A method for collaborative allocation of computing resources with quality of service guarantee

Abstract

The application provides a sensing and computing resource collaborative allocation method with service quality guarantee, and the main steps include: (1) constructing a sensing and computing integrated system architecture with service quality guarantee assisted by a UAV; (2) under the constraints of low latency and high reliability, constructing an optimization problem with the minimum system weighted average power consumption as the target; (3) on a large time scale, adopting a flexible actor-critic algorithm to design a trajectory and adjust a data packet loss probability; (4) on a small time scale, adopting a successive convex approximation and an alternating direction multiplier method to allocate a user frequency band resource; and (5) on the small time scale, adopting a low complexity method based on a maximum ratio transmission to allocate a UAV transmitting beamforming and a computing resource. Under the premise of guaranteeing a high reliable and low latency service quality of a user and meeting a target sensing performance, the application significantly reduces the system weighted average power consumption, and has good practical value in an industrial internet of things scene such as a smart grid.

Description

Technical Field

[0001] This invention relates to the field of 6G wireless communication technology, specifically to a collaborative allocation method for communication, sensing, and computing resources with quality of service assurance. It is applicable to integrated communication, sensing, and computing scenarios in industrial IoT such as smart grids and smart factories that require drone assistance and have deterministic latency and reliability quality of service requirements. Background Technology

[0002] With the accelerated evolution of sixth-generation (6G) mobile communication technology, the deep integration of communication, sensing, and computing has become one of the core characteristics of next-generation wireless networks. Integrated communication, sensing, and computing technology, through shared spectrum resources and hardware platforms, enables wireless nodes to simultaneously possess multi-dimensional capabilities such as data transmission, environmental awareness, and edge computing, significantly improving wireless resource utilization efficiency and overall system performance. However, terrestrial integrated communication, sensing, and computing wireless networks face several technical challenges. On the one hand, fixed terrestrial base stations struggle to dynamically adjust the spatial distribution of communication, computing, and sensing resources according to service needs. On the other hand, in complex environments such as remote mountainous areas and disaster sites, terrestrial base station signals are susceptible to obstruction and fading, leading to a significant decrease in communication reliability and sensing accuracy. These limitations have prompted researchers to focus on mobile aerial platforms to compensate for the shortcomings of terrestrial wireless networks.

[0003] Unmanned Aerial Vehicles (UAVs), with their unique advantages such as flexibility, rapid deployment, and on-demand coverage, have become an important component of the 6G integrated air-space-ground network. The combination of UAVs with integrated sensing and computing technologies can build flexible and efficient multi-functional intelligent nodes in the air. By dynamically adjusting their flight trajectories, UAVs can proactively optimize air-to-ground link channel conditions, providing users with customized sensing and computing services on demand. In smart grids, UAV-assisted applications have permeated multiple core aspects of power system operation and maintenance, playing a crucial role in line inspection, relay communication, and disaster emergency response.

[0004] In practical applications, industrial IoT scenarios such as smart grids not only require networks with data transmission, environmental sensing, and edge computing capabilities, but also demand deterministic ultra-low latency and ultra-high reliability quality of service guarantees for data links. Ultra-Reliable Low Latency Communication (URLLC) technology can effectively meet these requirements, demanding that data links provide extremely high reliability (e.g., packet loss probability less than 10%) under extremely short latency constraints (e.g., end-to-end latency less than 1ms). -5UAV-assisted sensing and computing systems with quality of service (QoS) guarantees require UAVs to provide communication, sensing, and computing services while meeting users' demands for ultra-low latency and ultra-high reliability QoS through dynamic adjustments to flight trajectories and resource allocation schemes. This effectively compensates for the limitations of terrestrial network channel conditions and insufficient coverage, and avoids excessive latency caused by long-distance data transmission. Therefore, integrating UAVs with sensing and computing technologies and introducing deterministic end-to-end latency and reliability guarantees for task processing has become a key technical path to improve the QoS of wireless networks.

[0005] In UAV-assisted sensory computing integrated systems with high reliability, low latency, and quality of service (QoS) guarantees, UAV trajectory design and system resource allocation still face numerous challenges. On one hand, due to the highly dynamic and random nature of the data volume generated by user-generated high-reliability, low-latency computing tasks, the update frequency of system resource allocation schemes is much faster than the update frequency of UAV trajectories. Therefore, how to rationally design UAV trajectories and dynamically allocate system resources at different time scales to ensure the low latency and high reliability requirements of user tasks, as well as target perception performance, remains a problem urgently needing research. On the other hand, existing UAV trajectory design and resource allocation methods have high computational complexity and cannot meet the real-time requirements of high-reliability, low-latency services for UAV trajectory design and system resource allocation methods. This invention proposes a sensory computing resource collaborative allocation method with QoS guarantees, which can effectively solve the above problems. Summary of the Invention

[0006] Technical Problem: To address the issues of existing UAV-assisted integrated sensing and computing systems, such as the failure to consider the randomness of ground user computing task arrivals, the inability to guarantee the latency and reliability quality of service (QoS) requirements of high-reliability, low-latency services, and the neglect of time-scale inconsistencies in trajectory design and resource allocation, as well as the excessively high computational complexity of optimization methods, this invention proposes a collaborative allocation method for sensing and computing resources with QoS guarantees. This method, while meeting the offloading requirements of high-reliability, low-latency computing tasks from various ground users and the target perception performance, aims to minimize the system's weighted average power consumption. It achieves rational design of UAV trajectories over large timescales and dynamic allocation of system resources over small timescales, improving system energy efficiency and ensuring low latency and high reliability for user computing task processing.

[0007] Technical Solution: The technical problem to be solved by this invention is to provide a collaborative allocation method for sensory computing resources with quality of service assurance. On the basis of ensuring the high reliability and low latency computing task offloading requirements of ground users and target perception performance, the method aims to minimize the weighted average power consumption of the system. On a large time scale, deep learning methods are used to solve the UAV trajectory design and data packet loss probability adjustment problem, and on a small time scale, low-complexity numerical optimization methods are used to solve the system resource allocation problem, thereby achieving efficient UAV trajectory design and system resource allocation on both time scales.

[0008] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows:

[0009] A collaborative allocation method for synergistic computing resources with quality of service assurance includes the following steps: constructing a UAV-assisted synergistic computing integrated system architecture with quality of service assurance; and constructing an optimization problem with the goal of minimizing the system's weighted average power consumption under constraints of maximum user computing task processing latency, maximum data packet loss probability, and target perception performance.

[0010] The optimization problem is decomposed into a large-scale UAV trajectory design and packet loss probability adjustment subproblem, and a small-scale system resource allocation subproblem. On the large-scale, a Soft Actor-Critic (SAC) algorithm is used for UAV trajectory design and dynamic adjustment of packet loss probability. On the small-scale, the system resource allocation subproblem is further decomposed into a user bandwidth allocation subproblem and a UAV transmission beamforming and computational resource allocation subproblem. Successive Convex Approximation (SCA) and Alternating Direction Method of Multipliers (ADMM) methods are used for user bandwidth allocation. A low-complexity method based on Maximum Ratio Transmission (MRT) is employed for UAV transmission beamforming and computational resource allocation.

[0011] Preferably, the UAV-assisted integrated sensing and computing system with quality of service assurance includes a UAV. One user, one target, and one fog node. The UAV is equipped with one edge computing server and two array elements. Uniform Linear Arrays (ULAs) are used for both signal transmission and reception. Both users and fog nodes are equipped with a single antenna. The service time of the UAVs is divided into equal segments. There are 1 time slot, and the length of each time slot is 1. The time slot set is Each time slot is further divided into There are 1 time frames, each time frame having a length of 1. The time frame set is .

[0012] Preferably, a probabilistic channel model is used to model the wireless communication channel gain between the user and the UAV, and between the UAV and the fog node.

[0013] Preferably, each user uses Frequency Division Multiple Access (FDMA) and finite block length coding to offload computation task data packets, defining... This represents the total bandwidth used for communication between the user and the UAV. In the... The first time slot Within a time frame, the user The achievable transmission rate (bits / second) between the UAV and the UAV can be expressed as:

[0014]

[0015] in Indicates user Occupied bandwidth Indicates user signal-to-noise ratio, It represents the Transmission Time Interval (TTI). For transmission of finite block length, the bit error rate is given. Represents the channel dispersion function. It is a complementary cumulative error function The inverse function of .

[0016] Preferably, the UAV transmits integrated sensing signals to simultaneously achieve data transmission to fog nodes and target perception. The first time slot within the [number] time slot Within each time frame, the UAV's transmitted signal is represented as:

[0017]

[0018] in For signals containing user computation task data, The signal contains target perception data. Indicates a dedicated sensing signal. , Vector for shaping the transmitted beam of communication signals. Transmit beamforming vectors for sensing signals.

[0019] Preferably, the fog node decodes user computing task data and perception data using Successive Interference Cancellation (SIC) technology.

[0020] Preferably, the UAV uses a frequency band that does not overlap with the user's for communication with fog nodes and target perception, with the bandwidth expressed as follows: The achievable transmission rates of user computing task data and perception data from UAVs to fog nodes are both modeled using finite block length transmission formulas, expressed as follows:

[0021]

[0022]

[0023] in and These represent the signal-to-noise ratios of the user data signal component and the sensed data signal component, respectively. and This represents the corresponding channel dispersion function. and These represent the transmission error rates of user computation task data and perception data, respectively.

[0024] Preferably, the signal-to-noise ratio of the echo signal from the UAV's transmitted signal reaching the target via the direct path and being reflected by the target is used as a sensing performance indicator, which can be expressed as:

[0025]

[0026] in This represents the gain matrix of the sensed echo signal. This represents the power spectral density of additive noise.

[0027] Preferably, the user The number of computational task data packets generated in each time frame is modeled as the mean. The Poisson process, the data size of a single data packet is (Bits). Each user uses a First-In-First-Served (FIFS) queue to buffer arriving packets. Definition Indicates user In the The first time slot Number of computation task data packets arriving within a time frame, time slot Internal users The update of the queue can be represented as

[0028]

[0029] Preferably, the maximum tolerable queuing delay for users in computing tasks is measured using the equivalent bandwidth theory. The maximum tolerable queuing delay violation probability is The required service rate is expressed as

[0030]

[0031] To meet the queue delay threshold With the probability of violation Users within each time frame The number of serviceable packets in the end queue should be no less than Therefore, time slot The A time frame from the user The number of packets that leave the queue can be expressed as The number of data packets received by the UAV is equal to the sum of the number of offload data packets from all users, which can be expressed as: .

[0032] Preferably, by jointly optimizing UAV trajectory design, adjusting data packet loss probability, and allocating system resources, an optimization problem is constructed with the goal of minimizing the weighted average power consumption of the system, while ensuring the latency and reliability of user computing task processing and target perception performance.

[0033]

[0034] in and Representing users respectively With UAV in time slot power consumption, and These represent the drone's launch power consumption and computing power consumption, respectively. Indicates user transmit power consumption, and These represent the computing resources allocated to the UAV for processing user computing task data and perception data locally, respectively. To optimize the set of variables, This is a set of system reliability parameters. and These are the user power consumption weight parameters and the UAV power consumption weight parameters, used to control the focus of power consumption optimization. and These represent the maximum power of the UAV and the user, respectively. To perceive the minimum threshold of signal-to-noise ratio This indicates the maximum flight speed of the UAV. and Indicates the start and end positions of the UAV. This indicates the maximum computing power of the drone's local edge computing server. This indicates the duration of each time slot on a large time scale. This represents the number of CPU cycles required to process 1 bit of user computation task data. This represents the amount of target perception data that the drone needs to process within each time frame.

[0035] Preferably, the optimization problem is decomposed into a UAV trajectory design and data packet loss probability adjustment subproblem on a large time scale and a system resource allocation problem on a small time scale.

[0036] Preferably, on a large time scale, the UAV trajectory design and packet loss probability adjustment subproblems are reconstructed as Markov processes. Definition Indicates the agent in the time slot The observed state space includes the UAV's current location, channel states with users, targets, and fog nodes, and queue backlog for each user at the start of the time slot; [Definition] Indicates the agent in the time slot The action space, which includes the UAV in the time slot Flight direction, flight speed, and transmission error rate , With queuing delay violation probability Percentage of total data packet loss Proportion; Definition Indicates the agent in the time slot The reward It is calculated by inverting the weighted average power consumption of the system after optimization on a small time scale and adding a penalty term based on trajectory constraints.

[0037] Preferably, the SAC algorithm is used to solve the subproblem of UAV trajectory design and data packet loss probability adjustment on a large time scale. Its training process mainly includes: network parameter initialization; agent interaction with the environment to gain experience; obtaining actions from the output of the stochastic policy network through reparameterization techniques; updating the policy network parameters and evaluation network parameters by maximizing the entropy function and minimizing the mean square error function based on stochastic gradient descent; and updating the temperature control parameters based on stochastic gradient descent.

[0038] Preferably, on a small time scale, the system resource allocation problem is further decomposed into mutually independent user frequency band allocation sub-problems and UAV transmit beamforming and computational resource allocation sub-problems.

[0039] Preferably, on a small time scale, in time slots Solving for each time frame using user power sets With frequency band allocation set This is a subproblem involving the allocation of user frequency bands to optimize variables. To solve this problem, in each iteration, we first... Its concave boundary is obtained by performing a first-order Taylor expansion. The optimization problem Replace with Then, the problem transforms into a convex problem. An auxiliary variable is introduced. Then, the convex problem is equivalently transformed. The augmented Lagrangian function of the transformed problem is constructed. And based on the ADMM method, the original variables are updated alternately. and Auxiliary variables and dual variables , No. The update step in the next iteration can be represented as follows:

[0040]

[0041] The original variable update is achieved through a binary search method, while the auxiliary variable update is based on a closed expression derived from the Karush-Kuhn-Tucker (KKT) conditions.

[0042] Preferably, on a small time scale, in time slots Solving for each time frame using the UAV transmit beamforming vector , , Computational resource allocation parameters , This paper presents a subproblem of UAV transmit beamforming and computational resource allocation with variable variability. To solve this problem with lower computational complexity, the transmit beamforming vector is first constructed using the MRT-based method:

[0043]

[0044]

[0045]

[0046] in , These are the channel gain vectors between the UAV and the fog node, and between the UAV and the target, respectively. , , These represent the transmit power of each signal component that needs further optimization.

[0047] The problem is rephrased as a power optimization and computational resource allocation problem, and a suboptimal feasible solution is obtained in three steps. The first step is to obtain the optimized solution using a binary search method. and The second step is to obtain the optimized result using a binary search method. and The third step is to obtain the optimized expression through a closed-form expression. Obtain the optimized , , Then, the transmit beamforming vector of the UAV can be calculated according to the beamforming vector formula based on MRT given above.

[0048] The synthetic computing resource collaborative allocation method with service quality assurance described in this invention, compared with the prior art, has the following technical advantages:

[0049] This invention presents a collaborative allocation method for sensory computing resources with quality of service assurance. Under the premise of ensuring low latency and high reliability service quality requirements for user computing tasks and meeting target perception performance, it aims to minimize the system's weighted average power consumption. On a large time scale, the SAC algorithm is used for UAV trajectory design and data packet loss probability adjustment. On a small time scale, SCA and ADMM methods are used for user frequency band resource allocation, and a low-complexity method based on MRT is used for UAV transmission beamforming and computing resource allocation. Compared with existing UAV trajectory design and resource allocation methods, this invention considers the inconsistency of UAV trajectory design and resource allocation on a time scale in high-reliability, low-latency scenarios, takes into account the randomness of the amount of computing task data generated by the user end, and incorporates the transmission error rate of finite block length coding and the probability of computing task queuing delay violation into the overall data packet loss probability as a reliability indicator. This is more in line with the actual characteristics of high-reliability, low-latency services and can achieve lower system weighted average power consumption in scenarios where computing tasks arrive randomly. When designing optimization methods for small time scales, this invention significantly reduces the computational complexity of user and UAV resource allocation optimization at small time scales by using the parallelizable and more efficient ADMM method, as well as the low-complexity beamforming and computational resource allocation method based on MRT that does not require iteration. This makes the invention have significant practical application value in industrial IoT scenarios such as smart grids and smart factories that require deterministic ultra-low latency and high reliability service quality assurance. Attached Figure Description

[0050] Figure 1 This is a diagram of a UAV-assisted integrated sensory computing system architecture with guaranteed service quality.

[0051] Figure 2This is a diagram of the system time slot partitioning structure based on dual time scales.

[0052] Figure 3 This is an implementation flowchart of a collaborative allocation method for synergistic computing resources with service quality assurance.

[0053] Figure 4 This is a flowchart illustrating the implementation of a user frequency band resource allocation method on a small time scale.

[0054] Figure 5 This is a flowchart illustrating the implementation of UAV transmission beamforming and computational resource allocation methods on a small timescale. Detailed Implementation

[0055] The specific embodiments of the present invention will be further described in detail below with reference to the accompanying drawings and the UAV-assisted integrated sensing and computing scenario.

[0056] like Figure 1 As shown, this example employs a drone-assisted integrated sensing and computing system architecture with quality of service assurance, which includes a UAV. One user, one target, and one fog node. The UAV is equipped with an edge computing server and two array elements. Uniform Linear Arrays (ULAs) are used for both signal transmission and reception. Both users and fog nodes are equipped with a single antenna. Figure 1 In the diagram, two types of arrows are used to indicate user computing task offloading signals and integrated communication and sensing signals, respectively: ground users transmit randomly generated computing task data packets to UAVs via offloading signals; UAVs generate integrated communication and sensing signals, one of which forwards user computing task data to fog nodes, and the other uses signal echoes to perceive ground targets.

[0057] like Figure 2 As shown, this example employs a dual-time-scale time-slot partitioning structure. The UAV's service time is divided into equal-length segments. There are 1 time slot, and the length of each time slot is 1. The time slot set is Each time slot is further divided into There are 1 time frames, each time frame having a length of 1. The time frame set is . Figure 2 In this invention, each time frame is divided into two stages according to the execution order: the control signaling transmission stage and the data transmission stage. The invention employs a two-layer time decoupling architecture: at the time slot level (large time scale), the UAV flight trajectory and reliability parameters are optimized; at the time frame level (small time scale), wireless and computing resources are dynamically allocated.

[0058] The position of the UAV within each time slot can be approximated as constant, represented as: ,in This represents the fixed flight altitude of the UAV. The user set is... ,user The position is represented as The target's location is represented as... The location of the fog node is represented as Within each time frame of each time slot, each user generates a random number of computation task data packets, which are first buffered in their respective computation task queues. In each time frame, the user offloads the computation task data packets from their queue to the UAV for processing. The UAV uses its local computing resources to process a portion of the user's computation task data packets, while simultaneously sending a signal to offload another portion of the data packets to the fog node. The echo of this transmitted signal, reflected by the target, is received by the UAV and used to acquire perception information about the target. To ensure low latency and high reliability of user computation task processing, as well as target perception performance, within the limits of power, spectrum, computing resources, and UAV maneuverability, and to minimize the system's weighted average power consumption, the following steps are required:

[0059] Step 1: Construct a UAV-assisted integrated communication and computing system architecture with quality of service assurance, and calculate the communication transmission rate between the UAV and the user, and between the UAV and the fog node, based on the short packet transmission formula. Calculate the signal-to-noise ratio of the echo of the UAV's transmitted signal reflected by the target.

[0060] (1) UAV and users The channel gain vector between them is modeled using a probabilistic channel model and can be expressed as:

[0061]

[0062] in and Representing UAV and user respectively There are probabilities of direct and indirect paths. Indicates UAV and user The distance between them. This is the reference gain when the transmitter and receiver are 1m apart. For users Between with UAV 3D array emission guide vector, For users The UAV follows a complex Gaussian distribution with mean 1 and covariance matrix equal to the identity matrix. The NLoS component.

[0063] To avoid mutual interference, each user employs a Frequency Division Multiple Access (FDMA) scheme to simultaneously offload computation task data packets. In time slots... The Each time frame, the UAV receives data from the user. The received signal-to-noise ratio of the signal can be expressed as:

[0064]

[0065] in For users The transmission power, The total bandwidth used for communication between the user and the UAV. For users The proportion of bandwidth used for communication with UAVs Let be the power spectral density of the noise.

[0066] To ensure low-latency data transmission, the communication data between the user and the UAV uses a finite block length coding scheme, and the impact of the transmission error rate cannot be ignored. A formula for the finite block length signal transmission rate is used to describe the transmission rate between the user and the UAV. In the... The first time slot Within a time frame, the user The achievable transmission rate (bits / second) between the UAV and the UAV can be expressed as:

[0067]

[0068] in It represents the Transmission Time Interval (TTI). For transmission of finite block length, the bit error rate is given. Represents the channel dispersion function. It is a complementary cumulative error function The inverse function of .

[0069] (2) The channel gain vector between the UAV and the fog node is modeled using a probabilistic channel model, and can be expressed as follows:

[0070]

[0071] in and These represent the probabilities of a direct path and an indirect path existing between the UAV and the fog node, respectively. This represents the distance between the UAV and the fog node. For UAV and fog node 3D array emission guide vector, The UAV and fog nodes follow a complex Gaussian distribution with mean 1 and covariance matrix equal to the identity matrix. The NLoS component.

[0072] In the The first time slot Within each time frame, the data transmitted by the UAV to the fog node includes both user computation task data and perception data. To avoid interference from user-transmitted signals, the UAV uses a frequency band that does not overlap with the user's for communication with the fog node and target perception; the bandwidth is [missing information]. The UAV transmits integrated sensing signals to simultaneously achieve data transmission to fog nodes and target perception. The first time slot within the [number] time slot Within each time frame, the UAV's transmitted signal is represented as:

[0073]

[0074] in For signals containing user computation task data, The signal contains target perception data. Indicates a dedicated sensing signal. , Vector for shaping the transmitted beam of communication signals. The beamforming vector is transmitted to sense the signal. After receiving the signal, the fog node first... Decoding is performed, taking into account the dedicated sensing signal. Since the signal is deterministic and the fog nodes are known, the dedicated sensing signal component can be removed from the received signal using Successive Interference Cancellation (SIC) technology. Therefore, the signal... The received signal-to-noise ratio can be expressed as

[0075]

[0076] After decoding to obtain the computation task data, the fog node uses SIC to... Eliminate and to Decoding is performed, therefore the signal The received signal-to-noise ratio can be expressed as

[0077]

[0078] The achievable transmission rate (bits / second) of UAV transmitting user computing task data and sensing data to fog nodes is modeled using the finite block length transmission formula, and is expressed as follows:

[0079]

[0080]

[0081] in , Represents the channel dispersion function. and These represent the transmission error rates of user computation task data and perception data, respectively.

[0082] (3) The UAV senses the target by transmitting and reflecting the signal echo through the direct-path link. The echo signal gain matrix between the UAV and the target can be expressed as follows:

[0083]

[0084] in This represents the transmission signal steering vector between the UAV and the target. The signal-to-noise ratio (SNR) of the UAV's perception of the target can be expressed as...

[0085]

[0086] Step 2: Based on the equivalent bandwidth theory, construct constraints on the processing latency and reliability of user computing tasks, and formulate an optimization problem with the objective of minimizing the system's weighted average power consumption. The joint optimization problem constructed in this step is... Figure 3 The initial optimization model.

[0087] (1) The generation process of data packets for each user terminal's computing task is modeled as mutually independent with an average rate of The Poisson arrival process of (data packets / time frames). The data size of each computation task data packet is... (bit), calculate the number of CPU cycles required to process 1 bit of data. Each user uses a First-In-First-Served (FIFS) queue to buffer arriving packets. (Time slot) Internal users The update of the queue can be represented as

[0088]

[0089] in Indicates user In the The first time slot Number of computation task data packets arriving within a time frame. User The equivalent bandwidth (data packets / second) of the queue arrival process can be expressed as:

[0090]

[0091] in The maximum tolerable queuing delay for the user. This represents the probability that the queuing delay exceeds the maximum tolerable queuing delay. This is to meet the queuing delay threshold. With the probability of violation Users within each time frame The number of serviceable packets in the end queue should be no less than Therefore, time slot The A time frame from the user The number of packets that leave the queue can be expressed as To satisfy this departure process, users in each time frame... The number of data packets transmitted between the UAV and the UAV needs to meet the following requirements.

[0092]

[0093] In the time slot The Within a given time frame, the number of data packets received by the UAV is equal to the sum of the number of offload data packets from all users, which can be expressed as: To ensure that all data packets received by the UAV in the previous time frame can be processed in the current time frame, the number of data packets processed by the UAV in each time frame needs to meet the following requirements.

[0094]

[0095] in Allocate computing resources to the UAV for local processing of user computing task data packets.

[0096] The total latency experienced by the user's computing task data packets can be expressed as: ,in This indicates the transmission delay of data packets that the user unloads to the UAV. This indicates the processing delay of the UAV after receiving the data packet. Both delays do not exceed the length of one time frame. Therefore, to ensure that the total delay is less than [a certain value], [further details are needed]. Set the maximum tolerable queuing delay of data packets in the user queue to Given the maximum tolerable queuing delay The total data packet loss probability during user data packet offloading and processing includes the transmission error probability and the queue delay violation probability. To ensure the reliability of user task processing, it needs to meet the following requirements.

[0097]

[0098] Considering the relatively low probability of violation under high-reliability, low-latency services, this constraint can be approximated by eliminating the product term between violation probabilities.

[0099]

[0100] (2) To ensure that the sensing data collected in the current time frame can be fully processed, the sensing data processing capability of the UAV needs to meet the following constraints.

[0101]

[0102] in This indicates that the UAV is allocated computing resources for processing sensing data locally. This represents the number of CPU cycles required to process 1 bit of sensor data. This represents the amount of sensing data (bits) that can be obtained within each time frame. Its value is determined by actual radar design parameters such as the sensing echo sampling frequency and the number of quantization bits.

[0103] Since the fog node first decodes the signal component corresponding to the user data packet after receiving the UAV's transmission signal, and then decodes the signal component corresponding to the sensing data, the transmission bit error rate of the sensing data can be expressed as: To ensure reliability, it needs to meet the following requirements. It can be approximated as

[0104]

[0105] (3) In the time slot Inside, users The average power consumption is mainly composed of the transmit power, expressed as:

[0106]

[0107] The power consumption parameter is expressed as In the time slot The first Within each time frame, the power consumption for signal transmission and computation at the UAV end are respectively expressed as follows:

[0108]

[0109]

[0110] UAV in time slots The flight power consumption can be expressed as

[0111]

[0112] in , , , , These are all aerodynamic parameters. For UAV in time slots Flight speed.

[0113] Therefore, UAV in time slots The average power consumption within can be expressed as

[0114]

[0115] in It is a fixed quantization parameter used to make the flight power consumption similar in value to other power consumption.

[0116] Considering the limited energy reserves of UAVs and users, by jointly optimizing UAV trajectory design and system resource allocation, and while ensuring the latency and reliability of user computing tasks and target perception performance, the following optimization problem is constructed with the objective of minimizing the weighted average power consumption of the system:

[0117]

[0118] in To optimize the set of variables, For the set of packet loss probabilities, and These are the user power consumption weight parameters and the UAV power consumption weight parameters, used to control the focus of power consumption optimization. and These represent the maximum power of the UAV and the user, respectively. To perceive the minimum threshold of signal-to-noise ratio This indicates the maximum flight speed of the UAV. and These represent the start and end positions of the UAV, respectively. This indicates the maximum computing power of the drone's local edge computing server. This indicates the duration of each time slot on a large time scale. This represents the number of CPU cycles required to process 1 bit of user computation task data. This represents the amount of target perception data that the drone needs to process within each time frame.

[0119] Step 3: Decompose the optimization problem into a large-scale UAV trajectory design and data packet loss probability adjustment sub-problem and a small-scale system resource allocation sub-problem. For example... Figure 3As shown, this invention employs a dual-timescale hierarchical iterative optimization: at the time slot level (large time scale), the SAC algorithm is used to optimize the UAV trajectory and reliability parameters, while at the time frame level (small time scale), fine-grained user and UAV resource allocation is performed.

[0120] (1) On a large time scale, the UAV trajectory design and data packet loss probability adjustment subproblem is expressed as:

[0121]

[0122] (2) On a small time scale, time slot The The subproblem of system resource allocation for a given time frame is represented as follows:

[0123]

[0124] in To optimize variables.

[0125] Step 4: On a large time scale, use the SAC algorithm to design UAV trajectories and adjust the data packet loss probability. For example... Figure 3 As shown in the upper part, this step involves the reinforcement learning agent interacting with the environment to output the UAV flight trajectory and data packet loss probability settings for the current time slot. The Markov process modeling and network parameter training process is as follows:

[0126] (1) The UAV trajectory design and data packet loss probability adjustment subproblem on a large time scale are equivalently transformed into a Markov Decision Process (MDP). The state space, action space and reward of the MDP are described below.

[0127] State space: definition Indicates the agent in the time slot The observed state space, which includes the UAV's current location, the channel states with the user, the target, and the fog nodes, and the queue backlog of each user at the start of the time slot, is represented as follows:

[0128]

[0129] Action space: definition These represent the UAV in the time slots. The flight direction and speed. From the optimization problem, it is easy to see that when all reliability constraints are equal, i.e. and To achieve the lowest average power consumption, therefore, it is defined as... They represent , , Percentage of total data packet loss proportion, satisfy , .definition Indicates time slot Action space:

[0130]

[0131] The actual transmission error rate and queuing delay violation probability can be calculated using the following method: , , as well as .

[0132] Reward function: Definition Indicates the agent in the time slot The reward will be a time slot The reward function is the sum of the inverse of the weighted average power consumption of the system after small-scale optimization and the penalty term based on trajectory constraints.

[0133]

[0134] The inversion of the system's weighted average power consumption can be achieved by time-slotting the data on a small time scale. Steps 5-7 within each time frame calculate the optimal power consumption of the UAV and the user, then perform a weighted summation, take the average, and invert the result to obtain the final power consumption. Indicates the weight of the penalty item. This represents a penalty term based on trajectory constraints, used to ensure that the UAV trajectory flies from the starting point to the destination. Specifically, it can be expressed as:

[0135]

[0136] in Indicates UAV in time slot Distance to the finish line This indicates the maximum distance a UAV can fly during the remaining service time.

[0137] (2) Based on the MDP described in (1), the training process of the SAC algorithm under a large time scale is specifically described as follows.

[0138] Initialization: The SAC algorithm mainly consists of five deep neural networks, including a policy network. Evaluation Network and and target evaluation network and At the beginning of training, the weight parameters of the deep neural network are... , , , , Perform random initialization.

[0139] Experience collection: A complete service cycle is defined as one round of SAC algorithm training, and each time slot within the service cycle is defined as one time step of SAC algorithm training. At each time step, the policy network... Receive the state space of the current time step Output the mean of the probability distribution of consecutive actions. With variance Then, the reparameterization technique is used to obtain the actual action at the current time step:

[0140]

[0141] in Let be a random variable that follows a standard normal distribution. This is the activation function.

[0142] System Execution Action Calculate rewards Then the environment is updated to the next time step. . tuple Store in the experience replay pool middle.

[0143] Network parameter update: from the experience replay pool Randomly selected from The loss function of the policy network is calculated using the following tuples:

[0144]

[0145] in Indicates temperature parameter, Represents action entropy. This indicates the evaluation value of the state and action output by the network.

[0146] The loss function for evaluating the network is calculated as follows:

[0147]

[0148] in

[0149]

[0150] in For reward discount parameters, This indicates the value of the target state action output by the target evaluation network.

[0151] Based on loss function and The network parameters are updated using the stochastic gradient descent method. , , The target evaluation network parameters are updated by taking a weighted average with the evaluation network parameters, and can be expressed as follows: ,in For smoothing parameters.

[0152] Temperature parameters The update is based on the following loss function and uses stochastic gradient descent:

[0153]

[0154] in It equals the negative value of the action space dimension.

[0155] By repeatedly performing experience collection and network parameter updates in multiple rounds, the SAC algorithm can eventually converge and a policy network can be obtained that can output reasonable UAV trajectories and packet loss probabilities based on the current time slot environment.

[0156] Step 5: At a smaller time scale, the system resource optimization problem for each time frame within each time slot is further decomposed into a user bandwidth allocation sub-problem and a UAV transmit beamforming and computational resource allocation sub-problem. For example... Figure 3 As shown in the lower half, these two sub-problems correspond to... Figure 4 and Figure 5 The detailed process includes the following steps:

[0157] (1) In time slot The One time frame, UAV position With packet loss probability This has already been obtained using the SAC algorithm in step 3. The user's power consumption is only related to the inter-user frequency band allocation variable. The user frequency band resource allocation subproblem can be expressed as:

[0158]

[0159] in , , It can be calculated by the equivalent bandwidth and the backlog of user task queues in the current time frame.

[0160] (2) In time slots The One time frame, UAV position With packet loss probability settings The power consumption of the UAV has been obtained using the SAC algorithm in step 3. The power consumption of the UAV is only related to the transmit beamforming and local computing resource allocation. The subproblem of UAV transmit beamforming and computing resource allocation can be expressed as:

[0161]

[0162] in It can be obtained by calculating the sum of user-unloaded computing task data in the previous time frame.

[0163] Step 6: Allocate user frequency band resources using the SCA method and ADMM algorithm. For example... Figure 4 As shown, the process includes: parameter initialization, first-order Taylor expansion to convexity problem, construction of augmented Lagrangian function, ADMM iterative update (binary search to update original variables, KKT closed-form solution to update auxiliary variables, update dual variables), two-level convergence determination, and finally output of optimal frequency band and power allocation. Specific steps include:

[0164] (1) Since the user transmit signal received by the UAV under the frequency division multiplexing scheme has a high signal-to-noise ratio, the user in the optimization problem will be optimized. Formula for short packet transmission rate of data transmitted to UAV In The term is approximately 1. For the approximated term... The second item in the text is about The first-order Taylor expansion yields its concave lower bound:

[0165]

[0166] in It is a constant. Let the Taylor expansion point be set as follows: The Taylor expansion point in the first iteration is set to... The Taylor expansion points in subsequent iterations are all derived from the optimization obtained in the previous iteration. .

[0167] In each iteration, the optimization problem will be... Replace with Then, the problem is transformed into a convex problem.

[0168] (2) To reduce the computational complexity of solving this convex problem, the ADMM method is adopted. Define auxiliary variables. The original problem is equivalently transformed into

[0169]

[0170] Construct the augmented Lagrangian function for this problem, expressed as:

[0171]

[0172] in As dual variables, The penalty factor is determined by the ADMM method, and iterative updates are performed according to the following steps. , , as well as .

[0173]

[0174] in The index represents the iteration number. The specific update method for each step is explained below.

[0175] First, solve the following: A parallel problem to update and :

[0176]

[0177] To address this issue, frequency band allocation parameters The optimal solution can be obtained by using a binary search method to solve the following problem.

[0178]

[0179] in

[0180]

[0181] User power consumption The optimal solution can be calculated as follows: .

[0182] Then, based on the Karush-Kuhn-Tucker (KKT) conditions, update as follows: :

[0183]

[0184] Finally, based on the ADMM standard update process, update as follows: :

[0185]

[0186] After multiple rounds of iterative updates, the algorithm converges and outputs the optimal value. and .

[0187] Step 7: Within each time slot, a low-complexity, non-iterative method is used for UAV transmission beamforming and computational resource allocation in each time frame. For example... Figure 5As shown, the process first uses MRT to construct the beamforming vector, and then solves it in three steps in a fixed order: the first step uses binary search to solve for the signal power of the sensing data transmission and the computational resources for local processing of the sensing data; the second step uses binary search to solve for the signal power of the user task data transmission and the computational resources for local processing of the task data; the third step uses closed-form solution to solve for the dedicated sensing signal power, and finally outputs the beamforming vector. The entire process is iterative and has low computational complexity. Specific steps include:

[0188] (1) Solving this problem using successive convex approximation and CVX toolbox will result in high complexity, which is not conducive to deployment in high reliability and low latency scenarios. Therefore, a low complexity method based on MRT is adopted to obtain a suboptimal feasible solution to the problem.

[0189] Due to the high flight altitude of the UAV, the channel gain parameters between it and the user and the target are dominated by the direct path component. In order to maximize the beam gain of the transmitted signal in both the fog node direction and the target direction, the transmit beamforming vector is first constructed using the MRT method:

[0190]

[0191]

[0192]

[0193] in , , These represent the transmit power of each signal component that needs further optimization.

[0194] Will and In and After approximating all values ​​to 1, the problem can be reformulated as a power optimization and computational resource allocation problem.

[0195]

[0196] in , , Next, the suboptimal feasible solution to this problem is obtained by following the procedure below.

[0197] First, solve the following problem to obtain the optimized result. and :

[0198]

[0199] This problem is a convex problem. By analyzing its KKT conditions, the optimal solution can be found. It can be represented as

[0200]

[0201] in The following equation can be obtained by solving the binary search method:

[0202]

[0203] Obtain the optimal solution Then, the optimal solution It can be represented as

[0204]

[0205] Secondly, solve the following problem to obtain the optimized result. and :

[0206]

[0207] This problem is also a convex problem. By analyzing its KKT conditions, its optimal solution... It can be represented as

[0208]

[0209] in The following equation can be obtained by solving the binary search method:

[0210]

[0211] Obtain the optimal solution Then, the optimal solution It can be represented as

[0212]

[0213] Finally, solving the following linear programming problem yields the optimized result.

[0214]

[0215] Optimal solution It can be represented as

[0216]

[0217] Obtain the optimized , , Then, the transmit beamforming vector of the UAV can be calculated according to the beamforming vector formula based on MRT given above.

[0218] This example focuses on a drone-assisted sensor-computing system with high reliability, low latency, and quality of service requirements. Based on 6G communication technology, reinforcement learning theory, and numerical optimization methods, it aims to minimize the system's weighted average power consumption. While ensuring low latency and high reliability for user computing tasks and target perception performance, a dual-timescale UAV trajectory design and system resource allocation architecture is constructed, and optimization schemes are proposed. At the large timescale, the SAC algorithm is used for UAV trajectory design and packet loss probability adjustment; at the small timescale, SAC and ADMM methods are used to solve the user bandwidth allocation problem, and a low-complexity method without iteration is used to solve the UAV transmit beamforming and computing resource allocation problem. The dual-timescale algorithm design significantly reduces the system's weighted average power consumption, ensures low latency and high reliability for user computing tasks and target perception performance, and has low computational complexity. This provides an efficient and reliable solution for trajectory optimization and collaborative resource allocation in drone-assisted sensor-computing systems with quality of service assurance.

[0219] The foregoing has described specific embodiments of this specification. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps recited in the claims may be performed in a different order than that shown in the embodiments and may still achieve the desired result. Furthermore, the processes depicted in the drawings do not necessarily require the specific or sequential order shown to achieve the desired result. In some embodiments, multitasking and parallel processing are possible or may be advantageous.

Claims

1. A method for collaborative allocation of synesthetic computing resources with quality of service assurance, characterized in that, Includes the following steps: Construct a drone-assisted integrated sensing and computing system architecture with quality of service assurance; In the system architecture, the total service time is divided into multiple large time scale units, and each large time scale unit is further divided into multiple small time scale units. Ground users generate computing tasks in each small time scale unit and offload them to the drone. The drone uses local computing resources to process part of the tasks, while transmitting integrated sensory signals to forward the remaining tasks to the fog node and perceive ground targets. Under the constraints that the end-to-end latency of user task processing does not exceed the maximum tolerable latency, the total probability of data packet loss does not exceed the maximum tolerable data packet loss probability, and the target perception signal-to-noise ratio is not lower than a preset threshold, a trajectory design and resource allocation optimization problem is constructed with the goal of minimizing the system weighted average power consumption. The optimization problem is decomposed into a large-scale UAV trajectory design and data packet loss probability adjustment sub-problem, and a small-scale system resource optimization problem. On a large time scale, reinforcement learning methods are used to solve the UAV trajectory design and reliability parameter adjustment subproblems to obtain the UAV flight trajectory and reliability parameters; On a small time scale, the system resource optimization problem is further decomposed into a first sub-problem and a second sub-problem, wherein the first sub-problem involves the allocation of transmission resources on the user side, and the second sub-problem involves the transmission beamforming and computing resource allocation on the UAV side. Solving the first subproblem on a small time scale yields a user transmission resource allocation scheme; Solving the second subproblem on a small time scale yields the UAV launch beamforming vector and computational resource allocation scheme; After completing the optimization of the large time scale unit, output the UAV trajectory and the resource allocation scheme for each small time scale unit; The solution of the first subproblem includes the use of successive convex approximation and alternating direction multiplier method; specifically, in each small time scale unit, the user transmission rate expression is convexly approximated to transform the original problem into a convex problem. After constructing the augmented Lagrangian function, the user transmission resource variables, auxiliary variables and dual variables are alternately updated until convergence. Solving the second subproblem involves employing a low-complexity method based on maximum ratio transmission. Specifically, this involves constructing a task data communication beamforming vector, a sensing data communication beamforming vector, and a dedicated sensing beamforming vector using the maximum ratio transmission method, transforming the beamforming problem into a transmit power optimization problem. A binary search is used to determine the sensing signal transmit power and sensing computing resources, as well as the task data transmit power and task computing resources. A closed-form solution is used to determine the dedicated sensing signal transmit power. The optimized power is then substituted into the maximum ratio transmission formula to obtain the final transmit beamforming vector.

2. The method according to claim 1, characterized in that, In the aforementioned drone-assisted integrated sensing and computing system architecture: The large time scale unit is a time slot, and the small time scale unit is a time frame; Users generate a random number of computation task data packets in each time frame, and each user uses a first-come, first-served queue to cache the computation task data packets they generate. The user adopts a finite block length coding method and, based on frequency division multiple access technology, simultaneously unloads computational task data to the UAV in each time frame; In each time frame, the drone uses its own computing resources to process a portion of the computing task data offloaded from the user, and sends a signal to further offload another portion of the computing task data to the ground fog node. The drone's transmission signals are used simultaneously to offload user computing task data packets to fog nodes and to sense ground targets.

3. The method according to claim 1, characterized in that, The service quality assurance mentioned includes: The end-to-end latency of user task data packet processing includes queuing latency, transmission latency, and processing latency; The total packet loss probability for user task packet processing includes the transmission error probability, the queuing delay violation probability, and the UAV transmission error probability.

4. The method according to claim 1, characterized in that, The reinforcement learning method is a flexible actor-commentator algorithm. The state space of the algorithm includes the current position of the UAV, the channel state, and the backlog of each user's computing task queue. The action space includes the UAV's flight speed, flight direction, and the proportion of each reliability parameter in the total data packet loss probability. The reward function is set as the sum of the inverse of the weighted average power consumption of the system after small time scale optimization and the penalty term based on trajectory constraints.

5. The method according to claim 4, characterized in that, The training process of the flexible actor-commentator algorithm includes: at the beginning of each large time scale unit, the agent outputs an action based on the current state, and the drone updates its position after executing the action; in each small time scale unit of each large time scale unit, the steps of solving the first and second sub-problems are executed to obtain the optimal power consumption of the user and the drone; the weighted average power consumption of the current large time scale unit is calculated and inverted as the reward, and the tuple consisting of state, action, reward, and next state is stored in the experience replay pool; random sampling is performed from the experience replay pool, the policy network parameters are updated by maximizing the entropy function, and the evaluation network parameters are updated by minimizing the loss function.

6. A sensor-computer integrated system with quality of service assurance, characterized in that, The method described in any one of claims 1-5 is used for UAV trajectory design and collaborative allocation of sensor computing resources to ensure the service quality of users’ high-reliability, low-latency services and meet the target perception performance requirements.

Images