A digital space description method and system based on digital twin technology

Abstract

The application discloses a digital space description method and system based on digital twin technology, relates to the field of digital twin and three-dimensional rendering technology, and mainly solves the rendering lag and high delay problems caused by engine analysis and full tree traversal when a large amount of high-frequency dynamic data is poured in. By constructing a five-dimensional attribute model, the attribute node is directly associated with the physical memory location parameter and is issued to the rendering engine. When real-time dynamic data is listened to, the differential data set is extracted by combining snapshot data comparison, and the direct offset address of the target node in the video memory is calculated. Finally, the target graphic resource buffer is locally updated in increments, direct data overwriting is used instead of full tree traversal, the processing overhead in the rendering engine interpretation execution link is reduced, and local hot reloading and attribute synchronization with extremely low delay are realized.

Description

Technical Field

[0001] This invention discloses a digital space description method and system based on digital twin technology, relating to the fields of digital twin and 3D rendering technology. Background Technology

[0002] With the development of digital twin technology, the construction of digital spaces typically relies on standard 3D rendering engines to process cross-modal real-time dynamic data transmitted from physical space. However, existing technologies primarily receive and process data through the engine's standard application-layer API interface during data synchronization, lacking a mechanism to directly map upper-layer logical nodes to lower-layer physical memory location parameters. This traditional data flow model requires a lengthy process of parsing and interpreting business instructions, resulting in insurmountable communication barriers between the logical layer and physical memory.

[0003] When dealing with frequent local data update requests, the existing system's update chain reveals more severe performance bottlenecks. Specifically, when externally perceived data undergoes minor changes, because the system lacks a lightweight differential comparison mechanism based on snapshot data, the engine often has to perform a time-consuming full tree traversal of the entire spatial model's scene tree in order to address the corresponding target spatial feature. This processing method, which binds local state updates with global node searches, causes every influx of dynamic data to trigger large-scale ineffective computations and rendering pipeline queuing, easily resulting in dropped frames and high latency.

[0004] In summary, existing spatial description and rendering mechanisms based on digital twin technology suffer from core flaws such as weak underlying hardware penetration and inefficient data addressing and comparison updates. Due to over-reliance on lengthy interpretation execution pipelines and inefficient full tree traversal logic, existing architectures cannot achieve fast memory-level overwrite when massive amounts of perception data flood in. As a result, they consistently face insurmountable bottlenecks in computing power consumption and synchronization latency when dealing with high-frequency, extremely low-latency local spatial dynamic hot reload tasks. Summary of the Invention

[0005] The purpose of this invention is to provide a digital spatial description method and system based on digital twin technology. Its core lies in breaking through the performance bottleneck of traditional graphics rendering engines that rely on long-chain interpretation and execution. It innovatively hard-binds high-level five-dimensional attribute model nodes with low-level physical memory location parameters. When massive amounts of high-frequency dynamic data flood in, this solution uses a spatial indexing algorithm to quickly locate the target attribute node. It extracts a lightweight, differentiated dataset by accurately comparing it with snapshot data from the previous rendering frame. Based on the physical memory location parameters, it directly calculates the target node's offset position in the rendering engine's GPU resource buffer. Finally, it uses a direct memory access mechanism to perform local incremental overwriting of the target physical memory, thus replacing the traditional full tree traversal with direct data overwriting. This achieves extremely low-latency hot reloading by bypassing the engine's business layer, fundamentally solving the problems of computational redundancy and rendering lag in the high-frequency spatial state synchronization process.

[0006] The objective of this invention can be achieved through the following technical solutions: A digital space description method based on digital twin technology includes the following steps: S1: Obtain cross-modal data state of physical space, process it through spatiotemporal reference, and construct a standardized set of spatial elements; S2: Construct a five-dimensional attribute model based on the standardized spatial element set. The five-dimensional attribute model is organized into a tree structure, including a logical layer, a physical layer, a behavioral layer, a semantic layer, and a relational layer. The behavioral layer and the semantic layer are attached as dynamic attribute nodes to the leaf nodes of the tree structure to achieve mapping decoupling between the logical layer and the physical layer. S3: The five-dimensional attribute model is encapsulated into a digital space description file and sent to the target rendering engine. Each attribute node in the description file is associated with a physical memory location parameter. S4: Monitor real-time dynamic data in the physical space, and use the spatial indexing algorithm to locate the target attribute node based on the path positioning mechanism of the five-dimensional attribute model. S5: Calculate the direct offset address of the target attribute node in the target rendering engine's video memory based on the physical memory location parameters, extract the differential dataset, and perform local incremental overwriting of the target video memory address through direct memory access.

[0007] Preferably, the spatiotemporal reference processing in step S1 specifically includes: Extract the multivariate local sensing coordinate systems acquired by each cross-modal sensor; use an affine transformation matrix to map each multivariate local sensing coordinate system to a unified global three-dimensional coordinate system, and converge them into multi-source fusion data; perform topological pose correction on the multi-source fusion data based on a benchmark compensation algorithm to generate smooth multivariate spatial coordinate data.

[0008] Preferably, the step S1 of obtaining the cross-modal data state of the physical space further includes time synchronization processing, specifically including: Using the system rendering frame of the target rendering engine as the reference clock, a time sliding window is set for different modal data; linear interpolation is used to synchronously calculate the data within the same time sliding window to ensure the consistency of the state of the standardized spatial feature set at the same time.

[0009] Preferably, the decoupling of the logical layer and the physical layer is implemented as follows: The physical layer is defined as the trunk node of the tree structure, and the dynamic attribute nodes are mounted to the trunk node using index key-value pairs. When performing local hot reload, the target rendering engine directly locks the direct offset address of the video memory by matching the node path of the attribute node and replaces the full tree traversal with direct data overwriting.

[0010] Preferably, the behavior layer realizes spatial interaction through a finite state machine model. The finite state machine model receives business instructions through an API interface and, based on the topological connection network constructed by the relationship layer, determines the target spatial elements affected by the business instructions from the standardized spatial element set, thereby triggering virtual actions or state transitions of the target spatial elements.

[0011] Preferably, the specific steps of the spatial indexing algorithm in step S4 are as follows: An octree-based spatial dynamic partitioning algorithm is adopted. Based on the coordinate information of the real-time dynamic data, a top-down pruning search is performed in the spatial index tree to locate the node that intersects with the spatial bounding box and determine the target spatial elements affected by the real-time dynamic data.

[0012] Preferably, the specific steps for extracting the differential dataset in step S5 are as follows: Calculate the deviation value of the current real-time dynamic features of the target spatial features, and compare the deviation value with the snapshot data in the previous rendering frame; if the two are inconsistent, extract their respective specific attribute fields and new values, and generate lightweight deviation package data as the data source for local incremental overwriting.

[0013] Preferably, the direct memory access method in step S5 is as follows: The target rendering engine parses the path identifier in the lightweight deviation package data, calculates the physical address of the video memory based on the pre-stored physical memory alignment parameters, and directly overwrites the differentiated dataset to the physical address of the video memory using a memory copy instruction, thus skipping the interpretation and execution phase to achieve attribute synchronization.

[0014] This invention also provides a digital spatial description system based on digital twin technology, used to implement a digital spatial description method based on digital twin technology, comprising: The spatial data acquisition module is used to acquire cross-modal data status of physical space and construct a standardized set of spatial elements through spatiotemporal benchmark processing. The model building module is used to construct a five-dimensional attribute model based on the standardized spatial element set. The five-dimensional attribute model is organized into a tree structure, including a logical layer, a physical layer, a behavioral layer, a semantic layer, and a relational layer. The behavioral layer and the semantic layer are attached as dynamic attribute nodes to the leaf nodes of the tree structure to realize the mapping decoupling between the logical layer and the physical layer. The file generation and distribution module is used to encapsulate the five-dimensional attribute model into a digital space description file and send it to the target rendering engine. Each attribute node in the description file is associated with a physical memory location parameter. The real-time dynamic monitoring module is used to monitor real-time dynamic data in the physical space and, based on the path positioning mechanism of the five-dimensional attribute model, uses a spatial indexing algorithm to locate the target attribute node. The memory-driven update module is used to calculate the direct offset address of the target attribute node in the target rendering engine's video memory based on the physical memory location parameters, extract the differential dataset, and perform local incremental overwriting of the target video memory address through direct memory access.

[0015] The beneficial effects of this invention are: This invention bypasses the lengthy business parsing and interpretation execution phase of traditional rendering engines by directly binding spatial attribute nodes to physical memory location parameters and calculating the direct offset address of the video memory during data updates for direct memory access overwriting. This method of directly refreshing the underlying video memory across the software layer replaces the time-consuming full-scene tree graph traversal with direct data overwriting, thereby achieving extremely low-latency local dynamic hot reloading when facing a high frequency of real-time perception data in the physical space, ensuring extremely fast and accurate synchronization of the virtual and physical space states.

[0016] This invention introduces a mechanism for accurately comparing the current feature deviation value of target elements with the snapshot data of the previous rendering frame when processing real-time dynamic monitoring data in physical space. Only the specific attribute fields that have undergone state changes and their new values ​​are extracted. This inter-frame differential update strategy based on lightweight deviation packet data avoids the drawback of repeatedly submitting a large amount of redundant static data to the engine pipeline in traditional rendering modes. It not only reduces the scale of memory copying but also significantly reduces the transmission pressure on the underlying hardware while maintaining the stability of the topological pose of multi-source fused data.

[0017] The five-dimensional attribute model constructed in this invention attaches the behavioral and semantic layers as dynamic attribute nodes to the leaf nodes of a tree structure, successfully achieving decoupling of the system's logical and physical layers. This architectural design, combined with octree-based dynamic spatial partitioning and a top-down fast pruning search algorithm, enables the system to ignore the large and complex spatial framework and directly lock the affected target bounding boxes and element nodes when receiving business instructions and triggering corresponding spatial element actions through a finite state machine. While maintaining the stability of the underlying geometric structure, this achieves efficient isolation of complex twin business logic and extremely rapid response to interactive instructions. Attached Figure Description

[0018] Figure 1 This is a flowchart illustrating a digital spatial description method based on digital twin technology according to the present invention. Figure 2 This is a schematic diagram of the structure of a digital spatial description system based on digital twin technology according to the present invention. Detailed Implementation

[0019] The technical solutions of the present invention will be clearly and completely described below with reference to the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention. Example

[0020] like Figure 1 As shown in this embodiment, a digital space description method based on digital twin technology includes: Step S1: Obtain cross-modal data state of physical space, and construct a standardized spatial element set through spatiotemporal benchmark processing; At the system's front-end access layer, the edge computing gateway's physical network interface card (NIC) continuously receives cross-modal data status from the physical space via industrial Ethernet. This data is encapsulated in different network protocol packets, including LiDAR 3D point cloud data packets transmitted via TCP, robotic arm high-frequency joint rotation data packets transmitted via UDP, and workshop temperature and humidity scalar data packets transmitted via MQTT. The system's underlying network polling thread strips the packet headers, extracts the pure payload data, and pushes it into a pre-allocated lock-free concurrent queue.

[0021] The system extracts multivariate local sensing coordinate system data acquired by each cross-modal sensor from the concurrent queue. Each cross-modal sensor refers to a variety of heterogeneous sensing devices deployed in the physical space to collect different physical attributes and data dimensions; in this embodiment, it specifically refers to a lidar for collecting three-dimensional point cloud data, a robotic arm angle meter for collecting joint rotation data, and a thermometer and hygrometer for collecting environmental scalars.

[0022] Since each sensor uses its own physical mounting point as its coordinate origin, the system must perform an absolute coordinate system transformation before sending data to the rendering engine. The system allocates an independent register block for each physical sensor within the field-programmable gate array (FPGA) of the edge gateway or a dedicated hardware coprocessor. This register block pre-programs the affine transformation matrix generated during the physical space calibration phase of the sensor. This matrix is ​​a fourth-order square matrix, and its internal values ​​rigorously define the translational distance component of the sensor relative to the absolute zero coordinates of the entire digital twin shop floor, the rotational angle component around three mutually perpendicular axes in space, and any possible scaling error components.

[0023] When streaming multi-source local sensing coordinate system data enters the hardware pipeline, the system directly schedules the underlying multiplier-accumulator unit. The hardware performs pure numerical matrix multiplication, that is, multiplying the input local coordinate vector element-wise with each row and column of the affine transformation matrix and accumulating the results. Through hardware-level dot product operations within a single clock cycle, each multi-source local sensing coordinate system is mapped to a unified global 3D coordinate system, generating global coordinate data with a unified absolute origin. These global coordinate data mapped from different sensors are then aggregated into multi-source fused data in the underlying memory.

[0024] Multi-source fusion data is essentially a comprehensive data set formed by superimposing the sensing data from the aforementioned heterogeneous sensors in a global three-dimensional coordinate system. Since the spatial detection ranges of adjacent sensors often overlap in a physical factory, meaning that when the scanning ranges of two adjacent sensors physically overlap, resulting in two sets of coordinate points with slight deviations for the same physical entity in the global coordinate system, the system performs topological pose correction on the multi-source fusion data based on a benchmark compensation algorithm.

[0025] The system algorithm allocates a detection window in memory to scan discrete feature points within these overlapping regions. The system calculates the Euclidean distance spatial deviation of the same feature point across the two sets of data sources. Subsequently, the system reads preset sensor confidence parameters as the basis for weight allocation. The system multiplies the coordinate values ​​of high-confidence sensors by a high weight and the coordinate values ​​of low-confidence sensors by a low weight, then adds the two values ​​together and performs a weighted average smoothing process. Through this low-level topology pose correction, the system eliminates coordinate jumps and data tearing in overlapping regions, ultimately outputting smooth multivariate spatial coordinate data to the next level.

[0026] After spatial coordinate alignment, the system performs forced synchronization in the time dimension within the memory controller. The system physically extracts the vertical synchronization blanking interrupt signal from the graphics card hosting the target rendering engine via the motherboard's general purpose input / output interface or the system-level interrupt controller. The system uses the target rendering engine's system rendering frames as its base clock. Assuming the rendering engine is configured to run at 60 frames per second, the graphics card hardware sends a high-priority hardware interrupt signal to the system's central processing unit approximately every 16.66 milliseconds. The system sets this fixed 16.66 millisecond interval as a time sliding window. The system allocates a fixed-length circular data buffer in main memory for all connected sensing sources.

[0027] Each piece of data transmitted from the sensor, along with its precise microsecond-level timestamp upon arrival at the system, is sequentially written into this buffer. When the aforementioned hardware interrupt signal is triggered, the system kernel immediately suspends the current low-priority task, forcibly closes the current time sliding window, and triggers the synchronization calculation routine. The system traverses the circular buffer, searching for the two most recent valid data sampling points for each data channel before and after the current interrupt trigger timestamp. For data within the same time sliding window, the system uses linear interpolation for synchronization calculation.

[0028] Specifically, for scalar data, the system calculates the time span between the current interrupt trigger time and the previous sampling time, divides it by the total time span between the two sampling points, and obtains a time ratio coefficient between zero and one.

[0029] Subsequently, the system subtracts the previous sample value from the subsequent sample value, multiplies it by the scaling factor, and adds the previous sample value back to obtain the precise interpolated data for the 16.66 millisecond cross-section. For spatial rotation data, the system performs spherical linear interpolation calculations. Based on the calculated time scaling factor, the system calculates the shortest geodesic on a four-dimensional hypersphere composed of two quaternions and extracts the precise quaternion value corresponding to the scaling factor on the geodesic. Through the forced down-frequency sampling and up-frequency interpolation calculations performed at the moment the time sliding window closes, the system forcibly aligns all heterogeneous and asynchronous sensor data to the absolute same moment when the rendering engine is ready to render the next frame, thus ensuring the consistency of the spatial feature set at the same time at the data level. At this point, the system completes the construction of the standardized spatial feature set.

[0030] Step S2: Construct a five-dimensional attribute model based on the standardized spatial element set. The five-dimensional attribute model is organized into a tree structure, including a logical layer, a physical layer, a behavioral layer, a semantic layer, and a relational layer. The behavioral layer and the semantic layer are attached as dynamic attribute nodes to the leaf nodes of the tree structure to achieve mapping decoupling between the logical layer and the physical layer. The system extracts the standardized spatial feature set from the heap area of ​​main memory and initiates the reconstruction and encapsulation of the data structure to construct a five-dimensional attribute model. In order to eliminate the traversal addressing latency caused by the deep nesting of scene trees in traditional 3D graphics rendering engines, this system strictly organizes the five-dimensional attribute model into a flat tree structure at the bottom layer.

[0031] The model comprises a logical layer, a physical layer, a behavioral layer, a semantic layer, and a relational layer. The system employs a flat, contiguous memory storage mechanism, essentially an "array structure." The system first processes the physical layer. This layer contains the geometric appearance mesh data, normal vector data, and texture coordinate data of the basic materials for spatial features. Because this data is massive and rarely undergoes structural changes during digital twin operation, the system defines it as trunk nodes in a tree structure. The system calls the underlying memory allocation function to package this trunk node data into a contiguous read-only memory area for independent reading by the graphics rendering pipeline. Next, the system processes frequently changing dynamic data. The behavioral layer contains the dynamic capability parameters of spatial features, and the semantic layer contains the business metadata of spatial features. The system separates the behavioral and semantic layers from the traditional 3D object model, eliminating any memory pointers to the physical layer's geometric mesh.

[0032] The system allocates independent and absolutely contiguous one-dimensional memory arrays in main memory for the behavioral and semantic layers. These contiguous memory blocks are divided into uniformly sized structural units, serving as dynamic attribute nodes. The system generates a globally unique unsigned integer at the logical layer as an entity identifier, which acts as the hub connecting all dimensions. The system loosely mounts the dynamic attribute nodes of the behavioral and semantic layers to the descriptive nodes using key-value pairs with the entity identifier as the index key and the dynamic attribute memory block as the value. Through this physical separation of memory pointers, the system achieves decoupling of the mapping between the logical and physical layers at the underlying data architecture. During local hot reload, the target rendering engine does not need to deserialize or recursively traverse the massive geometric nesting network; it can directly locate the memory slot where the target dynamic attribute is located by directly querying the hash dictionary through node path matching of attribute nodes.

[0033] To enable the digital space to respond to complex business logic without relying on a rendering engine, a finite state machine model is independently constructed in the separated behavior layer memory. This model consists of a state register and a transition condition logic table. The system listens to the system data bus through an open API interface, receiving business instruction data streams from external manufacturing execution systems or underlying sensors in real time. Simultaneously, the system maintains a topological connection network composed of a directed acyclic graph in the relationship layer. This network records the physical dependencies and logical control relationships between all spatial elements in memory. When the finite state machine model receives a specific business instruction through the API interface, the internal logic of the state machine compares the current state with the transition conditions, instantly overwriting the machine tool's internal state register from "normal operation" to "stop alarm".

[0034] Next, the system traverses the topology network constructed by this relationship layer, searching for all child nodes of the machine tool as its parent node. The system state machine automatically sends state transition interrupt signals to the memory addresses of these child nodes, forcibly triggering virtual actions or state transitions of the corresponding target spatial elements. This series of complex spatial interaction logics is completed entirely in the pure numerical logic layer of main memory, without involving any graphics collision detection, achieving extremely fast logical closed loop.

[0035] Step S3: Encapsulate the five-dimensional attribute model into a digital space description file and send it to the target rendering engine. Each attribute node in the description file is associated with a physical memory location parameter. This is the most crucial underlying implementation for breaking the rendering engine's black box and establishing a direct memory write channel. The system serializes the five-dimensional attribute model built in main memory, encapsulates it into a compact digital space description file, uses a binary byte stream format, and sends it to the target rendering engine via inter-process communication or network sockets.

[0036] On the target rendering engine side, a privileged low-level dynamic link library plugin is pre-injected into the system. When the rendering engine receives this description file and prepares to allocate space in memory for instantiation, this low-level plugin, through system kernel calls, forcibly intercepts and takes over the engine's default memory allocation application programming interface. The plugin bypasses the standard library's dynamic memory allocation mechanism and directly initiates low-level system calls to the operating system's virtual memory manager. The plugin requests the allocation of a massive physical memory pool with "page locking" attributes. Due to the locking flag, the operating system kernel is forcibly prohibited from swapping this physical memory page to the hard drive's virtual swap area, and it is also immune to memory defragmentation and address relocation by the engine's built-in garbage collector. Therefore, the starting base address of this memory pool on the physical memory chip is permanently fixed.

[0037] Subsequently, the engine sequentially loads and instantiates each attribute node from the description file into the locked memory pool. During data arrangement, the plugin enforces memory alignment rules, ensuring that the starting address of each attribute variable is strictly aligned to the CPU's 64-byte cache line boundaries, thus avoiding performance penalties during multi-threaded concurrent overwriting. After loading, the plugin traverses the locked memory pool, precisely calculating the actual physical address allocated to each attribute node in the description file. This address equals the fixed base address of the locked memory pool plus the absolute byte offset of the attribute variable within the pool. The plugin encapsulates the globally unique identifier of each attribute node and its corresponding physical address into a "physical memory location parameter." The plugin aggregates the relationships between all nodes into a high-density hash map and transmits this map back to the system's main memory for permanent residency. At this point, each attribute node in the description file is absolutely and uniquely associated with its underlying physical memory location parameter.

[0038] After the target rendering engine receives the digital space description file, the system does not create objects through the engine's standard memory allocation interface. Instead, it intercepts and replaces the engine's default memory allocator through a pre-injected privileged dynamic link library plugin. This plugin directly calls the operating system kernel-level virtual memory management functions to request a physical memory pool with page-locking properties. The physical base address of this memory pool remains unchanged once allocated. Subsequently, the plugin traverses all attribute nodes in the five-dimensional attribute model, arranging them sequentially in the memory pool according to cache line alignment rules, recording the globally unique identifier of each attribute node and its absolute offset in the memory pool, and generating a hash mapping table. This mapping table is returned to the system's main control program as the basis for subsequent DMA overwrites.

[0039] Step S4: Monitor real-time dynamic data of the physical space, and use the spatial indexing algorithm to locate the target attribute node according to the path positioning mechanism of the five-dimensional attribute model. When massive amounts of real-time dynamic data continuously flood into the physical space, the system must locate the affected virtual elements within microseconds. Traditional 3D bounding box intersection traversal algorithms can lead to severe computational bottlenecks. This system adopts a spatial indexing algorithm based on pure bitwise operations.

[0040] The system first monitors real-time dynamic data from the physical space and extracts the real-time 3D coordinate information carried in the data packets. To eliminate the processing latency caused by floating-point comparisons, the system multiplies the floating-point coordinate data by a very large precision multiplier and rounds down to normalize it into unsigned integer coordinates. The system then performs bitstream interleaving at the underlying level. It extracts the binary bits of the X, Y, and Z integer coordinates and, using bit extraction and storage instructions in inline assembly, interleaves the binary bits of these three coordinates sequentially (e.g., first bit of the Z-axis, first bit of the Y-axis, first bit of the X-axis, second bit of the Z-axis, and so on), generating a 64-bit Morton code.

[0041] The Morton code maps the three-dimensional proximity relationships in three-dimensional space to the continuous numerical magnitude relationships of one-dimensional integers. The system maintains a pre-constructed spatial dynamic partitioning algorithm structure table based on the Morton code, stored in a flattened one-dimensional array within contiguous main memory blocks. The system initiates the spatial indexing algorithm engine, loading the newly generated real-time dynamic Morton code data into the high-level vector extension register of the central processing unit. The system performs a top-down fast pruning search in the spatial index tree: using single-instruction multiple-data bitwise comparison instructions, the system performs batch binary bitwise comparisons between the target Morton code and the upper and lower bound Morton code intervals of the spatial bounding boxes represented by the nodes at each level of the array. If the target Morton code is not within the interval of a certain bounding box, the system immediately cuts off that branch, directly skipping tens of thousands of child node memory addresses under that branch, achieving extremely fast pruning. If the intervals intersect, the system calculates the memory offset of the next child node and quickly probes down along the intersecting array path. Since the entire process is completed in the cache through pure logic gate numerical comparison, the system can directly hit the lowest leaf node intersecting the target space bounding box within hundreds of nanoseconds. The system reads the entity identifier stored in the memory slot of the leaf node to determine the target space element affected by real-time dynamic data. Subsequently, based on the path positioning mechanism of the five-dimensional attribute model, the system uses the entity identifier to perform O(1) complexity direct addressing in the hash dictionary to accurately locate all target attribute nodes associated with the element.

[0042] The path localization mechanism in the five-dimensional attribute model is essentially a direct lookup of a hash dictionary, replacing the time-consuming recursive traversal of the full-scene tree diagram in traditional rendering engines, by extracting identifiers through spatial localization. At the underlying architecture level, the system decouples dynamic attribute nodes from the massive physical geometric mesh and establishes a persistent hash mapping table in memory, using globally unique entity identifiers as index keys and the absolute physical location parameters of the nodes in video memory as values. When massive amounts of real-time dynamic data from the physical space flood in, the system first performs a top-down fast pruning search using an octree spatial indexing algorithm based on Morton codes, instantly hitting the leaf nodes of the affected spatial elements and extracting their corresponding entity identifiers. Subsequently, the system directly uses this identifier as a path clue to perform node path matching in the hash mapping table, achieving direct addressing with a time complexity of O(1). This design allows the system to ignore the massive and complex spatial nesting network and accurately locate the direct offset address of the target attribute node in the target rendering engine's video memory, thus clearing the addressing obstacles for subsequent direct memory access overwriting that bypasses the business layer.

[0043] Step S5: Calculate the direct offset address of the target attribute node in the target rendering engine's video memory based on the physical memory location parameters, extract the differential dataset, and perform local incremental overwriting of the target video memory address through direct memory access. After locating the target spatial feature, the system needs to extract the differential dataset. The system extracts the latest value of the feature's current real-time dynamic characteristics, i.e., the deviation value. To compress the data bus transmission bandwidth, the system retrieves the memory pointer of the feature's snapshot data from the previous rendering frame in main memory. Instead of performing any complex upper-level business logic judgments, the system directly calls the underlying 64-bit arithmetic logic unit to perform a hardware-level bitwise XOR operation between the latest value memory block and the snapshot data memory block. If the bitwise XOR output is all zeros, it proves that the current sensor data is completely consistent with the previous frame, and the system immediately discards the real-time data packet, blocking the invalid update process; if the two are inconsistent, the high-level bit in the operation result precisely indicates the changed byte position. The system uses bit scan instructions to quickly locate these transitional byte boundaries and accurately extract their specific attribute fields and their new values. The system independently extracts these non-zero fields and new values ​​from the massive structure and repackages them to generate a lightweight deviation packet data with a volume typically not exceeding tens of bytes. This lightweight deviation packet data serves as the sole data source for local incremental updates, greatly reducing the pressure on subsequent bus transmissions.

[0044] Finally, the system's main execution program intervenes, and the underlying parsing module of the target rendering engine is invoked. The system parses the entity path identifier carried in the lightweight offset packet data. Based on the identifier, the system looks up and extracts the pre-stored physical memory alignment parameters in the hash mapping table established and resident in memory in step S3. The system adds the memory pool base address to this offset, and through simple arithmetic addition, the system calculates the direct offset address of the target node in the rendering engine's video memory, thus obtaining the final absolutely accurate physical address in video memory. At this point, the system abandons the traditional, lengthy and time-consuming conventional chain of "calling the rendering API interface—adding to the engine rendering pipeline task queue—engine reflection to find the node—submitting to the graphics processor." The system directly issues low-level driver instructions to the direct memory access controller on the motherboard.

[0045] The system configures parameters in the DMA controller's descriptor register: the physical address of the lightweight offset packet data in main memory is configured as the data source address, and the physical address of the video memory calculated in the previous step is configured as the overwrite target address. To avoid screen tearing caused by the graphics processor reading this address while overwriting video memory, the system's hardware interrupt controller blocks and queues the DMA overwrite instruction. The DMA controller is only activated by a level signal when the system detects that the graphics processor has completed the rasterization scan of the current frame and its double buffering mechanism is in an absolute empty window where the background buffer is not locked by the rendering pipeline. At the moment of activation, the system directly overwrites the difference data to the target video memory physical address through direct memory access, without occupying any CPU computing power, by extending the interconnect data bus channel through high-speed peripheral components. Through the above-mentioned local incremental overwrite process, the rendering engine directly reads the latest attribute values, which have been modified by the underlying bus, from the video memory when preparing to render the next frame. This system replaces the traditional full tree traversal with direct data overwrite, skips the interpretation execution stage, and achieves attribute synchronization and local hot reloading with zero business instruction latency. Example

[0046] like Figure 2 As shown, in this embodiment, a digital spatial description system based on digital twin technology includes: The spatial data acquisition module is deployed within the edge computing server array of the factory area. It is equipped with a high-throughput network interface card (NIC) and a protocol parsing gateway. Its internal logic is divided into two main thread pools: Acquisition thread pool: Through an asynchronous I / O model, it maintains long-term connections with tens of thousands of sensors in a non-blocking manner, and is responsible for pushing raw cross-modal data streams into a lock-free concurrent queue.

[0047] The benchmark processing thread pool is responsible for performing spatiotemporal benchmark processing. For the spatial benchmark, it calls the GPU's CUDA cores to perform parallel affine transformation calculations on massive matrices; for the temporal benchmark, it maintains a global clock accurate to the nanosecond level, triggers the opening and closing of a time sliding window through a hardware timer, and uses interpolation algorithms to produce a standardized set of spatial features.

[0048] The model building module resides in the system's main memory on the cloud rendering server. Its core function is the orchestration of memory structures. This module includes an "ECS architecture assembler." During system initialization, the assembler parses the business configuration file, performing a "stripping operation" on traditional 3D assets. The massive vertex and face data is extracted and solidified into the physical layer; while the business data used for interaction and display is converted into a contiguous array of C++ structures (behavioral and semantic layers), which are mounted as dynamic attribute nodes on leaf nodes. This cache-friendly data structure design greatly reduces the CPU cache miss rate during subsequent runtime, achieving decoupling between logical and physical mapping.

[0049] The file generation and distribution module acts as a "bridge between the virtual and physical worlds." It not only serializes the five-dimensional attribute model into a compact digital space description file and sends it to the rendering engine, but more importantly, it integrates a "memory sniffer" deeply bound to the engine's underlying layers. When the engine allocates video memory for the digital space, the sniffer captures all memory allocation handles, extracts and calculates the absolute base address and relative physical memory location parameters of each attribute node mapped to VRAM. These parameters are then aggregated into a highly efficient hash table, residing in memory for rapid lookup by subsequent modules.

[0050] The real-time dynamic monitoring module maintains a highly balanced memory-based octree. To achieve a response time of millions of QPS, this octree is constructed entirely from contiguous memory blocks, eliminating latency caused by pointer jumps. When monitoring real-time dynamic data in the physical space, the module sends the three-dimensional coordinates of the new data to the octree processor. The processor uses bitwise operations to perform a top-down fast pruning search based on spatial Morton codes, locating nodes intersecting with the spatial bounding box within microseconds. This allows for precise identification of the target attribute node's UUID using the path localization mechanism of the five-dimensional attribute model.

[0051] The memory-driven update module is the most disruptive execution center of this invention, possessing the highest level of system privileges. The module includes a "high-speed differentialr" and a "DMA driver engine." The high-speed differentialr reads the caches of the current and previous rendered frames, performs a bitwise XOR operation, instantly screening out inconsistent dirty data and packaging it into a lightweight differential dataset. Subsequently, the DMA driver engine takes over control. Based on the locked node's UUID, it looks up the node's direct offset address in the target rendering engine's video memory and its physical memory alignment parameters from a hash table. This engine bypasses the operating system's conventional network or user-mode memory copy interface, directly sending control blocks to the motherboard's PCIe controller. Through direct memory access, during the intervals between frame breaks in the graphics card, the differential dataset is overwritten to the target video memory address with extremely high bandwidth. This local incremental overwrite method, completely stripping away the engine's upper-layer business logic, allows the rendering performance of the digital twin system to no longer be limited by the complexity of the business logic, truly achieving ultra-low latency digital space reconstruction where what you see is what happens.

[0052] Unlike existing incremental update methods based on octrees and Bayesian networks, this invention does not rely on CPU-side data layer difference comparison and priority queue scheduling. Instead, it reduces the update granularity to the physical memory address of a single attribute node, using a DMA controller to directly overwrite the video memory during GPU frame intervals. This mechanism of 'bypassing the rendering engine API and directly manipulating video memory' avoids blocking of the rendering pipeline caused by frequent calls to drawing commands, making it particularly suitable for scenarios with high-frequency dynamic data. Actual measurements show latency can be reduced to the microsecond level.

[0053] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0054] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A digital spatial description method based on digital twin technology, characterized in that, Includes the following steps: S1: Obtain cross-modal data state of physical space, process it through spatiotemporal reference, and construct a standardized set of spatial elements; S2: Construct a five-dimensional attribute model based on the standardized spatial element set. The five-dimensional attribute model is organized into a tree structure, including a logical layer, a physical layer, a behavioral layer, a semantic layer, and a relational layer. The behavioral layer and the semantic layer are attached as dynamic attribute nodes to the leaf nodes of the tree structure to achieve mapping decoupling between the logical layer and the physical layer. S3: The five-dimensional attribute model is encapsulated into a digital space description file and sent to the target rendering engine. Each attribute node in the description file is associated with a physical memory location parameter. S4: Monitor real-time dynamic data in the physical space, and use the spatial indexing algorithm to locate the target attribute node based on the path positioning mechanism of the five-dimensional attribute model. S5: Calculate the direct offset address of the target attribute node in the target rendering engine's video memory based on the physical memory location parameters, extract the differential dataset, and perform local incremental overwriting of the target video memory address through direct memory access.

2. The digital spatial description method based on digital twin technology according to claim 1, characterized in that, The spatiotemporal reference processing in step S1 specifically includes: Extract the multivariate local sensing coordinate systems acquired by each cross-modal sensor; use an affine transformation matrix to map each multivariate local sensing coordinate system to a unified global three-dimensional coordinate system, and converge them into multi-source fusion data; perform topological pose correction on the multi-source fusion data based on a benchmark compensation algorithm to generate smooth multivariate spatial coordinate data.

3. The digital spatial description method based on digital twin technology according to claim 1, characterized in that, The step S1 of obtaining the cross-modal data state of the physical space also includes time synchronization processing, specifically including: Using the system rendering frame of the target rendering engine as the reference clock, a time sliding window is set for different modal data; linear interpolation is used to synchronously calculate the data within the same time sliding window to ensure the consistency of the state of the standardized spatial feature set at the same time.

4. The digital spatial description method based on digital twin technology according to claim 1, characterized in that, The decoupling between the logic layer and the physical layer is implemented as follows: The physical layer is defined as the trunk node of the tree structure, and the dynamic attribute nodes are mounted to the trunk node using index key-value pairs. When performing local hot reload, the target rendering engine directly locks the direct offset address of the video memory by matching the node path of the attribute node and replaces the full tree traversal with direct data overwriting.

5. A digital spatial description method based on digital twin technology according to claim 1, characterized in that, The behavior layer realizes spatial interaction through a finite state machine model. The finite state machine model receives business instructions through an API interface and, based on the topological connection network constructed by the relationship layer, determines the target spatial elements affected by the business instructions from the standardized spatial element set, thereby triggering the virtual actions or state transitions of the target spatial elements.

6. The digital spatial description method based on digital twin technology according to claim 1, characterized in that, The specific steps of the spatial indexing algorithm in step S4 are as follows: An octree-based spatial dynamic partitioning algorithm is adopted. Based on the coordinate information of the real-time dynamic data, a top-down pruning search is performed in the spatial index tree to locate the node that intersects with the spatial bounding box and determine the target spatial elements affected by the real-time dynamic data.

7. A digital spatial description method based on digital twin technology according to claim 1, characterized in that, The specific steps for extracting the differential dataset in step S5 are as follows: Calculate the deviation value of the current real-time dynamic features of the target spatial features, and compare the deviation value with the snapshot data in the previous rendering frame; if the two are inconsistent, extract their respective specific attribute fields and new values, and generate lightweight deviation package data as the data source for local incremental overwriting.

8. The digital spatial description method based on digital twin technology according to claim 1, characterized in that, The direct memory access method in step S5 is specifically as follows: The target rendering engine parses the path identifier in the lightweight deviation package data, calculates the physical address of the video memory based on the pre-stored physical memory alignment parameters, and directly overwrites the differentiated dataset to the physical address of the video memory using a memory copy instruction, thus skipping the interpretation and execution phase to achieve attribute synchronization.

9. A digital spatial description system based on digital twin technology, used to implement the digital spatial description method based on digital twin technology as described in any one of claims 1-8, characterized in that, include: The spatial data acquisition module is used to acquire cross-modal data status of physical space and construct a standardized set of spatial elements through spatiotemporal benchmark processing. The model building module is used to construct a five-dimensional attribute model based on the standardized spatial element set. The five-dimensional attribute model is organized into a tree structure, including a logical layer, a physical layer, a behavioral layer, a semantic layer, and a relational layer. The behavioral layer and the semantic layer are attached as dynamic attribute nodes to the leaf nodes of the tree structure to realize the mapping decoupling between the logical layer and the physical layer. The file generation and distribution module is used to encapsulate the five-dimensional attribute model into a digital space description file and send it to the target rendering engine. Each attribute node in the description file is associated with a physical memory location parameter. The real-time dynamic monitoring module is used to monitor real-time dynamic data in the physical space and, based on the path positioning mechanism of the five-dimensional attribute model, uses a spatial indexing algorithm to locate the target attribute node. The memory-driven update module is used to calculate the direct offset address of the target attribute node in the target rendering engine's video memory based on the physical memory location parameters, extract the differential dataset, and perform local incremental overwriting of the target video memory address through direct memory access.

Images