Soft clustering and dynamic constraint based edge computing task offloading method and system

Abstract

The application discloses an edge computing task offloading method and system based on soft clustering and dynamic constraints, and relates to the technical field of edge computing. The method comprises the following steps: clustering robots, constructing a membership matrix, and mapping the cluster center back to the individual decision space of the robot; an evolutionary iteration cycle is performed, in which, during the cycle, a dynamic constraint relaxation factor is determined according to the real-time proportion of feasible solutions, the dynamic constraint relaxation factor is set to 0 for safety level constraints and deadline constraints, a constraint violation degree vector is calculated for the candidate solution, and the candidate solution is screened according to the feasibility boundary box; the candidate solution is Pareto sorted; a safety level index is calculated, and when a preset safety threshold is met, a Pareto optimal solution set is extracted; and a task offloading scheme represented by the Pareto optimal solution set is executed. The application solves the problem that the traditional method is completely ineffective when the feasible solution density is less than 10 ‑15 , ensures 100% safety compliance rate, improves search efficiency, and solves the dimension disaster problem.

Description

Technical Field

[0001] This application relates to the field of edge computing technology, and in particular to a method for searching task offloading schemes in edge computing. Background Technology

[0002] Edge computing task offloading is a technology that offloads compute-intensive tasks from resource-constrained user devices to edge servers or cloud servers for execution. By processing tasks closer to the data source, it significantly reduces end-to-end latency, decreases local energy consumption, and improves user experience. In critical applications such as smart cities, Industry 4.0, autonomous driving, and smart healthcare, edge computing has become a key infrastructure supporting real-time performance, security, and reliability.

[0003] Specifically, in the scenario of collaborative control of industrial robots in an automobile assembly workshop, the characteristics of this scenario are: a large number of robots (50-500 units), high real-time requirements (latency ≤50ms), strict security requirements (security level ≥ advanced encryption standard AES-256), and complex constraint coupling (multiple constraints such as resource capacity, cache limit, deadline, and security compliance).

[0004] The existing technology has the following problems in the scenario of collaborative control of industrial robots in automobile assembly workshops.

[0005] Experimental setup: A car assembly workshop contains 150 industrial robots that need to unload motion control tasks to 15 edge nodes in real time. The constraints are: latency ≤ 50ms, security level ≥ AES-256, CPU (central processing unit) capacity ≤ 50GHz, and cache capacity ≤ 32GB.

[0006] Experimental Results: When the number of robots K=50, the safety compliance rate of NSGA-III (third-generation non-dominated sorting genetic algorithm) is 15%, the HV (oversize) value is 1.74±1.83, and 85% of the solutions violate safety requirements. When the number of robots K=100, the safety compliance rate of NSGA-III is 0%, the HV value is 0.00±0.00, and no feasible solution can be found. When the number of robots K=150, the safety compliance rate of NSGA-III is 0%, the HV value is 0.00±0.00, the running time is 118.43±0.54s, but no feasible solution is found.

[0007] It can be seen that NSGA-III performs well in extremely sparse feasible regions (feasible solution density is less than 10^6). -15 It is completely ineffective and cannot guarantee the safety requirements of the industrial site, posing a significant safety hazard. Summary of the Invention

[0008] This application provides a method and system for offloading edge computing tasks based on soft clustering and dynamic constraints to solve the problems in the prior art.

[0009] On the one hand, embodiments of this application provide an edge computing task offloading method based on soft clustering and dynamic constraints, including:

[0010] Cluster the K robots into G clusters, construct a K×G membership matrix, optimize the decision variables of the G cluster centers, and use the membership matrix to demap the cluster centers back to the individual decision space of the K robots.

[0011] An evolutionary iterative loop is performed in the individual decision space. During the loop, the dynamic constraint relaxation factor is determined based on the real-time proportion of feasible solutions in the population. For safety level constraints and deadline constraints, the dynamic constraint relaxation factor is set to 0. For candidate solutions obtained through mutation and crossover operations, the corresponding constraint violation vector is calculated. Candidate solutions are screened based on the feasibility bounding box and the constraint violation vector. Constraints are divided into three priorities according to their importance. For candidate solutions that pass the screening, it is checked whether they satisfy the first and second priority constraints. If they do, the candidate solutions that pass the screening are Pareto ranked according to the violation degree of the third priority constraint and the objective function value. For each candidate solution after Pareto ranking, the corresponding safety level index is calculated. For candidate solutions whose safety level index meets the preset safety threshold, the Pareto front is extracted according to the Pareto dominance criterion to obtain the Pareto optimal solution set.

[0012] The task offloading scheme represented by the Pareto optimal solution set is executed to achieve edge computing task offloading.

[0013] On the other hand, embodiments of this application also provide an edge computing task offloading system based on soft clustering and dynamic constraints, including:

[0014] The soft clustering dimension mapping module is used to cluster K robots into G clusters, construct a K×G membership matrix, optimize the decision variables of the G cluster centers, and use the membership matrix to demap the cluster centers back to the individual decision space of the K robots.

[0015] The dynamic constraint relaxation module is used to determine the dynamic constraint relaxation factor based on the real-time proportion of feasible solutions in the population during the evolutionary iteration loop in the individual decision space. For safety level constraints and deadline constraints, the dynamic constraint relaxation factor is set to 0.

[0016] The feasibility bounding box module is used to calculate the corresponding constraint violation vector for candidate solutions obtained through mutation and crossover operations, and to filter candidate solutions based on the feasibility bounding box and constraint violation vector;

[0017] The constraint priority-driven selection module is used to divide constraints into three priorities according to their importance. For the candidate solutions that pass the screening, it checks whether they satisfy the first and second priority constraints in turn. If they do, the candidate solutions that pass the screening are Pareto sorted according to the violation degree of the third priority constraint and the objective function value.

[0018] The safety-aware Pareto update module is used to calculate the corresponding safety level index for each candidate solution after Pareto sorting. For candidate solutions whose safety level index meets the preset safety threshold, the Pareto frontier is extracted according to the Pareto dominance criterion to obtain the Pareto optimal solution set.

[0019] The scheme execution module is used to execute the task offloading scheme represented by the Pareto optimal solution set to realize the offloading of edge computing tasks.

[0020] On the other hand, embodiments of this application also provide a computer storage medium storing multiple computer instructions for causing a computer to execute the above-described method.

[0021] On the other hand, embodiments of this application also provide an electronic device, including a memory and at least one processor;

[0022] This memory stores multiple computer instructions;

[0023] When the at least one processor executes multiple computer instructions, the at least one processor performs the method described above.

[0024] The edge computing task offloading method and system based on soft clustering and dynamic constraints in this application have the following advantages:

[0025] (1) To address the problem of extremely sparse feasible regions caused by the explosive growth in the number of robots, a search mechanism that can navigate efficiently in extremely sparse feasible regions is provided, solving the problem of feasible solution density being less than 10. -15 This addresses the problem of traditional methods failing completely. Experiments show that in the K=150 scenario, the security compliance rate of this application is 100%, while that of NSGA-III is 0%, filling the technological gap and providing excellent technical support for large-scale industrial edge computing deployment.

[0026] (2) To address the stringent requirements of hard constraints such as safety level and deadlines (e.g., industrial control tasks require a safety level ≥ AES-256 and a latency ≤ 50ms), a Dynamic Constraint Relaxation (DCR) mechanism is proposed. This mechanism constantly sets the relaxation factor for these two types of constraints to 0, ensuring that these constraints are never relaxed, thus guaranteeing a 100% safety compliance rate and meeting the stringent requirements of industrial applications for hard constraints such as safety level and deadlines. Experiments show that in the K=100 scenario, the safety compliance rate of this application is 100%, while the comparative algorithm NSGA-III only achieves 15%. In the K=150 scenario, NSGA-III drops to 0%.

[0027] (3) To address resource constraints (such as production line edge server CPU capacity ≤ 50GHz and memory ≤ 32GB), a Feasibility Bounding Box (FBB) mechanism is proposed to achieve solution space prediction with O(1) time complexity, discard obviously infeasible solutions, significantly reduce invalid fitness evaluations, reduce the proportion of invalid evaluations from over 80% to below 10%, and improve search efficiency by more than 8 times. In the K=150 scenario, the running time is 118.43±0.54s, which meets the industrial real-time requirements.

[0028] (4) To address the difficulty of optimizing high-dimensional decision spaces, a soft clustering dimensionality mapping (SCM) method is proposed. This method maps the decision variables of K robots to G cluster center spaces, achieving dimensionality reduction optimization of the high-dimensional decision space. The decision space is reduced from K×N dimensions to G×N dimensions, reducing the search space dimensionality by more than 80%, effectively solving the curse of dimensionality problem. It can still operate stably in extreme scales where K≥150. Experiments show that in the K=150 scenario, SCM dimensionality reduction reduces memory usage by 60% and improves convergence speed by 3-5 times.

[0029] (5) To address the issue of insufficient multi-objective trade-offs, a safety-aware Pareto update mechanism is proposed. This mechanism prioritizes safety level, filters solutions that do not meet safety compliance requirements, and extracts the Pareto front. Within the solution set that meets safety requirements, the front is extracted according to the traditional Pareto dominance criterion, achieving a deep trade-off among multiple objectives such as latency, energy consumption, and safety, and providing a rich set of Pareto solutions. In the K=150 scenario, this application outputs 20 Pareto solutions with an HV value of 1.28±0.28, providing diverse options for industrial scheduling.

[0030] (6) To address the issue of adaptability to dynamic environments, a lightweight optimization framework enables re-optimization to be completed in sub-second time, meeting real-time requirements. The method in this application has moderate complexity and can adapt to dynamic factors such as network bandwidth fluctuations of ±40%, node failure rates of 5%, and security threat evolution cycles of 30 minutes. Experiments show that under network fluctuation scenarios, the HV value decreases by <8%, while the security compliance rate remains at 100%. Attached Figure Description

[0031] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0032] Figure 1 This is an architecture diagram of the edge computing task offloading method provided in the embodiments of this application.

[0033] Figure 2 A flowchart of an edge computing task unloading method provided in an embodiment of this application.

[0034] Figure 3 This is an architecture diagram of soft clustering dimension mapping provided in the embodiments of this application.

[0035] Figure 4 A flowchart illustrating the feasibility boundary box provided for embodiments of this application. Detailed Implementation

[0036] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0037] Figure 1-4 The present application provides an architecture diagram and flowchart of an edge computing task offloading method based on soft clustering and dynamic constraints, as illustrated in the embodiments of this application. The embodiments of this application provide an edge computing task offloading method based on soft clustering and dynamic constraints, including:

[0038] S100: Cluster the K robots into G clusters, construct a K×G membership matrix, optimize the decision variables of the G cluster centers, and use the membership matrix to demap the cluster centers back to the individual decision space of the K robots.

[0039] For example, before formally executing the method proposed in this application, it is necessary to initialize the parameters to be used in the method. The initialized parameters include the following:

[0040] The number of robots, K, has a value range of 50 ≤ K ≤ 150;

[0041] The number of edge nodes N, with a value ranging from 5 to N to 15.

[0042] The number of task types, I, has a range of 3 ≤ I ≤ 5;

[0043] Population size P, recommended value is 100;

[0044] Maximum number of iterations MAX_FE, recommended value is 5000;

[0045] The recommended value for the number of clusters G is G=min(20, );

[0046] The relaxation coefficient β is recommended to be 2.0. Through three sets of experiments with K=50 / 100 / 150, it was verified that the hypervolume HV index is optimal and stable when β∈[1.8,2.2]. Therefore, the median value of 2.0 is adopted in this application embodiment to balance exploration and development.

[0047] Safety threshold SR threshold The recommended value is 0.9.

[0048] After parameter initialization, the decision variables of the K robots are mapped to the G cluster center spaces through the membership matrix θ, thereby achieving dimensionality reduction optimization of the high-dimensional decision space.

[0049] This method specifically includes the following sub-steps:

[0050] S110 uses a fuzzy C-means clustering algorithm to divide K robots into G clusters based on task similarity and resource requirements among robots.

[0051] The robot's feature vector includes: task computation requirements. Unit: Mcycles; Task data volume Unit: MB; Delay Constraint Unit: milliseconds (ms); Security level requirements The value is 1 L represents the total number of security levels. In the automotive assembly workshop industrial robot scenario of this application, L=3, indicating that it uses the national cryptographic standard SM4 (commercial cryptographic standard 4) encryption. =3.

[0052] S120, calculate the membership degree of each robot to each cluster, and construct K... G-dimensional membership matrix .

[0053] The formula for calculating membership degree is:

[0054]

[0055] in, Let be the membership degree of the k-th (k=1,2,…,K) robot to the i-th (i=1,2,…,G) cluster. ,and , and Let be the Euclidean distances from robot k to clusters i and j, respectively, and m be the ambiguity exponent, typically taken as 2. All This will form the membership matrix. .

[0056] S130 performs an evolutionary operation in the cluster space to optimize the decision variables of G cluster centers.

[0057] The decision variables at the cluster center include: unloading decisions for robots within the cluster. , where is an N-dimensional vector, represents the intra-cluster robot caching strategy. Let be an I-dimensional vector, representing the decision-making process for CPU frequency allocation within a cluster of robots. , where is an N-dimensional vector, represents the safety level of the robot within the cluster. , is an N-dimensional vector.

[0058] S140, through the membership matrix The cluster center solution is mapped back to the individual decision space of the K robots.

[0059] Regarding the uninstallation decision:

[0060]

[0061] Regarding caching strategies:

[0062]

[0063] Regarding CPU frequency allocation:

[0064]

[0065] Regarding security levels:

[0066]

[0067] in, The decision variable for unloading the k-th robot to the n-th (n=1,2,…,N) edge node (0 indicates no unloading, 1 indicates unloading). The unloading decision variable for the i-th cluster center at the n-th edge node ( (continuous values) The decision variable for using the t-th task type is cached for the k-th robot (0 indicates not using, 1 indicates using). The decision variable cached for the i-th cluster center using the t-th task type ( (continuous values) The CPU frequency allocation strategy (in GHz) for the k-th robot at the n-th edge node. The CPU frequency allocation strategy (in GHz) for the i-th cluster center at the n-th edge node. The safety level policy for the k-th robot at the n-th edge node (values) ), The security level strategy for the i-th cluster center at the n-th edge node (values) ).

[0068] Specifically, in the security level formula Ensure that the safety level of each robot meets at least its minimum safety requirements. This is an important part of the hard constraint protection mechanism, ensuring a 100% safety compliance rate.

[0069] The aforementioned soft clustering dimensionality mapping method achieves dimensionality reduction optimization of high-dimensional decision spaces, reducing the decision space from K×N dimensions to G×N dimensions, reducing the search space dimensionality by more than 80%, effectively solving the curse of dimensionality problem, and still operating stably even at extreme scales of K≥150. The method's convergence speed is improved by 3-5 times, and memory usage is reduced by 60%.

[0070] In the scenario where K=50 and N=5, the original decision space has 50×5=250 dimensions, while the reduced decision space has 20×5=100 dimensions, achieving a dimensionality reduction of 60.0%. In the scenario where K=100 and N=10, the original decision space has 100×10=1000 dimensions, while the reduced decision space has 20×10=200 dimensions, achieving a dimensionality reduction of 80.0%. In the scenario where K=150 and N=15, the original decision space has 150×15=2250 dimensions, while the reduced decision space has 20×15=300 dimensions, achieving a dimensionality reduction of 86.7%.

[0071] S200 involves an evolutionary iterative loop within the individual decision space. During this loop, a dynamic constraint relaxation factor is determined based on the real-time proportion of feasible solutions in the population. For safety level constraints and deadline constraints, the dynamic constraint relaxation factor is set to 0. For candidate solutions obtained through mutation and crossover operations, the corresponding constraint violation vector is calculated. Candidate solutions are then screened based on the feasibility bounding box and the constraint violation vector. Constraints are categorized into three priorities based on importance. For each candidate solution that passes the screening, it is checked whether it satisfies the first and second priority constraints. If it does, the candidate solutions that pass the screening are Pareto ranked according to the violation degree of the third priority constraint and the objective function value. For each candidate solution after Pareto ranking, a corresponding safety level index is calculated. For candidate solutions whose safety level index meets the preset safety threshold, the Pareto front is extracted according to the Pareto dominance criterion to obtain the Pareto optimal solution set.

[0072] For example, S200 includes the following sub-steps:

[0073] S210, Calculate the dynamic constraint relaxation factor based on the real-time proportion of feasible solutions in the population. .

[0074] Dynamic constraint relaxation factor The calculation formula is:

[0075]

[0076] in, Let be the dynamic constraint relaxation factor for the s-th constraint (s=1,2,3,4). The relaxation coefficient is... The real-time proportion of feasible solutions in the population:

[0077]

[0078] in, Let p be the feasibility score of the feasible solution, with a value in the range [0,1]. ≥0.9 indicates that the solution satisfies all constraints and meets the safety level, that is, the solution is a safe and compliant solution; Pop is the current evolutionary population, representing the set of all candidate solutions in the population, |Pop| is the population size, the recommended value is 100, representing the total number of candidate solutions in the population; This represents the number of safe and compliant solutions in the current population, i.e., the number of solutions with a feasibility score greater than or equal to 0.9.

[0079] Furthermore, for security level constraints and deadline constraints, set... =0, ensuring that these two types of constraints are never relaxed, while soft constraints (resource capacity constraints and cache constraints) can be appropriately relaxed according to the proportion of feasible solutions, improving search efficiency:

[0080]

[0081] Where cv' is the adjusted constraint violation degree, cv is the original constraint violation degree, cv≥0 indicates constraint violation, and cv=0 indicates constraint satisfaction; Let be the dynamic constraint relaxation factor for the s-th constraint, for hard constraints =0, for soft constraints ∈(0,1], The larger the value, the greater the relaxation of the constraint. This formula is obtained by dividing by (1+ Reducing the penalty for constraint violations allows the algorithm to explore solutions with minor violations of soft constraints in the early stages of the search, while allowing for more thorough iterations. It gradually decays to 0, and eventually all constraints are strictly satisfied.

[0082] This is the core method of this application, ensuring a 100% security compliance rate.

[0083] In the scenario of K=50 and N=5, the security compliance rate of the method in this application reaches 100%, while the comparative algorithm NSGA-III is only 15%; in the scenarios of K=100 and N=10 and K=150 and N=15, the security compliance rate of this application remains at 100%, while that of NSGA-III drops to 0%.

[0084] S220 performs mutation and crossover operations on feasible solutions in the population to generate candidate solutions.

[0085] The mutation operation uses polynomial mutation, as shown below:

[0086]

[0087] in, and These are the feasible solutions before and after the mutation, respectively. This represents the polynomial variation perturbation.

[0088] The crossover operation uses simulated binary crossover, as shown below:

[0089]

[0090] in, These are the feasible solutions, or candidate solutions, after the crossover operation. and These are two feasible solutions that participate in the crossover operation. is the distribution factor.

[0091] S230, using the Feasibility Bounding Box (FBB) to evaluate candidate solutions. Predict the time complexity and discard obviously infeasible solutions.

[0092] The Feasibility Boundary Box (FBB) mechanism specifically includes:

[0093] S231, initialize FBB boundary as .

[0094] S232, For each candidate solution, calculate its constraint violation vector. .

[0095] constraint violation degree vector The violation degree of each constraint is calculated as follows:

[0096] Violation of resource capacity constraints ,in, The actual CPU frequency (in GHz) allocated to the task of the k-th robot at the n-th edge node is determined by the CPU frequency allocation decision. It is jointly determined by the CPU scheduling strategy of edge node n; The total CPU capacity (in GHz) of the nth edge node represents the maximum computing resources available for task offloading on that edge node.

[0097] Cache constraint violation ,in, The total cache capacity of the nth edge node (in MB) represents the maximum cache space that the edge node can use to store task data.

[0098] Violation of deadline constraints ,in, Data transmission latency (unit: ms) depends on the size of the task data, network bandwidth, and network congestion level; The execution latency (in milliseconds) depends on the task's computational requirements and the allocated CPU frequency. The latency (in milliseconds) for secure processing depends on the complexity of the encryption / decryption algorithm for the security level (e.g., AES-128, AES-256, SM4). Higher security levels result in lower latency. The larger.

[0099] Degree of violation of security level constraints ,in, The minimum safety level requirement for the k-th robot task is determined by industrial safety standards and data sensitivity, and is relevant to the industrial scenario described in this application. =3.

[0100] S233, comparison If the boundary with FBB exists Any violation degree If (s=1,2,3,4) exceeds the boundary range of FBB, then the corresponding candidate solution is directly discarded.

[0101] The judgment condition is:

[0102]

[0103] in, The expression indicates existence, where s is a constraint index, s=1 indicates resource capacity constraint, s=2 indicates cache constraint, s=3 indicates deadline constraint, and s=4 indicates security level constraint. Let be the degree of violation of the s-th constraint; and 1.2 and 0.8 are the upper and lower bounds of the FBB boundary corresponding to the s-th constraint, respectively; 1.2 and 0.8 are boundary expansion factors, indicating that the upper bound is expanded upward by 20% and the lower bound is expanded downward by 20%.

[0104] Specifically, the time complexity of this judgment condition is O(s), which is independent of the number of robots K and the number of edge nodes N, and is therefore equivalent to O(1) time complexity. This is the core advantage of the FBB mechanism of this invention: by comparing simple vector ranges, it avoids expensive fitness evaluation and greatly improves search efficiency.

[0105] S234. For the candidate solutions that pass the screening, update the FBB boundary.

[0106] The updated formula is:

[0107]

[0108]

[0109] Then, the updated FBB boundary is expanded to increase the FBB's range. Specifically, the boundary is expanded to... Multiply by 0.9, Multiply by 1.1.

[0110] The aforementioned FBB mechanism achieves O(1) time complexity prediction, significantly reducing invalid evaluations. This is also the core approach of this application, reducing the proportion of invalid evaluations from... The above was reduced to The following methods can improve search efficiency by more than 8 times.

[0111] In the K=50 scenario, FBB filters 50% of obviously infeasible solutions, saving 2500 fitness evaluations; in the K=100 scenario, FBB filters 65% of invalid solutions, saving 3250 evaluations; and in the K=150 scenario, FBB filters 80% of invalid solutions, saving 4000 evaluations.

[0112] S235 performs a constraint priority-driven selection operation on the candidate solutions filtered by FBB, prioritizing the retention of candidate solutions with low violation of key constraints.

[0113] In the embodiments of this application, the four types of constraints are divided into three priorities according to their importance:

[0114] First priority: security level constraints, which must never be relaxed;

[0115] Second priority: Deadline constraint, which must also never be relaxed;

[0116] Third priority: resource capacity constraints and cache constraints.

[0117] For candidate solutions selected through the FBB (Free Best Practice) filter, first check if they meet the first priority constraint; if not, they are directly eliminated. The judgment condition is:

[0118]

[0119] For candidate solutions that satisfy the first priority constraint, further check whether they satisfy the second priority constraint. The judgment condition is:

[0120]

[0121] Among the candidate solutions that satisfy the first two levels of constraints, Pareto sorting is performed based on the violation degree of the third priority constraint and the objective function value.

[0122] The constraint priority-driven selection mentioned above ensures the priority satisfaction of key constraints, improves search efficiency, and avoids wasting computational resources on solutions that violate key constraints. This is particularly effective in extreme sparse feasible domain scenarios, which is also the core approach of this application.

[0123] S236, update the FBB boundary and record the constraint violation range of the latest candidate solution.

[0124] S237 performs a security-aware Pareto update, prioritizing security level, filtering out candidate solutions that do not meet security compliance requirements, and extracting the Pareto front.

[0125] Specifically, the Pareto update of security perception includes:

[0126] For each candidate solution, calculate its security level index SR:

[0127]

[0128] in, This is an indicator function.

[0129] SR is given the highest priority; only candidate solutions with an SR that meets a preset safety threshold (i.e., 0.9) are included in the Pareto dominance comparison. The judgment condition is:

[0130]

[0131] Among the candidate solutions that meet the safety requirements, the Pareto front is extracted according to the traditional Pareto dominance criterion.

[0132] The Pareto dominance criterion is expressed as:

[0133]

[0134] Where x and y are the non-Pareto dominant solution and the Pareto dominant solution, respectively. Indicates any, This indicates that it exists. and All are objective functions. In this application, there are three objective functions, namely p, q = 1, 2, 3:

[0135] The time delay objective function minimizes the total task completion time (in milliseconds) to meet industrial real-time requirements (e.g., time delay ≤ 50ms).

[0136] Energy consumption objective function: minimize the total energy consumption of task execution (unit: J) to meet the requirements of green computing and energy saving;

[0137] : Security risk objective function, minimize the security risk of the task (dimensionless, value range [0,1]), and ensure 100% security compliance rate.

[0138] The Pareto dominance criterion means that a solution x is dominated by a solution y if and only if the following two conditions are met simultaneously:

[0139] Condition 1: For all objective functions The value of the solution x on the objective function p is no worse than the value of the solution y, that is... ;

[0140] Condition 2: There exists at least one objective function. The value of the solution x on the objective function q is strictly better than the value of the solution y, that is... .

[0141] Output the Pareto optimal solution set S Ensure S All solutions satisfy security and compliance requirements.

[0142] The aforementioned Pareto update of security perception ensures a 100% security compliance rate, which is also the core approach of this application.

[0143] In scenarios with K=50, K=100, and K=150, the security compliance rate of this application is 100%, while that of NSGA-III is 15%, 0%, and 0%, respectively. Specifically, in the scenario with K=50 and N=5, the security compliance rate of this application is 100.0%, while the security compliance rate of the NSGA-III algorithm is 15.0%, representing an improvement of 567.0%. In the scenario with K=100 and N=10, the security compliance rate of this application remains 100.0%, while the security compliance rate of the NSGA-III algorithm is 0.0%. In the scenario with K=150 and N=15, the security compliance rate of this application remains 100.0%, while the security compliance rate of the NSGA-III algorithm is also 0.0%.

[0144] In experiments on hypervolume HV, under the scenario of K=50, N=5, the HV value of this application is 12.10, the HV standard deviation is 0.12, and the HV coefficient of variation is 0.9%, while the HV value of the NSGA-III algorithm is 1.74, representing an improvement of 596%, the HV standard deviation is 1.83, representing a reduction of 93.4%, and the HV coefficient of variation is 95%, representing a reduction of 99.1%. Under the scenario of K=100, N=10, the HV value of this application is 5.60, while the safety compliance rate of the NSGA-III algorithm is 0.00, indicating that no feasible solution can be found at all, representing an improvement of ∞. Under the scenario of K=150, N=15, the HV value of this application is 1.28, while the HV value of the NSGA-III algorithm is 0.00, also representing an improvement of ∞. This demonstrates the powerful search capability of the method in this application in extremely sparse feasible regions, solving the problem of feasible solution density below 10. -15 This addresses the problem that traditional methods completely fail.

[0145] S240: Determine if the maximum number of iterations has been reached. If not, return to step S200; otherwise, output the Pareto optimal solution set S. The Pareto optimal solution set S This represents an edge computing task offloading solution.

[0146] S300 executes the task offloading scheme represented by the Pareto optimal solution set to realize edge computing task offloading.

[0147] This application also provides an edge computing task offloading system based on soft clustering and dynamic constraints, including:

[0148] The soft clustering dimension mapping module is used to cluster K robots into G clusters, construct a K×G membership matrix, optimize the decision variables of the G cluster centers, and use the membership matrix to demap the cluster centers back to the individual decision space of the K robots.

[0149] The dynamic constraint relaxation module is used to determine the dynamic constraint relaxation factor based on the real-time proportion of feasible solutions in the population during the evolutionary iteration loop in the individual decision space. For safety level constraints and deadline constraints, the dynamic constraint relaxation factor is set to 0.

[0150] The feasibility bounding box module is used to calculate the corresponding constraint violation vector for candidate solutions obtained through mutation and crossover operations, and to filter candidate solutions based on the feasibility bounding box and constraint violation vector;

[0151] The constraint priority-driven selection module is used to divide constraints into three priorities according to their importance. For the candidate solutions that pass the screening, it checks whether they satisfy the first and second priority constraints in turn. If they do, the candidate solutions that pass the screening are Pareto sorted according to the violation degree of the third priority constraint and the objective function value.

[0152] The safety-aware Pareto update module is used to calculate the corresponding safety level index for each candidate solution after Pareto sorting. For candidate solutions whose safety level index meets the preset safety threshold, the Pareto frontier is extracted according to the Pareto dominance criterion to obtain the Pareto optimal solution set.

[0153] The scheme execution module is used to execute the task offloading scheme represented by the Pareto optimal solution set to realize the offloading of edge computing tasks.

[0154] This application also provides a computer storage medium storing a plurality of computer instructions for causing a computer to execute the above-described method.

[0155] This application also provides an electronic device, including a memory and at least one processor;

[0156] This memory stores multiple computer instructions;

[0157] When the at least one processor executes multiple computer instructions, the at least one processor performs the method described above.

[0158] In addition to the experiments described above on dynamic constraint relaxation, feasibility bounding boxes, soft clustering dimension mapping, and security-aware Pareto updates, this application also conducted experiments on data stability, multi-objective trade-off depth, and scalability.

[0159] In the data stability experiments, under the scenario of K=50 and N=5, the mean running time of the proposed method was 21.37s, the standard deviation was 7.01s, and the coefficient of variation was 32.8%, reflecting the algorithm's fluctuation in real-world scenarios. In contrast, the NSGA-III algorithm had a mean running time of 13.88s, a mean HV value of 12.10, a standard deviation of 0.28, and an HV coefficient of variation of 21.5%. Under the scenario of K=100 and N=10, the mean running time was 49.42s, the standard deviation was 21.96s, and the coefficient of variation was 44.4%, consistent with the nonlinear characteristics of high-dimensional optimization problems. Under the scenario of K=150 and N=15, the mean running time was 118.43s, the standard deviation was 0.54s, and the coefficient of variation was 0.5%, while the mean HV value was 1.28, the standard deviation was 0.28, and the HV coefficient of variation was 21.5%, demonstrating the excellent stability of the proposed method in large-scale scenarios.

[0160] The data difference between the two independent experiments, Benchmark and Comparative, was only 0.0%-0.5%, demonstrating the extremely high reproducibility and consistency of the experiments.

[0161] In the multi-objective tradeoff depth experiment, under the scenario of K=50 and N=5, the number of Pareto solutions in this application is 20, and the number of Pareto solutions in NSGA-III is 20. Although the number of solutions is the same, the solution quality of this application is far superior to that of the NSGA-III algorithm. Under the scenario of K=100 and N=10, the number of Pareto solutions in this application is 20, and the number of Pareto solutions in NSGA-III is 20. Although the number of solutions is the same, all solutions from the NSGA-III algorithm are infeasible. Under the scenario of K=150 and N=15, the number of Pareto solutions in this application is 20, and the number of Pareto solutions in NSGA-III is 20. Although the number of solutions is the same, all solutions from the NSGA-III algorithm are infeasible.

[0162] These Pareto solutions achieve a deep trade-off between latency, energy consumption, and security, providing diverse options for different application scenarios.

[0163] In scalability experiments, both the proposed method and the NSGA-III algorithm showed stable performance in scenarios with K=50 and N=5, demonstrating comparable scalability. In scenarios with K=100 and N=10, the proposed method remained stable, while the NSGA-III algorithm failed (HV=0). In scenarios with K=150 and N=15, the proposed method remained stable, and the NSGA-III algorithm also failed. This demonstrates that the proposed method remains stable even at the extreme scale of K=150, while the NSGA-III algorithm completely fails at this scale, proving the absolute advantage of the proposed method in large-scale scenarios. The proposed method maintains 100% security compliance across all scales, demonstrating its robustness.

[0164] exist Figure 1 The functions of each layer are described below.

[0165] The data received by the input layer includes: robot parameters (such as task data size, computational requirements, deadline, and safety level requirements), edge node information (such as CPU capacity, cache capacity, and network characteristics), constraints (resource capacity, cache limits, deadline, and safety compliance), and objective functions (latency, energy consumption, and safety risks).

[0166] The core algorithm layer performs soft clustering dimension mapping, dynamic constraint relaxation, feasibility bounding box, constraint priority-driven selection, and safety-aware Pareto update algorithm.

[0167] The output layer outputs Pareto optimal solution set, ensuring 100% safety and compliance. This solution set includes the unloading decision, caching strategy, CPU frequency allocation, and safety level for each robot.

[0168] The evaluation and verification layer is used to evaluate performance metrics (such as security compliance rate, hypervolume HV, runtime and number of Pareto solutions) and to verify the effectiveness of the algorithm.

[0169] The experimental data above can be summarized into Table 1-6 below.

[0170] Table 1 Experimental parameters for a medium-sized scenario

[0171]

[0172] Table 2 Experimental results for medium-sized scenarios

[0173]

[0174] Experimental data analysis:

[0175] The data stability is extremely high: the running time is 21.37±7.01s and the coefficient of variation is 32.8%, reflecting the algorithm fluctuations in real-world scenarios; the HV value is 12.10±0.12 and the coefficient of variation is 0.9%, indicating that the solution quality stability is extremely high.

[0176] Significant performance advantages: HV value 12.10 vs 1.74, an improvement of 596%, indicating that the quality of the solution in this application far exceeds that of NSGA-III.

[0177] Safety compliance advantage: 100% vs 15%, indicating that this application ensures hard constraints are met, while NSGA-III seriously violates safety requirements.

[0178] Table 3 Experimental parameters for large-scale scenarios

[0179]

[0180] Table 4 Experimental results for large-scale scenarios

[0181]

[0182] Experimental data analysis:

[0183] The data stability is extremely high: the running time is 49.42±21.96s and the coefficient of variation is 44.4%, which is consistent with the nonlinear characteristics of high-dimensional optimization problems; the HV value is 5.60±0.042 and the coefficient of variation is 0.75%, indicating that the solution quality stability is extremely high.

[0184] Performance advantage is decisive: HV value 5.60 vs 0.00, NSGA-III is completely ineffective, and no feasible solution can be found.

[0185] Security compliance advantage: 100% vs 0%, all NSGA-III solutions violate security requirements.

[0186] Extremely sparse feasible region: NSGA-III completely fails in the K=100 scenario (HV=0, safety=0%), proving that the feasible region density is less than 10. -15 .

[0187] Table 5 Experimental parameters for extreme scale scenarios

[0188]

[0189] Table 6 Experimental results for extreme scale scenarios

[0190]

[0191] Experimental data analysis:

[0192] Breakthrough improvement in data stability: runtime 118.43±0.54s, coefficient of variation 0.5%, extremely stable; HV value 1.28±0.28, coefficient of variation 21.5%, the small HV base leads to a large coefficient of variation.

[0193] Absolute performance advantage: HV value 1.28 vs 0.00, NSGA-III completely fails; security compliance rate 100% vs 0%, all solutions of NSGA-III violate security requirements.

[0194] Scalability verification: The method can still run stably at the extreme scale of K=150, proving that the method has good scalability.

[0195] For edge computing scenarios in smart factories, the method flow of this application is as follows.

[0196] Scenario Description: A car assembly workshop contains 150 industrial robots. Each robot needs to offload motion control tasks to 15 edge computing nodes in real time. Task constraints include: latency ≤ 50ms (ensuring real-time control), security level ≥ AES-256 (protecting industrial data security), CPU capacity ≤ 50GHz (edge ​​node resource limitations), and cache capacity ≤ 32GB (edge ​​node storage limitations). Objective functions include: minimizing task completion latency, minimizing task execution energy consumption, and minimizing security risks.

[0197] The method flow of this application is as follows:

[0198] Step 1: Robot clustering and membership calculation.

[0199] Feature vectors (task computation requirements, task data volume, latency constraints, and safety level requirements) of 150 robots were extracted. Fuzzy C-means clustering was used to divide the robots into 20 clusters, and the membership degree of each robot to each cluster was calculated to construct a 150×20 dimensional membership matrix. Through soft clustering, the high-dimensional decision space of 150×15=2250 dimensions was compressed to 20×15=300 dimensions, with a compression rate of 86.7%, effectively solving the curse of dimensionality problem.

[0200] Step 2: Initialize the dynamic constraint relaxation mechanism.

[0201] In accordance with industrial safety requirements, the relaxation factors for safety level constraints and deadline constraints are always set to 0 to ensure that these two types of hard constraints are never relaxed. The relaxation factors for resource capacity constraints and cache constraints are dynamically adjusted based on the proportion of feasible solutions, with an initial value of [value missing]. As the number of iterations increases, the value gradually decreases to 0, and eventually all constraints are strictly satisfied.

[0202] Step 3: Initialize the feasibility bounding box.

[0203] Initialize FBB boundaries as It covers all possible constraint violation ranges. The FBB bounding box only stores 2×4=8 floating-point numbers (4 constraints for the upper and lower bounds), with a space complexity of O(1) and extremely low memory usage, making it suitable for resource-constrained edge devices.

[0204] Step 4: Evolutionary Iteration Cycle.

[0205] Mutation and crossover operations are performed in the cluster space (20×15 dimensional) to generate new candidate solutions. For each candidate solution, its constraint violation vector (CV) (including four constraints: resource capacity, cache limit, deadline, and security compliance) is calculated. An O(1) time complexity prediction is performed using the FBB bounding box, discarding obviously infeasible solutions. In the K=150 scenario, the FBB mechanism filters out 80% of invalid solutions, saving 4000 fitness evaluations and improving search efficiency by more than 8 times.

[0206] Step 5: Constraint priority-driven selection.

[0207] Constraints are categorized into three priorities based on importance: first priority is security level constraints, second priority is deadline constraints, and third priority is resource capacity constraints and cache constraints. For candidate solutions selected through the FBB (Free Best-Beat) process, each solution is checked to ensure it meets the first two levels of constraints; solutions that do not are eliminated. Among the candidate solutions that meet the first two levels of constraints, Pareto ranking is performed based on the violation degree of the third-priority constraints and the objective function value, and the optimal solution is retained for the next generation.

[0208] Step 6: Safety perception Pareto update.

[0209] For each candidate solution, calculate its security level index SR ( (between), only

[0210] Only solutions that meet the safety compliance requirements are included in the Pareto dominance comparison. Safety level is prioritized first, filtering out solutions that do not meet safety compliance requirements and extracting the Pareto front. Within the solution set that meets the safety requirements, the front is extracted according to the traditional Pareto dominance criterion, outputting 20 Pareto solutions.

[0211] Step 7: Output the optimal solution set and execute.

[0212] Output the Pareto optimal solution set, ensuring that all solutions in the set meet safety and compliance requirements. In the K=150 scenario, the 20 Pareto solutions output by this application have a safety compliance rate of 100% and an HV value of 1.28±0.28, while the comparative algorithm NSGA-III has a safety compliance rate of 0% and an HV value of 0.00, indicating that no feasible solution can be found. The task offloading scheme represented by the Pareto optimal solution set is executed to achieve real-time collaborative control of 150 industrial robots, meeting industrial requirements such as latency ≤50ms and safety level ≥AES-256.

[0213] Through the above process, this application has successfully solved the core technical problems of extremely sparse feasible domain, difficulty in satisfying hard constraints, and difficulty in optimizing high-dimensional decision space in the edge computing scenario of smart factories, providing a reliable technical solution for large-scale industrial edge computing deployment.

[0214] Although preferred embodiments of this application have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of this application.

[0215] Obviously, those skilled in the art can make various modifications and variations to this application without departing from the spirit and scope of this application. Therefore, if such modifications and variations fall within the scope of the claims of this application and their equivalents, this application also intends to include such modifications and variations.

Claims

1. An edge computing task offloading method based on soft clustering and dynamic constraints, characterized in that, include: K robots are clustered into G clusters, a K×G membership matrix is ​​constructed, the decision variables of the G cluster centers are optimized, and the cluster centers are demapped back to the individual decision space of the K robots using the membership matrix. An evolutionary iterative loop is performed in the individual decision space. During the loop, a dynamic constraint relaxation factor is determined based on the real-time proportion of feasible solutions in the population. For safety level constraints and deadline constraints, the dynamic constraint relaxation factor is set to 0. For candidate solutions obtained through mutation and crossover operations, the corresponding constraint violation vector is calculated. The candidate solutions are screened based on the feasibility bounding box and the constraint violation vector. Constraints are divided into three priorities according to their importance. For the candidate solutions that pass the screening, it is checked whether they satisfy the first and second priority constraints. If they do, the candidate solutions that pass the screening are Pareto ranked according to the violation degree of the third priority constraint and the objective function value. For each candidate solution after Pareto ranking, a corresponding safety level index is calculated. For candidate solutions whose safety level index meets a preset safety threshold, the Pareto front is extracted according to the Pareto dominance criterion to obtain the Pareto optimal solution set. Among the three priorities, the first priority is the safety level constraint, the second priority is the deadline constraint, and the third priority is the resource capacity constraint and the cache constraint. The task offloading scheme represented by the Pareto optimal solution set is executed to achieve edge computing task offloading.

2. The edge computing task offloading method based on soft clustering and dynamic constraints according to claim 1, characterized in that, Based on the task similarity and resource requirements among robots, the fuzzy C-means clustering algorithm is used to divide K robots into G clusters.

3. The edge computing task offloading method based on soft clustering and dynamic constraints according to claim 1, characterized in that, Before calculating the constraint violation vector of the candidate solution, the feasibility bounding box is initialized, and then updated according to the constraint violation vector of the candidate solution that has passed the screening.

4. The edge computing task offloading method based on soft clustering and dynamic constraints according to claim 1, characterized in that, The method for filtering candidate solutions based on the feasibility bounding box and the constraint violation vector is as follows: Compare the constraint violation vector with the feasibility bounding box. If the constraint violation degree in the constraint violation vector exceeds the boundary of the feasibility bounding box, then the corresponding candidate solution is directly discarded.

5. The edge computing task offloading method based on soft clustering and dynamic constraints according to claim 1, characterized in that, After performing Pareto sorting, the boundaries of the feasibility bounding box are updated, and the latest constraint violation range of the candidate solution is recorded.

6. The edge computing task offloading method based on soft clustering and dynamic constraints according to claim 1, characterized in that, The Pareto dominance criterion is expressed as follows: Where x' and y' are the non-Pareto dominant solution and the Pareto dominant solution, respectively. Indicates any, This indicates that it exists. and For different objective functions.

7. A system applying the edge computing task offloading method based on soft clustering and dynamic constraints as described in any one of claims 1-6, characterized in that, include: The soft clustering dimension mapping module is used to cluster K robots into G clusters, construct a K×G membership matrix, optimize the decision variables of the G cluster centers, and use the membership matrix to demap the cluster centers back to the individual decision space of the K robots. The dynamic constraint relaxation module is used to determine the dynamic constraint relaxation factor based on the real-time proportion of feasible solutions in the population during the evolutionary iteration loop in the individual decision space. For safety level constraints and deadline constraints, the dynamic constraint relaxation factor is set to 0. The feasibility bounding box module is used to calculate the corresponding constraint violation vector for candidate solutions obtained through mutation and crossover operations, and to filter the candidate solutions based on the feasibility bounding box and the constraint violation vector. The constraint priority-driven selection module is used to divide constraints into three priorities according to their importance. For the candidate solutions that have passed the screening, it is checked in turn whether they satisfy the first and second priority constraints. If they do, the candidate solutions that have passed the screening are Pareto sorted according to the violation degree of the third priority constraint and the objective function value. The safety-aware Pareto update module is used to calculate the corresponding safety level index for each candidate solution after Pareto sorting, and extract the Pareto frontier for the candidate solutions whose safety level index meets the preset safety threshold according to the Pareto dominance criterion to obtain the Pareto optimal solution set. The scheme execution module is used to execute the task offloading scheme represented by the Pareto optimal solution set to realize the offloading of edge computing tasks.

8. A computer storage medium, characterized in that, The computer storage medium stores a plurality of computer instructions, which are used to cause the computer to perform the method described in any one of claims 1-6.

9. An electronic device, characterized in that, Includes memory and at least one processor; The memory stores multiple computer instructions; When the at least one processor executes the plurality of computer instructions, it causes the at least one processor to perform the method according to any one of claims 1-6.

Images