Ground target threat level evaluation method based on IDM-PSO-ELM

Abstract

The invention provides a ground target threat level evaluation method based on an IDM-PSO-ELM, and the method comprises the steps: selecting a target distance, a target speed, a course angle, an effective range, firepower configuration, detection capability, electronic interference capability, communication capability and viability as threat indexes, representing the threat degree of a target through employing a target threat level, and constructing a sample database; constructing a ground target threat level evaluation model by adopting an ELM network, and completing ELM network training based on the constructed sample database by taking the threat index as input and the target threat level as output; a dynamic multi-population particle swarm algorithm is introduced, an input layer weight and a hidden layer deviation of an ELM network are taken as particles, a mean square error of a predicted value and an actual value of the ELM network is taken as a fitness function, a dominant group, a disadvantage group and a mixed group are divided, and the positions of the particles are updated by adopting Levy flight for the dominant group. And updating particle positions by adopting a variation mode for the disadvantaged group, and introducing sine and cosine learning factors for the mixed group to update the particle positions, thereby obtaining an optimal ELM network for evaluating the threat level of the ground target. The method has higher target threat prediction precision.

Description

Technical Field

[0001] The invention belongs to the field of command and control, and in particular relates to a ground target threat level assessment method based on IDM-PSO-ELM. Background Art

[0002] The importance of ground target threat level assessment in modern warfare is self-evident. It provides commanders with clear strike priorities through quantitative analysis of the threat level of ground targets (such as tanks, infantry fighting vehicles, artillery, etc.), thereby improving resource allocation efficiency and decision-making quality. Threat assessment can not only help determine the enemy's most threatening targets, but also provide a basis for dynamic adjustment of tactics. With the application of artificial intelligence and big data technology, modern threat assessment systems can accurately identify and predict enemy targets in real time, further improving the intelligence and scientific nature of combat decisions.

[0003] At present, ground target threat assessment mainly relies on rule-based methods, mathematical models and artificial intelligence technology. Rule-based methods rely on expert experience to set rules, which makes it difficult to process complex battlefield data and has strong subjectivity and limitations. Although statistical and mathematical model methods can process multidimensional data, most of them assume that the relationship between targets is linear, ignoring the nonlinearity and complexity in reality. In recent years, machine learning and deep learning technologies have been widely used in threat assessment, but they face problems such as insufficient training data, black box nature of models and poor interpretability, which affects the credibility of assessment results and the transparency of decision-making.

[0004] In order to overcome the shortcomings of the existing methods, the present invention proposes a method for optimizing the extreme learning machine (ELM) based on the improved particle swarm algorithm (IDM-PSO). The method introduces a dynamic multi-population strategy into the traditional PSO method, divides the initial particle swarm into a dominant group, an inferior group and a mixed group, and adopts different speed update strategies for different populations. The position of the dominant group particles is updated by Levy flight and greedy algorithm; the particle speed and position of the inferior group are updated by Gaussian mutation and mixed particles; the speed of the mixed group particles is updated by sine and cosine learning factors, and the initial input weights and hidden layer bias of the ELM are optimized by the improved PSO algorithm. This new technology can provide fast and accurate threat assessment results for dynamically changing battlefield environments, and support the intelligence and efficiency of combat decisions. Summary of the invention

[0005] The object of the present invention is to provide a target threat assessment method based on IDM-PSO-ELM (Improved Dynamic Multi-swarm-Particle Swarm Optimization-Extreme Learning Machines).

[0006] The technical solution to achieve the purpose of the present invention is: a ground target threat level assessment method based on IDM-PSO-ELM, comprising the following steps:

[0007] Step 1: Select target distance, target speed, heading angle, effective range, firepower configuration, detection capability, electronic jamming capability, communication capability, and survivability as threat indicators, use target threat level to characterize the threat degree of the target, and build a sample database;

[0008] Step 2: Use the ELM network to build a ground target threat level assessment model, take the threat index as input and the target threat level as output, and complete the ELM network training based on the constructed sample database;

[0009] Step 3, introduce the dynamic multi-swarm particle swarm algorithm, take the input layer weights and hidden layer deviations of the ELM network as particles, take the mean square error between the predicted value and the actual value of the ELM network as the fitness function, divide the group into dominant group, inferior group and mixed group, use Levy flight to update the particle position for the dominant group, use mutation to update the particle position for the inferior group, and introduce sine and cosine learning factors to update the particle position for the mixed group, thereby obtaining the optimal ELM network for ground target threat level assessment.

[0010] Furthermore, in step 1, target distance, target speed, heading angle, effective range, firepower configuration, detection capability, electronic jamming capability, communication capability, and survivability are selected as threat indicators, and the target threat level is used to characterize the threat degree of the target, and a sample database is constructed, wherein:

[0011] The size of the target distance directly reflects the attack intention of the enemy combat unit. The shorter the distance, the greater the threat. The target distance is defined as {far, medium, close}.

[0012] The greater the target speed, the greater the threat to us. The target speed is defined as {fast, medium, slow}.

[0013] The smaller the heading angle of the enemy target, the greater the threat to our target. The heading angle is defined as {large, medium, small}.

[0014] The same type of combat platform may be equipped with different firepower units. The firepower threat of the platform is analyzed based on the number and type of weapons carried by the platform, and the firepower configuration is defined as {strong, relatively strong, general, relatively weak, weak};

[0015] The detection capability of enemy targets reflects the reconnaissance of our combat power configuration. The stronger the detection capability, the greater the threat to us. The detection capability is defined as {strong, relatively strong, average, relatively weak, weak}.

[0016] Electronic jamming capability refers to the ability of the enemy's combat platform to interfere with our various electronic equipment. The stronger the electronic jamming capability, the greater the threat of the enemy's target to us. The electronic jamming capability is defined as {strong, relatively strong, general, relatively weak, weak} five states;

[0017] Communication capability refers to the ability of communication, coordination and cooperation between target platforms, between platforms and reconnaissance and guidance equipment, and between platforms and command and control systems. Communication capability is defined as {strong, relatively strong, general, relatively weak, weak}.

[0018] Survivability refers to the ability of the target platform to resist detection, interference, interception and destruction. Survivability is defined as five states: {strong, relatively strong, average, relatively weak, weak}.

[0019] Furthermore, in step 1, target distance, target speed, heading angle, effective range, firepower configuration, detection capability, electronic jamming capability, communication capability, and survivability are selected as threat indicators, and the target threat level is used to characterize the threat degree of the target, and a sample database is constructed, wherein:

[0020] The threat level of the target is divided into six levels: low, lower, medium, high, higher, and very high, represented by numbers 1 to 6.

[0021] Furthermore, in step 2, the ELM network is used to construct a ground target threat level assessment model, with threat indicators as input and target threat level as output, and the ELM network training is completed based on the constructed sample database. The specific method is:

[0022] Assume that the i-th sample data is (x i ,t i ), where x i =[x i1 ,…,x im ] T , x i ∈R n ;t i =[t i1 ,…,t im ] T , t i ∈R n , x i represents the i-th ground target; x im represents the mth feature of the i-th ground target; t i Represents the threat level of the i-th target; for the ELM network consisting of L hidden layer nodes, it is expressed as:

[0023]

[0024] Where b irepresents the bias of the i-th hidden layer node; o j represents the activation output of the i-th target input at the j-th hidden layer node; w i represents the input weight; w i ·x j represents the inner product of the input weight and the target feature value; g(x) represents the activation function; β i represents the output weight;

[0025] The training goal of the ELM network is to minimize the output error, which can be expressed as:

[0026]

[0027] That is, there exists w i , β i and b i So that:

[0028]

[0029] It can be expressed as a matrix:

[0030] Hβ=T

[0031] In the formula, H represents the calculation output matrix of the hidden layer of the ELM network, where the i-th column corresponds to the output result of the i-th hidden layer node; β represents the output weight matrix; T represents the expected output matrix, that is, the threat level of the output target;

[0032] Expand to get:

[0033]

[0034] in:

[0035]

[0036] l represents the number of hidden layer nodes in the ELM network, and m represents the number of target threat level categories;

[0037] Train the ELM network to get the optimal solution and So that the following is true:

[0038]

[0039] This is equivalent to minimizing the loss function as follows:

[0040]

[0041] Solve the linear equation system Hβ=T and get the optimal solution of the loss function for:

[0042] Furthermore, in step 3, a dynamic multi-swarm particle swarm algorithm is introduced, with the input layer weights and hidden layer deviations of the ELM network as particles, and the mean square error between the predicted value and the actual value of the ELM network as the fitness function. The dominant group, the inferior group and the mixed group are divided, and the Lévy flight strategy is used to update the particle position for the dominant group, and the mutation method is used to update the particle position for the inferior group. The sine and cosine learning factors are introduced for the mixed group to update the particle position. Based on this, the optimal ELM network is obtained for ground target threat level assessment. The specific method is as follows:

[0043] Step 3.1, population initialization: Take the mean square error between the predicted value and the actual value of the ELM network as the fitness function and calculate the initial fitness value f(1) ;

[0044] Step 3.2, divide the population according to the fitness value: calculate the fitness value of all particles, and sort them according to the fitness value. The first N / 2 particles with smaller fitness values ​​are "dominant particles", and the last N / 2 particles with larger fitness values ​​are "inferior particles". Then, randomly select N / 6 particles from the "dominant particles" and "inferior particles" to form a population, called a "mixed group". The remaining "dominant particles" are called "dominant groups", and the remaining "inferior particles" are called "inferior groups". The particles are sorted and regrouped every time a new generation is updated.

[0045] Step 3.3, update the position and velocity of various swarm particles:

[0046] (1) Dominant group

[0047] The formula for updating the particle position using Levy flight is as follows:

[0048]

[0049] in, and x g (k) are the global optimal particle position and the global optimal particle position of the population after Levy flight update at the kth iteration, α is the step size control factor, which is 0.01, · represents the dot product, Levy(χ) is the random search path, and the flight trajectory is simulated by the Mantegna algorithm. Its mathematical expression is shown in the formula:

[0050]

[0051] Among them, the parameter χ is a random value in the interval (0,2) and is set to 1.5; μ and υ both obey the normal distribution and are defined as follows:

[0052]

[0053] The variance σ in the above formula is μ and σv Determined by the following formula:

[0054]

[0055] Where, Γ is the gamma function;

[0056] The evaluation strategy of the greedy algorithm is introduced to determine whether to update the optimal particle position. That is, the position is updated only when the updated position is better than the original position, otherwise the original position is retained. The implementation process is shown in the formula:

[0057]

[0058] in, is the particle position after the greedy algorithm updates, and f(·) represents the particle fitness function;

[0059] (2) Disadvantaged Group

[0060] The particle position is updated by mutation, and the mutation judgment is as follows:

[0061]

[0062] GV(x)=x×(1+N(0,1))

[0063] Among them, r and N(0,1) are random numbers between [0,1], x is the original parameter value, GV(x) is the mutated value, k is the current iteration number, K max is the maximum number of iterations. When the judgment condition is established, the Gaussian mutation operation is performed on the particles.

[0064] The position and velocity of the disadvantaged group particles are updated as follows:

[0065] v id (k+1)=ω·v id (k)+c1·r1·(p id (k)-x id (k))+c2·r2·(p gd (k)-x id (k))+c3·r3·(p mix (k)-x id (k))

[0066] x id (k+1)=x id (k)+ν id (k+1)

[0067] Among them, v id (k+1) is the velocity of the k+1th generation of the disadvantaged group particle, ω is the inertia weight, r1 and r2 are random numbers uniformly distributed between [0,1], and pid (k) is the historical best position of particle i in the previous k iterations, x id (k) is the position information of the particle at the kth iteration, p gd (k) is the global optimal position of the particle of the entire disadvantaged group in the first k iterations. c1 is the cognitive coefficient, c2 is the social coefficient, c3 is the hybrid learning factor, and r3 is a random number in the interval (0,1);

[0068] (3) Mixed group

[0069] The cosine learning factor and the sine learning factor are introduced to make the self-learning factor monotonically decrease and the population learning factor monotonically increase. The particle velocity update formula of the mixed group obtained in this way is:

[0070]

[0071] Step 3.4, update the local optimal position Pbest and the global optimal position Gbest of the particle;

[0072] Step 3.5, repeat steps 3.2 to 3.4 until the number of iterations k reaches the maximum;

[0073] Step 3.6, output the optimal input weight w and hidden layer bias b: repeat the iteration k times until convergence, minimize the mean square error between the predicted value and the actual value of the ELM network, that is, output the input layer weight and hidden layer bias when the fitness function is minimized, to form the optimized optimal ELM network;

[0074] Step 3.7, based on the optimal ELM network, evaluate the threat level of ground targets and obtain the target threat level evaluation results corresponding to each target.

[0075] Furthermore, the inertia weight adopts nonlinear inertia weight, and its weight factor is set to be larger in the early stage of the search; in the later stage of the search, its weight factor is set to be smaller, which is expressed as follows:

[0076]

[0077] Among them, ω max -ω min It represents the maximum inertia weight minus the minimum inertia weight, and its purpose is to control the variation of the inertia weight.

[0078] A ground target threat level assessment system based on IDM-PSO-ELM implements the ground target threat level assessment method based on IDM-PSO-ELM, realizes ground target threat level assessment based on IDM-PSO-ELM, and is divided into three modules to respectively perform steps 1 to 3.

[0079] A computer device comprises a memory, a processor and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the method for assessing the threat level of a ground target based on IDM-PSO-ELM is implemented to achieve the assessment of the threat level of a ground target based on IDM-PSO-ELM.

[0080] A computer-readable storage medium stores a computer program. When the computer program is executed by a processor, the ground target threat level assessment method based on IDM-PSO-ELM is implemented to achieve ground target threat level assessment based on IDM-PSO-ELM.

[0081] Compared with the prior art, the present invention has the following significant advantages: the ELM model is used to predict the threat level of battlefield targets, and the improved dynamic multi-population particle swarm algorithm is introduced to optimize the input weights and hidden layer deviations of the ELM. The algorithm model has a high prediction accuracy when identifying the target threat level, and the model has good real-time performance and generalization ability. BRIEF DESCRIPTION OF THE DRAWINGS

[0082] Figure 1 It is a general schematic diagram of the model of the ground target threat level assessment method based on IDM-PSO-ELM of the present invention.

[0083] Figure 2 This is a schematic diagram of the ELM network structure.

[0084] Figure 3 It is a schematic diagram of the mixed particle combination.

[0085] Figure 4 It is a schematic diagram of nonlinear inertia weight.

[0086] Figure 5 It is a flowchart of the ELM optimization process using the improved PSO algorithm.

[0087] Figure 6 It is a comparison chart of evolution curves.

[0088] Figure 7 It is the confusion matrix of simulation results of the improved model.

[0089] Figure 8 This is a comparison chart of algorithm improvements. DETAILED DESCRIPTION

[0090] The present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0091] Combination Figure 1The present invention is based on the ground target threat level assessment method of IDM-PSO-ELM, and adopts the ELM network to realize the data-based assessment method of target threat assessment to meet the real-time requirements of battlefield operations; introduces an improved dynamic multi-population particle swarm algorithm to optimize the ELM algorithm, uses the particle swarm algorithm to select the optimal weights and hidden layer deviations, and improves the global exploration and local development capabilities of the PSO algorithm through a multi-population strategy, so that the model has a more accurate target threat assessment result. The specific steps are as follows:

[0092] Step 1: Sample database construction

[0093] In order to accurately describe the threat characteristics of modern battlefields, the radar detection capabilities, equipment performance and intelligent combat characteristics of enemy main battle tanks, infantry fighting vehicles, armored vehicles and reconnaissance vehicles are comprehensively considered, and nine parameters including target distance, target speed, heading angle, effective range, firepower configuration, detection capability, electronic jamming capability, communication capability and survivability are selected as threat indicators to construct a ground target sample database;

[0094] 1) Target distance. The size of the target distance directly reflects the enemy's mobile vehicle's attack intention. The shorter the distance, the greater the threat. The target distance is defined as {far, medium, close}.

[0095] 2) Target speed. The greater the target speed, the greater the threat to our defense platform. The target speed is defined as {fast, medium, slow}.

[0096] 3) Heading angle. The smaller the heading angle of the enemy vehicle target, the greater the threat to us. The heading angle is defined as {large, medium, small}.

[0097] 4) Effective range. The effective range of a weapon refers to the distance at which the enemy's tank target can achieve the expected accuracy and power requirements when threatening to damage our side. The effective range is defined as {far, medium, and close}.

[0098] 5) Firepower configuration. Tank-type targets of the same type may be equipped with different firepower units, thus having different attack capabilities. Therefore, the firepower threat of tank-type targets can be analyzed based on the number and type of weapons carried by the tank-type targets, and the firepower configuration can be defined as {strong, relatively strong, general, relatively weak, weak}.

[0099] 6) Detection capability. In the face of electronic warfare, the enemy's detection capability reflects its reconnaissance of our combat force configuration, so as to form a new combat force deployment for our combat. The stronger its detection capability, the greater the threat to us. The detection capability is defined as {strong, relatively strong, general, relatively weak, weak}.

[0100] 7) Electronic jamming capability. Electronic jamming capability refers to the ability of enemy reconnaissance vehicles to interfere with our various electronic equipment and weapon systems. The stronger the jamming capability, the greater the threat to us. The electronic jamming capability is defined as {strong, relatively strong, general, relatively weak, weak}.

[0101] 8) Communication capability. Communication capability refers to the ability of enemy main battle tanks, tanks and reconnaissance vehicles, and tanks and command and control centers to communicate and coordinate with each other. Communication capability is of great significance for coordinated destruction of enemy targets, efficient determination of targets, and timely change of combat intentions and targets. Communication capability is defined as {strong, relatively strong, general, relatively weak, weak}.

[0102] 9) Survivability. Survivability refers to the ability of enemy combat vehicles to resist detection, interference, interception and destruction by means of false targets, decoys, stealth and other means. Survivability is defined as {strong, relatively strong, average, relatively weak, weak}.

[0103] After the index system is built, each index needs to be quantified. The rule of quantification is to divide the index into qualitative and quantitative indexes, and then use the corresponding processing method to quantify it. Among the above 9 threat indicators, target distance, target speed, heading angle and effective range are quantitative indicators; firepower configuration, detection capability, electronic jamming capability, communication capability and survivability are qualitative indicators. For the standardization processing using the range method, the quantification method is divided into benefit-based index quantification and cost-based index quantification, as follows:

[0104]

[0105] Among them, r i represents the i-th indicator to be quantified, a i It represents the quantitative processing result of the i-th indicator. For qualitative indicators, the scaling method is used to quantify them, that is, the decision maker uses precise quantitative evaluation language based on experience and evaluation language scale to reflect the quality of the evaluation language.

[0106] When the traditional neural network predicts the target threat level, the target threat value is used as the final output result. However, the battlefield situation data is inevitably random and fuzzy. Therefore, the method of using precise data representation is not intuitive enough and is not convenient for subsequent commanders to make quick and accurate decisions. Therefore, the present invention uses the target threat level to represent the threat level of the target. Considering the diversity of battlefield information and combining the expert knowledge base, the threat level of the target is divided into 6 levels: low, relatively low, medium, high, relatively high, and very high, represented by numbers 1 to 6. The threat level of each sample is evaluated in turn, and the results are shown in Table 1.

[0107] Step 2: Construction of ELM model

[0108] The battlefield threat index quantification parameters are used as the input of the ELM network, and the target threat level network output is used as the output. Based on this sample, the initial ELM network parameters are determined (the number of input layer nodes is 9, the number of output layer nodes is 6, and the number of hidden layer nodes is 30), and the network is trained to calculate the initial network output layer weight value.

[0109] Assume that the i-th ground target sample is (x i ,t i ), x i =[x i1 ,…,x im ] T , and x i ∈R n ;t i =[t i1 ,…,t im ] T , and t i ∈R n In the formula, x i represents the i-th ground target; x im represents the mth feature of the i-th ground target; t i Represents the threat level of the i-th target. For the ELM network consisting of L hidden layer nodes, it can be expressed as follows:

[0110]

[0111] Where b i represents the bias of the i-th hidden layer node; o j represents the activation output of the i-th target input at the j-th hidden layer node; w i represents the input weight; w i ·x j represents the inner product of the input weight and the target feature value; g(x) represents the activation function; β i Represents the output weight. Minimizing the output error is the training goal of the ELM network, which can be expressed as:

[0112]

[0113] That is, there exists w i , β i and b i So that:

[0114]

[0115] It can be expressed as a matrix:

[0116] Hβ=T

[0117] In the formula, H represents the calculation output matrix of the hidden layer of the ELM network, where the i-th column corresponds to the output result of the i-th hidden layer node; β represents the output weight matrix; T represents the expected output matrix, that is, the threat level of the output target. Expanding the above formula, we can get:

[0118]

[0119] in:

[0120]

[0121] l represents the number of hidden layer nodes in the ELM network, and m represents the number of categories of target threat levels.

[0122] Train the network to get the optimal solution and So that the following is true:

[0123]

[0124] Where i = 1,…,L is equivalent to the following minimization loss function:

[0125]

[0126] Solve the linear equation system Hβ=T and get the optimal solution of the loss function for:

[0127] Step 3: Improve the dynamic multi-population PSO algorithm to optimize the ELM model

[0128] The improved dynamic multi-swarm particle swarm algorithm is introduced to optimize the input layer weights and hidden layer biases of ELM, which makes up for the instability of the network caused by the random generation of ELM weights and hidden layer biases. At the same time, the multi-swarm strategy is introduced into the traditional PSO method, and the strategies such as Levy flight, hybrid particles, and sine-cosine learning factors are used to improve its global exploration and local development capabilities, overcoming the defects of swarm intelligence algorithms that are difficult to get rid of local optimality, low execution efficiency, and difficulty in balancing global and local search capabilities, and the model accuracy is higher.

[0129] The specific steps of IDM-PSO algorithm to optimize the ELM model are as follows:

[0130] Step 3.1, population initialization: Take the mean square error between the predicted value and the actual value of the ELM network as the fitness function and calculate the initial fitness value f(1).

[0131] Step 3.2, divide the population according to the fitness value: calculate the fitness value of all particles, and sort them according to the fitness value. The first N / 2 particles with smaller fitness values ​​are "dominant particles", and the last N / 2 particles with larger fitness values ​​are "inferior particles". Then randomly select N / 6 particles from the "dominant particles" and "inferior particles" to form a population, called a "mixed group". The remaining "dominant particles" are called "dominant groups", and the remaining "inferior particles" are called "inferior groups". The particles are sorted and regrouped every time a new generation is updated.

[0132] Step 3.3, update the position and speed of particles in various groups: in the process of dynamic grouping, the fitness value of particles in the "dominant group" is relatively small, and they need to inherit the position information of the optimal solution in the population and reduce the step size for a more detailed search; while the particles in the "disadvantaged group" should expand the "pace", while approaching the global extreme value, while exploring new optimal solutions around to improve mining capacity; as a local model, the "mixed group" contains both better particles and worse particles, which will dynamically adjust the diversity of the population, absorbing individual experience and sharing global information.

[0133] (1) Dominant group

[0134] The main reason why the particle swarm algorithm converges slowly in the later stage of optimization is that it is difficult to get rid of the current local extreme value, resulting in a decrease in accuracy. In order to enhance the ability of the dominant group to jump out of the local optimum, the Levy flight strategy is introduced, which has the characteristics of long-term random walk with a small step size and occasional sudden jumps in direction with a larger step size. The flight method enhances the activity and jumping ability of particles, expands the search range of particles, is conducive to enhancing particle diversity, avoiding the algorithm from falling into the local optimum, and can improve the convergence accuracy of the algorithm. The formula for updating the particle position by Levy flight is as follows:

[0135]

[0136] in, and x g (k) are the global optimal particle position and the global optimal particle position of the population after Levy flight update at the kth iteration, α is the step size control factor, generally taken as 0.01. Levy(χ) is the random search path, and · represents the dot product. Since Levy distribution is very complex and cannot be implemented, the Mantegna algorithm is currently used to simulate its flight trajectory, and its mathematical expression is shown in the formula:

[0137]

[0138] Among them, the parameter χ is a random value in the interval (0,2), usually 1.5; μ and v both obey the normal distribution and are defined as follows:

[0139]

[0140] The variance σ in the above formula is μ and σ v Determined by the following formula:

[0141]

[0142] Where Γ is the gamma function. At the same time, although Levy flight can make particles get rid of local optimality, it cannot guarantee that the updated particle position is better than the original position. Therefore, in order to avoid meaningless position updates, the present invention introduces a greedy algorithm evaluation strategy to determine whether to update the optimal particle position, that is, when the updated position is better than the original position, the position is updated, otherwise the original position is retained. The implementation process is shown in the formula:

[0143]

[0144] in, is the particle position after the greedy algorithm updates, and f(·) represents the particle fitness function.

[0145] (2) Disadvantaged Group

[0146] The "inferior group" has less information about particles with preservation value, and its population is far away from the optimal solution of the problem. The optimal solution can be searched in the entire space by mutation. Mutation can find various possible solution areas in the entire search space, and can also avoid premature convergence and increase the diversity of sub-populations. Mutation judgment is as follows:

[0147]

[0148] GV(x)=x×(1+N(0,1))

[0149] Among them, r and N(0,1) are random numbers between [0,1], x is the original parameter value, GV(x) is the mutated value, k is the current iteration number, K max is the maximum number of iterations. When the judgment condition is established, the executable formula performs Gaussian mutation operation on the particles. After multiple iterations, the probability of the judgment being established gradually decreases, and the probability of particles mutating also gradually decreases. The mutation operation can make particles mutate with a greater probability in the initial stage, expand the search range of particles in the solution space, and ensure the diversity of the particle swarm.

[0150] In addition, the concept of hybrid particles is introduced to change the traditional speed update formula, and the hybrid particles are recorded as p mix (k).p mix The dimension of (k) is a random mixture of the historical optimal values ​​of each particle, such as Figure 3 As shown in the figure: P1~P N —The current optimal position of N particles; P mix —From P1 to PN Particles randomly selected from each dimension are mixed.

[0151] The position and velocity update method of the disadvantaged group particles is obtained as follows:

[0152] v id (k+1)=ω·v id (k)+c1·r1·(p id (k)-x id (k))+c2·r2·(p gd (k)-x id (k))+c3·r3·(p mix (k)-x id (k))

[0153] x id (k+1)=x id (k)+ν id (k+1)

[0154] Among them, v id (k+1) is the velocity of the k+1th generation of the disadvantaged group particle, ω is the inertia weight, r1 and r2 are random numbers uniformly distributed between [0,1], and p id (k) is the historical best position of particle i in the previous k iterations, x id (k) is the position information of the particle at the kth iteration, p gd (k) is the global optimal position of the particles in the entire disadvantaged group in the first k iterations. c1 is the cognitive coefficient, c2 is the social coefficient, c3 is the hybrid learning factor, and r3 is a random number in the interval (0,1). Hybrid particles act as a traction factor to guide the speed update of particles, which can effectively solve the problem of falling into local optimality. At the same time, their own excellence also makes the particles evolve in a better direction. Through these two strategies, multi-directional random perturbations enhance the diversity of the search process to a certain extent. This optimization method is very suitable for particles in the "disadvantaged group" to conduct volatility searches in the solution space, deepening local learning.

[0155] (3) Mixed group

[0156] The mixed group is between the above two groups. Since the differences between individuals are large in the early stage of the algorithm, the focus is on their cognitive part, which can achieve multi-party communication. In the later stage, the algorithm focuses on strengthening the leadership of the global extreme value and prompting particles to gather around the optimal solution. Therefore, the cosine learning factor and the sine learning factor are introduced into the algorithm to make the self-learning factor monotonically decrease and the population learning factor monotonically increase. The speed update formula of the mixed group obtained in this way is:

[0157]

[0158] The inertia weight factor affects the performance of the PSO algorithm. A larger value is beneficial to the global search, while a smaller weight value is beneficial to speeding up the convergence of the algorithm, which is more beneficial for local detailed development. The inertia weight of the linear decreasing strategy adopted in the basic PSO algorithm cannot adaptively adjust the inertia weight according to the progress of the algorithm, resulting in the particle swarm being unable to balance the global search capability and the local search capability. Therefore, the weight factor can be set larger in the early stage of the search to help improve the global search capability; in the later stage of the search, the weight factor can be set smaller to help enhance the local development capability. This nonlinear inertia weight can effectively solve the problem of premature maturity and oscillation of the algorithm near the optimal solution. The specific expression is as follows:

[0159]

[0160] Among them, ω max -ω min It represents the maximum inertia weight minus the minimum inertia weight. The purpose is to control the variation of the inertia weight, so as to balance the exploration and development capabilities of the particle swarm.

[0161] Step 3.4, update the local optimal position Pbest and the global optimal position Gbest of the particle;

[0162] Step 3.5, repeat steps 3.2-3.4 until the number of iterations k reaches the maximum;

[0163] Step 3.6, output the optimal input weight w and hidden layer bias b: The improved PSO algorithm converges after repeated iterations k times, and minimizes the mean square error between the predicted value and the actual value of the ELM network, that is, the input layer weight and hidden layer bias output when the fitness function is minimized, to form the optimized optimal ELM model.

[0164] Step 3.7, based on the trained optimal ELM model, evaluate the threat level of ground targets and obtain the target threat level evaluation results corresponding to each target.

[0165] The present invention also proposes a ground target threat level assessment system based on IDM-PSO-ELM, implements the ground target threat level assessment method based on IDM-PSO-ELM, realizes ground target threat level assessment based on IDM-PSO-ELM, and executes steps 1 to 3 respectively in three modules.

[0166] A computer device comprises a memory, a processor and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the method for assessing the threat level of a ground target based on IDM-PSO-ELM is implemented to achieve the assessment of the threat level of a ground target based on IDM-PSO-ELM.

[0167] A computer-readable storage medium stores a computer program. When the computer program is executed by a processor, the ground target threat level assessment method based on IDM-PSO-ELM is implemented to achieve ground target threat level assessment based on IDM-PSO-ELM.

[0168] In summary, the present invention introduces an improved dynamic multi-population particle swarm algorithm to optimize the input layer weights and hidden layer deviations of ELM, making up for the instability of the network caused by the random generation of ELM network weights and hidden layer deviations. At the same time, the multi-population strategy is introduced into the traditional PSO method, and the strategies such as Levy flight, hybrid particles, and sine-cosine learning factors are used to improve its global exploration and local development capabilities, overcoming the defects of swarm intelligence algorithms that are difficult to get rid of local optimality, low execution efficiency, and difficulty in balancing global search capabilities and local search capabilities, and the model accuracy is higher.

[0169] Example 1

[0170] In order to verify the effectiveness of the solution of the present invention, the following simulation experiment is further described.

[0171] 1. Simulation conditions

[0172] In order to test the effectiveness of the target threat assessment model of IDM-PSO-ELM, the constructed sample data is divided into 70% training set and 30% test set, and the sample data is input into ELM, PSO-ELM and IDM-PSO-ELM models to compare the prediction results. At the same time, in order to verify the accuracy of the model, the evaluation results of the model are measured from the confusion matrix, precision, recall and F1-score. Some of the sample data are shown in Table 1.

[0173] Table 1 Training sample attribute information

[0174]

[0175] 2. Simulation content and result analysis

[0176] 2.1) Improve PSO algorithm performance

[0177] Figure 6 Compare the evolution curves of PSO and IDM-PSO, and analyze Figure 6It can be found that the improved dynamic multi-population PSO has higher global exploration and local development capabilities, which greatly improves the accuracy of the algorithm. When the PSO algorithm falls into the local optimum, the improved algorithm maintains the diversity of the population through dynamic population reorganization, making it less likely to fall into premature convergence when facing multi-peak problems, and through a variety of particle update strategies, the algorithm tends to global exploration in the early stage and focuses on local exploitation in the later stage, realizing the information flow between multiple populations and improving the accuracy of the algorithm optimization results.

[0178] 2.2) Model testing results

[0179] Figure 7 This is the model test result based on the IDM-PSO-ELM target threat assessment model. The threat level assessment result of the target test sample can be intuitively seen from the confusion matrix, and the result accuracy is 91.8%.

[0180] Figure 8 The results of the target threat level evaluation by ELM, PSO-ELM, GA-ELM and IDM-PSO-ELM models are shown in the figure. From the analysis of the above figure, it can be seen that the performance of ELM is affected by the initialization value because the weights and biases of the ELM model are randomly initialized. The prediction accuracy is only 56.2%, and the precision is only 61%. The prediction effect is much lower than the evaluation model using the optimization algorithm to improve ELM. The PSO-ELM model and the GA-ELM model greatly improve the accuracy and precision, but the recall rate is slightly insufficient, which means that the two models perform better in reducing false positives (False Positives), but may increase false negatives (False Negatives). The IDM-PSO-ELM model performs well in all four evaluation indicators, indicating that the algorithm does a very good job in balancing false positives and false negatives, while maintaining high accuracy (91.8%) and recall. In summary, the algorithm model proposed in the present invention has a high prediction accuracy when identifying the target threat level, and the model has good real-time and generalization capabilities.

[0181] The technical features of the above embodiments may be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0182] The above-described embodiments only express several implementation methods of the present application, and the descriptions thereof are relatively specific and detailed, but they cannot be understood as limiting the scope of the present application. It should be pointed out that, for a person of ordinary skill in the art, several variations and improvements can be made without departing from the concept of the present application, and these all belong to the protection scope of the present application. Therefore, the protection scope of the present application shall be subject to the attached claims.

Claims

1. A ground target threat level assessment method based on IDM-PSO-ELM, characterized in that: The following steps are involved: Step 1: Select target distance, target speed, heading angle, effective range, firepower configuration, detection capability, electronic jamming capability, communication capability, and survivability as threat indicators, use target threat level to characterize the threat degree of the target, and build a ground target sample library; Step 2: Use the ELM network to build a ground target threat level assessment model, take the threat index as input and the target threat level as output, and complete the ELM network training based on the constructed sample database; Step 3, introduce the dynamic multi-swarm particle swarm algorithm, take the input layer weights and hidden layer deviations of the ELM network as particles, take the mean square error between the predicted value and the actual value of the ELM network as the fitness function, divide the group into dominant group, inferior group and mixed group, use Levy flight to update the particle position for the dominant group, use mutation to update the particle position for the inferior group, and introduce sine and cosine learning factors to update the particle position for the mixed group, thereby obtaining the optimal ELM network for ground target threat level assessment.

2. The ground target threat level assessment method based on IDM-PSO-ELM according to claim 1 is characterized in that: Step 1: Select target distance, target speed, heading angle, effective range, firepower configuration, detection capability, electronic jamming capability, communication capability, and survivability as threat indicators, use target threat level to characterize the threat degree of the target, and build a sample database, where: The size of the target distance directly reflects the attack intention of the enemy combat unit. The shorter the distance, the greater the threat. The target distance is defined as {far, medium, close}. The greater the target speed, the greater the threat to us. The target speed is defined as {fast, medium, slow}. The smaller the heading angle of the enemy target, the greater the threat to our target. The heading angle is defined as {large, medium, small}. The same type of combat platform may be equipped with different firepower units. The firepower threat of the platform is analyzed based on the number and type of weapons carried by the platform, and the firepower configuration is defined as {strong, relatively strong, general, relatively weak, weak}; The detection capability of enemy targets reflects the reconnaissance of our combat power configuration. The stronger the detection capability, the greater the threat to us. The detection capability is defined as {strong, relatively strong, average, relatively weak, weak}. Electronic jamming capability refers to the ability of the enemy's combat platform to interfere with our various electronic equipment. The stronger the electronic jamming capability, the greater the threat of the enemy's target to us. The electronic jamming capability is defined as {strong, relatively strong, general, relatively weak, weak} five states; Communication capability refers to the ability of communication, coordination and cooperation between target platforms, between platforms and reconnaissance and guidance equipment, and between platforms and command and control systems. Communication capability is defined as {strong, relatively strong, general, relatively weak, weak}. Survivability refers to the ability of the target platform to resist detection, interference, interception and destruction. Survivability is defined as five states: {strong, relatively strong, average, relatively weak, weak}.

3. The ground target threat level assessment method based on IDM-PSO-ELM according to claim 1 is characterized in that: Step 1: Select target distance, target speed, heading angle, effective range, firepower configuration, detection capability, electronic jamming capability, communication capability, and survivability as threat indicators, use target threat level to characterize the threat degree of the target, and build a sample database, where: The threat level of the target is divided into six levels: low, lower, medium, high, higher, and very high, represented by numbers 1 to 6.

4. The ground target threat level assessment method based on IDM-PSO-ELM according to claim 1, characterized in that: Step 2: Use the ELM network to build a ground target threat level assessment model, with threat indicators as input and target threat level as output. Complete the ELM network training based on the constructed sample database. The specific method is as follows: For N targets (x i ,t i ), where x i =[x i1 ,…,x im ] T , x i ∈R n ; t i =[t i1 ,…,t im ] T , t i ∈R n , x i represents the i-th ground target; x im represents the mth feature of the i-th ground target; t i Represents the threat level of the i-th target; for the ELM network consisting of L hidden layer nodes, it is expressed as: Where b i represents the bias of the i-th hidden layer node; o j Represents the activation output of the i-th target input at the j-th hidden layer node; w i represents the input weight; w i ·x j represents the inner product of the input weight and the target feature value; g(x) represents the activation function; β i represents the output weight; The training goal of the ELM network is to minimize the output error, which can be expressed as: That is, there exists w i , β i and b i So that: It can be expressed as a matrix: Hβ=T In the formula, H represents the calculation output matrix of the hidden layer of the ELM network, where the i-th column corresponds to the output result of the i-th hidden layer node; β represents the output weight matrix; T represents the expected output matrix, that is, the threat level of the output target; Expand to get: in: l represents the number of hidden layer nodes in the ELM network, and m represents the number of target threat level categories; Train the ELM network to get the optimal solution and So that the following is true: This is equivalent to minimizing the loss function as follows: Solve the linear equation system Hβ=T and get the optimal solution of the loss function for:

5. The ground target threat level assessment method based on IDM-PSO-ELM according to claim 1, characterized in that: Step 3, introduce the dynamic multi-swarm particle swarm algorithm, take the input layer weights and hidden layer deviations of the ELM network as particles, take the mean square error between the predicted value and the actual value of the ELM network as the fitness function, divide the dominant group, the inferior group and the mixed group, use the Levy flight strategy to update the particle position for the dominant group, use the mutation method to update the particle position for the inferior group, and introduce the sine and cosine learning factor to update the particle position for the mixed group. Based on this, the optimal ELM network is obtained for ground target threat level assessment. The specific method is as follows: Step 3.1, population initialization: take the mean square error between the predicted value and the actual value of the ELM network as the fitness function and calculate the initial fitness value f(1); Step 3.2, divide the population according to the fitness value: calculate the fitness value of all particles, and sort them according to the fitness value. The first N / 2 particles with smaller fitness values ​​are "dominant particles", and the last N / 2 particles with larger fitness values ​​are "inferior particles". Then, randomly select N / 6 particles from the "dominant particles" and "inferior particles" to form a population, called a "mixed group". The remaining "dominant particles" are called "dominant groups", and the remaining "inferior particles" are called "inferior groups". The particles are sorted and regrouped every time a new generation is updated. Step 3.3, update the position and velocity of various swarm particles: (1) Dominant group The formula for updating the particle position using Levy flight is as follows: in, and x g (k) are the global optimal particle position and the global optimal particle position of the population after Levy flight update at the kth iteration, α is the step size control factor, which is 0.01, · represents the dot product, Levy(X) is the random search path, and the flight trajectory is simulated by the Mantegna algorithm. Its mathematical expression is shown in the formula: Among them, the parameter χ is a random value in the interval (0,2) and is set to 1.5; μ and υ both obey the normal distribution and are defined as follows: The variance σ in the above formula is μ and σ v Determined by the following formula: Where, Γ is the gamma function; The evaluation strategy of the greedy algorithm is introduced to determine whether to update the optimal particle position. That is, the position is updated only when the updated position is better than the original position, otherwise the original position is retained. The implementation process is shown in the formula: in, is the particle position after the greedy algorithm updates, and f(·) represents the particle fitness function; (2) Disadvantaged Group The particle position is updated by mutation, and the mutation judgment is as follows: GV(x)=x×(1+N(0,1)) Among them, r and N(0,1) are random numbers between [0,1], x is the original parameter value, GV(x) is the mutated value, k is the current iteration number, K max is the maximum number of iterations. When the judgment condition is established, the Gaussian mutation operation is performed on the particles. The position and velocity of the disadvantaged group particles are updated as follows: v id (k+1)=ω·v id (k)+c1·r1·(p id (k)-x id (k))+ c2·r2·(p gd (k)-x id (k))+c3·r3·(p mix (k)-x id (k)) x id (k+1)=x id (k)+ν id (k+1) Among them, v id (k+1) is the velocity of the k+1th generation of the disadvantaged group particle, ω is the inertia weight, r1 and r2 are random numbers uniformly distributed between [0,1], and p id (k) is the historical best position of particle i in the previous k iterations, x id (k) is the position information of the particle at the kth iteration, p gd (k) is the global optimal position of the particle of the entire disadvantaged group in the first k iterations. c1 is the cognitive coefficient, c2 is the social coefficient, c3 is the hybrid learning factor, and r3 is a random number in the interval (0,1); (3) Mixed group The cosine learning factor and the sine learning factor are introduced to make the self-learning factor monotonically decrease and the population learning factor monotonically increase. The particle velocity update formula of the mixed group obtained in this way is: Step 3.4, update the local optimal position Pbest and the global optimal position Gbest of the particle; Step 3.5, repeat steps 3.2 to 3.4 until the number of iterations k reaches the maximum; Step 3.6, output the optimal input weight w and hidden layer bias b: repeat the iteration k times until convergence, minimize the mean square error between the predicted value and the actual value of the ELM network, that is, output the input layer weight and hidden layer bias when the fitness function is minimized, to form the optimized optimal ELM network; Step 3.7, based on the optimal ELM network, evaluate the threat level of ground targets and obtain the target threat level evaluation results corresponding to each target.

6. The ground target threat level assessment method based on IDM-PSO-ELM according to claim 5 is characterized in that: The inertia weight adopts nonlinear inertia weight. In the early stage of search, the weight factor is set to be larger; in the later stage of search, the weight factor is set to be smaller. The expression is: Among them, ω max -ω min It represents the maximum inertia weight minus the minimum inertia weight, and its purpose is to control the variation of the inertia weight.

7. A ground target threat level assessment system based on IDM-PSO-ELM, characterized in that: The ground target threat level assessment method based on IDM-PSO-ELM as described in any one of claims 1 to 6 is implemented to realize the ground target threat level assessment based on IDM-PSO-ELM, and steps 1 to 3 are respectively performed in three modules.

8. A computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, the ground target threat level assessment method based on IDM-PSO-ELM according to any one of claims 1 to 6 is implemented to realize the ground target threat level assessment based on IDM-PSO-ELM.

9. A computer-readable storage medium having a computer program stored thereon, wherein when the computer program is executed by a processor, the method for ground target threat level assessment based on IDM-PSO-ELM according to any one of claims 1 to 6 is implemented to realize ground target threat level assessment based on IDM-PSO-ELM.

Images