Dynamic self-adaptive communication, sensing and calculation integrated method and system for space-air-ground integrated network

Abstract

The invention relates to a dynamic self-adaptive communication, sensing and calculation integrated method for an air-space-ground integrated network. The method comprises the following steps: S1, constructing an air-space-ground integrated network model fusing a low earth orbit satellite, an unmanned aerial vehicle and ground users / edge nodes; s2, establishing a cross-layer communication-perception-calculation integrated joint optimization model, and respectively forming two modes of binary unloading and partial unloading according to whether a task can be segmented or not; s3, for a binary unloading mode, realizing distributed solution and generating an unloading decision and a beam / power / bandwidth / computing resource scheme by using a dynamic adaptive prediction alternating direction multiplier (DAP-ADMM); and S4, aiming at a partial unloading mode, adopting a weighted minimum mean square error alternative optimization method, introducing a receiving equalizer, error weight and other auxiliary variables, executing block coordinate descent iteration, and outputting an unloading proportion, power distribution and a beam forming result. According to the DA-ISCC architecture and the optimization algorithm thereof provided by the invention, cross-domain calculation unloading with low time delay, high throughput and low overhead can be realized in a high-dynamic space-air-ground integrated network environment, and the DA-ISCC architecture and the optimization algorithm thereof can be widely applied to emergency communication, low-altitude logistics, ocean monitoring and other scenes.

Description

Technical Field

[0001] This invention belongs to the field of wireless communication network technology, specifically relating to a dynamic adaptive integrated communication, sensing, and computing method and system for integrated air-space-ground networks. Background Technology

[0002] With the rapid development of emerging applications such as the Internet of Things, autonomous driving, and extended reality, the demand for ubiquitous, high-performance communication, sensing, and computing services is growing. Traditional terrestrial networks have inherent limitations in terms of coverage, reliability, and deployment costs, making it difficult to meet the needs of future 6G networks for seamless global coverage and massive intelligent services.

[0003] Space-Air-Ground Integrated Networks (SAGINs), as a multi-layered heterogeneous network architecture, demonstrate enormous potential to provide wide-area, three-dimensional coverage by integrating high-orbit satellites, low-orbit satellites, high-altitude platforms, drones, and terrestrial cellular networks. However, this network architecture also brings new challenges: complex network topology, frequent changes in link states due to the high-speed movement of nodes (especially low-orbit satellites and drones), and time-varying and unstable wireless channel conditions.

[0004] Meanwhile, integrated sensing, communication, and computing (ISCC) has become a key technology for improving network resource utilization efficiency. By sharing hardware platforms and spectrum resources, ISCC enables network nodes to perform high-precision target perception (such as localization and tracking) and complex computational tasks while completing communication tasks. Applying ISCC technology to integrated air-space-ground networks can fully leverage its multi-platform and multi-angle collaborative advantages, providing strong support for various intelligent applications.

[0005] Achieving efficient task offloading in integrated air-space-ground networks is a current research hotspot. User equipment can offload computationally intensive and latency-sensitive tasks to UAVs or ground-based edge servers with greater computing power. Existing research mainly focuses on optimizing offloading decisions and resource allocation, such as computing resources and communication bandwidth, to minimize system energy consumption or latency. However, most of these methods are based on static or quasi-static network environment assumptions, ignoring the inherently high dynamic characteristics of integrated air-space-ground networks, such as the frequent switching of low-Earth orbit satellites due to orbital motion and the maneuvering flight of UAVs during missions. These dynamic changes can slow down the convergence speed of existing algorithms, or even prevent them from converging, making it difficult to guarantee service quality.

[0006] Furthermore, existing methods often fail to adequately optimize the joint operation of communication, sensing, and computation during modeling. For example, when allocating communication resources, their impact on sensing accuracy is not fully considered, and vice versa. This results in suboptimal resource allocation schemes, limiting the overall performance of the integrated communication, sensing, and computation system.

[0007] Therefore, there is an urgent need for a dynamic adaptive task offloading method that can adapt to the highly dynamic characteristics of integrated air-space-ground networks and can coordinate and optimize communication, sensing, and computing resources, in order to solve the problems of slow convergence and inability to guarantee service quality in dynamic environments in existing technologies. Summary of the Invention

[0008] To address the issues of slow convergence, insufficient robustness, and low efficiency of cross-layer resource collaboration in highly dynamic environments of integrated air-space-ground networks, this invention proposes a dynamic adaptive integrated communication-sensing-computing architecture for integrated air-space-ground networks. Joint optimization models are constructed under both binary offloading and partial offloading modes, and two efficient algorithms—the DAP-ADMM method and the weighted minimum mean square error method—are designed to achieve distributed solution, taking into account system utility, energy consumption, and sensing constraints.

[0009] This invention proposes a dynamic adaptive sensing and computing integration method for integrated air-space-ground networks, the method comprising:

[0010] S1. Construct an integrated air-space-ground network model that integrates low Earth orbit satellites, UAVs, and ground users / edge nodes, and uniformly represent the time-varying channels, node computing capabilities, energy budgets, and mission attributes of UAV-ground, low Earth orbit satellite-UAV, and backhaul links to support a hybrid access architecture of intra-cluster NOMA and inter-cluster SDMA.

[0011] S2. Establish a cross-layer communication-sensing-computing integrated joint optimization model. Based on whether the task is divisible, two modes are formed: binary offloading and partial offloading. With the goal of maximizing the weighted system utility, the task offloading and communication / sensing / computing resource allocation are jointly optimized, and power constraints, NOMA decoding order constraints, computing resource constraints, minimum signal-to-interference-plus-noise ratio constraints, and latency constraints are applied.

[0012] S3. For binary offloading mode, the Dynamic Adaptive Prediction Alternating Direction Multiplier Method (DAP-ADMM) is used. Through residual-aware penalty parameters and step size adaptive adjustment, local resource prediction driven by long short-term memory network and asynchronous communication update mechanism, the offloading decision and resource allocation are solved collaboratively, and the output is a scheme for beamforming, power control, bandwidth allocation and computational resource allocation.

[0013] S4. For partial offloading mode, a weighted minimum mean square error alternating optimization method is adopted, introducing auxiliary variables such as receiver equalizer and error weight, and performing block coordinate descent iteration to output the offloading ratio, power allocation and beamforming results.

[0014] Furthermore, the optimization model of S2 includes the following constraints and definitions:

[0015] S21. Communication and Backhaul Constraints: Effective Uplink Rate of User k Determined by intra-cluster NOMA and serial interference cancellation decoding, satisfying

[0016]

[0017] Backhaul link meets

[0018]

[0019] S22. Computing resource and power consumption constraints: CPU frequency allocated to drones / edge nodes satisfy And calculate power consumption

[0020]

[0021] S23, Perceptual Constraints: Perceptual SINR

[0022]

[0023] Must meet . in ;

[0024] S24, NOMA decoding order and physical layer constraints: satisfy the intra-cluster decoding order π and beam / power non-negativity, total power upper limit and inter-cluster beam interference suppression;

[0025] S25, Binary Unloading Variable and ; or the partial unloading rate variable r(·,k)≥0.

[0026] Furthermore, the binary unloading optimization problem P1 is defined as follows:

[0027]

[0028] And satisfy the aforementioned communication, backhaul, computation, sensing, and NOMA constraints, as well as the binary constraints; the partial offloading optimization problem P2 is defined as:

[0029]

[0030] All constraints in claim 2 are satisfied and r(·,k)≥0.

[0031] Furthermore, the DAP-ADMM of S3 includes:

[0032] S31. Variable relaxation and distributed decomposition: Relax the binary delta variable to [0,1], introduce consistency variables and Lagrange multipliers, construct an augmented Lagrange and divide the beam, power, offloading and computation subproblems according to nodes / clusters for parallel solution;

[0033] S32. Residual-aware penalty parameters and adaptive step size: ρ and step size η are dynamically updated based on the relationship between the original residual s(t) and the dual residual r(t). Increase ρ and decrease η when ρ is high, and vice versa, decrease ρ and moderately increase η.

[0034] S33, LSTM Local Resource Prediction: Using the local / neighbor decisions, channel and load timing of the past h time slots as input, predict the available computing power and key neighbor variables for the next time slot, which are used to replace expired information in local updates;

[0035] S34. Asynchronous Communication and Event-Triggered Synchronization: Each node performs N local updates before exchanging information once. Link congestion or topology disturbances trigger immediate synchronization. When the local processing delay exceeds a threshold... To maintain the stability of variables;

[0036] S35. Dual Update and Convergence: Update the multipliers according to the consistency error until the original and dual residuals are below the threshold ε or the maximum iteration is reached. .

[0037] Furthermore, the adaptive evolution of the penalty parameter is as follows:

[0038] ,

[0039] in, To initialize the penalty parameters, For the original residual, For dual residuals, The remaining time window for the task. This is the maximum allowed completion time. This strategy can be increased when node load is high and channel quality is poor. To accelerate convergence; to reduce the network's stability. To suppress oscillations.

[0040] Furthermore, the Long Short-Term Memory predictor outputs the available computing power range and feasible rate upper bound for the next time slot, which is used to prune the search space of the local subproblem and reduce cross-node communication; the predictor is pre-trained in the cloud and fine-tuned online by sliding the window to avoid overfitting and concept drift.

[0041] Furthermore, the weighted minimum mean square error alternating optimization of S4 includes:

[0042] S41, Weighted minimum mean square error equivalent: Converting rate maximization into minimizing the weighted mean square error by introducing a receiver equalizer u and error weight v;

[0043] S42, Block Coordinate Descent Update: In the t-th iteration, u(t) and v(t) are updated in a closed loop. Given (u, v), the convex quadratic problem of beamforming p(t+1) is solved, and given (p, u, v), the unloading ratio or allocation rate r(·,k) is updated.

[0044] S43. Introducing SINR Constraints The weighted minimum mean square error subproblem enables beamforming to improve communication efficiency while ensuring sensing accuracy.

[0045] S44. Using the objective function gain threshold Δ<ε as the convergence criterion, output the solution that satisfies the approximate KKT conditions.

[0046] Furthermore, the hybrid access adopts intra-cluster NOMA and inter-cluster SDMA: by constraining intra-cluster serial interference to eliminate decoding order π and inter-cluster beam approximately orthogonal design, cross-cluster interference is suppressed and spectral efficiency is improved; corresponding beam maps are generated in binary and partial offload modes respectively to adapt to user distribution.

[0047] Furthermore, the system utility is defined as follows:

[0048] ,

[0049] Where r(·,k) = f(·,k) / φ, f(·,k) represents the CPU frequency allocated to user k by the drone / edge, low-Earth orbit satellite, and cloud, respectively, and the weights are... The system is dynamically adjusted based on business priorities to ensure the computational speed and perceived reliability of high-priority tasks.

[0050] The present invention also relates to a dynamic adaptive communication and computing integrated system for an integrated air-space-ground network, comprising a computer module that applies the above-described method.

[0051] The proposed method exhibits fast convergence and high energy efficiency in dynamic integrated air-space-ground network environments: Under binary offloading, DAP-ADMM achieves residual stability to the order of 0.05 under topology perturbation conditions and has a convergence speed ≥60% higher than consensus ADMM, with average communication energy consumption lower than that of distributed baselines without asynchronicity / prediction; Under partial offloading, the weighted minimum mean square error converges within ≤100 iterations, and the average user computing rate remains at approximately 480–500 Mbit / s.

[0052] The method supports adaptive mode switching: when the remaining scheduling window ΔT is small or the link fluctuates drastically, DAP-ADMM is preferred to improve robustness; when the link is stable and can be finely scheduled, weighted minimum mean square error is preferred to improve system efficiency; the switching is based on a joint determination of the link stability index and the queue backlog threshold.

[0053] The deployment adopts a space-ground integrated cloud-edge collaborative control: the cloud side uniformly configures hyperparameters (long short-term memory window). (etc.) and service weight It is then distributed to drones / edge nodes; edge nodes are updated online and, in conjunction with event triggering, synchronously execute cross-layer resource orchestration and dynamic task routing, supporting hybrid access and collaborative computing for ground users and satellite users.

[0054] Beneficial effects

[0055] This invention proposes a converged sensing-computing-communication architecture for integrated air-space-ground networks and designs two corresponding optimization methods for computation offloading, providing a systematic solution for high-reliability, low-latency offloading of ISCC tasks in integrated air-space-ground networks. The specific technical effects of this invention are as follows:

[0056] At the architecture level, this invention proposes a dynamic adaptive architecture for ISCC in SAGINs (DA-ISCC) that integrates sensing, communication, and computing in integrated air-space-ground networks. It abstracts low-orbit satellites, UAVs, and ground edge nodes into a unified programmable resource pool, mathematically models the collaborative offloading process of cross-layer sensing and computing tasks, and realizes a unified scheduling model for sensing, communication, and computing in integrated air-space-ground networks for the first time.

[0057] To address the binary computation offloading problem, offloading decision-making and resource allocation are jointly modeled as a mixed integer programming problem with discrete variables and coupling constraints. A novel approach is proposed: the Dynamic Adaptive Predictive Alternating Direction Method of Multipliers (DAP-ADMM). This method employs a residual-aware dynamic step-size adjustment mechanism to maintain convergence stability under highly dynamic channels. A local resource prediction module based on LSTM is introduced to anticipate node load and suppress ineffective offloading. An asynchronous communication update strategy is designed to significantly reduce satellite link synchronization waiting time.

[0058] For the partial computational unloading problem, the original non-convex rate maximization problem is transformed into the equivalent weighted minimum mean square error (WMMSE) form. By updating auxiliary variables and optimizing block coordinates alternately, the optimal solution of Karush-Kuhn-Tucker (KKT) is efficiently approximated under the dual constraints of perception and computation.

[0059] In terms of experimental verification, on the self-constructed multi-user, multi-layer integrated air-space-ground simulation network, compared with traditional and consensus ADMM, DAP-ADMM can achieve stable convergence when the residual is less than 0.05, with a convergence speed improvement of more than 60% and a reduction of 1-2 orders of magnitude in the number of communication rounds; the weighted minimum mean square error method converges to a stable computing rate of more than 480 Mbit / s within 100 iterations, which is significantly better than existing heuristic algorithms.

[0060] Compared with existing methods, the DA-ISCC architecture and its optimization algorithm proposed in this invention can achieve "low latency, high throughput, and low overhead" cross-domain computing offloading in a highly dynamic integrated air-space-ground network environment, and can be widely applied to scenarios such as emergency communication, low-altitude logistics, and ocean monitoring. Attached Figure Description

[0061] Figure 1 This is a schematic diagram of the overall dynamic adaptive architecture of the sensor-computer fusion for the integrated air-space-ground network in this invention.

[0062] Figure 2 This is a diagram of the distributed binary offloading modular architecture based on DAP-ADMM in this invention;

[0063] Figure 3 This is a convergence curve of the DAP-ADMM method in this invention;

[0064] Figure 4 This is a comparison diagram of the beam patterns of NOMA and SDMA in the binary offloading scenario of this invention;

[0065] Figure 5 This is a comparison chart of the convergence performance of DAP-ADMM and the baseline method in this invention;

[0066] Figure 6 This is a schematic diagram illustrating the convergence performance under different timing strategies in this invention;

[0067] Figure 7 The dynamic penalty parameters of the DAP-ADMM method in this invention A schematic diagram of the adaptive evolution process;

[0068] Figure 8This is a schematic diagram comparing the average communication energy consumption under the dynamic air-space-ground integrated simulation network in this invention;

[0069] Figure 9 This is a schematic diagram comparing the residual convergence of DAP-ADMM and consensus ADMM with and without LSTM local resource prediction in this invention;

[0070] Figure 10 This is a schematic diagram of the convergence curve of the weighted minimum mean square error method in this invention;

[0071] Figure 11 This is a schematic diagram comparing the beam patterns of NOMA and SDMA in a partial unloading scenario of this invention. Detailed Implementation

[0072] The present invention will be further described below with reference to the accompanying drawings and specific embodiments, but this is not intended to limit the scope of the invention.

[0073] like Figure 1 As shown, this embodiment relates to a dynamic adaptive space-air-ground integrated architecture for sensing and computing. Unlike traditional space-air-ground converged architectures that only focus on communication links or hierarchical scheduling of tasks, the DA-ISCC architecture introduces a dynamic adaptive mechanism. It can intelligently adjust communication link configurations, sensing node collaboration strategies, and computing resource allocation methods based on real-time changes in the network environment, task load, and node resource status. This enables efficient collaboration of sensing and computing tasks among space (satellite), air (drone), and ground (edge / terminal) nodes. This architecture is specifically designed for the highly dynamic and heterogeneous needs of low-Earth orbit satellite communication and low-altitude economic scenarios. It integrates three core functions—communication, sensing, and computing—to construct an integrated system with physical layering and functional decoupling. The system as a whole consists of three physical layers and two types of heterogeneous networks, supporting cross-domain collaboration and dynamic task scheduling, improving information processing efficiency and service quality in complex and dynamic scenarios.

[0074] Based on the differences in physical entity deployment location and computing power, the framework constructs the following three-level hierarchical structure: the starry sky layer, the ground layer, and the cloud server layer.

[0075] (1) LEO-UAV Tier: This layer mainly consists of low-Earth orbit satellites and UAVs, and is responsible for handling communication and data offloading tasks in wide-area coverage and high-dynamic scenarios. Low-Earth orbit satellites are connected to ground stations through satellite-to-ground links to achieve fast transmission of low-latency data; UAVs offload data to low-Earth orbit satellites through UAV-satellite links, while receiving mission instructions and sensing information from the satellites.

[0076] (2) UAV-Terrestrial Tier: This layer consists of UAVs and ground-based IoT nodes, primarily responsible for the collection and preliminary processing of ground data. In this layer, UAVs not only act as communication nodes but also undertake sensing tasks, using beamforming technology and the Signal-to-Interference-plus-Noise Ratio (SINR) to achieve real-time monitoring of communication link quality. Ground-based IoT nodes connect via virtual links, sending the collected data to the UAVs, which then aggregate the data and upload it to a cloud server or low-Earth orbit satellite.

[0077] (3) Cloud Tier: This layer consists of cloud servers, which serve as the data processing and storage center for the entire framework. The cloud servers receive data uploaded from low-Earth orbit satellites and UAVs, perform in-depth analysis and processing, and feed the processing results back to ground users or satellite users. The cloud server layer is also responsible for coordinating task allocation and resource scheduling between layers to ensure the efficient operation of the entire system.

[0078] The framework comprises two main networks: a low-Earth orbit (LEO) satellite-drone network and a drone-terrestrial network. The LEO satellite-drone network is primarily responsible for communication and data offloading between LEO satellites and drones, while the drone-terrestrial network focuses on data interaction between drones and ground-based IoT nodes and base stations. This dual-network design fully leverages the wide coverage of LEO satellites and the flexibility of drones, while combining the high bandwidth and low latency of terrestrial networks to achieve the unique advantages of an integrated air-space-ground network.

[0079] Within this framework, the link design encompasses various types, including sensing signals, UAV communication links, satellite-to-ground communication links, feeder links, and fiber optic links. The user model includes two categories: ground users and satellite users. Ground users communicate with 5G base stations via terrestrial networks, with the base stations acting as relay nodes, forwarding signals to low-Earth orbit (LEO) satellites or terrestrial IoT edge nodes. Ground users can offload data to base stations, which then upload the data to LEO satellites or cloud servers. Satellite users, on the other hand, directly access the system via satellite networks, uploading data to LEO satellites, which then transmit it to ground stations or cloud servers via satellite-to-ground links.

[0080] This embodiment considers a multi-functional base station, equipped with Root antenna, service There are single-antenna users, and the index set is... This includes ground users and satellite users Furthermore, there exists a point located at an angle The goal, and Located at an angle The source of noise.

[0081] The base station receives three parts of the signal: the compute offload signal from the user, the target echo signal, and the interference signal from clutter sources. Satellite users directly offload data to LEO satellites, which then transmit the data via a satellite-to-ground channel. Forwarding signal.

[0082] For ground users Its transmitted signal is Power is The signal flow is The base station receives the following signal:

[0083]

[0084] Among users The channel modeling between the base station and the ground station is as follows:

[0085]

[0086] Includes path loss With small-scale fading .

[0087] For satellite users, signal transmission is divided into two stages: from the user to the LEO satellite, and then from the satellite to the base station. Due to the high-speed relative motion of the LEO satellite, its channel... Due to the effects of Doppler frequency shift, time-varying fading, and atmospheric attenuation, the equivalent channel of the composite baseband is:

[0088]

[0089] in This represents large-scale path loss. For users Doppler shift, For small-scale fading, It is a time-varying atmospheric attenuation factor.

[0090] The LEO satellite received the following signal:

[0091]

[0092] in For satellite user index set, For users Channel between the satellite and LEO satellite.

[0093] Satellites via satellite-to-ground links The signal is forwarded to the base station, therefore the base station receives the following signal:

[0094]

[0095] In addition, the base station also receives echo signals from the target being sensed and clutter sources:

[0096]

[0097] in Includes round-trip path loss With complex reflection coefficient , For angle The array response vector at that location.

[0098] When the base station sends a signal to the cloud server, the signal sent is:

[0099]

[0100] in For beamforming vectors, This is the information stream. The CS received signal is:

[0101]

[0102] in For base station-cloud channel, It is additive Gaussian noise. All transmitted signals and Assume that the variables are zero-mean, independent, and unity-power complex Gaussian variables.

[0103] In the proposed architecture, the performance of computational task offloading is determined not only by the computing power of the cloud server but also by the computing power of the edge computing server mounted on the low-Earth orbit satellite. Due to the high computing power and small size of the cloud server, the latency of downloading computation results from the server is typically ignored. In this framework, let... and These represent the sets of users performing computing tasks on the base station and the cloud server, respectively. Let this represent the set of users performing computational tasks on low-Earth orbit satellites. The computation cycles required to process one bit of data from each user depend on the nature of the task.

[0104] The computing rate of user k at the base station, i.e., the number of bits processed per second, can be expressed as:

[0105]

[0106] This represents the CPU cycle frequency (in cycles / s) allocated to user k on the BS side. Similarly, the computing rate of user k on the low-Earth orbit satellite can be expressed as:

[0107]

[0108] in, This represents the CPU cycle frequency allocated to user k on a low-Earth orbit satellite.

[0109] The computing speed of user k on the cloud server is given by the following formula:

[0110]

[0111] in, This represents the CPU cycle frequency allocated to user k at the cloud server. Considering that base stations are typically power-constrained, the computational power consumption at the base station should be taken into account, given by the following formula:

[0112]

[0113] in, This is a power factor related to the CPU architecture of the base station computing server. Furthermore, considering the causal relationship between computing speed and task offloading, computing speed... , and The following causal constraints must be met:

[0114]

[0115]

[0116]

[0117] Where B represents the channel bandwidth, The offloading rate of user k from the base station to the cloud server.

[0118] Therefore, causal constraints can be simplified to the following communication computation of causal constraints:

[0119]

[0120]

[0121] Next, the overall computational speed is given by the following formula:

[0122]

[0123] in, This represents the priority weight for each user. In binary offload mode, each user's computational task can only be executed entirely on a base station, low-Earth orbit satellite, or cloud server. At this time, satisfying... , and For partial offloading mode, tasks can be arbitrarily divided into several groups, executed on ground edge nodes, low-Earth orbit satellites, or cloud servers respectively. Define user sets. , as well as ,at this time However, in practical applications, computational tasks may be highly concentrated and cannot be divided, in which case a binary offloading mode is required.

[0124] Figure 2 illustrates the modular architecture of the DAP-ADMM method of this invention, which integrates three mature mechanisms for distributed task offloading in a dynamic integrated air-space-ground network.

[0125] (1) Dynamic step size adjustment. Standard ADMM uses fixed penalty parameters. This approach cannot handle link fluctuations and satellite transit intervals in SAGIN. A penalty adjustment mechanism based on residual error ratio and deadline awareness is introduced, expressed as follows:

[0126]

[0127] in and These are the original residual and the dual residual, respectively. For the remaining execution windows, The deadline is set. This strategy accelerates convergence under high load or unstable link conditions and suppresses oscillations under stable conditions, reducing complexity from... Reduce to .

[0128] (2) Local resource prediction based on LSTM. Due to resource fluctuations (such as LEO satellite visibility and UAV battery cycles), a time-series-based local feasibility prediction mechanism is introduced, using an LSTM network, as shown in the following expression:

[0129]

[0130] in Represents a node In time CPU or GPU load. Predicted set of available resources. Used to constrain the local sub-problem domain Filter out unreachable states and reduce redundant iterations.

[0131] (3) Asynchronous communication mechanism. The consensus mechanism ADMM requires all nodes to update synchronously in each iteration, which is difficult to achieve under long-distance, intermittent satellite links and asymmetric bandwidth conditions. The following asynchronous update mechanism is adopted:

[0132]

[0133] in For local latency, The threshold is used. To ensure consistency, timestamps are used to track variable versions and resolve conflicts. This design significantly reduces idle communication time and improves throughput and fault tolerance.

[0134] The pseudocode for DAP-ADMM is shown in Algorithm 1. In binary unloading mode, DAP-ADMM jointly optimizes unloading decisions. Calculation speed With beamforming vector To maximize effective computing speed .

[0135] Compared to traditional ADMM methods (which rely on centralized coordination or incur high global synchronization overhead), DAP-ADMM utilizes residual-aware dynamic step size adjustment to accelerate convergence under time-varying channel conditions, and combines LSTM local resource prediction to reduce iteration overhead caused by resource fluctuations and infeasible search states.

[0136] Effect verification

[0137] Figure 3 The convergence performance of DAP-ADMM in binary offloading mode is shown. The objective function and effective computational rate exhibit a non-linear growth trend. The objective value converges around the 260th iteration, while the effective rate increases rapidly in the first 50 iterations, then gradually slows down. Ultimately, the effective computational rate stabilizes between 400–450 Mbit / s, slightly lower than the target rate. This gap indicates that under strict offloading constraints, resource allocation flexibility is limited, leading to performance bottlenecks in later stages. This highlights the efficiency limitations of binary offloading, especially when the effective rate cannot reach the target value. Therefore, when using binary offloading in practical applications, a trade-off must be struck between system performance and computational efficiency.

[0138] like Figure 4 As shown, under the condition of a constant computing speed, the NOMA-assisted scheme in The main lobe is highly concentrated in the vicinity, while the side lobes are relatively weak, exhibiting excellent anti-interference performance. The overall beam pattern trend of the Space Division Multiple Access (SDMA) scheme is similar to that of NOMA; both can form a focused main lobe in the target direction. However, SDMA has a higher peak side lobe, weakening its interference suppression capability and thus affecting sensing performance. In contrast, NOMA exhibits stronger signal concentration and interference suppression capability in the main lobe direction.

[0139] Figure 5 The convergence performance of DAP-ADMM, standard ADMM, and consensus ADMM under dynamic network topologies was compared. In the simulation experiment, satellite switching and UAV trajectory movement were introduced between iterations 80 and 120 to simulate real-world topology changes. Standard ADMM performed well in the initial stage, but exhibited large residual fluctuations after topology changes, indicating its inability to adapt to changes due to fixed penalty parameters. Although consensus ADMM is a distributed method, it still exhibited instability within the same time period due to its fixed neighbor structure. In contrast, DAP-ADMM responded quickly through dynamic step size adjustment and asynchronous update mechanisms, maintaining a faster convergence speed. The final residual stabilized at around 0.05, which is an order of magnitude lower than standard ADMM and twice as low as consensus ADMM. Furthermore, DAP-ADMM improved convergence speed by more than 60%, demonstrating good robustness and adaptability.

[0140] Figure 6 The impact of three different step-size strategies on the residual convergence performance of DAP-ADMM was evaluated. This invention compares the following three strategies: fixed step size, a warm-up strategy, and the proposed dynamic strategy. The fixed step-size method converges slowly, with the residual only decreasing to approximately 0.4 within the first 50 iterations; the warm-up strategy (with a small initial penalty parameter that gradually increases) accelerates convergence in the mid-stage, eventually reducing the residual to approximately [missing value]. The dynamic strategy achieves faster and more stable convergence throughout the process, reducing the residual to less than 100 iterations. the following.

[0141] Figure 7 The penalty parameters under dynamic policies are further demonstrated. The evolutionary process. Observed. The adaptive adjustment mechanism adjusts the step size during the iteration process: a higher value is set initially to promote global convergence, and the value gradually stabilizes in subsequent stages to suppress oscillations. This adaptive mechanism also introduces the residual ratio and remaining task window as control terms, enabling the step size to self-adjust based on task urgency and link dynamics, thereby effectively mitigating common convergence instability or premature stalling problems under static parameter settings. Table 1 illustrates the channel modeling and system parameter settings for the space-air-ground integrated network.

[0142] Table 1 Channel Modeling and System Parameters for Space-Air-Ground Integrated Network

[0143]

[0144] Figure 8 The communication efficiency of different methods in a dynamic integrated air-space-ground network was further compared. This invention simulated a scenario where the number of nodes gradually increased from 10 to 60, introducing random UAV arrivals and departures and periodic low-Earth orbit satellite switching to simulate real topology changes. Fifty distributed computing offloading tasks were run under each configuration, and the average communication energy consumption per node in each iteration was recorded as an evaluation metric. As the number of nodes in the integrated air-space-ground network increased, the average communication energy consumption of all methods increased, due to the increased communication overhead caused by increased network density and link changes. Among all compared methods, DAP-ADMM, which employs an asynchronous communication mechanism, achieved the lowest energy consumption across all node sizes.

[0145] Figure 9 The residual convergence behavior of four schemes was compared: DAP-ADMM with / without LSTM local resource prediction, and consensus ADMM with / without prediction. Simulations considered the time-varying nature of node computational capabilities and zoomed in to observe convergence details within the 175–200 iteration range. Among all schemes, DAP-ADMM with LSTM prediction converged the fastest, with the final residual being lower than [value missing] within 200 iterations. In contrast, its unpredictable version converges more slowly, eventually stabilizing at approximately [value missing]. Similarly, LSTM prediction also improves the convergence behavior of consensus ADMM, but it is still superior to the DAP-ADMM series of methods. It is worth noting that consensus ADMM with LSTM prediction performs poorly in the early stages ( The convergence is faster, possibly due to the lack of step-size damping. However, as optimization progresses, DAP-ADMM rapidly surpasses consensus ADMM in residual magnitude and convergence stability. Within the 175–200 iteration range, the average residual of DAP-ADMM is more than 60% lower than that of consensus ADMM.

[0146] In the experimental setup, the LSTM model predicts the feasible resource set for the next time step by analyzing the CPU utilization over the past 5 time steps. The prediction results are used to dynamically constrain the solution space of local subproblems, effectively filtering out infeasible variable states. This pre-selection mechanism significantly reduces the search space and minimizes ineffective updates and communication overhead.

[0147] Figure 10 shows the convergence performance of the weighted minimum mean square error method under partial offloading mode. This method converges rapidly, with the computational speed increasing quickly with the number of iterations, eventually approaching the upper limit of 500 Mbit / s. The method converges to a stable high-speed solution in less than 100 iterations, demonstrating good convergence efficiency. This verifies the effectiveness of this method in approximating the optimal KKT solution, and it is particularly suitable for applications sensitive to latency and with high real-time requirements.

[0148] Figure 11 Beamforms of NOMA and SDMA schemes under partial offloading modes were compared. Similar to the binary offloading case, the NOMA-assisted scheme exhibits a more concentrated main lobe and stronger interference suppression capability. Although SDMA can also achieve main lobe focusing, its higher peak sidelobes result in weaker interference suppression capability in the clutter direction. This further demonstrates the superior adaptability of NOMA in densely sensed targets and complex interference environments.

[0149] While the invention has been described herein with reference to specific embodiments, it should be understood that these embodiments are merely examples of the principles and applications of the invention. Therefore, it should be understood that many modifications can be made to the exemplary embodiments, and other arrangements can be designed without departing from the spirit and scope of the invention as defined by the appended claims. It should be understood that different dependent claims and features described herein can be combined in ways different from those described in the original claims. It is also understood that features described in conjunction with individual embodiments can be used in other described embodiments.

Claims

1. A dynamic adaptive sensing and computing integration method for integrated air-space-ground networks, characterized in that, The method includes: S1. Construct an integrated air-space-ground network model that integrates low Earth orbit satellites, UAVs, and ground users / edge nodes, and uniformly represent the time-varying channels, node computing capabilities, energy budgets, and mission attributes of UAV-ground, low Earth orbit satellite-UAV, and backhaul links to support a hybrid access architecture of intra-cluster NOMA and inter-cluster SDMA. S2. Establish a cross-layer communication-sensing-computing integrated joint optimization model. Based on whether the task is divisible, two modes are formed: binary offloading and partial offloading. With the goal of maximizing the weighted system utility, the task offloading and communication / sensing / computing resource allocation are jointly optimized, and power constraints, NOMA decoding order constraints, computing resource constraints, minimum signal-to-interference-plus-noise ratio constraints, and latency constraints are applied. S3. For binary offloading mode, the Dynamic Adaptive Prediction Alternating Direction Multiplier Method (DAP-ADMM) is used. Through residual-aware penalty parameters and step size adaptive adjustment, local resource prediction driven by long short-term memory network and asynchronous communication update mechanism, the offloading decision and resource allocation are solved collaboratively, and the output is a scheme for beamforming, power control, bandwidth allocation and computational resource allocation. S4. For partial offloading mode, a weighted minimum mean square error alternating optimization method is adopted, introducing auxiliary variables such as receiver equalizer and error weight, and performing block coordinate descent iteration to output the offloading ratio, power allocation and beamforming results.

2. The dynamic adaptive sensing and computing integration method for integrated air-space-ground networks according to claim 1, characterized in that, The optimization model of S2 includes the following constraints and definitions: S21. Communication and Backhaul Constraints: Effective Uplink Rate of User k Determined by intra-cluster NOMA and serial interference cancellation decoding, satisfying Backhaul link meets S22. Computing resource and power consumption constraints: CPU frequency allocated to drones / edge nodes satisfy And calculate power consumption S23, Perceptual Constraints: Perceptual SINR Must meet . in ; S24, NOMA decoding order and physical layer constraints: satisfy the intra-cluster decoding order π and beam / power non-negativity, total power upper limit and inter-cluster beam interference suppression; S25, Binary Unloading Variable ∈{0,1} and ; or the partial unloading rate variable r(·,k)≥0.

3. The dynamic adaptive sensing and computing integration method for integrated air-space-ground networks according to claim 1 or 2, characterized in that, The binary unloading optimization problem P1 is defined as follows: And satisfy the aforementioned communication, backhaul, computation, sensing, and NOMA constraints, as well as the binary constraints; the partial offloading optimization problem P2 is defined as: All constraints in claim 2 are satisfied and r(·,k)≥0.

4. The dynamic adaptive sensing and computing integration method for integrated air-space-ground networks according to claim 1, characterized in that, The DAP-ADMM of S3 includes: S31. Variable relaxation and distributed decomposition: Relax the binary delta variable to [0,1], introduce consistency variables and Lagrange multipliers, construct an augmented Lagrange and divide the beam, power, offloading and computation subproblems according to nodes / clusters for parallel solution; S32. Residual-aware penalty parameters and adaptive step size: ρ and step size η are dynamically updated based on the relationship between the original residual s(t) and the dual residual r(t). Increase ρ and decrease η when ρ is high, and vice versa, decrease ρ and moderately increase η. S33, LSTM Local Resource Prediction: Using the local / neighbor decisions, channel and load timing of the past h time slots as input, predict the available computing power and key neighbor variables for the next time slot, which are used to replace expired information in local updates; S34. Asynchronous Communication and Event-Triggered Synchronization: Each node performs N local updates before exchanging information once. Link congestion or topology disturbances trigger immediate synchronization. When the local processing delay exceeds a threshold... To maintain the stability of variables; S35. Dual Update and Convergence: Update the multipliers according to the consistency error until the original and dual residuals are below the threshold ε or the maximum iteration is reached. .

5. The dynamic adaptive sensing and computing integration method for integrated air-space-ground networks according to claim 4, characterized in that, The adaptive evolution of the penalty parameter is as follows: ， in, To initialize the penalty parameters, For the original residual, For dual residuals, The remaining time window for the task. This is the maximum allowed completion time. This strategy can be increased when node load is high and channel quality is poor. To accelerate convergence; to reduce the network's stability. To suppress oscillations.

6. The dynamic adaptive sensing and computing integration method for integrated air-space-ground networks according to claim 4, characterized in that, The Long Short-Term Memory predictor outputs the available computing power range and feasible rate upper bound for the next time slot, which is used to prune the search space of local subproblems and reduce cross-node communication; the predictor is pre-trained in the cloud and fine-tuned online by sliding the window to avoid overfitting and concept drift.

7. The dynamic adaptive sensing and computing integration method for integrated air-space-ground networks according to claim 1, characterized in that, The weighted minimum mean square error alternating optimization of S4 includes: S41, Weighted minimum mean square error equivalent: Converting rate maximization into minimizing the weighted mean square error by introducing a receiver equalizer u and error weight v; S42, Block Coordinate Descent Update: In the t-th iteration, u(t) and v(t) are updated in a closed loop. Given (u, v), the convex quadratic problem of beamforming p(t+1) is solved, and given (p, u, v), the unloading ratio or allocation rate r(·,k) is updated. S43. Introducing SINR Constraints The weighted minimum mean square error subproblem enables beamforming to improve communication efficiency while ensuring sensing accuracy. S44. Using the objective function gain threshold Δ<ε as the convergence criterion, output the solution that satisfies the approximate KKT conditions.

8. The method according to claim 1 or 7, characterized in that, The hybrid access adopts intra-cluster NOMA and inter-cluster SDMA: by constraining intra-cluster serial interference to eliminate decoding order π and inter-cluster beam approximately orthogonal design, cross-cluster interference is suppressed and spectrum efficiency is improved; corresponding beam maps are generated in binary and partial offload modes to adapt to user distribution.

9. The dynamic adaptive sensing and computing integration method for integrated air-space-ground networks according to claim 1, characterized in that, The system utility is defined as Where r(·,k) = f(·,k) / φ, f(·,k) represents the CPU frequency allocated to user k by the drone / edge, low-Earth orbit satellite, and cloud, respectively, and the weights are... The system is dynamically adjusted based on business priorities to ensure the computational speed and perceived reliability of high-priority tasks.

10. A dynamic adaptive sensing and computing integrated system for integrated air-space-ground networks, characterized in that, It includes a computer module that applies the method described in any one of claims 1-9.

Images