A method of generating attack information based on multi-dimensional fusion

Abstract

The present invention relates to a method for generating strike information based on multi-element fusion, which belongs to the technical field of strike information generation, and solves the problems of complex user operation process, poor user experience, and poor scalability of strike assessment software in the prior art. The present invention integrates a strike assessment calculation model, a knowledge graph model, a pre-trained large language model, a knowledge base model, and an agent model based on a large language model development framework, performs intent recognition on user input text through a pre-trained large language model to obtain a functional mode that matches the intent, and obtains a return result of the running functional mode, generates a prompt word based on the return result, and inputs the prompt word into the pre-trained large language model to generate strike information. The present invention provides a strike information generation method that is simple to operate, highly accurate, and highly interpretable.

Description

Technical Field

[0001] The present invention relates to the technical field of attack information generation, and in particular to an attack information generation method based on multi-element fusion. Background Art

[0002] In the field of strike assessment, how to obtain strike information accurately and quickly is one of the key research issues. The existing acquisition process is to build a calculation model based on a database system composed of experimental data, simulation data and intelligence data combined with strike criteria and strike judgments. Users calculate the strike effect based on the target vulnerability analysis model and the power field calculation model, and then further use the existing software model to calculate the strike effect, and finally derive the final firepower strike plan based on the strike effect.

[0003] Although the existing strike assessment calculation software is relatively mature, and the required solution can be finally obtained through certain calculation steps, there are still two problems: First, the user threshold is high. Users need to accurately understand the content and operation of each calculation model, which requires rich prior knowledge. Users need to be proficient in target vulnerability knowledge and have certain experience in strategic plan formulation, or exhaustively calculate and analyze all attack parameters in the database, which will take a long time. Finally, the data of target vulnerability analysis and power field analysis are combined to analyze and calculate the best matching solution. The results need to be further sorted out to present a more intuitive and easy-to-understand solution form, and if there is not much experience, the number of trial and error will be relatively large. Second, the scalability of the solution is not strong. The strike information solution obtained by calculating for several specific types of targets is difficult to migrate to other targets. When the firepower composition changes, the above process needs to be fully executed again to obtain a new strike information solution. Summary of the invention

[0004] In view of the above analysis, an embodiment of the present invention aims to provide a method for generating strike information based on multi-element fusion, so as to solve the problems of complex user operation process, poor user experience and poor scalability of strike assessment software in the prior art.

[0005] The embodiment of the present invention provides a method for generating strike information based on multi-element fusion, comprising:

[0006] S1: constructing a large language model development framework, integrating a strike assessment calculation model, a knowledge graph model, a pre-trained large language model, a knowledge base model, and an agent model in the large language model development framework, thereby obtaining a strike information generation model;

[0007] S2: The attack information generation model receives the user input text, calls the pre-trained large language model to perform intent recognition on the user input text to obtain a functional mode that matches the intent, and executes the functional mode to obtain a return result. The functional mode includes: (1) agent model call mode; (2) knowledge base model query mode; (3) pre-trained large language model answer mode;

[0008] S3: The strike information generation model generates a prompt word based on the return result of the functional mode, and inputs the prompt word into the pre-trained large language model to generate strike information.

[0009] Based on the further improvement of the above method, the strike assessment calculation model includes a target vulnerability analysis model, a force field analysis model, a strike effect analysis model and a strike effect assessment model; the knowledge graph model includes entities and relationships between entities, and the entities include targets, attack parameters and attack platforms.

[0010] Based on the further improvement of the above method, the pre-trained large language model is obtained by fine-tuning the large language model based on the P-Tuning 2 algorithm or the Lora algorithm.

[0011] Based on the further improvement of the above method, when the functional mode is the intelligent model calling mode, the strike information generation model calls the intelligent model to generate a return result; when the functional mode is the knowledge base model query mode, the strike information generation model calls the knowledge base model to generate a return result; when the functional mode is the pre-trained large language model answering mode, the strike information generation model calls the pre-trained large language model to generate a return result.

[0012] Based on the further improvement of the above method, the strike information generation model calls the intelligent agent model to generate a return result, including: the intelligent agent model uses the Reasoning and Acting framework to control the execution order of each model in the strike assessment calculation model, and generates a return result based on the calculation results of each model.

[0013] Based on the further improvement of the above method, the return result generated based on the calculation results of each model includes: A1: obtaining the strike target, the intelligent agent model calls the target vulnerability analysis model to obtain the strike tree corresponding to the strike target; A2: the intelligent agent model calls the knowledge graph model to query and obtain the attack platform suitable for the strike target and its corresponding attack parameter data, the intelligent agent model calls the power field analysis model to perform power field analysis on the attack platform and the attack parameter data to obtain the available attack platform and its corresponding available attack parameters that meet the user's intention, and generates a combat plan based on the available attack platform and its corresponding available attack parameters; A3: the intelligent agent model calls the strike effect analysis model, and performs a power field analysis on the combat plan based on the strike tree. A strike effect analysis is performed to obtain the strike probability, strike time and strike cost corresponding to the combat plan; A4: the intelligent body model calls the strike effect evaluation model to determine whether the combat plan is effective. If the strike probability, the strike time and the strike cost all meet the preset conditions, the strike evaluation result is set to be valid, otherwise it is set to be invalid; A5: a pre-formulated firepower strike plan is formed based on the combat plan with a valid strike evaluation result and its corresponding strike probability, strike time and strike cost; A6: the intelligent body model calls the pre-trained large language model to determine whether the pre-formulated firepower strike plan can answer the user input text. If so, the names of each model in the strike evaluation calculation model and their calculation results are returned as results.

[0014] Based on the further improvement of the above method, when the functional mode is the knowledge base model query mode, the strike information generation model calls the knowledge base model to generate a return result including: cleaning, denoising and vectorizing the user input text to obtain a user vector sequence, calculating the similarity between the user vector sequence and the vector in the knowledge base model, obtaining a candidate vector sequence with a similarity higher than a first preset threshold, and letting the user judge whether the candidate vector sequence can answer the user input text. If so, the knowledge base model query result is used as the return result; if not, calling the knowledge graph model query to obtain the result as the return result.

[0015] Based on the further improvement of the above method, the strike information generation model generates a prompt word based on the return result of the functional mode, and the prompt word is input into the pre-trained large language model to generate the strike information, including: filling the model called in the functional mode, the return result and the user input text into the preset prompt word template to generate the prompt word, and the intelligent agent calls the pre-trained large language model to generate the strike information according to the prompt word.

[0016] Based on the further improvement of the above method, the preset prompt word template is implemented by using the PromptTemplate class, and the fields contained in the PromptTemplate class include the function mode, the return result and the user input text.

[0017] Based on further improvement of the above method, the vector similarity calculation method includes Faiss, Milvus or Pgvector.

[0018] Compared with the prior art, the present invention can achieve at least one of the following beneficial effects:

[0019] 1. The present invention proposes a method for generating strike information based on multi-element fusion, which integrates strike evaluation calculation model, knowledge graph model, pre-trained large language model, knowledge base model and intelligent agent model based on the large language model development framework, performs intent recognition on user input text through the pre-trained large language model to obtain a functional mode that matches the intent, and obtains the return result of the running functional mode, generates prompt words based on the return result, and inputs the prompt words into the pre-trained large language model to generate strike information. Compared with the prior art, the strike information generation method provided by the present invention is simple to operate, has high accuracy and strong interpretability.

[0020] 2. The present invention proposes a method for generating strike information based on multi-element fusion, in which a large language model is pre-trained to determine the functional mode based on user intent, without manually judging whether it is a computing task or a query task and then calling the corresponding functional mode, thereby improving the efficiency of obtaining strike information and being more convenient.

[0021] 3. The present invention proposes a method for generating strike information based on multi-element fusion. Users only need to input questions and related information to obtain more readable strike information through a large language model development framework, thereby helping users make decisions more conveniently and quickly.

[0022] In the present invention, the above-mentioned technical solutions can also be combined with each other to achieve more preferred combination solutions. Other features and advantages of the present invention will be described in the subsequent description, and some advantages can become obvious from the description, or can be understood by practicing the present invention. The purpose and other advantages of the present invention can be realized and obtained through the contents particularly pointed out in the description and the drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] The accompanying drawings are only used for the purpose of illustrating specific embodiments and are not to be considered as limiting the present invention. In the entire drawings, the same reference symbols represent the same components;

[0024] Figure 1A schematic diagram of a functional pattern matching process based on user input text by a strike information generation model in this application;

[0025] Figure 2 A schematic diagram of the agent model calling mode in this application;

[0026] Figure 3 A schematic diagram of calling a computational model for the intelligent agent model in this application;

[0027] Figure 4 A top view of the target aircraft in this application;

[0028] Figure 5 A side view of the target aircraft in this application;

[0029] Figure 6 A rear view of the target aircraft in this application;

[0030] Figure 7 A schematic diagram of the fuselage compartment division in this application;

[0031] Figure 8 This is a schematic diagram of the strike tree for a medium strike on the F8-ii target in this application;

[0032] Fig. 9 A schematic diagram of a knowledge base model query mode in this application;

[0033] Fig.10 A schematic diagram of the answer mode of the pre-trained large language model in this application;

[0034] Fig.11 A schematic diagram of generating strike information based on prompt words in this application;

[0035] Fig.12 This is a schematic diagram of the visual process of the strike information in this application;

[0036] Fig.13 FIG. 4 is a schematic diagram showing that the visualization result in this application is in image format. DETAILED DESCRIPTION

[0037] The preferred embodiments of the present invention are described in detail below in conjunction with the accompanying drawings, wherein the accompanying drawings constitute a part of this application and are used together with the embodiments of the present invention to illustrate the principles of the present invention, but are not used to limit the scope of the present invention.

[0038] A specific embodiment of the present invention discloses a method for generating strike information based on multi-element fusion, which specifically includes:

[0039] S1: Construct a large language model development framework, in which the attack assessment calculation model, knowledge graph model, pre-trained large language model, knowledge base model and agent model are integrated to obtain an attack information generation model, see Figure 1 , which gives a schematic diagram of the strike information generation model performing functional pattern matching based on user input text.

[0040] The strike assessment calculation model includes the target vulnerability analysis model, the force field analysis model, the strike effect analysis model and the strike effect assessment model.

[0041] The knowledge graph model includes entities and the relationships between entities. The entities include targets, attack parameters and attack platforms. The data architecture can be designed according to the attributes of each type of element in the entity to build a knowledge graph. Exemplarily, the attributes of the target include target name, target type, geometric dimensions, structural characteristics, etc. The attack parameters include projectile type, geometric shape and size, charge type and quantity, explosion parameters, etc. The attributes of the attack platform include geometric dimensions, structural characteristics, etc. The relationships between entities include the relationship between the target and the attack platform, and the relationship between the attack platform and the attack parameters; Exemplarily, for a target, its corresponding attack platform is the attack platform suitable for attacking the target, and the attack parameters corresponding to the attack platform are the attack parameters suitable for the attack platform.

[0042] The knowledge base model includes text files, video files and voice files, all of which are derived from public data in the field of strike assessment. Expert experience knowledge is compiled into the knowledge base model, such as the understanding of battlefield situation, the selection of strike targets, the use of manpower and firepower, etc., and standardized as an auxiliary planning tool.

[0043] Build an intelligent agent based on the large language model development framework, and write interface specifications for calling various models to implement model calling.

[0044] After the large language model is built, it can be fine-tuned using specific corpus. The fine-tuning work can be completed in the early preparation stage. Due to the limitation of hardware facilities, the P-Tuning 2 algorithm or the Lora algorithm can be used to achieve fine-tuning. The specific corpus is the text input by users in the field of strike assessment collected manually, and each user text is annotated with data. The annotation contents are: knowledge base query (KB), large model generation (LLM), and intelligent agent call (Agent);

[0045] Here are some specific predictions for illustration: {“Please query what are the attack modes of T-72 vehicle targets?”, “KB”}; {“Generate a description of the KK-level attack (lethal attack) of the T-72 vehicle”, “LLM”}; {“Please use the target vulnerability model to analyze the T-72 tank”, “Agent”}. The fine-tuned pre-trained large language model can understand the text input by the user more quickly and accurately, and extract the hidden intent from it.

[0046] Preferably, the LangChain framework is selected as the large language model development framework, and the ChatGLM3-6B model is selected as the pre-trained large language model.

[0047] S2: The attack information generation model receives the user input text, calls the pre-trained large language model to perform intent recognition on the user input text to obtain a functional mode that matches the intent, and executes the functional mode to obtain a return result. The functional mode includes: (1) agent model call mode; (2) knowledge base model query mode; (3) pre-trained large language model answer mode. Specifically including:

[0048] 1. When the function mode is the agent model calling mode, the attack information generation model calls the agent model to generate a return result. The agent model calling mode is as follows: Figure 2 shown.

[0049] The agent model uses the Reasoning and Acting framework to control the execution order of each model in the strike assessment calculation model, and generates return results based on the calculation results of each model. Figure 3 A schematic diagram of the agent model calling each computing model is given. The standard Reasoning and Acting (ReAct) process includes: thinking / action / action input / observation. Thought refers to thinking about how to solve the problem, action is a specific tool, action input is the parameter used by the tool, and observation is the result obtained after the tool is executed. This process can be repeated many times until the final answer is obtained.

[0050] Preferably, the execution order of each model obtained based on the ReAct framework is: target vulnerability analysis model, knowledge graph model, force field analysis model, strike effect analysis model and strike effect evaluation model.

[0051] Preferably, generating a return result based on the calculation results of each model specifically includes:

[0052] A1: Obtaining a strike target, the intelligent agent model calls the target vulnerability analysis model to obtain a strike tree corresponding to the strike target;

[0053] A2: The intelligent agent model calls the knowledge graph model to query and obtain the attack platform applicable to the attack target and its corresponding attack parameter data, the intelligent agent model calls the power field analysis model to perform power field analysis on the attack platform and the attack parameter data to obtain the available attack platform and its corresponding available attack parameters that meet the user's intention, and generates a combat plan based on the available attack platform and its corresponding available attack parameters;

[0054] A3: The intelligent agent model calls the strike effect analysis model, and performs strike effect analysis on the combat plan based on the strike tree to obtain the strike probability, strike time and strike cost corresponding to the combat plan;

[0055] A4: The intelligent agent model calls the strike effect evaluation model to determine whether the combat plan is effective. If the strike probability, the strike time and the strike cost all meet the preset conditions, the strike evaluation result is set to be effective, otherwise it is set to be invalid.

[0056] A5: A pre-formulated firepower strike plan is formed based on the effective combat plan based on the strike assessment result and the corresponding strike probability, the strike time required and the strike cost;

[0057] A6: The intelligent agent model calls the pre-trained large language model to determine whether the pre-formulated firepower strike plan can answer the user input text. If so, the names of each model in the strike evaluation calculation model and their calculation results are returned as results. In some cases, if it is determined that the firepower strike plan cannot answer the user input text, the default message "Sorry, the current intelligent agent model cannot answer your question" is returned.

[0058] For example, suppose that information about the approach of an enemy aircraft detected by a drone in front is received, and its current waypoint position is obtained as latitude 49:00:00.00n, longitude 128:30:24.58w, altitude 10000feet, and current speed 206m / s. The multiple images of the target aircraft captured are transmitted to the command center controllable by the rear operator, and the strike mission to be inflicted on the target is conveyed: carry out air strikes, long-range precision strikes, and cause moderate strikes.

[0059] In response to the news and strike missions from the front battlefield, the operator writes user input text in the following format:

[0060] Question: {Carry out a long-range precision strike on an enemy aircraft located at latitude 49:00:00.00n, longitude 128:30:24.58w, altitude 10000feet, and current speed 206m / s, causing a medium strike. Please formulate a fire strike plan.}

[0061] The present invention adopts a coarse-grained strike level classification standard, including: severe strikes correspond to level I strikes, which are manifested as the aircraft losing control within 30 seconds after being attacked; moderate strikes correspond to level II strikes, which are manifested as the aircraft losing control within 5 minutes after being attacked, or being unable to fly back to the base; light strikes correspond to level III strikes, which are manifested as the aircraft being unable to complete the mission after being attacked.

[0062] The pre-trained large language model performs intent recognition on the user input text and determines that the user's intent is "cause a moderate attack, please formulate a fire attack plan", and the user's purpose is "formulate a fire attack plan", so the functional mode that matches the intent is the intelligent agent model call mode, which obtains the fire attack plan based on the attack evaluation calculation model, and the intent key is "cause a moderate attack". Therefore, the attack information generation model calls the intelligent agent model to generate the return result.

[0063] Furthermore, the agent model calls the pre-trained large language model to determine the models needed to answer the user input text, including: target vulnerability analysis model, knowledge graph model, force field analysis model, strike effect analysis model, strike effect evaluation model and other computational analysis models related to strike evaluation, and determines the execution order of the target vulnerability analysis model, knowledge graph model, force field analysis model, strike effect analysis model and strike effect evaluation model based on the ReAct framework. The return results are generated based on the calculation results of each model, including:

[0064] A1: Get multiple images of the target aircraft. Figure 4 is a top view of the target aircraft, Figure 5 is the side view of the target aircraft, Figure 6 This is the rear view of the target aircraft. After identification, the target is identified as F8-ii. The target vulnerability analysis model first describes each component according to the cabin strike principle based on the strike target, and then models the strike target based on the description to generate a strike tree under the medium strike level, including:

[0065] First, the impact of component strikes on the target is analyzed: the key compartments of the target mainly include the radar compartment, cockpit, avionics system, fuel system, engine, control system and landing gear, etc. The strikes on each component of the target are described according to the compartment strike principle: the fuselage is the main structure of the aircraft, and the main avionics system and functional system of the aircraft are distributed throughout the fuselage. The fuselage is divided into multiple compartments from the nose to the tail. Figure 7 This is a schematic diagram of the fuselage compartment division.

[0066] The first compartment is the radome, which is made of cured glass fiber epoxy resin material. The penetration of fragments will not have a direct impact on its function, but will only affect part of the aerodynamic performance, causing a Level III strike. The overpressure deformation of the shock wave destroys the streamlined shape of the aircraft, affecting the aircraft speed, causing a Level III strike.

[0067] The second compartment contains the airborne radar and pitot tube. If the radar strikes, the aircraft will lose its detection and identification capabilities and will be unable to attack the target, resulting in a Level III strike. If the pitot tube strikes, the aircraft's speed will be uncontrollable and it will lose some of its flying capabilities, reaching a Level II strike.

[0068] The third compartment is the liquid-cooled avionics equipment compartment, where the fighter's main avionics system central processors are located. A small-area penetration can cause a Level III strike, while a large-area strike can directly cause the aircraft to lose control, resulting in a Level I strike.

[0069] The fourth compartment is the cockpit system and crew. The cockpit system is the control center of the aircraft. If the display screen is destroyed, the pilot will not be able to obtain flight data, causing the aircraft to lose control, resulting in a Level I damage. If the side joystick and throttle lever are broken by fragments, the aircraft will lose control, resulting in a Level I damage. If the pilot is hit, it may cause a Level III or even Level I damage depending on the injured part and the size of the wound. If the canopy is destroyed, the crew environment will deteriorate rapidly and flying will be difficult, resulting in a Level II damage.

[0070] The fifth compartment consists of the fuel tank and the front landing gear. The fuel tank was punctured and leaked oil, shortening the aircraft's range and causing a Level II hit. The fragments ignited the fuel tank, causing the aircraft to explode and disintegrate, causing a Level I hit. The front landing gear was hit, causing a Level II hit.

[0071] The sixth compartment is the air inlet and fuel tank. If the fuel tank is punctured and leaks oil, the aircraft's range will be shortened, causing a Level II hit; if the fuel tank explodes, it will cause a Level I hit; if the air inlet is damaged by overpressure, the inlet flow field will be distorted, and the aircraft's propulsion capability will be reduced, causing a Level III hit; if the air inlet is punctured by fragments, it will easily produce secondary fragments, causing the engine to inhale foreign matter, causing a Level III hit.

[0072] The seventh compartment contains the main fuel tank and the main landing gear. If the fuel tank is punctured, the aircraft's fuel supply is damaged, the flight range is greatly shortened, and this results in a Level II strike; if the fuel tank explodes, this results in a Level I strike; if the landing gear is hit, the aircraft cannot take off or land, and this results in a Level II strike.

[0073] The eighth compartment is a twin-engine compartment. If a single engine is hit, the aircraft will lose some of its maneuverability and can only fly by relying on flying skills. It is considered impossible to fly back to the base intact, resulting in a Level II hit. If both engines are hit at the same time, it will result in a Level I hit. If the engine is subdivided according to its composition, such as the compressor barrel, combustion chamber and turbine, the engine thrust will be significantly reduced, resulting in a Level III hit.

[0074] The functional strike of the wing is mainly composed of two parts: the strike of maneuverability and the strike of load-bearing performance. The flaps and ailerons of the aircraft realize the flexibility of the aircraft's maneuverability. The strike of flaps and ailerons will reduce the aircraft's flight lift and landing resistance, resulting in a level III strike. The aileron strike will cause the aircraft to lose flight control, resulting in a level I strike. The main load-bearing beam of the wing mainly bears the bending moment and lift generated during the flight of the aircraft. If the main beam is damaged and the wing is overloaded and broken, the aircraft will disintegrate in the air, resulting in a level I strike. At the same time, there are a large number of avionics sensors distributed on the aircraft's flaps. When the flaps are damaged, the aircraft will suffer a level III strike.

[0075] The horizontal tail is a fully movable tail wing, which is driven by a servo and a pivot. If the servo or pivot is hit, the horizontal tail will lose control, and it is believed that the aircraft will also quickly lose flight control, resulting in a Level II strike.

[0076] The vertical tail is composed of a stabilizer and a rudder. If any one of its components is hit, the aircraft's control ability will be seriously damaged, resulting in a Level II strike.

[0077] According to the compartment strike principle, the strike analysis description of the strike target is used to model the target vulnerability, and finally the target moderate strike tree structure of F8-ii is obtained as follows: Figure 8 shown.

[0078] A2: The intelligent agent model calls the knowledge graph model to query and obtain the attack platform and its corresponding attack parameter data suitable for the attack target, and the intelligent agent model calls the power field analysis model to perform power field analysis on the attack platform and the attack parameter data to obtain the available attack platform and its corresponding available attack parameters that meet the user's intention, and generates a combat plan based on the available attack platform and its corresponding available attack parameters, specifically including:

[0079] Through the query of the knowledge graph model, the attack platforms suitable for the current attack target and their corresponding attack parameter data are obtained as follows: (1) F-15 fighter jets, an air mobile launch platform, which supports the following types of missiles: AIM-7, AIM-9 and AIM-120; (2) 2S6 air defense vehicle, a ground mobile launch platform, which mainly includes two 30mm 2A38M anti-aircraft machine guns and eight 9M311 missiles. The double-barreled anti-aircraft machine guns have a maximum firing rate of 2,500 rounds / minute / barrel, an initial velocity of 960 m / s, a minimum range of 0.2 km, a maximum range of 4 km, and a maximum altitude of 3 km. The 9M311 is about 2.5 meters long and 150 mm in diameter. Its maximum flight speed is about Mach 2, and its average flight speed can reach Mach 1.7. Its maximum range is 8 km, and its maximum range after modernization is 15 km, with a maximum altitude of 3.5 km. After the power field analysis model is used for calculation and analysis, the available attack platforms and their corresponding available attack parameters that meet the user's intention are as follows: (1) F-15 fighter jets on an airborne mobile launch platform, using 4 AIM-120C missiles; (2) 2s6 air defense vehicle on a ground-based mobile launch platform, using two 2A38M anti-aircraft machine guns and 8 9M311 missiles. Based on the available attack platforms and their corresponding available attack parameters, the combat plans are generated as follows: Plan 1: Use a fighter formation, the host selects an F-15 fighter jet, mounts 4 AIM-120C missiles, and has four wingmen; Plan 2: Use an air defense system, using a 2s6 air defense vehicle on a ground-based mobile launch platform, loaded with two 30mm 2A38M anti-aircraft guns and 8 9M311 missiles.

[0080] A3: The intelligent agent model calls the strike effect analysis model, and performs strike effect analysis on the combat plan based on the strike tree to obtain the strike probability, strike time and strike cost corresponding to the combat plan. It is found that the strike probability of causing a moderate strike to the target in plan one is 85%, the time required is 20 minutes, and the cost is estimated to be 200,000; the strike probability of causing a moderate strike to the target in plan two is 90%, the time required is 1h40min, and the cost is estimated to be 250,000.

[0081] A4: The intelligent agent model calls the strike effect evaluation model to determine whether the combat plan is effective. If the strike probability, the strike time and the strike cost all meet the preset conditions, the strike evaluation result is set to be effective, otherwise it is set to be invalid.

[0082] Since the probability of achieving the target strike mission for both plans is greater than 80%, and the required time and cost are within the allowable range, the strike evaluation results of both plans are valid.

[0083] A5: A pre-formulated firepower strike plan is formed based on the effective combat plan based on the strike assessment result and its corresponding strike probability, strike time required and strike cost results.

[0084] A6: Agent thinking, that is, the agent model calls the pre-trained large language model to determine whether the pre-formulated firepower strike plan can answer the user input text. If so, the names of each model in the strike evaluation calculation model and its calculation results are returned as the result. In some cases, if it is determined that the firepower strike plan cannot answer the user input text, the default message "Sorry, the current agent model cannot answer your question" is returned.

[0085] When the various models and their calculation results obtained in the agent model call mode are used as the return result, the specific form of the returned result is: AGENT_RESULT="This is the result from [{tool_name1},{result1};{tool_name2},{result2};…]".

[0086] Exemplarily, the returned result for the agent model call mode is:

[0087] AGENT_RESULT= “This is the result from {target_analysis},{The targetis F8-ii, and this is the picture of the moderate damage(moderate_damage.png)};{ Knowledge Graph Model、Ammunition power fieldanalysis},{There are two usable platform for the task, one is F-15.json, the other is 2S6.json}; {damage_assessment},{This is the damage assessment resultof F-15(F-15.txt); this is the damage assessment result of 2S6(2S6.txt) ;{damage_effect},{"result": "1", "result": "1"}".

[0088] Among them, moderate_damage.png is the analysis result returned by the target vulnerability analysis model, that is, the strike tree of the target under moderate strike after the target analysis, which is returned in the form of a picture. The F-15.json file and the 2S6.json file are two combat plans that the force field analysis model considers suitable for the task. The F-15.txt file and the 2S6.txt file are files that are based on the strike effect analysis model to obtain the strike probability, strike time and strike cost of the two combat plans. "result": "1" indicates that the strike effect evaluation model determines that the current combat plan is valid.

[0089] 2. When the function mode is the knowledge base model query mode, the attack information generation model calls the knowledge base model to generate a return result, specifically including: cleaning, denoising and vectorizing the user input text to obtain a user vector sequence, calculating the similarity between the user vector sequence and the vector in the knowledge base model, obtaining a candidate vector sequence with a similarity higher than a first preset threshold, and letting the user judge whether the candidate vector sequence can answer the user input text. If so, the knowledge base model query result is used as the return result. If not, the knowledge graph model query is called to obtain the result as the return result. If the knowledge graph model still cannot answer the user's question, the default message "Sorry, the current knowledge base model query model cannot answer your question" is returned. The knowledge base model query mode is as follows: Fig. 9 shown.

[0090] Preferably, after the calculation results are sorted, the text segments corresponding to the first three vector sequences with the highest similarity are obtained by default as the return results of the knowledge base query, and the user can specify the number of similar vector sequences to be returned.

[0091] Exemplarily, the text may be vectorized by using a word embedding model m3e-base suitable for Chinese text, and the word embedding model is used to represent discrete natural language text as a continuous real number vector.

[0092] Exemplarily, for vector similarity calculation, the vector retrieval models embedded in LangChain can be used, including: Faiss, Milvus, and Pgvector. For the Faiss vector retrieval model, it is implemented using efficient compression and indexing techniques, supports multiple distance metrics and indexing algorithms, has fast data retrieval speed, and has high flexibility and scalability, but only supports two SDK languages, Python and JavaScript. When faster query speed and better scalability are required, Faiss can be selected. For the Milvus vector retrieval model, it supports the fusion of multiple data sources and multiple modal data, provides accurate similarity matching and efficient online learning functions, and supports streaming and batch processing of data. It supports multiple SDK languages ​​such as Python, Java, Go, C++, Node.js, and Ruby. When more functions and operations are required, or complex tasks such as multimodal tasks need to be processed, or higher accuracy is sought, Milvus can be selected. For the Pgvector vector retrieval model, it is an advanced technology that integrates two or more search algorithms to improve the relevance and accuracy of search results. When hybrid search (i.e., scalar filtering) needs to be implemented, Pgvector can be selected.

[0093] When the text segment obtained in the knowledge base model query mode is used as the return result, the return result format is: KB_RESULT = "This is the result from [{strike information text segment that meets user needs}]".

[0094] 3. When the function mode is the pre-trained large language model answer mode, the strike information generation model calls the pre-trained large language model to generate a return result. Fig.10 Schematic diagram of the answer mode for pre-trained large language models.

[0095] When the text obtained in the pre-trained large language model answer mode is used as the return result, the return result format is: LLM_RESULT = "This is the result from [{strike information text that meets user needs}]".

[0096] S3: the strike information generation model generates a prompt word based on the return result of the functional mode, and inputs the prompt word into the pre-trained large language model to generate strike information, including:

[0097] The model called in the functional mode, the returned result and the user input text are filled into a preset prompt word template to generate a prompt word, and the intelligent agent calls the pre-trained large language model to generate attack information according to the prompt word. Fig.11 A schematic diagram of generating strike information based on prompt words is given.

[0098] The preset prompt word template is implemented using the PromptTemplate class, and contains fields such as the function pattern pattern, the return result result, and the user input text input. Exemplarily, the function pattern pattern is one of the agent model call mode AGENT, the knowledge base model query mode KB, and the pre-trained large language model answer mode LLM. Exemplarily, the prompt word template Prompt=[AGENT / KB / LLM, AGENT_RESULT / KB_RESULT / LLM_RESULT, input].

[0099] Preferably, a thinking chain of the strike information generation model is constructed, wherein the input parameters of the thinking chain are the large language model and the prompt words, and the pre-trained large language model in the strike information generation model and the prompt words constructed in the above steps are input into the thinking chain to obtain the final strike information, specifically including: the thinking chain calls the pre-trained large language model to generate the final strike information according to the prompt words.

[0100] Exemplarily, the structure of the strike information generated by the intelligent model calling mode is: {attack platform, attack parameters and their quantity, strike probability, strike time required, strike cost, strike method and visual display method}, the strike method includes air-to-air, air-to-ground, ground-to-air, air-to-ship, shore-to-ship, ship-to-ship, ship-to-air, ship-to-ground, ship-to-submarine, submarine-to-ship, submarine-to-submarine, submarine-to-ground, anti-tank, etc., and the visual display method includes images and tables. Preferably, for the pre-formulated firepower strike plan 1: {F-15 fighter; 4 AIM-120C bombs; 4 wingmen; strike probability 85%; required time 20min; cost estimate 200,000; air-to-air; table}, for the pre-formulated firepower strike plan 2: {2s6 air defense vehicle, two 2A38M anti-aircraft guns and 8 9M311 bombs; strike probability 90%; required time 1h40min; cost estimate 250,000; ground-to-air; table}. After obtaining the strike information, you can call a pre-written program to visualize the strike information, such as Fig.12 As shown. For example, the image format is as Fig.13 As shown, it gives the means of causing damage to the target T-72 obtained by using the knowledge base model query mode when the target is T-72; the table format is shown in Table 1, which gives the firepower strike plan obtained by using the intelligent model call mode when the target is F8-ii.

[0101] Table 1 Fire attack plan in tabular format

[0102]

[0103] Compared with the prior art, the present embodiment provides a method for generating strike information based on multi-element fusion, which is simple to operate, highly accurate, and highly interpretable; the pre-trained large language model determines the functional mode based on the user's intention, without manually judging whether it is a computing task or a query task and then calling the corresponding functional mode, thereby improving the efficiency of obtaining strike information and being more convenient; the user only needs to input the question and its related information to obtain more readable strike information through the large language model development framework, thereby helping the user to make decisions more conveniently and quickly.

[0104] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by any technician familiar with the technical field within the technical scope disclosed by the present invention should be covered within the protection scope of the present invention.

Claims

1. A method for generating attack information based on multi-element fusion, characterized in that: include: S1: constructing a large language model development framework, integrating a strike assessment calculation model, a knowledge graph model, a pre-trained large language model, a knowledge base model, and an agent model in the large language model development framework, thereby obtaining a strike information generation model; The attack assessment calculation model includes a target vulnerability analysis model, a force field analysis model, an attack effect analysis model and an attack effect assessment model; the knowledge graph model includes entities and relationships between entities, and the entities include targets, attack parameters and attack platforms; the pre-trained large language model is obtained by fine-tuning the large language model based on the P-Tuning 2 algorithm or the Lora algorithm; S2: The attack information generation model receives the user input text, calls the pre-trained large language model to perform intent recognition on the user input text to obtain a functional mode that matches the intent, and executes the functional mode to obtain a return result. The functional mode includes: (1) agent model call mode; (2) knowledge base model query mode; (3) pre-trained large language model answer mode; Wherein, when the function mode is the agent model calling mode, the attack information generation model calls the agent model to generate a return result; The attack information generation model calls the agent model to generate a return result, including: the agent model uses the Reasoning and Acting framework to control the execution order of each model in the attack evaluation calculation model, and generates a return result based on the calculation results of each model; the generation of the return result based on the calculation results of each model includes: A1: Obtaining a strike target, the intelligent agent model calls the target vulnerability analysis model to obtain a strike tree corresponding to the strike target; A2: The intelligent agent model calls the knowledge graph model to query and obtain the attack platform applicable to the attack target and its corresponding attack parameter data, the intelligent agent model calls the power field analysis model to perform power field analysis on the attack platform and the attack parameter data to obtain the available attack platform and its corresponding available attack parameters that meet the user's intention, and generates a combat plan based on the available attack platform and its corresponding available attack parameters; A3: The intelligent agent model calls the strike effect analysis model, and performs strike effect analysis on the combat plan based on the strike tree to obtain the strike probability, strike time and strike cost corresponding to the combat plan; A4: The intelligent agent model calls the strike effect evaluation model to determine whether the combat plan is effective. If the strike probability, the strike time and the strike cost all meet the preset conditions, the strike evaluation result is set to be effective, otherwise it is set to be invalid. A5: A pre-formulated firepower strike plan is formed based on the effective combat plan based on the strike assessment result and the corresponding strike probability, the strike time required and the strike cost; A6: The intelligent agent model calls the pre-trained large language model to determine whether the pre-formulated firepower strike plan can answer the user input text. If so, the names of each model in the strike evaluation calculation model and their calculation results are returned as results; When the function mode is the knowledge base model query mode, the strike information generation model calls the knowledge base model to generate a return result; When the function mode is the pre-trained large language model answer mode, the attack information generation model calls the pre-trained large language model to generate a return result; S3: The strike information generation model generates a prompt word based on the return result of the functional mode, and inputs the prompt word into the pre-trained large language model to generate strike information.

2. The method for generating attack information based on multi-element fusion according to claim 1, characterized in that: When the functional mode is the knowledge base model query mode, the strike information generation model calls the knowledge base model to generate a return result, including: cleaning, denoising and vectorizing the user input text to obtain a user vector sequence, calculating the similarity between the user vector sequence and the vector in the knowledge base model, and obtaining a candidate vector sequence with a similarity higher than a first preset threshold, and the user determines whether the candidate vector sequence can answer the user input text. If so, the knowledge base model query result is used as the return result; if not, the knowledge graph model query is called to obtain the result as the return result.

3. According to the method for generating strike information based on multi-element fusion in claim 2, the strike information generation model generates a prompt word based on the return result of the functional mode, and inputs the prompt word into the pre-trained large language model to generate strike information, comprising: The model called in the functional mode, the returned result and the user input text are filled into a preset prompt word template to generate a prompt word, and the intelligent agent calls the pre-trained large language model to generate strike information according to the prompt word.

4. The method for generating attack information based on multi-element fusion according to claim 3 is characterized in that: The preset prompt word template is implemented using the PromptTemplate class, and the fields contained in the PromptTemplate class include the function mode, the return result, and the user input text.

5. The method for generating attack information based on multi-element fusion according to claim 4, characterized in that: The vector similarity calculation method includes Faiss, Milvus or Pgvector.

Images