Adaptive double time scale buoy satellite temperature-aware service deployment and task scheduling method

Abstract

The application discloses a kind of adaptive double time scale buoy satellite temperature perception service deployment and task scheduling method.First, facing buoy-satellite collaborative ocean edge computing network, the network model of underwater node, buoy node and low-orbit satellite collaborative processing task is constructed, and the edge node temperature model is established in combination with node computing power consumption, thermal resistance thermal capacity parameters and environmental temperature.Second, the self-attention mechanism is used to fuse service request, link availability, node temperature and short time scale operation cost, and the long time scale service deployment interval is dynamically determined.Finally, the improved DQN is used in the long time scale layer to generate temperature perception service deployment scheme, and the improved TD3 is used in the short time scale layer to obtain task scheduling scheme.The application can reduce the comprehensive system cost and node overheating risk, and improve the computing task processing efficiency of cross-domain ocean edge network.

Description

Technical Field

[0001] This invention belongs to the field of marine edge computing technology, specifically, it relates to an adaptive dual-timescale buoy satellite temperature sensing service deployment and task scheduling method. Background Technology

[0002] With the rapid development of applications such as smart ocean, offshore monitoring, underwater target identification, and marine emergency sensing, the computing tasks generated by underwater nodes are characterized by large data volumes, high computational complexity, and strict latency constraints. Due to the limited computing power, energy reserves, and communication range of underwater nodes themselves, relying solely on local processing or remote cloud center backhaul is insufficient to meet real-time processing requirements. Constructing a collaborative marine edge computing (MEC) network using surface buoy nodes and low-Earth-orbit (LEO) satellites, thus extending computing power to the edge of the ocean, has become an important way to improve the processing capabilities of offshore missions.

[0003] In buoy-satellite collaborative marine edge computing networks, tasks generated by underwater nodes typically depend on specific service types. These tasks can only be effectively processed when the corresponding service is deployed on a communicatively accessible buoy node or low-Earth orbit satellite. Therefore, service deployment determines the set of nodes capable of processing a task, while task scheduling determines which edge node actually executes the task. Service deployment is relatively stable and should not be updated frequently; however, task requests, link reachability, and node operating status change rapidly, requiring timely responses from task scheduling. Consequently, there is a significant difference in timescale between service deployment and task scheduling, necessitating collaborative optimization.

[0004] Meanwhile, buoy nodes and low-Earth orbit satellites generate computational power and cause node temperatures to rise during continuous task processing. In particular, low-Earth orbit satellites have limited heat dissipation, and continuous high-load computing can easily lead to temperature accumulation, resulting in thermal throttling, reduced processing capacity, and even equipment reliability risks. If task scheduling is based solely on link quality or computing power, tasks tend to concentrate on at least a few high-performance nodes, further exacerbating localized overheating and load imbalance. Therefore, in buoy-satellite collaborative marine edge computing scenarios, service deployment and task scheduling need to consider not only task processing costs and service availability but also node temperature status.

[0005] To address the underwater computing task offloading problem, existing technologies have proposed task offloading methods based on deep reinforcement learning. For example, the patent document "Underwater Computing Task Offloading Method Based on Improved Deep Deterministic Policy Gradient" (application date: December 4, 2024, application number: 202411768075.4, publication number: CN119255300A, authorization announcement number: CN119255300B, the content of this application can still be cited) constructs a three-layer network architecture of space-sea surface-underwater and uses an improved deep deterministic policy gradient algorithm to solve the underwater computing task offloading problem, so as to reduce the energy consumption of underwater sensor node task offloading; however, this scheme mainly focuses on the selection of task offloading mode and offloading quantity, without addressing the service deployment decisions on buoy nodes and low-orbit satellites, nor considering the impact of node temperature accumulation on task processing reliability.

[0006] Regarding the joint optimization of service deployment and task offloading, the patent document "Method, Apparatus, Computer Equipment and Medium for Joint Service Deployment and Task Offloading" (application date: March 30, 2021, application number: 202110337597.9, publication number: CN113010317A, the content of which can still be cited) considers different update frequencies for service deployment and task offloading, and makes service deployment and task offloading decisions separately based on a dual timescale framework; however, this scheme is mainly aimed at general mobile edge computing scenarios, does not consider cross-domain marine edge networks composed of underwater nodes, buoy nodes and low-orbit satellites, and does not adaptively determine the long-timescale service deployment interval based on service requests, link availability, node temperature and short-timescale operating costs.

[0007] In summary, existing technologies still have the following shortcomings: First, existing underwater mission offloading methods mostly focus on offloading decisions and energy consumption optimization, without considering the synergistic relationship between service deployment and mission scheduling; second, existing dual-timescale service deployment and mission offloading methods usually use fixed or preset timescales, lacking adaptive adjustment mechanisms for changes in system state; third, existing methods generally do not incorporate the temperature evolution of buoy nodes and low-Earth orbit satellites into the service deployment and mission scheduling process, which can easily lead to missions concentrating on a few nodes and causing overheating risks. Summary of the Invention

[0008] Purpose of the invention: To address the shortcomings of existing underwater mission offloading methods, such as the lack of coordination between service deployment and mission scheduling, the lack of adaptive timescale adjustment in existing dual-timescale service deployment and mission offloading methods, and the lack of sufficient consideration of node temperature accumulation and overheating risk in existing edge computing scheduling methods, this invention proposes an adaptive dual-timescale buoy satellite temperature sensing service deployment and mission scheduling method to overcome the deficiencies of the above technologies.

[0009] To achieve the above-mentioned objectives, the present invention employs the following technical solution:

[0010] An adaptive dual-timescale buoy satellite temperature sensing service deployment and task scheduling method includes the following steps:

[0011] S1: Construct a buoy-satellite collaborative marine edge computing network model and establish an edge node temperature model to obtain the underwater node's computing tasks, service deployment variables, path reachability status, and edge node temperature status vector; at the same time, divide the network running time into short-timescale time slots and long-timescale time slots, and establish an index mapping relationship between short-timescale time slots and long-timescale time slots;

[0012] S2: At each long timescale time slot boundary, based on the recent service request distribution, link availability statistics, edge node temperature status vector and the short timescale running cost in the previous long timescale interval, the current long timescale service deployment interval is dynamically determined using a self-attention mechanism.

[0013] S3: In the Based on the service deployment matrix of the previous long time-stamped time slot, the recent service request distribution, the node remaining storage vector, the edge node temperature status vector, and the current long time-stamped service deployment interval, an improved DQN is used to generate a temperature-aware service deployment scheme, and the service deployment scheme is maintained within the current long time-stamped service deployment interval.

[0014] S4: Within the current long-timescale service deployment interval, based on the service deployment scheme, the current task set, the path reachability status, and the edge node temperature status vector, an improved TD3 is used to obtain a task scheduling scheme. The short-timescale task running cost and temperature penalty are updated according to the task scheduling scheme. The short-timescale task running cost and temperature penalty are fed back to steps S2 and S3 to form the service deployment and task scheduling optimization results.

[0015] Furthermore, step S1 is specifically as follows:

[0016] S1-1: Constructing a buoy-satellite collaborative ocean edge computing network model:

[0017] A buoy-satellite collaborative ocean edge computing network, consisting of underwater nodes, buoy nodes, and low-Earth orbit (LEO) satellites, will be constructed in the target sea area. The LEO satellites will consist of... The set of buoy nodes is The underwater node set is Define the set of edge nodes with service deployment and task processing capabilities as Where any edge node is represented as ;

[0018] underwater nodes In short timescale slots The resulting computational task is represented as ,in, For the amount of task data, The number of CPU cycles required for the task. The type of service required for the task. The maximum tolerable latency for the task;

[0019] The system service collection is ,Serve The service image size is Edge nodes The storage capacity is Define service deployment variables ,in, Indicates service In long timescale slots Deployed on edge nodes ,otherwise ;

[0020] Define task scheduling variables ,in, Indicates underwater nodes In short timescale slots The generated tasks are scheduled to edge nodes. Execute, otherwise ;definition Indicates underwater nodes To edge nodes In short timescale slots Are there any available communication paths, where, Indicates underwater nodes To edge nodes In short timescale slots There is an available communication path, otherwise Tasks can only be scheduled to edge nodes where there is an available communication path and the services required by the task have been deployed, i.e., satisfying the requirement... and ;

[0021] Network runtime is divided into short-timescale slots. and long timescale time slot , No. Each long timescale time slot contains A short time-stamped time slot, and , ,in, , , Indicates the first The number of short timescale slots accumulated before the start of a long timescale slot; the It is dynamically determined by step S2 and serves as the time scale for long-timescale service deployment in step S3 and short-timescale task scheduling in step S4.

[0022] S1-2: Establish the temperature model for edge nodes:

[0023] For any edge node Define it in short timescale slots The temperature of the calculation unit is The ambient temperature is Thermal resistance is Heat capacity is The short timescale time slot length is The calculated power consumption is The temperature evolution of the edge nodes is then expressed as:

[0024] ,

[0025] in, For edge nodes In short timescale slots The computational power consumption;

[0026] Based on the dynamic voltage-frequency regulation model, edge nodes The computational power consumption is expressed as:

[0027] ,

[0028] in, For node idle power, For the effective switching capacitor coefficient, For nodes In short timescale slots The actual usage frequency of calculation, and , The maximum calculated frequency for a node;

[0029] Define the node's safe temperature threshold as The maximum safe operating temperature is And construct a temperature penalty term:

[0030] ;

[0031] Node temperature satisfies This yields any short-timescaled time slot. Edge node temperature state vector ; in the Take a long time-scaled time slot boundary, The corresponding edge node temperature state vector is represented as ;

[0032] Furthermore, step S2 is specifically as follows:

[0033] S2-1: Constructing a time-scale adjusted context sequence:

[0034] In each long timescale slot Boundary, obtain recent service request distribution Link availability statistics Edge node temperature state vector The average operating cost of the short timescale within the previous long timescale interval. And construct the context feature vector:

[0035] ,

[0036] in, Indicates service The proportion of requests within the previous long time interval;

[0037] Select the most recent Long timescale time slot Contextual features of the boundary, constructing a context sequence:

[0038] ,

[0039] When historical context features are insufficient When there are a few, use zero vectors or initial statistical features to fill in the blanks;

[0040] S2-2: Extracting time-scale modulated features based on self-attention mechanism:

[0041] context sequence Mapped to query matrices respectively Key matrix Sum matrix :

[0042] ,

[0043] in, , and These are the query mapping matrix, key mapping matrix, and value mapping matrix to be trained, respectively.

[0044] The attention representation is obtained by scaling the dot product attention:

[0045] ,

[0046] Pick The last row vector is used as a time-scale adjusted feature. ,in, For attention embedding dimensions, superscript This represents the matrix transpose operation, i.e. for Transpose of;

[0047] S2-3: Dynamically determine the deployment interval for long-timescaled services:

[0048] Preset candidate long time interval set Each candidate interval represents the number of short timescale slots contained within a long timescale slot; the adjustment feature is based on the timescale. Through time-scale adjustment strategy From the candidate long time-scaled interval set Select the current long-timescale service deployment interval:

[0049] ,

[0050] in, For candidate long time intervals, For the first The number of short timescale slots contained in a long timescale slot.

[0051] Furthermore, step S3 is specifically as follows:

[0052] S3-1: Deployment status of long-timescaled service:

[0053] In the Each long-time-scaled time slot boundary, based on the service deployment matrix of the previous long-time-scaled time slot. Recent service request distribution Node remaining storage vector and edge node temperature state vector Establish the deployment status of long-term time-marked services:

[0054] ,

[0055] in, ,and ;

[0056] S3-2: Determine the set of feasible candidate service deployment matrices:

[0057] Based on the long-timescaled service deployment status, a finite set of candidate service deployment matrices is generated:

[0058] ,

[0059] in, The number of candidate service deployment matrices, the first The candidate service deployment matrix is ​​represented as follows:

[0060] ,

[0061] in, Indicates the deployment matrix of candidate services China Service Whether deployed on edge nodes ;

[0062] Based on node storage capacity and predicted node temperature, the candidate service deployment matrix set is filtered to obtain a feasible candidate service deployment matrix set:

[0063] ,

[0064] in, Deploy matrix for candidate services Corresponding edge nodes The predicted temperature; when improving DQN to select service deployment actions, for those not belonging to The candidate service deployment matrix is ​​used to shield the services.

[0065] The predicted temperature is based on the current long-timescale time slot. The edge node temperature state vector at the boundary, the current long-timescale service deployment interval, the candidate service deployment matrix, the recent service request distribution, and the path reachability state are calculated; specifically, based on the recent service request distribution... Path reachability status and candidate service deployment matrix Determine the edge nodes under the candidate service deployment matrix. Candidate usage frequency The candidate computational power consumption is obtained based on the candidate computational frequency.

[0066] ;

[0067] Substituting the candidate computational power consumption into the edge node temperature model, we obtain:

[0068] ,

[0069] S3-3: Deployment scheme for improved DQN generation service based on temperature sensing:

[0070] When training the improved DQN, the service deployment cost, short-timescale task execution cost, and temperature penalty are combined to form the long-timescale reward:

[0071] ,

[0072] in, To normalize service deployment costs, To normalize the operating cost of short-term tasks, To normalize the short-timescale temperature penalty, , and To correspond to non-negative long-timescale weights, and The normalized service deployment cost is determined by the current service deployment scheme. Compared to the previous long-timescale time-slot service deployment matrix Service update volume, service image size The deployment transmission and write overhead are calculated and normalized; the normalized short-timescale task execution cost is obtained from the task scheduling matrix output in step S4. Current task set Path reachability status and the actual computation frequency used by edge nodes The path reachability state is calculated and normalized to obtain the path reachability state. Used to determine task transmission overhead and the actual computing frequency used by edge nodes. Used to determine task computational overhead; the normalized short-timescale temperature penalty is determined by the temperature penalty term in step S1-2. Calculated and normalized to obtain;

[0073] Deployment matrix of candidate services Construct a temperature-sensing Q-value prior:

[0074] ,

[0075] in, Deploy matrix for candidate services The corresponding normalized estimate of service update cost, To normalize the estimated temperature penalty, The cost of missing normalization services , and For the corresponding non-negative coefficients, and The normalized estimated service update cost is derived from the candidate service deployment matrix. Compared to the previous long-timescale time-slot service deployment matrix The service update volume is calculated; the normalized estimated temperature penalty is obtained from the predicted temperature obtained in step S3-2. Substituting the temperature penalty term from step S1-2 and normalizing it, the normalized service loss cost is obtained from the recent service request distribution. Candidate service deployment matrix And path reachability status determination, used to characterize the service matching loss that occurs when a recently requested service is not deployed on a reachable edge node;

[0076] Based on the temperature sensing Q-value prior Initialize the candidate service deployment matrix The value of the action; according to the first Mean temporal difference error of training rounds Adjusting the learning rate of the improved DQN:

[0077] ,

[0078] in, For the first Learning rate during rounds of training and These are the lower and upper bounds of the learning rate, respectively. It is a very small positive number;

[0079] In long timescale slots Boundary, from the set of feasible candidate service deployment matrices The candidate service deployment matrix with the highest action value is selected as the current temperature sensing service deployment scheme:

[0080] ,

[0081] in, To improve the action value function output by DQN, For network parameters; the aforementioned As the first The service deployment scheme is maintained within a long time-scaled service deployment interval.

[0082] Furthermore, step S4 is specifically as follows:

[0083] S4-1: Constructing the short-timescaled task scheduling state:

[0084] In the Within each long-time-scaled service deployment interval, maintain the service deployment plan obtained in step S3. Unchanged; in short timescale slots Based on the current task set Service Deployment Plan Path reachability status and edge node temperature state vector Construct short-timescaled task scheduling states:

[0085] ;

[0086] S4-2: Generating feasible task scheduling actions based on the improved TD3 algorithm:

[0087] A policy network employing an improved TD3 algorithm outputs a task scheduling preference matrix based on the short-timescale task scheduling state. ,in, Indicates underwater nodes Task scheduling to edge nodes Preference values;

[0088] The underwater nodes are determined based on path reachability, service availability, and node temperature constraints. The set of executable nodes:

[0089] ,

[0090] in, Represents edge nodes In long timescale slots Underwater nodes have been deployed The types of services required for the generated tasks ;

[0091] In the set of executable nodes The edge node with the highest task scheduling preference value is selected as the execution node:

[0092] ,

[0093] And obtain the task scheduling variables:

[0094] ,

[0095] This forms a short-timescale task scheduling matrix. ;

[0096] S4-3: Training and outputting a task scheduling scheme based on short-time-scaled rewards:

[0097] Construct a short-term reward based on task execution cost, temperature penalty, and residual constraint violation penalty:

[0098] ,

[0099] in, and These are the normalized short-timescale task execution cost and the normalized short-timescale temperature penalty, respectively. To implement the penalty term for violating normalized residual constraints, , and To correspond to non-negative short-timescale weights, and ;

[0100] In the improved TD3 algorithm training process, a dual-evaluation network is used to evaluate the value of short-timescaled actions, based on the aforementioned set of executable nodes. Perform a feasibility projection on the task scheduling preference matrix output by the policy network;

[0101] In each short timescale slot Output task scheduling matrix and the service deployment scheme Together, they form the optimized results of service deployment and task scheduling within the current long-term service deployment interval.

[0102] The advantages and technical effects of this invention are as follows:

[0103] First, this invention unifies the modeling of underwater nodes, buoy nodes, and low-Earth orbit satellites. It constrains the set of executable nodes by combining service deployment status and path reachability status, ensuring that task scheduling matches service availability and preventing tasks from being scheduled to edge nodes where the required services are not deployed or communication is unavailable. Second, it dynamically determines long-timescale service deployment intervals using a self-attention mechanism, allowing service deployment to adjust according to changes in service requests, link availability, node temperature, and short-timescale operating costs, reducing response lag or unnecessary updates caused by fixed timescales. Finally, it establishes an edge node temperature model and introduces temperature state constraints and temperature penalties into service deployment and task scheduling, enabling the system to perceive the thermal state of buoy nodes and low-Earth orbit satellites, reducing the risk of localized overheating caused by centralized task scheduling.

[0104] In summary, this invention can reduce system operating costs and node overheating risks, and improve the task processing efficiency and operational reliability of buoy-satellite collaborative marine edge computing networks. Attached Figure Description

[0105] Figure 1 This is an overall flowchart of one embodiment of the present invention;

[0106] Figure 2 This is a network architecture diagram of one embodiment of the present invention;

[0107] Figure 3 This is a schematic diagram illustrating the adaptive dual-timescale service deployment and task scheduling collaborative optimization according to an embodiment of the present invention;

[0108] Figure 4 This is a comparison chart of the normalized system costs of the method of the present invention and the comparative method in one embodiment of the present invention;

[0109] Figure 5 This is a comparison diagram of the highest node temperature between the method of the present invention and the comparative method in one embodiment of the present invention; Detailed Implementation

[0110] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments.

[0111] This embodiment proposes an adaptive dual-timescale buoy satellite temperature sensing service deployment and task scheduling method, the overall flowchart of which is as follows: Figure 1 As shown, it includes the following steps:

[0112] S1: Construct a buoy-satellite collaborative marine edge computing network model and establish an edge node temperature model to obtain the underwater node's computing tasks, service deployment variables, path reachability status, and edge node temperature state vector; simultaneously, divide the network runtime into short-timescale time slots and long-timescale time slots, and establish an index mapping relationship between short-timescale time slots and long-timescale time slots. The specific steps are as follows:

[0113] S1-1: Constructing a buoy-satellite collaborative ocean edge computing network model:

[0114] like Figure 2 As shown, a buoy-satellite collaborative ocean edge computing network is constructed in the target sea area, consisting of M=12 underwater nodes, B=3 buoy nodes, and S=6 low-Earth orbit satellites. The low-Earth orbit satellite set is as follows: The set of buoy nodes is The underwater node set is Define the set of edge nodes with service deployment and task processing capabilities as Where any edge node is represented as ;

[0115] underwater nodes In short timescale slots The resulting computational task is represented as ,in, For the amount of task data, The number of CPU cycles required for the task. The type of service required for the task. The maximum tolerable latency for the task;

[0116] The system service collection is ,Serve The service image size is Edge nodes The storage capacity is Define service deployment variables ,in, Indicates service In long timescale slots Deployed on edge nodes ,otherwise ;

[0117] Define task scheduling variables ,in, Indicates underwater nodes In short timescale slots The generated tasks are scheduled to edge nodes. Execute, otherwise ;definition Indicates underwater nodes To edge nodes In short timescale slots Are there any available communication paths, where, Indicates underwater nodes To edge nodes In short timescale slots There is an available communication path, otherwise Tasks can only be scheduled to edge nodes where there is an available communication path and the services required by the task have been deployed, i.e., satisfying the requirement... and ;

[0118] Network runtime is divided into short-timescale slots. and long timescale time slot , No. Each long timescale time slot contains A short time-stamped time slot, and , ,in, , , Indicates the first The number of short timescale slots accumulated before the start of a long timescale slot; the It is dynamically determined by step S2 and serves as the time scale for long-timescale service deployment in step S3 and short-timescale task scheduling in step S4.

[0119] S1-2: Establish the temperature model for edge nodes:

[0120] For any edge node Define it in short timescale slots The temperature of the calculation unit is The ambient temperature is Thermal resistance is Heat capacity is The short timescale time slot length is The calculated power consumption is The temperature evolution of the edge nodes is then expressed as:

[0121] ,

[0122] in, For edge nodes In short timescale slots The computational power consumption;

[0123] Based on the dynamic voltage-frequency regulation model, edge nodes The computational power consumption is expressed as:

[0124] ,

[0125] in, For node idle power, For the effective switching capacitor coefficient, For nodes In short timescale slots The actual usage frequency of calculation, and , The maximum calculated frequency for a node;

[0126] Define the node's safe temperature threshold as The maximum safe operating temperature is And construct a temperature penalty term:

[0127] ;

[0128] Node temperature satisfies This yields any short-timescaled time slot. Edge node temperature state vector ; in the Take a long time-scaled time slot boundary, The corresponding edge node temperature state vector is represented as ;

[0129] After completing step S1, the following method is adopted: Figure 3 The adaptive dual-timescale service deployment and task scheduling collaborative optimization framework shown dynamically determines the long-timescale service deployment interval, and generates service deployment schemes at the long-timescale layer and task scheduling schemes at the short-timescale layer.

[0130] S2: At the boundary of each long-timescale slot, based on the recent service request distribution, link availability statistics, edge node temperature state vector, and the short-timescale running cost within the previous long-timescale interval, the current long-timescale service deployment interval is dynamically determined using a self-attention mechanism. The specific steps are as follows:

[0131] S2-1: Constructing a time-scale adjusted context sequence:

[0132] In each long timescale slot Boundary, obtain recent service request distribution Link availability statistics Edge node temperature state vector The average operating cost of the short timescale within the previous long timescale interval. And construct the context feature vector:

[0133] ,

[0134] in, Indicates service The proportion of requests within the previous long time interval;

[0135] Select the most recent Contextual features of long-timescaled time slot boundaries are used to construct a context sequence:

[0136] ,

[0137] When historical context features are insufficient When there are a few, use zero vectors or initial statistical features to fill in the blanks;

[0138] S2-2: Extracting time-scale modulated features based on self-attention mechanism:

[0139] context sequence Mapped to query matrices respectively Key matrix Sum matrix :

[0140] ,

[0141] in, , and These are the query mapping matrix, key mapping matrix, and value mapping matrix to be trained, respectively.

[0142] The attention representation is obtained by scaling the dot product attention:

[0143] ,

[0144] Pick The last row vector is used as a time-scale adjusted feature. ,in, For attention embedding dimensions, superscript This represents the matrix transpose operation, i.e. for Transpose of;

[0145] S2-3: Dynamically determine the deployment interval for long-timescaled services:

[0146] Preset candidate long time interval set Each candidate interval represents the number of short timescale slots contained within a long timescale slot; the adjustment feature is based on the timescale. Through time-scale adjustment strategy From the candidate long time-scaled interval set Select the current long-timescale service deployment interval:

[0147] ,

[0148] in, For candidate long time intervals, For the first The number of short timescale slots contained in a long timescale slot.

[0149] S3: In the Based on the service deployment matrix of the previous long-timeslot, the recent service request distribution, the node remaining storage vector, the edge node temperature status vector, and the current long-timeslot service deployment interval, an improved DQN is used to generate a temperature-aware service deployment scheme, and the service deployment scheme is maintained within the current long-timeslot service deployment interval. The specific steps are as follows:

[0150] S3-1: Deployment status of long-timescaled service:

[0151] In the Each long-time-scaled time slot boundary, based on the service deployment matrix of the previous long-time-scaled time slot. Recent service request distribution Node remaining storage vector and edge node temperature state vector Establish the deployment status of long-term time-marked services:

[0152] ,

[0153] in, ,and ;

[0154] S3-2: Determine the set of feasible candidate service deployment matrices:

[0155] Based on the long-timescaled service deployment status, a finite set of candidate service deployment matrices is generated:

[0156] ,

[0157] in, The number of candidate service deployment matrices, the first The candidate service deployment matrix is ​​represented as follows:

[0158] ,

[0159] in, Indicates the deployment matrix of candidate services China Service Whether deployed on edge nodes ;

[0160] Based on node storage capacity and predicted node temperature, the candidate service deployment matrix set is filtered to obtain a feasible candidate service deployment matrix set:

[0161] ,

[0162] in, Deploy matrix for candidate services Corresponding edge nodes The predicted temperature; when improving DQN to select service deployment actions, for those not belonging to The candidate service deployment matrix is ​​used to shield the services.

[0163] The predicted temperature is based on the current long-timescale time slot. The edge node temperature state vector at the boundary, the current long-timescale service deployment interval, the candidate service deployment matrix, the recent service request distribution, and the path reachability state are calculated; specifically, based on the recent service request distribution... Path reachability status and candidate service deployment matrix Determine the edge nodes under the candidate service deployment matrix. Candidate usage frequency The candidate computational power consumption is obtained based on the candidate computational frequency.

[0164] ;

[0165] Substituting the candidate computational power consumption into the edge node temperature model, we obtain:

[0166] ,

[0167] S3-3: Deployment scheme for improved DQN generation service based on temperature sensing:

[0168] When training the improved DQN, the service deployment cost, short-timescale task execution cost, and temperature penalty are combined to form the long-timescale reward:

[0169] ,

[0170] in, To normalize service deployment costs, To normalize the operating cost of short-term tasks, To normalize the short-timescale temperature penalty, , and To correspond to non-negative long-timescale weights, and In this embodiment, we take =0.3、 =0.3、 =0.4;

[0171] Specifically, define the service update indicator variable:

[0172] ,

[0173] in, Indicates service In the A new long-timescale slot has been deployed to the edge node. In all other cases, the value is 0; based on the service update indicator variable, service image size, and deployment overhead, the service deployment cost is expressed as:

[0174] ,

[0175] in, Deploy services to edge nodes Equivalent deployment rate, For edge nodes The energy required to receive and write unit service data. and These are the service deployment latency weight and the service deployment energy consumption weight, respectively. , , In this embodiment, we take ;

[0176] For underwater nodes Scheduled to edge nodes The task, based on the amount of task data. Number of CPU cycles required for the task Path reachability status and the actual computation frequency used by edge nodes The task processing delay is obtained. Energy consumption for task processing The runtime cost of a short-timescaled task is expressed as:

[0177] ,

[0178] in, and These are the task latency weight and the task energy consumption weight, respectively. , , In this embodiment, we take ;

[0179] The short-timescale temperature penalty is determined by the temperature penalty term in step S1-2. Calculated;

[0180] The normalized service deployment cost, normalized short-timescale task operation cost, and normalized short-timescale temperature penalty are respectively determined by... , and Divide by the corresponding preset normalized reference value to obtain;

[0181] in, , and These are normalized reference values ​​for service deployment cost, short-duration task execution cost, and temperature penalty, respectively, and all are positive numbers; in this embodiment, Take the upper bound of the service deployment cost corresponding to all candidate new service deployments under the current simulation parameters. Take the upper bound of the task execution cost when all underwater nodes generate the maximum task under the current simulation parameters. ;

[0182] Deployment matrix of candidate services Construct a temperature-sensing Q-value prior:

[0183] ,

[0184] in, Deploy matrix for candidate services The corresponding normalized estimate of service update cost, To normalize the estimated temperature penalty, The cost of missing normalization services , and For the corresponding non-negative coefficients, and In this embodiment, we take =0.4、 =0.4、 =0.2; wherein, the normalized estimated service update cost is calculated according to the service deployment cost calculation method, taking into account the current service deployment scheme. Replace with candidate service deployment matrix The normalized estimated temperature penalty is then obtained based on the predicted temperature obtained in step S3-2. Substitute the temperature penalty term from step S1-2 and normalize it to obtain the normalized service missing cost; the normalized service missing cost is based on the recent service request distribution. Candidate service deployment matrix And path reachability status determination, used to characterize the service matching loss that occurs when a recently requested service is not deployed on a reachable edge node;

[0185] Based on the temperature sensing Q-value prior Initialize the candidate service deployment matrix The value of the action; according to the first Mean temporal difference error of training rounds Adjusting the learning rate of the improved DQN:

[0186] ,

[0187] in, For the first Learning rate during rounds of training and These are the lower and upper bounds of the learning rate, respectively. For a very small positive number, in this embodiment we take... =0.001;

[0188] In long timescale slots Boundary, from the set of feasible candidate service deployment matrices The candidate service deployment matrix with the highest action value is selected as the current temperature sensing service deployment scheme:

[0189] ,

[0190] in, To improve the action value function output by DQN, For network parameters; the aforementioned As the first The service deployment scheme is maintained within a long time-scaled service deployment interval.

[0191] S4: Within the current long-timescale service deployment interval, based on the service deployment scheme, the current task set, path reachability status, and edge node temperature status vector, an improved TD3 is used to obtain a task scheduling scheme. The short-timescale task running cost and temperature penalty are then updated according to the task scheduling scheme. The short-timescale task running cost and temperature penalty are fed back to steps S2 and S3 to form the service deployment and task scheduling optimization result. The specific steps are as follows:

[0192] S4-1: Constructing the short-timescaled task scheduling state:

[0193] In the Within each long-time-scaled service deployment interval, maintain the service deployment plan obtained in step S3. Unchanged; in short timescale slots Based on the current task set Service Deployment Plan Path reachability status and edge node temperature state vector Construct short-timescaled task scheduling states:

[0194] ;

[0195] S4-2: Generating feasible task scheduling actions based on the improved TD3 algorithm:

[0196] A policy network employing an improved TD3 algorithm outputs a task scheduling preference matrix based on the short-timescale task scheduling state. ,in, Indicates underwater nodes Task scheduling to edge nodes Preference values;

[0197] The underwater nodes are determined based on path reachability, service availability, and node temperature constraints. The set of executable nodes:

[0198] ,

[0199] in, Represents edge nodes In long timescale slots Underwater nodes have been deployed The types of services required for the generated tasks ;

[0200] In the set of executable nodes The edge node with the highest task scheduling preference value is selected as the execution node:

[0201] ,

[0202] And obtain the task scheduling variables:

[0203] ,

[0204] This forms a short-timescale task scheduling matrix. ;

[0205] S4-3: Training and outputting a task scheduling scheme based on short-time-scaled rewards:

[0206] Construct a short-term reward based on task execution cost, temperature penalty, and residual constraint violation penalty:

[0207] ,

[0208] in, and These are the normalized short-timescale task execution cost and the normalized short-timescale temperature penalty, respectively. To implement the penalty term for violating normalized residual constraints, , and To correspond to non-negative short-timescale weights, and In this embodiment, we take =0.35、 =0.35、 =0.3;

[0209] In the improved TD3 algorithm training process, a dual-evaluation network is used to evaluate the value of short-timescaled actions, based on the aforementioned set of executable nodes. Perform a feasibility projection on the task scheduling preference matrix output by the policy network;

[0210] In each short timescale slot Output task scheduling matrix and the service deployment scheme Together, they form the optimized results of service deployment and task scheduling within the current long-term service deployment interval.

[0211] Simulation comparisons were conducted using the method provided in this invention and two comparative methods in a buoy-satellite collaborative ocean edge computing scenario. The normalized system cost comparison results are as follows: Figure 4 As shown, the comparison results of the highest node temperatures are as follows: Figure 5 As shown. The method of this invention is referred to as "this invention"; the Fixed Two-Timescale Deployment and Scheduling (F-TDS) method uses a preset long-timescale service deployment interval and does not dynamically adjust it according to the system state; the Temperature-Unaware Adaptive Two-Timescale Deployment and Scheduling (TU-ATDS) method retains the adaptive timescale adjustment mechanism, but does not consider node temperature status and temperature penalty terms during service deployment and task scheduling.

[0212] This embodiment uses normalized system cost and node maximum temperature as performance evaluation indicators. The normalized system cost is obtained by weighted summation and normalization of service deployment cost, short-timescale task operation cost, and temperature penalty. The smaller the value, the better the overall system performance. The node maximum temperature represents the highest temperature among all buoy nodes and low-Earth orbit satellite nodes during the simulation period, and is used to measure the overheating risk of edge nodes.

[0213] This embodiment was simulated in MATLAB R2021a, with a fixed number of 12 underwater nodes. The average task data volume was set to 0.5 Mbit, 0.875 Mbit, 1.25 Mbit, 1.625 Mbit, and 2.0 Mbit. All methods used the same network topology, service request distribution, link reachability, node storage capacity, computing power, and temperature parameters to ensure the comparability of results. The main simulation parameters are listed in Table 1.

[0214] Table 1 - Simulation Parameters:

[0215] Sea area size 500 m × 500 m water depth 50 m Low Earth Orbit Satellite Altitude 1000 km Number of low-Earth orbit satellites S 6 Number of buoy nodes B 3 Number of underwater nodes M 12 Number of service types K 10 categories Short timescale slot length ( ) 10 s Candidate long-timescaled service deployment interval ( ) Task data volume ( ) 0.5–2 Mbit Number of CPU cycles required for the task ( ) cycles Maximum tolerable latency of the task ( ) 0.5–2 s Buoy node storage capacity ( ) 2–4 GB Low Earth Orbit satellite storage capacity ( ) 4–8 GB Maximum calculation frequency of buoy nodes ( ) 5–8 GHz Maximum computing frequency of low-Earth orbit satellites ( ) 8–12 GHz Node safety temperature threshold ( ) 70 ℃ Maximum safe operating temperature of the node ( ) 80 ℃ Length of attention history ( ) 4 Long timescale weights ( ) 0.3, 0.3, 0.4 Temperature sensing Q-value prior weights ( ) 0.4, 0.4, 0.2 Short timescale weights ( ) 0.35, 0.35, 0.3 Number of independent simulations 20 times

[0216] from Figure 4 It can be seen that the normalized system cost of the three methods generally increases with the increase of the average task data volume. This is because as the task data volume increases, the task transmission overhead, service matching pressure, and edge node computing and processing pressure increase simultaneously. F-TDS uses a fixed long-timescale service deployment interval, which is difficult to adapt to changes in service requests and link availability in a timely manner; although TU-ATDS can dynamically adjust the long-timescale service deployment interval, it does not consider the node temperature status, and tasks are still prone to be concentrated on some high-performance nodes, resulting in increased temperature penalties. In contrast, this invention can dynamically adjust the long-timescale service deployment interval by combining service requests, link availability, node temperature, and short-timescale operating costs, and generate temperature-aware service deployment schemes and task scheduling schemes, thus achieving a lower normalized system cost.

[0217] from Figure 5 It can be seen that the maximum node temperature of all three methods increases with the average task data volume. This is because as the task data volume increases, edge nodes need to bear higher transmission and computing processing pressure, and continuous computing load will cause node temperature accumulation. F-TDS, due to its fixed service deployment interval, struggles to alleviate load concentration in a timely manner; TU-ATDS lacks temperature-sensing constraints, making it easy for tasks to be continuously scheduled to nodes with better link conditions or stronger computing capabilities, causing the maximum node temperature to approach or even exceed the maximum safe operating temperature. In contrast, this invention introduces node temperature status and temperature penalty terms during service deployment and task scheduling, which can distribute task load and reduce node temperature accumulation and overheating risk.

[0218] In summary, this invention can reduce the normalized system cost in buoy-satellite collaborative marine edge computing networks, effectively suppress node temperature accumulation, and improve system task processing efficiency and operational reliability.

[0219] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit them. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions claimed by the present invention.

Claims

1. A method for deploying and scheduling adaptive dual-timescale buoy satellite temperature sensing services, characterized in that, Includes the following steps: S1: Construct a buoy-satellite collaborative marine edge computing network model and establish an edge node temperature model to obtain the underwater node's computing tasks, service deployment variables, path reachability status, and edge node temperature status vector; at the same time, divide the network running time into short-timescale time slots and long-timescale time slots, and establish an index mapping relationship between short-timescale time slots and long-timescale time slots; S2: At each long timescale time slot boundary, based on the recent service request distribution, link availability statistics, edge node temperature status vector and the short timescale running cost in the previous long timescale interval, the current long timescale service deployment interval is dynamically determined using a self-attention mechanism. S3: In the Based on the service deployment matrix of the previous long time-stamped time slot, the recent service request distribution, the node remaining storage vector, the edge node temperature status vector, and the current long time-stamped service deployment interval, an improved DQN is used to generate a temperature-aware service deployment scheme, and the service deployment scheme is maintained within the current long time-stamped service deployment interval. S4: Within the current long-timescale service deployment interval, based on the service deployment scheme, the current task set, the path reachability status, and the edge node temperature status vector, an improved TD3 is used to obtain a task scheduling scheme. The short-timescale task running cost and temperature penalty are updated according to the task scheduling scheme. The short-timescale task running cost and temperature penalty are fed back to steps S2 and S3 to form the service deployment and task scheduling optimization results.

2. The adaptive dual-timescale buoy satellite temperature sensing service deployment and task scheduling method as described in claim 1, characterized in that, The specific steps of S1 are as follows: S1-1: Constructing a buoy-satellite collaborative ocean edge computing network model: A buoy-satellite collaborative ocean edge computing network, consisting of underwater nodes, buoy nodes, and low-Earth orbit (LEO) satellites, will be constructed in the target sea area. The LEO satellites will consist of... The set of buoy nodes is The underwater node set is Define the set of edge nodes with service deployment and task processing capabilities as Where any edge node is represented as ; underwater nodes In short timescale slots The resulting computational task is represented as ,in, For the amount of task data, The number of CPU cycles required for the task. The type of service required for the task. The maximum tolerable latency for the task; The system service collection is ,Serve The service image size is Edge nodes The storage capacity is Define service deployment variables ,in, Indicates service In long timescale slots Deployed on edge nodes ,otherwise ; Define task scheduling variables ,in, Indicates underwater nodes In short timescale slots The generated tasks are scheduled to edge nodes. Execute, otherwise ;definition Indicates underwater nodes To edge nodes In short timescale slots Are there any available communication paths, where, Indicates underwater nodes To edge nodes In short timescale slots There is an available communication path, otherwise Tasks can only be scheduled to edge nodes where there is an available communication path and the services required by the task have been deployed, i.e., satisfying the requirement... and ; Network runtime is divided into short-timescale slots. and long timescale time slot , No. Each long timescale time slot contains A short time-stamped time slot, and , ,in, , , Indicates the first The number of short timescale slots accumulated before the start of a long timescale slot; the It is dynamically determined by step S2 and serves as the time scale for long-timescale service deployment in step S3 and short-timescale task scheduling in step S4. S1-2: Establish the temperature model for edge nodes: For any edge node Define it in short timescale slots The temperature of the calculation unit is The ambient temperature is Thermal resistance is Heat capacity is The short timescale time slot length is The calculated power consumption is The temperature evolution of the edge nodes is then expressed as: ， in, For edge nodes In short timescale slots The computational power consumption; Based on the dynamic voltage-frequency regulation model, edge nodes The computational power consumption is expressed as: ， in, For node idle power, For the effective switching capacitor coefficient, For nodes In short timescale slots The actual usage frequency of calculation, and , The maximum calculated frequency for a node; Define the node's safe temperature threshold as The maximum safe operating temperature is And construct a temperature penalty term: ； Node temperature satisfies This yields any short-timescaled time slot. Edge node temperature state vector ; in the Take a long time-scaled time slot boundary, The corresponding edge node temperature state vector is represented as .

3. The adaptive dual-timescale buoy satellite temperature sensing service deployment and task scheduling method as described in claim 1, characterized in that, Step S2 is as follows: S2-1: Constructing a time-scale adjusted context sequence: In each long timescale slot Boundary, obtain recent service request distribution Link availability statistics Edge node temperature state vector The average operating cost of the short timescale within the previous long timescale interval. And construct the context feature vector: ， in, Indicates service The proportion of requests within the previous long time interval; Select the most recent Long timescale time slot Contextual features of the boundary, constructing a context sequence: ， When historical context features are insufficient When there are a few, use zero vectors or initial statistical features to fill in the blanks; S2-2: Extracting time-scale modulated features based on self-attention mechanism: context sequence Mapped to query matrices respectively Key matrix Sum matrix : ， in, , and These are the query mapping matrix, key mapping matrix, and value mapping matrix to be trained, respectively. The attention representation is obtained by scaling the dot product attention: ， Pick The last row vector is used as a time-scale adjusted feature. ,in, For attention embedding dimensions, superscript This represents the matrix transpose operation, i.e. for Transpose of; S2-3: Dynamically determine the deployment interval for long-timescaled services: Preset candidate long time interval set Each candidate interval represents the number of short timescale slots contained within a long timescale slot; the adjustment feature is based on the timescale. Through time-scale adjustment strategy From the candidate long time-scaled interval set Select the current long-timescale service deployment interval: ， in, For candidate long time intervals, For the first The number of short timescale slots contained in a long timescale slot.

4. The adaptive dual-timescale buoy satellite temperature sensing service deployment and task scheduling method as described in claim 1, characterized in that, Step S3 is as follows: S3-1: Deployment status of long-timescaled service: In the Each long-time-scaled time slot boundary, based on the service deployment matrix of the previous long-time-scaled time slot. Recent service request distribution Node remaining storage vector and edge node temperature state vector Establish the deployment status of long-term time-marked services: ， in, ,and ; S3-2: Determine the set of feasible candidate service deployment matrices: Based on the long-timescaled service deployment status, a finite set of candidate service deployment matrices is generated: ， in, The number of candidate service deployment matrices, the first The candidate service deployment matrix is ​​represented as follows: ， in, Indicates the deployment matrix of candidate services China Service Whether deployed on edge nodes ; Based on node storage capacity and predicted node temperature, the candidate service deployment matrix set is filtered to obtain a feasible candidate service deployment matrix set: ， in, Deploy matrix for candidate services Corresponding edge nodes The predicted temperature; when improving DQN to select service deployment actions, for those not belonging to The candidate service deployment matrix is ​​used to shield the services. The predicted temperature is based on the current long-timescale time slot. The edge node temperature state vector at the boundary, the current long-timescale service deployment interval, the candidate service deployment matrix, the recent service request distribution, and the path reachability state are calculated; specifically, based on the recent service request distribution... Path reachability status and candidate service deployment matrix Determine the edge nodes under the candidate service deployment matrix. Candidate usage frequency The candidate computational power consumption is obtained based on the candidate computational frequency. ； Substituting the candidate computational power consumption into the edge node temperature model, we obtain: ， S3-3: Deployment scheme for improved DQN generation service based on temperature sensing: When training the improved DQN, the service deployment cost, short-timescale task execution cost, and temperature penalty are combined to form the long-timescale reward: ， in, To normalize service deployment costs, To normalize the operating cost of short-term tasks, To normalize the short-timescale temperature penalty, , and To correspond to non-negative long-timescale weights, and The normalized service deployment cost is determined by the current service deployment scheme. Compared to the previous long-timescale time-slot service deployment matrix Service update volume, service image size The deployment transmission and write overhead are calculated and normalized; the normalized short-timescale task execution cost is obtained from the task scheduling matrix output in step S4. Current task set Path reachability status and the actual computation frequency used by edge nodes The path reachability state is calculated and normalized to obtain the path reachability state. Used to determine task transmission overhead and the actual computing frequency used by edge nodes. Used to determine task computational overhead; the normalized short-timescale temperature penalty is determined by the temperature penalty term in step S1-2. Calculated and normalized to obtain; Deployment matrix of candidate services Construct a temperature-sensing Q-value prior: ， in, Deploy matrix for candidate services The corresponding normalized estimate of service update cost, To normalize the estimated temperature penalty, The cost of missing normalization services , and For the corresponding non-negative coefficients, and The normalized estimated service update cost is derived from the candidate service deployment matrix. Compared to the previous long-timescale time-slot service deployment matrix The service update volume is calculated; the normalized estimated temperature penalty is obtained from the predicted temperature obtained in step S3-2. Substituting the temperature penalty term from step S1-2 and normalizing it, the normalized service loss cost is obtained from the recent service request distribution. Candidate service deployment matrix And path reachability status determination, used to characterize the service matching loss that occurs when a recently requested service is not deployed on a reachable edge node; Based on the temperature sensing Q-value prior Initialize the candidate service deployment matrix The value of the action; according to the first Mean temporal difference error of training rounds Adjusting the learning rate of the improved DQN: ， in, For the first Learning rate during rounds of training and These are the lower and upper bounds of the learning rate, respectively. It is a very small positive number; In long timescale slots Boundary, from the set of feasible candidate service deployment matrices The candidate service deployment matrix with the highest action value is selected as the current temperature sensing service deployment scheme: ， in, To improve the action value function output by DQN, For network parameters; the aforementioned As the first The service deployment scheme is maintained within a long time-scaled service deployment interval.

5. The adaptive dual-timescale buoy satellite temperature sensing service deployment and task scheduling method as described in claim 1, characterized in that, Step S4 is as follows: S4-1: Constructing the short-timescaled task scheduling state: In the Within each long-time-scaled service deployment interval, maintain the service deployment plan obtained in step S3. Unchanged; in short timescale slots Based on the current task set Service Deployment Plan Path reachability status and edge node temperature state vector Construct short-timescaled task scheduling states: ； S4-2: Generating feasible task scheduling actions based on the improved TD3 algorithm: A policy network employing an improved TD3 algorithm outputs a task scheduling preference matrix based on the short-timescale task scheduling state. ,in, Indicates underwater nodes Task scheduling to edge nodes Preference values; The underwater nodes are determined based on path reachability, service availability, and node temperature constraints. The set of executable nodes: ， in, Represents edge nodes In long timescale slots Underwater nodes have been deployed The types of services required for the generated tasks ; In the set of executable nodes The edge node with the highest task scheduling preference value is selected as the execution node: ， And obtain the task scheduling variables: ， This forms a short-timescale task scheduling matrix. ; S4-3: Training and outputting a task scheduling scheme based on short-time-scaled rewards: Construct a short-term reward based on task execution cost, temperature penalty, and residual constraint violation penalty: ， in, and These are the normalized short-timescale task execution cost and the normalized short-timescale temperature penalty, respectively. To implement the penalty term for violating normalized residual constraints, , and To correspond to non-negative short-timescale weights, and ; In the improved TD3 algorithm training process, a dual-evaluation network is used to evaluate the value of short-timescaled actions, based on the aforementioned set of executable nodes. Perform a feasibility projection on the task scheduling preference matrix output by the policy network; In each short timescale slot Output task scheduling matrix and the service deployment scheme Together, they form the optimized results of service deployment and task scheduling within the current long-term service deployment interval.

Images