A water surface platform combat force generation method based on natural language processing and intelligent optimization algorithm

Abstract

The application discloses a water surface platform combat force generation method based on natural language processing and intelligent optimization algorithm, relates to the field of natural language processing, and combines intelligent search algorithms with optimization strategies, adopts an improved multi-population genetic algorithm, that is, utilizes multiple populations to perform parallel optimization search, and realizes diversified search strategies by assigning different control parameters to different populations; information exchange is performed between various populations through a migration operator, collaborative evolution is promoted, and convergence is judged by reserving optimal individuals of each generation and using the optimal individuals to maintain generation number, so that correctness and global optimality of a final solution are ensured. The application improves the quality and response speed of combat decision.

Description

Technical Field

[0001] The present invention relates to the field of natural language processing, and in particular to a method for generating surface platform combat forces based on natural language processing and intelligent optimization algorithms. Background Art

[0002] In modern warfare, the importance of force generation for surface forces is becoming increasingly prominent. Advanced software systems automatically generate appropriate ship combat platforms and weaponry configurations based on operational intent, mission requirements, and terrain constraints. Key technologies in this process include:

[0003] First, analysis of operational intent and mission requirements is the starting point for force generation. The system needs to fully understand the commander's operational intent and specific mission requirements, including but not limited to various combat types such as attack, defense, reconnaissance, and patrol. Using natural language processing (NLP) and semantic analysis technologies, the system can extract key information from the commander's orders and mission descriptions to provide guidance for subsequent force generation.

[0004] Analysis and simulation of terrain and environment are key steps in ensuring the rational deployment of forces. The system analyzes geographic environmental data for the combat area, including sea conditions, weather conditions, and topography. Through high-precision environmental simulation, the system can assess the impact of different environments on combat operations, thereby optimizing the selection of ships and weapons.

[0005] Based on this analysis, the system automatically generates ship and weapon models from a database. This database contains platform performance, weaponry technical specifications, and combat capabilities for each type of ship. Using a multi-objective optimization algorithm, the system comprehensively considers mission requirements, enemy threats, and environmental factors to automatically generate optimal ship and weapon configurations. These solutions must not only meet mission requirements but also strike a balance between tactical effectiveness, cost, and risk.

[0006] A search revealed that Chinese invention patent CN202011198177.9 provides a mission planning and simulation system and method for long-range air support aircraft. The battlefield force generation station is used for combat scenario configuration, movement simulation of enemy and friendly forces other than aircraft, and simulation process management. The mission planning and fire control solution station is used for situational threat assessment, mission planning, fire control solution, aircraft route planning and movement simulation, weapon launch control, effectiveness evaluation, and simulation display. The battlefield force generation station is connected to the mission planning and fire control solution station via Ethernet. Chinese invention patent CN202110277412.X relates to the field of intelligent simulation system technology. It is a simulation system and method based on a dual-logic layer agent, including an intelligence acquisition unit agent, an intelligence processing unit agent, a decision control unit agent, a terminal combat unit agent, an information network agent model, and a battlefield environment and force generation agent. Leveraging agent modeling concepts, this system simulates the functional and behavioral characteristics of system units by establishing system unit agent models. Multi-agent system interaction technology is then used to simulate the interactive behavior between system units. Chinese invention patent CN202211728527.7 discloses a distributed radar early warning simulation training system, including a guidance and control subsystem, an intelligence networking subsystem, a radar training simulator, a computer force generation system, and a data model center.

[0007] Therefore, in response to the demand for ship air defense and anti-missile force generation, the present invention designs a computer force generation method based on natural language processing and intelligent optimization algorithm to realize the generation of surface ship combat platforms. Summary of the Invention

[0008] In view of the above-mentioned defects of the prior art, the technical problem to be solved by the present invention is how to quickly and accurately automatically generate the optimal ship platform and weapon configuration plan based on combat intentions, mission requirements and environmental constraints.

[0009] To achieve the above objectives, the present invention provides a method for generating surface platform combat forces based on natural language processing and intelligent optimization algorithms. This method combines an intelligent search algorithm with an optimization strategy, adopts an improved multi-population genetic algorithm, that is, utilizes multiple populations for parallel optimization search, and implements diversified search strategies by assigning different control parameters to different populations; information is exchanged between the various populations through immigration operators to promote collaborative evolution, and by retaining the best individuals in each generation and using the best individuals to maintain algebraic judgment convergence, the correctness and global optimality of the final solution are ensured.

[0010] Furthermore, the method is based on the generation of a genetic algorithm, and the solution of the troop deployment plan is represented by a "chromosome". Through the selection, crossover, mutation and other operations of the "chromosome" from generation to generation, the optimal results are compared until the optimal "chromosome" is found.

[0011] Furthermore, the improved multi-population genetic algorithm comprises the following steps:

[0012] Step 1: Encode to form the initial population. Use integer permutation encoding method. There are m pre-defined positions. The chromosome is divided into m segments. Each segment corresponds to the number of a deployment position. If the number of troops and weapons n is less than m, then the first n positions are used to deploy troops and weapons, and the last mn positions are used for virtual deployment of troops and weapons.

[0013] Step 2: Calculate the fitness. |k1|k2|k3|…|km| is a chromosome, representing a deployment plan. Fi is the combat effectiveness corresponding to the plan. is the sum of the combat effectiveness of the population, and the fitness of the i-th deployment plan is

[0014] Step 3: Selection operation. Individuals with higher fitness, i.e., individuals with higher combat effectiveness, are selected from the population through the selection operator. The greater the fitness, the greater the probability of being selected, and the more offspring are produced in the next generation. These selected individuals are then used to reproduce the next generation to produce a new population. The probability of individual selection is calculated using the roulette wheel algorithm. The probability of each individual being selected is proportional to its fitness. The higher the fitness, the greater the probability of being selected.

[0015] Step 4: Crossover operation: determine whether each deployment scheme in the population crosses according to the crossover probability Pc, group the parent samples for the crossover operation into two groups, and randomly determine the crossover position for each group to perform the crossover operation;

[0016] Step 5: Evolutionary mutation operation: determine whether each deployment plan in the population is mutated based on the mutation probability Pm, then randomly determine two positions, swap their positions, and then calculate the combat effectiveness of each deployment plan of the new population after the mutation. If the combat effectiveness is improved, the mutation is effective, otherwise it is invalid.

[0017] Step 6: To improve local search capabilities, a reversal operation is performed after selection, crossover, and mutation. If the combat effectiveness of the new individual improves, the evolutionary reversal is successful and the next cycle begins. If the combat effectiveness of the new individual decreases, the original individual is restored. That is, the reversal is only accepted if the combat effectiveness improves after reversal; otherwise, the reversal is invalid. This process is called evolutionary reversal, and the operation process is the same as evolutionary mutation.

[0018] Step 7: Evolutionary reinsertion: First, reinsert the new population obtained after reversal into the parent population. Then, randomly replace the individuals with the highest combat effectiveness in the population after evolutionary reversal with the individuals in the reinserted population, thereby increasing the number of optimal individuals in each generation and improving the evolution rate.

[0019] Step 8. Improve the genetic algorithm and adopt steps 1-7. Different populations are assigned different control parameters to achieve different search purposes. The populations are connected through immigration operators to achieve multi-population co-evolution. The optimal individuals in each generation of evolution of various populations are preserved through artificial selection operators, and the optimal individual retention generation number is used as the basis for judging the convergence of the algorithm.

[0020] Furthermore, the force generation method should be flexibly and scientifically determined based on combat missions, enemy situation, our own situation, battlefield environment, and tactical requirements to ensure maximum combat effectiveness. A combat effectiveness model, a theater weighted employment model, and battlefield environment influencing factors should be established based on the red and blue ship equipment models. Multiple factors such as mission requirements and terrain environment should be accurately quantified into specific constraints to provide clear optimization boundaries and conditions for the algorithm.

[0021] Furthermore, the modeling of the constraint conditions mainly includes a combat effectiveness model, a battle zone weight evaluation model, and battlefield environment influencing factors.

[0022] Furthermore, the combat effectiveness model can be expressed as

[0023]

[0024] Among them, F is the sum of the combat effectiveness of the participating forces; α i is the impact factor of the battlefield environment on the deployment of the i-th type of troops and weapons; ri is the weight index of the i-th theater; Q i is the combat effectiveness index of the i-th type of troops and weapons; m is the total number of troops and weapons;

[0025] Q i =[lnC+ln(∑D+1)+lnE]ξ1ξ2ξ3

[0026] where Q i is the ship's combat effectiveness index; C is the maneuverability parameter; D is the firepower parameter; E is the target detection capability parameter; ξ1 is the survivability coefficient; ξ2 is the range coefficient; ξ3 is the electronic countermeasure capability coefficient;

[0027] (1) Mobility parameters

[0028]

[0029] Among them, n zatr is the maximum permissible overload; SEP is the turning rate, m / s;

[0030] (2) Firepower parameters

[0031] ∑D=D G +D1+D2+D3

[0032] Among them, ∑D is the total firepower parameter of the fighter; D G is the firepower parameter of the ship's gun; D1 is the parameter of a mounted ballistic missile; D2 is the parameter of another mounted cruise missile; D3 is the parameter of another mounted gliding missile;

[0033] D G Calculate according to the following formula

[0034]

[0035] Among them, rpm is the firing rate per minute, that is, the firing rate; V 0b is the initial velocity of the projectile; G b is the weight of the projectile; D gb is the caliber of the ship's gun; n is the number of ship's guns;

[0036] D1 is calculated as follows:

[0037]

[0038] Among them, R f is the maximum practical effective range; ΔH f To allow the total launch height difference; P K is the single-shot kill probability; n zmmax is the maximum overload of the missile;

[0039] The calculation method of D2 and D3 is the same as D1;

[0040] (3) Detection capability parameters

[0041] Only consider the performance parameters of shipborne radar

[0042]

[0043] Among them, L d is the maximum detection distance; m1 is the number of targets tracked simultaneously; m2 is the number of targets allowed to be attacked simultaneously;

[0044] (4) Viability coefficient

[0045]

[0046] Among them, l W L is the ship width; F is the overall length of the ship; E is the performance parameter of the shipborne radar;

[0047] (5) Range coefficient

[0048]

[0049] Among them, L maxibf It is the maximum range corresponding to the oil in the aircraft.

[0050] Furthermore, in the war zone weight evaluation model, the main factors used to measure the combat zone include: military value, political value, economic value and social value, and the evaluation sets include: extremely important, very important, important and general;

[0051] Let A be the set of factors for evaluating the importance of a theater: A = {military value a1, political value a2, economic value a3, social value a4}; let B be the set of evaluation results: B = {extremely important b1, very important b2, important b3, general b4}; by statistically obtaining the evaluation ratio of each factor of the theater's importance, we can obtain the single factor evaluation matrix C = c ij , where c ij Indicates that the evaluated war zone is from factor a i , let's look at the evaluation set b j Membership degree, i = 1, 2, 3, 4, j = 1, 2, 3, 4;

[0052] Let D be the weight vector of each factor: D = (d1, d2, d3, d4). In this problem, the weight of each factor in the factor set is determined to be D = (0.6, 0.2, 0.1, 0.1) according to the expert judgment method. The comprehensive evaluation result vector is obtained by weighted average: E = D × C = (e1, e2, e3, e4), where e i Indicates that after considering all factors, the evaluated battlefield is evaluated against the evaluation set b i The degree of membership of each level in

[0053] In order to obtain the weight index of the theater through the evaluation result vector E, let F = (1, 0.8, 0.4, 0.2) be the weight of each evaluation result in the evaluation set B, and the weight index of the evaluated theater is obtained by weighted average: r = F·E T ;

[0054] Finally, the importance indexes of the n combat areas are normalized between [0, 1] to obtain the weight distribution of the combat zones: R = [r1, r2, ..., r n ],satisfy Among them, r i is the weight index of the i-th war zone.

[0055] Furthermore, the battlefield environmental factors mainly include terrain, meteorology, hydrology, and electromagnetics, which have different degrees of influence on military operations. The three-dimensional real-life map can not only directly obtain geographic coordinates (x, y) and elevation (H), but also measure and query attribute information that affects troop deployment, such as water flow direction and velocity D, meteorological information P, visibility S, trafficability T, and protection F. That is, the terrain attributes of a certain location can be expressed as

[0056] R i (X i ,Y i ,H i ,D i ,P i ,S i ,T i ,F i )

[0057] Use It indicates the influence coefficient of hydrological information, average slope, visibility, trafficability, and protection capability on combat effectiveness. The value indicates the degree of influence on combat effectiveness. The larger the value, the greater the influence of the location on combat effectiveness. The specific value can be comprehensively evaluated based on data statistics. Therefore, It is expressed as the comprehensive impact factor of battlefield environment on troop deployment.

[0058] Furthermore, the method adopts conditional random field (CRF) for coarse segmentation, combines domain dictionary construction and word segmentation model prediction, and effectively trains and recognizes military terms.

[0059] Furthermore, the method performs part-of-speech tagging training based on a long short-term memory (LSTM) network with an attention mechanism to capture complex text context and part-of-speech features; by searching the output classification probabilities, the highest-scoring tagging sequence is obtained, thereby improving the parsing accuracy of military terminology and the system's natural language understanding capabilities.

[0060] This technology can quickly and accurately automatically generate the optimal ship platform and weapon configuration plan based on combat intent, mission requirements and environmental constraints, greatly reducing the uncertainty caused by manual operation and experience-based judgment, and significantly improving decision-making speed and combat effectiveness. These technological innovations not only improve the intelligence level of the system, but also enhance the system's adaptability in complex battlefield environments, fully meeting the urgent needs of modern naval warfare for efficient decision-making support systems. The technical solution adopted by the present invention has demonstrated excellent performance indicators in terms of force generation speed, configuration plan accuracy, algorithm efficiency and dynamic adjustment capabilities. Through the combination of optimization algorithms and parallel computing, the present invention can screen out the optimal solution from a large number of possible solutions in a very short time, ensuring the rationality and scientific nature of force configuration. In addition, the integration of NLP technology enables the system to accurately parse complex natural language commands, ensuring that the generation of combat plans is highly consistent with the commander's actual intentions. DETAILED DESCRIPTION

[0061] The present invention can be embodied in many different forms of embodiments, and the protection scope of the present invention is not limited to the embodiments mentioned herein.

[0062] 1. Force Generation Based on Genetic Algorithms

[0063] The basic principle of the genetic algorithm is to emulate the evolutionary law of "survival of the fittest" in the biological world, forming an adaptive global optimization probabilistic search algorithm. The troop deployment based on the genetic algorithm represents the solution of the troop deployment plan as a "chromosome". Through the selection, crossover, mutation and other operations on the "chromosome" from generation to generation, the optimal results are compared until the optimal "chromosome" is found.

[0064] 1. Encoding to form the initial population

[0065] An integer permutation encoding method is used. With m pre-selected positions, the chromosome is divided into m segments, i.e., the chromosome length is m, and each segment corresponds to a deployment position number. If the number of troops and weapons n is less than m, the first n positions are used for deployment, and the last mn positions are used for virtual deployment. For example, if 10 positions are pre-selected on the map, then: |5|10|3|1|6|7|4|2|8|9 is a valid chromosome. If only 5 troops and weapons need to be deployed, 5 virtual troops and weapons with very low firepower indexes are used to fill in the positions, establishing a mapping relationship between the codes and the deployment of troops and weapons.

[0066] 2. Calculate fitness

[0067] |k1|k2|k3|…|k m | is a chromosome, representing a deployment plan, F i is the combat effectiveness corresponding to this plan, is the sum of the combat effectiveness of the population (all plans), and the fitness of the i-th deployment plan is

[0068]

[0069] 3. Select an action

[0070] A selection operator selects individuals with higher fitness (FintV), or higher combat effectiveness, from the population. The greater the fitness, the greater the probability of selection, and the greater the number of offspring produced in the next generation. These selected individuals are then used to reproduce the next generation, creating a new population. The probability of individual selection is calculated using a roulette wheel algorithm. The probability of selection for each individual is proportional to its fitness, with individuals with higher fitness having a greater probability of selection.

[0071] 4. Crossover Operation

[0072] According to the crossover probability P c Determine whether each deployment scheme in the population crosses, determine the parent samples of the crossover operation and group them into twos, and randomly determine the crossover position for each group to perform the crossover operation.

[0073] 5. Evolution and mutation operations

[0074] According to the mutation probability P m Determine whether each deployment plan in the population has mutated. Then randomly determine two positions, swap them, and calculate the combat effectiveness of each deployment plan in the new population after the mutation. If the combat effectiveness improves, the mutation is effective, otherwise it is invalid. Ensure that the combat effectiveness of the new population after the mutation is increasing.

[0075] 6. Evolution reversal operation

[0076] In order to improve the local search capability, a reversal operation is performed after selection, crossover, and mutation.

[0077] If the combat effectiveness of the new individual is improved, the evolution is reversed successfully and enters the next cycle. If the combat effectiveness of the new individual decreases, the original individual is restored, that is, the combat effectiveness is improved after the reversal and it is accepted, otherwise the reversal is invalid. This process is the evolution reversal operation, and the operation process is the same as the evolution mutation operation.

[0078] 7. Evolutionary Reinsertion

[0079] First, the new population obtained after reversal is reinserted into the parent population. Then, the individuals with the highest combat effectiveness in the population after evolutionary reversal are randomly replaced with the individuals in the reinserted population, thereby increasing the number of optimal individuals in each generation and improving the evolution rate.

[0080] 8. Multi-population genetic algorithm

[0081] The calculation of population combat effectiveness based on genetic algorithm is a multivariate and nonlinear operation process. The calculation is complex and the results of each operation are slightly different, which exposes the problem of premature convergence of genetic algorithm. Multi-population genetics is an intelligent algorithm that optimizes and searches for multiple populations simultaneously according to the improved genetic algorithm in the previous section. Different populations are given different control parameters (crossover probability P c , mutation probability P m ), achieving different search objectives. Populations are connected through immigration operators, enabling multi-population co-evolution. The optimal individuals from each generation of evolution of each population are preserved through artificial selection operators, and the number of generations in which the optimal individuals are retained is used as the basis for judging the algorithm's convergence. Therefore, its optimal solution is the combined result of the co-evolution of multiple populations, and its solution is more accurate than that of the improved genetic algorithm.

[0082] The improved multi-population genetic algorithm enables the co-evolution of multiple populations, enabling the algorithm to quickly find the global optimal solution in a complex solution space. This improved multi-population genetic algorithm significantly improves the accuracy and efficiency of force generation, addressing the problems of traditional genetic algorithms that often fall into local optimality and slow convergence in complex combat environments, thereby improving the quality and response speed of combat decision-making.

[0083] 2. Establishment of the Combat Constraint Model

[0084] 1. Combat Effectiveness Model

[0085] Troop deployment should be determined flexibly and scientifically based on combat missions, enemy situation, our own situation, battlefield environment, and tactical requirements to ensure maximum combat effectiveness. When combat missions and enemy situation are relatively clear, the combat effectiveness model of troop deployment can be expressed as

[0086]

[0087] Among them, F is the sum of the combat effectiveness of the participating forces; α i is the impact factor of battlefield environment on the deployment of the i-th force and weapon; r i is the weight index of the ith war zone; Q i is the combat effectiveness index of the i-th type of troops and weapons; m is the total number of troops and weapons.

[0088] Q i =[lnC+ln(∑D+1)+lnE]ξ1ξ2ξ3

[0089] where Q i is the ship combat effectiveness index; C is the maneuverability parameter; D is the firepower parameter; E is the target detection capability parameter; ξ1 is the survivability coefficient; ξ2 is the range coefficient; ξ3 is the electronic countermeasure capability coefficient.

[0090] (1) Mobility parameters

[0091]

[0092] Among them, n zatr is the maximum permissible overload; SEP is the turning rate, m / s.

[0093] (2) Firepower parameters

[0094] ∑D=D G +D1+D2+D3

[0095] Among them, ∑D is the total firepower parameter of the fighter; D G is the firepower parameter of the ship's gun; D1 is the parameter of a mounted ballistic missile; D2 is the parameter of another mounted cruise missile; D3 is the parameter of another mounted gliding missile.

[0096] D G Calculate according to the following formula

[0097]

[0098] Among them, rpm is the firing rate per minute, that is, the firing rate; V 0b is the initial velocity of the projectile, m / s; G b is the weight of the projectile, g; D gb is the caliber of the naval gun (machine gun), in mm; n is the number of naval guns (machine guns) configured.

[0099] D1 is calculated as follows

[0100]

[0101] Among them, R f is the maximum practical effective range, km; ΔH f To allow the total launch height difference, km; P k is the single-shot kill probability; n zmmax . is the maximum overload of the missile.

[0102] The calculation method of D2 and D3 is the same as D1.

[0103] (3) Detection capability parameters

[0104] Only consider the performance parameters of shipborne radar

[0105]

[0106] Among them, L d is the maximum detection distance, km; m1 is the number of targets tracked simultaneously; m2 is the number of targets allowed to be attacked simultaneously.

[0107] (4) Viability coefficient

[0108]

[0109] Among them, L w is the ship width, m; L F is the total length of the ship, in meters; E is the performance parameter of the shipborne radar.

[0110] (5) Range coefficient

[0111]

[0112] Among them, L maxibf is the maximum range corresponding to the fuel in the aircraft, km.

[0113] 2. War Zone Weight Assessment Model

[0114] The key to regional air defense operations is to use limited air defense forces to protect multiple war zones of varying importance. Therefore, it is extremely important to judge the regional types and rank the importance of the regions.

[0115] According to the characteristics of combat air defense, the main factors that can be used to measure the combat area are: military value, political value, economic value and social value. The evaluation sets are: extremely important, very important, important and general.

[0116] Let A be the set of factors for evaluating the importance of a war zone: A = {military value (a1), political value (a2), economic value (a3), social value (a4)}; let B be the set of evaluation results: B = {extremely important (b1), very important (b2), important (b3), general (b4)}. By statistically obtaining the evaluation ratio of each factor of the war zone's importance, the single factor evaluation matrix C (c ij ), where c ij Indicates that the evaluated war zone is from factor a i , let's look at the evaluation set b j The membership degree is i=1,2,3,4,j=1,2,3,4.

[0117] Let D be the weight vector of each factor: D = (d1, d2, d3, d4). In this problem, the weight of each factor in the factor set is determined by the expert judgment method as D = (0.6, 0.2, 0.1, 0.1). The comprehensive evaluation result vector is obtained by weighted average: E = D × C = (e1, e2, e3, e4), where e i Indicates that after considering all factors, the evaluated battlefield is evaluated against the evaluation set b i The degree of membership of each level in.

[0118] In order to obtain the weight index of the theater through the evaluation result vector E, let F = (1, 0.8, 0.4, 0.2) be the weight of each evaluation result in the evaluation set B, and the weight index of the evaluated theater is obtained by weighted average: r = F·ET

[0119] Finally, the importance indexes of the n combat areas are normalized between [0, 1] to obtain the weight distribution of the combat zones: R = [r1, r2, ..., r n ],satisfy Among them, r i is the weight index of the i-th war zone.

[0120] 3. Factors affecting battlefield environment

[0121] The battlefield environment mainly includes terrain, meteorology, hydrology, and electromagnetics, which have different degrees of impact on military operations. A 3D real-world map can not only directly obtain geographic coordinates (x, y) and elevation (H), but can also measure and query attribute information that affects troop deployment, such as water flow direction and velocity (D), meteorological information (P), visibility (S), traffic (T), and protection (F). That is, the terrain attributes of a certain location can be expressed as

[0122] R i (X i ,Y i ,H i ,D i ,P i ,S i ,T i ,F i )

[0123] Use This value represents the influence of hydrological information, average slope, visibility, accessibility, and protection on combat effectiveness. The value indicates the degree of influence, with larger values ​​indicating a greater impact. The specific value can be comprehensively evaluated based on statistical data.

[0124] Therefore, the comprehensive factors affecting the battlefield environment on troop deployment are:

[0125] α i =1-(1-α Pi )×(1-α Di )×(1-α Si )×(1-α Ti )×(1-α Fi )

[0126] In summary, through precise constraint modeling, the force generation problem can be transformed into a multi-objective optimization problem. This technology simplifies the cost of force generation and simultaneously assigns appropriate weights to various elements of force generation, significantly improving the accuracy and rationality of force deployment. It ensures an optimal balance among multiple factors, such as tactical effectiveness, cost-effectiveness, and enemy threat, making operational decisions more scientific and effective.

[0127] III. Natural Language Processing

[0128] The overall framework of the domain word segmentation scheme is mainly divided into three modules: the CRF (Conditional Random Field) coarse segmentation module, the domain dictionary construction model, and the word segmentation model prediction module.

[0129] In feature selection, this patent selects and determines the following three types of features: character n-gram features, character category (word type) features, and word position marking features. Characters themselves are also a good feature for the CRF model. Commonly used character n-gram features include 1-gram features, 2-gram features, 3-gram features, etc. The larger the value of n, the more accurate and effective the sequence information extracted, but if n is too large, it is easy to cause the extracted feature data to be sparse. Therefore, the commonly used n-gram model in CRF generally takes 2-gram.

[0130] Texts in the military field are generally processed in a standardized manner, with relatively fixed character formats and usage methods, and relatively simple character categories. The character category refers to the word class to which the character belongs. According to the analysis of text characteristics, the character types in military field texts are relatively common and are generally similar to those in relatively regular corpora such as news corpora in terms of character categories. When training the CRF model, this character category is used as a feature, which is equivalent to the word class feature to which the character belongs.

[0131] The terms in the military field vary in length. Some are short, such as "base indicator", "volume collation", "zero error situation", etc., and some are long, such as "military equipment", "military information system", etc. These longer words are generally between 4 and 6 characters in length. When choosing the word position marking method, it is preferable to use a longer word position marking method, which has a higher segmentation accuracy for longer words. To compare the actual word segmentation effects of different word position markings, this article selects two word position marking methods: the 4-tag marking method (B, M, E, S) and the 6-tag marking method (B, Mi, M, E, S) (i = 1, 2,...).

[0132] In the CRF model, the feature template is the main basis for the model to extract various features. Different complexities of the formulated feature templates result in different numbers of feature functions, which in turn affect the training and testing efficiency of the CRF word segmentation model, and also affect the accuracy.

[0133] After completing the feature selection and definition of CRF and the formulation of the feature template, the labeled training corpus can be used to train the CRF word segmentation model.

[0134] 1) Expressions with specific meanings in the military field, such as words like "navy", "operations section", etc., can appear as an entity concept. Although they can be more finely segmented into "sea / army / " and "operations / section / ", this will make them two words without entity meanings.

[0135] 2) Abbreviations unique to the military field, such as "base / zhi," "joint / zhi," "three / three / four / determination," and "Eagle Strike-10." Directly splitting these words into single-word expressions, such as "base / zhi," "joint / zhi," "three / three / four / determination," and "Eagle Strike-10," will destroy the meaning of the original words.

[0136] 3) New concepts and terms in the military field, such as "'Four Have' Soldiers," "Air-Sea Integrated Battle," "Electromagnetic Spectrum Warfare," etc., can be added to the domain dictionary when the training corpus cannot be updated in a timely manner, so that the word segmentation model can accurately segment these new terms.

[0137] Compared with the word segmentation training corpus, there are many ways to obtain domain dictionaries. The domain dictionary for military text word segmentation constructed in this paper mainly comes from the following two aspects:

[0138] The first is to use term extraction tools to process a large amount of military corpus, and then form a domain dictionary through manual rule screening.

[0139] The second method is direct manual collection and addition, focusing on the latest weapon and equipment names, new concepts, and new terms that continue to emerge in the military field. Since these are newly defined terms, they appear relatively infrequently in the corpus, and extraction tools may not be able to effectively identify them. To ensure the timeliness of dictionary updates, direct manual addition can be used.

[0140] 1. Part-of-speech tagging based on attention LSTM

[0141] (1) Long Short-Term Memory Network

[0142] As a special type of recurrent neural network (RNN), LSTM effectively addresses the vanishing gradient problem of conventional RNNs by introducing memory cells and gates within each hidden layer unit to control the input and output flow of information. LSTM is more adept at processing sequential data such as natural language text and can model a wide range of contextual information within a sequence.

[0143] In the internal structure of the LSTM unit, there are three types of control gates used to control the input and output of information flow, each of which plays a different control role:

[0144] Input gate i is used to control the amount of information updated in the memory unit;

[0145] The forget gate f is used to control how much information of the memory unit at the previous moment is used;

[0146] The output gate o is used to control the amount of information output to the next hidden state.

[0147] Through the control of the above three gates and the storage function of the memory unit, the information inside the LSTM unit can be selectively passed.

[0148] For any time t, let the input vector x at that time be t , the hidden state vector at the previous moment is h t-1 LSTM can calculate h through internal information circulation and update t Hidden state at all times:

[0149] i t =σ(U i x t +W i h t-1 +b i )

[0150] f t =σ(U f x t +W f h t-1 +b f )

[0151] o t =σ(U o x t +W o h t-1 +b o )

[0152]

[0153] The meaning of each parameter is as follows:

[0154] c t Represents the status information of the memory unit;

[0155] Parameter matrix U i ,W i ,U f ,W f ,U o ,W o ,U c ,W c The weight matrices corresponding to different gates;

[0156] Parameter vector b i ,b f ,b o ,b c are the corresponding deviation vectors respectively;

[0157] Function σ and They are sigmoid and tanh functions respectively, and the symbol ⊙ represents the point-by-point multiplication operation between vectors.

[0158] The parameter update method of the LSTM network is similar to that of ordinary neural networks. At each moment, the backpropagation algorithm and the derivative chain rule are used to solve the partial derivatives of the objective function with respect to each parameter. This is called the BPTT algorithm, and the gradient descent method is still used to update the parameters.

[0159] In a standard LSTM network, information flows unidirectionally across time steps. The LSTM can only utilize information from past moments through reverse iterations, but cannot utilize information from future moments. In current common LSTM models, two LSTM layers with opposite time step directions are typically constructed in the hidden layer: a forward LSTM layer and a backward LSTM layer, thus forming a bidirectional LSTM (BLSTM).

[0160] The hidden layer consists of a forward LSTM and a backward LSTM, which are used to calculate the forward hidden state and the backward hidden state, respectively, which are then projected to the common output layer. Compared to unidirectional LSTM, because information in the bidirectional LSTM hidden layer flows in two opposite directions, the iterative calculation of the hidden unit at any moment can simultaneously obtain forward and backward historical information, modeling the previous and next contexts, making it more effective in feature extraction and representation of serialized data.

[0161] (2) Attention Mechanism

[0162] The attention mechanism is an effective probability weight assignment mechanism. By calculating attention probability weights at different moments, it allows word nodes that are highly relevant to the target word's annotation to receive greater attention and, consequently, greater probability weights, thereby optimizing the quality of the hidden layer's feature vectors. By assigning different probability weights to hidden units in a neural network, the attention mechanism enables the hidden layer to focus on feature information that is more beneficial to the classification task while reducing attention to redundant information. Thus, within the same context sequence, hidden layers incorporating the attention mechanism can effectively capture specific contextual information, further optimizing the quality of the extracted features.

[0163] Let the network input vector be x i , the hidden state vector is h i , then in the network model with the attention mechanism, the new hidden state vector s is composed of the initial hidden state vector h at each moment i Jointly decided, the calculation formula is as follows:

[0164]

[0165] Among them, α i represents the initial hidden state h iThe weights relative to the new hidden layer are calculated as follows:

[0166]

[0167] e i =vtanh(wh i +b)

[0168] Among them, e i The energy value of the hidden state at the i-th moment is mainly composed of the hidden state vector h at that moment i w and v are weight matrices, and b is the corresponding bias term. The process corresponding to the formula realizes the transformation from the initial hidden layer to the new attention layer. The weight coefficient α corresponding to the hidden layer at each moment i Reflects its influence on the current output.

[0169] 2. Part-of-speech tagging

[0170] (1) Magnetic labeling model

[0171] The part-of-speech tagging model incorporates an attention mechanism based on BLSTM. It consists of three main parts: the input layer, the hidden layer, and the output layer. The hidden layer consists of a unidirectional LSTM layer, a bidirectional LSTM layer, and an attention layer, which are described below.

[0172] 1) Input layer. The word vector matrix M is formed through pre-training, and each word is indexed in the word vector matrix to convert it into its corresponding word vector form x t , as the input of the BLSTM network.

[0173] 2) Hidden layer. The calculation is mainly divided into three steps:

[0174] The unidirectional LSTM contains only one hidden layer in one direction, and the input vector x is t and the hidden state vector h at the previous moment t-1 , calculate the current hidden state h t The bidirectional LSTM contains a forward layer and a reverse layer, and the hidden state vector of the forward layer at the current moment needs to be calculated separately. and the reverse hidden state vector

[0175]

[0176] Where m is the hidden unit dimension. The LSTM() function represents the nonlinear transformation of the LSTM network.

[0177] Step 2: Calculate the BLSTM hidden layer based on the LSTM forward and reverse hidden states.

[0178] The forward hidden state vector is summed up by weighted summation. and the reverse hidden state vector Perform linear combination to obtain the hidden layer vector h of BLSTM t ∈R m×1

[0179]

[0180] Step 3: According to the attention mechanism, assign probability weights to the BLSTM hidden layer and calculate the new attention hidden layer. Since BLSTM contains forward and reverse layers, it is necessary to consider the forward hidden state at the same time. and reverse hidden state In this paper, we use the aggregated hidden state vector h t , calculate the hidden layer energy e at that moment t :

[0181] e t =V2tanh(W2h t +b2)

[0182] Then, according to the energy value of the hidden state vector at each moment, the attention probability weight corresponding to the hidden state at that moment is calculated:

[0183]

[0184] Finally, the hidden state and the corresponding probability weight at each moment are multiplied and accumulated to obtain the new attention hidden layer vector s t ∈R m×1 :

[0185]

[0186] 3) Output layer. Use the sofmax function to calculate the label probability distribution on the time annotation set:

[0187] y t =softmax(W3s t +b3)

[0188] For the output vector y t The kth (k=1,2,…,L) dimension (y t ) k , and its calculation formula is:

[0189]

[0190] Let the output vector y t for:

[0191] y t =(y1,y2,...,y L )

[0192] (2) Training model

[0193] Given a training set where x (i) Represents the i-th sentence in the training set Indicates the part-of-speech tag sequence corresponding to the sentence n represents the sentence length, and N represents the number of samples in the training set. In model training, the log-likelihood loss function is used, plus the L2 regularization term. The objective function is specifically defined as follows:

[0194]

[0195] During the training process, the mini-batch gradient descent method is used, and the batch size is set to k.

[0196] (3) Tag sequence decoding

[0197] BLSTM outputs the probability of the part-of-speech tag sequence at each moment, giving a probability distribution based on the tag set. A typical direct decoding method searches directly from the output classification probabilities to obtain the tag sequence with the highest score.

[0198] For the training set Where the i-th sentence x (i) The corresponding part-of-speech sequence is y (i) The score of the labeled sequence is:

[0199]

[0200] The goal is to find the maximum scoring sequence from this set:

[0201]

[0202] y∈Y(x (i) )

[0203] The structured loss function is defined as follows:

[0204]

[0205] In summary, the use of natural language processing technology enables the system to analyze the operational intent and specific mission requirements contained in commanders' orders, including but not limited to various combat types such as attack, defense, reconnaissance, and patrol. Therefore, when establishing a combat constraint model, relevant parameters can be adjusted to create a constraint model that best meets the commander's operational needs, providing guidance for subsequent force generation.

[0206] The preferred embodiments of the present invention have been described in detail above. It should be understood that numerous modifications and variations based on the concepts of the present invention are possible without inventive effort by those skilled in the art. Therefore, any technical solution that can be derived by one skilled in the art through logical analysis, reasoning, or limited experimentation based on the concepts of the present invention and the prior art should be within the scope of protection defined by the claims.

Claims

1. A method for generating surface platform combat forces based on natural language processing and intelligent optimization algorithm, characterized in that: Combining intelligent search algorithms with optimization strategies, the improved multi-population genetic algorithm utilizes multiple populations for parallel optimization search, and implements diversified search strategies by assigning different control parameters to different populations. Populations exchange information through immigration operators to promote co-evolution, and by retaining the best individual in each generation and using the best individual to maintain algebraic convergence, the correctness of the final solution and global optimality are ensured. The improved multi-population genetic algorithm comprises the following steps: Step 1: Encode to form the initial population. Use integer permutation encoding method. There are m pre-defined positions. The chromosome is divided into m segments. Each segment corresponds to the number of a deployment position. If the number of troops and weapons n is less than m, then the first n positions are used to deploy troops and weapons, and the last mn positions are used for virtual deployment of troops and weapons. Step 2: Calculate the fitness. |k1|k2|k3|…|km| is a chromosome, representing a deployment plan. Fi is the combat effectiveness corresponding to the plan. is the sum of the combat effectiveness of the population, and the fitness of the i-th deployment plan is ; Step 3: Selection operation. Individuals with higher fitness, i.e., individuals with higher combat effectiveness, are selected from the population through the selection operator. The greater the fitness, the greater the probability of being selected, and the more offspring are produced in the next generation. These selected individuals are then used to reproduce the next generation to produce a new population. The probability of individual selection is calculated using the roulette wheel algorithm. The probability of each individual being selected is proportional to its fitness. The higher the fitness, the greater the probability of being selected. Step 4: Crossover operation: determine whether each deployment scheme in the population crosses according to the crossover probability Pc, group the parent samples for the crossover operation into two groups, and randomly determine the crossover position for each group to perform the crossover operation; Step 5: Evolutionary mutation operation: determine whether each deployment plan in the population is mutated based on the mutation probability Pm, then randomly determine two positions, swap their positions, and then calculate the combat effectiveness of each deployment plan of the new population after the mutation. If the combat effectiveness is improved, the mutation is effective, otherwise it is invalid. Step 6: To improve local search capabilities, a reversal operation is performed after selection, crossover, and mutation. If the combat effectiveness of the new individual improves, the evolutionary reversal is successful and the next cycle begins. If the combat effectiveness of the new individual decreases, the original individual is restored. That is, the reversal is only accepted if the combat effectiveness improves after reversal; otherwise, the reversal is invalid. This process is called evolutionary reversal, and the operation process is the same as evolutionary mutation. Step 7: Evolutionary reinsertion: First, reinsert the new population obtained after reversal into the parent population. Then, randomly replace the individuals with the highest combat effectiveness in the population after evolutionary reversal with the individuals in the reinserted population, thereby increasing the number of optimal individuals in each generation and improving the evolution rate. Step 8: Improve the genetic algorithm and use steps 1-7. Different populations are assigned different control parameters to achieve different search purposes. The populations are connected through immigration operators to achieve multi-population co-evolution. The optimal individuals in each generation of evolution of each population are saved through artificial selection operators. The number of generations of optimal individuals is used as the basis for judging the convergence of the algorithm. The modeling of constraint conditions mainly includes combat effectiveness model, battle zone weight evaluation model and battlefield environment influencing factors; The combat effectiveness model can be expressed as Among them, F is the sum of the combat effectiveness of the participating forces; is the impact factor of the battlefield environment on the deployment of the i-th type of troops and weapons; ri is the weight index of the i-th theater; is the combat effectiveness index of the i-th type of troops and weapons; m is the total number of troops and weapons; in is the ship's combat effectiveness index; C is the maneuverability parameter; D is the firepower parameter; E is the target detection capability parameter; is the viability coefficient; is the range coefficient; is the electronic countermeasure capability coefficient.

2. The method for generating surface platform combat forces based on natural language processing and intelligent optimization algorithm according to claim 1, characterized in that: (1) Mobility parameters in, is the maximum permissible overload; SEP is the turning rate, m / s; (2) Firepower parameters in, is the total firepower parameter of the fighter; is the naval gun firepower parameter; Parameters of a mounted ballistic missile; Parameters of another cruise missile mounted; Parameters of another type of gliding missile mounted; Calculate according to the following formula in, is the rate of fire per minute, i.e. the rate of fire; is the initial velocity of the projectile; is the weight of the projectile; is the caliber of the ship's gun; n is the number of ship's guns; D1 is calculated as follows: in, is the maximum practical effective range; To allow for the total height difference of the launch; is the single-shot kill probability; is the maximum overload of the missile; The calculation method of D2 and D3 is the same as D1; (3) Detection capability parameters Only consider the performance parameters of shipborne radar in, is the maximum detection distance; The number of targets tracked simultaneously; The number of targets allowed to be attacked at the same time; (4) Viability coefficient in, is the ship's width; is the overall length of the ship; E is the performance parameter of the shipborne radar; (5) Range coefficient in, It is the maximum range corresponding to the oil in the aircraft.

3. The method for generating surface platform combat forces based on natural language processing and intelligent optimization algorithm according to claim 1, characterized in that: The force generation method should be determined flexibly and scientifically based on combat missions, enemy situation, our own situation, battlefield environment, and tactical requirements to ensure maximum combat effectiveness. A combat effectiveness model, a theater weighted recruitment model, and battlefield environment influencing factors should be established based on the red and blue ship equipment models. Mission requirements and terrain environmental factors should be accurately quantified into specific constraints to provide clear optimization boundaries and conditions for the algorithm.

4. The method for generating surface platform combat forces based on natural language processing and intelligent optimization algorithm according to claim 1, characterized in that: In the war zone weight evaluation model, the main factors used to measure the combat zone are: military value, political value, economic value and social value, and the evaluation sets are: extremely important, very important, important and general; Let A be the set of factors for evaluating the importance of a war zone: A={military value , political value , economic value , social value }; Let B be the set of judgment results: B={extremely important Very important ,important ,generally }; By statistically obtaining the proportion of each factor in the importance of the war zone, a single factor evaluation matrix is ​​obtained ,in Indicates the factors of the evaluated war zone , let's look at the evaluation set Membership degree, i=1,2,3,4, j=1,2,3,4; Let D be the weight vector of each factor: D=( , , ), in this problem, according to the expert judgment method, the weight of each factor in the factor set is determined to be D=(0.6, 0.2, 0.1, 0.1), and the comprehensive evaluation result vector is obtained by weighted average: E=D×C=( , , ),in, Indicates that after considering all factors, the evaluated war zone has a positive impact on the evaluation set. The degree of membership of each level in In order to obtain the weight index of the theater through the evaluation result vector E, let F = (1, 0.8, 0.4, 0.2) be the weight of each evaluation result in the evaluation set B, and the weight index of the evaluated theater is obtained by weighted average: ; Finally, the importance indexes of the n combat areas are normalized between [0, 1] to obtain the weight distribution of the combat zones: ,satisfy ,in, is the weight index of the i-th war zone.

5. The method for generating surface platform combat forces based on natural language processing and intelligent optimization algorithm according to claim 1, characterized in that: The battlefield environmental factors mainly include terrain, meteorology, hydrology, and electromagnetics, which have different degrees of influence on military operations; the three-dimensional real-life map can not only directly obtain geographic coordinates (x, y) and elevation (H), but also measure and query the attribute information of water flow direction and velocity D, meteorological information P, visibility S, accessibility T, and protection F that affect the deployment of troops. That is, the terrain attributes of a certain location can be expressed as Use 、 、 、 、 It indicates the influence coefficient of hydrological information, average slope, visibility, trafficability, and protection capability on combat effectiveness. The value indicates the degree of influence on combat effectiveness. The larger the value, the greater the influence of the location on combat effectiveness. The specific value can be comprehensively evaluated based on data statistics. Therefore, It is expressed as the comprehensive impact factor of battlefield environment on troop deployment.

6. The method for generating surface platform combat forces based on natural language processing and intelligent optimization algorithm according to claim 5, characterized in that: The method performs part-of-speech tagging training based on a long short-term memory (LSTM) network with an attention mechanism to capture complex text context and part-of-speech features; by searching the output classification probabilities, the tagging sequence with the highest score is obtained, thereby improving the parsing accuracy of military terminology and the system's natural language understanding capabilities.

Images