An ai optimization system and method of cloud-edge integrated distributed computing architecture

Abstract

The application relates to the technical field of distributed computing and artificial intelligence optimization, in particular to an AI optimization optimization system and method of a cloud-edge integrated distributed computing architecture, which comprises an environment perception unit, a global optimization unit, a gradient decision unit, a hierarchical execution unit and a compensation effective unit; the environment perception unit collects edge node network bandwidth; the global optimization unit aggregates historical data to generate a global data set with a topology label; a full-quantity gradient matrix is generated through deep reinforcement learning; the gradient decision unit dynamically retains key gradients according to bandwidth; residual gradient influence factors are calculated and edge re-optimization is triggered; the hierarchical execution unit executes local optimization by taking the residual gradient as a constraint; accurate solutions and space compensation values are generated by fusing global data; and the compensation effective unit issues compensation values to update the edge model. The system balances the efficiency and precision under dynamic bandwidth, reduces the cloud and edge deviation, and improves the distributed scene optimization reliability.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of distributed computing and artificial intelligence optimization, in particular to an AI optimization system and method of a cloud-edge integrated distributed computing architecture. BACKGROUND

[0002] Distributed computing and artificial intelligence optimization is an important technology. Under the background of the rapid development of Internet of Things and edge computing, the cloud-edge integrated architecture realizes efficient processing and intelligent decision-making of massive terminal data through the cooperation of cloud global planning and edge local processing, which is of great significance to improve computing resource utilization, reduce network delay and ensure real-time application experience. This technology is widely used in intelligent manufacturing, smart city and other fields. Traditional single cloud or edge computing mode has been difficult to meet the complex optimization needs of large-scale distributed scenarios.

[0003] However, the existing cloud-edge integrated computing architecture has the core problem of imbalance between computing efficiency and optimization accuracy in the AI optimization process. The traditional scheme adopts a fixed gradient transmission strategy and does not dynamically adjust the gradient data volume according to the real-time network bandwidth. When the bandwidth is insufficient, full gradient transmission can easily increase the delay and even cause data loss. Meanwhile, simple gradient reduction can reduce the optimization accuracy due to the lack of key information. In addition, there is a lack of effective use of residual gradient and edge local re-optimization mechanism, which causes deviation between cloud global optimization and edge local execution, further affecting the accuracy of the final decision. These problems superimpose to make the system have poor adaptability in dynamic network environment, which cannot efficiently cope with bandwidth fluctuations and cannot guarantee the reliability of the optimization result. It is difficult to meet the dual requirements of real-time and accuracy in large-scale distributed scenarios. In order to solve this problem, we provide an AI optimization system and method of a cloud-edge integrated distributed computing architecture. SUMMARY

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

[0005] 1. Since the traditional scheme uses fixed gradient transmission and does not adjust the data volume according to the real-time bandwidth, it causes delay or precision reduction. Therefore, the gradient decision unit of the present application dynamically calculates the gradient reservation quantity according to the bandwidth, reserves the key gradient, and balances the transmission efficiency and optimization accuracy.

[0006] 2. Since the traditional scheme lacks residual gradient utilization and edge re-optimization mechanism, it causes deviation between the cloud and the edge. Therefore, the hierarchical execution unit of the present application uses the residual gradient as a constraint for local optimization to generate a compensation value, which can reduce the deviation and improve the decision accuracy.

[0007] To achieve the above object, an AI optimization optimization system of a cloud-edge integrated distributed computing architecture is provided, comprising:

[0008] An environment perception unit is deployed on an edge computing node and collects network bandwidth values and computing load data of terminal equipment in real time through hardware sensors;

[0009] A global optimization unit is deployed on a cloud and aggregates historical running data of multiple edge computing nodes, generates a global data set containing cross-node associated features, and runs a deep reinforcement learning model for multi-objective iterative calculation to generate a full gradient matrix;

[0010] A gradient decision unit obtains the full gradient matrix of the current iteration period from the global optimization unit, calculates the gradient reservation number 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 reserved gradient set, N is the gradient reservation number, eliminates the reserved gradient set from the full gradient matrix, and the remaining gradients form a residual gradient set, calculates the ratio of the matrix norm of the residual gradient set to the full gradient matrix norm, and judges and outputs the edge re-optimization instruction according to the ratio;

[0011] A hierarchical execution unit receives the edge re-optimization 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 and the global data set to generate an accurate solution, and calculates a spatial compensation value according to the accurate solution and the suboptimal solution;

[0012] A compensation effective unit issues the spatial compensation value to the edge computing node for optimization optimization.

[0013] The second object of the present application is to provide a method for implementing an AI optimization optimization system of a cloud-edge integrated distributed computing architecture comprising any one of the above, comprising the following steps:

[0014] S1, the edge node collects the network bandwidth value of the terminal equipment in real time through the hardware sensor, the cloud receives the multi-node historical running data, eliminates the clock offset, extracts the cross-node task dependency chain and resource competition feature, and generates a global data set with topology label;

[0015] S2, generate a reserved gradient set based on the current bandwidth value dynamic mapping gradient reservation ratio, perform proximity clustering and reorganization on the remaining gradients to generate a spatial continuous residual set, and calculate a residual influence factor by hierarchical weighted fusion of full / residual norm, and output a re-optimization instruction by dynamically adjusting the threshold value combined with the load data;

[0016] S3, split the residual gradient set into constraint groups according to the spatial dimension, perform constraint satisfaction verification on the edge node to output a suboptimal solution, fuse the suboptimal solution and the global associated features, and generate an accurate solution that satisfies the global constraints through Monte Carlo random path search.

[0017] S4, the Euclidean distance of the accurate solution and the suboptimal solution in the vector space is calculated as a space compensation value, is issued to the edge node after scaling according to a preset learning rate, and is used for updating the local model parameter through gradient back propagation for optimization and optimization.

[0018] Compared with the prior art, the present application has the following beneficial effects:

[0019] 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 the bandwidth, establishes a mapping relationship with the gradient reservation ratio, 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 accuracy from being reduced caused by simple deletion by accurately retaining the core gradient, and realizes the dynamic balance of transmission efficiency and optimization accuracy under different bandwidth conditions.

[0020] 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 to generate a suboptimal solution at the edge node, and then generates an accurate solution by fusing the global data of the cloud and calculates the space compensation value. The residual information is fully utilized to fill the execution deviation between the cloud and the edge, which greatly improves the accuracy of the final decision and reduces the optimization error caused by gradient fragmentation.

[0021] 3. The time stamp calibration eliminates the node clock offset, extracts the cross-node task dependency and resource competition characteristics, generates a global data set with topology markers, provides comprehensive correlation information for optimization, dynamically adjusts the load and accuracy weight based on the Pareto frontier, adopts polarity block parallel computing to improve iteration efficiency, and enhances the adaptability of the model to complex distributed scenarios. At the same time, the space compensation value is issued to the edge to update the parameters by the compensation effective unit, forming a closed loop optimization, further ensuring the stability and reliability of the system in a dynamic network environment, meeting the dual demands of real-time and accuracy in intelligent manufacturing, smart city and other scenarios. BRIEF DESCRIPTION OF DRAWINGS

[0022] Figure 1 is the overall block diagram of the present application;

[0023] Figure 2 is the overall flowchart of the present application.

[0024] The meanings of the various labels in the figure are as follows:

[0025] 1. Environment perception unit; 2. Global optimization unit; 3. Gradient decision unit; 4. Hierarchical execution unit; 5. Compensation effective unit. DETAILED DESCRIPTION

[0026] The technical solutions in the embodiments of the present application will be clearly and completely described 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, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative work fall within the scope of the present application.

[0027] The present application provides an AI optimization system of a cloud-edge integrated distributed computing architecture. Please refer to Figure 1 as shown, comprising:

[0028] The environment perception unit 1 is deployed on the edge computing node and collects the network bandwidth value and the computing load data of the terminal device in real time through the hardware sensor.

[0029] The global optimization unit 2 is deployed on the cloud and aggregates the historical running data of multiple edge computing nodes to generate a global data set containing cross-node associated features, and runs a deep reinforcement learning model for multi-objective iterative calculation to generate a full gradient matrix.

[0030] When the global optimization unit 2 generates the global data set containing cross-node associated features, the spatio-temporal alignment feature fusion is performed:

[0031] After the environment 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 associated information, the global optimization unit 2 needs to generate a global data set containing cross-node associated features, and this process is realized through spatio-temporal alignment feature fusion. In order to ensure the security and integrity of data transmission, the historical running data of each edge computing node is received through the encryption channel of the cloud. These data include the computing load, task execution time, resource occupation of the node, etc., which are recorded in real time by the edge node and uploaded regularly. The encryption channel can prevent data from being tampered with or leaked during transmission, ensuring the reliability of the original data. Since the clocks of different edge nodes may be offset, resulting in non-uniformity of the time dimension of the data, the historical running data needs to be timestamp calibrated. Specifically, by synchronizing the unified time reference of each node, the timestamps of all data are converted to standard time in the same time zone, eliminating the clock offset between nodes. This can enable subsequent analysis to be performed in a unified time framework, avoiding misjudgment of associated features due to different time synchronization, and ensuring the time consistency of cross-node data. After calibration, the task dependency chain generated when multiple nodes cooperate to complete a task needs to be identified. The task dependency chain refers to the sequential dependency relationship formed when multiple edge nodes cooperate to complete the same task. For example, node A completes data preprocessing, then node B can perform model inference, and the output of node B is the premise for node C to execute decision-making. The specific identification process is as follows:

[0032] The calibrated historical operation data is traversed, the cross-node task records with the same task identification are extracted, and a directed acyclic graph is constructed according to the chronological order of the task start and end time and the correlation of the task input and output, such as the input of node B containing the output identification of node A. The nodes in the graph represent task steps, and the edges represent the dependency relationship between the steps, that is, the task dependency chain is formed. This operation can clearly present the cooperation process of the cross-node task and provide a structural basis for extracting resource competition and event association features. The resource competition coefficient between devices and the event trigger sequence are extracted from the task dependency chain. The resource competition coefficient is an index for measuring the degree of competition for the same type of resource by different nodes when performing dependent tasks. The specific calculation is as follows:

[0033] The number of requests and the number of conflicts for the same resource by the nodes participating in the dependency chain within the same time period are counted, and the ratio of the number of conflicts to the number of requests is the resource competition coefficient. The event trigger sequence is a sequence of key events (such as node A task completion triggering node B start, node C resource shortage triggering node A pause) in the task state change in the dependency chain in chronological order. By scanning the state log of the task dependency chain, the event type and trigger node information are extracted in chronological order to form a sequence. The resource competition 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 cooperation and help the model capture the cross-node event association rule. To reduce the data dimension and highlight the key features, the resource competition coefficient, the event trigger sequence, and the original historical operation data are dimensionally compressed. Specifically, the principal component analysis method is used:

[0034] The three types of data are converted into a unified dimensional feature vector. By calculating the characteristic value and the characteristic vector, the principal components with a variance contribution degree of the top 80% are selected as the compressed features to reduce the dimension. This operation can reduce the data amount while preserving the main information and improve the subsequent model calculation efficiency. According to the structural relationship of the task dependency chain, the compressed features are added with topological markers to generate a global data set with topological markers. The topological markers can preserve the spatial correlation information between nodes, so that the data set contains not only the compressed key features but also reflects the structural relationship between cross-node tasks, providing comprehensive association features for the deep reinforcement learning model and improving the accuracy and pertinence of global optimization.

[0035] When the global optimization unit 2 runs the deep reinforcement learning model for multi-objective iterative calculation, the Pareto gradient collaborative optimization is implemented:

[0036] After generating the global data set with topology labels, to achieve the balance and efficient calculation of multi-objective optimization, the global optimization unit 2 needs to implement the Pareto gradient collaborative optimization when running the deep reinforcement learning model. Pareto gradient collaborative optimization refers to the collaborative regulation of gradient updates based on the Pareto front in the multi-objective optimization process. That is, the optimal solution set in which multiple objectives cannot be improved simultaneously, and improving one objective will cause the deterioration of another objective. It ensures to find the optimal balance between multiple targets such as calculation load and optimization accuracy. The global data set is input into the deep reinforcement learning model. Since the model needs to consider both calculation load and optimization accuracy during iteration, and these two targets 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:

[0037] The model calculates the position of the current solution in the Pareto frontier every iteration, and if it finds that the response delay of some edge nodes is caused by too high calculation load, the precision weight is reduced and the calculation load weight is increased to prioritize system running efficiency, and if it finds that the optimization precision is insufficient and the deviation of the optimization result exceeds the allowed range, the precision weight is increased and the calculation load weight is reduced to prioritize result accuracy. Through this dynamic adjustment, the model can find an adaptive target priority in different scenarios, avoiding system imbalance caused by single target optimization. After determining the weight proportion of calculation load and precision, in order to improve the gradient update efficiency and ensure the synergy of multi-objective optimization, the polarity block parallel computing technology is used. Since the sign (positive or negative) of the gradient reflects the direction of parameter adjustment, the same sign gradient represents that the parameters need to be adjusted in the same direction and can be updated independently without affecting each other's optimization direction, and the different sign gradient represents that the parameters need to be adjusted in the opposite direction, and if updated alone, it may cause optimization target conflict. Therefore, the gradient is first classified according to the sign. Specifically, the gradient vectors of each layer output of the model are traversed, all positive gradients are classified into one category, and negative gradients are classified into another category. Then, the same sign gradients are aggregated into continuous storage blocks according to their positions in the matrix. For these same sign gradient blocks, an asynchronous update method is used, i.e. each block can be updated independently without waiting for other blocks to complete. This can make full use of parallel computing resources and speed up the update. For different sign gradients, a synchronous update method is used, i.e. all different sign gradients need to be calculated at the same time node and then updated uniformly to avoid parameter adjustment confusion caused by separate updates. For example, if the gradient of a certain parameter is positive and the gradient of another associated parameter is negative, synchronous updating can ensure that the adjustment amplitudes of the two parameters match each other and do not disrupt the balance between multiple targets. After each iteration, the updated gradient vectors of each layer are sequentially spliced in the order of network level from the input layer to the hidden layer and then to the output layer to form a complete gradient matrix, and a polarity label is marked for each gradient in the matrix, i.e. a full-amount gradient matrix with polarity labels is generated. The polarity label can clearly record the adjustment direction of the gradient, providing a directional basis for the gradient decision unit 3 to filter key gradients according to the bandwidth, ensuring that the retained gradients can not only reflect the core information of 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.

[0038] The gradient decision unit 3 obtains the full-amount gradient matrix of the current iteration period from the global optimization unit 2, calculates the number of gradients to be retained according to the network bandwidth value, sorts the gradients in the full-amount gradient matrix in descending order of absolute value, retains the first N gradients to form a retained gradient set, N is the number of gradients to be retained, removes the retained gradient set from the full-amount gradient matrix, and the remaining gradients form a residual gradient set. Calculate the ratio of the matrix norm of the residual gradient set to the matrix norm of the full-amount gradient matrix, and determine the output edge re-optimization instruction according to the ratio;

[0039] The gradient decision unit 3 performs bandwidth interval dynamic mapping when calculating the gradient reservation quantity according to the network bandwidth value:

[0040] 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:

[0041] 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 the transmission efficiency needs to be prioritized at low bandwidth and more gradient information can be considered at high bandwidth, 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 in which 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 the nearest integer. The gradient reservation quantity is obtained, 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. At low bandwidth, the amount of data transmitted is reduced to reduce latency, and at high bandwidth, more gradients are reserved to ensure optimization accuracy, achieving a dynamic balance between efficiency and accuracy and providing a reasonable quantity basis for subsequent gradient selection.

[0042] The gradient decision unit 3 generates the residual gradient set by implementing spatial continuity reorganization:

[0043] After the gradient decision unit 3 determines the gradient reservation quantity and generates the reserved gradient set, in order to avoid the influence of the scattered distribution of the remaining gradients on the efficiency of the subsequent edge re-optimization, spatial continuity reorganization needs to be performed on the residual gradient set, and the reserved gradient set is removed from the full-quantity gradient matrix, that is, the first N gradients with the largest absolute values, to obtain the remaining residual gradients. In order to grasp the distribution characteristics of these residual gradients, the spatial distribution state needs to be scanned, and the specific process is as follows:

[0044] The full-quantity gradient matrix is regarded as a two-dimensional grid, the rows represent the network levels, and the columns represent different parameter positions at the same level. Each residual gradient is marked with its physical coordinates in the grid, such as (3, 5) for the 3rd row and 5th column. By traversing the entire matrix through a sliding window (the window size is set to 3x3 grid), the number of residual gradients contained in each window is counted, and the connection between adjacent windows is recorded. If there are shared residual gradient coordinates between two windows, it is determined as a continuous block, otherwise it is a discrete block. In this way, the continuous regions and isolated points in the residual gradients are identified. This can accurately locate the spatial distribution form of the residual gradients, provide a basis for determining whether reorganization is needed, and ensure that no scattered gradient information is missed. A tolerance threshold is set, which is 15% of the total number of residual gradients, that is, up to 15% of the residual gradients are allowed to be in a discrete state. When the number of discrete gradient blocks exceeds the threshold, for example, there are 100 residual gradients in total 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:

[0045] Take each discrete residual gradient coordinate as a seed point, calculate the Manhattan distance between it and other residual gradients (such as the distance between coordinates (3, 5) and (3, 6) is 1, and the distance between (4, 5) is 1), and classify the gradients with a distance less than or equal to 2 into the same cluster, that is, allow adjacent or diagonal adjacent gradients to form a cluster. For isolated points that still exist after clustering, the distance between them and other gradients is greater than 2, then find the nearest continuous block around them, and distribute the gradient values of the isolated points to the adjacent blocks according to the distance weight through coordinate mapping, and finally form multiple spatially continuous gradient blocks. For example, the original residual gradients have discrete points at (2, 3), (2, 5), and (3, 4). After clustering, (2, 3) and (3, 4) are classified into one class because the distance between them is 2, and (2, 5) is assigned to the nearest (3, 4) block because it has no nearby neighbors. The reorganization forms a continuous region. This operation can integrate scattered residual gradients into continuous blocks, avoid calculation gaps when local optimization is performed on edge nodes due to scattered gradient distribution, improve the effectiveness of residual gradients as constraint terms, and through scanning the spatial distribution to identify discrete blocks and starting the clustering engine to reorganize continuous regions, the full information of residual gradients is preserved, and the spatial structure is optimized, so that the subsequent hierarchical execution unit 4 can more efficiently convert the residual gradients into constraint conditions, and enhance the accuracy and stability of local optimization of edges.

[0046] The gradient decision unit 3 calculates the norm ratio of the residual gradient set and the full-amount gradient matrix, and performs hierarchical weighted fusion:

[0047] After completing the spatial continuity reorganization of the residual gradient set, in order to quantify the degree of influence of the quantized residual gradient on global optimization, the gradient decision unit 3 needs to calculate the norm ratio of the residual gradient set and the full-amount gradient matrix through hierarchical weighted fusion, and then output the residual influence factor. The Frobenius norm of the full-amount gradient matrix of each convolution layer and the modified Manhattan norm of the corresponding layer of the residual gradient set need to be calculated respectively. The Frobenius norm is a norm used to measure the size of a matrix, and the calculation method is 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-amount gradient in that convolution layer. The modified Manhattan norm is an optimization of the traditional Manhattan norm (i.e. the sum of the absolute values of all elements in the matrix). By assigning spatial weights to elements at different positions, it better fits the continuous distribution characteristics of the residual gradient after spatial reorganization. When calculating, the absolute value of the elements in the residual gradient set is first taken, 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-amount gradient and the residual gradient from the overall scale and spatial distribution, respectively, providing basic data for subsequent fusion. Since the convolution layers gradually extract more abstract and critical features from shallow to deep, the influence of residual gradient in deep layers is often more significant than in shallow layers. Therefore, after obtaining the norms of each layer, fusion weights need to be assigned according to the depth of the convolution layer to achieve differentiated fusion of features at different levels. Specifically, for shallow convolution layers (1st-2nd layers), an equal weight strategy is adopted, i.e. the Frobenius norm of the full-amount gradient and the modified Manhattan norm of the residual gradient each account for 50% of the weight. This is because the shallow features are relatively basic and the influence of the residual gradient is balanced. For deep convolution layers (3rd layer and above), an exponentially increasing weight is adopted, with a base weight of 0.5 and an increase of 0.1 for each deeper layer (the residual norm weight of the 3rd layer is 0.6, the 4th layer is 0.7, and so on, with a maximum of 0.9). Through this weight allocation, the influence of deep residual gradient on global optimization can be highlighted, making the fusion result more consistent with the logical rules of feature extraction. After completing the weight allocation, the full-amount gradient norm and the residual gradient norm of each layer are weighted and summed, i.e. full-amount norm x (1-residual weight) + residual norm x residual weight, to obtain the fusion norm of each layer. The global full-amount norm total is obtained by summing the full-amount gradient Frobenius norms of all convolution layers, and the global fusion norm total is obtained by summing the fusion norms of all convolution layers. Finally, the ratio of the global fusion norm total to the global full-amount norm total is calculated, which is the residual influence factor. For example, if the global full-amount norm total is 100 and the global fusion norm total is 30, the residual influence factor is 0.3. By calculating the norm ratio in this way, the importance differences of features at different levels are considered, and the spatial continuity characteristics of the residual gradient are reflected through the modified Manhattan norm, making the output residual influence factor more accurately reflect the actual influence of the residual gradient on global optimization, providing reliable quantitative basis for subsequent judgment of whether to trigger edge re-optimization, effectively improving the scientificity and pertinence of gradient decision.

[0048] When the gradient decision unit 3 outputs the edge re-optimization instruction according to the ratio, it implements load adaptive decision:

[0049] After calculating the residual influence factor, in order to ensure that the output of the edge re-optimization instruction meets the influence 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, and real-time obtains the edge computing load data collected by the environment perception unit 1. These data include CPU utilization, memory occupancy, task queue length and other key indicators of the edge node, which can comprehensively reflect the current operation pressure of the edge node, provide a basis for subsequent load level division, and establish a negative coupling relationship between the load level and the trigger threshold. The specific process is:

[0050] The edge computing load is divided into three levels of low, medium and high (CPU utilization below 30% is low load, 30%-70% is medium load, and above 70% is high load), and the corresponding trigger threshold is matched for each load level (the initial trigger threshold is set to 0.5, which means that the residual influence factor needs to exceed this value to trigger re-optimization), and the negative coupling relationship is:

[0051] The higher the load level, the higher the trigger threshold. For example, when the load is high, the trigger threshold is adjusted to 0.7, when the load is medium, it remains 0.5, and when the load is low, it is adjusted to 0.3. This is because when the load of the edge node is high, triggering re-optimization easily may exacerbate the operation pressure, while when the load is low, the threshold can be relaxed to respond more sensitively to the residual influence. Through this dynamic adjustment, the trigger condition is adapted to the carrying capacity of the edge node, avoiding node overload caused by re-optimization operation. After establishing the negative coupling relationship, the residual influence factor calculated in the foregoing is compared with the dynamically adjusted trigger threshold, and the relationship between the two is:

[0052] The trigger threshold dynamically changes with the load level, and the residual influence factor reflects the influence degree of the residual gradient on the global optimization. Only when the residual influence 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 has the ability to bear the re-optimization operation. At this time, the edge re-optimization instruction is generated. Otherwise, if the factor does not exceed the threshold, it means that the residual influence is small or the edge node load is too high to perform re-optimization, and the cloud computing mode is maintained, that is, the cloud dominates the global optimization, and the edge node only executes the decision result issued by the cloud without local re-optimization. The edge re-optimization instruction specifically refers to the instruction indicating the edge node to start the local optimization program, including the residual gradient set, constraint conditions and other key information, which is used to trigger the hierarchical execution unit 4 to perform local optimization calculation based on the residual gradient as the constraint. After the instruction is generated, the edge node will call the local lightweight model to carry out optimization calculation based on the residual gradient, so as to compensate for the deviation between the cloud and the edge by using the residual information on the premise of considering the self-load, and output the instruction according to the comparison result by dynamically adjusting the threshold through real-time acquisition of load data and establishment of negative coupling relationship. It not only ensures that the edge re-optimization can be triggered in time when the residual influence is significant, but also avoids the situation that the edge node cannot undertake the re-optimization task due to high load, so that the whole decision-making process is more suitable for the actual running state of the edge node, and the adaptability and reliability of the system in the dynamic environment are further improved.

[0053] The gradient decision unit 3 performs gradient stability verification before outputting the instruction:

[0054] Before determining whether to output the edge re-optimization instruction according to the load adaptive decision, to avoid misjudgment of the decision due to abnormal fluctuations of the residual impact factor, the gradient decision unit 3 also needs to perform gradient stability verification. This step can further filter transient interference and ensure the reliability of the instruction output. The system automatically records the fluctuation trajectory of the residual impact factor in the historical iteration process, that is, stores the residual impact factor value obtained each time in the order of iteration periods to form a continuous change curve. By tracing the trajectory, the stability of the residual impact factor can be observed directly, and data basis for judging whether there is abnormal fluctuation is provided. A tolerance window is set (for example, the fluctuation amplitude of the residual impact factor in the last three iteration periods does not exceed ±0.1). When it is detected that the fluctuation amplitude exceeds the tolerance window (the residual impact factors of the last three iterations are 0.4, 0.6, and 0.3, and the fluctuation amplitude 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, the singular value decay analysis of the full gradient matrix needs to be started. The singular value decay analysis reflects the important characteristic components of the matrix by calculating the singular values of the full gradient matrix and observing the decay trend of the singular values with the number of iterations to judge the overall stability of the gradient matrix. If the singular values decay too quickly, it means that the gradient information is lost quickly, which may affect the optimization accuracy. When performing singular value decay analysis, a decay curve is generated, with the horizontal axis representing the number of iterations and the vertical axis representing the size of the singular value. A safety baseline is also preset. If the analysis finds that the slope of the decay curve exceeds the safety baseline (that is, the absolute value of the slope is greater than 0.05 / iteration), it indicates that the stability of the full gradient matrix has been destroyed, and the fluctuation of the residual impact factor is not accidental. At this time, regardless of the result of the load adaptive decision, the edge re-optimization instruction will be output to compensate for the deviation caused by the unstable gradient through local optimization of the edge. By recording the fluctuation trajectory, setting the tolerance window to trigger singular value analysis, and determining whether to output the instruction according to the slope of the decay curve, the gradient stability verification as a supplementary mechanism of the load adaptive decision can effectively identify the abnormal state of the gradient system and avoid optimization deviation caused by insufficient stability. The output of the edge re-optimization instruction takes into account both the load state and the overall reliability of the gradient system, further improving the logical closed loop of the gradient decision.

[0055] The hierarchical execution unit 4 receives the edge re-optimization 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 accurate solution, and calculates a spatial compensation value according to the accurate solution and the suboptimal solution.

[0056] When the hierarchical execution unit 4 inputs the residual gradient set as an inequality constraint term into the deep reinforcement learning model, constraint transformation and local optimization are performed.

[0057] After the gradient decision unit 3 outputs the edge re-optimization instruction, in order to fully utilize the residual gradient information and combine it with 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, to facilitate step-by-step verification of constraint satisfaction, it is divided into sub-constraint groups according to spatial dimensions. Specifically, a 3×3 grid is used as the basic unit, and the grid size can be adaptively adjusted according to the gradient matrix dimension. The residual gradients within each grid constitute a sub-constraint group. If a continuous block spans multiple grids, it is split according to the grid boundary and marked as an associated sub-group. This method of spatial grid division makes constraint verification more efficient. This targeted approach, along with the labeling of associated subgroups to avoid information fragmentation caused by splitting, represents an innovative optimization of traditional constraint grouping methods. It transforms each sub-constraint into inequality constraints. For a single subgroup, the modified Manhattan norm of the residual gradient within that group is used as the benchmark, resulting in "gradient deviation for local optimization ≤ benchmark value × 1.2" (where 1.2 is a tolerance coefficient allowing for a certain range of deviation). For associated subgroups, an additional cross constraint, "gradient deviation difference between subgroups ≤ associated benchmark value × 0.8," is added to ensure consistency of constraints in the associated region. This transformation converts the spatial characteristics of the residual gradient into quantifiable constraints, providing clear boundaries for local optimization.

[0058] Subsequently, a lightweight model is run on the edge nodes for local optimization computation. This lightweight model is a simplified version of the cloud-based deep reinforcement learning model, retaining the core optimization logic but reducing the number of network layers to minimize the objective function, such as the weighted sum of resource consumption and optimization bias. It iterative solutions are performed based on the aforementioned inequality constraints. After each iteration, the constraint satisfaction is verified by calculating the deviation between the current solution and each sub-constraint group. If the deviation of all individual sub-groups is ≤ the baseline value × 1.2, and the cross-deviation of related sub-groups is ≤ the associated baseline value × 0.8, then the constraints are considered satisfied; otherwise, the model parameters are adjusted and the iteration is repeated. This process ensures that local optimization always occurs within the residual gradient constraint range, avoiding solution divergence. When all constraint groups meet the tolerance conditions, the solution at this point is output as the suboptimal solution. This suboptimal solution reflects the optimal result of the edge node under local constraints, but may have limitations due to a lack of global information.

[0059] To integrate a global perspective, the suboptimal solution and cross-node association features (task dependency chains, resource contention coefficients) from the global dataset are input into the Monte Carlo sampling engine, which generates a candidate solution set through random path search.

[0060] The engine starts from a suboptimal solution, randomly generates 100 optimization paths within the range defined by the global characteristics, calculates the objective function value of each path, and selects the top 30% of the high-quality paths corresponding to the solution as the candidate solution. 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.

[0061] Subsequently, a candidate solution that satisfies all global constraints is selected from the candidate solution set as the accurate solution. The global constraints include conditions such as "cross-node resource conflict rate ≤ 5%" and "task response delay ≤ preset threshold" that reflect the overall optimization goal. The Euclidean distance of the accurate solution and the suboptimal solution in the vector space is calculated, which is the straight-line distance of the parameter vectors corresponding to the two solutions. This distance is the spatial compensation value, which reflects the degree of deviation between local optimization and global optimization. Through constraint transformation, the residual gradient is accurately applied. Local optimization combined with global sampling generates an accurate solution, and the compensation value is obtained by quantifying the deviation through the Euclidean distance. This not only takes advantage of the local computing power of the edge node, but also incorporates the global perspective of the cloud, providing a scientific basis for the subsequent compensation effective unit 5 to update the edge model, effectively reducing the optimization deviation between the cloud and the edge, and improving the overall optimization accuracy of the system.

[0062] The compensation effective unit 5 sends the spatial compensation value to the edge computing node for optimization.

[0063] In the present application, the environment 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 topology markers, and a full gradient matrix is generated through deep reinforcement learning. The gradient decision unit 3 dynamically retains key gradients according to bandwidth, calculates residual gradient influence factors and triggers edge re-optimization, the hierarchical execution unit 4 executes local optimization with residual gradient as a constraint, generates accurate solution and spatial compensation value by fusing global data, and the compensation effective unit 5 sends the compensation value to update the edge model. This 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.

[0064] The second purpose of the present application is to provide a method for implementing an AI optimization optimization system including the above-mentioned any one kind of cloud edge integrated distributed computing architecture, comprising the following steps:

[0065] S1, the edge node collects the network bandwidth value of the terminal device in real time through the hardware sensor, and the cloud receives the historical running data of multiple nodes, extracts the cross-node task dependency chain and resource competition characteristics after eliminating the clock offset, and generates a global data set with topology markers;

[0066] S2, based on the current bandwidth value dynamically mapping gradient reservation ratio to generate a reserved gradient set, performing proximity clustering and reorganization on the remaining gradients to generate a spatially continuous residual set, hierarchically weighting and fusing full residual norm to calculate residual influence factor, combining load data to dynamically adjust threshold to output re-optimization instructions;

[0067] S3, split the residual gradient set into sub-constraint groups according to spatial dimensions, perform constraint satisfaction verification on edge nodes to output suboptimal solutions, fuse suboptimal solutions with global correlation features, and generate accurate solutions that satisfy global constraints through Monte Carlo random path search;

[0068] S4, calculate the Euclidean distance of the accurate solution and the suboptimal solution in the vector space as the spatial compensation value, scale it by the preset learning rate after being sent to the edge node, and update the local model parameters through gradient back propagation for optimization.

[0069] The basic principles, main features and advantages of the present application are shown and described above. It should be understood by those skilled in the art that the present application is not limited by the above examples, and the above examples and descriptions in the specification are only preferred examples of the present application and are not intended to limit the present application. Without departing from the spirit and scope of the present application, various changes and improvements can be made to the present application, and these changes and improvements all fall within the scope of the present application. The scope of protection of the present application is defined by the appended claims and their equivalents.

Claims

1. An AI optimization optimization system of a cloud-edge integrated distributed computing architecture, characterized in that, The method comprises the following steps: An environment perception unit (1) is deployed at an edge computing node to collect network bandwidth values and computing load data of terminal devices in real time through hardware sensors; A global optimization unit (2) is deployed in the cloud to aggregate historical running data of multiple edge computing nodes, generate a global data set containing cross-node associated features, and run a deep reinforcement learning model for multi-objective iterative calculation to generate a full gradient matrix; A gradient decision unit (3) obtains the full gradient matrix of the current iteration period from the global optimization unit (2), calculates the gradient reservation number 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 reserved gradient set, removes the reserved gradient set from the full gradient matrix, and calculates the ratio of the matrix norm of the residual gradient set to the matrix norm of the full gradient matrix to determine whether to output an edge re-optimization instruction; A hierarchical execution unit (4) receives the edge re-optimization 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 accurate solution, and calculates the Euclidean distance of the accurate solution and the suboptimal solution in the vector space, i.e. the straight line distance of the parameter vectors corresponding to the two solutions, which is the spatial compensation value; A compensation enforcement unit (5) sends the spatial compensation value to the edge computing node for optimization. 2.The AI optimization system of the cloud-edge integrated distributed computing architecture of claim 1, wherein, When the global optimization unit (2) generates a global data set containing cross-node associated features, it performs spatio-temporal alignment feature fusion: The historical running data of each edge computing node is received through an encrypted channel in the cloud, and the historical running data is timestamp calibrated to eliminate clock offset between nodes. After calibration, the task dependency chain generated during cross-node device collaboration tasks is identified, the resource competition coefficient and event trigger sequence between devices are extracted, and the resource competition coefficient, event trigger sequence and original historical running data are dimensionally compressed to generate a global data set with topological markers. 3.The AI optimization system of the cloud-edge integrated distributed computing architecture of claim 2, wherein, When the global optimization unit (2) runs a deep reinforcement learning model for multi-objective iterative calculation, it implements Pareto gradient collaborative optimization: After inputting the global data set into the deep reinforcement learning model, the proportion of the calculation load weight and the precision weight is dynamically adjusted based on the Pareto frontier constraint, the polarity block parallel computing technology is used to aggregate gradients of the same sign into continuous storage blocks for asynchronous updating, and gradients of different signs are updated synchronously. After iteration, the gradient vectors are spliced in the order of network levels to generate a full gradient matrix with polarity labels.

4. The AI optimization system of claim 3, wherein, When the gradient decision unit (3) calculates the gradient reservation number according to the network bandwidth value, it performs bandwidth interval dynamic mapping: The network bandwidth value is divided into three dynamic intervals of low, medium and high, a corresponding relationship between the intervals and the gradient reservation proportion is established, a preset proportion coefficient is called according to the interval where 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 reservation number is output after taking the integer. 5.The AI optimization system of the cloud-edge integrated distributed computing architecture of claim 1, wherein, When the gradient decision unit (3) generates the residual gradient set, it implements spatial continuity reorganization: After removing the reserved gradient set from the full gradient matrix, the spatial distribution of the remaining gradients is scanned. When the number of discrete gradient blocks exceeds the tolerance threshold, the proximity clustering engine is started. The spatially continuous residual gradient set is generated by re-aggregating the gradients based on their physical location coordinates in the matrix.

6. The AI optimization system of claim 5, wherein, 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 convolution layer of the full gradient matrix and the modified Manhattan norm of the corresponding layer of the residual gradient set are calculated respectively. According to the depth of the convolution layer, the fusion weight is allocated. The shallow layer uses equal weight strategy, and the deep layer uses exponential increasing weight. The two layer norms are weighted and summed to calculate the ratio, and the residual influence factor is output.

7. The AI optimization system of claim 6, wherein, When the gradient decision unit (3) outputs the edge re-optimization instruction based on the ratio, it implements load adaptive decision: Real-time acquisition of edge computing load data collected by the environment perception unit (1), establishment of a negative coupling relationship between load level and trigger threshold, comparison of the residual influence factor with the dynamically adjusted threshold, generation of an edge re-optimization instruction when the factor is greater than the threshold, otherwise maintain the cloud computing mode.

8. The AI optimization system of claim 7, wherein, The gradient decision unit (3) performs gradient stability verification before outputting the instruction: Record the fluctuation trajectory of the residual influence factor in the history iteration. When the continuous fluctuation amplitude exceeds the tolerance window, start the singular value decay analysis of the full gradient matrix. If the decay curve slope exceeds the safety baseline, the edge re-optimization instruction is forced to output. 9.The AI optimization system of the cloud-edge integrated distributed computing architecture of claim 1, wherein, When the hierarchical execution unit (4) inputs the residual gradient set into the deep reinforcement learning model as an inequality constraint term, it performs constraint transformation and local optimization: Divide the residual gradient set into sub-constraint groups according to the spatial dimension, convert each group into an inequality constraint condition, run a lightweight model on the edge node to perform local optimization calculation, and verify the constraint satisfaction degree after each iteration. When all constraint groups meet the tolerance condition, output the suboptimal solution. Input the suboptimal solution and the cross-node associated features in the global data set into the Monte Carlo sampling engine to generate a candidate solution set through random path search. Select the candidate solution that satisfies all global constraints as the exact solution. Calculate the Euclidean distance between the exact solution and the suboptimal solution in the vector space, and output the spatial compensation value.

10. A method for implementing an AI optimization system comprising a cloud-edge integrated distributed computing architecture according to any one of claims 1-9. The steps include: S1, the edge node collects the network bandwidth value of the terminal device in real time through the hardware sensor, and the cloud receives the historical running data of multiple nodes, extracts the cross-node task dependency chain and resource competition features after removing the clock offset, and generates a global data set with topology markers; S2, generate a reserved gradient set based on the current bandwidth value dynamic mapping gradient reservation ratio, perform proximity clustering and reorganization on the remaining gradients to generate a spatially continuous residual set, calculate the residual influence factor by hierarchical weighted fusion of full / residual norm, and output the re-optimization instruction by dynamically adjusting the threshold combined with the load data; S3, split the residual gradient set into sub-constraint groups according to the spatial dimension, perform constraint satisfaction verification on the edge node to output the suboptimal solution, fuse the suboptimal solution and the global associated features, and generate the exact solution that satisfies the global constraints through Monte Carlo random path search; S4, calculate the Euclidean distance of the precise solution and the suboptimal solution in the vector space as the space compensation value, send it to the edge node after scaling by the preset learning rate, and update the local model parameters through gradient back propagation for optimization.

Images