Intelligent algorithm for battlefield situation real-time analysis and fire control sensor operation recommendation based on time sequence classification
Patent Information
- Application Number
- CN202411440436.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-10-15
- Publication Date
- 2025-01-07
- Estimated Expiration
- Not applicable · inactive patent
AI Technical Summary
[0004]目前，国内外空战领域学者针对空战飞行员火控和飞行员传感器操作推荐提出了多种方法：基于专家系统的方法，通过专家经验构建专家知识库，将已知专家知识显式建模，从而构建战场态势特征，但是这种方法受限于专家知识，灵活性较差，规则库可能不完善；基于优化方法的决策分析，该类方法的共同点在于通过各类态势评估方法将各种态势信息转化为标准的优化模型，这种方法对于无人机对抗博弈问题效果很好，但是面对有人机战场实时态势分析不够灵活；基于深度学习的方法，通过RNN/CNN网络以及Transformer等深度学习方法进行决策建模，但是目前空战数据较少，数据标签处理标注过程漫长，无法满足深度学习算法的数据量需求，而且深度模型解释性较差，在推荐过程中人机交互性较差
[0007]首先进行特征工程，将原始数据中的经纬高特征转换为东北天坐标，利用卡尔曼滤波、小波变换方法处理数据中的噪点，再进行战场态势敌我双方交互信息特征构造，然后对多维高阶特征进行特征选取，将有效特征进行滑动窗口处理，提取其时域特征，再利用集成学习模型LightGBM对火控、传感器操作关键点进行识别，针对行为关键点利用态势战术库进行匹配以完成战场实时火控、传感器操作推荐。最后通过增益、覆盖度等方式衡量每个特征的重要性，利用基于lime的方法解释算法，在特征重要程度排名的基础上对特征进一步筛选以减少特征维度获得轻量级模型，使得火控、传感器操作推荐具有可解释性。
Smart Images
Figure BSA0000299758870000021 
Figure BSA0000299758870000022 
Figure BSA0000299758870000032
Abstract
Description
Technical field:
[0001] The present invention relates to the field of intelligent algorithms, and specifically to the fields of time series analysis, data mining, visualization and interpretability. It provides real-time fire control and sensor operation recommendations to pilots on the battlefield, and helps with battlefield situation awareness and decision-making assistance.
[0002] In order to recommend the sensor operation and weapon control behavior of the pilot, determine the initial battlefield situation information corresponding to the behavior, and deeply explore the objective situation factors under the pilot's behavioral decision, there have been a lot of studies on the real-time analysis of battlefield situation and the fire control sensor operation recommendation methods, such as decision-making based on various tactical action libraries, decision-making based on optimization methods, and decision-making based on deep learning methods. However, the above methods have some disadvantages when dealing with real-time battlefield fire control and sensor operation recommendations: various methods based on tactical action libraries are based on expert knowledge and a large amount of air combat experience, but they cannot solve the current problems of small battlefield data volume and long sample label calibration process, and most methods are applied to offline scenarios and cannot make real-time fire control and sensor operation recommendations; existing deep learning models usually require large-scale labeled data for training to achieve high performance, and obtaining sufficient actual combat data may be a challenge, and they are usually considered black box models, and it is difficult to explain their decision-making process. Background technology:
[0003] Real-time battlefield situation analysis is a key application in the military and defense fields. It requires the combination of multiple background technologies to achieve real-time situation monitoring, analysis and decision-making. Among them, air combat pilot fire control and pilot sensor operation are crucial in battlefield situation analysis. Modeling the association between real-time battlefield situation information and pilot fire control and pilot sensor operation is an effective intelligent algorithm processing method. How to model battlefield situation characteristics is a key issue in the field of air combat at home and abroad. How to construct high-order features in limited data feature dimensions is also a challenge in the field of data mining.
[0004] At present, scholars in the field of air combat at home and abroad have proposed a variety of methods for air combat pilot fire control and pilot sensor operation recommendations: expert system-based methods, which build an expert knowledge base through expert experience, explicitly model known expert knowledge, and thus construct battlefield situation characteristics. However, this method is limited by expert knowledge, has poor flexibility, and the rule base may be imperfect; decision analysis based on optimization methods, the common point of this type of method is that various situation information is converted into a standard optimization model through various situation assessment methods. This method works well for UAV confrontation game problems, but it is not flexible enough for real-time situation analysis of manned battlefields; deep learning-based methods, through RNN / CNN networks and Transformer and other deep learning methods for decision modeling, but currently there is less air combat data, the data label processing and annotation process is long, and it cannot meet the data volume requirements of deep learning algorithms, and the deep model has poor interpretability, and the human-computer interaction is poor in the recommendation process. In military applications, the transparency and interpretability of decisions are usually crucial. The present invention is an intelligent algorithm for recommending battlefield fire control sensor operations. It extracts features by processing the original data exported by tacview, constructs high-order battlefield situation features by feature transformation and feature construction, and selects appropriate features through verification and evaluation. It uses an integrated learning model algorithm to identify key behavior points of fire control and sensor operations, and then recommends behaviors by matching the situation and tactics library. It analyzes the importance of features to explain their influence in the model, making the model interpretable. Summary of the invention:
[0005] The present invention proposes an intelligent algorithm for real-time analysis of battlefield situation based on time series classification and recommendation of pilot fire control sensor operation, which makes real-time fire control and sensor operation recommendations to pilots on the battlefield, thus providing assistance for battlefield situation perception and auxiliary decision-making.
[0006] In order to recommend the pilot's sensor operation and weapon control behavior, determine the initial battlefield situation information corresponding to the behavior, and deeply explore the objective situation factors under the pilot's behavioral decision-making, the present invention provides an intelligent algorithm for real-time analysis of battlefield situation and recommendation of fire control sensor operation based on time series classification.
[0007] First, feature engineering is performed to convert the latitude and longitude features in the original data into northeast celestial coordinates. Kalman filtering and wavelet transform methods are used to process the noise in the data. Then, the interactive information features of the battlefield situation between the enemy and us are constructed. Then, multi-dimensional high-order features are selected, and the effective features are processed with sliding windows to extract their time domain features. The integrated learning model LightGBM is used to identify the key points of fire control and sensor operations. The situation tactical library is used to match the key points of behavior to complete the real-time battlefield fire control and sensor operation recommendations. Finally, the importance of each feature is measured by gain, coverage, etc., and the lime-based method is used to interpret the algorithm. Based on the ranking of feature importance, the features are further screened to reduce the feature dimensions to obtain a lightweight model, making the fire control and sensor operation recommendations interpretable. Specific implementation method:
[0008] The present invention provides an intelligent algorithm for real-time analysis of battlefield situation based on time series classification and recommendation of pilot fire control sensor operation, comprising the following steps:
[0009] Step 1: Obtain battlefield flight parameters and fire control data sets, use the DCS World simulation platform to simulate the battlefield situation and obtain data sets, merge the obtained flight parameter data with avionics data to obtain the final data set. The data set contains the following original features; the aircraft's three-dimensional coordinate data A LLA = [Longitude, latitude, Altitude], attitude angle information [Roll, Pitch, Yall], heading angle Heading, true airspeed TAS, calibrated airspeed CAS, etc. The sampling frequency in the experiment is 20HZ, that is, each sampling point is separated by 0.05 seconds.
[0010] Step 2: Use the above-obtained data set to randomly divide the 20 battle data into training set, validation set, and test set in a ratio of 1:1:1 and perform 3-fold cross validation. Use the training set data to train the model, select the optimal effect feature set on the validation set data, and test the model performance on the test set data.
[0011] Step 3: The longitude, latitude and altitude involved in the present invention need to be converted into coordinates in the Northeast Sky Coordinate System. The Northeast Sky Coordinate System (ENU, East-North-Up) is a local coordinate system commonly used in aerospace, geographic information systems and other fields. To convert longitude, latitude and altitude into ENU coordinates, a reference point (longitude, latitude and altitude) is first required, and then the ENU coordinates of the target point are calculated based on this reference point.
[0012] The specific conversion steps are as follows:
[0013] (1) Select a reference point: Select a longitude, latitude, and altitude as a reference point.
[0014] (2) Calculate the difference:
[0015] Δ longitude = target point longitude - reference point longitude
[0016] Δ dimension = target point dimension - reference point dimension
[0017] ΔAltitude = Target point altitude - Reference point altitude
[0018] (3) Convert to Cartesian coordinates: Use the following formula to convert the longitude and latitude difference into Cartesian coordinates (x, y, z):
[0019] x = R * Δ longitude * cos (reference point latitude)
[0020] y = R * Δ dimension
[0021] z = Δ altitude where R is the mean radius of the Earth.
[0022] (4) Conversion to ENU coordinates: Use the following formula to convert Cartesian coordinates to ENU coordinates:
[0023] E=-x*sin(reference point longitude)+y*cos(reference point longitude)
[0024] N = -x*sin(reference point dimension)*cos(reference point longitude)-y*sin(reference point dimension)*sin(reference point longitude)+z*cos(reference point dimension)
[0025] U = x*cos(reference point dimension)*cos(reference point longitude)+y*cos(reference point dimension)*sin(reference point longitude)+z*sin(reference point dimension)
[0026] After the above steps, the original data is converted into standard data through coordinate system transformation and noise filtering.
[0027] In an air combat environment, the flight parameter data information obtained by the radar may contain noise points and data missing problems, which will be very unfavorable for real-time judgment of fire control and sensor key points, so the original data needs to be standardized. In addition to transferring all maneuvering data to a unified coordinate system through coordinate system conversion, it is also necessary to use Kalman filtering and wavelet transform to reduce noise points and missing values in the data.
[0028] Step 4: Based on the battlefield situation domain knowledge, construct high-level battlefield situation information such as various speeds, azimuth angles and their change rates, target entry angles and their change rates according to the original data of our aircraft attitude information and enemy aircraft position information. The following is the specific calculation formula:
[0029] The velocity vector of our aircraft is calculated by the following formula:
[0030]
[0031] Where v is the velocity, x is the three-axis coordinate, and t is the sampling time. Calculate the velocity of the three-dimensional coordinate axis [v x , v y , v z ];
[0032]
[0033] where v x ,v y ,v z is the three-axis speed, and the aircraft flight speed is calculated;
[0034]
[0035] where v x ,v y ,v z is the three-axis speed, and the azimuth of our aircraft is calculated.
[0036]
[0037] where v x ,v y , v z is the three-axis speed, and the aircraft pitch angle is calculated;
[0038]
[0039] Roll is the roll angle in the original data, Pitch is the pitch angle of the aircraft, and the aircraft inclination angle is calculated;
[0040]
[0041] Where GS represents ground speed, Pitch represents pitch angle, TSA represents true airspeed, and the entry angle characteristics are calculated.
[0042]
[0043] ROS=atan 2 (v z , Ground Distance)-atan 2 (v z ,Air Distance)
[0044] where v zis the vertical speed, Ground Distance is the horizontal distance between the enemy and our aircraft, and Air Distance is the spatial distance between the enemy and our aircraft. The line of sight angular velocity of the enemy aircraft is calculated;
[0045]
[0046] where v x ,v y are the x-axis and y-axis speeds, and the aircraft heading angle is calculated;
[0047]
[0048] Where angle is the pitch angle or heading angle, and the rate of change of the attitude angle is calculated.
[0049]
[0050] where v z v x v y are the speeds in the z, x, and y directions respectively, and the enemy aircraft's climb angle is calculated.
[0051] Step 5: Divide the time window and extract the basic unit of time series data. Through the above-mentioned feature construction, the battlefield situation information at a certain moment can be reflected, but the time series characteristics of the flight parameter data and avionics data of enemy and friendly fighters during flight cannot be reflected. For time series classification problems, delineating time windows to extract features can effectively capture the characteristics of data within a period of time. The time series features extracted by the present invention include: mean, variance, standard deviation, maximum value, minimum value, difference between maximum and minimum values, and coefficient of variation, seven-dimensional time series features.
[0052] The size of the time window in the present invention is a hyperparameter, and an optimal time window needs to be found through the results of experiments with different time window sizes. Taking the key point judgment of radar frequency behavior as an example, the LightGBM model is used to select the optimal time window size by comparing the accuracy of the key behavior point judgment of fire control and each sensor operation under different windows.
[0053] Step 6: Use the LightGBM model to judge the key behavior points of fire control and sensors. The LightGBM model is currently a more mainstream model for ensemble learning. Compared with other decision tree-based algorithms, LightGBM uses a "growth by leaf nodes" strategy, which allows fewer splits during training, thereby reducing memory and computing costs. At the same time, due to the split point selection strategy, less memory is required, and highly parallel training is supported. It can make full use of multi-core CPUs and distributed computing resources, improve training speed, and has built-in support for category features. It does not require unique hot encoding or label encoding, reducing the complexity of feature engineering. LightGBM uses a decision tree-based framework, and the algorithm flow is as follows:
[0054] First, set up battlefield situation fire control and sensor data sets When i = 0, indicating that this is not a critical behavior point. i =1 means this moment is a critical behavior point.
[0055] The first step is to initialize the parameters: specify the number of trees to be trained, specify the learning rate as 0.01, the learning rate is a parameter that controls the degree of influence of each iterative update of the tree on the model, a smaller learning rate will make the model more stable, and a larger learning rate will make the model converge faster, specify the depth of the tree as 4, specify the loss function as logarithmic loss, specify the ratio of features used in each tree training as 0.8 to control the randomness of the model and prevent overfitting, and specify the minimum number of samples contained in the leaf nodes of each tree as 20.
[0056] The second step is to initialize the weak classifier as:
[0057]
[0058] Among them, G0(x) is the initial weak classifier, which will become a strong classifier after multiple iterations.
[0059] The third step is to iterate. For the mth iteration, m=1,2...,M:
[0060] For each sample i, i=1,2...,n, get the value of the first tree:
[0061]
[0062] Using the loss function L(y i , F m-1 (x1)) calculates the negative gradient value and approximates it as the residual:
[0063]
[0064] The negative gradient value of the first tree can be calculated.
[0065] The fourth step is to use the histogram algorithm to discretize the continuous floating-point features into k discrete values, and construct a Histogram with a width of k. The training data is traversed to count the cumulative statistics of each discrete value in the histogram. When performing feature selection, it is only necessary to traverse and find the optimal segmentation point based on the discrete values of the histogram.
[0066] On top of the histogram algorithm, a leaf-wise algorithm with depth limit is used.
[0067] (p m , f m ,v m )=arg min L(F m-1 (x i )·split(p,f,v),Y)
[0068] The regression tree can be expressed as W q(x) , q∈{1, 2, .., J}, where J represents the number of leaves q represents the decision rule of the tree w is a vector representing the sample weight of the leaf node. Therefore, LightGBM will be trained in an additive form at the tth step, as follows:
[0069]
[0070] The above results are difficult to calculate, and the objective function is quickly approximated using Newton's method. For simplicity, after removing the constant term, the formula can be transformed into the following form:
[0071]
[0072] where g i and They represent the first-order and second-order gradient statistics of the loss function respectively. Let I j represents the sample set of leaf j, then the above formula can be transformed into the following form:
[0073]
[0074] For a given tree structure q(x), the optimal leaf weight score for each leaf node is and The extreme value of can be solved as follows:
[0075]
[0076]
[0077] in It can be regarded as a scoring function to measure the quality q of the tree structure. Finally, the objective function after adding the split is:
[0078]
[0079] Among them I L and I R They are sample sets for the left and right branches, respectively. Different from traditional GBDT-based techniques (such as XGBoost and GBDT), the LightGBM used in the present invention grows the tree vertically, while other algorithms grow the tree horizontally, which makes LightGBM an effective method for processing large-scale data and features.
[0080] Step 7: Use Shap to explain the model. Shap is an interpretable algorithm library that calculates the shapley of each feature to obtain the contribution of the feature to the model output. The formula for calculating the contribution value is:
[0081]
[0082] Where g is the explanation model, M is the input feature dimension, and t is whether the corresponding feature exists. is a constant term, is the Shapley value of each feature. Since the integrated learning model LightGBM used is a decision tree model, all features exist, and the formula can be rewritten as:
[0083]
[0084] Through the above calculations, we can obtain the importance of each feature in the judgment of the key behavior points of fire control and sensors for the judgment results, and then ablate the model, and finally select the most important features for judging the key behavior points of fire control and sensor operation to ensure the interpretability of the model.
[0085] Step 8: Enter the situation and tactics library for matching. After determining the key behavior points of fire control and sensor operation, enter the situation and tactics library and use the method of combining rule matching based on Minkowski distance and distance matching to obtain the most similar operation behavior for recommendation.
[0086] Step 9: During the test, the data of the test set data_test is passed into the model optimized by the above steps. In the process of judging the key behavior points of fire control and sensors, the accuracy (Accuracy, acc) and the comprehensive evaluation score F1 (F1_score) are used as the main evaluation indicators. At the same time, in order to more comprehensively evaluate the performance of the model, the recall rate (Recall, R) and the precision rate (Precision, Pre) are used as auxiliary evaluation indicators. In addition, since the present invention is suitable for real-time analysis of battlefield situations, the algorithm time overhead must also be considered, so the model running time and the situation tactical library matching algorithm consumption time are introduced as considerations at the same time.
[0087] Among them, accuracy, recall, precision, F1_score and model running time are expressed as follows:
[0088] Accuracy: Accuracy = (true positive examples + true negative examples) / total number of samples;
[0089] Recall: Recall = True positive examples / (True positive examples + False negative examples);
[0090] Precision: Precision = True Positives / (True Positives + False Positives);
[0091] F1_score: F1 score = 2*(precision*recall) / (precision+recall);
[0092] Model running time: the time taken to complete a key behavior point judgment.
Claims
1. A method for dividing time windows, extracting basic units of time series data, and constructing features can reflect battlefield situation information at a certain moment, but cannot reflect the time series characteristics of flight parameter data and avionics data of enemy and friendly fighters during flight. For time series classification problems, delimiting time windows to extract features can effectively capture the characteristics of data within a period of time. The time series features extracted by the present invention include: The seven-dimensional time series characteristics are mean, variance, standard deviation, maximum value, minimum value, difference between maximum and minimum values, and coefficient of variation. The size of the time window in the present invention is a hyperparameter. It is necessary to find an optimal time window through the results of experiments with different time window sizes. Taking the judgment of key points of radar frequency behavior as an example, the LightGBM model is used to select the optimal time window size by comparing the accuracy of the key behavior points of fire control and each sensor operation under different windows.
2. A method for judging the key behavior points of fire control and sensors using the LightGBM model. The LightGBM model is currently a mainstream model for integrated learning. Compared with other decision tree-based algorithms, LightGBM uses a "growth by leaf nodes" strategy, which allows fewer splits during training, thereby reducing memory and computing costs. At the same time, due to the split point selection strategy, less memory is required, and highly parallel training is supported. It can make full use of multi-core CPUs and distributed computing resources, improve training speed, and has built-in support for category features. It does not require one-hot encoding or label encoding, reducing the complexity of feature engineering.