Intention-driven cloud-edge-end heterogeneous computing power collaborative scheduling system and method

Abstract

The application relates to the technical field of computing power cooperative scheduling management, and specifically discloses a cloud-edge-end heterogeneous computing power cooperative scheduling system and method based on intention driving, which obtains intention semantic vectors, video stream data, buffer occupancy rates and heterogeneous computing power state data, generates an intra-frame intention heat map by fusion when the buffer is overloaded, splits the video stream into core intention and non-core background data streams, and estimates computing power consumption. The computing power consumption is input into a game cycle model, intention retention rate and overflow risk are iteratively calculated, the feature retention range is reduced when the risk is too high, and secondary task computing power is borrowed for compensation when the retention rate is insufficient, until a double-optimal state is reached. The dynamic balance of system load and information value is achieved. Finally, a scheduling decision matrix is generated, the core intention original drawing is transmitted, and redundant background is discarded, effectively solving the problem of key information loss in a resource-limited scene, and significantly improving the response efficiency and business fidelity of the cloud-edge-end cooperative system.

Description

Technical Field

[0001] This invention belongs to the field of computing power collaborative scheduling and management technology, and relates to an intent-driven cloud-edge-device heterogeneous computing power collaborative scheduling system and method. Background Technology

[0002] As the security monitoring field evolves towards ultra-high definition and intelligence, 4K video streams, as the core carrier of on-site details and key information, are crucial for accurately identifying target features and reconstructing the truth of events due to their high-resolution image quality. Meanwhile, the buffer, as a key storage node in the video acquisition and transmission link, directly determines the continuity of data flow and the real-time response capability of the system through its space utilization and throughput rate, serving as the cornerstone for ensuring the stable operation of video services.

[0003] It is worth noting that the high bitrate and large data volume characteristics of 4K video streams present an inherent trade-off with limited physical buffer resources. In sudden scenarios or network congestion, the instantaneous influx of massive amounts of video data can easily cause a sharp rise in buffer levels, triggering an overflow risk. This dynamic imbalance between data supply and storage capacity necessitates that the system make scheduling decisions within a very short time to prevent data stream interruptions or system crashes, making it a critical challenge to be addressed in video transmission control.

[0004] However, existing buffer overflow handling strategies often lack a deep understanding of the value of video content. Traditional methods rely heavily on first-in-first-out (FIFO) or coarse-grained frame-level discarding mechanisms, mechanically removing frames based solely on the order of data storage or space occupancy, while ignoring significant differences in semantic information within video frames. This lack of intent-driven processing makes it highly susceptible to misjudging high-value frames containing key target features as redundant data and discarding them when resources are scarce. This results in the permanent loss of valuable information at critical moments, failing to meet the refined requirement of "protecting all that should be protected" for core data in intelligent scenarios. Summary of the Invention

[0005] In view of the problems existing in the prior art, the present invention provides an intent-driven cloud-edge-device heterogeneous computing power collaborative scheduling system and method to solve the above-mentioned technical problems.

[0006] To achieve the above and other objectives, the technical solution adopted by the present invention is as follows: The first aspect of this invention provides an intent-driven cloud-edge-device heterogeneous computing power collaborative scheduling method, the method comprising: Acquire intent semantic vectors, video stream data, buffer occupancy data, and heterogeneous computing power status data; when the buffer occupancy data is greater than a preset occupancy threshold, perform calculations on the intent semantic vectors and video stream data to generate an intra-frame intent heatmap; Based on the intra-frame intent heatmap, the video stream data is split into a core intent data stream and a non-core background data stream; based on the core intent data stream, the non-core background data stream, and heterogeneous computing power status data, the expected computing power consumption is estimated. The expected value of computing power consumption is input into the game theory cycle model for iteration. The game theory cycle model includes the following iterative steps: Generate an estimated intent retention rate and a buffer overflow risk index based on the expected computing power consumption; If the buffer overflow risk index is greater than the preset risk threshold, reduce the feature retention range of the core intent data stream to update the expected value of computing power consumption, and return to the step of generating the estimated intent retention rate and buffer overflow risk index. If the buffer overflow risk index is not greater than the preset risk threshold and the estimated intent retention rate is less than the preset retention value, extract the secondary task computing power compensation parameters from the heterogeneous computing power status data to update the expected computing power consumption value, and return to the step of generating the estimated intent retention rate and the buffer overflow risk index. If the buffer overflow risk index is not greater than the preset risk threshold and the estimated intention retention rate is not less than the preset retention value, the current expected value of computing power consumption is converted into a scheduling decision matrix and the iteration is terminated. According to the scheduling decision matrix, the core intent data stream is processed with original-quality transmission, and the non-core background data stream is processed with dimensionality reduction and discarding.

[0007] A second aspect of the present invention provides an intent-driven cloud-edge-device heterogeneous computing power collaborative scheduling system, the system comprising: Intent heatmap calculation module: acquires intent semantic vector, video stream data, buffer occupancy data, and heterogeneous computing power status data; when the buffer occupancy data is greater than the preset occupancy threshold, it performs calculations on the intent semantic vector and video stream data to generate an intra-frame intent heatmap; The computing power consumption calculation module: Based on the intra-frame intent heatmap, the video stream data is split into core intent data stream and non-core background data stream; based on the core intent data stream, non-core background data stream and heterogeneous computing power status data, the expected value of computing power consumption is estimated; The computing power iteration loop module inputs the expected value of computing power consumption into the game loop model for iteration. The game loop model includes the following iteration steps: Generate an estimated intent retention rate and a buffer overflow risk index based on the expected computing power consumption; If the buffer overflow risk index is greater than the preset risk threshold, reduce the feature retention range of the core intent data stream to update the expected value of computing power consumption, and return to the step of generating the estimated intent retention rate and buffer overflow risk index. If the buffer overflow risk index is not greater than the preset risk threshold and the estimated intent retention rate is less than the preset retention value, extract the secondary task computing power compensation parameters from the heterogeneous computing power status data to update the expected computing power consumption value, and return to the step of generating the estimated intent retention rate and the buffer overflow risk index. If the buffer overflow risk index is not greater than the preset risk threshold and the estimated intention retention rate is not less than the preset retention value, the current expected value of computing power consumption is converted into a scheduling decision matrix and the iteration is terminated. According to the scheduling decision matrix, the core intent data stream is processed with original-quality transmission, and the non-core background data stream is processed with dimensionality reduction and discarding.

[0008] As described above, the intention-driven cloud-edge-device heterogeneous computing power collaborative scheduling system and method provided by the present invention have at least the following beneficial effects: 1. This invention constructs a multi-dimensional end-to-end perception foundation by acquiring intent semantic vectors, video stream data, buffer occupancy data, and heterogeneous computing power status data. When the buffer occupancy data exceeds a preset occupancy threshold, it uses intent semantic vectors and video stream data to perform cross-modal operations to generate an intra-frame intent heatmap, thereby achieving accurate mapping from underlying physical data to high-level semantic features. This process effectively overcomes the drawbacks of traditional video transmission that ignores business intent and blindly transmits all data, enabling the system to accurately identify and lock key target areas in the image at the data source, providing a reliable semantic basis for subsequent accurate video stream processing in resource-constrained environments.

[0009] 2. This invention constructs a game-theoretic loop model with iterative steps. Based on the intra-frame intent heatmap, video stream data is split into core intent data streams and non-core background data streams. Based on the expected value of computing power consumption, the estimated intent retention rate and buffer overflow risk index are dynamically generated. Through a nonlinear feedback mechanism that reduces the feature retention range when the risk is overloaded and plunders the computing power compensation parameters of secondary tasks when the retention rate is insufficient, a dynamic balance between system stability and business fidelity is achieved. This closed-loop game logic completely solves the problem that traditional methods are rigid in dealing with computing power fluctuations or network congestion, and are prone to causing the loss of key features or system crashes. It significantly improves the system's adaptive adjustment capability to sudden pressure in heterogeneous computing environments.

[0010] 3. This invention terminates the iteration when the buffer overflow risk and intent retention rate reach the dual-optimization standard. The current expected value of computing power consumption is transformed into a scheduling decision matrix, and the core intent data stream is processed with original-quality transmission according to the matrix, while non-core background data stream is processed with dimensionality reduction and discarding. This strategy not only effectively avoids the ineffective occupation of valuable network bandwidth and computing resources by non-critical background data, but also ensures high-fidelity transmission of core intent information, greatly improving the operating efficiency and decision response speed of the intelligent video analysis system, and providing a new technical path for refined resource scheduling in cloud-edge-device collaborative scenarios. Attached Figure Description

[0011] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0012] Figure 1 This is a schematic diagram showing the connections between the steps of the method of the present invention.

[0013] Figure 2 This is a schematic diagram illustrating the logical connection when the buffer overflow risk index exceeds a preset risk threshold in this invention.

[0014] Figure 3 This is a schematic diagram showing the connections of the various modules in the system of the present invention. Detailed Implementation

[0015] The following description, in conjunction with the implementation of this invention, is merely an example and illustration of the concept of this invention. Those skilled in the art can make various modifications or additions to the specific embodiments described, or use similar methods to replace them, as long as they do not deviate from the inventive concept or exceed the scope defined in these claims, all of which should fall within the protection scope of this invention.

[0016] In traditional cloud-edge-device video transmission systems, unified data scheduling strategies and static buffer management rules cannot adapt to the dynamic changes in network bandwidth fluctuations and heterogeneous computing resources. When the buffer occupancy rate of the video data stream surges non-linearly in sudden scenarios, the system cannot establish a dynamic correlation between the semantic value of the data and the underlying transmission resources, leading to a mismatch between scheduling decisions and actual business intentions. This static transmission mechanism reduces the timeliness of critical information transmission, causing the video stream acquired by the receiving end to contain a large amount of redundant background data and lose core intent features, ultimately affecting the accuracy and real-time performance of intelligent analysis results.

[0017] For example, in 4K high-definition security monitoring scenarios, sudden high-frequency events can cause the buffer occupancy rate to surge from 30% to the 95% overflow threshold in an instant. At this point, the uplink bandwidth can only support data throughput at 30% of the original resolution. Traditional systems still employ first-in-first-out (FIFO) or simple frame-level discarding strategies, failing to distinguish semantic differences within video frames. This results in the latest keyframes containing the suspect's facial features being directly discarded, while earlier, meaningless background frames occupy valuable transmission queues. The backend intelligent analysis platform, lacking core feature data, cannot trigger alarms, and the generated monitoring reports are either blank or contain false alarms.

[0018] If the above problems are not addressed, the indiscriminate discarding of critical business data will cause the intelligent sensing system to lose its ability to capture core targets under sudden pressure, resulting in a severe semantic loss disaster. The disconnect between scheduling strategies and intended value will exacerbate the idle computing power and network congestion at edge nodes, and may even trigger cascading system crashes. The lack of a game-theoretic feedback mechanism in resource allocation will also lead to an imbalance in the utilization of heterogeneous computing resources, causing high-priority tasks to be blocked due to computing power starvation, ultimately forming a negative feedback loop that affects the overall robustness and business continuity of the cloud-edge-device collaborative system.

[0019] To address the aforementioned issues, this application first considers establishing a dynamic correlation mechanism between data semantic value and physical resource status. Traditional systems use static rules to handle buffer overflows, resulting in the loss of core intent data along with background data. To resolve this, this application attempts to generate an intra-frame intent heatmap by performing cross-modal operations on intent semantic vectors and video stream data, splitting the video stream into core intent and non-core background data streams. Further analysis reveals that simply splitting data streams is insufficient to handle complex and variable computational bottlenecks; a game-theoretic iterative model is needed to iteratively calculate the intent retention rate and overflow risk index using the expected value of computational consumption. By designing a dual feedback adjustment mechanism of feature range reduction and computational compensation, the scheduling decision matrix adaptively converges with system pressure, thereby resolving the contradiction between transmission congestion and intent fidelity.

[0020] After introducing the basic concept of the present invention, the embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Example

[0021] Please see Figure 1-2 As shown, the intent-driven cloud-edge-device heterogeneous computing power collaborative scheduling method includes the following steps: Acquire intent semantic vectors, video stream data, buffer occupancy data, and heterogeneous computing power status data; if the buffer occupancy data is not greater than a preset occupancy threshold, extract full-frame pixel data from the video stream data and push the full-frame pixel data to the regular transmission queue to execute the original image transmission task; if the buffer occupancy data is greater than a preset occupancy threshold, perform calculations on the intent semantic vectors and video stream data to generate an intra-frame intent heatmap.

[0022] Preferably, generating an intra-frame intent heatmap includes: When the buffer occupancy rate is greater than the preset occupancy rate threshold, feature dimensionality reduction extraction calculation is performed on the intention semantic vector to generate a spatial attention weight matrix to characterize the target perception intensity. The spatial attention weight matrix is ​​used to perform feature mapping operation with the pixel array contained in the video stream data to generate the global saliency distribution matrix of the corresponding video stream frame. Pixel coordinates with values ​​greater than a preset activity threshold in the global saliency distribution matrix are selected, spatial clustering is performed on the pixel coordinates to form coordinate intervals of the region of interest, and an intra-frame intent heatmap is generated based on the coordinate intervals of the region of interest.

[0023] Preferably, generating the global saliency distribution matrix of the corresponding video stream frame includes: The video stream data is divided into grids according to a preset division scale to construct multiple grid regions and the basic feature pixel vectors corresponding to each grid region. Based on the spatial structure alignment rules, the attention weight values ​​that map to the spatial arrangement of each grid region are extracted from the spatial attention weight matrix. The attention weight values ​​are multiplied by the corresponding basic feature pixel vectors to calculate the regional saliency feature scores for each grid region. According to the original spatial coordinate order of each grid region in this pixel matrix, all regional saliency feature scores are matrix-merged to generate an intermediate saliency matrix. Obtain the maximum extreme value among the values ​​contained in the intermediate significance matrix, and use the maximum extreme value as the denominator to perform division normalization operation on all values ​​in the intermediate significance matrix in turn, and finally generate a global significance distribution matrix that maps to the preset numerical range.

[0024] In one specific embodiment, the system first acquires intent semantic vectors, video stream data, buffer occupancy data, and heterogeneous computing power status data in real time through the video acquisition module and the system monitoring link. When performing transmission logic judgments, a buffer occupancy threshold is pre-set. The threshold value is recommended to be set between 0.60 and 0.85. In order to ensure the continuity of image quality while preventing data packet loss to the greatest extent, this embodiment preferably uses 0.75 as the threshold for triggering hierarchical transmission. When the real-time monitored buffer occupancy data is not greater than 0.75, the system determines that the current network link and cache redundancy are sufficient, directly extracts the full-frame pixel data of the video stream data and pushes it to the regular transmission queue to perform the original image transmission task.

[0025] When the buffer occupancy rate is greater than 0.75 and a congestion warning is triggered, the system initiates an intent-driven feature extraction process. First, it performs feature dimensionality reduction extraction on the high-dimensional intent semantic vector. Then, using a pre-defined mapping transformation matrix, it projects the semantic features to the spatial dimension, generating a spatial attention weight matrix to characterize the target perception intensity. The calculation formula is as follows: ,in The generated spatial attention weight matrix has elements ranging from [0,1] and is used to quantify the importance of different spatial locations to the expression of intent. The input is a 1×N dimensional intent semantic vector; The preset N×(H×W) dimensional feature projection matrix is ​​obtained by pre-training based on a large number of historical intent and visual saliency association samples, and is used to map abstract semantics onto H×W grid regions of the image; It is the bias vector; This is the activation function.

[0026] Subsequently, the system performs feature mapping operations on the spatial attention weight matrix and the pixel array contained in the video stream data to generate a global saliency distribution matrix for the corresponding video stream frame. This process first divides the pixel array into grids according to a preset division scale (preferably 16×16 pixels as a grid unit to balance computational accuracy and processing overhead), constructs multiple grid regions, and extracts the basic feature pixel vector corresponding to each grid, that is, the average vector of pixel grayscale values ​​in that region. Then, based on the spatial structure alignment rules, the weight values ​​of the corresponding coordinates are extracted from the spatial attention weight matrix, and the regional saliency is calculated using the following formula: ,in This is the regional saliency feature score for the (i,j)th grid region, with the same dimensions as pixel intensity; These are the corresponding attention weights; Let be the brightness value of the p-th pixel within the grid; K is the total number of pixels in a single grid. Based on this, the system performs matrix concatenation according to the original spatial coordinate order of the grid to generate an intermediate saliency matrix, and uses a maximum normalization algorithm to generate the final global saliency distribution matrix, as shown in the formula: ,in This is the global significance distribution matrix. The MAX function is used to extract the maximum extreme value term in the matrix as the denominator. This normalization process eliminates the absolute brightness deviation under different lighting backgrounds, so that the matrix values ​​are uniformly mapped to the relative intensity range of [0,1].

[0027] Finally, the system selects pixel coordinates with values ​​greater than a preset activity threshold from the global saliency distribution matrix, performs spatial clustering on the discrete coordinates using a density clustering algorithm, identifies the closed region with the highest pixel density and determines the coordinate range of its bounding rectangle, forming the coordinate interval of the region of interest, and generates an intra-frame intent heatmap accordingly.

[0028] It should be added that when the buffer occupancy rate does not exceed a preset threshold, the system extracts full-frame pixels from the video stream data and pushes them to the regular transmission queue for original image transmission, thus ensuring the integrity of image quality when resources are sufficient. Once the buffer occupancy rate exceeds the preset threshold, the system switches to generating an intra-frame intent heatmap. The core principle of this is to focus on key areas of interest in the image through the collaborative calculation of intent semantic vectors and video stream data, thereby reducing data processing overhead and maintaining analytical efficiency.

[0029] Specifically, the steps for generating an intra-frame intent heatmap follow a progressive logical sequence: First, feature dimensionality reduction is performed on the intent semantic vector to generate a spatial attention weight matrix. This step simplifies high-dimensional intent information into an intensity representation that can be mapped to the spatial domain. Then, this matrix is ​​used for feature mapping with the video stream pixel matrix. By multiplying the weights by the pixel features, the saliency score of each grid region is obtained, and these are then concatenated into an intermediate saliency matrix. This process ensures spatial alignment between intent information and visual data. Next, the intermediate matrix is ​​normalized to generate a global saliency distribution matrix, ensuring all values ​​fall within a preset range for subsequent threshold filtering. Finally, significant pixel coordinates are selected based on an activity threshold, and spatial clustering is used to form coordinate intervals for areas of interest, ultimately generating the heatmap. Each step depends on the output of the previous step: dimensionality reduction provides structured weights for mapping, mapping operations transform weights into visual saliency, normalization ensures data comparability, and clustering aggregates spatially continuous regions based on the saliency distribution. When resources are limited, an intent-driven attention mechanism prioritizes processing the parts of the video that are relevant to the user's intent, reducing transmission and computational load while preserving semantically critical information. This process not only improves the system's adaptability in heterogeneous computing environments but also visually reveals the focus of intent through heatmaps, enhancing the efficiency and targeting of video analysis.

[0030] Based on the intra-frame intent heatmap, the video stream data is split into a core intent data stream and a non-core background data stream; the expected value of computing power consumption is estimated based on the core intent data stream, the non-core background data stream, and heterogeneous computing power status data.

[0031] Preferably, the video stream data is split into a core intent data stream and a non-core background data stream based on the intra-frame intent heatmap, including: Analyze the spatial distribution structure of the intra-frame intent heatmap and extract the set of closed boundary contour coordinates used to delineate the preset high-heat pixel intervals; The current video stream data is restored and unfolded into a complete original two-dimensional layer pixel array; Using the set of closed boundary contour coordinates, a spatial mask cutting operation is performed on the pixel array of the original two-dimensional layer. The internal pixel clusters surrounded by the set of closed boundary contour coordinates are extracted as the focus feature target point array, and the peripheral environment pixel clusters other than the focus feature target point array are extracted separately as blank background base map point arrays. Based on the original continuous frame arrangement order of the video stream data, the multi-frame focused feature target point array generated within the preset reading period is encapsulated and spliced ​​sequentially according to the original time sequence to generate a core intent data stream with an independent structure. Following the same continuous frame arrangement order, the corresponding multiple blank background base map dot matrix within the preset reading period are synchronously dimensionality reduced and stitched together to generate a non-core background data stream after the key feature information has been extracted.

[0032] Preferably, the expected value of computing power consumption is estimated, including: Read the amount of key feature bits in the core intent data stream and the amount of base map bits in the non-core background data stream; The calculation of key feature bit data volume is transformed into core computing requirements with high resource ratio, and the calculation of basic base map bit data volume is transformed into edge dimensionality reduction transformation requirements with low resource ratio. Analyze heterogeneous computing power status data to extract the current idle computing power quota parameters of each independent processing node distributed in the cloud, edge and terminal, as well as the available bandwidth parameters of the network links connecting the nodes at all levels. The core computing demand and the edge dimensionality reduction transformation demand are used as the dividend, and the current idle computing power quota parameter is used as the divisor to perform a division quantization allocation operation to obtain the local computing cost score associated with each independent processing node. The throughput load is calculated by combining the key feature bit data volume and the base map bit data volume with the available bandwidth parameters of the network link, and the remote network transmission consumption score of the corresponding overall data flow and migration overhead is generated. According to the preset system computing network integration evaluation ratio weight, the local computing cost score and the remote network transmission consumption score are weighted and accumulated to generate the expected value of computing power consumption.

[0033] Preferably, the logic for associating the local computation cost scores of each independent processing node is as follows: According to the preset node task allocation mapping relationship, the core computing demand and the edge dimensionality reduction transformation demand are assigned to the corresponding independent processing nodes, and the computing power demand values ​​assigned to the same node are accumulated to generate the node task carrying capacity associated with each independent processing node. For each independent processing node, extract the corresponding current idle computing power quota parameter, set the node task capacity as the dividend and the current idle computing power quota parameter as the divisor, perform a division operation, and calculate the node computing power load ratio that represents the current resource squeezing ratio. Extract the preset penalty weights that characterize the sensitivity of computing power, and perform fixed-point multiplication of the node computing power load ratio with the preset penalty weights to calculate the local computing cost score of each independent processing node.

[0034] It should be added that content layering of the video stream based on the generated intra-frame intent heatmap aims to differentiate data according to semantic importance and provide a quantitative basis for subsequent resource scheduling decisions. Specifically, firstly, spatial masking is performed on the original video frames based on the closed boundary contours representing high attention in the heatmap. This step is based on the premise that the heatmap is generated driven by intent semantics, and its high-heat areas strictly correspond to key targets in the video spatially. Therefore, by segmenting through this contour, the "focused feature target matrix" representing the core intent and its complement "blank background matrix" can be accurately separated at the pixel level. The above matrix from multiple consecutive frames is then encapsulated according to its original temporal order, naturally yielding the core intent data stream and the non-core background data stream. This process ensures that the continuity of dynamic intent in the temporal dimension is maintained.

[0035] After data splitting, estimating the expected computational power consumption is essential for adaptive resource scheduling. First, the data volume (bits) of the two data streams is converted into corresponding processing requirements. This is based on a reasonable model: core intent data typically requires complex analysis (such as target recognition), thus its bit volume corresponds to a high-weight core computational requirement; while non-core background data can be simplified (such as dimensionality reduction and compression), thus its bit volume corresponds to a low-weight edge transformation requirement. Next, heterogeneous computational power status data is introduced as constraints, including the idle computational power and network bandwidth of each node. Core computation and edge transformation requirements are allocated to the corresponding nodes. By calculating the ratio of "node task load" to "current idle computational power," the node computational power load ratio is obtained, multiplied by a preset penalty weight to reflect the non-linear overhead caused by resource competition, ultimately generating a local computation cost score. Simultaneously, the throughput of data volume and available bandwidth is measured to generate a remote network transmission consumption score. Finally, based on the system's trade-off preference for computation and communication (preset weights), the two are weighted and fused to generate a unified expected computational power consumption value. The system systematically quantifies the characteristics of data processing tasks into the requirements for computing cycles and network bandwidth in heterogeneous environments, and provides a comparable total cost index through weighted integration, thereby providing an objective quantitative basis for determining whether to execute or how to schedule the processing task.

[0036] In summary, by deriving data partitions from semantic heatmaps and then establishing a predictive model from data partitions to resource consumption, the system can not only identify the content that should be prioritized, but also assess in advance the system overhead required to process that content. This lays a key foundation for making efficient and adaptive transmission or analysis decisions in resource-constrained scenarios such as buffer overload.

[0037] In one specific embodiment, the system traverses the spatial heatmap matrix and sets a preset heat differentiation threshold. The threshold is recommended to be set within the dimensionless probability range of 0.60 to 0.85. To avoid falsely removing valid edge features, this embodiment preferably uses 0.75. The system uses a contour line extraction algorithm to obtain all heat values ​​greater than or equal to... The system extracts the coordinates of the outermost pixels of the connected components to form a closed boundary contour coordinate set. After restoring and unfolding the current video stream data into a complete original two-dimensional layer pixel array, the system constructs a binary space mask M(x,y) on the layer based on this closed boundary contour coordinate set. The coordinates M(x,y) located inside the boundary are assigned a value of 1, and the coordinates located outside the boundary are assigned a value of 0.

[0038] Spatial segmentation is performed using this mask. The original pixel matrix is ​​multiplied by the mask using an outer product, extracting the enclosed internal pixel clusters losslessly as focused feature target points. Simultaneously, the mask is inverted and multiplied with the original layer, extracting the surrounding environment pixel clusters separately as blank background point maps. Subsequently, the system introduces a preset reading cycle based on the original continuous frame arrangement of the video stream. To ensure smooth inter-frame flow while accumulating sufficient data slices, the period is recommended to be set within an adjustable range of 100 to 300 milliseconds. In this embodiment, 200 milliseconds is preferred. The multi-frame focused feature target point array within this period is aggregated and stitched together according to its original time sequence to generate a core intent data stream with a complete spatiotemporal structure. For the multi-frame blank background map point array, the spatial resolution and frame rate are synchronously reduced and stitched together using the mean pooling algorithm to generate a non-core background data stream that is highly compressed and stripped of key feature information.

[0039] Next, the system estimates the expected computing power consumption based on the previously separated core intent data stream, non-core background data stream, and heterogeneous computing power status data. First, it reads the amount of key feature bits contained in the core intent data stream in real time. and the amount of base map bit data included in the non-core background data stream. Both are measured in bits. Since the core features require high-dimensional neural network semantic reasoning, while the background image only needs basic image reconstruction encoding, the system uses a differential computing power consumption mapping model to calculate the computational requirements, as shown in the formula. as well as ,in This represents the core computing requirements, expressed in floating-point operations. The core operation complexity operator represents the computational cost required for each bit of critical data, and its value is obtained based on the depth of the pre-loaded inference model in the background; similarly, To meet the demand for edge dimension reduction transformation, The background operation complexity operator.

[0040] The system then analyzes the current heterogeneous computing network environment and extracts the current idle computing power quota parameters of each independent processing node in the cloud, edge data center, and local terminal. The unit is floating-point operations per second, and the available bandwidth of the network links for communication between nodes at each level is measured simultaneously. The unit is bits per second. Based on the preset node task allocation mapping relationship, the system assigns the core computing requirements and edge dimensionality reduction transformation requirements to the corresponding independent processing nodes. It then performs a scalar accumulation operation on all computing power requirements assigned to the same physical node k to generate the node task capacity associated with that node. Based on this, the node task capacity is set as the divisor, and the current idle computing power quota parameter is set to the preset task response time window. The product is set as the divisor to perform division quantization allocation operation, and the formula is: This formula, a variation of the service utilization formula in classical queuing theory, innovatively introduces a mandatory time constraint term. (Unit: seconds), transforming the denominator from pure rate to the "maximum absolute computing power capacity of the system" within that time window, thus converting the dynamically uncertain physical concurrency process into a static absolute occupancy measurement; the calculated... The node computing power load ratio represents the current resource occupancy rate. To mitigate the overheating and downtime threats caused by full load on certain nodes (such as battery-powered low-power edge devices), the system extracts a pre-set penalty weight representing the degree of computing power sensitivity. The weight is a dimensionless coefficient, set to 1.0 by default for nodes with constant power supply, and dynamically set to 1.5 to 2.5 for sensitive devices such as power-constrained terminals. The system performs a fixed-point multiplication of the node's computing power load ratio with the corresponding node's preset penalty weight to directly generate the local computation cost score associated with each independent processing node. .

[0041] After completing the evaluation of the local computational dimension, the system simultaneously transfers the amount of key feature bit data extracted earlier. and the amount of base map bit data Combined as the total circulating numerator, and the available bandwidth parameter of the network link. and preset transmission delay tolerance threshold Combined, throughput load is calculated using the following formula: ,in The unit is seconds, thus generating a remote network transmission consumption score for the overall data flow and migration overhead. Finally, the system calculates the network integration evaluation weights according to preset system weights (e.g., setting separate calculation weights). =0.6 and transmission weight =0.4), and perform a weighted summation and integration operation on the average of the local computation cost scores of all scheduled systems and the remote network transmission consumption scores, that is... Finally, the expected value of computing power consumption is generated by summarizing. . This represents the average score of local computation cost.

[0042] The expected value of computing power consumption is input into the game theory cycle model for iteration. The game theory cycle model includes the following iterative steps: Generate an estimated intent retention rate and a buffer overflow risk index based on the expected computing power consumption; If the buffer overflow risk index is greater than the preset risk threshold, reduce the feature retention range of the core intent data stream to update the expected value of computing power consumption, and return to the step of generating the estimated intent retention rate and buffer overflow risk index. If the buffer overflow risk index is not greater than the preset risk threshold and the estimated intent retention rate is less than the preset retention value, extract the secondary task computing power compensation parameters from the heterogeneous computing power status data to update the expected computing power consumption value, and return to the step of generating the estimated intent retention rate and the buffer overflow risk index. If the buffer overflow risk index is not greater than the preset risk threshold and the estimated intention retention rate is not less than the preset retention value, the current expected value of computing power consumption is converted into a scheduling decision matrix and the iteration is terminated. According to the scheduling decision matrix, the core intent data stream is processed with original-quality transmission, and the non-core background data stream is processed with dimensionality reduction and discarding.

[0043] Preferably, an estimated intent retention rate and a buffer overflow risk index are generated based on the expected computing power consumption, including: Input the expected value of computing power consumption into a pre-set performance depreciation mapping matrix, and extract the intention feature depreciation coefficient corresponding to the current resource consumption state; The key feature bit data volume constituting the core intent data stream is obtained, and the key feature bit data volume is multiplied with the intent feature conversion coefficient to perform dimensionality reduction processing, and the estimated retained feature data volume under the current computing power conditions is calculated. The estimated retention feature data amount is used as the dividend and the key feature bit data amount is used as the divisor to perform a numerical division operation to calculate and generate the estimated intention retention rate. The expected value of computing power consumption is substituted into the preset queuing time estimation rule to perform resource mapping and deduction, and the expected cumulative processing latency parameter corresponding to the current node allocation architecture is generated. Obtain the video queue write rate parameter associated with the current terminal device, multiply and integrate the video queue write rate parameter with the expected processing cumulative latency parameter to obtain the queue prediction backlog increment parameter generated within the latency period; Extract the buffer occupancy data that is monitored synchronously during the video stream data acquisition stage, and add and merge the queue predicted backlog increment parameter with the buffer occupancy data to generate a buffer overflow risk index.

[0044] Preferably, when the buffer overflow risk index exceeds a preset risk threshold, the feature retention range of the core intent data stream is reduced to update the expected value of computing power consumption, including: The risk overflow difference parameter is obtained by subtracting the buffer overflow risk index from the preset risk threshold. The risk spillover difference parameter is linked with the preset convergence adjustment mapping table for a correlation query and matching operation to extract the feature truncation compensation coefficient corresponding to the risk spillover difference parameter. Extract a preset activity threshold, and perform a multiplication amplification conversion process on the preset activity threshold and the feature truncation compensation coefficient to generate a dynamic intent truncation threshold for shrinking the target perception boundary; A secondary filtering screening is performed on the pixel data contained in the core intent data stream based on the dynamic intent truncation threshold, removing edge-related pixel clusters whose feature significance is less than the dynamic intent truncation threshold, thereby compressing the core intent data stream into a dimensionality-reduced core intent data stream. Read the amount of updated key feature bits corresponding to the core intent data stream of dimensionality reduction, and recalculate based on the amount of updated key feature bits, non-core background data stream and heterogeneous computing power status data to generate the expected value of updated computing power consumption. Extract the expected value of updated computing power consumption to generate state machine jump instructions. Based on the state machine jump instructions, trigger the system execution pointer to backtrack, and force the current round of computing process to return to the extraction and generation steps of the estimated intention retention rate and buffer overflow risk index before execution.

[0045] When the buffer overflow risk index is no greater than the preset risk threshold and the estimated intention retention rate is less than the preset retention value, the specific operation logic is as follows: The retention rate difference parameter is calculated by subtracting the preset retention value as the minuend and the estimated intended retention rate as the subtrahend. The retention rate difference parameter is used to perform query matching operations on the pre-built computing power borrowing mapping table to determine the compensation computing power quota required to make up for the lack of current intent features; Based on the compensation computing power quota, the heterogeneous computing power status data is addressed and traversed for retrieval, parsing and suspending the secondary background execution tasks currently at the bottom layer, thereby extracting the secondary task computing power compensation parameters that have reached the compensation computing power quota indicators. Obtain the expected computing power consumption value within the current verification cycle, perform addition and superposition operation on the expected computing power consumption value and the computing power compensation parameter of the secondary task, and obtain the updated expected computing power consumption value after superimposing the computing power margin; Based on the expected value of updated computing power consumption, process reverse backtracking control instructions are constructed and issued. Through process reverse backtracking control instructions, the execution pointer of the computing node is forced to jump, guiding the system to return to the processing stage of generating the estimated intention retention rate and buffer overflow risk index.

[0046] Preferably, the current expected computing power consumption is transformed into a scheduling decision matrix, including: Extract the convergence target signal generated by comparing the current buffer overflow risk index and the estimated intention retention rate, and capture and lock the current expected value of computing power consumption from the cache space based on the convergence target signal; The locked expected value of computing power consumption is substituted into the preset heterogeneous node mapping dimension space to decompose the data protocol and extract the distributed computing power quota vectors of each independent processing node corresponding to different spatial levels. Extract the set of node topology coordinate parameters of the global computing network, and perform a two-dimensional arrangement, fusion and alignment operation on all the distributed computing power quota vectors according to the set of node topology coordinate parameters to construct a scheduling decision matrix that matches the current network physical structure. Extract the scheduling decision matrix and compile it to generate the corresponding system state solidification control instructions. Based on the system state solidification control instructions, intercept and forcibly shut down the computing resource evaluation loop state machine associated with the task, thereby physically severing the jump link of parameter reverse backtracking and completing the termination iteration action of the entire calculation cycle.

[0047] It should be added that this step generates an estimated intent retention rate and a buffer overflow risk index based on the expected computing power consumption. These two are the core criteria for evaluating the feasibility of the decision. The intent retention rate is calculated by mapping the core data volume to resource depreciation, simulating the proportion of intent information that can be retained under the estimated resource consumption. The buffer overflow risk index is calculated by converting the estimated processing latency into the queue backlog increment and adding it to the current occupancy rate, predicting the buffer overflow risk that may be caused by executing this scheduling scheme.

[0048] Subsequently, the model enters a conditional judgment and iterative loop based on the above two indicators, with its logical connection following the closed-loop control principle of "problem identification - targeted adjustment - re-evaluation". In the first case, when the buffer overflow risk index exceeds the threshold, it indicates that the current solution may lead to system instability. In this case, the feature selection threshold is dynamically increased to reduce the feature retention range of the core intent data stream. This operation directly reduces the amount of core data that needs to be processed, thereby lowering the expected value of the updated computing power consumption, aiming to sacrifice a small amount of intent information to prioritize ensuring the system does not overflow. In the second case, when the risk is controllable but the estimated intent retention rate is insufficient, it indicates that the system has a safety margin but the processing quality is substandard. In this case, secondary tasks are suspended to extract the computing power compensation parameters for secondary tasks and add them to the original expected value. This is equivalent to dynamically allocating more computing resources to the current task, which is expected to improve the updated intent retention rate. In the third case, when the risk is controllable and the retention rate meets the standard, it indicates that a feasible solution satisfying both quality and stability constraints has been found, and the iteration terminates. Finally, a scheduling decision matrix is ​​generated based on the expected value of computing power consumption at convergence. This matrix is ​​essentially a specific allocation scheme of resources among heterogeneous nodes, and based on this, differentiated fidelity preservation and dimensionality reduction processing is performed on core and non-core data streams.

[0049] The entire model is constructed as a feedback system with the expected value of computing power consumption as the state variable. Whether narrowing the scope or leveraging computing power, both are adjustments to this state variable. After adjustment, the model must return to the starting point and recalculate the risk and retention rate indicators based on the new state value to verify the effect of the adjustment. This process repeats itself, like a game, until a stable solution is found that makes both the system state (risk and retention rate) fall within an acceptable range. This ensures that the final decision is the result of multiple rounds of trade-offs and verification, achieving an optimal trade-off between intent analysis quality and system operational stability under given constraints, thereby improving the adaptability and robustness of the video analytics system in complex environments.

[0050] In one specific embodiment, the system will The input is a nonlinear performance degradation mapping function pre-fitted based on massive historical streaming test data, used to calculate the intention feature reduction coefficient. The specific formula is as follows ,in The baseline computing power requirements for maintaining 100% feature resolution in a video stream. This is a preset inflection point parameter for computing power threshold, reflecting the critical point of hardware performance degradation. A value between 0.6 and 0.8 is generally recommended; in this embodiment, 0.75 is preferred to ensure that preventative dimensionality reduction is triggered when computing power utilization reaches 75%. The smoothing adjustment constant, which is a dimensionless coefficient for controlling the steepness, is preferably 10, and e is a natural constant.

[0051] Subsequently, the system acquires the amount of key feature bits that constitute the core intent data stream in the current frame image. Its unit is Mbit, and it is compared with Perform multiplication and dimensionality reduction to calculate the estimated amount of feature data to be retained under the current computing power conditions. Dividing the two yields the dimensionless estimated retention rate. To assess network congestion risk, the system will... Substituting the preset queuing time estimation rules, the expected cumulative processing delay is deduced. This calculation formula is a variation of the M / M / 1 model in queuing theory, adapted to the bounded memory characteristics of edge computing. Specifically: ,in This represents the total amount of data currently backed up in the queue. This represents the instruction forwarding conversion rate of the current node, measured in Mbit / GOP. It indicates the amount of bitstream data that can be processed per GOP of computing power. Measured in real-time by a front-end probe, the system uses this division combined with constant conversion to derive an absolutely precise latency in milliseconds. Next, the system obtains the video queue write rate through the underlying socket interface. Its unit is Mbit / ms, multiplied by the predicted queue backlog increment within the delay period. Ultimately, the system extracts the amount of data already occupied in the initial buffer within the RAM segment, which is synchronously monitored via hardware interrupts. Add the two together and then compare them with the device's physical maximum buffer capacity. Divide them to get the buffer overflow risk index. The dimension is percentage.

[0052] In the first decision branch, the buffer overflow risk index calculated above... Greater than the preset risk threshold In such cases, this threshold is typically set as a safe red line for the system's tolerable physical memory level, recommended to be between 75% and 85%. In this embodiment, to cope with sudden large-scale background traffic, it is preferably set to 80%. It is necessary to forcibly reduce the feature retention range of the core intent data stream to alleviate memory pressure and thus update the expected computing power value. The system will then use the actual risk value... Subtract the preset risk threshold Obtain the risk spillover difference parameter Next, the difference parameter is used to calculate the dynamic truncation threshold; specifically, the difference is input into the exponential compensation formula. Among them To truncate the step size coefficient, As the overload acceleration factor, the characteristic cutoff compensation coefficient is derived. .

[0053] Then, the system's preset activity detection benchmark threshold is extracted. This refers to the bottom-level boundary of pixel saliency filtering, which is usually based on the historical frame difference method. The residual mean is set to a default grayscale gradient of 15, and the conversion is performed using a product formula. A dynamic intent truncation threshold is generated. Using this threshold as a guideline, the system performs a second Sobel traversal on the current core intent's raw data stream, calculating the spatial gradient magnitude of each pixel cluster as its feature saliency. When the absolute value of the feature saliency of a local region is less than this dynamic intent truncation threshold, the truncation threshold is determined. At that time, it is marked as a redundant edge-related pixel and a zeroing-out operation is performed. The system reads the updated key feature bit data of the dimensionality reduction stream and re-runs the Markov chain resource evaluation logic to generate a completely new expected value for updated computing power consumption. Finally, the system uses the scheduler module built into the microcontroller to capture the updated expected value and generate a state machine jump instruction, forcing the computation execution pointer to perform a reverse backtracking pop operation in the memory stack, forcing the computation stream to return to the retention rate and risk calculation steps described at the beginning and retesting and tuning based on the new parameters.

[0054] If no overflow error occurs, then the buffer overflow risk index is within the range. Not exceeding the preset risk threshold, and the estimated intention retention rate. Less than the preset retention value In situations where high-resolution video transmission for core purposes such as remote conferencing or autonomous driving is required, a preset retention value is typically defined between 0.85 and 0.90. The system then initiates a compensation logic to leverage computing power to maintain image quality. The system performs subtraction to derive the retention rate difference parameter. Using this request, the difference input computing power borrowing function is used to derive the compensation computing power quota needed to compensate for the current intention's lack of resources. ,in This represents the system's maximum theoretical computing power. To compensate for losses in actual scheduling, a damping coefficient of 1.5 is used to borrow computing power. After defining the metrics, the system operation kernel thread performs address traversal to retrieve the underlying task stack. A frozen context state mechanism is used to temporarily parse and suspend secondary background tasks with execution priority below level C (such as log rotation and regular background system heartbeat detection) at the current timestamp. From these tasks, the corresponding secondary task computing power compensation parameters that have met the quota are released and reclaimed. Subsequently, the system numerically adds the expected value of the original computing power consumption in this round to the computing power released through this borrowing, that is... The expected value of the updated computing power after superimposed compensation computing power is obtained. The scheduler randomly sends a reverse backtracking control level signal to the main computing process, directly manipulating the CPU's program counter to force the instruction execution flow to jump back to the entry point of the processing stage that generates the estimated retention rate and risk index. This process is repeated until the borrowed computing power is sufficient to make the image quality retention rate exceed the target threshold.

[0055] Finally, when the optimal solution is reached—that is, when the calculated buffer overflow risk index is no greater than a preset threshold and the estimated intention retention rate fully meets or exceeds the preset retention value—the system terminates the game-theoretic closed-loop process and transforms the parameters into a fixed scheduling decision matrix. The system's underlying arbitrator compares the results and outputs a Boolean signal indicating convergence, triggering a data phase-locked loop to physically lock the expected computing power consumption value at the top layer of the current memory stack. Subsequently, based on the protocol parsing engine's layer-by-layer decomposition of the spatial requirements mapped from the overall expected computing power consumption value, it decomposes these requirements according to the communication protocol into dedicated computing power quota vectors responsible for decoding, rendering, network transmission, and other tasks associated with multiple heterogeneous stream processing chip cores. At this point, the communication bus management module is invoked to extract the set of node interconnection topology coordinate parameters of the global edge computing physical network. Through tensor product, these planarized one-dimensional computing power quota vectors, along with their mapped coordinate axes, undergo a many-to-many two-dimensional arrangement and fusion operation, thereby assembling and constructing an N×N-dimensional scheduling decision matrix in main memory that perfectly matches the actual hardware physical link architecture. Example

[0056] like Figure 3 As shown, the cloud-edge-device heterogeneous computing power collaborative scheduling system based on intent-driven architecture includes an intent heatmap calculation module, a computing power consumption calculation module, and a computing power iteration loop module. The various modules are connected via wired and / or wireless connections to enable data transmission between them; Intent heatmap calculation module: acquires intent semantic vector, video stream data, buffer occupancy data, and heterogeneous computing power status data; when the buffer occupancy data is greater than the preset occupancy threshold, it performs calculations on the intent semantic vector and video stream data to generate an intra-frame intent heatmap; The computing power consumption calculation module: Based on the intra-frame intent heatmap, the video stream data is split into core intent data stream and non-core background data stream; based on the core intent data stream, non-core background data stream and heterogeneous computing power status data, the expected value of computing power consumption is estimated; The computing power iteration loop module inputs the expected value of computing power consumption into the game loop model for iteration. The game loop model includes the following iteration steps: Generate an estimated intent retention rate and a buffer overflow risk index based on the expected computing power consumption; If the buffer overflow risk index is greater than the preset risk threshold, reduce the feature retention range of the core intent data stream to update the expected value of computing power consumption, and return to the step of generating the estimated intent retention rate and buffer overflow risk index. If the buffer overflow risk index is not greater than the preset risk threshold and the estimated intent retention rate is less than the preset retention value, extract the secondary task computing power compensation parameters from the heterogeneous computing power status data to update the expected computing power consumption value, and return to the step of generating the estimated intent retention rate and the buffer overflow risk index. If the buffer overflow risk index is not greater than the preset risk threshold and the estimated intention retention rate is not less than the preset retention value, the current expected value of computing power consumption is converted into a scheduling decision matrix and the iteration is terminated. According to the scheduling decision matrix, the core intent data stream is processed with original-quality transmission, and the non-core background data stream is processed with dimensionality reduction and discarding.

[0057] It should be noted that the interval and threshold sizes are set for ease of comparison. The size of the threshold depends on the amount of sample data and the base number set by those skilled in the art for each set of sample data, as long as it does not affect the proportional relationship between the parameter and the quantized value. Furthermore, the above formulas are all dimensionless calculations, and the formulas are derived from software simulations using a large amount of collected data to obtain the most recent real-world results. The preset parameters in the formulas are set by those skilled in the art according to the actual situation.

[0058] It should be understood that in the various embodiments of this application, the order of the above-mentioned processes does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.

[0059] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0060] In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. An intent-driven cloud-edge-device heterogeneous computing power collaborative scheduling method, characterized in that, include: Acquire intent semantic vectors, video stream data, buffer occupancy data, and heterogeneous computing power status data; When the buffer occupancy rate is greater than the preset occupancy threshold, the intent semantic vector and video stream data are processed to generate an intra-frame intent heatmap. Based on the intra-frame intent heatmap, video stream data is split into a core intent data stream and a non-core background data stream. The expected value of computing power consumption is estimated based on the core intent data stream, non-core background data stream, and heterogeneous computing power status data. The expected value of computing power consumption is input into the game theory cycle model for iteration. The game theory cycle model includes the following iterative steps: Generate an estimated intent retention rate and a buffer overflow risk index based on the expected computing power consumption; If the buffer overflow risk index is greater than the preset risk threshold, reduce the feature retention range of the core intent data stream to update the expected value of computing power consumption, and return to the step of generating the estimated intent retention rate and buffer overflow risk index. If the buffer overflow risk index is not greater than the preset risk threshold and the estimated intent retention rate is less than the preset retention value, extract the secondary task computing power compensation parameters from the heterogeneous computing power status data to update the expected computing power consumption value, and return to the step of generating the estimated intent retention rate and the buffer overflow risk index. If the buffer overflow risk index is not greater than the preset risk threshold and the estimated intention retention rate is not less than the preset retention value, the current expected value of computing power consumption is converted into a scheduling decision matrix and the iteration is terminated. According to the scheduling decision matrix, the core intent data stream is processed with original-quality transmission, and the non-core background data stream is processed with dimensionality reduction and discarding.

2. The intention-driven cloud-edge-device heterogeneous computing power collaborative scheduling method according to claim 1, characterized in that, Generate an intra-frame intent heatmap, including: When the buffer occupancy rate is greater than the preset occupancy rate threshold, feature dimensionality reduction and extraction calculation are performed on the intent semantic vector to generate a spatial attention weight matrix. The spatial attention weight matrix is ​​used to perform feature mapping operation with the pixel array contained in the video stream data to generate the global saliency distribution matrix of the corresponding video stream frame. Pixel coordinates with values ​​greater than a preset activity threshold in the global saliency distribution matrix are selected, spatial clustering is performed on the pixel coordinates to form coordinate intervals of the region of interest, and an intra-frame intent heatmap is generated based on the coordinate intervals of the region of interest.

3. The intention-driven cloud-edge-device heterogeneous computing power collaborative scheduling method according to claim 2, characterized in that, Generate the global saliency distribution matrix for the corresponding video stream frames, including: The video stream data is divided into grids according to a preset division scale to construct multiple grid regions and the basic feature pixel vectors corresponding to each grid region. Based on the spatial structure alignment rules, the attention weight values ​​that map to the spatial arrangement of each grid region are extracted from the spatial attention weight matrix. The attention weight values ​​are multiplied by the corresponding basic feature pixel vectors to calculate the regional saliency feature scores for each grid region. According to the original spatial coordinate order of each grid region in this pixel matrix, all regional saliency feature scores are matrix-merged to generate an intermediate saliency matrix. Obtain the maximum extreme value among the values ​​contained in the intermediate significance matrix, and use the maximum extreme value as the denominator to perform division normalization operation on all values ​​in the intermediate significance matrix in turn, and finally generate a global significance distribution matrix that maps to the preset value range.

4. The intention-driven cloud-edge-device heterogeneous computing power collaborative scheduling method according to claim 1, characterized in that, Based on intra-frame intent heatmaps, video stream data is split into a core intent data stream and a non-core background data stream, including: Analyze the spatial distribution structure of the intra-frame intent heatmap and extract the set of closed boundary contour coordinates used to delineate the preset high-heat pixel intervals; The current video stream data is restored and unfolded into a complete original two-dimensional layer pixel array; Using the set of closed boundary contour coordinates, a spatial mask cutting operation is performed on the pixel array of the original two-dimensional layer. The internal pixel clusters surrounded by the set of closed boundary contour coordinates are extracted as the focus feature target point array, and the peripheral environment pixel clusters other than the focus feature target point array are extracted separately as blank background base map point arrays. Based on the original continuous frame arrangement order of the video stream data, the multi-frame focused feature target point array generated within the preset reading period is encapsulated and spliced ​​sequentially according to the original time sequence to generate a core intent data stream with an independent structure. Following the same continuous frame arrangement order, the corresponding multiple blank background base map dot matrix within the preset reading period are synchronously dimensionality reduced and stitched together to generate a non-core background data stream after the key feature information has been extracted.

5. The intention-driven cloud-edge-device heterogeneous computing power collaborative scheduling method according to claim 1, characterized in that, The estimated expected computing power consumption includes: Read the amount of key feature bits in the core intent data stream and the amount of base map bits in the non-core background data stream; The calculation of key feature bit data volume is transformed into the core computational requirement of high resource ratio attribute, and the calculation of basic base map bit data volume is transformed into the edge dimensionality reduction transformation requirement of low resource ratio attribute. Analyze heterogeneous computing power status data to extract the current idle computing power quota parameters of each independent processing node distributed in the cloud, edge and terminal, as well as the available bandwidth parameters of the network links connecting the nodes at all levels. The core computing demand and the edge dimensionality reduction transformation demand are used as the dividend, and the current idle computing power quota parameter is used as the divisor to perform a division quantization allocation operation to obtain the local computing cost score associated with each independent processing node. The throughput load is calculated by combining the key feature bit data volume and the basic base map bit data volume with the available bandwidth parameters of the network link, and the remote network transmission consumption score of the corresponding overall data flow and migration overhead is generated. According to the preset system computing network integration evaluation ratio weight, the local computing cost score and the remote network transmission consumption score are weighted and accumulated to generate the expected value of computing power consumption.

6. The intention-driven cloud-edge-device heterogeneous computing power collaborative scheduling method according to claim 5, characterized in that, The logic for associating the local computation cost scores of each independent processing node is as follows: Based on the preset node task allocation mapping relationship, the core computing demand and the edge dimensionality reduction transformation demand are assigned to the corresponding independent processing nodes, and the computing power demand values ​​assigned to the same node are accumulated to generate the node task carrying capacity associated with each independent processing node. For each independent processing node, extract the corresponding current idle computing power quota parameter, set the node task load as the dividend and the current idle computing power quota parameter as the divisor, perform a division operation, and calculate the node computing power load ratio. Extract the preset penalty weights that characterize the sensitivity of computing power, and perform fixed-point multiplication of the node computing power load ratio with the preset penalty weights to calculate the local computing cost score of each independent processing node.

7. The intention-driven cloud-edge-device heterogeneous computing power collaborative scheduling method according to claim 1, characterized in that, Based on the expected computing power consumption, an estimated intent retention rate and a buffer overflow risk index are generated, including: Input the expected value of computing power consumption into a pre-set performance depreciation mapping matrix, and extract the intention feature depreciation coefficient corresponding to the current resource consumption state; The key feature bit data volume constituting the core intent data stream is obtained, and the key feature bit data volume is multiplied with the intent feature conversion coefficient to perform dimensionality reduction processing, and the estimated retained feature data volume under the current computing power conditions is calculated. The estimated retention feature data amount is used as the dividend and the key feature bit data amount is used as the divisor to perform a numerical division operation to calculate and generate the estimated intention retention rate. The expected value of computing power consumption is substituted into the preset queuing time estimation rule to perform resource mapping and deduction, and the expected cumulative processing latency parameter corresponding to the current node allocation architecture is generated. Obtain the video queue write rate parameter associated with the current terminal device, multiply and integrate the video queue write rate parameter with the expected processing cumulative latency parameter to obtain the queue prediction backlog increment parameter generated within the latency period; Extract the buffer occupancy data that is monitored synchronously during the video stream data acquisition stage, and add and merge the queue predicted backlog increment parameter with the buffer occupancy data to generate a buffer overflow risk index.

8. The intention-driven cloud-edge-device heterogeneous computing power collaborative scheduling method according to claim 7, characterized in that, If the buffer overflow risk index exceeds a preset risk threshold, reduce the feature retention range of the core intent data stream to update the expected computing power consumption, including: The risk overflow difference parameter is obtained by subtracting the buffer overflow risk index from the preset risk threshold. The risk spillover difference parameter is linked with the preset convergence adjustment mapping table for a correlation query and matching operation to extract the feature truncation compensation coefficient corresponding to the risk spillover difference parameter. Extract a preset activity threshold, and perform a multiplication amplification conversion process on the preset activity threshold and the feature truncation compensation coefficient to generate a dynamic intent truncation threshold for shrinking the target perception boundary; A secondary filtering screening is performed on the pixel data contained in the core intent data stream based on the dynamic intent truncation threshold, removing edge-related pixel clusters whose feature significance is less than the dynamic intent truncation threshold, thereby compressing the core intent data stream into a dimensionality-reduced core intent data stream. Read the amount of updated key feature bits corresponding to the core intent data stream of dimensionality reduction, and recalculate based on the amount of updated key feature bits, non-core background data stream and heterogeneous computing power status data to generate the expected value of updated computing power consumption. Extract the expected value of updated computing power consumption to generate state machine jump instructions. Based on the state machine jump instructions, trigger the system execution pointer to backtrack, and force the current round of computing process to return to the extraction and generation steps of the estimated intention retention rate and buffer overflow risk index before execution.

9. The intention-driven cloud-edge-device heterogeneous computing power collaborative scheduling method according to claim 7, characterized in that, When the buffer overflow risk index is no greater than the preset risk threshold and the estimated intention retention rate is less than the preset retention value, the specific operation logic is as follows: The retention rate difference parameter is calculated by subtracting the preset retention value as the minuend and the estimated intended retention rate as the subtrahend. The retention rate difference parameter is used to perform query matching operations on the pre-built computing power borrowing mapping table to determine the compensation computing power quota required to make up for the lack of current intent features; Based on the compensation computing power quota, the heterogeneous computing power status data is addressed and traversed for retrieval, parsing and suspending the secondary background execution tasks currently at the bottom layer, thereby extracting the secondary task computing power compensation parameters that have reached the compensation computing power quota indicators. Obtain the expected computing power consumption value within the current verification cycle, and perform an addition and superposition operation on the expected computing power consumption value and the computing power compensation parameter of the secondary task to obtain the updated expected computing power consumption value after superimposing the computing power margin. Based on the expected value of updated computing power consumption, process reverse backtracking control instructions are constructed and issued. Through process reverse backtracking control instructions, the execution pointer of the computing node is forced to jump, guiding the system to return to the processing stage of generating the estimated intention retention rate and buffer overflow risk index.

10. The intention-driven cloud-edge-device heterogeneous computing power collaborative scheduling method according to claim 7, characterized in that, The current expected computing power consumption is transformed into a scheduling decision matrix, including: Extract the convergence target signal generated by comparing the current buffer overflow risk index and the estimated intention retention rate, and capture and lock the current expected value of computing power consumption from the cache space based on the convergence target signal; The locked expected value of computing power consumption is substituted into the preset heterogeneous node mapping dimension space to decompose the data protocol and extract the distributed computing power quota vector of each independent processing node corresponding to different spatial levels. Extract the set of node topology coordinate parameters of the global computing network, and perform a two-dimensional arrangement, fusion and alignment operation on all the distributed computing power quota vectors according to the set of node topology coordinate parameters to construct a scheduling decision matrix that matches the current network physical structure.

11. An intent-driven cloud-edge-device heterogeneous computing power collaborative scheduling system, characterized in that, It is implemented based on the intent-driven cloud-edge-device heterogeneous computing power collaborative scheduling method according to any one of claims 1-10, including: Intent heatmap calculation module: acquires intent semantic vector, video stream data, buffer occupancy data, and heterogeneous computing power status data; when the buffer occupancy data is greater than a preset occupancy threshold, it performs calculations on the intent semantic vector and video stream data to generate an intra-frame intent heatmap; The computing power consumption calculation module: Based on the intra-frame intent heatmap, the video stream data is split into core intent data stream and non-core background data stream; based on the core intent data stream, non-core background data stream and heterogeneous computing power status data, the expected value of computing power consumption is estimated; The computing power iteration loop module inputs the expected value of computing power consumption into the game loop model for iteration. The game loop model includes the following iteration steps: Generate an estimated intent retention rate and a buffer overflow risk index based on the expected computing power consumption; If the buffer overflow risk index is greater than the preset risk threshold, reduce the feature retention range of the core intent data stream to update the expected value of computing power consumption, and return to the step of generating the estimated intent retention rate and buffer overflow risk index. If the buffer overflow risk index is not greater than the preset risk threshold and the estimated intent retention rate is less than the preset retention value, extract the secondary task computing power compensation parameters from the heterogeneous computing power status data to update the expected computing power consumption value, and return to the step of generating the estimated intent retention rate and the buffer overflow risk index. If the buffer overflow risk index is not greater than the preset risk threshold and the estimated intention retention rate is not less than the preset retention value, the current expected value of computing power consumption is converted into a scheduling decision matrix and the iteration is terminated. According to the scheduling decision matrix, the core intent data stream is processed with original-quality transmission, and the non-core background data stream is processed with dimensionality reduction and discarding.

Images