Research on cascading failure mechanisms has long been a focal point in complex network reliability. Percolation models rely on network topology to analyze connectivity changes but neglect dynamic factors such as node load. Avalanche models simulate rapid local-to-global failure spread through chain-reaction rules but inadequately account for network heterogeneity. Load-capacity coupling models assume nodes have finite capacity; when load exceeds capacity, nodes fail and redistribute load to neighbors. The seminal work by Motter and Lai [12] established the foundation for this class of models, with subsequent extensions incorporating tunable parameters for load redistribution.

However, these models assume static, homogeneous nodes, failing to capture SAGIN’s cross-layer heterogeneity. Recent studies address this but remain single-layer focused: Liu et al. [13] analyzed satellite failures; Zhang and Du [14] developed a low Earth orbit (LEO) satellite dynamic model; Wang et al. [15] investigated Unmanned Aerial Vehicle (UAV) failures due to energy limits. These single-layer approaches cannot capture how failures propagate across space, air, and ground domains—a gap our multi-factor coupled model addresses by integrating computational, transmission, and energy loads across all layers with link-type dependent propagation probabilities.

Frequent reconfigurations from satellite motion and user mobility challenge cascading failure analysis. Valdez et al. [16] reviewed cascading failures in dynamic networks; Xiao et al. [17] emphasized need for time-varying SAGIN models.

Zhang et al. [18] proposed a LEO satellite dynamic model incorporating orbital mechanics. Xu et al. [19] examined cascading effects in UAV swarms. However, these address dynamics within single layers only. While advance single-layer dynamics, they do not extend to multi-layer SAGIN where space, air, and ground dynamics interact and amplify cascading effects—motivating our time-varying topology modeling across all layers.

In recent years, intelligent methods have been increasingly applied to network fault management. Cui et al. [20] employed multi-agent reinforcement learning for resource allocation in UAV networks, achieving dynamic optimization of communication and computation resources. Chen et al. [21] proposed a node collaborative defense mechanism based on game theory, formulating fault propagation as a non-cooperative game between attackers and defenders. However, these target isolated measures without integrated early warning, isolation, and self-healing. The fragmentation in existing defenses limits system-wide resilience—directly motivating our closed-loop “early warning-isolation-self-healing” framework that coordinates prevention, suppression, and recovery phases.

Satellites operate under strict resource limits, often overlooked. Dakic et al. [22] analyzed vehicle-to-everything (V2X) reliability. Fang et al. [23] proposed DRL-driven resource allocation. Al-Zahrani et al. [24] reviewed 5th-Generation (5G) optimization. However, these prioritize performance over resilience, and AI-based methods remain computationally intensive. The impracticality of complex approaches on resource-constrained platforms motivates our lightweight linear programming formulation solved via greedy approximation, ensuring deployability on satellites and aerial platforms.

As summarized in Table 1, existing research lacks a unified approach to SAGIN’s heterogeneity, dynamism, and resource constraints. These gaps directly motivate our four contributions: a multi-factor cascading model addressing heterogeneity, a time-varying topology framework addressing dynamics, a closed-loop defense addressing coordination, and a lightweight optimization addressing resource constraints—all building upon and extending prior work to the SAGIN context.

The space layer comprises LEO, medium Earth orbit (MEO), and high Earth orbit (HEO) satellites delivering wide-area coverage and long-distance communications. LEO satellites offer low latency and high bandwidth but require constellation networking for global coverage; MEO satellites balance coverage and latency for regional needs; HEO satellites provide extensive regional coverage with higher latency and limited bandwidth. Satellite nodes face challenges including frequent link handovers and power constraints due to orbital mechanics.

The aerial layer consists of drones and high-altitude airships acting as relay hubs between space and ground layers, enhancing regional communication and supporting edge computing. Drones offer flexible deployment and rapid repositioning, while airships provide extended endurance and wide coverage. Both face reliability fluctuations due to weather and energy constraints. Inter-layer communication is enabled by gateway nodes such as satellite ground stations and multi-mode UAVs, explicitly modeled in our simulation. The ground layer comprises 5G base stations, core network equipment, and end-user devices. Base stations provide ample computing power, stable links, and low latency; core network equipment manages resource allocation and data routing as the operational hub [25], end-user devices—IoT devices, smartphones, and vehicles—act as service initiators and recipients. Due to limited ground coverage, wide-area coordination depends on the space-air layer.

Topological Dynamism: Satellite motion, aerial platform repositioning, and mobile ground users cause frequent link handovers and topology reconfigurations, creating a highly dynamic SAGIN topology characterized by time-varying connectivity [26].

Topological Heterogeneity: Space, air, and ground nodes differ significantly in physical characteristics, resource capabilities [27], and communication protocols. Satellite links suffer from free-space loss and rain fade, causing fluctuations in delay, bandwidth, and loss rate; air links are affected by weather and platform vibrations; ground links operate in stable environments with ample bandwidth and low latency. These disparities across link types increase the complexity of network coordination and fault propagation analysis. The topology is a hybrid of three-layer hierarchical architecture with intra-layer mesh connectivity.

Building upon foundational load-capacity coupling models, we extend the classical framework to accommodate SAGIN heterogeneity. The classical model defines node failure when load exceeds capacity, triggering load redistribution to neighbors. However, this model assumes static topology and homogeneous nodes, limiting its applicability to SAGIN. Our improved model incorporates multi-dimensional load definitions, link-type dependent propagation probabilities, and dynamic load evolution to capture SAGIN-specific characteristics. Node load is defined as the weighted sum of computational task load, data transmission load, and energy consumption load, calculated as (1)Li=w1Li,comp+w2Li,trans+w3Li,energybased on SAGIN node resource characteristics. The coefficients w1, w2, and w3 are determined by node type and application scenarios using the Analytic Hierarchy Process (AHP), a multi-criteria decision-making method that systematically combines expert knowledge and network operational objectives. For instance, satellite nodes, with limited power budgets, are assigned higher energy consumption load coefficients (w3>w1,w2), while ground base stations, with stable power supply, prioritize computational and data transmission loads (w1,w2>w3). This approach ensures that weight selection is grounded in domain expertise rather than arbitrary assignment. Node capacity, representing the maximum load a node can handle, is defined as the minimum of computational capacity, transmission capacity, and energy capacity, (2)Ci=min(Ci,comp,Ci,trans,Ci,energy)

Ci,comp denotes the computational capacity of node i, measured by its maximum supported CPU utilization; Ci,trans represents the transmission capacity of node i, quantified by its maximum supported bandwidth occupancy; Ci,energy indicates the energy capacity of node i, defined by the load threshold corresponding to the minimum remaining energy ratio.

The initial failure node is selected based on a node failure probability model. The failure probability Pfail,i of a node depends on its operational time t, resource consumption rate ri, and external environmental interference intensity si, defined as (3)Pfail,i=1−e−αt−βri−γsiwhere α, β, and γ are scaling factors. Through random sampling, the initial failure node set is determined according to each node’s failure probability. When node i fails, its load Li migrates to adjacent healthy nodes following these rules: First, adjacent nodes are ranked by link priority coefficient pi,j (which correlates with link bandwidth, transmission delay, and reliability). Then, the load is evenly distributed to adjacent healthy nodes in descending order of priority, with the allocation ratio to node j calculated as (4)pi,j/∑j∈Nipi,j

If the remaining capacity of adjacent healthy nodes is insufficient to handle the migrated load, the remaining load continues to migrate to the next-level adjacent nodes.

To account for inter-layer link attenuation, the fault propagation probability Pi,j is defined as the probability of a node i’s fault propagating to node j, given by(5)Pi,j=Lj,receivedCj⋅ηi,jwhere, Lj,received represents the received migration load at node j, while ηi,j is the link attenuation coefficient, which varies with the link type. The attenuation coefficients decrease progressively from inter-satellite links to space-ground links, with terrestrial links exhibiting the lowest attenuation.

To describe the dynamic changes of load over time (reflecting business growth or sudden traffic surges), the load dynamic evolution formula is introduced:(6)Li(t+1)=Li(t)+ΔLi,external(t)+∑j∈NiΔLj→i(t)where DeltaLi,external denotes the externally added load, and sumj∈NiΔLj→i represents the migrated load from neighboring node j, load transfer during fault propagation.

The recovery process of faulty nodes can be described by a time and resource-based recovery probability model:(7)Precover,i(t)=1−exp⁡(−λit)where λi is the recovery rate coefficient positively correlated with node resources (e.g., computing power, energy), and t is the fault duration.

Early warning threshold formula provides a quantitative early warning trigger condition for fault prediction based on multi-source information fusion.(8)Alert if: L^i(t+Δt)>θ⋅CiorQ^i,j(t+Δt)<Qthwhere L^i is the predicted load, θ is the safety factor, Q^i,j is the predicted link quality, and Qth is the quality threshold. A summary of the key variables and their notations is provided in Table 2.

The prevention stage optimizes resource allocation according to the importance of the node, evaluated by means of a composite metric that combines centrality of the between, centrality of the degree and functional dependency. Blockchain-enabled collaborative frameworks have shown promise in securing resource coordination across distributed logistics networks [28], inspiring our trusted resource allocation design. Betweenness centrality reflects routing significance, degree centrality indicates connectivity density, and functional dependency quantifies inter-node functional reliance. The node importance index Ii=aBi+bDi+cFi uses weights a, b, and c determined by the AHP, a multi-criteria decision method that quantifies the relative importance of different centrality metrics without introducing prohibitive computational overhead. Resources are allocated proportionally to these indices. High-importance hub nodes, including high-orbit satellites and core ground base stations, receive priority allocation of bandwidth, computing power, and energy to enhance capacity and fault tolerance, alongside dedicated redundant backups such as standby satellites and redundant links. Regular nodes have their resource ratios dynamically adjusted based on service demand and current load to maintain network load balancing.

The suppression stage implements proactive load diversion triggered by fault precursor detection. A multi-source fusion prediction model collects real-time node status, link performance, and topology evolution data. Long Short-Term Memory time-series analysis identifies precursor features, predicting potential faults and locating at-risk nodes and links ahead of occurrence. Upon detecting precursors, dynamic load diversion is activated. Partial loads from at-risk nodes are preemptively migrated to adjacent healthy nodes, with paths selected based on node capacity, link quality, and service priority. Idle unmanned aerial vehicles are deployed to fault-prone areas to establish temporary links, facilitating load offloading and coverage maintenance. A link priority mechanism prioritizes core services for bandwidth allocation; non-core services may be temporarily suspended to contain fault propagation and protect critical network functionality.

The system achieves rapid fault identification via link quality monitoring and node heartbeat feedback. Link reliability is measured by transmission latency, bandwidth, and packet loss rate. Node failures are detected through periodic heartbeat exchanges with multi-node collaborative verification. Algorithm 1 formalizes this process.

Upon fault detection, the system reconstructs topology through a greedy strategy, prioritizing nearby healthy nodes with sufficient residual capacity. For satellite layers, pre-planned reconnection procedures based on orbital predictions are applied. Core link failures trigger redundant link activation with dynamic bandwidth adjustment. Algorithm 2 details the reconstruction process.

This reconstruction problem can be viewed as a linear program: maximize successfully rerouted flows subject to node capacity constraints. The greedy algorithm provides an approximate solution with significantly lower overhead than exact solvers.

Algorithm 1 has time complexity O(N+E); Algorithm 2 achieves O(Nlog⁡N). Both have space complexity O(N+E), meeting real-time requirements on resource-constrained platforms.

This experiment employs NS-3 as the primary simulation tool for network-level modeling, supporting heterogeneous node configuration, mobility patterns, and link-level performance simulation. MATLAB is integrated for offline data processing, model solving, and visualization. The co-simulation operates iteratively: NS-3 generates trace files containing link delays, packet loss rates, and throughput at each time step; MATLAB reads these traces, computes node loads and failure probabilities using Eqs. (1)–(8), and determines defense actions; updated configuration parameters are then fed back to NS-3 for the next simulation cycle. Table 3 shows the network scale settings. This modular approach requires no custom API-level interfacing, using standard file I/O for data exchange, ensuring ease of replicability. The combination of NS-3 and MATLAB ensures both authenticity and flexibility while meeting the experimental verification requirements.

The node attribute parameter settings are shown in Table 4.

The link parameter settings are shown in Table 5.

All experiments are based on three defined failure scenarios: LEO satellite node failure (simulated by disabling network interface), ground base station overload (CPU utilization exceeds threshold), and UAV swarm disconnection (introducing high path loss for air-ground links). These are implemented in NS-3 through custom modules that modify node status or environmental conditions at specified times. We conducted validation experiments on cascading failure propagation characteristics to analyze the relationship between topology density, load threshold, and both the propagation speed and impact scope of cascading failures. Different topology densities and load thresholds were configured, and multiple simulation experiments were performed across three typical failure scenarios. The time required for failure propagation to reach a stable state and the proportion of failed nodes were recorded.

For the density analysis in Figs. 2 and 3, we scaled this baseline by factors of 0.5, 1.0, and 1.5 to generate low (density = 5), medium (density = 10), and high (density = 15) densities. This variable-controlling approach allows us to isolate the effect of network scale on fault propagation while preserving SAGIN’s hierarchical structure.

As illustrated in Fig. 2, under identical load thresholds, higher topology density accelerates failure propagation and expands its scope. For ground base station failure at load threshold 1.6, increasing density from 5 to 10 to 15 reduces stabilization time from 18 to 12 to 8 time units, while the failed node proportion rises from 35% to 55% to 70%. Tighter connectivity creates more propagation paths and accelerates load migration, facilitating rapid failure diffusion.

Conversely, as shown in Fig. 3, under fixed topology density, higher load thresholds slow failure propagation and reduce its scope. For LEO satellite failure at density 10, raising the threshold from 1.2 to 1.6 to 2.0 increases stabilization time from 10 to 15 to 22 time units, while the failed node proportion drops from 65% to 45% to 25%. Higher thresholds enhance node bearing capacity, enabling nodes to absorb more migrated load and reduce secondary failures.

We validated our model against actual failure scenarios using LEO satellite failure. Our model achieved prediction errors of 5%–8% for stabilization time and 4%–7% for failed node proportion, compared to 20%–25% and 15%–20% for CASCADE. For an actual stabilization time of 15 time units, our model predicted 14–16 units vs. CASCADE’s 11–19 units; for an actual failed node proportion of 45%, our model predicted 42%–48% vs. CASCADE’s 36%–54%.

We further conducted comparative analyses between this model and CASCADE as well as Zhang and Du LEO-specific model [14], covering three scenarios (LEO satellites, ground base stations, and drone swarms). Fig. 4 presents a quantitative comparison of prediction errors among the three models. Building on the load-capacity framework, our model incorporates multi-dimensional loads and cross-layer propagation, capturing interdependencies across space, air, and ground. For LEO satellite failure, our model achieves errors of 5%–8% (stabilization time) and 4%–7% (failed node proportion), outperforming CASCADE (20%–25%, 15%–20%) and Zhang and Du [14] (12%–15%, 10%–12%). Similar improvements across all scenarios demonstrate generalizability.

We evaluated the performance of the Early Warning-Isolation-Self-Healing closed-loop collaborative defense strategy from three dimensions: failure recovery rate, communication interruption duration and resource overhead. Two experimental groups were set up for comparison, one adopting the proposed defense strategy and the other using traditional defense strategies of static redundancy backup and static load balancing. Simulation experiments were carried out across three typical failure scenarios, with all relevant performance metrics recorded for analysis.

As shown in Figs. 5 and 6, our strategy achieved failure recovery rates of 85%–90%, compared to 60%–65% for traditional strategies. For example, in the UAV cluster disconnection scenario, our strategy recovered 88% of failed nodes and links, whereas the traditional strategy recovered only 62%. Furthermore, as shown in Fig. 7, our strategy achieved average communication interruption durations of 5–8 time units, significantly outperforming the 15–20 time units observed with traditional strategies. In telemedicine applications, for instance, our strategy constrained communication interruptions to within 6 time units, satisfying the stringent latency requirements for remote surgical guidance, whereas traditional strategies’ prolonged interruptions risked compromising healthcare service continuity. Regarding resource overhead, our strategy reduced bandwidth occupation and energy consumption by 20%–25% compared to traditional approaches. Specifically, in satellite node failure scenarios, our strategy decreased satellite energy consumption by 22%, effectively enhancing satellite node endurance. The resource utilization rate are shown in Fig. 8.

These experimental results demonstrate that our designed defense and self-healing algorithms significantly outperform traditional strategies in terms of failure recovery capability, communication continuity assurance, and resource utilization efficiency, substantiating their marked performance advantages.