AI optimization system and method for cloud-side integrated distributed computing architecture

Abstract

The invention relates to the technical field of distributed computing and artificial intelligence optimization, in particular to an AI optimization system and method for a cloud-edge integrated distributed computing architecture, and the system comprises an environment sensing unit, a global optimization unit, a gradient decision-making unit, a hierarchical execution unit and a compensation effective unit. The global optimization unit aggregates historical data to generate a global data set with a topological mark, a full gradient matrix is generated through deep reinforcement learning, the gradient decision unit dynamically retains a key gradient according to a bandwidth, calculates a residual gradient influence factor and triggers edge re-optimization, and the hierarchical execution unit executes local optimization by taking the residual gradient as a constraint. Global data are fused to generate a precise solution and a space compensation value, the compensation effective unit issues the compensation value to update the edge model, the efficiency and precision under dynamic bandwidth are balanced by the system, the cloud and edge deviation is reduced, and the distributed scene optimization reliability is improved.

Description

Technical Field

[0001] The present invention relates to the field of distributed computing and artificial intelligence optimization technology, and specifically to an AI optimization system and method for a cloud-edge integrated distributed computing architecture. Background Art

[0002] Distributed computing and artificial intelligence optimization is an important technology. Against the backdrop of the rapid development of the Internet of Things and edge computing, the cloud-edge integrated architecture achieves efficient processing and intelligent decision-making of massive terminal data through the collaboration of global coordination in the cloud and local processing at the edge nodes. It is of great significance to improving computing resource utilization, reducing network latency, and ensuring real-time application experience. This technology is widely used in smart manufacturing, smart cities and other fields. The traditional single cloud or edge computing model can no longer meet the complex optimization needs of large-scale distributed scenarios.

[0003] However, the existing cloud-edge integrated computing architecture has a core problem of imbalance between computing efficiency and optimization accuracy in the AI ​​optimization process. The traditional solution adopts a fixed gradient transmission strategy and does not dynamically adjust the amount of gradient data according to the real-time network bandwidth of the edge node. When the bandwidth is insufficient, the full gradient transmission is likely to lead to increased latency and even data loss. Simply pruning the gradient will reduce the optimization accuracy due to the lack of key information. At the same time, the lack of effective utilization of residual gradients and local re-optimization mechanism at the edge leads to deviations between global optimization on the cloud and local execution at the edge, further affecting the accuracy of the final decision. The combination of these problems leads to poor adaptability of the system in dynamic network environments. It can neither efficiently cope with bandwidth fluctuations nor ensure the reliability of the optimization results. It is difficult to meet the dual requirements of real-time and accuracy in large-scale distributed scenarios. To solve this problem, we provide an AI optimization system and method for a cloud-edge integrated distributed computing architecture. Summary of the Invention

[0004] The purpose of the present invention is to provide an AI optimization system and method for a cloud-edge integrated distributed computing architecture to solve the problems raised in the above background technology.

[0005] 1. Because traditional solutions use fixed gradient transmission and do not adjust data volume according to real-time bandwidth, resulting in delays or reduced accuracy, this case uses a gradient decision unit to dynamically calculate the number of retained gradients based on bandwidth, retaining key gradients and balancing transmission efficiency and optimization accuracy.

[0006] 2. Because traditional solutions lack residual gradient utilization and edge reoptimization mechanisms, resulting in deviations between the cloud and the edge, this case uses the residual gradient as a constraint through hierarchical execution units to perform local optimization and generate compensation values, which can reduce deviations and improve decision accuracy.

[0007] To achieve the above objectives, an AI optimization system with a cloud-edge integrated distributed computing architecture is provided, including: The environmental perception unit is deployed on the edge computing node and collects the network bandwidth value and computing load data of the terminal device in real time through hardware sensors; The global optimization unit is deployed in the cloud and aggregates historical operating data from multiple edge computing nodes to generate a global dataset containing cross-node correlation features. It then runs a deep reinforcement learning model for multi-objective iterative calculations to generate a full gradient matrix. The gradient decision unit obtains the full gradient matrix of the current iteration cycle from the global optimization unit, calculates the number of retained gradients based on the network bandwidth value, sorts the gradients in the full gradient matrix in descending order of absolute value, retains the first N gradients to form a retained gradient set, where N is the number of retained gradients, removes the retained gradient set from the full gradient matrix, and the remaining gradients form a residual gradient set. The ratio of the matrix norm of the residual gradient set to the matrix norm of the full gradient is calculated, and the output edge reoptimization instruction is determined based on the ratio. The hierarchical execution unit receives the edge reoptimization instruction, inputs the residual gradient set as an inequality constraint term into the deep reinforcement learning model, performs local optimization calculation and outputs a suboptimal solution, fuses the suboptimal solution with the global data set to generate an exact solution, and calculates a spatial compensation value based on the exact solution and the suboptimal solution; The compensation effectiveness unit sends the space compensation value to the edge computing node for optimization.

[0008] A second object of the present invention is to provide a method for implementing an AI optimization system including a cloud-edge integrated distributed computing architecture as described above, comprising the following steps: S1. Edge nodes collect the network bandwidth value of terminal devices in real time through hardware sensors. The cloud receives historical operation data of multiple nodes, removes clock offsets, extracts cross-node task dependency chains and resource competition characteristics, and generates a global dataset with topology labels. S2: Dynamically map the gradient retention ratio based on the current bandwidth value to generate a retained gradient set, perform proximity clustering on the remaining gradients to generate a spatially continuous residual set, calculate the residual impact factor by layered weighted fusion of the full amount / residual norm, and dynamically adjust the threshold based on the load data to output a re-optimization instruction; S3: Split the residual gradient set into sub-constraint groups according to the spatial dimension, perform constraint satisfaction verification at the edge nodes to output a suboptimal solution, fuse the suboptimal solution with the global correlation features, and generate an exact solution that satisfies the global constraints through Monte Carlo random path search; S4. Calculate the Euclidean distance between the exact solution and the suboptimal solution in the vector space as the spatial compensation value, send it to the edge node, and scale it according to the preset learning rate. Update the local model parameters through gradient backpropagation for optimization.

[0009] Compared with the prior art, the present application has the following beneficial effects: 1、The environment perception unit collects the edge node network bandwidth in real time, the gradient decision unit divides the bandwidth into three dynamic intervals of low, medium and high according to this, establishes a mapping relationship with the gradient reservation ratio, and retains the key gradient according to the gradient absolute value sorting and eliminates the redundant data. This mechanism not only avoids the delay and loss caused by full transmission when the bandwidth is insufficient, but also prevents the precision decline caused by simple reduction by accurately retaining the core gradient, so as to realize the dynamic balance of transmission efficiency and optimization precision under different bandwidth conditions.

[0010] 2、The gradient decision unit identifies the influence degree of residual gradient through norm ratio calculation, triggers the edge re-optimization instruction, and the hierarchical execution unit converts the residual gradient into a constraint term. The local optimization is performed at the edge node to generate a suboptimal solution, and then the accurate solution is generated by fusing the global data in the cloud to calculate the spatial compensation value. The execution deviation between the cloud and the edge is filled by fully utilizing the residual information, which greatly improves the accuracy of the final decision and reduces the optimization error caused by gradient fragmentation.

[0011] 3、The time stamp calibration is used to eliminate node clock offset, the cross-node task dependency and resource competition characteristics are extracted, the global data set with topology label is generated to provide comprehensive correlation information for optimization, the load and precision weight are dynamically adjusted based on the Pareto frontier, the polarity block parallel calculation is adopted to improve the iteration efficiency, and the adaptability of the model to complex distributed scenarios is enhanced. At the same time, the spatial compensation value is issued by the compensation effective unit to update the edge parameters, forming a closed loop optimization, further ensuring the stability and reliability of the system in the dynamic network environment, and meeting the dual demands of real-time and accuracy in intelligent manufacturing, smart city and other scenarios. BRIEF DESCRIPTION OF DRAWINGS

[0012] Figure 1 The overall block diagram of the present application is shown in the figure; Figure 2 The overall flowchart of the present application is shown in the figure.

[0013] The meanings of various labels in the figure are as follows: 1、environment perception unit; 2、global optimization unit; 3、gradient decision unit; 4、hierarchical execution unit; 5、compensation effective unit. DETAILED DESCRIPTION

[0014] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0015] The present invention provides an AI optimization system with cloud-edge integrated distributed computing architecture. Figure 1 As shown, including: Environmental perception unit 1, deployed at the edge computing node, collects network bandwidth and computing load data of terminal devices in real time through hardware sensors; Global optimization unit 2, deployed in the cloud, aggregates historical operating data from multiple edge computing nodes to generate a global dataset containing cross-node correlation features, and runs a deep reinforcement learning model for multi-objective iterative calculations to generate a full gradient matrix; When the global optimization unit 2 generates a global dataset containing cross-node correlation features, it performs spatiotemporal alignment feature fusion: After the environmental perception unit 1 collects the network bandwidth of the edge node, in order to ensure that the cloud can perform global optimization based on comprehensive and relevant information, the global optimization unit 2 needs to generate a global data set containing cross-node correlation features. This process is achieved through feature fusion with spatiotemporal alignment. To ensure the security and integrity of data transmission, the historical operation data of each edge computing node is received through the encrypted channel of the cloud. This data includes the computing load of the node, task execution time, resource occupancy, etc., which are recorded in real time by the edge node and uploaded regularly. The encrypted channel can prevent the data from being tampered with or leaked during transmission, ensuring the reliability of the original data. Since the clocks of different edge nodes may deviate, resulting in inconsistent data time dimensions, it is necessary to timestamp calibrate the historical operation data. Specifically, by synchronizing the unified time base of each node, the timestamps of all data are converted to the standard time of the same time zone, eliminating the clock offset between nodes. This allows subsequent analysis to be performed within a unified time framework, avoiding misjudgment of associated features due to time asynchrony, and ensuring the time consistency of cross-node data. After calibration, it is necessary to identify the task dependency chain generated when cross-node devices collaborate on tasks. The task dependency chain refers to the sequential dependency relationship formed when multiple edge nodes collaborate to complete the same task. For example, node B can only perform model inference after node A completes data preprocessing, and the output of node B is the prerequisite for node C to execute the decision. The specific identification process is as follows: Traverse the calibrated historical operation data and extract cross-node task records with the same task identifier. Based on the order of task start and end times and the correlation between task inputs and outputs, such as the input of node B containing the output identifier of node A, construct a directed acyclic graph. The nodes in the graph represent task steps, and the edges represent the dependencies between steps, forming a task dependency chain. This operation can clearly present the collaborative process of cross-node tasks and provide a structural basis for extracting resource competition and event correlation features. Extract the resource competition coefficient and event trigger sequence between devices from the task dependency chain. The resource competition coefficient is an indicator that measures the degree to which different nodes compete for the same type of resources when executing dependent tasks. The specific calculation is: The number of requests and conflicts for the same resource by nodes participating in the dependency chain within the same time period is counted. The ratio of the number of conflicts to the number of requests is the resource contention coefficient. The event trigger sequence is a chronological record of key events of task status changes in the dependency chain (e.g., the completion of node A's task triggers the start of node B, and insufficient resources at node C trigger the pause of node A). By scanning the status log of the task dependency chain and sorting by timestamp, event types and triggering node information are extracted to form a sequence. The resource contention coefficient can reflect the degree of resource conflict between nodes and provide a basis for optimizing resource allocation. The event trigger sequence can reflect the dynamic response relationship in task collaboration and help the model capture the event correlation rules across nodes. To reduce data dimensions and highlight key features, the resource contention coefficient, event trigger sequence, and original historical operation data are dimensional compressed using the principal component analysis method: The three types of data are converted into feature vectors of unified dimension. By calculating the eigenvalues ​​and eigenvectors, the principal components with the top 80% of variance contribution are selected as compressed features to achieve dimensionality reduction. This operation can reduce the amount of data while retaining the main information, improve the efficiency of subsequent model calculations, and add topological labels to the compressed features according to the structural relationship of the task dependency chain to generate a global data set with topological labels. The topological labels can retain the spatial correlation information between nodes, so that the data set contains both compressed key features and reflects the structural relationship across nodes, providing comprehensive correlation features for the deep reinforcement learning model and improving the accuracy and pertinence of global optimization.

[0016] When the global optimization unit 2 runs the deep reinforcement learning model for multi-objective iterative calculations, it implements Pareto gradient collaborative optimization: After generating a global dataset with topological labels, in order to achieve a balance and efficient calculation of multi-objective optimization, the global optimization unit 2 needs to implement Pareto gradient collaborative optimization when running the deep reinforcement learning model. Pareto gradient collaborative optimization refers to the coordinated regulation of gradient updates based on the Pareto front in the multi-objective optimization process. That is, multiple objectives cannot be improved at the same time, and improving one objective will lead to the deterioration of the optimal solution set of other objectives. It ensures that the optimal balance is found between multiple objectives such as computational load and optimization accuracy. The global dataset is input into the deep reinforcement learning model. Since the model needs to take into account both computational load and optimization accuracy during the iteration process, and these two objectives often conflict, it is necessary to dynamically adjust the weight ratio of the two based on the Pareto front constraint. The specific process is as follows: Each time the model iterates, it calculates the position of the current solution on the Pareto front. If it is found that the computational load is too high, causing response delays at some edge nodes, the accuracy weight is reduced and the computational load weight is increased, prioritizing system efficiency. If it is found that the optimization accuracy is insufficient and the deviation of the optimization result exceeds the allowable range, the accuracy weight is increased and the computational load weight is reduced, prioritizing result accuracy. Through this dynamic adjustment, the model can find the appropriate target priority in different scenarios and avoid system imbalance caused by single-target optimization. After determining the weight ratio of computational load to accuracy, polarity block parallel computing technology is required to improve gradient update efficiency and ensure the synergy of multi-objective optimization. Since the sign of the gradient (positive or negative) reflects the direction of parameter adjustment, gradients with the same sign indicate that the parameters need to be adjusted in the same direction and can be updated independently without affecting each other's optimization direction. Gradients with different signs indicate that the parameters need to be adjusted in opposite directions. If updated separately, it may lead to optimization target conflicts. Therefore, the gradients are first classified by sign. Specifically, the gradient vectors output by each layer of the model are traversed, all positive gradients are classified into one category and negative gradients into another category. Then, gradients with the same sign are aggregated according to their position in the matrix. For these consecutive storage blocks, an asynchronous update method is used. That is, each block can independently perform update calculations without waiting for other blocks to complete. This can fully utilize parallel computing resources and speed up the update speed. For gradients with different signs, since their adjustment directions are opposite, a synchronous update method is required. That is, all gradients with different signs must be calculated at the same time node before being updated uniformly to avoid parameter adjustment confusion caused by separate updates. For example, if the gradient of a parameter is positive and the gradient of another related parameter is negative, synchronous updating can ensure that the adjustment amplitudes of the two are matched and do not disrupt the balance between multiple objectives. After each iteration, the updated gradient vectors of each layer are concatenated in sequence according to the network layer, from the input layer to the hidden layer to the output layer, to form a complete gradient matrix. Each gradient in the matrix is ​​marked with its original sign, that is, a full gradient matrix with polarity labels is generated. The polarity labels can clearly record the adjustment direction of the gradient, providing a directional basis for the subsequent gradient decision unit 3 to screen key gradients according to the bandwidth. This ensures that the retained gradients can not only reflect the core information of the parameter adjustment, but also maintain the synergy of multi-objective optimization, thereby improving the iteration efficiency and optimization effect of the deep reinforcement learning model as a whole.

[0017] Gradient decision unit 3 obtains the full gradient matrix of the current iteration cycle from global optimization unit 2, calculates the number of retained gradients based on the network bandwidth value, sorts the gradients in the full gradient matrix in descending order of absolute value, retains the first N gradients to form a retained gradient set, where N is the number of retained gradients, removes the retained gradient set from the full gradient matrix, and the remaining gradients form a residual gradient set. The ratio of the matrix norm of the residual gradient set to the matrix norm of the full gradient is calculated, and the output edge reoptimization instruction is determined based on the ratio. The gradient decision unit 3 performs bandwidth interval dynamic mapping when calculating the gradient reservation quantity according to the network bandwidth value: After the global optimization unit 2 generates the full-quantity gradient matrix with polarity labels, in order to dynamically adjust the gradient transmission quantity according to the real-time network condition, the gradient decision unit 3 needs to perform bandwidth interval dynamic mapping to calculate the gradient reservation quantity. In order to accurately divide the network bandwidth state, the network bandwidth value is divided into three dynamic intervals of low, medium and high. The specific process is as follows: The network bandwidth fluctuation data in the last 5 minutes is collected by the environment perception unit 1, and the average value and standard deviation of the bandwidth in this period are calculated. The upper limit of the low bandwidth interval is the average value minus 1.5 times the standard deviation, and the value below this is low bandwidth. The lower limit of the high bandwidth interval is the average value plus 1.5 times the standard deviation, and the value above this is high bandwidth. The interval between the two is the medium bandwidth interval. This dynamic division method based on real-time fluctuation characteristics can avoid the problem that fixed thresholds cannot adapt to network fluctuations, making the interval division more in line with the actual bandwidth state. The corresponding relationship between the interval and the gradient reservation ratio is established. Considering that low bandwidth needs to prioritize transmission efficiency and high bandwidth can consider more gradient information, the low bandwidth interval corresponds to a gradient reservation ratio of 30%, i.e. 30% of the key gradients in the full-quantity gradient are reserved, the medium bandwidth interval corresponds to a reservation ratio of 60%, and the high bandwidth interval corresponds to a reservation ratio of 90%. At the same time, a preset ratio coefficient is associated with each interval. This corresponding relationship not only reflects the adaptation logic of bandwidth and gradient transmission quantity, but also provides a clear quantitative basis for subsequent calculations. Then, according to the interval where the current network bandwidth value is located, the preset ratio coefficient is called. Specifically, the gradient decision unit 3 receives the current bandwidth value from the environment perception unit 1 in real time, compares it with the threshold values of the three divided intervals, and if the current value is below the upper limit of the low bandwidth, it is determined to be in the low bandwidth interval and the 0.3 ratio coefficient is automatically called. If it is in the medium bandwidth range, the 0.6 coefficient is called. If it is higher than the lower limit of the high bandwidth, the 0.9 coefficient is called. This calling process is automatically completed through the binding mechanism of interval threshold and coefficient without human intervention, ensuring the real-time response of the calling process. The ratio coefficient called is multiplied by the total number of elements in the full-quantity gradient matrix, and the result is rounded to obtain the gradient reservation quantity. For example, if the full-quantity gradient matrix contains 1000 elements and the current bandwidth is in the medium bandwidth interval, 1000 x 0.6 = 600, i.e. 600 gradients are reserved. This way, the number of gradients reserved can be dynamically adjusted according to the bandwidth, reducing the amount of data transmitted to reduce latency in low bandwidth, and reserving more gradients to ensure optimization accuracy in high bandwidth, achieving a dynamic balance between efficiency and accuracy and providing a reasonable quantity basis for subsequent gradient selection.

[0018] The gradient decision unit 3 generates the residual gradient set by implementing spatial continuity reorganization: After the gradient decision unit 3 determines the number of retained gradients and generates a retained gradient set, in order to avoid the residual gradients from being scattered and affecting the efficiency of subsequent edge reoptimization, it is necessary to perform spatial continuity reorganization on the residual gradient set. The retained gradient set, that is, the first N gradients with the largest absolute values, is removed from the full gradient matrix to obtain the remaining residual gradients. In order to understand the distribution characteristics of these residual gradients, their spatial distribution state needs to be scanned. The specific process is as follows: The full gradient matrix is ​​considered as a two-dimensional grid, with rows representing network layers and columns representing different parameter positions at the same layer. The physical coordinates of each residual gradient in the grid are marked, such as (3,5) in the 3rd row and the 5th column. The entire matrix is ​​traversed by a sliding window (the window size is set to a 3×3 grid), and the number of residual gradients contained in each window is counted. At the same time, the connection between the residual gradients between adjacent windows is recorded. If two windows have shared residual gradient coordinates, they are judged as continuous blocks, otherwise they are discrete blocks. In this way, continuous areas and isolated points in the residual gradients are identified. This can accurately locate the spatial distribution of the residual gradients and provide a basis for determining whether reorganization is needed, ensuring that scattered gradient information is not missed. A tolerance threshold is set, which is 15% of the total number of residual gradients, that is, a maximum of 15% of the residual gradients are allowed to be in a discrete state. When the number of discrete gradient blocks found in the scan exceeds the threshold, for example, there are 100 residual gradients and 20 discrete blocks, which exceeds the 15% threshold, the proximity clustering engine is started for reorganization. The working process of the proximity clustering engine is as follows: Take each discrete residual gradient coordinate as the seed point, calculate its Manhattan distance with other residual gradients (such as the distance between coordinates (3,5) and (3,6) is 1, and the distance to (4,5) is 1), and classify gradients with a distance less than or equal to 2 into the same cluster, that is, allow gradients that are adjacent to each other or diagonally to form a cluster. For isolated points that still exist after clustering, if the distance to other gradients is greater than 2, find the nearest continuous block around it, and distribute the gradient value of the isolated point to the adjacent block according to the distance weight through coordinate mapping, and finally form multiple spatially continuous gradient blocks. For example, the original residual gradient has discrete points at (2,3), (2,5), and (3,4). After clustering, (2,3) and (3,4) are classified into one category with a distance of 2. (2,5) is assigned to the block where the nearest (3,4) is located because it has no nearby neighbors. They are reorganized into continuous areas. This operation can integrate scattered residual gradients into continuous blocks, avoiding computational faults during local optimization of edge nodes due to scattered gradient distribution, and improving the effectiveness of residual gradients as constraints. By scanning the spatial distribution to identify discrete blocks and starting the clustering engine to reorganize continuous areas, all the information of the residual gradient is retained and its spatial structure is optimized, so that the subsequent hierarchical execution unit 4 can more efficiently convert the residual gradient into constraint conditions, thereby enhancing the accuracy and stability of local optimization of the edge.

[0019] When the gradient decision unit 3 calculates the norm ratio of the residual gradient set to the full gradient matrix, it performs hierarchical weighted fusion: After completing the spatial continuity reorganization of the residual gradient set, in order to quantify the influence of the residual gradient on the global optimization, the gradient decision unit 3 needs to calculate the norm ratio of the residual gradient set and the full gradient matrix through hierarchical weighted fusion, and then output the residual influence factor. It is necessary to calculate the Frobenius norm of each convolution layer of the full gradient matrix and the modified Manhattan norm of the corresponding layer of the residual gradient set respectively. The Frobenius norm is a norm used to measure the size of the matrix. It is calculated as the square root of the sum of the squares of all elements in the matrix, which can reflect the overall energy scale of the full gradient in the convolution layer. The modified Manhattan norm is a traditional Manhattan norm (that is, the absolute value of all elements in the matrix). The optimization of the sum of the values ​​​​is carried out by assigning spatial weights to elements at different positions, which is more in line with the continuous distribution characteristics of the residual gradient after spatial reorganization. When calculating, the absolute value of the elements in the residual gradient concentration is first taken, and then multiplied by the corresponding spatial weight and summed to obtain the modified Manhattan norm. The calculation of these two norms can quantify the characteristics of the full gradient and the residual gradient from the perspective of overall scale and spatial distribution, respectively, and provide basic data for subsequent fusion. Since the convolution layer gradually extracts more abstract and more critical features from shallow to deep, the influence of the residual gradient in the deep layer is often more significant than that in the shallow layer. Therefore, after obtaining the norm of each layer, the fusion weight needs to be assigned according to the depth of the convolution layer to achieve different layers. Differentiation fusion of high-level features. Specifically, for shallow convolutional layers (layers 1-2), an equal weight strategy is adopted, that is, the Frobenius norm of the full gradient and the modified Manhattan norm of the residual gradient each account for 50% of the weight. This is because shallow features are relatively basic and the influence of residual gradients is more balanced. For deep convolutional layers (layer 3 and above), an exponentially increasing weight is adopted, setting the basic weight to 0.5, and increasing the weight by 0.1 for each deeper layer (the residual norm weight of the third layer is 0.6, the fourth layer is 0.7, and so on, and the maximum does not exceed 0.9). Through this weight distribution, the influence of deep residual gradients on global optimization can be highlighted, making the fusion effect more balanced. The result is more in line with the logical law of feature extraction. After completing the weight allocation, the full gradient norm and the residual gradient norm of each layer are weightedly summed, that is, the full norm × (1-residual weight) + residual norm × residual weight, to obtain the fusion norm of each layer. The full gradient Frobenius norm of all convolutional layers is summed to obtain the global full norm sum, and the fusion norm of all convolutional layers is summed to obtain the global fusion norm sum. Finally, the ratio of the global fusion norm sum to the global full norm sum is calculated. This ratio is the residual influence factor. For example, if the global full norm sum is 100 and the global fusion norm sum is 30, the residual influence factor is 0.3. By calculating the norm ratio in this hierarchical weighted fusion approach, we not only take into account the importance differences of features at different levels, but also reflect the spatial continuity characteristics of the residual gradient through the modified Manhattan norm. This enables the output residual influence factor to more accurately reflect the actual impact of the residual gradient on the global optimization, providing a reliable quantitative basis for subsequent judgment on whether to trigger edge reoptimization, and effectively improving the scientificity and pertinence of gradient decision-making.

[0020] When the gradient decision unit 3 outputs the edge reoptimization instruction based on the ratio, it implements the load adaptive decision: After calculating the residual impact factor, in order to ensure that the output of the edge reoptimization instruction is consistent with the impact degree of the residual gradient and adapts to the actual carrying capacity of the edge node, the gradient decision unit 3 needs to implement load adaptive decision-making and obtain the edge computing load data collected by the environment perception unit 1 in real time. These data include key indicators such as the CPU utilization rate, memory occupancy rate, and task queue length of the edge node. These indicators can fully reflect the current computing pressure of the edge node, provide a basis for subsequent load level classification, and establish a negative coupling relationship between the load level and the trigger threshold. The specific process is as follows: The edge computing load is divided into three levels: low, medium, and high (CPU utilization below 30% is low load, 30%-70% is medium load, and above 70% is high load). At the same time, a corresponding trigger threshold is matched for each load level (the initial trigger threshold is set to 0.5, indicating that the residual impact factor must exceed this value to trigger re-optimization). The negative coupling relationship is reflected as follows: The higher the load level, the higher the trigger threshold. For example, the trigger threshold is raised to 0.7 under high load, maintained at 0.5 under medium load, and lowered to 0.3 under low load. This is because when the edge node load is high, easily triggering reoptimization may increase computing pressure. When the load is low, the threshold can be relaxed to respond more sensitively to residual effects. Through this dynamic adjustment, the trigger condition is adapted to the carrying capacity of the edge node, avoiding node overload caused by reoptimization operations. After establishing the negative coupling relationship, the residual impact factor calculated above is compared with the dynamically adjusted trigger threshold. The relationship between the two is as follows: The trigger threshold changes dynamically with the load level, and the residual impact factor reflects the degree of influence of the residual gradient on the global optimization. Only when the residual impact factor exceeds the trigger threshold under the corresponding load level, it indicates that the influence of the residual gradient is significant enough and the edge node is capable of carrying the re-optimization operation. At this time, the edge re-optimization instruction is generated. On the contrary, if the factor does not exceed the threshold, it means that the residual impact is small or the edge node load is too high and it is not suitable for re-optimization. The cloud computing mode is maintained, that is, the cloud dominates the global optimization, and the edge node only executes the decision results issued by the cloud, and does not perform local re-optimization. The edge re-optimization instruction specifically refers to the instruction to instruct the edge node to start the local optimization program, which includes the residual gradient set and constraint conditions. Key information such as the residual gradient is used to trigger the hierarchical execution unit 4 to perform local optimization calculations based on the residual gradient as a constraint. When the instruction is generated, the edge node will call the local lightweight model and perform optimization calculations based on the residual gradient. Under the premise of taking into account its own load, the residual information is used to make up for the deviation between the cloud and the edge. By obtaining load data in real time, establishing a negative coupling relationship to dynamically adjust the threshold, and outputting instructions based on the comparison results, it is ensured that the edge re-optimization can be triggered in time when the residual impact is significant, and the situation where the edge node is unable to undertake the re-optimization task due to excessive load is avoided. The entire decision-making process is more in line with the actual operating status of the edge node, further improving the adaptability and reliability of the system in a dynamic environment.

[0021] The gradient decision unit 3 performs a gradient stability check before outputting the instruction: Before determining whether to output the edge reoptimization instruction based on the load adaptive decision, in order to avoid misjudgment due to abnormal fluctuations in the residual influence factor, the gradient decision unit 3 also needs to perform a gradient stability check. This link can further filter out instantaneous interference and ensure the reliability of the instruction output. The system will automatically record the fluctuation trajectory of the residual influence factor during the historical iteration process, that is, store the residual influence factor value obtained each time in the order of the iteration cycle to form a continuous change curve. By tracing this trajectory, the stability of the residual influence factor can be intuitively observed, providing a data basis for judging whether there is abnormal fluctuation and setting a tolerance window (for example, the residual influence factor is within 3 consecutive iteration cycles). The fluctuation range of the impact factor does not exceed ±0.1). When it is detected that the continuous fluctuation range exceeds this tolerance window (the residual impact factors of three consecutive iterations are 0.4, 0.6, and 0.3 respectively, and the fluctuation range exceeds ±0.1), it indicates that the influence state of the residual gradient is unstable and there may be potential changes that have not been captured. At this time, it is necessary to start the singular value decay analysis of the full gradient matrix. The singular value decay analysis is to judge the overall stability of the gradient matrix by calculating the singular values ​​of the full gradient matrix, reflecting the important characteristic components of the matrix, and observing its decay trend with the number of iterations. If the singular value decays too quickly, it means that the gradient information loss is accelerated, which may affect the optimization accuracy. During singular value decay analysis, a decay curve is generated, with the number of iterations on the horizontal axis and the singular value size on the vertical axis. A safety baseline is also preset. If the analysis reveals that the slope of the decay curve exceeds the safety baseline (i.e., the absolute value of the slope is greater than 0.05 / iteration), the stability of the full gradient matrix has been compromised and the fluctuations in the residual influencing factors are not accidental. In this case, regardless of the load adaptive decision result, edge reoptimization instructions are forcibly issued to compensate for the deviation caused by gradient instability through local edge optimization. By recording fluctuation trajectories, setting a tolerance window to trigger singular value analysis, and determining whether to force instructions based on the slope of the decay curve, gradient stability verification, as a supplementary mechanism for load adaptive decision-making, effectively identifies abnormal conditions in the gradient system and avoids optimization deviations caused by insufficient stability. This ensures that the output of edge reoptimization instructions not only takes into account the load status but also ensures the overall reliability of the gradient system, further improving the logical closed loop of gradient decision-making.

[0022] The hierarchical execution unit 4 receives the edge reoptimization instruction, inputs the residual gradient set as the inequality constraint term into the deep reinforcement learning model, performs local optimization calculation and outputs a suboptimal solution, fuses the suboptimal solution with the global data set to generate an exact solution, and calculates the spatial compensation value based on the exact solution and the suboptimal solution; When the hierarchical execution unit 4 inputs the residual gradient set as an inequality constraint term into the deep reinforcement learning model, it performs constraint transformation and local optimization: After the gradient decision unit 3 outputs the edge reoptimization instruction, in order to make full use of the residual gradient information and combine the global features to generate accurate optimization results, the hierarchical execution unit 4 needs to perform constraint transformation and local optimization. Since the residual gradient set forms multiple continuous blocks after spatial continuity reorganization, in order to facilitate step-by-step verification of constraint satisfaction, it is divided into sub-constraint groups according to the spatial dimension. Specifically, the 3×3 grid is used as the basic unit. The grid size can be adaptively adjusted according to the gradient matrix dimension. The residual gradient in each grid constitutes a sub-constraint group. If a continuous block spans multiple grids, it is split according to the grid boundary and marked as an associated subgroup. This method of dividing by spatial grids can make constraint verification more efficient. The targeted labeling of associated subgroups can avoid information fragmentation caused by splitting. It is an innovative optimization of the traditional constraint grouping method. Each group of sub-constraints is converted into inequality constraints. For a single subgroup, the modified Manhattan norm of the residual gradient within the group is used as the benchmark value, which is converted into "gradient deviation of local optimization ≤ benchmark value × 1.2" (1.2 is the tolerance coefficient, allowing a certain range of deviation); for associated subgroups, an additional cross-constraint of "gradient deviation difference between subgroups ≤ associated benchmark value × 0.8" is added to ensure the constraint consistency of the associated area. Through this conversion, the spatial characteristics of the residual gradient are converted into quantifiable constraints, providing clear boundaries for local optimization.

[0023] Afterward, a lightweight model is run on the edge node for local optimization. This lightweight model is a simplified version of the cloud-based deep reinforcement learning model. It retains the core optimization logic but reduces the number of network layers. The goal is to minimize an objective function, such as the weighted sum of resource consumption and optimization deviation. Based on the aforementioned inequality constraints, an iterative solution is performed. After each iteration, the constraint satisfaction is verified by calculating the deviation between the current solution and each sub-constraint group. If the deviations of all individual sub-groups are ≤ the benchmark value × 1.2, and the cross-deviations of the associated sub-groups are ≤ the associated benchmark value × 0.8, the constraints are considered satisfied. Otherwise, the model parameters are adjusted and the iteration is repeated. This process ensures that the local optimization is always carried out within the residual gradient constraints, preventing solution divergence. When all constraint groups meet the tolerance conditions, the solution at that time is output as the suboptimal solution. This suboptimal solution reflects the optimal result of the edge node under the local constraints, but may have limitations due to the lack of global information.

[0024] To integrate the global perspective, the suboptimal solution and the cross-node correlation features (task dependency chain, resource competition coefficient) in the global dataset are input into the Monte Carlo sampling engine, and a candidate solution set is generated through random path search: The engine takes the suboptimal solution as the starting point and randomly generates 100 optimization paths within the range limited by global features. Each path corresponds to a set of parameter adjustment plans, calculates the objective function value of each path, and selects the solutions corresponding to the top 30% of high-quality paths as candidate solutions. This process can expand the optimization space based on the suboptimal solution and introduce global correlation information to make up for the shortcomings of local optimization.

[0025] Subsequently, a candidate solution that satisfies all global constraints is selected from the candidate solution set as the exact solution. Global constraints include conditions reflecting the overall optimization goal, such as "cross-node resource conflict rate ≤ 5%" and "task response delay ≤ preset threshold." The Euclidean distance between the exact solution and the suboptimal solution in vector space is calculated. This distance, that is, the straight-line distance between the corresponding parameter vectors of the two solutions, is the spatial compensation value, and its magnitude reflects the degree of deviation between local optimization and global optimization. Constraint transformation enables the precise application of residual gradients. Local optimization combines global sampling to generate an exact solution, and the deviation is then quantified using Euclidean distance to obtain the compensation value. This not only leverages the local computing advantages of edge nodes but also incorporates the global perspective of the cloud. This provides a scientific basis for the subsequent compensation validation unit 5 to update the edge model, effectively reducing the optimization deviation between the cloud and edge, and improving the overall optimization accuracy of the system.

[0026] The compensation validation unit 5 sends the space compensation value to the edge computing node for optimization.

[0027] In the present invention, the environmental perception unit 1 collects the network bandwidth of the edge node, the global optimization unit 2 aggregates historical data to generate a global data set with topological labels, generates a full gradient matrix through deep reinforcement learning, the gradient decision unit 3 dynamically retains the key gradient according to the bandwidth, calculates the residual gradient influencing factor and triggers edge reoptimization, the hierarchical execution unit 4 uses the residual gradient as a constraint to perform local optimization, fuses the global data to generate an accurate solution and a spatial compensation value, and the compensation effectiveness unit 5 sends the compensation value to update the edge model. The system balances efficiency and accuracy under dynamic bandwidth, reduces the deviation between the cloud and the edge, and improves the reliability of distributed scene optimization.

[0028] A second object of the present invention is to provide a method for implementing an AI optimization system including any one of the above-mentioned cloud-edge integrated distributed computing architectures, comprising the following steps: S1. Edge nodes collect the network bandwidth value of terminal devices in real time through hardware sensors. The cloud receives historical operation data of multiple nodes, removes clock offsets, extracts cross-node task dependency chains and resource competition characteristics, and generates a global dataset with topology labels. S2: Dynamically map the gradient retention ratio based on the current bandwidth value to generate a retained gradient set, perform proximity clustering on the remaining gradients to generate a spatially continuous residual set, calculate the residual impact factor by layered weighted fusion of the full amount / residual norm, and dynamically adjust the threshold based on the load data to output a re-optimization instruction; S3: Split the residual gradient set into sub-constraint groups according to the spatial dimension, perform constraint satisfaction verification at the edge nodes to output a suboptimal solution, fuse the suboptimal solution with the global correlation features, and generate an exact solution that satisfies the global constraints through Monte Carlo random path search; S4. Calculate the Euclidean distance between the exact solution and the suboptimal solution in the vector space as the spatial compensation value, send it to the edge node, and scale it according to the preset learning rate. Update the local model parameters through gradient backpropagation for optimization.

[0029] The above shows and describes the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The above embodiments and descriptions are merely preferred examples of the present invention and are not intended to limit the present invention. Various changes and improvements may be made to the present invention without departing from the spirit and scope of the present invention. Such changes and improvements fall within the scope of the present invention. The scope of protection claimed in the present invention is defined by the appended claims and their equivalents.

Claims

1. An AI optimization system with a cloud-edge integrated distributed computing architecture, characterized by: include: The environment perception unit (1) is deployed at the edge computing node and collects the network bandwidth value and computing load data of the terminal device in real time through hardware sensors; The global optimization unit (2) is deployed in the cloud, aggregates the historical operation data of multiple edge computing nodes, generates a global data set containing cross-node correlation features, and runs the deep reinforcement learning model to perform multi-objective iterative calculations to generate a full gradient matrix; The gradient decision unit (3) obtains the full gradient matrix of the current iteration cycle from the global optimization unit (2), calculates the number of gradient retention according to the network bandwidth value, sorts the gradients in the full gradient matrix in descending order of absolute value, retains the first N gradients to form a retained gradient set, where N is the number of gradient retention, removes the retained gradient set from the full gradient matrix, and the remaining gradients form a residual gradient set. The ratio of the matrix norm of the residual gradient set to the matrix norm of the full gradient is calculated, and the output edge reoptimization instruction is judged based on the ratio; The hierarchical execution unit (4) receives the edge reoptimization instruction, inputs the residual gradient set as an inequality constraint term into the deep reinforcement learning model, performs local optimization calculation and outputs a suboptimal solution, fuses the suboptimal solution with the global data set to generate an exact solution, and calculates a spatial compensation value based on the exact solution and the suboptimal solution; The compensation validation unit (5) sends the spatial compensation value to the edge computing node for optimization.

2. The AI ​​optimization system for cloud-edge integrated distributed computing architecture according to claim 1 is characterized in that: When the global optimization unit (2) generates a global dataset containing cross-node correlation features, it performs spatiotemporal alignment feature fusion: The historical operation data of each edge computing node is received through an encrypted channel in the cloud, and the timestamp of the historical operation data is calibrated to eliminate the clock offset between nodes. After the calibration is completed, the task dependency chain generated when cross-node devices collaborate on tasks is identified, and the resource competition coefficient and event trigger sequence between devices are extracted. The resource competition coefficient, event trigger sequence and original historical operation data are dimensionally compressed to generate a global data set with topology labels.

3. The AI ​​optimization system for cloud-edge integrated distributed computing architecture according to claim 2 is characterized in that: When the global optimization unit (2) runs the deep reinforcement learning model to perform multi-objective iterative calculations, it implements Pareto gradient collaborative optimization: After the global dataset is input into the deep reinforcement learning model, the ratio of the computational load weight to the precision weight is dynamically adjusted based on the Pareto front constraint. Polarity block parallel computing technology is used to aggregate gradients with the same sign into continuous storage blocks for asynchronous update, and gradients with different signs are updated synchronously. After the iteration, the gradient vectors are spliced ​​in the order of the network layers to generate a full gradient matrix with polarity labels.

4. The AI ​​optimization system for cloud-edge integrated distributed computing architecture according to claim 3 is characterized in that: When the gradient decision unit (3) calculates the number of gradient reserves according to the network bandwidth value, it performs dynamic bandwidth interval mapping: The network bandwidth value is divided into three dynamic intervals: low, medium, and high. A corresponding relationship between the interval and the gradient retention ratio is established. The preset ratio coefficient is called according to the interval in which the current network bandwidth value is located. The coefficient is multiplied by the total number of elements in the full gradient matrix, and the gradient retention number is output after rounding.

5. The AI ​​optimization system for cloud-edge integrated distributed computing architecture according to claim 1 is characterized in that: When the gradient decision unit (3) generates the residual gradient set, spatial continuity reorganization is performed: After removing the retained gradient set from the full gradient matrix, the spatial distribution of the remaining gradients is scanned. When it is detected that the discrete gradient block exceeds the tolerance threshold, the proximity clustering engine is started to re-aggregate the gradient according to its physical position coordinates in the matrix to generate a spatially continuous residual gradient set.

6. The AI ​​optimization system for cloud-edge integrated distributed computing architecture according to claim 5 is characterized in that: When the gradient decision unit (3) calculates the norm ratio of the residual gradient set to the full gradient matrix, it performs hierarchical weighted fusion: The Frobenius norm of each convolutional layer of the full gradient matrix and the modified Manhattan norm of the corresponding layer of the residual gradient set are calculated respectively. The fusion weight is assigned according to the depth of the convolutional layer. The shallow layer adopts an equal weight strategy, and the deep layer adopts an exponentially increasing weight. The ratio of the two layers is calculated after weighted summation, and the residual influence factor is output.

7. The AI ​​optimization system for cloud-edge integrated distributed computing architecture according to claim 6 is characterized in that: The gradient decision unit (3) implements load adaptive decision when outputting edge reoptimization instructions based on the ratio judgment: The edge computing load data collected by the environmental perception unit (1) is acquired in real time, a negative coupling relationship between the load level and the trigger threshold is established, the residual impact factor is compared with the threshold after dynamic adjustment, and an edge reoptimization instruction is generated when the factor is greater than the threshold, otherwise the cloud computing mode is maintained.

8. The AI ​​optimization system for cloud-edge integrated distributed computing architecture according to claim 7 is characterized in that: The gradient decision unit (3) performs a gradient stability check before outputting an instruction: Record the fluctuation trajectory of the residual influencing factors in historical iterations. When the continuous fluctuation amplitude exceeds the tolerance window, start the singular value attenuation analysis of the full gradient matrix. If the slope of the attenuation curve exceeds the safety baseline, force the output of the edge reoptimization instruction.

9. The AI ​​optimization system for cloud-edge integrated distributed computing architecture according to claim 1 is characterized in that: When the hierarchical execution unit (4) inputs the residual gradient set as an inequality constraint term into the deep reinforcement learning model, it performs constraint transformation and local optimization: The residual gradient set is divided into sub-constraint groups according to the spatial dimension. Each group is converted into an inequality constraint condition. A lightweight model is run on the edge node for local optimization calculation. The constraint satisfaction is verified after each iteration. When all constraint groups meet the tolerance conditions, a suboptimal solution is output. The suboptimal solution and the cross-node correlation features in the global dataset are input into the Monte Carlo sampling engine. A candidate solution set is generated through random path search. The candidate solution that meets all global constraints is selected as the exact solution. The Euclidean distance between the exact solution and the suboptimal solution in the vector space is calculated, and the spatial compensation value is output.

10. A method for implementing an AI optimization system comprising a cloud-edge integrated distributed computing architecture as described in any one of claims 1 to 9, characterized in that: The following steps are involved: S1. Edge nodes collect the network bandwidth value of terminal devices in real time through hardware sensors. The cloud receives historical operation data of multiple nodes, removes clock offsets, extracts cross-node task dependency chains and resource competition characteristics, and generates a global dataset with topology labels. S2: Dynamically map the gradient retention ratio based on the current bandwidth value to generate a retained gradient set, perform proximity clustering on the remaining gradients to generate a spatially continuous residual set, calculate the residual impact factor by layered weighted fusion of the full amount / residual norm, and dynamically adjust the threshold based on the load data to output a re-optimization instruction; S3: Split the residual gradient set into sub-constraint groups according to the spatial dimension, perform constraint satisfaction verification at the edge nodes to output a suboptimal solution, fuse the suboptimal solution with the global correlation features, and generate an exact solution that satisfies the global constraints through Monte Carlo random path search; S4. Calculate the Euclidean distance between the exact solution and the suboptimal solution in the vector space as the spatial compensation value, send it to the edge node, and scale it according to the preset learning rate. Update the local model parameters through gradient backpropagation for optimization.

Images