LLM-Driven Modeling and Decision Support Methods for Cross-Domain Collaborative Mission Systems

Abstract

Cross-domain formations composed of Unmanned Aerial Vehicles (UAVs) and Unmanned Surface Vessels (USVs) are critical for maritime defense but face significant challenges in countering complex aerial threats and developing flexible, collaborative strategies. Addressing the limitations of traditional decision support systems in semantic understanding and dynamic adaptation, this paper proposes a novel Large Language Model (LLM)-driven decision support framework grounded in the Department of Defense Architecture Framework (DoDAF). By integrating Retrieval-Augmented Generation (RAG) with a domain-specific knowledge base, the framework enhances the LLM’s ability to align natural-language directives with standardized DoDAF view models, effectively mitigating hallucinations in tactical generation. The proposed framework coordinates a closed-loop process, using Petri net-based static logic verification to ensure structural consistency and Monte Carlo-based dynamic effectiveness evaluation to optimize the selection of kill chains. Experimental validations in a simulated UAV-USV maritime defense scenario demonstrate that the framework achieves 96.6% entity accuracy and 100% format compliance in model generation. In comparison, the generated cooperative kill chains significantly outperform non-cooperative methods by improving interception efficacy by approximately 26.08% under saturation attack conditions. This study develops an automated, interpretable workflow that transforms unstructured situational understanding into decision reporting, significantly enhancing the efficiency and reliability of cross-domain collaborative mission planning.

1. Introduction

Currently, cross-domain formations composed of Unmanned Aerial Vehicles (UAVs) and Unmanned Surface Vessels (USVs) face increasingly severe aerial threats and complex, ever-changing battlefield environments when conducting maritime defense missions. Strengthening the collaborative mission capabilities between UAVs and USVs is of significant strategic importance for improving the overall mission effectiveness of the formation. However, the lack of a unified top-level planning framework at the tactical command support level presents numerous challenges to the generation of intelligent collaborative combat solutions.

In recent years, AI technology has advanced rapidly, evolving from simple perceptual intelligence to higher-level cognitive intelligence, playing a crucial role across a range of complex scenarios. In particular, the emergence of Large Language Models (LLMs) have demonstrated robust intent understanding, logical reasoning, and generalization capabilities, providing new insights into solving complex decision-making problems in cross-domain collaborative tasks. In UAV-USV cross-domain collaborative scenarios, the high dimensionality of the defense space and the strong adversarial nature of the mission lead to an exponential increase in decision-making difficulty. Traditional decision support systems based on rules or single optimization algorithms have seen development across various fields [1,2], but they often struggle to cope with dynamic changes and cannot understand the semantics of complex command intentions. Therefore, introducing an LLM-driven intelligent decision-making framework can not only assist in analyzing ambiguous battlefield situations but also establish a semantically consistent interaction mechanism across different platforms, thus providing a core driving force for generating flexible and efficient collaborative solutions.

To address the aforementioned challenges of semantic understanding and dynamic adaptability in existing cross-domain mission systems, this study proposes a novel LLM-driven decision-support framework. This approach synergizes the generative capabilities of Artificial Intelligence with the rigorous structural standards of the Department of Defense Architecture Framework (DoDAF). By integrating Retrieval-Augmented Generation (RAG) with domain-specific knowledge, we aim to provide a transparent, adaptive, and verifiable solution for complex engineering environments. The main contributions of this paper are summarized as follows:

1.

An innovative systems engineering framework for transparent modeling is proposed. We architect a novel integration pipeline that fuses LLMs with the DoDAF standard. By leveraging RAG to align natural-language directives with standardized architectural views, this method effectively bridges the semantic gap. Importantly, we strictly distinguish the framework’s interpretability into two dimensions: the intrinsic interpretability of the underlying reasoning pipeline, guaranteed by the rigorous structural mapping of DoDAF and formal Petri net verification, and the post hoc explainability of the outputs, provided by the LLM-generated reasoning traces. This dual-layered transparency mitigates the “hallucination” risks inherent in standalone generative models.

2.

A hybrid computing framework is established for adaptive decision making. We construct a closed-loop verification mechanism that combines Petri nets for static logic validation with Monte Carlo simulations for dynamic effectiveness evaluation. This hybrid approach enables the system to handle the uncertainty of stochastic battlefield environments and ensures that generated kill chains are both logically sound and operationally effective.

3.

The practical application efficacy is validated in multidisciplinary scenarios. We apply the proposed method to a cross-domain collaborative interception mission involving heterogeneous UAV and USV formations. Experimental results demonstrate that this AI-empowered system improves interception efficiency and adaptability compared to traditional non-collaborative approaches, providing a verified reference for UAV-USV maritime defense mission systems.

The remainder of this paper is organized as follows: Section 2 reviews related work on intelligent decision support and DoDAF applications, Section 3 details the proposed LLM-driven modeling framework and the hybrid verification mechanism, Section 4 presents the experimental setup and analyzes the simulation results in a cross-domain maritime defense scenario, and finally, Section 5 concludes the paper.

2. Related Work

2.1. DoDAF System Modeling and Analysis Methods

2.1.1. Overview of DoDAF Methodology

The System of Systems (SoS) is an important concept in systems engineering. A mission system is composed of several independent and interacting equipment systems, and these interactions are also an important part of the system [3]. Designing a decision support system for a mission system requires first establishing an understanding of the mission scenario and the mission system itself; therefore, research on mission system modeling and analysis methods is necessary. Existing typical mission system modeling and analysis methods mainly include the DoDAF, the Ministry of Defense Architectural Framework (MoDAF), and the Unified Architecture Framework (UAF), among which DoDAF and UAF are the most widely used.

1.

DoDAF method.

In the field of mission system research, the DoDAF is currently the most widely adopted methodology for modeling and analyzing the composition, functions, and interactions of military systems. The framework has evolved through several major iterations, with the latest version, DoDAF 2.02, released in January 2015 [4]. Since then, its military applications have been extensively explored. In 2019, Posherstnik et al. [5] proposed a DoDAF-based framework for joint command and control, focusing on electromagnetic spectrum interoperability. Subsequently, Beery et al. [6] applied DoDAF to defense engineering management in 2021 to guide the development of system architecture products. In 2022, Xu et al. [7] introduced a modeling method for the precision strike process in joint anti-carrier operations, utilizing SysML to describe the system architecture. Zhuo [8] later proposed a concept model construction method based on DoDAF in 2023, developing event description and state space models for a territorial sea defense system, which were validated through simulation. Zhao et al. [9] implemented the modeling and simulation of an unmanned landing combat system, completing the campaign-level joint landing concept design utilizing the All Viewpoint (AV) and Operational Viewpoint (OV) models of DoDAF. Similarly, Zhao et al. [10] analyzed the anti-drone swarm combat architecture, identifying command relationships, information flows, and state transitions. In 2024, Zhao et al. [11] proposed an iterative design method combining DoDAF with ExtendSim executable models, demonstrating its feasibility in an anti-submarine warfare context. In summary, DoDAF has demonstrated a high level of maturity in mission system modeling and concept design.

2.

UAF method.

The Unified Architecture Framework (UAF) is an architectural design methodology for modeling relationships between organizations and systems, released by the Object Management Group (OMG) in 2017. In 2021, Hause et al. [12] proposed a UAF-based cybersecurity defense modeling method, defining defense objectives and their integration into defensive operations through a set of security viewpoints. In 2023, Lu et al. [13] introduced a UAF-based architecture modeling method for the air defense and anti-missile combat system of a carrier strike group. This study primarily focused on the modeling process and combat requirements but lacked a detailed discussion on engineering applications. Dong [14] implemented the modeling and simulation of air traffic control scenarios by defining functional domains based on UAF. Furthermore, Martin et al. [15] proposed a UAF-based method to model the mission analysis process in combat planning, achieving the description of combat mission concepts and entity attributes.

A comparative analysis of the aforementioned methods reveals that the DoDAF method is relatively mature in combat system modeling and analysis applications. Its comprehensive view models can effectively characterize relevant elements of various complex mission scenarios. Although the UAF method is relatively novel, its primary applications focus on non-mission scenarios. In the context of mission system modeling, UAF cannot currently construct multi-dimensional multi-view models directly, which affects its reliability and accuracy. Consequently, this paper adopts the DoDAF method to model and analyze the defense system of the UAV-USV formation.

2.1.2. Introduction to DoDAF Views

The U.S. Department of Defense has released several different versions of the DoDAF framework [16], which have constructed standardized representation methods for the architecture and related elements [17], ensuring the consistency of the construction rules, views and equipment descriptions of the architecture model, and ensuring that the subsystems within the system can achieve integration and interoperability, thereby meeting the mission element coordination requirements in the context of cross-domain joint mission. The latest version of the DoDAF architecture framework 2.02 [18] defines eight views for the architecture description: All Viewpoint (AV), Capability Viewpoint (CV), Data and Information Viewpoint (DIV), Operational Viewpoint (OV), Project Viewpoint (PV), Services Viewpoint (SvcV), Standards Viewpoint (StdV), and Systems Viewpoint (SV). Each view also contains 52 functional models [19].

Table 1. DoDAF Viewpoints and Model Definitions.

All Viewpoint (AV)	Project Viewpoint (PV)	Data & Info Viewpoint (DIV)	Standards Viewpoint (StdV)
AV-1: Overview and Summary Information	PV-1: Project Portfolio Relationships	DIV-1: Conceptual Data Model	StdV-1: Standards Profile
AV-2: Integrated Dictionary	PV-2: Project Timelines	DIV-2: Logical Data Model	StdV-2: Standards Forecast
	PV-3: Project to Capability Mapping	DIV-3: Physical Data Model
Capability Viewpoint (CV)	Services Viewpoint (SvcV)	Operational Viewpoint (OV)	Systems Viewpoint (SV)
CV-1: Vision	SvcV-1: Services Context Description	OV-1: High-Level Operational Concept Graphic	SV-1: Systems Interface Description
CV-2: Capability Taxonomy	SvcV-2: Services Resource Flow Description	OV-2: Operational Resource Flow Description	SV-2: Systems Resource Flow Description
CV-3: Capability Phasing	SvcV-3a: Systems-Services Matrix	OV-3: Operational Resource Flow Matrix	SV-3: Systems-Systems Matrix
CV-4: Capability Dependencies	SvcV-3b: Services-Services Matrix	OV-4: Organizational Relationships Chart	SV-4: Systems Functionality Description
CV-5: Capability to Organizational Development Mapping	SvcV-4: Services Functionality Description	OV-5a: Operational Activity Decomposition Tree	SV-5a: Op. Activity to Systems Functionality Traceability Matrix
CV-6: Capability to Operational Activities Mapping	SvcV-5: Operational Activity to Services Traceability	OV-5b: Operational Activity Model	SV-5b: Op. Activity to Systems Traceability Matrix
CV-7: Capability to Services Mapping	SvcV-6: Services Resource Flow Matrix	OV-6a: Operational Rules Model	SV-6: Systems Resource Flow Matrix
	SvcV-7: Services Measures Matrix	OV-6b: State Transition Description	SV-7: Systems Measures Matrix
	SvcV-8: Services Evolution Description	OV-6c: Event-Trace Description	SV-8: Systems Evolution Description
	SvcV-9: Services Tech & Skills Forecast		SV-9: Systems Tech & Skills Forecast
	SvcV-10a: Services Rules Model		SV-10a: Systems Rules Model
	SvcV-10b: Services State Transition Description		SV-10b: Systems State Transition Description
	SvcV-10c: Services Event-Trace Description		SV-10c: Systems Event-Trace Description

Definitions derived from DoDAF Version 2.02 [18].

shows the DoDAF view and functional model definitions.

2.2. LLM-Driven Multi-View Modeling Analysis of System

The DoDAF framework does not restrict the application combination of views and functional models. Based on the aforementioned DoDAF views and functional models, a set of relevant elements involving weapon systems, such as mission operations, equipment systems, and system capabilities, can be constructed as needed. Selecting appropriate views and models enables the representation of relationships within the mission system, which is a crucial aspect of system-wide missions. The ability to quickly select and combine views during mission execution, building a view set adapted to the mission, is one of the key reasons for introducing LLM on top of DoDAF.

In recent years, Cognitive AI, also known as Knowledge-based AI or Symbolic AI, has been successfully used to enhance multi-agent system-based simulations, enabling them to make tactical decisions in dynamic environments [20]. LLM enables communication between humans and artificial intelligence systems in military applications. This includes publishing theories and objectives in natural language, enabling AI agents to understand and execute complex instructions [21]. Chatbots and conversational AI systems, such as Copilot and Gemini, have been used for decision support in command-and-control processes, demonstrating their potential to provide accurate and comprehensive information [22].

The integration of LLM with the DoDAF framework has also proven feasible. In 2025, Topcu et al. [23] limited the dialogue scenario to communication with DoD-related personnel using prompts. LLM can output defense-related terminology based on the questioner’s questions, including DoDAF-related military data and decision results. In the same year, Ormo [24] input rich, structured data sources related to DoDAF into an LLM, equipped with integrated organizational and system models, enabling the AI agent to understand organizations and related systems, achieving AI-driven view model creation while reducing the workload required to create specific task prompts. Therefore, LLM-driven decision-making scheme generation has great potential. However, LLM currently only has natural language understanding and reasoning capabilities; it still cannot achieve logical verification of view models or effectiveness evaluation of kill chains. A framework driven by LLM is needed to support the complete closed-loop process from task issuance to scheme generation and verification.

3. LLM-Driven DoDAF System Modeling and Decision Generation Framework

3.1. Overview of the Framework

The LLM-driven DoDAF system modeling and decision generation framework is shown in Figure 1.

As illustrated in Figure 1, upon receiving situational information for the mission scenario, the LLM synthesizes this input with information retrieved via Retrieval Augmented Generation (RAG) to perform reasoning. It infers the required view models and invokes external modules to generate them. Following the modeling phase, the LLM continues to orchestrate the modules to perform static and dynamic verification, including Petri net-based logical verification, temporal logic consistency checks, kill chain effectiveness assessment, and model completeness verification. The results are fed back to the LLM for secondary reasoning and natural language report generation, culminating in the selection of the optimal kill chain and the formulation of a comprehensive decision scheme.

This LLM-centric decision framework highly automates the generation and verification of mission plans by leveraging RAG-enhanced inputs, the LLM’s semantic understanding and logical reasoning, and the optimized orchestration of functional modules. This significantly reduces the time required for view modeling and decision generation. Furthermore, the reliance on specialized small models for verification and evaluation ensures the traceability of the decision process.

3.2. Prompt Engineering

In the proposed framework, the LLM does not autonomously determine decision logic or system behavior. Instead, all outputs are strictly constrained by predefined system workflows, formal architectural rules, and module invocation protocols.

Contextual information, including background knowledge, the complete decision-making workflow, and interface specifications of functional modules, is exposed to the LLM through a structured interaction interface. This interface enables the model to interpret user requests within a bounded and well-defined system context, rather than relying on unconstrained natural language reasoning [25].

Prior to task execution, the LLM is provided with a fixed set of system-level inputs, including workflow descriptions, module invocation schemas, and structured output templates. These elements serve as communication mechanisms between the system and the model, rather than constituting the core methodological contribution. Illustrative examples of such interface configurations are provided in Appendix A for clarity and reproducibility.

3.3. Situation- and Resources-Based DoDAF Modeling

The proposed automated view model generation method, driven by a LLM and executed by external modeling modules, is fundamentally based on establishing a mechanism for mapping and transmitting unstructured situational semantics to structured model parameters.

This mechanism begins with the LLM comprehending prompts and employing RAG to analyze natural-language scenario descriptions deeply. It then autonomously plans the necessary view types based on situational characteristics—for instance, constructing OV-1 to depict command relationships and OV-5b to represent temporal features. Subsequently, the LLM orchestrates interactions with specialized modeling modules in a logical sequence. The LLM transforms identified mission entities and logical relationships into standardized intermediate-format parameters. Via API interfaces, it drives external modeling modules to automatically render graphical elements and optimize layout, ultimately generating system architecture views that comply with DoDAF specifications. The following section details the logic of view construction, along with the functions and application scenarios of each view.

3.3.1. UAV-USV System Modeling Steps and Implementation Methods

In the early stage of system architecture design, it is essential first to formulate a preliminary concept for the mission scenario. Starting with system requirements analysis, the process involves capturing and analyzing requirements. Based on these results, system functional analysis and modeling are performed to establish the composition relationships between the mission system and each subsystem. Finally, functional models—such as mission actions, system interfaces, and data flows—are generated to finalize the modeling design of the entire UAV-USV formation defense and counterattack mission system [26]. Therefore, the system modeling process is categorized into four steps: (1) requirement capture; (2) requirement analysis; (3) functional analysis; and (4) functional modeling. Following this four-step framework [27], this chapter extracts key functional models relevant to the maritime defense and counterattack mission system from DoDAF views, analyzed as follows:

1.

Requirement Capture: Construction of System All-Viewpoint Model. In the requirement capture stage, using the AV-1 Overview and Summary Information model and based on the mission scenario and system concept, the top-level vision of the system mission is developed. This covers the system mission intent, thereby forming an intuitive understanding of system requirements.

2.

Requirement Analysis: Construction of System Capability Model. Based on the system overview formed in the requirement capture stage, the system capabilities and taxonomy required to support the mission concept are further decomposed using the CV-1 Capability Vision model and the CV-2 Capability Taxonomy model. Subsequently, the CV-4 Capability Dependencies model is used to analyze and define the capability dependencies among systems within the mission architecture.

3.

Functional Analysis: Construction of System Mission View Model. In the functional analysis stage, the conceptual design of the maritime defense and counterattack mission system is first completed based on the OV-1 High-Level Operational Concept Graphic [28]. Subsequently, relationships between mission resources are established via the OV-2 Mission Resource Flow Description model. On this basis, mission actions, rules, and effects are defined using the OV-5b, OV-6a, and OV-6b models, respectively, while mission event tracking is conducted using the OV-6c model.

4.

Functional Modeling: Construction of System View Model Upon completion of the aforementioned processes, the component analysis of the auxiliary decision-making system is achieved through functional models such as the SV-1 Systems Interface Description, SV-2 Systems Resource Flow Description, and SV-4 Systems Functionality Description under the System View. The entire process is illustrated in Figure 2.

3.3.2. Construction of System Panoramic View Model

In the requirement capture phase, the concept and scope of the maritime defense and counterattack mission system were established based on the AV-1 Overview and Summary Information model. The summary content of the AV-1 model proposed in this paper is presented in

Table 2. Overview of the Maritime Air Defense and Anti-Missile Mission System AV-1.

Category	Description
Background	To strengthen the UAV-USV formation defensive counterattack mission capability, the networked collaboration of C2 systems (auxiliary decision-making systems), sensor systems, interception systems, and electronic countermeasure systems is utilized. This achieves intelligent target threat assessment, formation mission analysis, and COA (Course of Action) generation, thereby completing the kill chain closed-loop and constructing a novel distributed, networked formation defensive counterattack mission system.
Objective	Centered on constructing UAV-USV collaborative kill chains, the objective is to maximize the utilization of sensor systems and interception systems within the formation, enhancing the formation’s situational awareness capabilities and improving the interception success rate against threat targets.
Constraints	The utilization of weaponry and equipment must strictly adhere to tactical rules and mission doctrines. Building upon intelligent decision support, it is also necessary to improve the efficiency and convenience of the interaction between the commander and the auxiliary decision-making system.
Models	Select relevant functional models under the All Viewpoint (AV), Capability Viewpoint (CV), Operational Viewpoint (OV) (referred to as Mission View), and Systems Viewpoint (SV).

.

3.3.3. Construction of System Capability View Model

According to the CV-1 Capability Vision model, the system-level capability requirement for the maritime defense and counterattack mission system is to counter aerial threats by establishing a regional defense architecture that integrates “detection, tracking, guidance, and strike.” Considering the composition of the UAV-USV formation, the maritime defense and counterattack capability requirements are decomposed into five system-level capabilities based on the CV-2 Capability Taxonomy model: Command and Decision Capability, Target Detection Capability, Air Defense Operations Capability, Electronic Countermeasures Capability, and Cross-Domain Coordination Capability. By further decomposing these system-level capabilities, the subsystem-level capability requirements within the architecture are derived. Figure 3 illustrates the Capability Taxonomy Model for the maritime defense and counterattack mission system.

1.

The Command and Decision Capability refers to the integrated formation-level command capability provided by the auxiliary decision-making system. At the subsystem level, primarily, it requires Situational Data Processing Capability to ensure the accuracy of situational information fusion. Secondly, it must provide an essential Target Threat Assessment Capability. Subsequently, it enables Target Intent Identification based on target characteristics. On this basis, it facilitates Formation Mission Analysis and COA Generation. This procedure constitutes a critical decision-making process constrained by the OODA loop, thereby enabling a closed-loop operation at the formation level of the kill chain.

2.

Target Detection Capability refers to the ability to detect and track targets provided by shipborne or airborne sensors (specifically radars). Shipborne detection and tracking capabilities are typically supplied by radars mounted on Unmanned Surface Vessels (USVs). Airborne detection and tracking capabilities are provided by Unmanned Airborne Early Warning (UAEW) aircraft and Unmanned Combat Aerial Vehicles (UCAVs).

3.

Defense Operations Capability is constituted by USVs and UAVs equipped with interception weaponry. The effectiveness of target neutralization is determined jointly by the guidance and strike capabilities of the weapons. Guidance serves as a prerequisite for completing target destruction within the kill chain. Since the defensive counterattack mission of the UAV-USV formation can be executed through cross-domain sea-air coordination, UAEWs, UCAVs, and USV radars can all provide guidance functions. This introduces complexity into cross-domain collaborative decision-making and COA generation.

4.

Electronic Countermeasures Capability primarily comprises Electronic Jamming Capability and Electronic Defense Capability. Electronic Jamming Capability refers to the support jamming provided by Unmanned Electronic Attack (UEA) aircraft. Electronic Defense Capability refers to the ability to guarantee information transmission assurance among various mission groups or units within the formation.

5.

Cross-Domain Coordination Capability consists of Networked Communication Capability and Data Sharing Capability. Networked Communication Capability is typically determined by the data link capabilities deployed on individual platforms within the formation. Data Sharing Capability refers to the information-sharing ability within the formation driven by the network communication system. It encompasses functions such as collaborative data processing and dissemination, enabling the formation to possess collaborative detection, situational sharing, and target tracking capabilities.

Given these capabilities’ definitions, it is essential to further specify the dependency relationships among subsystem-level capabilities in accordance with the CV-4 Capability Dependencies model, as illustrated in Figure 4.

During the construction phase of the Capability Viewpoint, leveraging the CV-2 Capability Taxonomy and the subsystem-level CV-4 Capability Dependencies facilitates the precise definition of requirements and logical relationships among capabilities within the architecture. This establishes a critical foundation for the subsequent development of the Operational Viewpoint and the Data and Information Viewpoint models.

3.3.4. Construction of System Mission View Model

Constructing the Operational Viewpoint (OV) models is a critical approach for comprehensively describing operational elements, including mission concepts, organizational relationships, activity dependencies, and operational rules. Consequently, this section first defines the operational concept of the entire mission system based on the OV-1 High-Level Operational Concept Graphic. Subsequently, it characterizes operational resources and their flow relationships within the architecture using the OV-2 Operational Resource Flow Description. Following this, the OV-4 Organizational Relationships Chart is constructed to detail the interrelationships among various subsystems. Finally, the attributes of specific operational activities are further specified based on the OV-5b Operational Activity Model, the OV-6a Operational Rules Model, and the OV-6b State Transition Description.

1.

OV-1 High-Level Operational Concept Graphic. The maritime defense and counterattack mission system proposed in this paper focuses on the UAV-USV formation. Designed for defensive counterattack mission scenarios, it enables intelligent target threat assessment, formation mission analysis, and COA (Course of Action) generation through networked collaboration among C2 systems (auxiliary decision-making systems) [29], sensor systems, interception systems, and electronic countermeasure systems. This establishes a UAV-USV formation defense and counterattack mission architecture characterized by intelligent decision support. Figure 5 illustrates the OV-1 High-Level Operational Concept Graphic for this mission system.

The OV-1 High-Level Operational Concept Graphic designed in this section is tailored for the maritime defense and counterattack scenario of the UAV-USV formation. It primarily incorporates key subsystem elements involved in the scenario, including the UAV Launch Platform, USV Formation, Unmanned Airborne Early Warning (UAEW) Formation, Unmanned Combat Aerial Vehicle (UCAV) Formation, and Unmanned Electronic Attack (UEA) Formation. The hostile targets are predominantly identified as aerial threats, such as missiles or aircraft.

2.

OV-2 Operational Resource Flow Description. Further analysis based on the OV-1 High-Level Operational Concept Graphic yields the OV-2 Operational Resource Flow Description. This model characterizes the resource-interaction requirements among subsystem nodes within the mission architecture. Systematically identifying and analyzing these resource interaction requirements provides a critical basis for the requirements justification and interface analysis of the auxiliary decision-making system.

Table 3. OV-2 Operational resource flow model for maritime defense and counterattack mission system.

Resource Flow ID	Resource Flow Type	Source Node	Target Node
SitnFlow01	Situational Flow	UCAV	FOC
SitnFlow02	Situational Flow	USV	FOC
SitnFlow03	Situational Flow	UAEW	FOC
SitnFlow04	Situational Flow	UEA	FOC
CtrlFlow01	C2 Flow (Command & Control)	FOC	UCAV
CtrlFlow02	C2 Flow	FOC	USV
CtrlFlow03	C2 Flow	FOC	UAEW
CtrlFlow04	C2 Flow	FOC	UEA
CtrlFlow05	C2 Flow	UCAV	USV
CtrlFlow06	C2 Flow	UAEW	USV

presents the OV-2 Operational Resource Flow Description.

In this section, resource flows are categorized into Situational Flows and Command and Control (C2) Flows. Situational Flows primarily comprise feedback information derived from target detection, whereas C2 Flows are typified by command resources such as operational directives and guidance data.

3.

OV-4 Organizational Relationships Chart Model. The OV-4 Organizational Relationships Chart primarily illustrates the types, quantities, and compositional hierarchies of subsystems within the architecture. Figure 6 defines the organizational relationships for the maritime defense and counterattack mission. Typically, the composition of the UAV-USV formation is determined based on mission requirements prior to assembly.

Conventional UAV-USV formations are generally structured around the Flight Ops Center. They comprise USV, UCAV, UAEW, and UEA formations, thereby forming a joint combat force integrated with diverse equipment.

Table 4. Example of UAV-USV Formation Configuration under the Maritime Defense and Counterattack Mission System.

Type	Number of Formations	Units per Formation	Total Units
Flight Ops Center	1	1	1
USV Formation	1	4	4
UCAV Formation	1	12	12
UAEW Formation	1	4	4
UEA Formation	1	4	4

presents an example of such a defense and counterattack formation configuration, omitting other non-primary combat elements.

4.

OV-5b Operational Activity Model. Based on the OV-1 High-Level Operational Concept Graphic, OV-2 Operational Resource Flow Description, and OV-4 Organizational Relationships Chart, the OV-5b Operational Activity Model is constructed. This model reflects the operational activity processes of each subsystem element within the mission architecture. First, a checklist of basic feasible activities for each subsystem is derived from the OV-3 Operational Resource Flow Matrix, as presented in

Table 5. List of Basic Feasible Activities based on OV-5b.

Type	Activity ID	Activity Description	Quantity
Command and Flight Ops Center	FOCAct01	Situational Data Processing	6
	FOCAct02	Target Threat Assessment
	FOCAct03	Target Intent Identification
	FOCAct04	Formation Mission Analysis
	FOCAct05	COA Generation
	FOCAct06	COA Dissemination
USV Formation	USVAct01	USV Target Detection	6
	USVAct02	USV Target Tracking
	USVAct03	Shipborne Missile Launch
	USVAct04	Shipborne Missile Initial Guidance
	USVAct05	Shipborne Missile Mid-course Guidance
	USVAct06	Shipborne Missile Terminal Guidance
UEA Formation	UEAAct01	UEA Electronic Reconnaissance	2
	UEAAct02	UEA Electronic Jamming
UCAV Formation	UCAVAct01	UCAV Target Detection	7
	UCAVAct02	UCAV Target Tracking
	UCAVAct03	Air-to-Air Missile Launch
	UCAVAct04	Air-to-Air Missile Initial Guidance
	UCAVAct05	Air-to-Air Missile Mid-course Guidance
	UCAVAct06	Air-to-Air Missile Terminal Guidance
	UCAVAct07	Shipborne Missile Mid-course Guidance
UAEW Formation	UAEWAct01	UAEW Target Detection	4
	UAEWAct02	UAEW Target Tracking
	UAEWAct03	Shipborne Missile Mid-course Guidance
	UAEWAct04	Air-to-Air Missile Mid-course Guidance

. Subsequently, based on this checklist, the comprehensive OV-5b Operational Activity Model for the architecture is developed [30].

In Figure 7, the activities listed under each swimlane represent the operational activities of that specific subsystem, while the connecting lines define the sequential logic and tactical rule constraints between activities. This model designates the UAV Flight Ops Center as the Command and Control node, equipped with an auxiliary decision-making system for air defense and anti-missile missions. Upon initiation of the system-level Air Defense Early Warning Mission, each subsystem node enters the operational process in accordance with tactical rules. Initially, the USV Formation, UCAV Formation, and UAEW Formation conduct target detection operations, while the UEA Formation executes electronic reconnaissance operations. Once a target is detected, the subsystem nodes report target data to the C2 node. The auxiliary decision-making system at the C2 center then executes a series of specific actions, including Situational Data Processing, Target Threat Assessment, Target Intent Analysis, Formation Mission Analysis, and COA Generation. Following the commander’s confirmation of the engagement and the COA, the approved Course of Action (COA) is disseminated to the respective subsystem nodes. Figure 7 illustrates the OV-5b Operational Activity Model of the combat system.

Table 6. Target Neutralization Kill Chain Combinations.

Kill Chain ID	Kill Chain Sequence	Collaborative
KillChain01	USVAct01 → USVAct02 → USVAct03 → USVAct04 → USVAct05 → USVAct06 → Target Neutralized	No
KillChain02	UCAVAct01 → UCAVAct02 → UCAVAct03 → UCAVAct04 → UCAVAct05 → UCAVAct06 → Target Neutralized	No
KillChain03	USVAct01 → USVAct02 → USVAct03 → USVAct04 → UCAVAct05 → USVAct06 → Target Neutralized	Yes
KillChain04	USVAct01 → USVAct02 → USVAct03 → USVAct04 → UAEWAct03 → USVAct06 → Target Neutralized	Yes
KillChain05	UCAVAct01 → UCAVAct02 → UCAVAct03 → UCAVAct04 → UAEWAct03 → UCAVAct06 → Target Neutralized	Yes
KillChain06	UAEWAct01 → UAEWAct02 → USVAct03 → USVAct04 → USVAct05 → USVAct06 → Target Neutralized	Yes
KillChain07	UAEWAct01 → UAEWAct02 → UCAVAct03 → UCAVAct04 → UCAVAct05 → UCAVAct06 → Target Neutralized	Yes
KillChain08	UAEWAct01 → UAEWAct02 → USVAct03 → USVAct04 → UCAVAct07 → USVAct06 → Target Neutralized	Yes
KillChain09	UAEWAct01 → UAEWAct02 → USVAct03 → USVAct04 → UAEWAct03 → USVAct06 → Target Neutralized	Yes
KillChain10	UAEWAct01 → UAEWAct02 → UCAVAct03 → UCAVAct04 → UAEWAct04 → UCAVAct06 → Target Neutralized	Yes

presents the Kill Chain Combinations for target neutralization. As shown, multiple kill chains exist based on the collaborative relationships within the current mission system. KillChain01 and KillChain02 represent self-contained kill loops executed by the USV Formation and UCAV Formation, respectively. KillChain03 through KillChain10 represent coordinated kill chains completed jointly by the USV, UCAV, and UAEW formations.

The kill chains currently presented do not account for target allocation in multi-target scenarios. In the event of a hostile saturation attack, the resulting complex situation would lead to a sharp increase in the number of potential kill chains. Therefore, selecting the optimal kill chain combination to formulate the formation’s Mission COA is a critical objective of COA generation and optimization. Building on the identified kill chain combinations, it is necessary to define further the rules governing the activities within these chains using the OV-6a Operational Rules Model. Additionally, the OV-6b State Transition Description is used to determine the key elements for kill-chain evaluation, thereby supporting the requirements input for the auxiliary decision-making system design process [31].

5.

OV-6a Operational Rules Model and OV-6b State Transition Description. The objective of constructing the OV-6a Operational Rules Model is to provide a formal definition and description of operational activities. For any given operational activity $Act$, the model encompasses the following elements: activity attributes $Attribute$, operational resource requirements $ActRes$, preconditions $ActS$, termination conditions $ActE$, execution status $ActStatus$, input events $ActInput$, output events $ActOutput$, evaluation events $ActEva$, and the activity description function $f\left(Act\right)$. To simplify state transition processing, state transition parameters are integrated into the Action Output Event $ActOutput$, while the evaluation metric system is incorporated into the Action Evaluation Event $ActEva$. Specifically, the Action Attribute $Attribute$ comprises the Action Identifier $ActID$, Start Time $ActStime$, End Time $ActEtime$, Duration $ActDurtime$, Start Position $ActPos$, and End Position $ActEPos$. Mission Resource Requirements include Weapon Requirements $ActResWeapon$ and Sensor Requirements $ActResSensor$. The Action Preconditions $ActS$ encompass the Execution Status of Preceding Actions $PreActStatus$ and the Current Status of the Hostile Target $TarPreStatus$. The Action Termination Condition $ActE$ is defined as the Achieved State of the Target $TarEndStatus$. The Action Input Event $ActInput$ serves as the trigger condition for the action. The Action Output Event $ActOutput$ represents output events generated during the operational process, including the Execution Effect $ActEffect$, Weapon Resource Consumption $ActResWeaponDe$, and Sensor Resource Consumption $ActResSensorDe$. The Action Evaluation Event $ActEva$ constitutes the set of metrics constructed for operational effectiveness evaluation, including Communication Latency $ActTC$, Other Latencies $ActTZ$, Timeliness Evaluation $ActWT$, Accuracy Evaluation $ActWP$, Target Acquisition Assessment $ActP1$, and Damage Assessment $ActP2$. Consequently, the formal definition of the Mission Action $Act$ can be represented by Equations (1)–(7).

$$\begin{array}{cc}\hfill Act=& \langle Attribute,ActRes,ActS,ActE,ActStatus,\hfill \\ \hfill & ActInput,ActOutput,ActEva,f\left(Act\right)\rangle \hfill \end{array}$$

(1)

$$\begin{array}{cc}\hfill Attribute=& \langle ActID,ActStime,ActEtime,ActDurtime,ActSPos,ActEPos\rangle \hfill \end{array}$$

(2)

$$\begin{array}{cc}\hfill ActRes=& \langle ActResWeapon,ActResSensor\rangle \hfill \end{array}$$

(3)

$$\begin{array}{cc}\hfill ActS=& \langle PreActStatus,TarPreStatus\rangle \hfill \end{array}$$

(4)

$$\begin{array}{cc}\hfill ActE=& \langle TarEndStatus\rangle \hfill \end{array}$$

(5)

$$\begin{array}{cc}\hfill ActOutput=& \langle ActEffect,ActResWeaponDe,ActResSensorDe\rangle \hfill \end{array}$$

(6)

$$\begin{array}{cc}\hfill ActEva=& \langle ActTC,ActTZ,ActWT,ActWP,Act{P}_1,Act{P}_2\rangle \hfill \end{array}$$

(7)

$f\left(Act\right)$ denotes the description function of the operational activity, typically comprising the activity-effect function, duration function, and causal state function. These components are synthesized based on mission equipment performance and tactical rules during the COA generation process.

6.

The OV-6b State Transition Description Module. The OV-6b utilizes a graphical method to depict the evolution of mission states in response to diverse mission events, thereby clarifying the sequence of operational activities. As a complement to the OV-5b model, the OV-6b model constructed herein describes how the state transitions of the maritime formation anti-missile mission are driven by various mission events. Figure 8 illustrates the complete state transition process for the UAV-USV formation’s maritime defense and counterattack mission, delineating the entire sequence from target approach to target neutralization.

In summary, based on the aforementioned mission rule model and state transition model, the mission action elements and characteristics of the mission system are defined, thereby constructing a system mission view model that serves as a reference for the construction and analysis of maritime defense and counterattack system models.

3.4. Decision-Making Workflow Method Based on Modules Driven by LLM

3.4.1. RAG Enhanced Situational Understanding and Decision Generation

Based on the comprehensive definition and analysis of the DoDAF Operational Viewpoint (OV) models discussed previously. At the same time, LLMs possess strong logical reasoning and scenario adaptation capabilities. However, the mission environment in the specific vertical domain of UAV-USV maritime defense and counterattack is highly complex. Since the operational situation is replete with specialized, non-public terminology, specific mission doctrine, and strict command-and-control norms, LLMs often struggle to achieve accurate semantic alignment and are prone to domain-knowledge hallucinations. Therefore, this paper introduces the Retrieval-Augmented Generation method to construct an external knowledge enhancement framework [32]. This framework integrates relevant mission manuals and expert knowledge bases, linking them to DoDAF-structured data in real time. By constraining the LLM’s reasoning boundaries, the system facilitates the transition from general semantic understanding to professional mission decision-making, thereby significantly enhancing the model’s depth of understanding and the reliability of decision-making in complex cross-domain mission scenarios.

First, a Domain-specific Knowledge Base is constructed for the UAV-USV formation defense mission. This knowledge base is derived from multi-source data through document cleaning, preprocessing, and structuring, covering three core data categories: DoDAF architectural specifications, mission doctrines and rule sets, and equipment performance parameters. The DoDAF specification data defines the meta-model constraints and construction rules for each view model (such as the OV-1 High-Level Operational Concept Graphic and OV-2 Operational Resource Flow Description), ensuring that the generated models conform to standard architectural specifications. The mission doctrine base records the formation’s Rules of Engagement (ROE), coordinated flight/navigation specifications, and various contingency plans. The equipment performance parameter base contains key entity attributes such as radar detection range, blind zones, interceptor weapon range, and kill probability. The system utilizes a text embedding model to transform the aforementioned unstructured and semi-structured text into high-dimensional vectors, storing them in a vector database to form an external knowledge storage space for efficient retrieval.

Secondly, during the decision-making phase, a dynamic retrieval and context fusion mechanism based on semantic similarity is established. When the LLM receives a situational description or command instruction, the system first vectorizes the input text. It uses the Cosine Similarity algorithm to retrieve the k most relevant knowledge fragments from the vector database. The retrieval process can be formally represented as

$$\begin{array}{c}\hfill R=Retrieve(q,D,k)\end{array}$$

(8)

where q is the query vector, D is the domain knowledge base, and the $Retrieve$ function returns the set of the k most relevant document fragments $R=\{d_1,d_2,...,d_k\}$. Subsequently, through Prompt Engineering, the retrieved domain knowledge R is deeply fused with the original mission instruction x to construct an Augmented Prompt. This mechanism not only provides the LLM with real-time contextual information but also limits the model’s inference boundaries and output format through In-Context Learning. Based on this, the LLM performs inference, accurately identifying mission intent and mapping it to standardized DoDAF view elements. This effectively suppresses the “hallucination” phenomenon common in general large models within specialized domains, ensuring the tactical rationality and standardization of the generated mission decisions.

Table 7. Configuration details of the RAG system.

Embedding	Chunk Strategy	Indexing Design	Retrieval Depth k	Context Window Budget	Failure- Handling Rules
BGE-Large- En-v1.5	Chunk size: 512; Overlap: 64	FAISS	5	2048 tokens	Threshold < 0.65

shows the core hyperparameters and architecture configuration of the RAG module. By explicitly defining the chunking strategy, retrieval depth, and context budget, the system ensures that the prompt input to the LLM is stable and well-defined. Furthermore, the system incorporates a failure-handling rule: when the retrieved text’s similarity falls below a set threshold, the system will refuse to perform knowledge fusion. This mechanism effectively prevents the system from forcibly retrieving information when relevant tactical guidelines are lacking, thereby significantly reducing the risk of injecting irrelevant information or noise into the decision-making process.

3.4.2. Module-Based Model and Kill Chain Assessment

After achieving deep semantic understanding of the situation using RAG, the LLM serves as the core controller to drive external functional modules, compensating for large language models’ limitations in interpretability and complex logic verification. This process operates in a mode that coordinates different modules based on the LLM’s logical decisions. First, the LLM performs parametric parsing on the enhanced text output by RAG, extracting key information such as mission type, target characteristics, and available resource constraints, converting them into standardized API call instructions. The system then calls the Scheme Generation Module, which retrieves candidate paths suitable for the current situation from a predefined kill chain set (as shown in Table 6), based on the OV-5b Operational Activity decomposition tree and the OV-6a Operational Rules Model [33]. For example, regarding long-range high-speed targets, the module prioritizes matching cross-domain Cooperative Kill Chains that include UAEW cooperative detection. It then populates specific platform IDs and time window parameters, thereby instantiating the abstract mission intent into an executable Mission Course of Action (COA).

The generated kill chain plan must undergo rigorous static logic verification to ensure the mission sequence is self-consistent with tactical rules. To this end, the framework introduces a Petri net based Static Logic Verification mechanism. The system automatically converts the OV-5b operational activity sequence generated by the LLM into a Petri net (PN) model using a mapping algorithm, and then performs a full state-space search on the model using PN reachability analysis. The system focuses on detecting whether problems such as Deadlocks, Resource Conflicts, or Timing Violations exist within the mission process. For instance, if the verification detects that a USV node has executed a launch action without obtaining fire control authorization, the system captures this logical error. It generates a feedback prompt, driving the LLM to correct the plan until it passes the logical consistency check.

To achieve automated static verification, the structural elements of the DoDAF Operational Activity Model must be mathematically mapped to a Petri net. A formal Petri net is a directed bipartite graph, denoted by a tuple $PN=(P,T,F,W,M_0)$. The mapping protocol under this framework is shown in

Table 8. Mapping relationships between Petri net elements and OV-5b models.

Element	Description
Places (P)	Represents mission resources, preconditions, or specific operational states (e.g., “Target Detected”, “Guidance Handover Ready”).
Transitions (T)	Directly corresponds to the operational activities defined in OV-5b (e.g., “USV Target Tracking”, “Missile Launch”).
Directed Arcs (F)	Maps to the control flows and data flows in OV-5b, defining the causal sequence between states and actions.
Tokens (M)	Represents the dynamic instantiation of mission events or the availability of resources.

.

Based on this mapping, unstructured combat logic is transformed into a computable state-space model, thereby enabling static verification of the combat model.

After passing static verification, the candidate plan enters the Dynamic Effectiveness Evaluation stage to select the optimal decision using quantitative metrics. This paper constructs a multidimensional kill chain effectiveness evaluation metric system that comprehensively considers interception success rate, resource consumption cost, and time response speed. The evaluation module uses Monte Carlo simulation, combined with real-time performance metrics from various sensors and weapon systems, to predict and calculate the execution effects of each candidate kill chain. Finally, the system ranks all candidate COAs based on the calculation results. It outputs the mission plan with optimal effectiveness as the final decision-support recommendation, thereby realizing a complete closed loop from qualitative understanding of mission intent to quantitative effectiveness evaluation.

3.4.3. Comprehensive Decision-Making Report Generation

Upon the completion of logic verification and kill chain effectiveness evaluation, the framework proceeds to the final Decision Generation phase, as illustrated in Figure 9. At this stage, the core task of the LLM is to transform underlying simulation data and logical states into a comprehensive decision report endowed with tactical semantics.

Additionally, to optimize the commander’s interactive experience and reduce cognitive load in the mission environment, the system adopts an intuitive side-by-side view design on the terminal interface. One side of the interface presents the underlying DoDAF architecture view and Monte Carlo data, while the other side displays a natural-language decision report generated by an LLM. In the Reasoning Trace module of the report, the system automatically highlights key tactical indicators (such as the optimal kill chain effectiveness score and resource shortage warnings). After quickly reading and verifying the key logic, the commander can directly issue the optimal COA by clicking the Approve button on the interface; if tactical intervention is required, natural language commands (such as “switch to alternative 2”) can be entered directly in the dialog box, and the system will respond instantly and refresh the simulation results. This minimalist interactive design ensures the interpretability of the decision-making process while effectively implementing a human-centered supervision mechanism.

The process begins with a context fusion mechanism that maps the quantitative indicators output by the dynamic effectiveness evaluation module—such as capture probability ($Act{P}_1$) and communication duration ($ActTC$)—to the operational elements defined within the DoDAF views. The system-generated evaluation summary lists the performance parameters of candidate kill chains, providing commanders with an objective, quantitative basis for situational judgment regarding the robustness of different schemes in complex mission environments.

Subsequently, based on the ranking results of the comprehensive effectiveness function, the system explicitly presents an Optimal Kill Chain Recommendation in the report. To address the post hoc explainability issues inherent in traditional algorithmic decision-making, the report generated by this framework includes a comprehensive Reasoning Trace. This section clearly elucidates the rationale behind the recommendation, detailing how the system balanced trade-offs under multiple constraints. For instance, the system might state that “Although the theoretical hit probability of KillChain03 is marginally lower than that of KillChain05, it is determined to be the optimal solution for the current situation due to the superior anti-jamming capabilities of its composite guidance mode and a response time that aligns with the current interception window”. This interpretable reporting output mechanism bridges the cognitive gap between machines and human commanders, significantly improving the operational credibility of the decision support system [34].

Finally, the report generates prospective Tactical Suggestions. By integrating mission doctrines and historical cases from the RAG knowledge base, the LLM extends beyond mere firepower allocation to provide auxiliary guidance, including risk warnings and contingency plan preparations. These suggestions cover status alerts for key resources (e.g., “USV-02 unit ammunition reserves are critically low”) and the preparation of backup options (e.g., “It is recommended to pre-position KillChain07 as a supplementary engagement link”). The final comprehensive report is presented in structured text, allowing commanders to query details or adjust the Course of Action (COA) using natural language, thereby achieving an effective interface between data calculation results and actual command operations.

In summary, the design principles and implementation process of the LLM-driven UAV-USV formation defense modeling and decision generation framework are detailed above. Initially grounded in the DoDAF architecture framework, the study completes requirement capture, capability analysis, and the construction of the Operational Viewpoint (OV) for maritime defense and counterattack missions from a top-level perspective. The rigorous definition of the OV-5b Operational Activity Model and the OV-6a Operational Rules Model provides a standardized, structured description for cross-domain collaborative missions in complex battlefield environments.

Building on this foundation, the framework leverages RAG to support the LLM’s semantic understanding of specialized scenarios, effectively addressing the knowledge gaps and hallucination issues common to large models in specific vertical domains. Subsequently, based on an LLM-driven collaborative decision-making mechanism involving both large and small models, the framework integrates Petri-net-based Static Logic Verification and multi-dimensional Dynamic Effectiveness Evaluation modules to validate the views’ logic and assess the performance of the constructed kill chains.

Ultimately, the framework implements a closed-loop process encompassing unstructured situational understanding, parameterized kill-chain generation, logical consistency verification, and comprehensive decision report output. This significantly improves the scientific rigor and interpretability of the decision schemes, laying a solid theoretical and technical foundation for the specific scenario simulation and effectiveness verification to be conducted in the next chapter.

4. Experiment and Case Study

This section focuses on the comprehensive testing and verification of the established maritime formation defense and counterattack architecture model. Grounded in specific mission scenarios, the validation covers the complete architectural workflow, including LLM-based situational understanding, the collaboration between static and dynamic logic analysis modules, and decision report generation [35,36].

4.1. Mission Scenario

To verify the generalization ability and robustness of this framework in complex, unpredictable battlefield environments and to overcome the limitations of a single fixed scenario, this section introduces a constrained randomized scenario-generation mechanism. By simulating the dynamic fluctuations in battlefield resources and enemy threats, we conduct a comprehensive statistical evaluation of the system’s decision-making effectiveness under conditions of uncertainty. In cross-domain collaborative missions, available resources and target strike scale are often highly uncertain. To avoid generating invalid scenarios that violate tactical logic during randomization, this framework introduces constraints during randomization generation. Based on the tactical settings, the generation of the blue force threat parameters follows these constraints:

$$\begin{array}{cc}\hfill M_{Total}& =N_{UCAV}\times P_{unit}\hfill \end{array}$$

(9)

$$\begin{array}{cc}\hfill W_{attack}& =P_{unit}\hfill \end{array}$$

(10)

where $M_{Total}$ represents the total number of incoming missiles, $N_{UCAV}$ represents the total number of Blue Force UCAVs, $P_{unit}$ represents the missile load per unit, and $W_{attack}$ represents the number of attack waves. The system first randomly generates basic variables within a reasonable range and then mathematically derives the relevant dependent variables. The specific generation boundaries and coupling rules are shown in

Table 9. Generation rules for combat scenario entities.

Faction	Entity Object	Value Range	Generation Rules
Blue Force	Number of UCAV formations	$[1,4]$	Total combat aircraft: $N_{UCAV}=F_{blue}\times U_{blue}$
	UCAVs per formation	$[2,6]$	Attack waves: $W_{attack}=P_{blue}$
	Missile payload per UCAV	$[2,8]$	Total incoming missiles: $M_{Total}=N_{UCAV}\times P_{blue}$
Red Force	Number of USVs	$[2,6]$	Interception resource matching constraint:
	Number of Launch & Recovery Centers	$[1,2]$	Total interceptor capacity (USV + UCAV payload) must satisfy:
	Number of UCAV formations	$[1,4]$	$Capacit{y}_{red}\ge M_{Total}\times \alpha$
	UCAVs per formation	$[2,6]$	To ensure meaningful confrontation, set threshold $\alpha \in [0.8,1.2]$.
	Number of UAEWs	$[1,3]$	Eliminate extreme invalid scenarios.

.

Under this mechanism, the LLM-driven decision-making center dynamically maps underlying entities to kill chain templates based on randomly generated available assets in each scenario, thereby generating a COA in real time under uncertainty conditions.

4.2. LLM Situational Understanding and View Modeling

The experiment was initiated by inputting the mission scenario designed in Section 4.1 into the system in natural language. In response to the descriptions of “multi-wave,” “saturation attack,” and various heterogeneous platforms in the text, the system activated a situational understanding process using RAG.

The system first vectorized the input text to retrieve relevant entries from the domain knowledge base regarding “UAV-USV formation coordinated air defense”, “wave-by-wave interception tactics”, and the equipment performance of both Red and Blue forces. Subsequently, utilizing prompt templates, the system fused this domain knowledge with key information from the original scenario to construct an enhanced prompt for the LLM. The LLM performed parametric extraction on the input, successfully transforming the unstructured battlefield description into structured situational data. It explicitly identified the Blue Force’s intent as “a continuous saturation attack of XX missiles across X waves”. It catalogued the Red Force’s available defensive resources, achieving precise semantic alignment of the battlefield situation.

Based on this structured situational understanding, the LLM then drove external modeling tools to construct the view. First, based on the extracted Red Force composition (including UAV Launch and Recovery Platforms, USVs, and various aerial formations), the LLM invoked the view modeling module interface to automatically generate the OV-1 High-Level Operational Concept Graphic, accurately mapping the physical deployment and command relationships of each combat node.

Furthermore, addressing the temporal characteristics of the Blue Force’s “wave-by-wave attack,” the LLM guided the generation of the OV-5b Operational Activity Model. This model structured defensive operations into four iterative “detection-allocation-interception” closed loops, each corresponding to a specific attack wave. The generated model’s logic showed that Red Force defensive resources were optimally allocated across the four time windows, ensuring interception coverage for all XX incoming targets. This validated the LLM’s capability to transform complex tactical temporal logic into standard models compliant with DoDAF specifications.

4.3. Evaluation of Model Generation Accuracy

To quantitatively verify the performance of the LLM-driven framework for transforming natural-language tactical instructions into standardized DoDAF view models, this section designs a comparative experiment. A mission set was constructed, comprising 20 tactical instructions of varying complexity. This set covers three scenarios: “single-instruction tasks”, “multi-wave complex instructions”, and “multi-party collaborative missions”.

Qwen3-8B was selected as the core inference engine for the framework. It was deployed via the Alibaba Cloud Model Studio Bailian platform, with experiments conducted locally via API calls. Compared to Qwen2 and Qwen2.5, the Qwen3-8B model demonstrates significant improvements in instruction following, long-context generation, understanding structured data (e.g., tables), and generating structured outputs from natural language.

We compared the proposed method against an unoptimized general large language model. The evaluation metrics cover the following three dimensions:

Entity Accuracy: This metric evaluation framework accurately extracts and maps the capabilities of all combat units (such as USVs, UCAVs, and UAEWs) from unstructured instructions. Its formal definition is the ratio of correctly identified entities to the total number of entities in the scenario. The formula is as follows:

$$\begin{array}{c}\hfill Ac{c}_{ent}={\displaystyle \frac{N_{correct}}{N_{total}}}\times 100\%\end{array}$$

(11)

where $N_{correct}$ is the number of entities correctly identified without hallucination, and $N_{total}$ represents the total number of entities explicitly defined in the task instructions.

Tactical Logic Validity: This metric assesses whether the generated OV-5b operational sequence strictly conforms to tactical doctrine and timing constraints. Instead of relying on manual annotation, we utilize a Petri net-based static verification module for objective evaluation. A sequence is considered valid only if it successfully passes all reachability and deadlock-free checks. The formula is as follows:

$$\begin{array}{c}\hfill TLV={\displaystyle \frac{1}{M}}\sum _{i=1}^{M}L(\mathrm{P}{\mathrm{N}}_{\mathrm{Verify}}\left(S_i\right)=\mathrm{True})\times 100\%\end{array}$$

(12)

where M is the total number of generated tactical instructions, $S_i$ is the generated mission sequence, and the indicator function $L(·)$ returns 1 when sequence $S_i$ passes Petri net verification ($\mathrm{PN}\_\mathrm{Verify}$), otherwise it returns 0. This metric ensures that the generated intermediate data can be seamlessly integrated into downstream simulation engines.

Compliance is not checked manually; an automated script rigorously verifies it. This script parses the LLM output against a predefined JSON schema, confirming the presence of all required DoDAF attributes, proper label closure, and the absence of phantom data fields.

4.3.1. Comparative Evaluation of Tactical Instruction Generation

In this experiment, Ours refers to the RAG-enhanced LLM with Constraints proposed in this paper, and Qwen-8b is a general model without a RAG knowledge base and format constraints. Experimental results are shown in

Table 10. Comparative Experimental Results on Tactical Instruction Generation.

Scenario Type	Method	Entity Accuracy (%)	Logic Validity (%)	Format Compliance (%)
Simple Instruction	Qwen-8b	94.5	86.8	81.2
	Ours	99.5	99.0	100.0
Complex Multi-wave	Qwen-8b	78.2	64.5	58.4
	Ours	98.0	95.5	100.0
Multi-party Engagement	Qwen-8b	54.6	38.2	45.5
	Ours	92.4	89.2	100.0
Average	Qwen-8b	75.8	63.2	61.7
	Ours	96.6	94.6	100.0

.

As shown in Table 10, the proposed framework exhibits significant performance advantages in all test scenarios. While the unoptimized baseline model performs reasonably well on simple single-instruction tasks, achieving an entity accuracy of 94.5%, its performance drops sharply as tactical complexity increases. Particularly in multi-party engagement scenarios, the baseline model suffers from severe forgetting when handling long contexts, frequently confusing entity IDs between red and blue parties or ignoring temporal dependencies, resulting in entity accuracy and logical validity dropping to 54.6% and 38.2%, respectively. Conversely, Ours maintains high robustness in complex multi-node, multi-wave interactions, not only maintaining logical validity at 89.2% in complex scenarios but also achieving an overall average entity accuracy of 96.6% and logical validity of 94.6%.

Furthermore, regarding format compliance, which determines system engineering usability, the baseline model, due to frequent generation of unclosed labels or hallucinations, has an average compliance rate of only 61.7%, making it completely unparseable by subsequent simulation engines. This framework consistently maintained 100% format compliance across various complex scenarios, achieving zero-error integration with the underlying simulation system.

In summary, this framework demonstrated comprehensive advantages in handling highly complex cross-domain collaborative instructions. To further analyze and quantify the specific contributions of each core component to these performance improvements, the next section will conduct targeted ablation experiments.

4.3.2. Ablation Study and Component Analysis

To verify the independent contributions and synergistic effects of each core component in this framework, and to rigorously explain the performance improvement compared to the baseline model, we conducted ablation experiments, the results of which are shown in

Table 11. Ablation study of different components on model performance.

Method	Entity Accuracy (%)	Logic Validity (%)	Format Compliance (%)
Baseline (Qwen3-8B)	75.8	63.2	61.7
Baseline + RAG	92.4	75.4	63.5
Baseline + Schema	74.1	62.8	100.0
Baseline + Petri net	76.2	88.5	62.0
Baseline + RAG + Schema	93.1	78.5	100.0
Baseline + RAG + Petri net	94.5	92.0	64.1
Baseline + RAG + Schema + Petri net	96.6	94.6	100.0

.

The baseline model, Qwen3-8B, often exhibits domain hallucination when handling complex professional instructions and struggles to consistently output standard formats, achieving entity accuracy of only 75.8%, logical validity of only 63.2%, and format compliance of only 61.7%. Experiments show that each introduced component specifically addresses specific engineering pain points: the RAG module alone provides the model with accurate domain knowledge and injects tactical rules, effectively mitigating the hallucination phenomenon and significantly improving entity accuracy to 92.4%; the schema constraint alone enforces output standardization during the output phase, standardizing the consistency between natural language and machine-readable simulation data, ensuring 100% format compliance; and the Petri net static verification module alone, with its deterministic deadlock and conflict feedback mechanism, actively forces the model to correct tactical defects, significantly increasing logical validity to 88.5%.

Furthermore, the complete framework demonstrates a high degree of synergy and complementarity. For example, the logical verification of Petri nets must be based on the standard-formatted data output generated by the schema constraints. Ultimately, the complete architecture integrating all components achieves globally optimal performance, fully demonstrating the necessity and scientific validity of the complete closed-loop design of extraction, constraint, and verification presented in this paper.

4.4. Static Analysis of Petri Net-Based Models

Based on the constructed view models, the LLM invokes the static verification module to assess their logical validity. Leveraging the Petri net framework, this module establishes a formal mapping centered on the OV-5b Operational Activity Model—which encapsulates complex logical dependencies—to verify the model’s logical structure. The Petri net model derived from the OV-5b model is illustrated in Figure 10.

The module performs a rationality verification on the transformed Petri net model. By analyzing the state transition characteristics of each combat node derived from the diagram, the following conclusions are drawn:

• Each Petri net model is safe and bounded;
• Every marking (state) in the Petri net model is reachable;
• Each Petri net model is live and deadlock-free.

Table 12. Static simulation results of maritime formation defense system.

Model Verification Metrics	Conflict Verification	Cycle Verification	Deadlock Verification
Verification Result	Interpretable conflicts exist	No cycles	No deadlocks

shows Static simulation results of maritime formation defense system. The diagram indicates the existence of conflicts within the Petri net model; specifically, after “Target Detected”, a single place simultaneously triggers two transitions. Similar conflicts occur regarding the “Dissemination of Combat Action Plan” and “Target Indication”. However, these are correlated logical conflicts (non-destructive). Upon receiving information such as detection results, the Command Center enters the combat command state, while the formation simultaneously enters the target tracking state, ensuring no transition omissions. The states of “Target Detected”, “COA Dissemination”, and “Target Indication” possess sufficient information resources to support the triggering of multiple transitions. At this stage, the LLM performs natural-language editing based on results from specialized small models. It outputs them in the final report, allowing the commander at the Command Center to make subjective judgments to resolve the situation. Based on the internal analysis described above, the module returns the results to the LLM, which then generates the natural language output and invokes the next module.

The static simulation experiments demonstrate that although the proposed LLM-driven maritime formation defense and counterattack system involves diverse architectural views, complex operational activities, and heterogeneous information flows, the system architecture exhibits structural cohesion and optimal resource utilisation. The identified local conflicts are interpretable and resolvable, and the information dissemination process is fluid. Crucially, the model is proven to be deadlock-free, thereby satisfying the validity principles of system architecture modeling.

Furthermore, to evaluate the computational feasibility of the framework for handling complex cross-domain collaborative scenarios, we selected the extreme boundary scenario with the highest concurrency from 100 random scenarios, namely a 6-wave saturation attack by 64 missiles from the Blue Force, and performed a state-space quantization analysis. The Petri net model instantiated for this extreme scenario contains 245 places and 188 transitions. The depth-first search of the reachable graph generated 45,210 reachable markers and 112,450 directed edges. The complete traversal of the entire state space and deadlock detection took only 1.2 s on an Intel i7 processor. This extreme test strongly demonstrates that within the set tactical boundaries, the Petri net model generated by this system effectively avoids the “state explosion” problem, ensuring that each dynamically generated COA of the LLM can be rigorously and in real-time formally verified when facing highly uncertain and complex battlefield inputs.

4.5. Dynamic Model Testing

The dynamic verification process comprises two primary components. The first component focuses on verifying the logical-temporal consistency of operational events based on the OV-5b and OV-6b models. The second component involves a comparative effectiveness analysis between cooperative and non-cooperative kill chains. To execute these tasks, the LLM invokes the corresponding modules to perform targeted verification.

4.5.1. Temporal Consistency Analysis

To verify the logical-temporal consistency of operational events, the module first uses the Systems Modeling Language (SysML) to construct a state machine diagram for the OV-6b model. Subsequently, based on the logical framework of the OV-5b model established in the preceding view modeling module, it constructs a sequence diagram to describe the model’s behavior formally. After constructing the SysML views, the module uses its internal simulator to perform model initialization, generate executable software code, and run dynamic simulations. Finally, a comparative analysis is conducted between the operational flow derived from the OV-6b model and the logic defined in the OV-5b model.

Figure 11 presents the event sequence diagram generated by the module based on the OV-6b model. A comparative analysis was conducted against the OV-5b modeling data previously ingested by the view modeling module. The results are as follows: Interaction Consistency: The content and sequence of key interactions in Figure 11 specifically detection, command dissemination, and missile guidance—align perfectly with the operational flows depicted in Figure 7. Command Logic: In Figure 11, the FOC functions as the central command node, generating and disseminating directives. The USV, UCAV, UAEW, and UEA formations execute combat missions after exchanging information with the platform. This overall logic and temporal sequencing are consistent with Figure 7. Guidance Handover: Figure 11 depicts the USV formation interacting with the UAEW formations to execute the guidance authority handover. Similarly, the UCAV formation interacts with the UAEW formation to transfer guidance control. These temporal and logical sequences mirror those illustrated in Figure 7. EW Operations: In Figure 11, the UEA executes detection and electronic jamming tasks. It interacts with the Launch and Recovery Platform once to receive mission assignments, maintaining temporal and logical consistency with Figure 7. Based on the aforementioned analysis, the constructed OV-6b simulation sequence diagram demonstrates strict logical and temporal consistency with the OV-5b model. The module returns these analytical results to the LLM, which then synthesises the data and generates the final output in natural language.

4.5.2. Kill Chain Synergy Advantage Assessment

After confirmation of logical-temporal consistency, the LLM-driven module performs the kill chain effectiveness evaluation. Grounded in the operational scenario, the module performs a comparative analysis of various metrics between cooperative and non-cooperative kill chains [37]. The constituent parameters of the kill chain metrics are presented in

Table 13. Constituent parameters of kill chain metrics.

Parameter	Compliance Delay	Other Delay	Timeliness	Accuracy	Acquisition	Damage	Weight
Notation	$ActTC$	$ActTZ$	$ActWT$	$ActWP$	$Act{P}_1$	$Act{P}_2$	$w_1,w_2,w_3$

, where the calculation formula for the Weighted Score ($WS$) is defined as follows:

$$\begin{array}{c}\hfill WS=\left(1-{\left({\displaystyle \frac{ActTC+ActTZ}{ActWT}}\right)}^4\right)\times w_l+ActWP\times w_2+Act{P}_1\times Act{P}_2\times w_3\end{array}$$

(13)

where $w_l$, $w_2$, and $w_3$ are set by the commander based on the actual situation, and their values range from 0 to 1, with $w_1+w_2+w_3=1$. The weights reflect the relative importance attached to time, accuracy, and operational longitude under different combat missions. To eliminate the bias that may be introduced by this subjective weight assignment and to verify the reliability of the decision-making model, this paper will conduct a rigorous sensitivity analysis of the weight set in Section 4.7.

To quantify the performance of different decision-making models under uncertainty, 2000 Monte Carlo simulation experiments were conducted for each of the 100 randomly generated scenarios mentioned above within the module. The experimental results are shown in

Table 14. Performance comparison of different kill chain modes.

Kill Chain Type	Average $\mathit{WS}$	$\mathit{WS}$ Standard Deviation	Average Compliance Delay (ms)	${\mathit{ActP}}_1$
Non-cooperative Mode	0.618	0.045	8.42	0.63
Cooperative Mode	0.752	0.018	6.98	0.84

. Kill chains 01 and 02 are non-cooperative, while kill chains 03–10 are cooperative, with specific links shown in Table 6. Analysis of the table data shows that, without optimizing the parameters of cooperative kill chains, their kill-chain metric is significantly better than that of non-cooperative kill chains, demonstrating the superiority of cooperative kill chains in maritime air defense scenarios.

Data show that, under varying battlefield constraints, the cooperative kill chain generated by this framework not only significantly outperforms the traditional non-cooperative mode in terms of average weighted score, but also has a much lower standard deviation in effectiveness. This demonstrates that traditional single-domain kill chains are highly susceptible to fluctuations in the battlefield environment. At the same time, the LLM-driven cross-domain cooperative decision-making framework can effectively mitigate the negative impact of battlefield uncertainty by providing cross-domain guidance and adaptive resource scheduling via schemes that dynamically introduce UAEW, thereby demonstrating superior system robustness.

To further verify the defensive reliability of this framework under extreme conditions, we statistically analyzed the overall target interception rate across 100 randomly generated adversarial scenarios. Given a sufficiently large sample and the approximate normality of the sample mean, we calculated a 90% confidence interval (CI) to estimate the lower bound of the system’s operational effectiveness, in accordance with statistical standards. The statistical results are shown in

Table 15. Comparison of interception rate statistics between different kill chain modes.

Kill Chain Type	Average Interception Rate	Interception Rate Range	90% Confidence Interval (CI)	Effectiveness Degradation Rate
Non-cooperative	0.625	0.438–0.705	[0.608, 0.642]	15.3%
Cooperative	0.788	0.712–0.855	[0.776, 0.800]	4.1%

.

The effectiveness decay rate in the table refers to the decrease in interception rate relative to the lowest baseline strike under extreme saturation attacks. In the test set covering multi-scale uncertainty attacks, the average interception rate of the cooperative kill chain reached 0.788. More statistically significant, at a 90% confidence level, the lower bound of the interception rate confidence interval for the cooperative mode is still significantly higher than the upper bound of the non-cooperative mode. Furthermore, when facing extreme saturation attacks, the interception effectiveness of the non-cooperative mode showed a significant decay, while the cooperative mode driven by this framework only decayed by 4.1%.

Upon completion of the effectiveness evaluation, the resulting metric data is fed back to the LLM to facilitate kill chain comparison and selection. Subsequently, leveraging the data returned by the module, the LLM performs a comprehensive assessment to formulate traceable inferences, which are then synthesised and output in natural language.

4.6. Model Completeness Check

After dynamic verification, the LLM invokes the module to perform a completeness check of the architecture model, using the DoDAF architectural data as the baseline. As an intermediary, the LLM retrieves data meta-models relevant to the DoDAF model’s completeness requirements—specifically Operational Activities, Operational Information, and Performers—from the view modeling module. To facilitate this verification, completeness rules are formulated based on meta-model relationships. Specifically, the rules mandate that every Operational Activity must have at least one instance and be assigned to at least one Performer. Furthermore, information exchange must include both the transmission and reception ends (i.e., a closed loop). In the context of the mapping matrix, this manifests as row and column completeness, implying the absence of null rows or columns. For data completeness verification, a binary association matrix is constructed. Taking the relationship between Operational Activities and Performers as an example, Matrix $\mathbf{A}$ is established to quantify the associations between data entities, where the element values are defined as follows:

$$\begin{array}{c}\hfill a_{ij}=\left\{\begin{array}{cc}1,\hfill & \mathrm{if}{b}_i\mathrm{is}\mathrm{associated}\mathrm{with}{c}_j\hfill \\ 0,\hfill & \mathrm{otherwise}\hfill \end{array}\right.\end{array}$$

(14)

where $b_i$ and $c_j$ denote elements belonging to the Operational Activity data set B and the Performer data set C, respectively (i.e., $b_i\in B,c_j\in C$). The required association relationships between Operational Activities and Performers are presented in

Table 16. Association Relationships between Operational Activities and Performers.

Operational Activity	Performer
Issuance of Defense Alert Mission	UAV Launch and Recovery Platform
Target Detection	USV Formation, UCAV Formation, UAEW Formation, UEA Formation
Target Tracking	USV Formation, UCAV Formation, UAEW Formation
Generation and Dissemination of COA	UAV Launch and Recovery Platform
Electronic Jamming	UEA Formation
Cooperative Missile Guidance Task 1	UCAV Formation, UAEW Formation
Cooperative Missile Guidance Task 2	UCAV Formation, UAEW Formation
Cooperative Missile Guidance Task 3	USV Formation
Target Neutralization	USV Formation, UCAV Formation, UAEW Formation, UEA Formation

.

The overlap matrix is as follows:

$$\begin{array}{c}\hfill A=\left[\begin{array}{ccccc}1& 0& 0& 0& 0\\ 0& 1& 1& 1& 1\\ 0& 1& 1& 1& 0\\ 0& 1& 1& 1& 0\\ 1& 0& 0& 0& 0\\ 0& 0& 0& 0& 1\\ 0& 0& 1& 1& 0\\ 0& 0& 1& 1& 0\\ 0& 1& 0& 0& 0\\ 0& 1& 1& 1& 1\end{array}\right]\end{array}$$

(15)

The rows, arranged from top to bottom, represent the following: Issuance of Defense Alert Mission, Target Detection, Target Tracking, Generation and Dissemination of COA, Electronic Jamming, Cooperative Missile Guidance Task 1, Cooperative Missile Guidance Task 2, Cooperative Missile Guidance Task 3, and Target Neutralization. The columns, arranged from left to right, represent the following: UAV Launch and Recovery Platform, USV Formation, UCAV Formation, UAEW Formation, and UEA Formation.

Based on the completeness rules, the binary association matrix in Equation (11) is verified. Let $M_i$ and $N_j$ denote the count of ones in the i-th row and j-th column, respectively. The analysis yields the following: Row Completeness Verification:

$$\begin{array}{c}\hfill \left\{\begin{array}{c}{M}_1=M_5=M_6=M_9=1\hfill \\ M_7=M_8=2\hfill \\ M_3=M_4=3\hfill \\ M_2=M_{10}=4\hfill \end{array}\right.\end{array}$$

(16)

The values $M_1$ through $M_{10}$ are all greater than or equal to 1, thereby satisfying the row completeness criterion.

Column Completeness Verification:

$$\begin{array}{c}\hfill \left\{\begin{array}{c}{N}_1=2\hfill \\ N_5=3\hfill \\ N_2=5\hfill \\ N_3=N_4=6\hfill \end{array}\right.\end{array}$$

(17)

Regarding the column completeness verification, values $N_1$ through $N_5$ are all greater than 1, satisfying the column completeness criterion. Consequently, the binary association matrix A, constructed based on the Operational Activities and Performers of the maritime formation defense and counterattack mission, complies with the completeness rules. The verification logic for the remaining datasets follows the same principle, and all satisfy the completeness requirements. Finally, the module transmits the verification results back to the LLM, which interprets the data and generates the natural language output.

Through the implementation of the aforementioned comprehensive workflow, the LLM aggregates RAG-based common mission decision rules, situation-driven DoDAF view modeling data, static and dynamic model verification results, kill chain effectiveness evaluation metrics, and model completeness verification outcomes.

In summary, the experimental results in the above sections demonstrate that while this framework significantly improves the quality of cross-domain collaborative decision-making, its “modeling-verification-evaluation” closed loop inevitably introduces a certain computational latency, with a single end-to-end time of approximately 12 s. However, this system is positioned for formation-level command and control (C2) and decision support. In a typical cross-domain air defense early warning phase, commanders usually have a tactical decision-making window of tens of seconds to several minutes. Therefore, the aforementioned computational overhead is within the timing requirements of a C2 system. The framework delivers high-quality decision-making solutions with lower latency, demonstrating strong engineering feasibility.

4.7. Decision Making Under Uncertainty and System Robustness

In complex cross-domain collaborative missions, the system faces not only the objective uncertainty of battlefield environment fluctuations, but also the cognitive uncertainty of commanders’ subjective tactical preferences, as well as the potential underlying failure risk of the LLM-driven architecture. To comprehensively evaluate the effectiveness and reliability of this framework in mission-support contexts, this section will conduct an in-depth analysis of three aspects: parameter sensitivity, statistical convergence, and system failure modes.

4.7.1. Sensitivity Analysis of Decision Weights

In the integrated effectiveness evaluation model, the baseline weight allocation is based on the current naval air defense tactical doctrine, with the highest value assigned to the time-related weights. To verify the robustness of the decision model to subjective parameters, we conducted a weight sensitivity test, setting two tactical preferences: high-precision weight ($w_2=0.5$) and high-damage weight ($w_3=0.5$). The test results show that although drastic fluctuations in weights caused changes in the effectiveness evaluation scores for each kill chain, the overall ranking of the cross-domain cooperative kill chain consistently remained first across all three preferences, and the score difference relative to the optimal single-domain non-cooperative kill chain remained above 12%. This demonstrates that the cross-domain cooperative scheme recommended in this framework is robust to interference from subjective weight settings.

4.7.2. Monte Carlo Convergence and Uncertainty Intervals

To eliminate random noise in the dynamic evaluation, we conducted convergence tracking across 2000 Monte Carlo iterations to assess the reasonableness. The results show that when the number of iterations reaches 1500, the variance of the interception rate converges to ${10}^{-4}$, ensuring that the simulation results are in an absolute statistical steady state. Across 100 random scenarios, we calculated the 90% confidence interval (CI) for interception effectiveness. The CI for the cooperative mode is [0.776, 0.800], while that for the non-cooperative mode is [0.608, 0.642]. After accounting for underlying uncertainty perturbations, the lower limit of effectiveness for the cooperative mode remains significantly higher than the upper limit for the non-cooperative mode, providing statistical validation of our method’s effectiveness.

4.7.3. Failure Modes and Defense Mechanisms

To address the inherent vulnerabilities of LLM systems that rely on prompt engineering and external module orchestration in mission-critical environments, this framework incorporates multiple defense and error correction mechanisms at the architectural design level to handle the following four typical potential failure modes:

1.

Injection via corrupted sources. As a decision-support framework focused on specific domain, the DoDAF specification and the tactical doctrine database in RAG are physically isolated in offline, read-only storage. External situational input is strictly limited to parameterized standard data streams, effectively blocking malicious prompt injection and knowledge source pollution from the physical architecture.

2.

Retrieval errors. The RAG module often retrieves mismatched tactical orders due to semantic offsets, leading the LLM to generate incorrect action mappings. However, this framework does not directly use the RAG output as the execution command; instead, it uses it as input to generate the OV-5b model. The model logic is verified using a Petri net verification module. Any behavior that violates tactical timing due to retrieval errors will trigger a deadlock or conflict error during the static verification phase, forcing the LLM to re-infer.

3.

Schema-conforming but semantically wrong. During the experimental phase, we discovered that LLM can generate fake kill chains that perfectly conform to the JSON Schema but are tactically absurd. Static format validation cannot intercept such special hallucinations. At this point, the dynamic performance evaluation module serves as a filter. Such semantically flawed fake kill chains will be assigned extremely low anomalous performance scores ($WS$) in the underlying simulation engine due to weapon mismatch or an extremely low probability of damage. They will be automatically eliminated from the scheme ranking and will not appear in the final recommendation.

4.

Over-trust in generated recommendations. To prevent commanders from unquestioningly trusting system outputs under Automation Bias, this framework strictly adheres to a Human-in-the-loop design, as seen in the Commander Confirmed node in OV-6b. The system output is not automatically executed code, but rather a decision report containing a detailed reasoning trace. The report not only provides quantitative rankings but also explicitly lists resource consumption risks and alternative schemes, forcing commanders to obtain final human approval based on the chain of evidence.

5. Conclusions

This study systematically investigates the modeling and decision-making challenges of UAV-USV cross-domain formations by establishing a comprehensive framework that fuses the DoDAF architecture with LLM-driven intelligence. We constructed a rigorous top-level system design based on DoDAF views to standardise complex collaborative missions. We introduced a RAG-enhanced LLM mechanism that successfully bridges the gap between natural language understanding and professional tactical modeling, effectively mitigating domain-specific hallucinations. Through the integration of a Petri net-based static analysis module and a dynamic effectiveness evaluation system, the framework ensures that the generated Courses of Action (COAs) are not only logically consistent and deadlock-free but also tactically superior, as evidenced by a substantial increase in the average interception rate from 0.625 to 0.788 in cooperative scenarios. Ultimately, this research distinguishes and achieves both the intrinsic interpretability of the formal modeling pipeline and the post hoc explainability of the decision outputs. By providing commanders with an explainable decision report containing quantitative effectiveness rankings and natural-language reasoning traces, the framework creates a highly credible, closed-loop decision support system that significantly enhances the scientific rigour of tactical choices.

6. Future Work

While the framework proposed in this paper has been effectively validated in tactical-level maritime cross-domain collaborative air defense scenarios, further refinement in scalability and scenario generalization is needed to apply it to more complex mission scenarios. Future research will focus primarily on the following two directions:

Firstly, addressing the scalability challenges for theater-level operations. As the scale of collaborative assets grows to hundreds, centralized processing methods face dual bottlenecks: LLM context overflow and Petri net static validation state explosion. To address this, we plan to introduce a hierarchical multi-agent framework. The system will employ a top-level command agent to perform global task decomposition, and distributed tactical agents will construct local DoDAF sub-views in parallel. Simultaneously, combined with modular Petri Nets (MPNs), validated local views will be encapsulated as macro-transitions, thereby reducing the dimensionality of the state space while achieving large-scale, efficient, and parallel validation.

Secondly, extending the framework’s universality of scenarios. Current assessments primarily focus on air and sea defense scenarios. Future work will extend this large-scale model-driven systems engineering framework to other typical mission environments. By conducting generalization tests under more diverse resource constraints and tactical doctrines, the robustness and effectiveness of the decision support model across different cross-domain mission systems will be further improved.