AI-Based Satellite Anomaly Detection Using Fused Multi-View Remote Sensing Data: Technologies and Prospects

Abstract

In recent years, satellites have been widely employed in diverse domains such as Earth observation, weather forecasting, and space exploration, generating large-scale data through remote sensing and telemetry. However, these data have primarily been used to analyze surface changes or atmospheric phenomena, while their application to anomaly detection of satellite systems themselves has remained limited. Traditionally, anomaly detection has relied on onboard sensors and ground control centers through Failure Detection, Isolation, and Recovery (FDIR) techniques. Recently, however, increasing attention has been given to detecting anomalous states through cross-observation among multiple cooperative satellites and between satellites and ground stations.In particular, for mission-dedicated satellite constellations such as space-based solar power systems, the necessity of artificial intelligence (AI)-based analysis leveraging remote sensing and telemetry data has been emphasized. This review comprehensively examines recent research trends and representative case studies on AI-based satellite anomaly detection using remote sensing data and provides perspectives on future directions.

1. Introduction

Over the past several decades, artificial satellites have propelled scientific and industrial progress through diverse missions in Earth observation, lunar and Martian exploration, and deep-space investigation. Central to these achievements is remote sensing, which has matured into a foundational capability that yields vast volumes of imagery, orbital ephemerides, electromagnetic signals, and meteorological data. Despite this richness, the predominant emphasis of remote sensing has been on observing physical and environmental phenomena of the Earth system. Its potential as a primary source for diagnosing the health and anomalies of the satellite system itself remains comparatively underexploited.

Anomaly detection refers to identifying unexpected states that deviate from nominal patterns. In satellite operations, anomalies span a wide spectrum—from component-level faults (e.g., short circuits, thermal excursions, battery irregularities) to system-level events (e.g., attitude instability, orbit drift, propulsion malfunctions, payload data abnormalities) (Fejjari et al., 2025). Failure to timely detect and mitigate such conditions risks not only single-satellite mission loss but also degraded performance across an entire satellite constellation. Traditionally, anomalies are managed by Failure Detection, Isolation, and Recovery (FDIR) pipelines that rely on on-board sensor telemetry and ground-segment analysis (Jalilian et al., 2017; Castel et al., 2006). While effective for localized, single-asset diagnosis, these approaches often depend heavily on ground intervention and may be ill-suited to time-critical or multi-asset cooperative operations.

The limitations of traditional FDIR are pronounced in modern architectures such as space solar power satellite (SSPS) constellations—comprising tens to hundreds of spacecraft forming a coherent power-generation and transmission structure—and mega-constellations exemplified by Starlink. In these regimes, operators must detect not only individual spacecraft faults but also emergent, constellation-level phenomena arising from inter-satellite coupling. Rapid insight into relative orbital changes, formation-flying position errors, ground-station coverage gaps, latency, and link losses is essential, yet immediate, holistic responses are difficult to achieve with conventional, single-vehicle FDIR (Iiyama et al., 2024; Ren et al., 2021). This motivates an alternative paradigm: cross-asset, cross-segment remote sensing for anomaly detection, wherein satellites observe one another—and the ground segment observes the collective—so that anomalies are inferred from mutual remote measurements across space and ground.

Recent advances in artificial intelligence (AI) provide powerful tools to realize this shift. To navigate stringent on-orbit compute constraints while exploiting distributed data, federated learning enables spacecraft to train models locally and aggregate knowledge globally, enhancing constellation-wide detection without centralizing raw data (Atefrad and Karami, 2025). In parallel, large language models (LLMs) and agentic AI are emerging as decision-support substrates for autonomous interpretation of heterogeneous remote-sensing streams and for synthesizing cooperative response strategies among spacecraft and ground assets (Maranto, 2024; Mousist, 2025; Gao et al., 2025).

Scope and contributions. This review synthesizes recent research on: (i) the structural limits of traditional FDIR in the context of distributed, cooperative space systems; (ii) AI-enabled anomaly detection pipelines that fuse remote-sensing products (imagery, orbital, environmental data) with conventional telemetry; and (iii) emerging operational paradigms for constellation-scale anomaly awareness in SSPS and formation-flying missions. We delineate methodological trends, identify gaps in current practice, and outline promising directions for integrating mutual remote observations, federated learning, and agentic autonomy into next-generation satellite health management.

2. Materials and Methods

Early detection of faults, malfunctions, and environment-induced off-nominal states is a decisive factor for mission success. Historically, satellite anomaly detection has centered on FDIR algorithms and systems, drawing on (i) analysis of on-board sensor telemetry, (ii) ground-segment diagnosis and recovery procedures, and (iii) statistical or physics-based evaluation against known failure modes (Tafazoli, 2009). While these methods have underpinned reliable operations since the dawn of the space age, the emergence of mega-constellations and cooperative multi-satellite missions exposes clear gaps and motivates looking beyond the spacecraft’s interior to external observations.

2.1. Traditional Approaches to On-Board Sensor–Based Diagnosis

Spacecraft carry diverse sensors to continuously monitor electrical power, thermal control, propulsion, and attitude control subsystems. In classical practice, engineers specify nominal operating ranges; telemetry that breaches these thresholds is flagged as anomalous. This approach is simple and fast, but vulnerable to (i) novel failure modes and compound anomalies, (ii) sensor aging and drift that masquerade as faults, and (iii) context-dependent phenomena that defy fixed thresholds.

For instance, a drop in solar array output can be misattributed to power-system degradation when the root cause is an attitude-control error, altering the sun-pointing geometry—as reported in early NASA operations (Lewis Spacecraft Mission Failure Investigation Board, 1998). Moreover, purely internal monitoring struggles to infer externally manifested states—e.g., structural damage, partial deployment, or mechanical hang-ups—and recovery often defaults to redundancy switching or safe-mode holds pending ground intervention, hindering rapid response.

Fig. 1. Attitude Determination and Control System (ADCS) mode and anomaly of Lewis satellite (Lewis Spacecraft Mission Failure Investigation Board, 1998).

2.2. Traditional Approaches to Ground-Segment–Led Final Response

A second pillar is ground-based analysis of downlinked telemetry and tracking data. By comparing predicted states (e.g., orbit from dynamics models) with observations, operators can identify deviations such as orbit drift (Xu et al., 2025). Ground teams also watch attitude, link margin, and RSSI, latency, and subsystem health to trigger recovery playbooks.

However, this paradigm faces latency and visibility constraints: during out-of-view periods the system is effectively blind; even in contact, bandwidth and scheduling can limit diagnostic depth, and longer delays push mitigation later in time (Diro et al., 2024). Decision-making often relies heavily on human expertise; as data volumes and constellation sizes scale, operator overload becomes a structural risk. As of 2025, the number of satellites in orbit (including those reentering) is reported at 13,883 (Orbit.ing-now.com, 2025), exceeding the global fleet of passenger aircraft, and projects like Starlink require supervising thousands of spacecraft—conditions under which manual, ground-centric monitoring cannot support timely, system-wide anomaly response.

Traditional thresholding and physics-based detection methods (e.g., Keplerian/perturbed-orbit predictions vs. measured ephemerides; thermal balance models vs. measured temperatures) remain valuable for transparency and physical plausibility. Yet real space environments introduce interactions and disturbances—inter-satellite effects, solar flares, radiation, micrometeoroids, long-term structural fatigue—that are difficult to capture fully. As mission complexity rises, model mismatch and false assessments become more likely.

Fig. 2. Typical communication between a ground telemetry station and an on-orbit spacecraft (Liu et al., 2023).

2.3. Integrated Limitations of Legacy Methods

Across these practices runs a common thread: they are optimized for single-spacecraft, localized diagnosis. Today’s landscape—mega-constellations and tightly coordinated formations—reveals several fundamental limitations:

• Scale of data: telemetry and remote-sensing outputs from thousands of spacecraft exceed the throughput and attention budgets of traditional pipelines.

• Emergent, multi-asset anomalies: inter-satellite coupling (e.g., formation-flying orbital interference), constellation-level behaviors, and environment-driven collective effects are hard to observe from a single-asset viewpoint.

• Latency and responsiveness: ground-centric loops inherently incur communication gaps and delays, impeding on-orbit, time-critical mitigation.

• Limited adaptability: thresholds and fixed models struggle with previously unseen or poorly characterized failure modes.

Consequently, an FDIR regime that hinges on on-board scalar thresholds and ground-driven decisions is unlikely to meet future anomaly-detection requirements. The path forward is to (i) exploit mutual remote measurements—satellite-to-satellite and satellite-to-ground—so anomalies can be inferred from external observables as well as internal telemetry, (ii) adopt AI-enabled pattern discovery at constellation scale, and (iii) architect distributed, on-orbit autonomy so monitoring, detection, and first-line response are executed with real-time effectiveness across the fleet.

3. Results

Despite the constraints of legacy approaches, recent research shows that fusing multi-view remote-sensing data with AI methods can close critical gaps and enable finer-grained, more autonomous detection at a constellation scale. Here, we adopt an expanded notion of remote sensing that includes not only a spacecraft’s on-board telemetry but also satellite-to-Earth, satellite-to-satellite, and Earth-to-satellite observations. Such data expose off-nominal patterns that single-asset, interior-only monitoring frequently misses. This section reviews representative applications and summarizes the associated technical trends.

3.1. Satellite-to-Earth Observations for Self-Health Awareness

Downlinked high-resolution imagery, typically exploited for Earth analytics, can also be mined to infer the health of the observing spacecraft and its payloads. Abnormalities in downlink signal quality, payload artifacts, or spatiotemporal inconsistencies in the image stream may indicate issues spanning the radio frequency (RF) chain, focal-plane electronics, thermal regulation, or attitude control (Wang et al., 2021). Modern deep learning, particularly convolutional neural networks (CNNs), is a well-known architecture of artificial neural networks widely used for image-based detection tasks. It has proven effective in detecting structural defects and deformations directly in imagery, while time-series image analysis captures gradual drifts that unfold over long campaigns. Recent reports include U-Net–based models for solar-array fault detection and CNN pipelines that flag image-borne anomalies indicative of upstream hardware degradations (Rudro et al., 2024). These techniques complement internal FDIR by revealing externally manifested failure signatures (e.g., partial deployment, thermal gradients seen via radiometry, stray-light patterns) that may not breach on-board scalar thresholds yet portend system-level risk.

Fig. 3. Example of AI-based automatic anomaly detection using remote sensing images (Wang et al., 2021).

Key technical trends: (i) multi-modal fusion of imagery with attitude/thermal telemetry to improve attribution; (ii) self-supervised pretraining to mitigate label scarcity; and (iii) uncertainty-aware inference, i.e., explicitly modeling aleatoric/epistemic uncertainty (e.g., Bayesian/ensemble methods with calibrated confidence) to manage illumination, viewing-geometry, and environmental variability and to support risk-aware decisions.

3.2. Satellite-to-Satellite Observations for Relative-State Anomaly Detection

Mutual observation among spacecraft enables early detection of formation-level off-nominals. Monitoring relative position and attitude exposes orbit dispersion, geometry collapse, and coupling effects that are hard to perceive from a single vehicle. For example, gyro anomalies can be inferred from relative navigation residuals during formation flight (Shakouri and Assadian, 2018). NASA’s Starling Formation-Flying Optical Experiment (StarFOX) demonstrated angles-only, vision-based relative navigation across a four-CubeSat cluster, analyzing relative-orbit changes to flag non-nominal dynamics (Kruger et al., 2024).

Fig. 4. Monitoring a formation with three vehicles and a CubeSat inspector spacecraft (Caruso et al., 2022).

Beyond peer-to-peer sensing, the inspector spacecraft concept positions a dedicated vehicle to image and diagnose others on-orbit—useful for deployment verification, damage assessment, and proximity-operations safety (Caruso et al., 2022). Policy and procurement momentum is notable; for instance, in August 2025, the German Aerospace Center announced a pre-procurement notice for a rapidly fieldable inspection satellite aimed at in-orbit object assessment (Deutsche Raumfahrtagentur im DLR e.V., 2025). Key technical trends: (i) collaborative perception and distributed filtering across the swarm; (ii) graph-based learning over inter-satellite measurements; and (iii) on-board, event-triggered compute to limit downlink burden while preserving real-time responsiveness.

3.3. Earth-to-Satellite Observations for External Anomaly Cues

While still emergent for routine health management, ground-based optical sensing developed for space situational awareness (SSA) increasingly supports anomaly inference at scale. Wide-field telescopes and tracking systems can highlight atypical orbital maneuvers, fragmentation signatures, or photometric changes correlated with attitude or configuration anomalies (Al-Hourani and Manzoor, 2025). As catalogs grow and cadence increases, these feeds can provide independent, cross-segment evidence that augments on-board/space-borne views—particularly when contact windows are sparse or contested.

As shown in Fig. 5, experimental analyses demonstrate that satellite light-curve measurements derived from ground-based optical photometry are effective for attitude-state estimation. The results indicate that reliability and discriminative power vary with illumination and viewing geometry, underscoring the value of ground optical sensors in providing external cues of attitude/configuration anomalies that complement on-board observations.

Fig. 5. The initial and final illumination conditions used in the light curve scenario. L represents the illumination vector (Blacketer et al., 2022).

4. Discussion

Space missions operate in extreme, low-observability environments, making the robustness and trustworthiness of FDIR pipelines central to mission assurance. Recent AI-driven remote-sensing anomaly detection spans multiple data types (imagery, RF, telemetry, orbital ephemerides) and task goals (fault discovery, isolation, prognosis), with method selection closely tied to signal characteristics and operational constraints.

While promising, near-term operationalization of AI in space systems faces practical constraints: data scarcity (few labeled anomalies and distribution shift across missions), on-board computational limits (power, memory, and real-time latency budgets that bound model size and complexity), and certification and assurance challenges (verification/validation, traceability, and safety case construction for mission-critical deployments). Over the longer term, we anticipate these constraints will ease as larger curated datasets emerge, space-qualified compute improves, and standards and qualification pathways for learning-enabled systems mature.

CNNs for imagery: In image-centric scenarios, CNNs are now widely used to detect externally manifest faults and configuration issues. By learning spatial features, CNNs can visually expose structural deviations and payload artifacts that may remain invisible to scalar telemetry or orbital parameters. Reports increasingly confirm effectiveness for both internal (e.g., focal-plane electronics) and external (e.g., partial deployments, panel damage) conditions, with U-Net–style architectures frequently employed for part-level localization (Rudro et al., 2024).

Sequence models for time series: For orbit and telemetry time series, recurrent neural network (RNN)/long short-term memory (LSTM) and Transformer families dominate. LSTMs capture temporal dependence and are well-suited to progressive off-nominals such as orbit divergence or attitude instability (Li et al., 2023). More recently, Transformer-based forecasters learn nominal dynamics and flag anomalies via prediction-error residuals, including demonstrations on orbit-determination workflows (Ré et al., 2024).

Reconstruction-based methods: Autoencoders (AEs) learn nominal behavior and detect anomalies as reconstruction-error excursions, showing strong results on satellite telemetry in unsupervised regimes (Zhao et al., 2024). These approaches are attractive where labeled faults are scarce and model footprints must be compact for SmallSats (Jayeprokash and Gonski, 2025).

4.1. Constellation-Scale Anomaly Detection

Despite clear operational demand, published AI demonstrations at a constellation scale remain comparatively limited. A forward-leaning architecture mirrors safety-critical fleets on Earth (e.g., EVs): lightweight on-board diagnostic kernels integrated into flight software to (i) detect incipient faults; (ii) identify candidate root causes early; (iii) separate concurrent faults; and (iv) forecast time-to-failure—all while operating under tight compute and power budgets. Such autonomous on-orbit diagnostics and prognostics reduce ground dependence, improve decision latency under sparse contacts, and increase fleet-level resilience and service continuity.

Space-based solar power (SBSP) constellations epitomize the need. These systems—actively studied in the US, UK, and Japan—envision tens to hundreds of spacecraft forming a coherent structure for power aggregation and beam control (Glaser, 1968; Flournoy, 2011). Beyond single-asset health, operators must detect interaction-driven anomalies (array geometry loss, phasing instabilities) and safety-critical events during power beaming. Progress hinges on distributed autonomy (task recognition, formation management, beam steering) and—in the longer term—on-orbit assembly/maintenance, all of which benefit from embedded AI for rapid sense-plan-act cycles (Rodgers et al., 2024).

Fig. 6. Functional decomposition of space-based solar power (SBSP) Design Reference Systems (Rodgers et al., 2024).

Fig. 7. Autonomous swarm navigation of NASA’s Starling (NASA, 2023).

Large operational constellations offer early evidence. For BeiDou, LSTM-based detectors leveraging orbital and signal-quality metrics reportedly reduced orbit error by up to 20.8% (Cai et al., 2024). NASA’s Starling Swarm Mission (SSM) demonstrated on-orbit Distributed Spacecraft Autonomy (DSA) with crosslinks, vision-based relative/absolute navigation, and autonomous maneuver planning/execution. Within Starling, StarFOX validated angles-only relative navigation and constellation-aware orbit solutions—key ingredients for reducing ground reliance and enabling on-board anomaly response (NASA, 2023).

4.2. Single-Spacecraft Demonstrations

At the individual-vehicle level, numerous flight and operational studies show AI’s practical value:

• Early AE telemetry monitoring: A decade ago, the University of Tokyo applied autoencoders to spacecraft health monitoring, outperforming Principal Component Analysis (PCA) on anomaly separation (Sakurada and Yairi, 2014).

• On-board AI at scale (OPS-SAT): European Space Agency (ESA)’s OPS-SAT has hosted real-time on-orbit anomaly detection (AD) pipelines; the community OPS-SAT-AD benchmark supports reproducible evaluation. Additional on-board workloads (e.g., multi-frame super-resolution) foreshadow cluster- and swarm-class Earth Observation (EO) operations with embedded detection. In parallel, the Gaia program explored Machine Learning (ML)-assisted triage of anomalous records within its high-throughput science pipeline (Roberti et al., 2021). Collectively, these efforts demonstrate the feasibility of shifting first-line anomaly detection and quality control to the edge, reducing downlink burden and shortening the sense–decide–act loop.

• SMAP benchmark lineage: NASA/MIT’s Soil Moisture Active Passive (SMAP) mission (launched 2015) catalyzed open telemetry anomaly-detection benchmarks, triggering LSTM-led work followed by Autoencoder and Generative Adversarial Network (GAN) variants. These methods improved scientific data quality control and eased ground post-processing loads (Hundman et al., 2018).

Fig. 8. On-board anomaly detection using deep learning (KP LABS, 2022).

• Stripe-noise detection in EO imagery: For Gaofen-2, linear-target modeling of sensor-induced stripe artifacts coupled with YOLOv4 enhancements and a line-object loss delivered fast, large-volume detection—useful as a pre-restoration (destriping) quality gate across multi-sensor archives (Li et al., 2022).

• Physics-informed diagnostics (Antarctic Demonstrator for the Advanced Particle-astrophysics Telescope [ADAPT]/Electrical Power System [EPS]: Building on NASA’s ADAPT power-system datasets, physics-informed AI embeds EPS constraints (power conservation; current/voltage limits) to refine nominal distributions and detect out-of-distribution states, improving robustness and explainability under noisy, sparse data—enhancing deployability in flight operations (Cena et al., 2024).

Historically, anomaly detection emphasized internal sensor and link-signal analysis. Over the last decade, however, studies leveraging imagery and EO data have risen sharply: Shah et al. (2022) estimate ~2,500 EO-based anomaly-detection studies pre-2019, increasing to ~3,500 in 2020–2021 (≈+40%). This shift indicates a broadening from platform-centric sensing to payload-data–driven detection with modern AI, expanding both coverage (external manifestations) and semantic richness (fine-grained attribution). Table 1 synthesizes the aforementioned cases and supplements the characteristics of the required AI models.

Table 1. Summary of applications of AI methods for anomaly detection

5. Conclusions

Building on the preceding cases, we foresee a decisive shift in the data substrate for satellite anomaly detection: while telemetry remains foundational—capturing power, thermal, propulsion, and attitude signals with high diagnosticvalue—remote-sensing products (imagery, multi-spectral/hyperspectral, Sythetic Aperture Radar [SAR], AR, RF/photometry) will account for a growing share of actionable cues as systems migrate to formations, mega-constellations, and large collaborative structures. Sequence models (RNNs/Transformers) should further refine real-time deviation sensing on noisy, intermittent time series. In contrast, CNN/segmenter pipelines and self-supervised pretraining will extend image-driven detection to subtle structural drifts and environment-coupled behaviors that internal sensors under-represent. SAR- and spectroscopy-aware models, ideally coupled with physics priors, will add sensitivity to material and thermal states correlated with hardware degradation.

A central trajectory is a multi-modal fusion as the default design. Cross-evidence attribution—e.g., corroborating an EPS power drop with high-resolution imagery of solar-array damage or occlusion—improves diagnostic confidence and operator trust. Incorporating explainability (XAI) and calibrated uncertainty turns heterogeneous observations into causal narratives (what changed, where, and why), reduces false positives, and supports auditability across mission phases (Nazat et al., 2024).

At the constellation scale, federated learning enables fleets to learn from distributed observations without centralizing raw data, easing bandwidth pressure and privacy concerns while sharing anomaly priors across diverse vehicles and orbits. On top of perception and prediction, LLMs and agentic AI will increasingly serve as planning and triage layers, interpreting alerts, synthesizing hypotheses, recommending mitigations, and coordinating multi-asset responses within brief contact windows and limited human oversight.

Coupling these capabilities with digital twins creates a closed loop between “as-designed/as-expected” and “as-flown/as-observed” states. Streaming residuals between twin forecasts and fused measurements provide early warning, what-if analysis, and safer on-orbit reconfiguration, while physics-informed constraints stabilize learning in non-stationary environments.

Key open problems remain. Data scarcity and label sparsity call for self-supervision, weak supervision, and simulation-to-flight domain adaptation. On-board constraints demand co-design of models and compute—quantization, sparsity, event-triggered inference—suited to flight hardware. Robustness requires uncertainty calibration and physics-aware priors to manage illumination geometry, radiation, and distribution shift. Finally, rigorous operations-aligned evaluation is essential: benchmarks and KPIs should measure operator workload reduction, mean time to detect/mitigate, false-alarm costs, and fleet-level resilience—not merely offline accuracy.

Overall, the research vector converges on four intertwined outcomes: refined telemetry-based detection with modern time-series inference; rapid expansion of remote-sensing–based detection that directly observes structural and functional degradation; pervasive multi-modal fusion for accuracy, attribution, and trust; and autonomy at the edge, where federated learning, LLMs/agentic AI, and digital twins enable fleets to detect, diagnose, and initiate response with minimal ground latency.

This review synthesized the application space of AI for satellite anomaly detection across telemetry and remote-sensing information streams. Traditional FDIR—centered on on-board thresholds and ground-led analysis—has served single-asset missions well. Yet, its limitations are exposed by mega-constellations and large cooperative missions such as space-based solar power.

Recent advances demonstrate practical value across methods—CNNs for imagery, RNNs/Transformers for time series, autoencoders for unsupervised detection, federated learning for fleet-wide sharing, XAI for interpretability, and LLM/agentic layers for decision support. Notably, the proportion of studies exploiting EO/imagery has risen sharply (e.g., from roughly 20–25% pre-2015 to ≥55–60% during 2020–2023), signaling a paradigm shift beyond telemetry-only pipelines toward payload-data–driven health management.

Operational exemplars reinforce this trend: ESA OPS-SAT and Sentinel-2 have validated on-orbit, image-centric anomaly workflows; NASA Starling has demonstrated formation-flight autonomy and angles-only relative navigation for anomaly-aware operations; commercial fleets (e.g., Starlink, Planet Labs) report AI-assisted monitoring for link quality and sensor fault triage. Collectively, these efforts indicate a transition from research prototypes to embedded operational capability. Looking ahead, we anticipate progress along four axes:

• Refinement of telemetry-based detection with modern sequence models and uncertainty-aware inference

• Rapid growth of remote-sensing–based detection for direct observation of structural/functional degradation

• Multimodal fusion to raise accuracy, attribution, and explainability

• Autonomy at the edge, where LLMs, agentic AI, federated learning, and digital twins enable fleets to detect, diagnose, and initiate response with minimal ground latency.

In sum, AI-enabled fusion of telemetry and remote sensing will be central to next-generation satellite operations—improving stability, reliability, and resilience for both constellation services and long-duration exploration alike—and offering a concrete roadmap for both research communities and flight programs building the future space infrastructure.

Author Contributions

Conceptualization: Lee HH; Data curation, Formal analysis: Lee HH, Kim JW; Funding acquisition: Lee HH; Investigation, Methodology: Lee HH, Kim JW; Project administration: Lee HH; Supervision: Lee HH; Validation: Lee HH, Kim JW; Writing–original draft, Writing–review & editing: All authors.

Conflicts of Interest

No potential conflict of interest relevant to this article was reported.

Funding

This research was supported by the Industrial Technology Alchemist Project(R&D) funded by the Ministry of Trade, Industry and Energy of Korea in 2025 through the Korea Institute for Advancement of Technology (KIAT)(No. 07472986).

Data Availability Statement

Data sharing is not applicable to this article as no new data were created or analyzed in this study.