A semantic search and distributed scheduling method and system for immersive scene resources

Abstract

This invention discloses a semantic search and distributed scheduling method for immersive scene resources, comprising the following steps: S1, decoupling and encoding of multi-dimensional semantic features of the scene based on a pre-trained language model; S2, asset retrieval based on multimodal contrastive learning and physical attribute completion based on GCN; S3, incremental hash comparison and dynamic bandwidth optimization scheduling based on an edge computing environment; S4, engine architecture alignment and metadata self-repair algorithm based on a directed acyclic graph (DAG). Simultaneously, a semantic search and distributed scheduling system for immersive scene resources is also disclosed, including: an NLP intent modeling and decoupling engine: mapping long text to a feature matrix; a cross-modal retrieval and attribute graph hub: connecting internal and external databases; a global bandwidth and edge scheduler: data block deduplication and dynamic bandwidth high-load balancing computation; and an engine-side dependency stream repair plugin: DAG topology analysis and underlying metadata path overwriting. This invention can significantly reduce the cost of acquiring materials and improve deployment stability.

Description

Technical Field

[0001] This invention belongs to the fields of data retrieval, cloud computing and virtual reality / mixed reality (XR) resource management technology, specifically relating to a semantic search and distributed scheduling method and system for immersive scene resources. Background Technology

[0002] With the explosive growth of metaverse, spatial computing, and XR technologies, the demand for high-quality 3D scene assets has increased dramatically across various industries. However, the existing processes for 3D resource retrieval, acquisition, and scene construction present the following significant technical challenges: Low search accuracy and intent understanding gap: Traditional 3D resource search relies heavily on discrete keyword (tag) matching, which cannot understand the abstract "scene-level" descriptions of long user texts (such as "oppression" and "Tang Dynasty style"), and lacks a cross-modal joint representation mechanism of text domain and 3D geometry / material domain.

[0003] Resource fragmentation and attribute missing: 3D models required for commercial scenarios often lack physical and interactive attributes (such as colliders, material hardness, coefficient of friction, etc.). Simply retrieving visual models cannot be directly used for XR haptic and physical interaction development.

[0004] High deployment threshold and fragile dependency chain: Industrial-grade 3D resources are huge (often in the GB range) and have extremely strict relative path dependencies when imported into large 3D engines (such as Unreal Engine). When combining resources across libraries, it is very easy to break the internal reference pointers between models (Mesh), materials, and textures, resulting in large areas of "gray models" and "texture misalignment" in the engine, and a high failure rate in scene deployment. Summary of the Invention

[0005] The purpose of this invention is to provide a semantic search and distributed scheduling method and system for immersive scene resources, which can greatly reduce the cost of acquiring materials and improve deployment stability.

[0006] To achieve the above objectives, the present invention provides: 1. A semantic search and distributed scheduling method for immersive scene resources, comprising the following steps: S1, Scene Multi-Dimensional Semantic Feature Decoupling and Encoding Based on Pre-trained Language Model: The system receives the user's natural language scene description X, and uses a large language model with an attention mechanism for feature decoupling, mapping the input text into three independent high-dimensional semantic feature vector spaces: entity vectors... Environmental atmosphere vector With interaction expectation vector The final fusion output is a scene-level text feature description matrix. This serves as the query input for subsequent cross-modal retrieval; S2, Asset Retrieval Based on Multimodal Contrast Learning and Physical Attribute Completion Based on GCN: Constructing a text-3D cross-modal joint embedding space, and calculating the text feature description matrix output from S1 in a distributed vector database. With 3D asset feature matrix High-precision recall is achieved using cosine similarity between the models; simultaneously, a "material-physics" knowledge graph is introduced for the recalled visual 3D models. ,in, Represents a set of nodes with physical and tactile attributes. The set of edges representing the relationships between attributes is used to automatically complete the missing physical and tactile attributes using a graph convolutional neural network (GCN), thereby generating a set of 3D asset resource list files F to be downloaded; S3, Incremental hash comparison and dynamic bandwidth optimization scheduling based on edge computing environment: Divide the resource manifest file set F determined in S2 into several fixed-length data blocks. Extract the hash values ​​of each fixed-length data block. Incremental deduplication comparison is performed with the target terminal's local cache tree. For missing data blocks, a network flow scheduling model based on the edge node status is constructed to complete the physical transmission of all 3D assets to the terminal. S4, Engine Architecture Alignment and Metadata Self-Repair Algorithm Based on Directed Acyclic Graph (DAG): After the 3D asset transfer is complete, for the 3D assets recalled in S2 and transferred to the terminal via S3, the engine-side automated plugin parses the underlying binary header of these files and abstracts their internal dependencies into a directed acyclic graph. ,in Represents an asset file node. This represents the reference relationship edges between files; when a break in the reference path caused by cross-library transfer is detected, i.e. If the pointer points to a null pointer, the system will initiate an adaptive path repair algorithm.

[0007] As a further aspect of the present invention: the formula for calculating the attention distribution in the feature decoupling process in S1 is expressed as follows: ; in, The attention weight score between the i-th word and the j-th word; The query vector representing the i-th word in the sequence; The key vector representing the j-th word in the sequence; This represents the key vector of the m-th word in the sequence; T is used as a superscript here to represent the transpose of a matrix or vector. The feature dimension representing the key vector. Used to scale the dot product result to prevent gradient vanishing; This represents the total word length of the input natural language sequence; These represent specific word position indices within the input sequence.

[0008] As a further aspect of the present invention: The text feature description matrix is ​​calculated in S2. With 3D asset feature matrix The formula for calculating the cosine similarity between them is expressed as: ; in, represents the cross-modal comprehensive similarity score between natural language text and target 3D asset; k represents the enumeration index of semantic feature dimensions, with values ​​including entity, environment, and interaction dimensions; The preset importance weight coefficient represents the k-th semantic feature dimension; The feature vector representing the text in the k-th dimension; The feature vector representing the 3D asset in the k-th dimension; This represents the L2 norm of the solution vector.

[0009] As a further aspect of this invention: the graph convolutional neural network GCN in S2 automatically completes the missing physical and tactile attributes, and the node update formula is expressed as: ; in, This represents the hidden feature representation of the material node p in the (l+1)th network layer of the knowledge graph. Represents a non-linear activation function; q represents the set of neighboring nodes that are directly connected to node p in the graph; q represents the index of a particular neighboring node in the set of neighboring nodes. This represents the edge weight normalization constant between node p and node q, used to balance node information of different degrees; This represents the learnable weight matrix in the l-th layer of the graph convolutional network; represents the hidden feature representation of neighbor node q in the l-th network layer; l represents the current index of the layer depth of the graph convolutional network.

[0010] As a further aspect of this invention: the network flow scheduling model construction method in S3 is as follows: It has There are 10 available edge nodes, and the available network bandwidth of node y is 10 ... The network latency to the target terminal is Define a binary variable for the resource allocation matrix. The objective function is to minimize the total download time. ; The constraints are as follows: This ensures that each data block is processed by one and only one node. y represents the total number of available edge nodes participating in the scheduling; n represents the index of the edge node; x represents the total number of fixed-length data blocks that are divided; and y represents the index of the fixed-length data block. Representing resource allocation decision variables, if data blocks The value is 1 if the data is assigned to the edge node y for transmission, and 0 otherwise. Represents data blocks The actual file size in bytes; This represents the current available transmission bandwidth of edge node y; The weighted penalty coefficient representing the balance between bandwidth transmission time and physical delay time; Represents the physical network latency from edge node y to the target terminal; By solving the objective function using a greedy algorithm or linear programming, extremely fast parallel scheduling between P2P and edge nodes can be achieved.

[0011] As a further aspect of this invention: In S4, a dual verification mechanism using Levenstein distance and the globally unique identifier (GUID) is employed to perform a nearest neighbor path search within the local Content directory, expressed by the formula: ; in, Represents the Levenstein edit distance between two string paths; This represents the original reference path of the broken or invalid element discovered during the 3D asset analysis process; This represents the candidate repair paths obtained by traversing the local cache directory in the terminal. This represents the cost of inserting a character into a string. This represents the cost of deleting a character from a string. This represents the cost of replacing a character in a string. Filter out After finding the correct local path with the smallest and matching GUID, the plugin forcibly overwrites the relative offset in the file's hexadecimal header, rebuilds the correct dependency tree, and ensures that materials and textures are 100% automatically restored in the 3D engine.

[0012] To achieve the above objectives, the present invention also provides a semantic search and distributed scheduling system for immersive scene resources for implementing the above method, comprising: NLP Intent Modeling and Decoupling Engine: Includes text processing components and a multi-dimensional embedding model for mapping long texts into feature matrices; Cross-modal retrieval and attribute graph hub: connects internal and external databases, and has a built-in multimodal similarity calculation engine and GCN physical attribute derivation graph; Global Bandwidth and Edge Scheduler: Responsible for SHA-256 hash deduplication of data blocks, dynamic bandwidth high load balancing calculation, and parallel stream transmission control; Engine-side dependency stream repair plugin: Resides within the terminal engine and is responsible for performing DAG topology analysis and underlying metadata path overwriting.

[0013] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention utilizes attention mechanisms and joint embedding spaces to quantify abstract "atmosphere, lighting, and interactive intent" into computable mathematical vectors, breaking through the single keyword matching mechanism and achieving high recall of overall resources at the "scene level".

[0014] This invention introduces attribute completion for the first time in the visual asset distribution stage. Based on graph neural networks, it automatically infers tactile / physical parameter tables through material types, breaking down the barriers from "visual assets" to "XR interactive assets".

[0015] This invention transforms the traditional unordered HTTP download into an operations research-based allocation problem with time minimization as the objective function. Combined with hash increment comparison, it greatly saves backbone network bandwidth usage for cross-regional transmission.

[0016] This invention addresses the vulnerability of asset pointers within the Unreal Engine by innovatively proposing edit distance search and automatic redirection in the underlying hexadecimal header, thus eradicating the persistent problems of "texture loss" and "material breakage" common in 3D industrial collaboration.

[0017] This invention significantly lowers the professional threshold; operators only need to interact with the system using natural language, and the system can automatically complete the construction of a concrete, high-precision 3D scene from an abstract concept. The success rate of scene deployment is close to 100%, and the path self-healing algorithm eliminates the path errors often made during manual import, eliminating the "red error message" and "gray model" phenomena in Unreal Engine. Deployment time is shortened exponentially, and thanks to hash deduplication and edge-optimized scheduling, the time for acquiring and assembling large GB-level XR scene resources is reduced from several days to minutes. Attached Figure Description

[0018] Figure 1This is a flowchart of the semantic search and distributed scheduling method for immersive scene resources according to the present invention.

[0019] Figure 2 This is a block diagram of the semantic search and distributed scheduling system for immersive scene resources of the present invention. Detailed Implementation

[0020] The present invention will be further illustrated by the following examples.

[0021] like Figure 1 As shown, a semantic search and distributed scheduling method for immersive scene resources includes the following steps: S1, Scene Multi-Dimensional Semantic Feature Decoupling and Encoding Based on Pre-trained Language Model: The system receives the user's natural language scene description X, and uses a large language model with an attention mechanism (such as BERT or its variants) for feature decoupling, mapping the input text into three independent high-dimensional semantic feature vector spaces: entity vectors... Environmental atmosphere vector With interaction expectation vector The final fusion output is a scene-level text feature description matrix. This serves as the query input for subsequent cross-modal retrieval.

[0022] Specifically, the formula for calculating the attention distribution in the feature decoupling process is expressed as follows: ; This includes the following steps: S1, Scene Multi-Dimensional Semantic Feature Decoupling and Encoding Based on Pre-trained Language Model: The system receives the user's natural language scene description X, and uses a large language model with an attention mechanism (such as BERT or its variants) for feature decoupling, mapping the input text into three independent high-dimensional semantic feature vector spaces: entity vectors... Environmental atmosphere vector With interaction expectation vector The final fusion output is a scene-level text feature description matrix. This serves as the query input for subsequent cross-modal retrieval.

[0023] Specifically, the formula for calculating the attention distribution in the feature decoupling process is expressed as follows: ; Specifically, a text-3D cross-modal joint embedding space is constructed, and the text feature description matrix output from S1 is calculated in a distributed vector database. With 3D asset feature matrix High-precision recall is achieved using cosine similarity between the two pairs of samples, expressed by the formula: ; in, represents the cross-modal comprehensive similarity score between natural language text and target 3D asset; k represents the enumeration index of semantic feature dimensions, with values ​​including three dimensions: entity (ent), environment (env), and interaction (int); The preset importance weight coefficient represents the k-th semantic feature dimension; The feature vector representing the text in the k-th dimension; The feature vector representing the 3D asset in the k-th dimension; The L2 norm (modulus) of the solution vector; Meanwhile, a "material-physics" knowledge graph was introduced for the recalled visual 3D models. ,in, Represents a set of nodes with physical and tactile attributes. The set of edges representing the relationships between attributes is used to automatically complete the missing physical and tactile attributes using a graph convolutional neural network (GCN), thereby generating a set of 3D asset resource list files F to be downloaded.

[0024] Specifically, the Graph Convolutional Neural Network (GCN) automatically completes the missing physical and tactile attributes, and the node update formula is expressed as follows: ; in, This represents the hidden feature representation of the material node p in the (l+1)th network layer of the knowledge graph. Represents a non-linear activation function; q represents the set of neighboring nodes that are directly connected to node p in the graph; q represents the index of a particular neighboring node in the set of neighboring nodes. This represents the edge weight normalization constant between node p and node q, used to balance node information of different degrees; This represents the learnable weight matrix in the l-th layer of the graph convolutional network; represents the hidden feature representation of neighbor node q in the l-th network layer; l represents the current index of the layer depth of the graph convolutional network.

[0025] S3, Incremental hash comparison and dynamic bandwidth optimization scheduling based on edge computing environment: Divide the resource manifest file set F determined in S2 into several fixed-length data blocks. Extract the hash values ​​of each fixed-length data block. Incremental deduplication comparison is performed with the target terminal's local cache tree. For missing data blocks, a network flow scheduling model based on the edge node status is constructed to complete the physical transmission of all 3D assets to the terminal.

[0026] Specifically, the network flow scheduling model is constructed as follows: It has There are 10 available edge nodes, and the available network bandwidth of node y is 10 ... The network latency to the target terminal is Define a binary variable for the resource allocation matrix. The objective function is to minimize the total download time. ; The constraints are as follows: This ensures that each data block is processed by one and only one node. y represents the total number of available edge nodes participating in the scheduling; n represents the index of the edge node; x represents the total number of fixed-length data blocks that are divided; and y represents the index of the fixed-length data block. Representing resource allocation decision variables, if data blocks The value is 1 if the data is assigned to the edge node y for transmission, and 0 otherwise. Represents data blocks The actual file size in bytes; This represents the current available transmission bandwidth of edge node y; The weighted penalty coefficient representing the balance between bandwidth transmission time and physical delay time; Represents the physical network latency from edge node y to the target terminal; By solving the objective function using a greedy algorithm or linear programming, extremely fast parallel scheduling between P2P and edge nodes can be achieved.

[0027] S4, an engine architecture alignment and metadata self-repair algorithm based on a directed acyclic graph (DAG): After the 3D assets are transferred, for the 3D assets (whose underlying file format is .uasset or .gltf) recalled in S2 and transferred to the terminal via S3, the engine-side automation plugin parses the underlying binary header of these files and abstracts their internal dependencies into a directed acyclic graph. ,in Represents an asset file node. This represents the reference relationship edges between files; when a break in the reference path caused by cross-library transfer is detected, i.e. If the pointer points to a null pointer, the system will initiate an adaptive path repair algorithm.

[0028] Furthermore, S4 utilizes a dual verification mechanism of Levenshtein distance and globally unique identifier (GUID) to perform nearest neighbor path search within the local Content directory. The formula is as follows: ; in, Represents the Levenstein edit distance (minimum number of operation steps) between two string paths; This represents the original reference path of the broken or invalid element discovered during the 3D asset analysis process; This represents the candidate repair paths obtained by traversing the local cache directory in the terminal. This represents the cost of inserting a character into a string. This represents the cost of deleting a character from a string. This represents the cost of replacing a character in a string. Filter out After finding the correct local path with the smallest and matching GUID, the plugin forcibly overwrites the relative offset in the file's hexadecimal header, rebuilds the correct dependency tree, and ensures that materials and textures are 100% automatically restored in the 3D engine.

[0029] A semantic search and distributed scheduling system for immersive scene resources to implement the above method includes: NLP Intent Modeling and Decoupling Engine: Includes text processing components and a multi-dimensional embedding model for mapping long texts into feature matrices; Cross-modal retrieval and attribute graph hub: connects internal and external databases, and has a built-in multimodal similarity calculation engine and GCN physical attribute derivation graph; Global Bandwidth and Edge Scheduler: Responsible for SHA-256 hash deduplication of data blocks, dynamic bandwidth high load balancing calculation, and parallel stream transmission control; Engine-side dependency stream repair plugin: Resides within the terminal engine and is responsible for performing DAG topology analysis and underlying metadata path overwriting.

[0030] A specific implementation example is as follows: A cultural tourism company needs to quickly switch tourists to a "bustling market scene in the style of the Tang Dynasty" in its large-space VR experience hall. The operator inputs the description in the system front end: "Generate a bustling market street in the style of the Tang Dynasty, with warm-toned lighting and a prosperous atmosphere. There are interactive wooden vendors and paper lanterns fluttering in the wind on both sides of the street." 1) Semantic decoupling of scenario requirements After receiving the description, the NLP model uses a Self-Attention network to perform weight allocation and feature decoupling calculations, outputting a feature matrix. : Entity vector space ( High-dimensional representations extracted: [Tang Dynasty street architecture, ancient vendors, paper lanterns].

[0031] Environment vector space ( Extract and map: Warm Lighting preset (Warm Lighting RGB:255,180,100); "Bustling" is mapped to a crowd white noise sound effect resource.

[0032] Interactive vector space ( ) Analyze the skeletal animation requirements (lantern swaying in the wind) and the physical collision requirements.

[0033] 2) Cross-modal vector association and GCN attribute completion Cross-modal similarity calculation: The system performs this calculation using a self-built vector library. Calculation. High-resolution data "Tang Dynasty Street Combination Model Asset_ID_098" and "High-Precision Street Vendor Asset_ID_102" were identified.

[0034] GCN Physical Parameter Derivation: Extract the visual material label (Albedo / Roughness determined as Wood) from the "street vendor" model. Input this label into the "Material-Physics" knowledge graph, and propagate the formula through node features. The model deduces the missing physical matrix configuration and generates an associated JSON: {"Friction_coefficient": 0.65, "Hardness_GPa": 0.05, "Restitution": 0.2}, providing a realistic mechanical basis for the VR glove's subsequent grasping interaction.

[0035] 3) Edge scheduling based on objective function After generating the final resource list (approximately 8GB in total size): Incremental deduplication: The 8GB data is divided into 1MB blocks $c_i$, and the SHA-256 hash is extracted. A tree-structured comparison is performed with the local library of the experience center, and common basic textures (such as tile normal maps) are matched, removing 3GB of redundancy.

[0036] Bandwidth optimization solution: 5GB of resources remain to be downloaded. The system obtains the status data (bandwidth) of 3 surrounding edge CDN nodes. ,Delay Substitute the parameters into the scheduling objective function. A linear programming solution was used. The solution guided the system to start 64 concurrent TCP threads to simultaneously pull the corresponding file slices from the optimal edge node, compressing the original 40-minute download process to 3.5 minutes.

[0037] 4) DAG-based UE structure alignment and metadata self-healing After the file is distributed to the local computer, the engine-side automation plugin starts working.

[0038] Topology analysis: The plugin parses the imported model and constructs a dependency graph via the Python API. A broken edge was found in the `Vendor_Mesh.uasset` file. The original creator's referenced texture path pointer was left in `D: / Old_Project / Textures / Wood_BaseColor.png`, which is invalid on the current machine.

[0039] Edit Distance Addressing and Header Overwriting: The plugin traverses the local standardized Content directory and uses formulas Calculate string similarity and locate the newly downloaded file in the local directory: / Game / TangDynasty / Materials / Wood_BaseColor.png.

[0040] Subsequently, the plugin enables the underlying byte stream mode, locates the hexadecimal header area of ​​the ExportMap / ImportMap in the .uasset file, and overwrites the original pointer offset with the calculated new relative path address.

[0041] Successful loading: After the dependency tree is repaired, the engine automatically instantiates the Actor in the Level. Tourism operators only need to click "Confirm" to instantly create a 3D Tang Dynasty bustling market scene with no missing materials or errors and complete preset physical tactile parameters.

Claims

1. A semantic search and distributed scheduling method for immersive scene resources, characterized in that, Includes the following steps: S1, Scene Multi-Dimensional Semantic Feature Decoupling and Encoding Based on Pre-trained Language Model: The system receives the user's natural language scene description X, and uses a large language model with an attention mechanism for feature decoupling, mapping the input text into three independent high-dimensional semantic feature vector spaces: entity vectors... Environmental atmosphere vector With interaction expectation vector The final fusion output is a scene-level text feature description matrix. This serves as the query input for subsequent cross-modal retrieval; S2, Asset Retrieval Based on Multimodal Contrast Learning and Physical Attribute Completion Based on GCN: Constructing a text-3D cross-modal joint embedding space, and computing the text feature description matrix output from S1 in a distributed vector database. With 3D asset feature matrix High-precision recall is achieved using cosine similarity between the models; simultaneously, a "material-physics" knowledge graph is introduced for the recalled visual 3D models. ,in, Represents a set of nodes with physical and tactile attributes. The set of edges representing the relationships between attributes is used to automatically complete the missing physical and tactile attributes using a graph convolutional neural network (GCN), thereby generating a set of 3D asset resource list files F to be downloaded; S3, Incremental hash comparison and dynamic bandwidth optimization scheduling based on edge computing environment: Divide the resource manifest file set F determined in S2 into several fixed-length data blocks. Extract the hash values ​​of each fixed-length data block. Incremental deduplication comparison is performed with the target terminal's local cache tree. For missing data blocks, a network flow scheduling model based on the edge node status is constructed to complete the physical transmission of all 3D assets to the terminal. S4, Engine Architecture Alignment and Metadata Self-Repair Algorithm Based on Directed Acyclic Graph (DAG): After the 3D asset transfer is complete, for the 3D assets recalled in S2 and transferred to the terminal via S3, the engine-side automated plugin parses the underlying binary header of these files and abstracts their internal dependencies into a directed acyclic graph. ,in Represents an asset file node. This represents the reference relationship edges between files; when a break in the reference path caused by cross-library transfer is detected, i.e. If the pointer points to a null pointer, the system will initiate an adaptive path repair algorithm.

2. The semantic search and distributed scheduling method for immersive scene resources according to claim 1, characterized in that, The formula for calculating the attention distribution in the feature decoupling process in S1 is expressed as: ； in, The attention weight score between the i-th word and the j-th word; The query vector representing the i-th word in the sequence; The key vector representing the j-th word in the sequence; This represents the key vector of the m-th word in the sequence; T is used as a superscript here to represent the transpose of a matrix or vector. The feature dimension representing the key vector. Used to scale the dot product result to prevent gradient vanishing; This represents the total word length of the input natural language sequence; These represent specific word position indices within the input sequence.

3. The semantic search and distributed scheduling method for immersive scene resources according to claim 1, characterized in that, Calculate the text feature description matrix in S2. With 3D asset feature matrix The formula for calculating the cosine similarity between them is expressed as: ； in, represents the cross-modal comprehensive similarity score between natural language text and target 3D asset; k represents the enumeration index of semantic feature dimensions, with values ​​including entity, environment, and interaction dimensions; The preset importance weight coefficient represents the k-th semantic feature dimension; The feature vector representing the text in the k-th dimension; The feature vector representing the 3D asset in the k-th dimension; This represents the L2 norm of the solution vector.

4. The semantic search and distributed scheduling method for immersive scene resources according to claim 3, characterized in that, In S2, the GCN convolutional neural network automatically completes the missing physical and tactile attributes. The node update formula is expressed as: ； in, This represents the hidden feature representation of the material node p in the (l+1)th network layer of the knowledge graph. Represents a non-linear activation function; q represents the set of neighboring nodes that are directly connected to node p in the graph; q represents the index of a particular neighboring node in the set of neighboring nodes. This represents the edge weight normalization constant between node p and node q, used to balance node information of different degrees; This represents the learnable weight matrix in the l-th layer of the graph convolutional network; represents the hidden feature representation of neighbor node q in the l-th network layer; l represents the current index of the layer depth of the graph convolutional network.

5. The semantic search and distributed scheduling method for immersive scene resources according to claim 1, characterized in that, How to construct a network flow scheduling model in S3: It has There are 10 available edge nodes, and the available network bandwidth of node y is 10 ... The network latency to the target terminal is Define a binary variable for the resource allocation matrix. The objective function is to minimize the total download time. ； The constraints are as follows: This ensures that each data block is processed by one and only one node. y represents the total number of available edge nodes participating in the scheduling; n represents the index of the edge node; x represents the total number of fixed-length data blocks that are divided; and y represents the index of the fixed-length data block. Representing resource allocation decision variables, if data blocks The value is 1 if the data is assigned to the edge node y for transmission, and 0 otherwise. Represents data blocks The actual file size in bytes; This represents the current available transmission bandwidth of edge node y; The weighted penalty coefficient representing the balance between bandwidth transmission time and physical delay time; Represents the physical network latency from edge node y to the target terminal; By solving the objective function using a greedy algorithm or linear programming, extremely fast parallel scheduling between P2P and edge nodes can be achieved.

6. The semantic search and distributed scheduling method for immersive scene resources according to claim 1, characterized in that, S4 utilizes a dual verification mechanism of Levenstein distance and the globally unique identifier (GUID) to perform nearest neighbor path search within the local Content directory. The formula is as follows: ； in, Represents the Levenstein edit distance between two string paths; This represents the original reference path of the broken or invalid element discovered during the 3D asset analysis process; This represents the candidate repair paths obtained by traversing the local cache directory in the terminal. This represents the cost of inserting a character into a string. This represents the cost of deleting a character from a string. This represents the cost of replacing a character in a string. Filter out After finding the correct local path with the smallest and matching GUID, the plugin forcibly overwrites the relative offset in the file's hexadecimal header, rebuilds the correct dependency tree, and ensures that materials and textures are 100% automatically restored in the 3D engine.

7. A semantic search and distributed scheduling system for immersive scene resources for implementing the semantic search and distributed scheduling method for immersive scene resources according to any one of claims 1-6, characterized in that, include: NLP Intent Modeling and Decoupling Engine: Includes text processing components and a multi-dimensional embedding model for mapping long texts into feature matrices; Cross-modal retrieval and attribute graph hub: connects internal and external databases, and has a built-in multimodal similarity calculation engine and GCN physical attribute derivation graph; Global Bandwidth and Edge Scheduler: Responsible for SHA-256 hash deduplication of data blocks, dynamic bandwidth high load balancing calculation, and parallel stream transmission control; Engine-side dependency stream repair plugin: Resides within the terminal engine and is responsible for performing DAG topology analysis and underlying metadata path overwriting.

Images