for extracting LiDAR point cloud data. It takes a time index as input and returns the corresponding point cloud frame. This is a 3D point cloud to 2D image plane projection operator. This is an image data extraction function that takes a time index as input and returns the corresponding image frame. This is the power coefficient of the spatial residual, used to strengthen the penalty for the total loss of samples with high residuals; This is the state residual weighting balancing coefficient, used to adjust the relative importance of spatial residuals and state residuals in the overall loss. Its value ranges from [0.1, 1]. This is a function for extracting device status data. It takes a time index as input and returns the corresponding status frame. The reference frame index reconstruction function for the state sequence is obtained through interpolation or fitting, and the target is aligned with the actual state frame. This is the power coefficient of the state residual, used to enhance the impact of state alignment error on the total loss. and The value range is [1.5, 2.5]. The normalized frame parameters, ranging from [0, N−1], serve as the independent variables for unified sampling and mapping; The upper limit of the normalized frame domain is determined by the total number of target frames;

[0076] According to the preset sampling interval, the mapping function is called sequentially to output the point cloud frame, image frame, and status frame at the corresponding time, forming a candidate synchronization frame triplet; specifically, it is indexed by frame. The mapping function is called sequentially to extract data at the corresponding time points, thus initially forming a set of multi-source candidate synchronization frame triples: ;

[0077] Threshold comparison is performed on the residuals of point cloud and image projection in candidate synchronization frame triples on the projection plane. For residuals exceeding the threshold, interpolation is used to correct the mapping parameters until the residual requirements are met. Specifically, the power value of the residual between point cloud and image projection is calculated. If it exceeds the threshold Then, the mapping function is fine-tuned using linear interpolation between adjacent frame indices to eliminate alignment anomalies: ,like Then interpolation correction ;

[0078] Output all corrected synchronization frame triples to generate a synchronization frame set; specifically, summarize all corrected frame triples and output the final synchronization frame set. .

[0079] The present invention is further configured such that step S2 includes:

[0080] For the synchronized point cloud, an adjacency weight matrix based on distance and normal similarity is constructed. Topological noise is adaptively removed using the graph Laplacian inverse filter operator. A multi-layer voxel grid is created using a baseline voxel scale and a geometric series multiplication factor. Within each grid layer, representative points are selected based on a local point density priority principle to form a multi-resolution voxel set. Specifically, for each pair of point cloud samples... , Calculation based on distance power decay and the angle between the normals Adjacency weight of cosine : , The value range of is [2, 4]. The adjacency weight decay scale is set according to the point cloud range, and then the graph Laplacian matrix is ​​constructed. and with adaptive strength Perform inverse filtering to remove topology noise: , It is the identity matrix. An adaptive intensity function for extracting LiDAR point cloud data. The value range is [0.1, 1]; then, based on the benchmark scale... and multiplication factor Generate multi-voxel scale: Multiplication factor The value range is [1.2, 2]. The quantization function is used to select the densest representative points in each voxel grid layer, and the multi-resolution voxel set is output. : , For quantization functions;

[0081] In the frequency domain, low-frequency and high-frequency components are sequentially extracted from synchronized images using multi-harmonic windows and then superimposed with exponential fusion weights to generate full-spectrum features. These full-spectrum features are then excited by a second-order Laplacian operator and multiplied by a hypersurface function with gradient magnitude to obtain the enhanced image. Specifically, the synchronized images are first processed... Perform Fourier transform According to the multi-order Bessel harmonic window With power weight Extract and reconstruct full-spectrum features: , The harmonic window order ranges from [3, 6], and the power weight is... The values ​​range from [1,3], and then an inverse transform is performed to generate a fused image. : Then, the second-order Laplace operator is applied and multiplied by the gradient magnitude. Power and Enhancement Factor Hypersurface excitation yields enhanced images with prominent high-frequency details. : Among them, power weight The value range is [1,3], and the enhancement coefficient is... The value range is [0.5, 2]. The texture gradient excitation power has a value range of [1,3].

[0082] A time delay coupling matrix is ​​generated by mapping the time delay difference of each channel to a power value for the device state sequence. The Euclidean distance penalty factor is adjusted using the time delay coupling matrix. A fuzzy clustering algorithm is used to introduce a coupling penalty in the membership update. The iteration continues until the membership degree and the cluster center converge, thus obtaining the state membership degree. Specifically, the time delay power value is calculated for any two channel sampling points. Mapping difference : Time delay power The value range is [1,3], and is expressed as a power of the difference in state values. With decay scale Constructing the coupling matrix Then apply the coupling penalty. Introducing the state-cluster center distance metric (Power-order p>1), and update the membership degree using a fuzzy C-mean with coupling penalty based on the power-order membership degree exponent m>1. Continue until convergence, then output the state membership matrix. : Time delay power Power of the difference between state values The value range is [1,3], and the attenuation scale is... Based on the state value range setting, p is the distance metric power, with a value range of [2,4], and the power membership index m has a value range of [1.5,3].

[0083] Adaptive topological filtering based on distance and normal similarity effectively eliminated noise while preserving the skeleton structure of ground features. Multi-layer voxel meshes precisely balanced overall and local details. Image frequency domain multi-harmonic fusion and hypersurface texture excitation effectively enhanced full-spectrum details. Fuzzy clustering with time-delay coupling and coupling penalties accurately characterized the correlation of device states. These combinations significantly improved the quality and multimodal coupling capability of the three-source data, laying a solid foundation for subsequent high-precision digital twin modeling.

[0084] The present invention is further configured such that step S3 includes:

[0085] Feature mapping is performed on the multi-resolution voxel set, corresponding enhanced image region, and state membership of each spatial unit to generate the configurational spine spectrum, holographic texture spectrum, and pulse entropy state spectrum. The invention further involves constructing a sparse topological Laplacian operator for the multi-resolution voxel set of each spatial unit, and cyclically applying high-order filtering to different frequency bands using a Chebyshev polynomial spectral filtering basis. The configurational spine spectrum is generated through power-law regularization of the spectral coefficients and skeleton weight allocation. Specifically, for the ... Constructing a sparse adjacency matrix from multi-resolution voxel sets of spatial units (Based on distance and normal similarity), adjacency weight The logic is derived from the above logical calculations, which will not be elaborated here, forming the graph Laplace operator. : , Let be a degree matrix, with diagonal elements as follows: , For the first Within the space unit Adjacency weights between points; spectral basis recursively generated using Chebyshev polynomials. : For node feature vectors (Voxel represents point coordinates and normal) Apply higher-order filtering; multiply the output of each order filter by the skeleton weights. (follow) (decreasing), accumulated to form a configurational spine spectrum : Emphasizing the core structure frequency bands and the weighting of the skeleton. Following a decreasing sequence, ;

[0086] Complex rotatable wavelet decomposition is performed on the enhanced image region corresponding to each spatial unit. Rotation scale invariance mapping is achieved through Fourier-Mellin transform, and holographic texture spectrum is constructed through superelliptical phase reconstruction and amplitude saturation operation; specifically, for the ... Unit-enhanced image region Perform complex rotatable wavelet decomposition (M-scale, N-direction) Fourier-Mellin transform (FM) is applied to extract rotation- and scale-invariant amplitude spectra, and each amplitude spectrum is then sorted by power. ∈[1,3] superposition; then reconstruct the wavelet phase using a superelliptical phase function. (Spherical Ellipse) performs saturation mapping to generate a holographic texture spectrum. : , For the first Level Scale A directional complex wavelet, 1≤M≤5, 4≤N≤16;

[0087] The fuzzy membership vector and time-delay coupling sequence of each spatial unit are input into the Tsallis non-additive entropy mapping, combined with time-delay jitter power filtering and normalization mapping, to output the pulse entropy state spectrum; specifically, for the ... Meta-membership vector Input Tsallis nonadditive entropy mapping (entropy order) ∈(0,2]), quantize the nonlinear entropy of multi-channel clustering; then, the time delay coupling difference between each center is used. (power) Multiplying the values ​​of ∈[1,3] and summing them with the entropy value yields the pulse entropy state spectrum that balances non-additive entropy and time delay coupling. : , For the first Sample pair Fuzzy membership degree of cluster centers;

[0088] For any two spatial units, attenuation calculation is performed based on the higher-order norm distance of the configuration spine spectrum and the holographic texture spectrum. The sum of the pulse entropy state spectrum is set as the amplification factor after entropy increase modulation, and the holographic entropy membrane weight is obtained through exponential mapping.

[0089] The holographic entropy membrane weights are element-wise multiplied with a power-law decay factor based on the spatial distance between unit centers to form a cross-modal node connection weight matrix; specifically, for any unit pair (i,j), the weights are calculated by scaling the power-law distance between the difference in the configuration spine spectrum and the difference in the texture spectrum. The attenuated states are superimposed and multiplied by a power of the sum of the pulse entropy spectrum. Generate holographic entropy membrane weights: , The power of spectral difference attenuation, with values ​​ranging from [2,4], represents the power distance across the scale. Set according to the range of spectral difference. The value range is [1,3]; and the distance from the center of space. The power-law decay factor is combined to obtain the edge weight matrix. : , This is the spatial decay power, with a value range of [1, 3]. The spatial attenuation scale is set according to the park's boundaries;

[0090] The edge weight matrix is ​​normalized by degree, and the normalized result, along with the configuration spine spectrum, holographic texture spectrum, and pulse entropy state spectrum, is set as structured input to construct a cross-modal semantic spine mapping network. Specifically, longitude normalization forms a normalized matrix. : , The degree matrix, along with node features Together they constitute the input of the semantic spine mapping network.

[0091] The present invention is further configured such that step S4 includes:

[0092] For the mapping features of spatial units in the semantic spine mapping network, mapping vector groups are generated through query mapping, key mapping, and value mapping. Based on the high-order norm distance between mapping vectors and the learnable gating signal, adaptive association coefficients are calculated within the multi-head adaptive association coding. Specifically, the original node features are mapped into query, key, and value vectors to provide a basis for attention calculation. The original node features, including configurable spine spectrum, holographic texture spectrum, and pulse entropy spectrum, are mapped to three different vector spaces to measure the similarity between nodes and obtain message content.

[0093] An adaptive correlation coefficient drives the intra-head aggregation of corresponding value mapping vectors, and a separable convolutional confluence is applied to the aggregation results of all heads to form a node embedding after multiple rounds of information propagation. Specifically, each attention head automatically generates a correlation coefficient based on the higher-order norm distance between the query and the key and a learnable gating signal, which is used to aggregate the value vector within the head. Each "head" adaptively calculates attention weights through the following mechanism: calculating the attention weights for the query vector and each key vector. norm ( >2), enhance the difference measurement between highly heterogeneous nodes, which is generated by the concatenation result of query and key by a small neural network, and multiplied by the distance metric after Sigmoid activation to achieve nonlinear gated modulation; each head uses the obtained attention coefficient to weight its corresponding value vector set to obtain the intra-head aggregate vector;

[0094] In the output of each propagation layer, the embedding of the previous layer is fused by the residual ratio to achieve the superposition of feature order. Specifically, the aggregation results of each head are fused by separable convolution and the embedding of the previous layer is merged by the residual ratio to achieve the superposition of information order in multiple rounds. The aggregation vectors of all heads are first concatenated in the channel dimension and then fused channel by channel by separable convolution, while maintaining lightweight and efficiency. The embedding of the previous layer is introduced and fused into the current output by a learnable ratio to retain low-order features and accelerate gradient propagation, thereby achieving the continuous superposition of feature order.

[0095] The final layer embedding obtained through multiple rounds of information propagation undergoes linear mapping and dynamic normalization processing using channel recalibration and temperature coefficients to amplify the normalized embedding amplitude and obtain the hierarchical semantic token. Specifically, after linear transformation, the final layer embedding is adaptively adjusted for its numerical range using channel recalibration and dynamic normalization of the temperature coefficients, and the normalized result is amplified to output the hierarchical semantic token. The final layer embedding is multiplied by the mapping matrix and biased for dimensionality reduction / increase and preparation for normalization. Learnable scaling and bias factors are used to recalibrate the activations of each channel, highlighting important channel information. The recalibrated vector is divided by a learnable temperature coefficient to dynamically control the normalization strength, balancing stability and sensitivity. Finally, it is multiplied by a token enhancement coefficient. This ensures that the final token has sufficient expressive power in subsequent generative AI networks.

[0096] The present invention is further configured such that step S5 includes:

[0097] Latent variables are generated by initializing and mapping all hierarchical semantic tokens. The tokens are then concatenated and a scene drift vector is appended to form a conditional context. Specifically, each hierarchical semantic token is applied to a learnable initialization mapping, extending the low-dimensional token encoding to the latent variable space to form an initial noise state. All tokens are concatenated sequentially, and a scene drift vector is appended to the end. The scene drift vector is a learnable vector describing the global viewpoint or time drift effect, providing macroscopic background information for the generation process. Furthermore, the scene drift vector is a learnable global offset encoding used to simulate the overall migration effect of the viewpoint or time dimension, enhancing the spatiotemporal coherence of the generated scene. The scene drift vector first generates a spatiotemporal offset code from the time frame index and the spatial unit center coordinates to capture the global viewpoint or time migration trend. Then, a gating network is used to perform an overall summary of all hierarchical semantic tokens to extract semantic offset information. Finally, the spatiotemporal offset code and semantic offset are merged and mapped to the latent variable dimension to form a global drift vector that includes both spatiotemporal continuity and semantic conditions.

[0098] In each iteration, the multimodal alignment loss gradient is calculated based on the self-attention mapping of latent variables and conditional context. Gradient descent is performed on the latent variables with an adaptive gradient step size, and high-order noise injection is completed by combining the dynamic noise ratio generated by the surge tuner according to the current residual mode and tuning parameters. Specifically, in each iteration, the current latent variables and conditional context are used together as the query and key / value input. The various dimensions of the latent variables and token semantics are aligned through multi-head self-attention. The multimodal alignment loss gradient is calculated to reflect the inconsistency between the current denoising result and the token condition. Gradient descent is performed on the latent variables with an adaptive gradient step size (dependent on the historical gradient modes and convergence rate), and high-order noise injection is completed by combining the dynamic noise ratio generated by the surge tuner to maintain the diversity of generation and local details.

[0099] After each iteration, the normalized power ratio of the difference before and after the latent variable is evaluated. If it is lower than the preset convergence threshold, the iteration is stopped. Specifically, at the end of the iteration, the power ratio of the difference before and after the latent variable is calculated and normalized. If the ratio is lower than the preset convergence threshold, it indicates that the denoising process has fully aligned the conditional context with the generation details, and the iteration is terminated to avoid over-denoising or oscillation.

[0100] The convergent latent variables and semantic tokens at each level are set as inputs. 3D geometry and texture information are recovered through decoding mapping. A focus fusion operator is then used to add key token information, outputting a set of 3D scene fragments. Specifically, the final latent variables and all levels of semantic tokens are input into the decoding mapping network, and 3D geometry and texture data are recovered through multi-layer learnable mapping. A focus fusion operator is introduced during the decoding process. This operator dynamically enhances the decoding of specific regions based on token importance, ensuring that key semantic information is prominently expressed in the fragments. The output is a high-fidelity, structurally coherent set of 3D scene fragments that conforms to the conditions of the original semantic tokens. Furthermore, the focus... In the decoding stage, the fusion operator enhances the feature mapping of key sub-regions based on token activation, improving the consistency between local details and overall semantics. The focus fusion operator first scores the similarity between all semantic tokens and the current decoding query vector to determine the relative importance of each token in the current generation task. Then, based on the mapping relationship between tokens and spatial points, the scores are mapped to spatial focus masks to highlight the regions that should be reconstructed. In the decoding network, the original features are weighted and amplified with adjustable gain parameters to further enhance the information of high-scoring regions. Finally, the fusion result is injected into each decoding level or output stage to achieve fine-grained focusing and detail enhancement of key semantic segments.

[0101] The present invention is further configured such that step S6 includes:

[0102] For a 3D scene fragment at a target voxel point, query vectors and key vectors are generated through query mapping and key mapping. Specifically, at each target voxel point, two sets of linear mappings are performed on the local attributes of all scene fragments and their corresponding hierarchical semantic tokens. One set generates a query vector, representing the current reference requirement at that point; the other set generates a key vector for each fragment, representing the information characteristics that the fragment can provide. Local attributes include geometric coordinates, normal vectors, and texture values.

[0103] Based on the higher-order norm difference between the query vector and the key vector, an exponential function is used for distance decay, combined with a global coupling coefficient to generate unnormalized attention weights. Specifically, the higher-order norm difference between the key vector and the query vector of each segment is calculated, and the difference result is decayed using an exponential function, supplemented by a global coupling coefficient for multiplication, to obtain the initial attention weight of each segment for that voxel point. This highlights segments that are highly matched to the query, while smoothly suppressing irrelevant or noisy information based on the coupling coefficient. Furthermore, the global coupling coefficient is a learnable scaling parameter used for distance decay across all voxels and segments. To maintain consistency in information fusion, the entropy activity of the entire network is obtained by averaging the impulse entropy spectrum of all spatial units, which measures the degree of state coupling and nonlinear distribution. The maximum eigenvalue of the normalized Laplacian matrix of the semantic spine mapping network is calculated to characterize the global connectivity and information diffusion capability of the network. The network activity is nonlinearly amplified by a learnable power, and the connectivity index is multiplied after being mapped by an adjustable exponent to obtain the initial coupling strength. This strength is mapped by a nonlinear function and truncated to a preset upper limit to generate a global coupling coefficient for uniformly adjusting the attention weights among all nodes. The global coupling coefficient reflects both the overall uncertainty of the network at the state coupling level and the connectivity characteristics of the network topology, enabling attention calculation to adaptively amplify or converge according to global semantics and structure, thereby improving the efficiency and robustness of cross-modal fusion.

[0104] Power regularization and exponential mapping are applied to the unnormalized weights to obtain normalized attention coefficients. Specifically, the unnormalized weights are first power regularized (to amplify large weights and further reduce small weights), and then normalized by exponential mapping and sum division, so that the sum of the weights of all segments at the same voxel is one, ensuring that the fusion result is energy-conserving and interpretable.

[0105] The absolute values ​​of voxel attribute values ​​for 3D scene fragments are taken and amplified by a power of the attention coefficient. Then, Minkowski pooling is performed on all fragment values. Weighted fusion is achieved through power-law summation and corresponding root operations, outputting the fused voxel field. Specifically, the absolute value of the attribute value for each fragment at that voxel point is first taken (to eliminate the influence of sign), then amplified by a power of the normalized weights to reflect the importance of high-weight fragments. Subsequently, Minkowski pooling is performed on all amplified attribute values—that is, the weighted values ​​are summed by a power and the corresponding reciprocal root is taken to highlight peak contributions while preserving the distribution pattern. The final fused attribute for that voxel point is obtained after pooling.

[0106] The present invention is further configured such that step S7 includes:

[0107] A global deviation scalar is constructed by integrating the powers of the fused voxel field with the real-time point cloud projection error and the image projection error along the entire park space. Specifically, the global deviation is obtained by summing the powers of the fused voxel field with the real-time perceived point cloud projection error and the image projection error at each spatial point, and then performing a double integral along the entire park space. This scalar comprehensively reflects the degree of geometric and visual inconsistency between the digital twin model and the physical scene, providing a global metric for self-correction.

[0108] The adaptive entropy pulse strength is calculated by applying an adaptive hyperbolic mapping to the global deviation scalar and combining it with an entropy-increasing modulation term. Specifically, the global deviation scalar is input into a hyperbolic tangent mapping to compress the deviation value to a controllable range in a nonlinear manner. Then, a modulation term is generated according to a learnable entropy-increasing function—this term amplifies or suppresses the mapping based on the entropy-increasing trend of the deviation value, and is multiplied with the hyperbolic mapping result to form the adaptive entropy pulse strength. This strength responds to the global deviation while also taking into account the uncertainty of the deviation distribution, and is used to balance the update pace and stability.

[0109] The exponentially decaying update rate of the mapping function and generator parameters is adjusted by the adaptive entropy pulse intensity, and the parameters are updated element-by-element by driving the deviation gradient. Specifically, in each update of the mapping function and generator parameters, the adaptive entropy pulse intensity is introduced as an adjustment factor for the exponential decay rate. Specifically, the deviation gradient is calculated for the mapping and generator parameters based on the global deviation, and the update rate is controlled by combining exponential decay with the entropy pulse intensity, achieving fine-tuning of the parameters element-by-element, thus achieving efficient convergence between global and local conditions.

[0110] The iteration stops when the power-normalized difference between the global biases of two consecutive rounds is compared with a preset convergence threshold, and the final digital twin model is output after the convergence condition is met. Specifically, after each closed-loop update, the global bias is recalculated and normalized after taking a high power of the difference with the previous round's bias to obtain the update convergence index. Iteration stops only when this index is lower than the preset convergence threshold, and the final digital twin model is output, ensuring that the self-correction is both sufficient and not excessive.

[0111] The invention is further configured to include: exporting the converged digital twin voxel field and material data as rendering vertex and texture resources, and mapping them to the scene graph structure of the visualization interaction platform; specifically, after the digital twin model undergoes closed-loop self-correction and meets the convergence condition, the final three-dimensional voxel field and its associated material information need to be converted into platform resources that can directly drive real-time rendering and interaction, including voxel field meshing, material parameter texturing, and scene graph structure encapsulation; voxel field meshing: performing Marching Cubes or Dual on the converged voxel field. The Contouring algorithm generates continuous triangular mesh vertices and faces along a specified threshold; it applies edge folding and quadtree simplification strategies to overly dense or fine regions, ensuring that the mesh topology retains key details while meeting real-time rendering performance requirements; it automatically generates vertex normals and tangents based on the coordinate difference of adjacent vertices for subsequent lighting and PBR material rendering; material parameter texturing: it extracts material properties such as color, roughness, and metallicity from each unit of the voxel field; it uses an automated tiling algorithm to generate seamless UV coordinates for the mesh, including ABF++ or LSCM; it bakes material properties to one or more sets of textures (Diffuse, Normal, Metalness, Roughness) and optimizes texture resolution to balance quality and memory; scene graph structure encapsulation: in the visualization interaction platform, mesh and texture resources are encapsulated as scene nodes, each containing geometric mesh references, material texture references, and parent-child hierarchical relationships between nodes; it attaches a unique identifier and attribute dictionary to each node to support runtime interaction (selection, highlighting, attribute extraction); and it generates a resource list required by the platform. The Manifest lists all meshes, textures, and dependencies, facilitating asynchronous loading and version management.

[0112] Through the aforementioned resource export and encapsulation process, high-precision digital twin models can be seamlessly integrated into mainstream visualization platforms while maintaining rendering performance and detail fidelity. Automated meshing and simplification ensure high frame rates during interaction, PBR mapping and precise normals enhance visual realism, and the scene graph structure provides powerful attribute query and event response capabilities, offering users a smooth, intuitive, and scalable interactive experience.

[0113] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented, in whole or in part, as a computer program product. The computer program product includes one or more computer instructions or computer programs. When the computer instructions or computer programs are loaded or executed on a computer, all or part of the processes or functions described in the embodiments of this application are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that includes one or more sets of available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium. A semiconductor medium can be a solid-state drive.

[0114] It should be understood that the term "and / or" in this article is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, or B existing alone. A and B can be singular or plural. Additionally, the character " / " in this article generally indicates an "or" relationship between the preceding and following related objects, but it can also represent an "and / or" relationship. Please refer to the context for a more accurate understanding.

[0115] In this application, "at least one" means one or more, and "more than one" means two or more. "At least one of the following" or similar expressions refer to any combination of these items, including any combination of single or multiple items. For example, at least one of a, b, or c can mean: a, b, c, ab, ac, bc, or abc, where a, b, and c can be single or multiple.

[0116] It should be understood that in the various embodiments of this application, the order of the above-mentioned processes does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.

[0117] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0118] Those skilled in the art will understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0119] In the several embodiments provided in this application, it should be understood that the disclosed system can be implemented in other ways. For example, the device embodiments described above are merely illustrative. For instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between devices or units may be electrical, mechanical, or other forms.

[0120] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0121] In addition, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0122] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0123] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for digital twin modeling of a park based on generative AI technology, characterized in that, include: S1: Perform spatiotemporal mapping and unified sampling on LiDAR point cloud, panoramic image and device status data to generate a synchronous frame set; S2: Apply topological adaptive filtering and multi-resolution resampling to the synchronized point cloud to obtain a voxel set; perform frequency domain harmonic fusion and texture enhancement on the synchronized image to obtain an enhanced image; construct a time delay coupling matrix from the device status data and perform fuzzy clustering to obtain the state membership degree. S3: Perform feature mapping on the multi-resolution voxel set, corresponding enhanced image region and state membership of each spatial unit to generate configurational spine spectrum, holographic texture spectrum and pulse entropy state spectrum; For any two spatial units, attenuation calculations are performed based on the high-order norm distance between the configurational spine spectrum and the holographic texture spectrum. The sum of the pulse entropy state spectrum is set as an amplification factor after entropy increase modulation, and the holographic entropy membrane weight is obtained through exponential mapping. The holographic entropy membrane weight is then multiplied element-wise with the power-law attenuation factor based on the spatial distance between unit centers to form the cross-modal node connection weight matrix. The connection weight matrix is ​​then normalized in degree, and the normalized result, along with the configurational spine spectrum, the holographic texture spectrum, and the pulse entropy state spectrum, is set as structured input to construct a cross-modal semantic spine mapping network. S4: Using the semantic spine mapping network as input, a graph attention network is used to perform multiple rounds of information propagation and then linear mapping and normalization to generate hierarchical semantic tokens; S5: Generate latent variables by initializing and mapping all levels of semantic tokens, and then concatenate the tokens and attach the scene drift vector to form a conditional context. In each iteration, the multimodal alignment loss gradient is calculated based on the self-attention mapping of latent variables and conditional context. Gradient descent is performed on the latent variables with an adaptive gradient step size. High-order noise injection is completed by combining the surge tuner with the dynamic noise ratio generated according to the current residual mode and tuning parameters. After each iteration, the normalized power ratio of the difference between the latent variables before and after is evaluated. The iteration stops when it is lower than the preset convergence threshold. The convergent latent variables and semantic tokens at each level are set as inputs. The 3D geometry and texture information is recovered by decoding the mapping. The key information of the tokens is added by the focus fusion operator to output a set of 3D scene fragments. S6: Weighted fusion of 3D scene fragments in voxel space based on cross-attention kernels; S7: Construct a global deviation scalar by integrating the power of the fused voxel field with the real-time point cloud projection error and the image projection error along the park space; calculate the adaptive entropy pulse intensity by applying an adaptive hyperbolic mapping to the global deviation scalar and combining it with an entropy increase modulation term; The exponential decay update rate of the mapping function and generator parameters is adjusted by the adaptive entropy pulse intensity, and the parameters are updated element by element by driving the bias gradient. By comparing the power-normalized difference of the global bias in two consecutive rounds with the preset convergence threshold, the iteration stops after the convergence condition is met, and the final digital twin model is output.

2. The method for digital twin modeling of a park based on generative AI technology according to claim 1, characterized in that, Step S1 includes: The raw timestamps of LiDAR point cloud, panoramic image and device status data are linearly normalized to a unified frame domain. Construct point cloud image alignment mapping function and state alignment mapping function, and solve for the optimal parameters of the mapping function by minimizing the residual between spatial point cloud and projected image and the residual between state features and frame index; According to the preset sampling interval, the mapping function is called in sequence to output the point cloud frame, image frame and status frame at the corresponding time, forming a candidate synchronization frame triplet; Threshold comparison is performed on the residuals of point clouds and images on the projection plane in the candidate synchronization frame triplet. For residuals exceeding the threshold, the mapping parameters are corrected by interpolation to meet the residual requirements. Output all corrected synchronization frame triples to generate a synchronization frame set.

3. The method for digital twin modeling of a park based on generative AI technology according to claim 1, characterized in that, Step S2 includes: For the synchronous point cloud, an adjacency weight matrix based on distance and normal similarity is constructed. Topological noise is eliminated with adaptive intensity using the graph Laplacian inverse filter operator. Multi-layer voxel grids are divided with a reference voxel scale and a geometric series multiplication factor. Representative points are selected in each layer of the grid according to the principle of local point density priority to form a multi-resolution voxel set. The low-frequency and high-frequency components of the synchronous image are extracted sequentially in the frequency domain according to the multi-order harmonic window and superimposed according to the exponential fusion weight to generate full-spectrum features. The full-spectrum features are excited by the second-order Laplacian operator and multiplied by the hypersurface function of the gradient magnitude to obtain the enhanced image. The device state sequence is mapped to the power of the delay difference of each channel to generate a delay coupling matrix. The Euclidean distance penalty factor is adjusted by the delay coupling matrix. The coupling penalty is introduced into the membership update by using the fuzzy clustering algorithm. The iteration continues until the membership degree and the cluster center converge to obtain the state membership degree.

4. The method for digital twin modeling of a park based on generative AI technology according to claim 1, characterized in that, A sparse topological Laplacian operator is constructed for the multi-resolution voxel set of each spatial unit, and a high-order filter is applied cyclically to different frequency bands using the Chebyshev polynomial spectral filtering basis. The configuration spine spectrum is generated by power regularization of the spectral coefficients and skeleton weight allocation. Complex rotatable wavelet decomposition is performed on the enhanced image region corresponding to each spatial unit. Rotation scale invariance mapping is achieved through Fourier-Mellin transform. Holographic texture spectrum is constructed through superelliptical phase reconstruction and amplitude saturation operation. The fuzzy membership vector and time-delay coupling sequence of each spatial unit are input into the Tsallis non-additive entropy mapping, and combined with time-delay jitter exponential filtering and normalization mapping, the pulse entropy state spectrum is output.

5. The method for digital twin modeling of a park based on generative AI technology according to claim 1, characterized in that, Step S4 includes: For the mapping features of spatial units in the semantic spine mapping network, mapping vector groups are generated through query mapping, key mapping and value mapping; based on the high-order norm distance between mapping vectors and the learnable gating signal, adaptive correlation coefficients are calculated in multi-head adaptive correlation coding; The head-level aggregation of corresponding value mapping vectors is driven by adaptive correlation coefficients, and separable convolution is applied to the aggregation results of all heads to form node embeddings after multiple rounds of information propagation. In the output of each propagation layer, the embedding of the previous layer is fused by the residual ratio to achieve feature order superposition; The final layer embedding obtained through multiple rounds of information propagation is linearly mapped and combined with channel recalibration and dynamic normalization of temperature coefficient to amplify the normalized embedding amplitude and obtain the hierarchical semantic token.

6. The method for digital twin modeling of a park based on generative AI technology according to claim 1, characterized in that, Step S6 includes: For a 3D scene fragment at the target voxel point, a query vector and a key vector are generated through query mapping and key mapping; Distance decay is performed using an exponential function based on the difference in the higher-order norms of the query vector and the key vector, and unnormalized attention weights are generated by combining the global coupling coefficient. Power regularization and exponential mapping are applied to the unnormalized weights to obtain normalized attention coefficients. The absolute values ​​of the voxel attributes of the 3D scene fragments are taken and amplified by powers of the attention coefficient. Then, Minkowski pooling is performed on all fragment values. Weighted fusion is achieved through power accumulation and corresponding root operations, and the fused voxel field is output.

7. The method for digital twin modeling of a park based on generative AI technology according to claim 1, characterized in that, Also includes: The digital twin voxel field and material data that have reached the convergence condition are exported as rendering vertex and texture resources, and mapped to the scene graph structure of the visualization and interactive platform.

Images