CN115310508A - Damage probability calculation method based on machine learning classifier - Google Patents
Damage probability calculation method based on machine learning classifier Download PDFInfo
- Publication number
- CN115310508A CN115310508A CN202210583315.8A CN202210583315A CN115310508A CN 115310508 A CN115310508 A CN 115310508A CN 202210583315 A CN202210583315 A CN 202210583315A CN 115310508 A CN115310508 A CN 115310508A
- Authority
- CN
- China
- Prior art keywords
- damage probability
- damage
- algorithm
- calculation
- target
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F17/00—Digital computing or data processing equipment or methods, specially adapted for specific functions
- G06F17/10—Complex mathematical operations
- G06F17/17—Function evaluation by approximation methods, e.g. inter- or extrapolation, smoothing, least mean square method
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F17/00—Digital computing or data processing equipment or methods, specially adapted for specific functions
- G06F17/10—Complex mathematical operations
- G06F17/18—Complex mathematical operations for evaluating statistical data, e.g. average values, frequency distributions, probability functions, regression analysis
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N20/00—Machine learning
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Computational Mathematics (AREA)
- Mathematical Analysis (AREA)
- Mathematical Optimization (AREA)
- Pure & Applied Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- Databases & Information Systems (AREA)
- Algebra (AREA)
- Operations Research (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Evolutionary Biology (AREA)
- Life Sciences & Earth Sciences (AREA)
- Probability & Statistics with Applications (AREA)
- Artificial Intelligence (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Evolutionary Computation (AREA)
- Medical Informatics (AREA)
- Computing Systems (AREA)
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
Abstract
The invention discloses a damage probability calculation method based on a machine learning classifier, which comprises the following steps: establishing a damage probability calculation model of the fragment warhead to the target; according to the damage probability calculation model, carrying out a large amount of random calculation aiming at all random parameters influencing the calculation result; the damage probability results in a large amount of example data are approximate to two types of 0 and 1 or three types of 0,0.5 and 1; training the above example data by using a classifier to obtain a damage probability prediction model; and calculating the average single damage probability of the fragment warhead to the target under the condition of specific bullet intersection by using a damage probability prediction model. Compared with the conventional method for calculating the damage probability, the damage probability calculating method based on the machine learning classifier not only ensures certain accuracy of calculation precision, but also reduces the calculation amount of the algorithm and greatly shortens the calculation time.
Description
Technical Field
The invention belongs to the technical field of damage assessment, and particularly relates to a damage probability calculation method based on a machine learning classifier.
Background
In order to accurately evaluate the damage effect of the fragment warhead on the target and facilitate the cooperation research of the warfare, a corresponding damage probability calculation model needs to be established on the basis of the target vulnerability analysis.
The common damage probability represents the probability of damaging the target by the warhead, the calculation amount of the existing damage probability calculation method of the target by the fragment warhead is too large, a single result can be calculated in MATLAB (matrix laboratory) for complex conditions by hundreds of seconds or even thousands of seconds, and the calculation amount is exponentially increased along with the increase of the number of fragments. When further to obtain the average single-shot damage probability, the Monte Carlo method is adopted to carry out 1000 random repeated calculations, the time consumption is calculated in days. Therefore, a calculation method for rapidly calculating the damage probability of the fragment warhead to the target on the basis of ensuring certain accuracy of the result is needed to be found.
Disclosure of Invention
The invention aims to overcome the defects and provides a damage probability calculation method based on a machine learning classifier, which comprises the following steps: establishing a damage probability calculation model of the fragment warhead on the target; according to the damage probability calculation model, a large number of random calculations are carried out on all random parameters influencing the calculation result based on a Monte Carlo method, and each parameter in each calculation is randomly taken; the damage probability results in a large amount of example data obtained in the last step are approximate to two types of 0 and 1 or three types of 0,0.5 and 1; training the example data by using a classifier in machine learning, and applying an optimization iteration process aiming at a part of classifier algorithms to improve the performance of the classifier so as to finally obtain a damage probability prediction model; and calculating the average single damage probability of the fragment warhead to the target under the condition of specific bullet intersection by using the obtained damage probability prediction model. Compared with the conventional method for calculating the damage probability, the damage probability calculating method based on the machine learning classifier not only ensures certain accuracy of calculation precision, but also reduces the calculation amount of the algorithm and greatly shortens the calculation time.
In order to achieve the above purpose, the invention provides the following technical scheme:
a damage probability calculation method based on a machine learning classifier comprises the following steps:
according to the three-dimensional model of the target, performing vulnerability analysis on the component parts of the target; according to the vulnerability analysis result, establishing a damage probability calculation model of the fragment warhead to the target;
according to the damage probability calculation model, aiming at all random parameters influencing the calculation result, carrying out repeated random calculation for more than m times based on the Monte Carlo method to obtain a large amount of damage probability example data aiming at different damage levels; each example data comprises a group of independent variable random parameter arrays and damage probability result values of different levels corresponding to the random parameter arrays; m is more than 10000;
a large amount of damage probability arithmetic data are arranged into a prediction variable set and a response set; the prediction variable set is a set of random parameter arrays in all the example data, and the response set is a damage probability set with different levels corresponding to the random parameter arrays; in the response set, according to the actual value of the damage probability in the [0,1] interval, the damage probability data is approximately valued into 0 or 1 or 0,0.5 and 1;
training a prediction variable set serving as input and a response set serving as output by utilizing a classifier algorithm in machine learning to obtain damage probability prediction models aiming at different damage levels;
under the condition of a specific bullet intersection, setting related random parameters in the damage probability prediction model, namely, part of random parameters which can be related to the specific bullet intersection condition in an input random parameter array as fixed values, carrying out repeated random calculation for more than n times by adopting a Monte Carlo method based on the damage probability prediction model, and predicting to obtain the average single-shot damage probability of the fragment warhead to the target under the specific bullet intersection condition, wherein the rest random parameters in the input random parameter array are still random values within a specific range; n is more than 500.
Further, according to the actual value of the damage probability in the [0,1] interval, the specific method for approximating the damage probability data to 0 or 1 or 0,0.5 and 1 is as follows:
classifying the actual value of the damage probability by a rounded mathematical approximation, wherein the actual value of the damage probability is approximately 0 when the damage probability is below 0.5 and is approximately 1 when the damage probability is above 0.5; or approximately 1 when the damage probability is 0.25 or less, approximately 0.5 when the damage probability is 0.25 or more and 0.75 or less, and approximately 1 when the damage probability is 0.75 or more.
Further, the classifier algorithm is one of a logistic regression algorithm, a naive Bayes algorithm, a decision tree algorithm, a discriminant analysis algorithm, a support vector machine algorithm, a nearest neighbor algorithm or an ensemble learning algorithm.
Furthermore, in the classifier algorithm, the accuracy of the prediction result is adjusted through the misclassification cost; the basis of finishing the training of the classifier algorithm is the maximum training time or the maximum iteration times;
the validation method of the classifier comprises cross validation, leave-out validation or no validation.
Further, the random parameter includes: the method comprises the following steps of (1) detonating position of a fragment warhead, incoming angle of a missile, pitch angle of the missile, trajectory angle of the missile, yaw angle of the missile, miss distance of the missile, final speed of the missile, and included angle between final speed vector of the missile and the ground or guidance precision probability error;
the entrance angle of the missile is randomly selected in the range of 0-360 degrees according to the average distribution;
the ballistic angle of the missile is randomly distributed within the range of-10 degrees to 0 degrees;
the yaw angle of the missile is randomly taken in an atmosphere of-2 degrees to 2 degrees according to uniform distribution;
the final speed of the missile is randomly selected from 200m/s to 300m/s according to uniform distribution;
the vertical direction miss distance of the missile is randomly selected between-1 m and 1m according to normal distribution.
Furthermore, according to the damage probability calculation model, when repeated random calculation based on the Monte Carlo method is carried out for more than m times aiming at all random parameters influencing the calculation result, input constants are required to be predetermined, wherein the constants comprise the physical parameters of the warhead, the flight parameters of the warhead, the detection distance of the fuze and the detonation delay; the physical parameters of the warhead comprise the structure and the number of prefabricated fragments, the equivalent TNT loading amount of the warhead and the loading coefficient of the warhead, and the flight parameters of the warhead comprise the slip angle of the warhead.
Further, the damage probability calculation model is as follows:
wherein, N or For the number of non-redundant parts, k is the number of redundant parts, P or,i A probability of damage to each non-redundant component; m is a group of 1 ，M 2 ，…，M k The number of redundant components of each redundant component group; p is and,j The damage probability of each part in the redundant part group; the target components comprise a plurality of groups of redundant components and a plurality of non-redundant components, wherein when any one of the redundant components in the group of redundant components is damaged, the target is judged to be damaged, and when any one of the non-redundant components is damaged, the target is judged to be damaged.
Further, the different damage levels include: m-level task damage, F-level fire control damage, and K-level catastrophic damage; the damage of the M-level task indicates that the target functional component can not fully exert the due function, and the obstacle removal needs 1 to 24 hours; f-level fire control damage represents that a target functional part loses the function capability, and the obstacle can be eliminated after 1 to 7 days and nights; a K-level catastrophic failure means that the target weapon system loses operational capability and is not economically feasible to repair or repair the damage.
Furthermore, the parameters required to be set in the decision tree algorithm include the maximum classification number, the classification criterion and the alternative decision splitting;
the parameters required to be set in the discriminant analysis algorithm comprise a covariance structure;
the parameters required to be set in the support vector machine algorithm comprise a kernel function, a frame constraint level and a kernel scale mode;
the parameters required to be set in the nearest neighbor algorithm comprise the number of neighbor points, distance measurement and distance weight;
the parameters required to be set in the ensemble learning algorithm comprise an ensemble method, a learner type, a maximum split number and the number of learners.
Further, when the classifier algorithm is other than the logistic regression algorithm, the optimizer is adopted to perform iterative optimization on each adjustment parameter of the classifier algorithm, and an optimized damage probability prediction model is obtained after dozens of iterations.
Compared with the prior art, the invention has the following beneficial effects:
1) The method adopts the damage probability prediction model to replace the original damage probability calculation model, greatly reduces the calculation amount and improves the calculation efficiency by more than 20 times;
2) Compared with the result obtained by carrying out Monte Carlo calculation through the damage probability calculation model, the average single damage probability obtained by carrying out Monte Carlo calculation through the damage probability prediction model of the method can achieve the accuracy of more than 75 percent on average;
3) The method carries out approximate value taking on the damage probability, improves the classification efficiency on the premise of ensuring the accuracy and avoids huge calculation amount.
Drawings
Fig. 1 is a flowchart of a damage probability calculation method based on a machine learning classifier according to the present invention.
Detailed Description
The features and advantages of the present invention will become more apparent and appreciated from the following detailed description of the invention.
The word "exemplary" is used exclusively herein to mean "serving as an example, embodiment, or illustration. Any embodiment described herein as "exemplary" is not necessarily to be construed as preferred or advantageous over other embodiments. While the various aspects of the embodiments are presented in drawings, the drawings are not necessarily drawn to scale unless specifically indicated.
Firstly, a three-dimensional model of the target is established, vulnerability analysis is carried out on all parts of the target, and then a damage probability calculation model of the fragment warhead on the target is established. Further, random calculation is carried out more than 10000 times on the basis of a Monte Carlo method according to random parameters such as the bullet characteristics and the intersection conditions which affect the calculation results, and each random parameter in each calculation is randomly valued according to the self characteristics. The damage probability results in a large amount of arithmetic example data obtained in the last step are approximated to be two types of 0 and 1 or three types of 0,0.5 and 1 according to the distribution condition of the damage probability results on the [0,1] interval; the example data is trained by using a classifier in machine learning, an optimization iterative process is applied to a part of classifier algorithms to improve the performance of the classifier, and finally a damage probability prediction model is obtained, wherein the damage probability can be predicted to be 0 or 1 (or 0,0.5 or 1) according to values of random parameters. The obtained damage probability prediction model is used for replacing the original damage probability calculation model, and the Monte Carlo method is adopted again to calculate the average single damage probability of the warhead to the target under the condition of specific bullet-and-eye interaction. The damage probability prediction model can be used for calculating the damage probability under the condition of intersecting different bullet eyes, the calculated amount is small, and the huge calculated amount of the damage probability calculation model is avoided. Therefore, the method can obviously shorten the time required for calculating the average single-shot damage probability on the basis of ensuring certain accuracy of the damage probability calculation result.
As shown in fig. 1, the method for calculating damage probability based on a machine learning classifier of the present invention adopts a classifier algorithm in machine learning to train to obtain a damage probability prediction model, so as to calculate the damage probability of a fragment warhead to a target (such as an air-defense missile system) under a specific condition, and comprises the following steps:
s101: establishing a damage probability calculation model of the fragment warhead to the target;
s201: according to the damage probability calculation model, aiming at all random parameters influencing the calculation result, such as the tail end flying speed, the entrance angle, the pitch angle, the miss distance and the like of the warhead of the target, more than 10000 times of repeated random calculation is carried out based on the Monte Carlo method, and a large amount of damage probability calculation example data aiming at different damage levels are obtained. In each calculation, random parameters are randomly taken within a certain range according to certain probability distribution, specifically, for example, the entry angle of the missile relative to a target is randomly taken within the range of 0-360 degrees according to uniform distribution, the channel angle of the missile is randomly taken within the range of-10-0 degrees according to uniform distribution, the yaw angle of the missile is randomly taken within the atmosphere of-2 degrees according to uniform distribution, the final speed of the missile is randomly taken within the range of 200-300 m/s according to uniform distribution, the miss distance of the missile in the vertical direction is randomly taken within the range of-1 m according to normal distribution, and the like;
s301: and adjusting a large amount of damage probability calculation example data (more than 10000 groups) obtained by calculation into a prediction variable and a response set, namely all random input parameter sets of the calculation examples and K-level, F-level and M-level damage probability sets which are correspondingly output, and approximately taking the damage probability data of each level as 0 or 1 or 0,0.5 and 1 according to the distribution condition of the damage probability in a [0,1] interval.
The damage probability of the warhead to the target in the damage probability calculation model is real number in a [0,1] interval, but according to the specific warhead performance and the vulnerability analysis and geometric modeling condition of the target, the results of most damage probabilities are 0 and 1 or 0,0.5 and 1. Therefore, here, the damage probability is classified by mathematical approximation of "rounding off", that is, 0.5 or less is approximately 0,0.5 and 1 or more, or 0.25 or less is approximately 1,0.25 and 0.75 or less is approximately 0.5,0.75 and 1 or more.
S301: training the above example data by using a classifier algorithm in machine learning, applying an optimization iterative process aiming at a part of classifier algorithms to improve the performance of the classifier, and finally obtaining a damage probability prediction model aiming at different damage levels, wherein the damage probability prediction model can predict whether the damage probability is 0 or 1 (or 0,0.5 or 1) according to values of various random parameters, so that huge calculation amount in the damage probability calculation model is avoided;
the classifier mainly adopts several algorithms of logistic regression, naive Bayes, decision trees, discriminant analysis, support vector machines, nearest neighbor and ensemble learning and the like.
The data set substituted into the classifier mainly comprises a predictive variable (namely an independent variable) and a response part, wherein the predictive variable consists of a plurality of vectors, and each vector represents the value of a certain input parameter in 10000 times of calculation; the response (i.e., dependent variable) is a vector representing the damage probability output for each of 10000 calculations. Therefore, the damage probabilities at three different levels are three different responses, and are respectively combined with the predictive variables to form a data set to be carried into the classifier.
The verification method of the classifier mainly comprises three types of cross verification, leave-out method verification and no-verification.
The misclassification cost is one of adjustable parameters of the classifier, taking the case that the damage probability is 0 or 1 as an example, the influence of misclassifying 1 as 0 and misclassifying 0 as 1 is different, and can be embodied by the misclassification cost, and generally, the misclassification cost is appropriately adjusted according to a specific prediction result and an actual conforming condition.
Different algorithms can be selected in the classifier, which have different adjustment parameters, such as:
the decision tree algorithm comprises the maximum classification number (more than 2 positive integers), classification criteria (a kini diversity index, a dichotomy rule and the like), alternative decision splitting and the like;
the discriminant analysis algorithm has a covariance structure (diagonal or full);
a kernel function (linear, gaussian and the like), a frame constraint level, a kernel scale mode and the like in the support vector machine algorithm;
the nearest neighbor algorithm comprises the number of neighboring points (positive integer), distance measurement (Euclidean, city block and the like), distance weight (equal distance, inverse distance and the like) and the like;
the ensemble learning algorithm includes an ensemble method (bag, adaBoost, RUSBoost, and the like), a learner type, a maximum split number, the number of learners, and the like.
In order to improve the accuracy of the classifier algorithm, optimizers such as Bayes and the like are adopted to further optimize the adjustment parameters of the classifier algorithm, and the better machine learning classification model is finally obtained after dozens of iterations by continuously adjusting parameter values through iteration.
The classifier algorithm can use iterative optimization except a logistic regression algorithm, and relevant adjustment parameters include optimizer selection (bayesian optimization, network search and the like, wherein bayesian optimization is more commonly used), acquisition functions (expected improvement per second, expected improvement and the like), iteration times (generally integers more than 30 times), learning rate (generally 0.001 to 1), calculation time and the like.
The classifier is adopted to train the samples, the algorithms can be used, and the ensemble learning algorithm applying the iterative optimization method is a relatively good choice. The training result of the classifier mainly adopts accuracy as a basis for measuring the performance of the classifier, wherein the accuracy is the proportion of the predicted correct sample number to the total sample number. However, for the distribution of some example data (for example, the damage probability result 1 is significantly more than 0, or vice versa), the prediction result of the classifier may be completely biased to the side with a larger number (i.e., result 1), where the accuracy may be higher, but the prediction model is not reasonable, and then the misclassification cost needs to be adopted for adjustment, so that the training process of the classifier is more inclined to the side with a smaller number to predict more accurately. The basis for the classifier training to end is mainly determined by the maximum training time or the number of iterations.
S401: by applying the damage probability classifier prediction model obtained in the last step, under the condition of a specific bullet intersection, a part of random parameters are fixed according to physical reality and can be 1 or a plurality of random parameters, the specific fixed parameters are determined by the physical reality or research requirements, such as an entry angle, a pitch angle and the like, the Monte Carlo method is adopted again to carry out more than 500 times of repeated random calculation, and after data statistics, the average single-shot damage probability of the warhead to the target under the condition of the specific bullet intersection is obtained (the single-shot damage probability is output by the damage probability prediction model through single calculation, if the Monte Carlo method is adopted to carry out multiple times of random calculation, a plurality of single-shot damage probabilities can be obtained, and the average single-shot damage probability is obtained after averaging). By changing the fixed parameter types, the average single damage probability of the warhead to the target under different conditions can be conveniently calculated, for example, the entrance angle of the fixed missile is 270 degrees, the final speed of the fixed missile is 200m/s, the trajectory angle of the fixed missile is-5 degrees, and the like.
Further, the specific method for establishing the damage probability calculation model of the fragment warhead to the target comprises the following steps:
step 1: establishing a three-dimensional model of the target;
the establishment of the target three-dimensional model mainly refers to the equivalence and simplification of all main parts of the target into surface units, wherein the surface units comprise quadrilateral node coordinates of all the surface units, information such as thickness and material of the surface units and the like, and a digital model is formed.
Step 2: analyzing vulnerability of each target part;
the vulnerability analysis of each target part is mainly carried out according to the theory in the target vulnerability (Beijing university of science university publishers, li facing east, etc.), vulnerability coefficients are set for each main part of the target to represent the damage difficulty degree of each part under the action of certain fragment penetration or explosion shock waves, the value of the vulnerability coefficients is equal to the ratio of the vulnerable area of the outer surface of the part to the presented area, the vulnerability coefficients are assigned to the surface unit of each part in the three-dimensional model to be used for the damage probability calculation in the step 5, and the target three-dimensional model with the surface unit containing vulnerability information is obtained so far and is the basis for calculating the damage probability of the fragment to the surface unit in the bullet intersection process.
And 3, step 3: establishing a damaged tree model for damaging the target;
the damage degree of the target is generally divided into three levels according to target vulnerability: m-level damage, F-level damage, K-level damage;
wherein M-level damage is task damage, corresponding to the condition that the target functional component can not fully exert the due function, and the maintenance personnel need 1 to 24 hours to remove the obstacle; the F-grade damage, namely the fire control damage, corresponds to the failure of the function capability of a target functional part, and can eliminate the obstacle only after 1 to 7 days and nights by a special maintenance unit; the K-level damage, namely the destruction damage, is economically infeasible to repair or repair the damage corresponding to the loss of the operational capability of a target weapon system;
analyzing the influence of damage on all parts on the function of the whole system, for example, destructively damaging both a radar antenna and a shelter can cause destructive damage on an air defense missile system, finally respectively forming three levels of damaged trees, and providing non-combined or non-redundant parts and combined or redundant parts existing in each level of damaged trees;
and 4, step 4: establishing a bullet and target intersection model of the warhead and the target;
establishing a bullet and target intersection model of the fragment warhead and the target, which specifically comprises the following contents:
1) And establishing a fragment field and an overpressure field under a warhead coordinate system (the field is a physical term and refers to the distribution condition of an object in space, the fragment field refers to the distribution condition of each fragment in space, including position and velocity vector information and the like, and the overpressure field refers to the distribution condition of the overpressure of the explosion shock wave in space). In the aspect of a fragment field, the space distribution, the mass distribution and the speed distribution of fragments after static explosion of a warhead are calculated through the warhead charging, the prefabricated fragments and a shell structure, a corresponding space analysis model is established, and a random fragment field under a coordinate system of the warhead, namely information such as the mass, the shape coefficient and the speed vector of each fragment, is obtained. In the aspect of an overpressure field, calculating the overpressure delta p of the shock wave peak value generated by the explosion of the fragment warhead m And positive pressure specific impulse i:
in the formula of omega e Equivalent loading amount; omega is the charge amount under the cased charge; alpha is the charge coefficient; r is 0 /r f Is the ratio of the initial radius of the shell to the crushing radius; gamma is the polytropic index of the detonation products; Δ p m The peak value of shock wave generated when the explosive charges explode in an infinite air area is expressed in MPa; m is e Is equivalent TNT equivalent in kg; r is the distance from the explosion point to the center of the explosion and is m; c is a dimensionless constant, and i has the unit Pa · s;
2) Establishing a bullet-target intersection model of the fragment warhead and the target, mainly establishing a fragment field and an overpressure field under a target coordinate system through an intersection relation of the warhead and the target, judging whether each fragment in the fragment field intersects with each surface unit of the target or not, if so, judging whether the fragments can penetrate through the surface units, and calculating whether the overpressure field can damage each specific surface unit of the target or not. The meeting relation between the fragmentation field and the target is determined by the terminal attitude of the warhead and the terminal speed direction of the warhead.
And 5: and establishing a damage probability calculation model of the fragment warhead to the target.
Calculating the damage probability of the fragment warhead to the target, specifically comprising the following steps:
1) Calculating the damage probability of the fragmentation field to the target component:
and calculating the damage probability of each part of the target, wherein the parts are regarded as a whole to be calculated, and the damage probability of the parts is calculated by taking the sum of the number of the fragments penetrating on all surface elements of the parts as the effective damage fragments of the whole parts. Taking all fragments penetrating through the t-th part in the part as effective damage fragments of the part, and calculating the damage probability of the t-th part in the part according to the following formula:
p t ＝1-e -ξ·n
where ζ is the vulnerability coefficient and n is the total number of fragments penetrated by the tth part;
the method comprises the following steps of calculating the damage probability of the second part according to the non-redundancy of each surface, namely, the whole part is considered to be damaged as long as one surface in a surface unit forming the part is damaged, wherein the damage probability of the second part in the part is calculated according to the following formula:
in the formula p m Is the damage probability of the mth surface unit in the s-th part, and R is the total number of the surface units of the s-th part;
the other type of parts are units with a part of redundant faces, that is, if a certain percentage (for example, 20%) of the total number of face units are destroyed (the damage probability is 1), the part is considered to be destroyed (the damage probability is 1);
2) Calculating the damage probability of the overvoltage field to the target component:
firstly, calculating the overpressure of the shock wave received by a certain surface unit, and if the overpressure of the shock wave is greater than the pressure bearing threshold of the surface unit, considering the overpressure damage probability of the surface unit as 1;
the damage criterion of the shock wave is calculated according to the following formula:
(ΔP-P * )(I-I * )≥K
wherein, delta P and I are respectively the overpressure at the shock wave peak value and the impulse at the shock wave ratio, P * And I * Respectively shock wave critical overpressure and shock wave critical specific impulse, K being a constant, depending on the vulnerability of the target. When the above equation is satisfied, the target bin is considered 100% likely to be destroyed. The damage probability of the rupture field and the overpressure field on a certain surface element is the maximum value and is the final damage probability of the surface element, and then the damage probability of the component is further calculated by referring to the content in the section 1) according to the damage condition of the surface element.
3) Calculating a total damage probability of a system
According to the target vulnerability analysis theory, two damage calculation conditions exist in each part forming the target, one is that the target is damaged when at least one part in the non-combined or non-redundant parts is damaged, and the other is that the target is damaged when all the combined or redundant parts are damaged. Then the total damage probability of the warhead to the target system can be respectively calculated according to the damage probability of each part of the target and the analysis results of the damage trees of the three levels (the total damage probability is divided into three levels and is respectively and independently calculated), and the formula is as follows:
wherein, N or The number of bins without redundant parts, k the number of groups of redundant parts, P or,i For the probability of failure of each non-redundant component, M 1 ，M 2 ，…，M k Respectively, the number of redundant parts in each redundant part group, P and,i Is the damage probability of each part in the redundant part group.
Further, a specific method for performing more than 10000 times of repeated random calculations on all random parameters affecting the calculation result based on the monte carlo method is as follows:
step 1: determining a constant which needs to be input when calculating the damage probability, namely a constant quantity when performing random calculation;
the constants to be input when calculating the damage probability are mainly: the physical parameters of the warhead, such as the structure and the number of the prefabricated fragments, the equivalent TNT explosive loading amount of the warhead, the explosive loading coefficient of the warhead and the like; flight parameters of the warhead, such as the slip angle of the warhead; and the detection distance and detonation delay of the fuse.
And 2, step: determining random parameters which need to be randomly calculated by using a Monte Carlo method when calculating the damage probability;
the random parameters that need to be randomly calculated by using the monte carlo method when calculating the damage probability mainly include: the detonation position of the fragment warhead, the entrance angle, the pitch angle, the miss distance, the terminal velocity, the included angle between the terminal velocity vector and the ground, and the guidance precision probability error, namely the circle probability error.
And 3, step 3: based on a damage probability calculation model, on the basis of determining the constants in the step 1, a Monte Carlo method is used for performing more than 10000 times of repeated random calculations on all random parameters in the step 2, and finally a large amount of damage probability calculation example data (more than 10000 groups) can be obtained. In each calculation, random parameters are randomly taken within a certain range according to certain probability distribution, specifically, for example, the entry angle of the missile relative to a target is randomly taken within the range of 0-360 degrees according to average distribution, the channel angle of the missile is randomly taken within the range of-10-0 degrees according to uniform distribution, the yaw angle of the missile is randomly taken within the atmosphere of-2 degrees according to uniform distribution, the final speed of the missile is randomly taken within the range of 200-300 m/s according to uniform distribution, the miss distance of the missile in the vertical direction is randomly taken within the range of-1 m according to normal distribution, and the like.
Further, the damage probability results in a large amount of example data obtained by the monte carlo random calculation are approximate to two types of 0 and 1 or three types of 0,0.5 and 1 according to the distribution condition of the damage probability results in the [0,1] interval, and the specific method comprises the following steps:
the damage probability of the warhead to the target in the damage probability calculation model is real number in a [0,1] interval, but according to the specific warhead performance and the vulnerability analysis and geometric modeling condition of the target, most damage probability results are often near 0 and 1 or near 0,0.5 and 1. Therefore, here, the damage probability is classified by mathematical approximation of "rounding off", that is, 0.5 or less is approximately 0,0.5 and 1 or more, or 0.25 or less is approximately 1,0.25 and 0.75 or less is approximately 0.5,0.75 and 1 or more.
After the processing, the damage probability example data can be analyzed by using a classifier algorithm in machine learning, and a prediction model is extracted.
Further, the specific method for training the above-mentioned a large amount of damage probability example data subjected to result approximation processing by using a classifier algorithm of machine learning to generate the damage probability prediction model is as follows:
step 1: determining an algorithm, a data set structure and a verification method adopted when classifier algorithm training data adopting a machine learning method is adopted;
training data using a classifier algorithm of a machine learning method first selects an algorithm and defines a dataset structure. The classifier mainly comprises several algorithms such as logistic regression, naive Bayes, decision trees, discriminant analysis, support vector machines, nearest neighbor and ensemble learning. The mathematical principles of the algorithms are different, the analysis performance of the algorithms on various specific problems and data is also different, and various algorithms are needed to be adopted for analysis to find out the relatively optimal algorithm.
The data set substituted into the classifier mainly comprises a predictive variable (namely an independent variable) and a response part, wherein the predictive variable consists of a plurality of vectors, and each vector represents the value of a certain input parameter in 10000 times of calculation; the response (i.e., the dependent variable) is a vector representing the damage probability output results (0 and 1, or 0,0.5 and 1) for each of 10000 calculations. Therefore, the damage probabilities at three different levels are three different responses, and are respectively brought into the classifier together with the prediction variables to form a data set.
The machine learning analysis process of the classifier needs to select a verification method for analyzing and verifying the performance of the prediction model on the data set, and the verification method generally comprises three types of cross verification, leave-out verification and no verification. Wherein, the cross validation is to divide a data set into X parts, then take one part as a test set, take the rest X-1 parts as a training set, and circulate the steps until each part is tested, and cross validation is usually performed by 10 folds; the leave-out method verification means that a data set is directly divided into two mutually exclusive sets, one set is used as a training set, the other set is used as a test set, and errors are tested on the test set after a model is trained on the training set and used as estimation of generalization errors. Typically, 10-fold cross-validation is selected.
And 2, step: determining each parameter when classifier training data of a machine learning method is adopted;
the parameters to be set when the machine learning method is adopted and the classifier training data is based on the classifier mainly comprise the following parameters:
the misclassification cost is one of adjustable parameters of the classifier, taking the situation that the damage probability is 0 or 1 as an example, the influence of misclassifying a 1 as a 0 and misclassifying a 0 as a 1 is different, and can be embodied through the misclassification cost, and the misclassification cost is generally properly adjusted according to a specific prediction result and an actual conforming situation;
different algorithms in the classifier have different adjustment parameters, specifically for example:
the decision tree algorithm comprises a maximum classification number (more than 2 positive integers), classification criteria (a Gini diversity index, a dichotomy rule and the like), alternative decision splitting and the like;
the discriminant analysis algorithm has a covariance structure (diagonal or full);
a kernel function (linear, gaussian and the like), a frame constraint level, a kernel scale mode and the like in the support vector machine algorithm;
the nearest neighbor algorithm comprises the number of neighboring points (positive integer), distance measurement (Euclidean, city block and the like), distance weight (equal distance, inverse distance and the like) and the like;
the ensemble learning algorithm includes an ensemble method (bag, adaBoost, RUSBoost, and the like), a learner type, a maximum split number, the number of learners, and the like.
Subsequent example data training can be performed based on the classifier by adopting the parameters.
And 3, step 3: inputting more than 10000 groups of damage probability example data obtained by random calculation in S201 into a classifier tool for machine learning to train;
1) More than 10000 sets of damage probability arithmetic example data obtained by random calculation in the S201 are adjusted into an input part and an output part, namely all input parameter sets of arithmetic examples and K-level, F-level and M-level damage probability sets which are output correspondingly;
2) The input data set and the output data set are input into classifier tool software for various machine learning to be trained (K-class training, F-class training and M-class training respectively), and usable tools include Matlab and Python and the like.
And 4, step 4: through continuously carrying out classifier training on the damage probability example data for multiple times (dozens of times), adjusting part of learning parameters, such as changing a training algorithm or misclassification cost, and the like, a damage probability prediction model with the highest accuracy (the ratio of the number of correctly predicted samples to the total number of samples) can be finally found. By utilizing the model, the corresponding K-level, F-level and M-level damage probabilities under different input conditions can be predicted according to the values of all parameters, so that huge calculation amount in the damage probability calculation model is avoided.
And 5: furthermore, each parameter of the classifier algorithm can be optimized by adopting an optimizer, so that the accuracy is improved;
in order to improve the accuracy of the classifier algorithm, optimizers such as Bayes and the like can be further adopted to optimize each parameter of the classifier algorithm, and a better machine learning classification prediction model is finally obtained after dozens of iterations by continuously adjusting parameter values through iteration.
The classifier algorithm can use iterative optimization except for a logistic regression algorithm, and relevant adjustment parameters include optimizer selection (bayesian optimization, network search and the like, wherein bayesian optimization is more commonly used), the number of learners (generally integers of more than 10), an acquisition function (expected improvement, expected improvement and the like per second), iteration times (generally integers of more than 30 times), a learning rate (generally 0.001 to 1) and the like.
The classifier is adopted to train the samples, the algorithms can be used, and the ensemble learning algorithm applying the iterative optimization method is a relatively good choice. The training result of the classifier mainly adopts accuracy as a basis for measuring the performance of the classifier, wherein the accuracy is the proportion of the predicted correct sample number to the total sample number. However, for the distribution of some example data (for example, the damage probability result 1 is significantly more than 0, or vice versa), the prediction result of the classifier may be completely biased to the side with a larger number (i.e., result 1), where the accuracy may be higher, but the prediction model is not reasonable, and then the misclassification cost needs to be adopted for adjustment, so that the training process of the classifier is more inclined to the side with a smaller number to predict more accurately.
Further, the specific method for calculating the average single damage probability of the warhead to the target under the specific bullet-and-eye interaction condition by using the damage probability prediction model obtained in S301 and adopting the monte carlo method is as follows:
step 1: determining bullet intersection conditions required by calculation according to user requirements;
the calculation of the required specific bullet meeting condition is determined by the actual requirements of a user of the method, under the condition, a part of random parameters such as an entrance angle, a pitch angle and the like are fixed according to physical actual or research requirements, other parameter value taking methods are unchanged and are still random values within a specific range, and an input data set corresponding to the bullet meeting condition is obtained.
And 2, step: and (2) applying the obtained damage probability prediction model, carrying out more than 500 times of repeated random calculation by adopting a Monte Carlo method aiming at the input data set in the step (1), counting data to obtain the average single damage probability of the fragment warhead to the target under the bullet meeting condition in the step (1), wherein other random parameter value taking methods except the specific bullet meeting condition are unchanged in each calculation. Therefore, 10000 times of repeated random calculation are carried out on the damage probability calculation model, and the damage probability prediction model obtained by classifier training replaces the original damage probability calculation model, and the calculation amount is small, so that the huge calculation amount of the damage probability calculation model is avoided. Furthermore, the average single-shot damage probability of the warhead to the target under different conditions can be conveniently calculated by changing the bullet-and-target intersection condition and the corresponding input data set (namely changing the fixed parameter type).
Compared with the result obtained by directly carrying out Monte Carlo calculation through the damage probability prediction model, the average single damage probability obtained by carrying out Monte Carlo calculation through the damage probability prediction model can reach the accuracy of more than 75 percent on average, and meanwhile, the method obviously reduces the calculated amount and improves the calculation efficiency by more than 20 times.
The invention has been described in detail with reference to specific embodiments and illustrative examples, but the description is not intended to limit the invention. Those skilled in the art will appreciate that various equivalent substitutions, modifications or improvements may be made to the technical solution of the present invention and its embodiments without departing from the spirit and scope of the present invention, which fall within the scope of the present invention. The scope of the invention is defined by the appended claims.
Those skilled in the art will appreciate that those matters not described in detail in the present specification are well known in the art.
Claims (10)
1. A damage probability calculation method based on a machine learning classifier is characterized by comprising the following steps:
according to the three-dimensional model of the target, performing vulnerability analysis on the component parts of the target; according to the vulnerability analysis result, establishing a damage probability calculation model of the fragment warhead to the target;
according to the damage probability calculation model, aiming at all random parameters influencing the calculation result, carrying out repeated random calculation for more than m times based on the Monte Carlo method to obtain a large amount of damage probability example data aiming at different damage levels; each example data comprises a group of random parameter arrays and damage probabilities of different levels corresponding to the random parameter arrays; m is more than 10000;
a large amount of damage probability example data are arranged into a prediction variable set and a response set; the prediction variable set is a set of random parameter arrays in all the example data, and the response set is a damage probability set with different levels corresponding to the random parameter arrays; in the response set, according to the actual value of the damage probability in the [0,1] interval, the damage probability data is approximately valued into 0 or 1 or 0,0.5 and 1;
training a prediction variable set serving as input and a response set serving as output by utilizing a classifier algorithm in machine learning to obtain damage probability prediction models aiming at different damage levels;
setting related random parameters in the damage probability prediction model as fixed values under the condition of specific bullet meeting, carrying out repeated random calculation for more than n times by adopting a Monte Carlo method based on the damage probability prediction model, and predicting to obtain the average single-shot damage probability of the fragment warhead to the target under the condition of the specific bullet meeting; n is more than 500.
2. The damage probability calculation method based on the machine learning classifier as claimed in claim 1, wherein the specific method for approximating the damage probability data to 0 or 1 or 0,0.5 and 1 according to the actual value of the damage probability in the [0,1] interval is as follows:
classifying the actual value of the damage probability by a rounded mathematical approximation, wherein the actual value of the damage probability is approximately 0 when the damage probability is below 0.5 and is approximately 1 when the damage probability is above 0.5; or approximately 1 when the damage probability is 0.25 or less, approximately 0.5 when the damage probability is 0.25 or more and 0.75 or less, and approximately 1 when the damage probability is 0.75 or more.
3. The method of claim 1, wherein the classifier algorithm is one of a logistic regression algorithm, a naive Bayes algorithm, a decision tree algorithm, a discriminant analysis algorithm, a support vector machine algorithm, a nearest neighbor algorithm, or an ensemble learning algorithm.
4. The method of claim 1, wherein in the classifier algorithm, the accuracy of the predicted result is adjusted by the cost of misclassification; the basis of finishing the training of the classifier algorithm is the maximum training time or the maximum iteration times;
the validation method of the classifier comprises cross validation, leave-out validation or no validation.
5. The method of claim 1, wherein the random parameters comprise: the detonation position of the fragment warhead, the entrance angle of the missile, the pitch angle of the missile, the trajectory angle of the missile, the yaw angle of the missile, the miss distance of the missile, the final speed of the missile, the included angle between the final speed vector of the missile and the ground or the guidance precision probability error;
the entrance angle of the missile is randomly selected in the range of 0-360 degrees according to the average distribution;
the ballistic angle of the missile is randomly taken within the range of-10 degrees to 0 degrees according to uniform distribution;
randomly taking the yaw angle of the missile in an atmosphere of-2 degrees to 2 degrees according to uniform distribution;
the final speed of the missile is randomly selected from 200m/s to 300m/s according to uniform distribution;
the vertical direction miss distance of the missile is randomly selected between-1 m and 1m according to normal distribution.
6. The damage probability calculation method based on the machine learning classifier as claimed in claim 5, wherein according to the damage probability calculation model, when the repeated random calculation based on the Monte Carlo method is performed for more than m times for all the random parameters affecting the calculation result, input constants are predetermined, wherein the constants include the physical parameters of the warhead, the flight parameters of the warhead, the detection distance of the fuze and the initiation delay; the physical parameters of the warhead comprise the structure and the number of the prefabricated fragments, the equivalent TNT (trinitrotoluene) explosive loading amount of the warhead and the explosive loading coefficient of the warhead, and the flight parameters of the warhead comprise the slip angle of the warhead.
7. The method of claim 1, wherein the damage probability calculation model is:
wherein N is or Number of non-redundant parts, k number of redundant parts, P or,i A probability of damage to each non-redundant component; m 1 ，M 2 ，…，M k The number of redundant components of each redundant component group; p and,j The damage probability of each part in the redundant part group; the target components comprise a plurality of groups of redundant components and a plurality of non-redundant components, wherein when any one of the redundant components in the group of redundant components is damaged, the target is judged to be damaged, and when any one of the non-redundant components is damaged, the target is judged to be damaged.
8. The method of claim 1, wherein the different damage levels comprise: m-level task damage, F-level fire control damage, and K-level catastrophic damage; the damage of the M-level task indicates that the target functional component can not fully exert the due function, and the obstacle removal needs 1 to 24 hours; f-level fire control damage represents that a target functional part loses the function capability, and the obstacle can be eliminated after 1 to 7 days and nights; a K-level catastrophic failure means that the target weapon system loses operational capability and is not economically feasible to repair or repair the damage.
9. The method of claim 3, wherein the parameters required to be set in the decision tree algorithm include maximum classification number, classification criteria and alternative decision splitting;
the parameters required to be set in the discriminant analysis algorithm comprise a covariance structure;
the parameters required to be set in the support vector machine algorithm comprise a kernel function, a frame constraint level and a kernel scale mode;
the parameters required to be set in the nearest neighbor algorithm comprise the number of adjacent points, distance measurement and distance weight;
the parameters required to be set in the ensemble learning algorithm comprise an ensemble method, a learner type, a maximum split number and the number of learners.
10. The damage probability calculation method based on the machine learning classifier as claimed in claim 9, wherein when the classifier algorithm is other than the logistic regression algorithm, the optimizer is used to perform iterative optimization on each adjustment parameter of the classifier algorithm, and an optimized damage probability prediction model is obtained after tens of iterations.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202210583315.8A CN115310508A (en) | 2022-05-25 | 2022-05-25 | Damage probability calculation method based on machine learning classifier |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202210583315.8A CN115310508A (en) | 2022-05-25 | 2022-05-25 | Damage probability calculation method based on machine learning classifier |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN115310508A true CN115310508A (en) | 2022-11-08 |
Family
ID=83855251
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202210583315.8A Pending CN115310508A (en) | 2022-05-25 | 2022-05-25 | Damage probability calculation method based on machine learning classifier |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN115310508A (en) |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN116958607A (en) * | 2023-09-20 | 2023-10-27 | 中国人民解放军火箭军工程大学 | Data processing method and device for target damage prediction |
-
2022
- 2022-05-25 CN CN202210583315.8A patent/CN115310508A/en active Pending
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN116958607A (en) * | 2023-09-20 | 2023-10-27 | 中国人民解放军火箭军工程大学 | Data processing method and device for target damage prediction |
| CN116958607B (en) * | 2023-09-20 | 2023-12-22 | 中国人民解放军火箭军工程大学 | Data processing method and device for target damage prediction |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN111080108B (en) | Data-driven weapon equipment combat effectiveness evaluation index screening method and system | |
| Ryan et al. | Machine learning methods for predicting the outcome of hypervelocity impact events | |
| CN115310508A (en) | Damage probability calculation method based on machine learning classifier | |
| CN114036610A (en) | Penetration depth prediction method based on data enhancement | |
| CN115630257A (en) | Blasting funnel volume prediction method | |
| CN114818471A (en) | Damage probability calculation method based on machine learning neural network | |
| CN107273914B (en) | Efficient fusion identification method based on information source self-adaptive dynamic selection | |
| Olivieri et al. | Simulations of satellites mock-up fragmentation | |
| CN110852574B (en) | Target threat assessment method and medium based on improved grey target theory | |
| US20230110939A1 (en) | Device and method for generating damage matrix based on warhead fragment data and target vulnerable area data | |
| Boltenkov et al. | Devising a method for improving the efficiency of artillery shooting based on the Markov model | |
| CN113988305B (en) | Machine learning-based penetration algorithm verification method and system | |
| CN114490582A (en) | Auxiliary rapid decision-making method and system for war field ammunition | |
| Ding et al. | Sensor placement optimization based on an improved inertia and adaptive particle swarm algorithm during an explosion | |
| CN115618222A (en) | Prediction method of tunneling response parameters | |
| CN113837644A (en) | Equipment combat effectiveness and contribution rate integrated evaluation method based on grey correlation | |
| Hou et al. | Adaptive fuze-warhead coordination method based on BP artificial neural network | |
| Liang et al. | ASE: Anomaly Scoring Based Ensemble Learning for Imbalanced Datasets | |
| CN105868536A (en) | Test data based health state assessment method for solid rocket engine | |
| CN117892637B (en) | Target plate breakdown thickness prediction method and device based on joint network model | |
| AU2021102797A4 (en) | Real-time cleaning method for maritime non-combat targets and system therefor | |
| CN111091865B (en) | Method, device, equipment and storage medium for generating MoRFs prediction model | |
| CN109866931B (en) | Airplane throttle control method based on self-encoder | |
| Buhrii et al. | Adversarial attacks on object detection systems | |
| Wang et al. | Improved classification model for peptide identification based on self-paced learning |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination |