A data visualization management system based on digital twinning

Abstract

This invention relates to the fields of digital twin, data visualization, and graphics processing technology, specifically a data visualization management system based on digital twins, comprising: an initialization module for dividing a digital twin 3D scene into a spatial mesh and constructing a visual tensor tree, binding data source addresses, vertex buffer objects, texture objects, and video memory quotas to nodes; an entropy field calculation module for calculating information entropy and generating a dynamic entropy pressure matrix; a resource pressure differential redistribution module for reclaiming resources from low-entropy pressure nodes and allocating them to high-entropy pressure nodes; and an adaptive rendering execution module for updating vertex buffer objects and texture objects according to the redistributed quotas, and rendering high- and low-entropy pressure nodes using first and second polygon meshes respectively, so as to maintain video memory occupancy and / or rendering computation load below a preset load threshold; this invention improves the real-time performance of resource migration and rendering updates in abnormal areas.

Description

Technical Field

[0001] This invention relates to the fields of digital twin, data visualization and graphics processing technology, specifically a data visualization management system based on digital twin. Background Technology

[0002] Data visualization management technology based on digital twins is a commonly used operation monitoring method in industrial scenarios. It can integrate equipment status data in physical space into a three-dimensional scene and display the device model, operating parameters, alarm information and trend charts in a linked manner, with multi-dimensional spatial expression capabilities and operation status presentation capabilities. With the development of the Industrial Internet of Things and 3D visualization technology, digital twin platforms have gradually evolved from static model displays to dynamic visualization systems that integrate real-time time-series data of multiple modalities such as temperature, pressure, vibration, flow rate, and gas concentration. However, when faced with large-scale scenarios and high-frequency concurrent data access, existing digital twin data visualization methods are still easily limited by graphics processor memory capacity, rendering load, and resource allocation mechanisms. This makes it difficult to achieve differentiated rendering of abnormal and stable areas on demand, thereby affecting the display clarity of key equipment and the overall efficiency of visualization management. Summary of the Invention

[0003] To address the aforementioned technical problems, this invention provides a data visualization management system based on digital twins. Specifically, the technical solution of this invention includes: The initialization module is used to divide the digital twin 3D scene into a spatial mesh and construct a visual tensor tree mapped to computer memory. The nodes of the visual tensor tree are bound to data source addresses, vertex buffer objects and texture objects, and are allocated graphics processor memory address blocks, initial graphics processor memory quotas and rendering instruction quotas. The initial graphics processor memory quotas of each node are combined to form the current graphics processor memory allocation matrix. The entropy field calculation module is used to receive real-time time-series data streams from the data source address, calculate the information entropy value of the real-time time-series data streams corresponding to each node using a sliding time window at the central processing unit, and calculate the dynamic entropy pressure corresponding to each node based on the information entropy value through a preset memory mapping function. The dynamic entropy pressure of each node is combined to generate a dynamic entropy pressure matrix with the same dimension as the current graphics processor memory allocation matrix. The resource pressure differential reallocation module is used to calculate the residual matrix between the current graphics processor memory allocation matrix and the dynamic entropy pressure matrix. If the corresponding element value in the residual matrix is ​​greater than zero, the node is determined to be a low-entropy pressure node. If the corresponding element value in the residual matrix is ​​equal to zero, the node is determined to be a stable node and its current graphics processor memory allocation matrix and rendering instruction quota remain unchanged. If the corresponding element value in the residual matrix is ​​less than zero, the node is determined to be a high-entropy pressure node, and the graphics processor memory address block and rendering instruction quota of the low-entropy pressure node are reclaimed. The reclaimed graphics processor memory address block and rendering instruction quota are then redistributed to the high-entropy pressure node to generate a reallocated memory quota. The adaptive rendering execution module is used to update the vertex buffer objects and texture objects of the digital twin 3D scene on the graphics processor side according to the reallocation of video memory quota. Specifically, when the node is a high-entropy pressure node, a first polygon mesh rendering is performed, and when the node is a low-entropy pressure node, a second polygon mesh rendering is performed, so as to keep the video memory utilization and / or rendering computation load of the graphics processor below a preset load threshold.

[0004] Preferably, the initialization module includes: a spatial resolution unit, used to resolve the spatial coordinate matrix of the digital twin 3D scene and divide the spatial coordinate matrix into a spatial grid containing multiple levels; and a tensor tree construction unit, used to convert the spatial grid into an octree-structured visual tensor tree and bind a preset device interface pointer as a data source address to the leaf node of the visual tensor tree, wherein the leaf node is associated with a data visualization chart for displaying real-time time-series data streams, and an initial graphics processor memory quota is allocated to each leaf node, wherein the initial graphics processor memory quota includes the upper limit of the number of vertices allocated and the texture resolution level.

[0005] Preferably, the entropy field calculation module includes: a data receiving unit, used to monitor the input / output throughput of the device interface corresponding to the device interface pointer, and collect real-time time-series data streams at a preset sampling frequency; an entropy value calculation unit, used to extract the probability distribution characteristics of the real-time time-series data streams within a sliding time window, and calculate the Shannon entropy value as the information entropy value based on the probability distribution characteristics; and a matrix generation unit, used to perform weighted calculations by combining the Shannon entropy value with preset node spatial importance weights or data source type weights to generate the dynamic entropy pressure of each node, and to generate a dynamic entropy pressure matrix by combining the dynamic entropy pressures of each node.

[0006] Preferably, the resource pressure differential redistribution module includes: a residual comparison unit, used to extract the current graphics processor memory allocation matrix and perform subtraction processing between the current graphics processor memory allocation matrix and the dynamic entropy pressure matrix to obtain a residual matrix; The virtual pressure transmission unit is used to identify nodes with corresponding element values ​​greater than zero in the residual matrix as high residual nodes and nodes with corresponding element values ​​less than zero as low residual nodes. It also establishes a virtual pressure transmission tree by combining the hierarchical structure of the visual tensor tree with the residual matrix, and determines the path for resources to migrate from low-entropy pressure nodes to high-entropy pressure nodes based on the pressure gradient direction from high residual nodes to low residual nodes in the virtual pressure transmission tree as the direction of gravity. The memory remapping unit is used to preempt the graphics processor memory address blocks of low-entropy pressure nodes and remap the pointers of the released graphics processor memory address blocks to high-entropy pressure nodes to generate reallocated memory quotas.

[0007] Preferably, the adaptive rendering execution module includes: an instruction compilation unit for compiling reallocated video memory quotas into batch rendering instructions at the graphics processor level; and a state update unit for executing the batch rendering instructions to update vertex buffer objects and texture objects on the graphics processor side.

[0008] Preferably, the state update unit is further configured to: determine that a node is a high-entropy-pressure node when the dynamic entropy pressure of a node is greater than or equal to a preset entropy pressure threshold, and load the preset first level of detail model data into the vertex buffer object for first polygon mesh rendering; and determine that a node is a low-entropy-pressure node when the dynamic entropy pressure of a node is less than a preset entropy pressure threshold, and forcibly replace the model data corresponding to the node with the preset second level of detail proxy model data for second polygon mesh rendering.

[0009] Preferably, the system further includes: a thread scheduling module, used to allocate the highest priority CPU computing thread to the data visualization chart bound to the high-entropy-pressure node when the node is a high-entropy-pressure node; and a baseline update module, used to extract the rendering performance data output by the adaptive rendering execution module and feed the rendering performance data back to the initial graphics processor memory quota to update the preset historical memory allocation baseline.

[0010] Preferably, the adaptive rendering execution module further includes: a visual focusing unit, used to reduce the texture resolution level corresponding to the low-entropy pressure node and apply a preset blur shader when rendering the second polygonal mesh, and to increase the texture resolution level corresponding to the high-entropy pressure node and apply a preset highlight shader when rendering the first polygonal mesh.

[0011] Preferably, when constructing a visual tensor tree, the tensor tree construction unit stores the hierarchical relationship of each node as a pointer structure in memory, and associates it with the corresponding graphics processor memory address block through the pointer structure.

[0012] Preferably, the real-time time-series data stream is a multimodal device status data stream collected by an Internet of Things (IoT) sensor network deployed in physical space.

[0013] Compared with the prior art, the present invention has the following beneficial effects: 1. This system divides the digital twin 3D scene into a spatial grid and constructs a visual tensor tree mapped to computer memory during the initialization phase. It binds the data source address, vertex buffer object, texture object, graphics processor memory address block, initial graphics processor memory quota and rendering instruction quota to each node. This enables the formation of an addressable, recyclable and traceable structured correspondence between the scene, device and graphics resources, thereby providing unified underlying support for local resource recycling, memory remapping and local redrawing in large-scale industrial scenarios. 2. By storing the hierarchical relationship of each node as a pointer structure in memory and associating it with the corresponding graphics processor memory address block, the overhead of number lookup and cross-table location can be reduced during high-frequency resource scheduling, thereby improving the real-time performance of resource migration and rendering updates in abnormal areas. 3. By allocating the highest priority CPU computing thread to the data visualization charts bound to high-entropy pressure nodes in the thread scheduling, this system can keep the refresh of charts such as trend charts, spectrum charts and alarm timelines of abnormal nodes synchronized with the improvement of 3D model accuracy, thereby avoiding the situation where the model is clearly displayed but the chart update is lagging behind. 4. This system extracts the rendering performance data output by the adaptive rendering execution module and feeds it back to the initial graphics processor memory quota to update the historical memory allocation baseline. This allows the initial resource allocation to be gradually corrected from a static constant to a historical baseline that closely reflects the actual operating rules, thereby improving the adaptability of resource allocation in subsequent operating cycles and the overall visualization management effect. Attached Figure Description

[0014] The present invention will be further explained below with reference to the accompanying drawings and embodiments: Figure 1 This is a schematic diagram of a data visualization management system based on digital twins, provided as an embodiment of the present invention. Detailed Implementation

[0015] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to specific embodiments.

[0016] A data visualization management system based on digital twins includes: an initialization module, used to divide the digital twin 3D scene into a spatial grid and construct a visual tensor tree mapped to computer memory, wherein the nodes of the visual tensor tree are bound to data source addresses, vertex buffer objects and texture objects, and are allocated graphics processor memory address blocks, initial graphics processor memory quotas and rendering instruction quotas, and the initial graphics processor memory quotas of each node are combined to form the current graphics processor memory allocation matrix; The entropy field calculation module is used to receive real-time time-series data streams from the data source address, calculate the information entropy value of the real-time time-series data streams corresponding to each node using a sliding time window at the central processing unit, and calculate the dynamic entropy pressure corresponding to each node based on the information entropy value through a preset memory mapping function. The dynamic entropy pressure of each node is combined to generate a dynamic entropy pressure matrix with the same dimension as the current graphics processor memory allocation matrix. The resource pressure differential reallocation module is used to calculate the residual matrix between the current graphics processor memory allocation matrix and the dynamic entropy pressure matrix. If the corresponding element value in the residual matrix is ​​greater than zero, the node is determined to be a low-entropy pressure node. If the corresponding element value in the residual matrix is ​​equal to zero, the node is determined to be a stable node and its current graphics processor memory allocation matrix and rendering instruction quota remain unchanged. If the corresponding element value in the residual matrix is ​​less than zero, the node is determined to be a high-entropy pressure node, and the graphics processor memory address block and rendering instruction quota of the low-entropy pressure node are reclaimed. The reclaimed graphics processor memory address block and rendering instruction quota are then redistributed to the high-entropy pressure node to generate a reallocated memory quota. The adaptive rendering execution module is used to update the vertex buffer objects and texture objects of the digital twin 3D scene on the graphics processor side according to the reallocation of video memory quota. Specifically, when the node is a high-entropy pressure node, a first polygon mesh rendering is performed, and when the node is a low-entropy pressure node, a second polygon mesh rendering is performed, so as to keep the video memory utilization and / or rendering computation load of the graphics processor below a preset load threshold.

[0017] This embodiment provides a data visualization management mechanism based on digital twins, such as... Figure 1 As shown; specifically, the digital twin scheduling platform of a large-scale refining and chemical park is the main scenario. The park includes crude oil tank area, hydrocracking unit, power station and transmission pipeline. The platform needs to simultaneously access high-frequency equipment data such as temperature, pressure, vibration, flow rate and combustible gas concentration, and continuously display equipment status, alarm area and trend chart in the three-dimensional scene. Specifically, during the initialization phase, the system spatially segments the entire 3D scene and constructs a visual tensor tree in memory. The spatial coordinates, data source address, and video memory quota contained in each node of the visual tensor tree constitute the tensor structure. The outermost layer of the park can be regarded as the root node, and then subdivided into device area nodes, unit device nodes, and leaf nodes. Each node records three types of basic information: first, the data source address of the corresponding device or area; second, the vertex buffer object and texture object associated with the node; and third, the video memory address block, initial video memory quota, minimum reserved quota, and rendering instruction quota allocated to the node. For ease of explanation, assume that only four leaf nodes are selected for illustrative quantization at the current moment, namely the reactor node. Compressor node Heat exchanger nodes and utility tunnel nodes If the initial video memory allocation for the four nodes is set to 8, 6, 4, and 2 resource units respectively, then the current video memory allocation matrix formed at this moment is denoted as... After completing one round of resource recycling and reallocation, the updated reallocation memory quota matrix is ​​denoted as follows: ; During operation, the entropy field calculation module continuously receives data streams from each node and calculates the information entropy value on the central processing unit side according to a sliding time window; for example, using the most recent 10 seconds as a window, if... If the pressure data fluctuates beyond a preset fluctuation threshold within the window, and the dispersion of the discretized probability distribution is greater than a preset dispersion threshold, then its entropy value is greater than a preset high entropy threshold; if If the temperature change rate of the pipe gallery is lower than the preset change threshold within a preset time period and the probability distribution is concentrated, then its entropy value is low. Continuing with the exemplary quantization deduction, let the normalized entropy values ​​calculated from the four nodes be 0.9, 0.7, 0.3, and 0.1, respectively, and then map them to a dynamic entropy pressure matrix using a preset memory mapping function.

[0018] The mapping function can be a linear function or a piecewise function. For example, it can map the entropy value in the range of 0 to 1 to the quota expectation range of 1 to 10 to reflect the current demand intensity of the node for video memory and rendering capabilities. The resource pressure differential reallocation module compares the current memory allocation matrix with the dynamic entropy pressure matrix to generate a residual matrix:

[0019] Under this definition, a negative residual indicates insufficient allocation, and the corresponding node is classified as a high-entropy pressure node; a positive residual indicates relative over-allocation, and the corresponding node tends to be considered a low-entropy pressure node; in this case, , Areas identified as requiring additional resources , Areas identified as where resources can be released; System Recycling and A portion of the video memory address block and its corresponding rendering instruction quota are reclaimed, for example, one resource unit from each block, resulting in a total of two transferable resource units; these two resource units are then remapped to... and This updates the reallocated video memory quota to: or

[0020] Specifically, it depends on the pre-set minimum retention quota policy, that is, during resource reclamation, the remaining video memory quota of any low-entropy pressure node must not be lower than its corresponding minimum retention quota; the updated quota not only changes the video memory usage, but also simultaneously affects the model precision, texture level and batch processing instruction number that the node can use. The adaptive rendering execution module updates vertex buffer objects and texture objects on the graphics processor side according to the reallocated quota. For high-entropy pressure nodes, the first polygon mesh is called for high-precision rendering, such as displaying compressor blade casing details, vibration hot spots, and local curves. For low-entropy pressure nodes, the second polygon mesh is called for low-precision rendering, such as displaying large-area stable pipe corridors as simplified geometry or low-resolution texture blocks. In this way, the system does not uniformly degrade the entire park, but concentrates limited graphics resources on areas where the data change rate is greater than the preset fluctuation threshold, thereby keeping the video memory usage and rendering load stably below the preset threshold. Furthermore, as an optional anomaly handling mechanism, when a node fails to receive a valid data stream within a certain time window, its information entropy value can be set to the most recent valid value or to a preset steady-state baseline value to avoid meaningless frequent upgrades and downgrades due to temporary network jitter. When the overall entropy value of all nodes rises and the recovered resources are insufficient to meet the needs of all high-entropy pressure nodes, the system prioritizes nodes with higher alarm levels or higher spatial importance, while maintaining suboptimal accuracy for the remaining nodes. When a low-entropy pressure node has been reduced to the minimum guaranteed memory quota, its resources will no longer be deprived to prevent the appearance of completely invisible critical devices in the scenario. For example, during nighttime operation in the refining and chemical park, most of the equipment in the pipe gallery is in a steady state, except for the reactor inlet pressure of the hydrocracking unit, which suddenly oscillates and the compressor bearing temperature rises synchronously. At this moment, the system calculates that the difference between the dynamic entropy pressure of the nodes where the reactor and compressor are located is greater than the preset comparison threshold and higher than the surrounding nodes. Therefore, it automatically recovers some of the video memory and rendering instructions of the heat exchanger group and the remote pipe gallery and puts them into the abnormal area. What the dispatcher sees on the large screen is: the abnormal device area model is displayed at the first resolution, the chart refresh rate is higher than the preset base frequency, while the stable area is automatically simplified; the purpose of this step is to directly convert the intensity of data change into the intensity of graphics resource allocation, so as to realize on-demand rendering of digital twin scenes and priority display of anomalies; Furthermore, to ensure consistency in the judgment criteria throughout the text, the residual matrix in this embodiment is uniformly defined according to the method of subtracting the dynamic entropy pressure matrix from the current graphics processor's memory allocation matrix, i.e., written as ;in, This represents the graphics processor memory allocation matrix at the current moment. This represents the dynamic entropy pressure matrix obtained by entropy value mapping at the same time, while This represents the residual matrix after subtracting the two; to ensure dimensional consistency in mathematical operations, the dynamic entropy pressure matrix involved in the subtraction is... All element values ​​in the matrix must be strictly converted to the allocation matrix beforehand using a memory mapping function. Same graphics processor memory resource unit level; Under this definition, a negative residual indicates resource shortage, corresponding to high-entropy pressure candidate nodes that need to be replenished; a positive residual indicates resource surplus, corresponding to low-entropy pressure candidate nodes that can be recycled; a residual close to zero indicates that the current quota is basically matched with demand; the aforementioned This is only intended to illustrate that different allocation results exist under different minimum retention quota strategies. In actual implementation, a redistribution method with a minimum retention quota is preferred, ensuring that low-entropy pressure nodes entering the second polygon mesh rendering still retain positive resource quotas and remain visible. For example, it is preferable to use... This type of result ensures that the description of low-entropy pressure nodes performing second polygon mesh rendering in this embodiment is consistent with the resource protection strategy.

[0021] Furthermore, the initialization module includes: a spatial resolution unit, used to resolve the spatial coordinate matrix of the digital twin 3D scene and divide the spatial coordinate matrix into a spatial mesh containing multiple levels; and a tensor tree construction unit, used to convert the spatial mesh into an octree-structured visual tensor tree and bind a preset device interface pointer as a data source address to the leaf node of the visual tensor tree. The leaf node is associated with a data visualization chart for displaying real-time time-series data streams, and each leaf node is allocated an initial graphics processor memory quota, which includes the upper limit of the number of vertices and the texture resolution level.

[0022] This embodiment provides a tensor initialization step for digital twin 3D scenes. Specifically, in the aforementioned refining park scene, if only a static model list and equipment coding table are used for binding, the correspondence between rendering objects and data sources will become loose when the number of devices increases to tens of thousands. Subsequently, it will be impossible to meet the preset computation time constraints to perform fine-grained memory reclamation and local redrawing. Therefore, this embodiment further introduces a hierarchical spatial mesh and an octree-style visual tensor tree, so that the scene, devices and resource quotas establish a one-to-one structured relationship during the initialization stage. Specifically, the spatial resolution unit reads the spatial coordinate matrix of the 3D scene; this matrix can be understood as the bounding box information of each device model in the scene, for example, the coordinate range of the reactor area is... to , to , to The coordinate range of the utility tunnel area is to , to , to The system first uses the entire surrounding space of the park as a first-level grid, and then further subdivides areas where the equipment density is greater than the preset density threshold or the business importance meets the preset level. To facilitate understanding, a simplified exemplary scenario can be constructed: the root grid. Covering the entire park, the secondary grid is divided into Tank farm, Reaction zone Power Zone and Utility tunnel area; will Further subdivided into reactor, compressor, heat exchangers and Valve assembly; thus, the spatial grid naturally forms a spatial organizational structure containing multiple levels; The tensor tree building unit transforms the spatial grid into an octree structure; each node not only stores its spatial boundaries but also binds a device interface pointer to the interface instance that collects data from that area; for example, The corresponding data source address can point to the memory handle of the reactor pressure acquisition service. It can point to the compressor vibration acquisition interface; in addition to binding the data source, the leaf node is also associated with one or more data visualization charts, such as trend curves, threshold bar charts or multi-parameter dashboards, so that the corresponding real-time monitoring results can be directly displayed when the local area is zoomed in. In the initial resource allocation, each leaf node is assigned two core parameters: a maximum number of vertices and a texture resolution level. For example, the reactor node is allocated a maximum of 120,000 vertices and a texture level of 2048; the compressor node is allocated a maximum of 100,000 vertices and a texture level of 1024; and the pipe gallery node, due to its regular geometric structure and lower attention rating than the preset baseline, can be allocated only a maximum of 30,000 vertices and a texture level of 512. In this way, even if resource reallocation does not occur later, the initial state is not an even distribution, but already has basic business differentiation. As an anomaly handling mechanism, if a device model lacks precise coordinate information, the system can degenerate by using its parent region's grid to accommodate it, and bind a data interface to that parent node. The model can then be split downwards after subsequent models are completed. If a leaf node corresponds to multiple sensor interfaces, the multiple interface pointers can be encapsulated into an interface array and attached to the same node for aggregation during subsequent entropy calculations. If a region exists in space but no sensors are deployed yet, the node can still retain geometric information and a minimum memory quota, with its data source address pointing to a null placeholder object to prevent tree structure breaks. For example, during the modeling and deployment phase of the refining and chemical industrial park, after engineers import the 3D model of the hydrocracking unit into the platform, the system first divides the space according to three levels: unit, cell, and equipment. Then, the reactor, compressor, heat exchanger, and valve group are attached as leaf nodes to the corresponding areas. Each leaf node is initialized by binding a data interface pointer, a trend chart component, and initial vertices and texture levels. In this way, when a subsequent reactor node malfunctions, the platform can directly locate its spatial position, data source, and graphic resource boundaries without having to perform a global search. The purpose of this step is to provide an addressable, reclaimable, and traceable underlying structure for subsequent entropy field calculations, resource stripping, and local rendering.

[0023] Furthermore, the entropy field calculation module includes: a data receiving unit, used to monitor the input / output throughput of the device interface corresponding to the device interface pointer, and collect real-time time-series data streams at a preset sampling frequency; an entropy value calculation unit, used to extract the probability distribution characteristics of the real-time time-series data streams within a sliding time window, and calculate the Shannon entropy value as the information entropy value based on the probability distribution characteristics; and a matrix generation unit, used to perform weighted calculations by combining the Shannon entropy value with preset node spatial importance weights or data source type weights to generate the dynamic entropy pressure of each node, and to generate a dynamic entropy pressure matrix by combining the dynamic entropy pressures of each node.

[0024] This embodiment provides an information entropy calculation step based on a sliding time window. Specifically, in the aforementioned scenario, if the determination of whether to increase rendering accuracy is based solely on the original numerical value, misjudgments may occur: for example, if the values ​​of certain temperature sensors are consistently greater than the upper limit of the preset normal range, and the variance is less than the preset variance threshold within a preset time period, their monitoring significance may not be higher than that of a vibration sensor with normal but drastic fluctuations. Therefore, this embodiment introduces the uncertainty of data changes as the core driving quantity, and uses the data flow entropy value to replace the single-point value as the basis for resource scheduling. Specifically, the data receiving unit first monitors the input / output throughput of each device interface and collects time-series data according to a preset sampling frequency. Taking compressor vibration data as an example, if the sampling frequency is 10 times per second, 100 sampled values ​​can be generated within a 10-second sliding window. To facilitate entropy calculation, these sampled values ​​can be discretized into several buckets according to numerical intervals. For example, if the probability of the vibration amplitude falling into intervals A, B, C, and D is 0.1, 0.2, 0.5, and 0.2 respectively, then the probability distribution dispersion within this window is greater than a preset dispersion threshold, and the entropy value is greater than a preset high entropy threshold. If almost all 100 sampled values ​​of another temperature node fall into the same interval, the probability distribution is close to... If the entropy value is less than the preset low entropy threshold, then its entropy value is less than the preset low entrop The entropy calculation unit calculates the Shannon entropy based on the above probability distribution, and then obtains the information entropy. For example, in order to perform quantitative deduction, the reactor pressure node can be set to have an entropy of 0.88, the compressor vibration node to have an entropy of 0.81, the heat exchanger temperature node to have an entropy of 0.26, and the remote lighting circuit current node to have an entropy of 0.12. The matrix generation unit does not directly use this entropy value for memory allocation, but continues to add the node spatial importance weight or data source type weight. For example, the spatial importance of the core process area is set to 1.2, and that of the auxiliary facility area is set to 0.8; the weight of pressure and combustible gas data source types is set to 1.3, and the weight of lighting current is set to 0.6; thus, the dynamic entropy pressure of the reactor pressure node can be expressed as... The compressor vibration node can be represented as The heat exchanger temperature node is approximately The number of lighting circuit nodes is approximately The dynamic entropy pressure of each node can be combined in the order of the tree structure to form a dynamic entropy pressure matrix. The node spatial importance here reflects the coupling between spatial location and business criticality, while the data source type weight reflects the difference in the impact of different sensors on safe production. The introduction of both can prevent the system from allocating resources only according to the fluctuation amplitude. For example, even if the lighting current of an edge area fluctuates for a duration less than a preset time threshold, its visual priority should not be higher than the pressure fluctuation of the core reactor. Furthermore, in order to enable the weighted dynamic entropy pressure in this embodiment to be used for both memory mapping and threshold determination, the matrix generation unit can perform unified dimension processing when outputting the dynamic entropy pressure. Specifically, any one of the following methods can be used: preset upper limit truncation, segmented mapping, or interval normalization, to organize the weighted dynamic entropy pressure into a preset determination interval before sending it to the subsequent module. For example, if the system sets the threshold judgment range to 0 to 1, and the maximum value of the original dynamic entropy pressure after weighting all nodes within a certain window is 1.37, then 1.37 can be mapped to 1.00, 1.07 to approximately 0.78, 0.23 to approximately 0.17, and 0.06 to approximately 0.04. Subsequent threshold comparisons, memory mapping, and node sorting will all be based on this unified scale. This avoids inconsistencies in judgment criteria caused by using amplified dynamic entropy pressure in the previous stage and unamplified entropy values ​​in the subsequent stage. Furthermore, the extraction of probability distribution features can employ fixed-bucket-width discretization, quantile-based discretization, or business threshold-based discretization. If a certain type of data has a clear process range, such as pressure alarm upper and lower limits, vibration intensity levels, or gas concentration safety zones, then the business threshold-based discretization method is preferred to make the entropy value change more closely reflect the actual production semantics. If a stable business threshold is lacking, then fixed-bucket-width or quantile-based discretization is used. In this way, the source path of entropy value, the weight superposition path, and the dimensionless path form a complete closed loop, which facilitates stable execution of subsequent resource scheduling. Additional notes: When the number of samples within a window is less than the preset threshold, the system can extend the window length or reuse the probability distribution of the previous window to prevent entropy distortion caused by a small number of samples; when a node accesses multiple modal data at the same time, the entropy value of each modality can be calculated separately and then a weighted sum or the maximum value can be taken to adapt to different business preferences; when throughput monitoring finds that the interface congestion rate is greater than the preset congestion threshold, the system can appropriately reduce the sampling frequency of low-importance nodes and prioritize the sampling integrity of high-importance nodes. For example, in a hydrocracking unit, the reactor inlet pressure and compressor vibration both exhibit high-frequency fluctuations during the night shift, while the heat exchanger outlet temperature remains relatively stable, and the lighting current shows only sporadic fluctuations. The platform statistically analyzes their distribution characteristics within the same sliding window and generates a dynamic entropy-pressure matrix by combining the importance of the process zone and the weight of the sensor type. Thus, although individual edge nodes may also experience numerical fluctuations, abnormal process zones will still receive higher priority in resource allocation. The purpose of this step is to uniformly convert multi-source heterogeneous real-time data into comparable and quantifiable dynamic pressure, thereby providing direct input for subsequent memory redistribution.

[0025] Furthermore, the resource pressure differential reallocation module includes: a residual comparison unit, used to extract the current graphics processor memory allocation matrix and perform subtraction processing between the current graphics processor memory allocation matrix and the dynamic entropy pressure matrix to obtain a residual matrix; a virtual pressure transmission unit, used to determine nodes with corresponding element values ​​greater than zero in the residual matrix as high residual nodes and nodes with corresponding element values ​​less than zero as low residual nodes, and to establish a virtual pressure transmission tree based on the residual matrix, and to determine the path for resource migration from low entropy pressure nodes to high entropy pressure nodes based on the pressure gradient direction from high residual nodes to low residual nodes in the virtual pressure transmission tree as the gravity direction; and a memory remapping unit, used to strip the graphics processor memory address blocks of low entropy pressure nodes and remap the pointers of the released graphics processor memory address blocks to high entropy pressure nodes to generate reallocated memory quotas.

[0026] This embodiment provides a memory pressure redistribution step based on residual gradients. Specifically, after the entropy field calculation is completed, if the system simply adds quotas to all high-entropy nodes, the probability of the total memory exceeding the upper limit will be higher than the preset probability threshold. If a global replacement is performed after simply sorting by entropy value, the spatial proximity in the tree structure may be ignored, leading to frequent cross-regional resource relocation and increased remapping overhead. Therefore, this embodiment uses a residual matrix and a virtual pressure transmission tree to achieve controllable and traceable resource migration. Specifically, the residual comparison unit first extracts the current memory allocation matrix and the dynamic entropy compression matrix and then calculates the difference; assuming that at a certain moment, the current memory allocation matrix of the six nodes is written as:

[0027] The dynamic entropy pressure matrix is ​​written as:

[0028] The residual matrix obtained by subtracting the two is:

[0029] In this example, the first two nodes represent resource scarcity, and the last three nodes represent relatively abundant resources; the virtual pressure transmission unit then embeds these nodes into the parent-child relationships of the tensor tree; for example, and Belonging to the same reaction region subtree, , It belongs to the heat exchange zone subtree. , It belongs to the remote pipe gallery subtree; the system constructs the pressure transmission direction along the tree structure, so that resources are preferentially migrated between spatially adjacent or same-level nodes to reduce memory fragmentation and batch update resource consumption; To facilitate understanding, a simplified deduction can be used: The residual is 2. The residual is 1. The residual is 1. and Each is missing 2 resource units; the system first starts with and , Belonging to the same process area Recover 2 resource units, then retrieve from a location within a preset distance threshold. Recover 1 resource unit, according to the preset resource replenishment rules, from a remote location. One resource unit is recycled; the resulting migration order is not arbitrary, but constrained by the direction of gravity in the virtual pressure transmission tree. The so-called direction of gravity essentially represents the priority path for resource migration from the surplus area to the shortage area; after the released video memory address blocks are reclaimed, the video memory remapping unit re-points the pointers of these address blocks to the vertex buffer objects and texture objects corresponding to the high entropy pressure nodes, realizing the switching of resource owners, without having to request new video memory from the operating system or graphics driver layer every time. Furthermore, to avoid ambiguity in the description of residual values ​​and node categories, this embodiment adopts the following determination criteria: when the node residual is positive, it means that the current allocation is greater than the dynamic entropy pressure requirement corresponding to the node, and the node is included in the low entropy pressure candidate set of recyclable resources; when the node residual is negative, it means that the current allocation is less than the dynamic entropy pressure requirement corresponding to the node, and the node is included in the high entropy pressure candidate set of resources to be supplemented. When the node residual is equal to zero or the absolute value of the residual is less than the preset differential threshold, the node is considered not to migrate and only the existing resource state is retained. Therefore, the direction from high residual node to low residual node in the embodiment can be specifically implemented as: establishing a resource transmission direction from positive residual node with higher value and abundant resources to negative residual node with lower value and scarce resources. Furthermore, memory remapping does not strip all resources from low-entropy nodes at once, but rather proceeds step by step according to a preset minimum migration granularity. This minimum migration granularity can correspond to a memory address block, a texture page, or a set of batch instruction quotas. Using the six-node example above, if the minimum migration granularity is set to one resource unit, the system will complete one round of migration from... , Towards , After migration, the updated local residuals will be recalculated; if a node has met the requirements or its absolute value of the difference is less than the preset micro-difference threshold, then migration will not continue; this can avoid the target node occupying too many resources in a short time due to one-time overshoot, and can also avoid the source node being over-reclaimed. This memory address block pointer remapping differs from ordinary parameter rewriting. Its advantage lies in preserving the consistency of underlying address management. Especially when multiple nodes in a single frame are simultaneously upgrading or downgrading, direct remapping is often more stable than destruction and reconstruction, and is also more suitable for batch processing. As an exception handling mechanism, when a low-entropy pressure node has insufficient reclaimable resources, the system only reclaims the portion above its minimum retention quota; when multiple high-entropy pressure nodes compete for the same batch of reclaimable resources, they can be allocated according to alarm level, spatial importance, or time sequence; when the size of the video memory address block does not perfectly match the requirements of the target node, adjacent address blocks can be concatenated or allocated according to the closest principle. If this still cannot be satisfied, the system will fall back to the medium precision model instead of directly jumping to the highest precision model to avoid rendering jitter. For example, in a refining and chemical park, the reactor area and compressor area simultaneously experience anomalies within the same minute, while the heat exchange area and the remote pipe gallery remain in a steady state. The platform first identifies the shortage area and the surplus area based on the residual matrix, and then, based on the spatial adjacency relationship in the tensor tree, prioritizes recovering memory from the heat exchange area within the same unit to replenish the reactor and compressor. Only when resources within the same unit are insufficient will resources continue to migrate to the remote pipe gallery. In this way, the resource flow is consistent with the process space distribution; the purpose of this step is to transform the memory reclamation and remapping process from global preemption to directional local migration through residual and pressure transmission relationship, so as to reduce the computational overhead of reallocation and improve rendering continuity. Furthermore, to maintain consistency with the terminology used in the aforementioned embodiments, the terms "high residual node" and "low residual node" in this document are used only to describe the magnitude of the residual values ​​and their propagation direction, and are not directly equivalent to the names of "high entropy pressure node" and "low entropy pressure node" themselves. More specifically, the larger the positive residual value, the higher its resource surplus, and the more preferential it is as the starting point for resource outflow; the smaller the negative residual value, the higher its resource shortage, and the more preferential it is as the endpoint for resource inflow. Therefore, within the same decision period, it is preferable to first determine the resource surplus side and the resource shortage side based on the residual matrix, and then classify them into low-entropy pressure candidate nodes and high-entropy pressure candidate nodes respectively based on the results, so as to ensure that the meaning of the residual gradient terminology and the node category terminology in the whole text are in one-to-one correspondence.

[0030] Furthermore, the adaptive rendering execution module includes: an instruction compilation unit for compiling reallocated video memory quotas into batch rendering instructions at the graphics processor level; and a state update unit for executing the batch rendering instructions to update vertex buffer objects and texture objects on the graphics processor side.

[0031] This embodiment provides a rendering execution step that sinks the redistribution result down to the graphics processor execution layer. Specifically, after the aforementioned redistribution stage is completed, if the upper-layer application still submits rendering commands node by node, a large number of discrete calls will be generated, leading to an increase in instruction interaction overhead between the processor and the graphics processor, thereby weakening the technical effect brought by redistribution. Therefore, this embodiment compiles the quota change into the underlying batch rendering instructions, and then uniformly updates the corresponding vertex buffer objects and texture objects. Specifically, the instruction compilation unit first reads the reallocated memory quota of each node and converts it into resource update batches that can be executed by the graphics processor. For example, if a reactor node needs to be upgraded from medium precision to high precision, its batch processing may include: expanding the vertex buffer object capacity, loading a higher resolution texture, enabling additional normal maps, and refreshing local graph textures; if a gallery node needs to be downgraded, its batch processing may include: switching to a simplified mesh, unloading a high-resolution texture, and replacing it with a low-resolution texture placeholder. For ease of explanation, we can assume there are four update tasks within the same frame: To improve the accuracy of the reactor, To improve the precision of the compressor, To reduce the precision of the heat exchanger, To reduce the accuracy of the utility tunnel; the system will... , Merge into high-priority batches ,Will , Merge into low priority batches The data is sent to the graphics processor in a preset order. After the state update unit executes the batch processing instructions, it actually replaces or rebinds the vertex buffer objects and texture objects on the graphics processor side. The update here is not only a change in visual appearance, but also includes the synchronization of underlying resource reference relationships. For example, when a reactor node is allocated to a new memory address block, its vertex buffer object's index pointer and texture sampling pointer must be rewritten synchronously to ensure that subsequent drawing calls can correctly read data from the new address. After the update is completed, the graphics processor can start drawing the current frame or the next frame according to the new resource state. Additional notes: When individual resource updates fail in a batch, such as when high-resolution textures have not yet been uploaded, the node can temporarily maintain the intermediate state of the previous frame to avoid rendering interruption; when the batch size exceeds the acceptable limit for a single frame, the system can postpone low-priority downgraded tasks to the next frame to ensure the high-entropy compression node's precision update first; when address contention occurs during the switching between old and new vertex buffer objects, the new buffer object can be read-only attached first, and then the old buffer object can be released to ensure that no null references occur within a single frame; For example, on the large screen of the refining and chemical park dispatch, after the reactor and compressor are identified as abnormal core areas, the system merges their precision-up tasks into a high-priority batch, while merging the precision-down tasks of the stable area into another batch; the graphics processor first updates the high-precision mesh and texture of the abnormal area, and then releases the excess resources of the stable area; the result seen by the dispatcher on the screen is that the details of the abnormal equipment switch to the set resolution of the first polygon mesh rendering within a few frames, while the background area switches to the second polygon mesh rendering resolution through a preset transition algorithm, without the whole screen flickering; The purpose of this step is to stably and with low computational overhead implement the quota redistribution results from the upper layer to the graphics hardware execution plane, thereby achieving consistency between resource scheduling and actual rendering output.

[0032] Furthermore, the state update unit is also used to: determine that a node is a high-entropy-pressure node when the dynamic entropy pressure of a node is greater than or equal to a preset entropy-pressure threshold, and load the preset first level of detail model data into the vertex buffer object for first polygon mesh rendering; and determine that a node is a low-entropy-pressure node when the dynamic entropy pressure of a node is less than a preset entropy-pressure threshold, and forcibly replace the model data corresponding to the node with the preset second level of detail proxy model data for second polygon mesh rendering.

[0033] This embodiment provides a model detail level switching step based on a dynamic entropy pressure threshold. Specifically, in the aforementioned rendering process, if vertex and texture resources are only fine-tuned based on quota values ​​without setting clear hierarchical switching rules, some nodes will repeatedly be in a semi-upgraded or semi-downgraded state between adjacent frames, resulting in screen jitter and frequent resource switching. Therefore, this embodiment further introduces an entropy pressure threshold to determine whether a node has high or low entropy pressure, and calls the first detail level model and the second detail level proxy model respectively. Specifically, an entropy pressure threshold can be preset. ,For example If the dynamic entropy pressure of a node is greater than or equal to the threshold, then the node is classified as a high entropy pressure node and the first level of detail model data is loaded for it. The model can include the complete device appearance, subdivided components, fine textures, and local state animations; if the dynamic entropy pressure of a node is lower than the threshold, the node is classified as a low entropy pressure node and forcibly replaced with the second level of detail proxy model data; the proxy model can be a low-polygon simplified mesh, a bounding box model, a wireframe, or a texture replacement model. To illustrate this with an example scenario, suppose there are currently three nodes. , , Their dynamic entropy pressures are 0.82, 0.74, and 0.31, respectively; then The node is identified as a high-entropy pressure node, and the first level of detail model is loaded. and Nodes identified as low-entropy pressure nodes, among which Although it's close to the threshold, it's still processed with low entropy compression, loading the second level of detail proxy model; if in the next sliding window... If the value is increased to 0.78, it will switch to the first level of detail model during the next state update; through this explicit binary branch, the graphics pipeline can more stably select specific model resources. In practice, a forced replacement method is used. That is, once a node falls into the low entropy pressure range, the original high-detail model is no longer retained. Instead, a preset proxy model is directly used to replace it, thereby releasing the corresponding vertex and texture resources and allocating the corresponding graphics processor memory address block to the high-entropy pressure node. Unlike the traditional progressive model reduction based on camera distance, the switching basis in this embodiment is the data semantic strength. Furthermore, in this embodiment, the threshold comparison is preferably performed on the dynamic entropy pressure after uniform dimensional processing; that is, if the matrix generation unit has already performed interval normalization, segmented mapping or upper limit truncation on the weighted dynamic entropy pressure, the state update unit directly performs threshold comparison on the result under the uniform scale. If the output of the previous stage is still an amplified dimensional value, the state update unit first converts it to a decision range consistent with the threshold, and then decides whether to call the first level of detail model or the second level of detail proxy model. This ensures that the decision criteria of the high-entropy pressure node / low-entropy pressure node are consistent with the calculation criteria of the previous stage entropy field, avoiding the misunderstanding caused by the previous stage dynamic entropy pressure value being greater than 1 while the subsequent stage threshold is fixed in the range of 0 to 1. Furthermore, model switching is preferably not triggered directly by a single instantaneous window result, but can be combined with a continuous window confirmation strategy; for example, when a node satisfies the entry into high entropy pressure condition for two consecutive sliding windows, it officially switches to the first level of detail; when it satisfies the exit from high entropy pressure condition for two consecutive sliding windows, it falls back to the second level of detail. For boundary nodes approaching the threshold, the decision to switch immediately can be made by combining the batch queue congestion level, remaining memory, and whether the node is in an alarm state. This preserves the determinism of threshold determination while reducing frequent upgrades and downgrades caused by short-term noise fluctuations. As an anomaly handling mechanism, to prevent nodes from repeatedly jumping around the threshold, the entry threshold and exit threshold can be set differently. For example, the threshold for entering high entropy pressure can be 0.78, and the threshold for exiting high entropy pressure can be 0.72, forming a hysteresis zone. When the proxy model is not yet ready, the previous low-level model can be temporarily maintained to avoid blank rendering. When a node is below the entropy pressure threshold but is a critical device that must be visible for a long time according to legal security requirements, the minimum level of detail retention can be set instead of completely reducing to the simplest proxy model. For example, during the failure propagation process of a hydrocracking unit, the dynamic entropy pressure of the reactor node reaches 0.91, the compressor node reaches 0.84, and the remote pipe gallery node is only 0.18. Based on this, the platform switches the reactor and compressor to the first level of detail, displaying the surface details of the equipment, key components, and local trend maps; and switches the pipe gallery to the second level of detail proxy model, retaining only the outline and basic texture. In this way, the abnormal area in the center of the screen will be clearly highlighted, while the background steady-state area will be automatically simplified. The purpose of this step is to transform the continuously changing entropy pressure into a stable and executable model switching instruction through thresholding, thereby improving the determinism of rendering decisions. Furthermore, to maintain consistency with the resource pressure redistribution process, the determination of high-entropy pressure nodes and low-entropy pressure nodes in this embodiment is used for model detail level selection, and its input is preferably dynamic entropy pressure after uniform dimension processing; while whether resources need to be recycled or replenished, and how to establish the recycling path, are still determined by the residual matrix and residual gradient; in other words, the dynamic entropy pressure threshold mainly answers which detail level should be used for rendering of the node, and the residual matrix mainly answers whether the resources of the node should be increased or decreased and from where to migrate. In actual execution, the state update unit can first receive the resource update result corresponding to the reallocated memory quota, and then decide whether to load the first level of detail model or the second level of detail proxy model based on the dynamic entropy pressure threshold. When a node has a negative residual but the dynamic entropy pressure has not yet exceeded the upgrade threshold, it can be given priority to replenish basic resources but not upgraded to the first level of detail until its dynamic entropy pressure meets the threshold condition before the model switch is completed. In this way, the resource reallocation decision and the model switch decision can cooperate with each other in their uses without conflicting with each other.

[0034] Furthermore, the system also includes: a thread scheduling module, used to allocate the highest priority CPU computing thread to the data visualization chart bound to the high-entropy pressure node when the node is a high-entropy pressure node; and a baseline update module, used to extract the rendering performance data output by the adaptive rendering execution module and feed the rendering performance data back to the initial graphics processor memory quota to update the preset historical memory allocation baseline.

[0035] This embodiment provides a thread-priority scheduling and historical baseline update mechanism for high-entropy pressure nodes. Specifically, in the aforementioned scheme, although memory and model accuracy have been tilted towards abnormal regions, if the graph calculation thread still shares the same priority as ordinary nodes, it is possible that the 3D model is clearly displayed, but the corresponding trend chart and fluctuation spectrum chart are updated lagging behind. In addition, if the initial memory quota remains unchanged for a long time, and the system reallocates from the same basic state each time, it is difficult to gradually adapt to real working conditions. Therefore, this embodiment further adds thread scheduling and baseline write-back. Specifically, after identifying a high-entropy pressure node, the thread scheduling module will assign the data visualization charts bound to that node to the highest priority CPU computing thread. Taking compressor vibration monitoring as an example, the charts bound to it may include time-domain waveforms, spectrum graphs, and alarm timelines. Since the computational load of the spectrum graph exceeds the preset computational load threshold, if it continues to queue in the ordinary thread pool, it may cause an anomaly to occur before the chart is refreshed. By binding to a priority thread, the chart computation task of this node can be executed before that of the low-entropy pressure node. A simplified deduction is as follows: Suppose the thread pool contains threads , , Reactor chart task Compressor chart task This belongs to the high-entropy pressure task and the heat exchanger diagram task. And the task of drawing diagrams If it is a normal task, the system will prioritize... and Assigned to , ,Will , They are queued in a regular thread pool to await execution; this allows the visualization charts of abnormal nodes to keep pace with the 3D rendering. The baseline update module collects performance data after each rendering round, such as single-frame rendering time, peak memory usage, number of batch processing instructions, and number of model switching, and feeds these results back into the initial memory quota to form a new historical allocation baseline. For example, if the system repeatedly finds that the reactor area always needs to borrow resources from neighboring areas in abnormal scenes, its initial vertex limit can be directly increased in the next running cycle. If the remote pipe gallery maintains low entropy pressure for a long time and the degradation has no impact on the overall monitoring, its initial texture level can be appropriately reduced. In this way, the initial quota is no longer a static constant, but a historical baseline with learning and correction capabilities. As an exception handling mechanism, when the highest priority thread is fully loaded, the system can split the high-entropy pressure graph task into a base graph and an extended graph, prioritizing the calculation of the base graph and deferring the calculation of the extended graph. When the rendering performance data deviates abnormally from the norm, for example, due to driver upgrades or temporary network failures, the baseline update module can avoid writing occasional anomalies into the baseline by using a moving average or removing outliers. When a node does not have enough historical samples recently, it will maintain its original initial quota and update it only when the sample size is sufficient. For example, during a nighttime fluctuation event in the refining park, after the reactor and compressor were identified as high-entropy pressure nodes, their trend charts and spectrum charts were updated immediately by the highest priority thread, while the ordinary statistical charts of the heat exchanger and pipe gallery were updated later. After the event ended, the platform recorded that the reactor area needed to increase the texture level and vertex limit three times in a row under high load, so it automatically raised the historical baseline of the area during the next round of initialization. The purpose of this mechanism is to make abnormal areas not only more clearly rendered, but also to update the charts faster, and to gradually make the initial resource allocation closer to the actual operating rules through performance feedback.

[0036] Furthermore, the adaptive rendering execution module also includes: a visual focusing unit, used to reduce the texture resolution level corresponding to low-entropy pressure nodes and apply a preset blur shader when rendering the second polygonal mesh, and to increase the texture resolution level corresponding to high-entropy pressure nodes and apply a preset highlight shader when rendering the first polygonal mesh.

[0037] This embodiment provides a visual focusing step that works in conjunction with resource reallocation. Specifically, when only model accuracy is upgraded or downgraded, although the internal resources of the system have been tilted, the operator may not be able to immediately perceive the focus on the display terminal. Especially in complex industrial scenarios, a large number of devices have similar geometric outlines, and the difference in the number of polygons alone is not enough to form a significant attention guide. Therefore, this embodiment further superimposes texture resolution adjustment and shader differentiation processing on the rendering layer. Specifically, when the visual focus unit uses the second polygon mesh for rendering at low-entropy pressure nodes, it not only reduces the number of vertices, but also simultaneously reduces the texture resolution level and applies a preset blur shader. For example, the tube gallery node that originally used a 1024 resolution texture can be reduced to a 256 resolution and a light Gaussian blur or edge softening effect can be applied to reduce its visual saliency weight to below the preset saliency threshold. Conversely, when rendering high-entropy pressure nodes using the first polygon mesh, the system increases the texture resolution level and applies a preset highlight shader, such as increasing local brightness, enhancing edge contrast, overlaying alarm tones, or enhancing normal reflection details; a microscopic deduction is as follows: Let the node... This is a high-entropy compression node, its texture level is increased from 512 to 2048, and a highlight shader is enabled; node As a low-entropy compression node, its texture level is reduced from 1024 to 256, and a blur shader is enabled; the result is that, at the same viewpoint, The surface nameplate, connecting flange, and localized hot spots are clearer, while This results in a low-contrast, low-detail background effect; This processing is not simply a visual effect, but is coupled with a resource deprivation mechanism: the reason why downgraded nodes are blurred is because their texture resources and shading overhead are actively reduced; the reason why enhanced nodes are highlighted is because they have obtained an additional quota and loaded higher-quality rendering resources; thus, the resource scheduling result is directly translated into a change in the focus of the visual layer. As an anomaly handling mechanism, when the spatial distance between multiple high-entropy pressure nodes is less than the preset minimum spacing threshold, the upper limit of the highlight intensity can be limited to avoid overexposed areas exceeding the preset area threshold; when low-entropy pressure nodes belong to the environmental background but need to maintain basic recognizability, the blur shader only reduces high-frequency details without completely erasing the outline; when the operator manually locks a certain area as the object of interest, the system can apply minimum texture level and minimum sharpness constraints to that area to prevent it from being excessively blurred due to short-term low-entropy pressure. For example, in the fault handling interface of the refining and chemical park, abnormal equipment near the compressor is upgraded to high texture resolution and has a bright edge effect, while the remote stable pipe gallery and auxiliary facilities are downgraded to low texture resolution and have a blurred color applied. Even in a panoramic view, the dispatcher can intuitively feel that the visual center of gravity is concentrated on the abnormal device. The purpose of this step is to convert the background resource tilt into a visually focused result that can be perceived by the front end, thereby improving the identification efficiency of abnormal areas.

[0038] Furthermore, when constructing a visual tensor tree, the tensor tree construction unit stores the hierarchical relationship of each node as a pointer structure in memory, and associates it with the corresponding graphics processor memory address block through the pointer structure.

[0039] This embodiment provides a pointer association structure between visual tensor tree nodes and video memory address blocks. Specifically, after the aforementioned octree is constructed, if the tree node relationship is only stored in the logical number table, while the video memory address blocks are stored separately in another resource table, then frequent number lookups and cross-table positioning are required when performing high-frequency video memory remapping, which can easily lead to performance bottlenecks. Therefore, this embodiment unifies the node hierarchy relationship and the video memory address block association relationship into a memory pointer structure. Specifically, each tree node retains at least a parent node pointer, a set of child node pointers, and a resource pointer in memory; the resource pointer further points to the graphics processor memory address block description structure currently occupied by the node; for ease of explanation, we can assume that the node... This represents the compressor region, whose parent node is the device region node. The child nodes include bearing nodes. and casing nodes resource pointer Point to video memory address block to ; Thus, when When a node is identified as a high-entropy pressure node and needs expansion, the system can follow... , , , The pointer relationships between nodes allow for quick location of reclaimable resources in adjacent nodes, and the new memory address block pointer is directly attached to them. Instead of having to re-traverse the entire scene resource table; This structure is particularly effective in implementing memory remapping; for example, if the heat exchanger node is stabilized. resource pointer Originally pointed to the address block to Furthermore, this portion of the address block is reclaimed and transferred to the compressor node. Then the system can Update to point to both to and to Or, after splicing, point to a new contiguous description block; at the same time, The address block is rewritten to its backup address; the whole process is mainly manifested as pointer update rather than object reconstruction; thus, tree structure traversal, resource reclamation and rendering object update can be completed within the same addressing system. As an exception handling mechanism, when a node's child nodes have not yet initialized their memory pointers, they can temporarily inherit the shared resource pointers of their parent nodes and then split them after local subdivision. When a node is deleted or a model is taken offline, the association between its resource pointers and memory address blocks should be removed first, and then the node's memory should be reclaimed to avoid the formation of dangling pointers. When there are not enough contiguous memory address blocks, the resource pointers can point to a chained description structure to record multiple scattered address segments. For example, in a refining park, if a compressor node needs to improve rendering accuracy due to abnormal vibration, the system searches down the parent node of the unit area to find the resource pointer of the adjacent heat exchanger node and finds that there is reclaimable video memory. Therefore, the reclaimed address block is directly attached to the resource pointer of the compressor node. Since the hierarchical relationship and resource relationship between nodes have been fixed in memory through the pointer structure, this process can be completed within the time period that meets the preset real-time response time constraint. The purpose of this step is to provide fast addressing and low computational cost remapping capabilities for high-frequency resource scheduling, thereby supporting digital twin rendering scenarios with high real-time requirements.

[0040] Furthermore, the real-time time-series data stream is a multimodal device status data stream collected by an IoT sensor network deployed in the physical space.

[0041] This embodiment provides a data access method suitable for multimodal IoT sensor networks. Specifically, in industrial digital twin scenarios, if only a single type of sensor data is accessed, entropy field calculation and resource scheduling can only reflect single-dimensional changes, which may easily overlook the complex coupling characteristics of device status. Therefore, in this embodiment, the accessed data stream is preferably a multimodal device status data stream collected by an IoT sensor network deployed in physical space. Specifically, multimodal data can include at least two of the following: pressure, temperature, vibration, flow rate, current, valve position, gas concentration, and video analysis results. In the context of a refining and chemical industrial park, reactors can simultaneously access pressure and temperature data, compressors can simultaneously access vibration, current, and bearing temperature data, and pipe racks can access combustible gas concentration and video smoke recognition results. To illustrate the scenario, assume that the compressor node receives three modalities of data within a 10-second window, with corresponding entropy values ​​of vibration (0.82), current (0.35), and bearing temperature (0.76). The system can aggregate these data according to preset weights, for example, vibration weight 0.5, current weight 0.2, and temperature weight 0.3. The aggregated node information entropy can then be written as: If spatial importance is added, the final dynamic entropy pressure of the node can be obtained. It can be seen that multimodal input can more comprehensively reflect the real device state, rather than relying on the occasional fluctuations of a single sensor. Various data interfaces in the IoT sensor network can be connected to the platform via industrial Ethernet, wireless acquisition gateways, or edge computing nodes. The platform does not require all modes to have the exact same sampling frequency, but allows high-frequency vibration data and low-frequency temperature data to be aligned within a unified sliding window. The alignment method can be resampling, interpolation, or taking values ​​based on the most recent timestamp to form a unified window statistical basis. As an anomaly handling mechanism, when a certain modality's data is temporarily missing, the system can calculate the aggregate entropy value using only the remaining modalities and reduce the reliability coefficient of the result according to the number of missing modalities; when a certain modality's data quality is abnormal, such as continuously outputting fixed values ​​or values ​​beyond the range, it can be marked as a distorted modality and removed from the current window to avoid polluting the overall entropy pressure; when time-consuming modalities such as video analysis cannot be updated in every window, they can be refreshed at a lower frequency and the previous result can be reused between two windows; For example, in the early stage of a compressor failure in a refining and chemical park, the vibration data showed high-frequency jitter, the bearing temperature rose slowly, and the current curve also showed a slight deviation. The platform simultaneously acquired these three types of data from the Internet of Things sensor network and aggregated them to form a node dynamic entropy pressure that is more reliable than a single vibration entropy value. As a result, the node will receive additional rendering resources and higher visualization priority earlier. The purpose of this step is to improve the realism and robustness of entropy field calculations through multimodal device state data streams, thereby enhancing the responsiveness of digital twin systems to complex industrial anomalies.

[0042] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A data visualization management system based on digital twinning, characterized by, The system includes: an initialization module, used to divide the digital twin 3D scene into a spatial grid and construct a visual tensor tree mapped to computer memory, wherein the nodes of the visual tensor tree are bound to a data source address, a vertex buffer object and a texture object, and are allocated with a graphics processor memory address block, an initial graphics processor memory quota and a rendering instruction quota, and the initial graphics processor memory quota of each node is combined to form a current graphics processor memory allocation matrix; The entropy field calculation module is used to receive real-time time-series data streams from the data source address, calculate the information entropy value of the real-time time-series data streams corresponding to each node using a sliding time window at the central processing unit, and calculate the dynamic entropy pressure corresponding to each node based on the information entropy value through a preset memory mapping function, and generate a dynamic entropy pressure matrix with the same dimension as the current graphics processor memory allocation matrix by combining the dynamic entropy pressures of each node. The resource pressure differential reallocation module is used to calculate the residual matrix between the current graphics processor memory allocation matrix and the dynamic entropy pressure matrix. If the corresponding element value in the residual matrix is ​​greater than zero, the node is determined to be a low-entropy pressure node. If the corresponding element value in the residual matrix is ​​equal to zero, the node is determined to be a stable node and its current graphics processor memory allocation matrix and rendering instruction quota remain unchanged. If the corresponding element value in the residual matrix is ​​less than zero, the node is determined to be a high-entropy pressure node, and the graphics processor memory address block and rendering instruction quota of the low-entropy pressure node are reclaimed. The reclaimed graphics processor memory address block and rendering instruction quota are then redistributed to the high-entropy pressure node to generate a reallocated memory quota. An adaptive rendering execution module is used to update the vertex buffer object and the texture object of the digital twin 3D scene on the graphics processor side according to the reallocated video memory quota. In the case that the node is the high-entropy pressure node, a first polygon mesh rendering is performed, and in the case that the node is the low-entropy pressure node, a second polygon mesh rendering is performed, so as to keep the video memory utilization rate and / or rendering computing load of the graphics processor below a preset load threshold.

2. The data visualization management system based on digital twinning according to claim 1, wherein, The initialization module includes: a spatial resolution unit, used to resolve the spatial coordinate matrix of the digital twin 3D scene and divide the spatial grid containing multiple levels based on the spatial coordinate matrix; Tensor tree construction unit is used to convert the spatial grid into the visual tensor tree of octree structure, and bind a preset device interface pointer as the data source address to the leaf node of the visual tensor tree. The leaf node is associated with a data visualization chart for displaying the real-time time-series data stream, and an initial graphics processor memory quota is allocated to each leaf node. The initial graphics processor memory quota includes the upper limit of the number of vertices allocated and the texture resolution level.

3. The data visualization management system based on digital twinning of claim 2, wherein, The entropy field calculation module includes: a data receiving unit, used to monitor the input and output throughput of the device interface corresponding to the device interface pointer, and to collect the real-time time-series data stream at a preset sampling frequency; The entropy calculation unit is used to extract the probability distribution characteristics of the real-time time series data stream within the sliding time window, and calculate the Shannon entropy value as the information entropy value based on the probability distribution characteristics; the matrix generation unit is used to combine the Shannon entropy value with preset node spatial importance weights or data source type weights for weighted calculation to generate the dynamic entropy pressure of each node, and generate the dynamic entropy pressure matrix by combining the dynamic entropy pressures of each node.

4. The data visualization management system based on digital twinning of claim 3, wherein, The resource pressure differential redistribution module includes: a residual comparison unit, used to extract the current graphics processor memory allocation matrix and perform a difference operation between the current graphics processor memory allocation matrix and the dynamic entropy pressure matrix to obtain the residual matrix; The virtual pressure transmission unit is used to determine nodes with corresponding element values ​​greater than zero in the residual matrix as high residual nodes and nodes with corresponding element values ​​less than zero as low residual nodes. It also establishes a virtual pressure transmission tree by combining the hierarchical structure of the visual tensor tree with the residual matrix, and determines the path for resources to migrate from the low-entropy pressure node to the high-entropy pressure node based on the pressure gradient direction from the high residual node to the low residual node in the virtual pressure transmission tree as the gravity direction. The memory remapping unit is used to preempt the graphics processor memory address block of the low-entropy pressure node and remap the pointer of the released graphics processor memory address block to the high-entropy pressure node to generate the reallocated memory quota.

5. The data visualization management system based on digital twinning of claim 1, wherein, The adaptive rendering execution module includes: an instruction compilation unit for compiling the reallocated video memory quota into batch rendering instructions at the graphics processor level; and a state update unit for executing the batch rendering instructions to update the vertex buffer object and the texture object at the graphics processor.

6. A data visualization management system based on digital twins according to claim 5, characterized in that, The state update unit is further configured to: determine that the node is a high-entropy-pressure node when the dynamic entropy pressure of the node is greater than or equal to a preset entropy-pressure threshold, and load the preset first level of detail model data into the vertex buffer object for the first polygon mesh rendering; and determine that the node is a low-entropy-pressure node when the dynamic entropy pressure of the node is less than the preset entropy-pressure threshold, and forcibly replace the model data corresponding to the node with the preset second level of detail proxy model data for the second polygon mesh rendering.

7. A data visualization management system based on digital twins according to claim 2, characterized in that, The system further includes a thread scheduling module, used to allocate the highest priority central processing unit computing thread to the data visualization chart bound to the high entropy pressure node when the node is the high entropy pressure node; The baseline update module is used to extract the rendering performance data output by the adaptive rendering execution module and feed the rendering performance data back to the initial graphics processor memory quota in order to update the preset historical memory allocation baseline.

8. A data visualization management system based on digital twins according to claim 2, characterized in that, The adaptive rendering execution module further includes: a visual focusing unit, used to reduce the texture resolution level corresponding to the low-entropy pressure node and apply a preset blur shader when rendering the second polygonal mesh, and to increase the texture resolution level corresponding to the high-entropy pressure node and apply a preset highlight shader when rendering the first polygonal mesh.

9. A data visualization management system based on digital twins according to claim 2, characterized in that, When constructing the visual tensor tree, the tensor tree construction unit stores the hierarchical relationship of each node as a pointer structure in memory, and associates it with the corresponding graphics processor memory address block through the pointer structure.

10. A data visualization management system based on digital twins according to claim 1, characterized in that, The real-time time-series data stream is a multimodal device status data stream collected by an Internet of Things (IoT) sensor network deployed in physical space.

Images