Target intention recognition method based on data and knowledge combined driving

Abstract

The invention relates to the field of target intention prediction, and relates to a target intention recognition method based on knowledge and data combined driving, and the method comprises the following steps: 1, collecting behavior variable data of an air flying target, and carrying out the data cleaning, thereby obtaining x; 2, expert knowledge is obtained through a Delphi method, and a confidence rule base is constructed; step 3, inputting the data into a confidence rule base module to obtain an embedded representation xe of expert knowledge, and splicing the embedded representation of the expert knowledge and the data to form lt; x, xegt; all the data are input into the neural network module; and step 4, iteratively optimizing parameters of the belief rule base and the neural network in a collaborative optimization mode until the model converges. According to the method, the effect of expert knowledge in the intention recognition process is fully utilized, and the accuracy of the model is improved. And high intention recognition accuracy can still be obtained under the condition that the data volume is not large.

Description

Technical Field

[0001] The present invention relates to the field of target intention prediction, and relates to a method for identifying target intention jointly driven by knowledge and data. Background Art

[0002] In the field of target intention recognition, the rule base method, as one of the early applied technologies, structures historical experience and knowledge into a rule base and uses reasoning technology to realize the recognition of target intention, marking an important transformation from purely relying on manual judgment to computer automated processing. This method has significant milestone significance for promoting the development of target intention recognition technology. He Yu et al. successfully realized the recognition of target intention by transforming the knowledge of domain experts into a series of rules and combining the belief rule base and evidence reasoning technology. Wang Haibin, Liu Zuandong et al. improved the belief rule base by using the fuzzy mathematics method. The results show that the model after fuzzy processing has enhanced robustness, especially in dealing with uncertain data. In addition to the belief rule base, other researchers have also explored diverse technical means such as evidence reasoning and template matching for target intention recognition. For example, scholars such as Chen Li used Dynamic Bayesian Networks to predict the immediate actions of targets and combined template matching technology to predict future combat intentions. Zhou Zhiqiang et al. proposed an improved evidence reasoning method, which considered the weights of different evidences and combined basic probability assignment and evidence to achieve accurate recognition of missile intentions.

[0003] Although these methods based on expert knowledge have been proven effective in long-term practice, they show limitations under the requirements of rapid response in modern battlefields. These methods highly rely on the input of expert knowledge and require a large amount of preliminary work to sort out and structure expert knowledge, which limits their application in dynamic environments. In addition, these methods usually have fewer parameters and limited learning ability, and do not fully utilize the advantages of data-driven decision-making.

[0004] With the continuous progress of artificial intelligence technology, the capabilities of computers in data processing, integration, and analysis have been significantly enhanced. This progress has greatly promoted the "digital intelligence" process of the battlefield environment. Against this backdrop, machine learning and deep learning technologies have played a crucial role in the military field. Machine learning methods introduce a new pattern recognition paradigm that does not rely on experts' prior knowledge. Instead, by inputting a large amount of training data into the model, the model can automatically learn and capture the underlying patterns in the data. In specific research results, Jiang Jiayun et al. proposed an innovative two-layer recognition model. This model first combines the support vector machine and genetic algorithm for preliminary target intention recognition, and then uses the hidden Markov model to further identify the final intention of the target. In addition, Wang and Hu Zhiyong respectively used the XGBoost and random forest models to identify target intentions and both achieved satisfactory results.

[0005] The relevant academic literature 1 (Wang H B, Guan Xin, Yi Xiao, et al. Intention Recog-nition Method Based on Improved Fuzzy Belief Rule[J]. Journal of Electronics&Information Technology, 2023, 45(03): 941-948) proposed an air target intention recognition method based on fuzzy inference. This method constructs a fuzzy inference model and formulates fuzzy inference rules by defining the characteristics of the heading change trend and combining the state information such as the height, speed, and heading of the target as well as expert knowledge. Specifically, using the multi-input single-output characteristic of the fuzzy inference model, the flight speed, height, distance, relative heading angle, and turning trend of the target are used as inputs, and the intention of the target (such as attack, penetration, search, and evacuation) is output. The feasibility and effectiveness of this method are verified through simulation experiments, and the results show that this method can accurately infer the target intention without a large amount of prior knowledge.

[0006] Academic literature 2 (TENG F, GUO X, SONG Y, et al. An Air Target Tactical Intention Recognition Model Based on Bidirectional GRU With Attention Mechanism [J / OL]. IEEE Access, 2021, 9: 169122-169134. DOI: 10.1109 / ACCESS.2021.3135495.) The air target tactical intention recognition model based on BiGRU and attention mechanism proposed has significant innovation in aspects such as feature extraction, network structure design, and application of attention mechanism. This model can effectively handle the dynamic and time-varying characteristics of target states, improving the accuracy and efficiency of air target tactical intention recognition. The experimental results verify the superior performance of the model, making it have broad application prospects in the intention recognition task under complex battlefield environments.

[0007] Academic literature 3 (ZHAO Liang, SUN Peng, ZHANG Jieyong, et al. A Fast Air Target Intention Recognition Method for Difficult-to-Classify Samples [J]. Journal of Air Force Engineering University, 2024, 25(01): 76-82) performs well in dealing with imbalanced data and difficult-to-classify samples. Through the combination of sliding window estimation and spatio-temporal convolutional network, it can effectively handle the imbalance and complexity of data, improving the generalization ability and recognition efficiency of the model. The introduction of self-attention mechanism further enhances the model's ability to capture key features, thus improving the classification accuracy. However, the limitation of this method is its high dependence on data preprocessing. The setting of sliding window parameters requires multiple experiments and adjustments, which increases the debugging difficulty of the model. In addition, the complexity of spatio-temporal convolutional network and self-attention mechanism may lead to higher training time and computational cost of the model, which may limit its efficiency and scalability in practical applications.

[0008] However, these machine learning-based methods have certain limitations in practical applications. They highly rely on the quality and quantity of training data. When the scale of training data is small or contains more noise, the performance of the model often fails to reach the expected effect. In actual combat environments, due to the high complexity of battlefield situations, the data collected by detection devices usually has uncertainty, and the amount of data that can be obtained is limited in specific situations. These factors all limit the wide application of machine learning methods.

[0009] Facing the above problems, it is particularly important to explore the application of algorithms that combine data-driven and knowledge-driven approaches in the field of air combat target intention recognition. Such algorithms can integrate the advantages of expert knowledge and large-scale datasets, improve the robustness of the model in dealing with uncertain and noisy data, and thus achieve more accurate and reliable target intention prediction in complex and changing battlefield environments. In view of this, this paper proposes the BRB-NN algorithm, which combines the advantages of the belief rule base and neural network, and makes full use of expert knowledge and historical data for reasoning and judgment.

[0010] The above three patents basically focus on using deep learning techniques to select relevant variables for PM2.5 concentration prediction. However, the related patents have the following problems: Patent 1 selects too many variables, which easily leads to too high algorithm complexity and insufficient model practicality; the GCN adopted by Patent 2 is a calculation method for the entire graph, and the node features of the entire graph need to be updated in one calculation, which is insufficient for the complex and changing PM2.5 prediction; Patent 3 only considers the PM2.5 time series and does not utilize the spatial information causing PM2.5 pollution, resulting in limited prediction accuracy when facing long-step spatio-temporal prediction.

[0011] A scientific and reasonable method for identifying the intention of air targets is the premise for our side to make good tactical deployments. The existing technologies have the following problems:

[0012] 1. Dependence on expert knowledge: Traditional expert system methods (such as rule-based methods) highly rely on the input of expert knowledge. This not only requires a large amount of preliminary work to organize and structure expert knowledge, but also is limited by the experience scope and subjective judgment of individual experts;

[0013] 2. Poor adaptability: Traditional rule-based methods usually have fewer parameters and limited learning ability, and it is difficult to adapt to rapidly changing data environments and complex and changing battlefield situations;

[0014] 3. Dependence on a large amount of data: These methods highly rely on the quality and quantity of training data. When the scale of training data is small or contains more noise, the performance of the model often fails to reach the expected effect. Summary of the Invention

[0015] Aiming at the defects existing in the prior art, the purpose of the present invention is to provide a method for identifying target intention based on the BRB-NN data and knowledge joint-driven algorithm, to provide a new idea for solving the problem of air combat target intention recognition, combine the belief rule base algorithm with the deep learning algorithm, and improve the accuracy of air combat target intention recognition.

[0016] To achieve the above purpose, the technical solution of the present invention is as follows:

[0017] The method for identifying the intention of air targets based on data and knowledge joint driving is as follows:

[0018] Step 1: Collect the behavioral variable data of airborne targets and perform data cleaning to obtain x;

[0019] Step 2: Obtain expert knowledge through the Delphi method and construct a belief rule base;

[0020] Step 3: Input the data into the belief rule base module to obtain the embedded representation x of expert knowledge. e , concatenate the embedded representation of expert knowledge and the data to form <x, x e >, and input them into the neural network module together;

[0021] Step 3.1: First, for each rule in the belief rule base, the premise conditions of the rule usually contain no less than 2 attributes. Compare the feature values of the data with the antecedent attributes of the rule one by one.

[0022] Step 3.2: Match the fuzzified data with each rule and calculate the activation weight of each rule. Determine the calculation method of the activation weight w k of the k-th rule, δ i represents the importance degree of the i-th antecedent attribute relative to other attributes, and θ k is the rule weight, indicating the importance of the k-th rule relative to other rules. represents the matching degree between the input data and the j-th reference value of the i-th conditional attribute in the k-th rule. The formula

[0023] 1 and formula 2 convert the matching degree into the activation weight of each rule in the rule base;

[0024]

[0025] After obtaining the activation weight of each rule, use the Dempster combination method to aggregate the consequent information of each rule, so as to obtain the inference result. The formulas of the Dempster combination method are as shown in formulas 3 and 4:

[0026]

[0027] where β j represents the confidence degree of the j-th result. The closer it is to 1, the higher the confidence degree of the j-th result; the closer it is to 0, the lower the confidence degree of the j-th result. β j is the embedded representation x of knowledge e .

[0028] Step 3.3: The knowledge embedding vector x e is already a numerical vector. The processed original data vector x and the knowledge embedding vector x eConcatenate them in a certain order to form a new vector <x, x e >.

[0029] Finally, use the concatenated vector <x, x e > as the input of the neural network module to provide a data basis for subsequent model training and intent recognition tasks.

[0030] Step 4: Iteratively optimize the parameters of the belief rule base and the neural network in a collaborative optimization manner until the model converges.

[0031] Preferably, Step 1 further includes the following steps:

[0032] Step 1.1: Select appropriate domain experts in the digital field, with the number controlled between 3 and 7;

[0033] Step 1.2: Let each expert independently provide their own knowledge in the field of target intent recognition and organize it in the form of a belief rule base;

[0034] Step 1.3: Let each expert evaluate the knowledge provided by other experts and adjust their own belief rule base according to the belief rule base provided by other experts;

[0035] Step 1.4: Repeat the above operations repeatedly until the opinions of the experts reach an agreement, and organize the final and consistent belief rule base.

[0036] Preferably, Step 2 specifically includes the following steps:

[0037] Step 2.1: Calculate the matching degree between the data and each rule in the belief rule base module, and then calculate the activation degree of the rule;

[0038] Step 2.2: Calculate the intent distribution of each piece of data through the activation degree and the consequent of the belief rule base, and represent this distribution as a vector x e ;

[0039] Step 2.3: Concatenate the original data and the knowledge embedding vector to form <x, x e > and input them into the neural network module together.

[0040] Preferably, Step 4 specifically includes the following steps:

[0041] Step 4.1, fix the parameters of the belief rule base module, and optimize the neural network module of the model using the backpropagation algorithm. Before starting the optimization, ensure that the weights and biases of the neural network have been randomly initialized. Input the training data into the neural network and perform forward propagation calculations according to the network structure. Each layer of neurons performs weighted summation based on the input data and the weights and biases of the current layer, and obtains the output through the activation function. These outputs will be used as the input for the next layer until the prediction result is output by the last layer. The operation logic of each layer is shown in Formulas 5 and 6:

[0042] Z = W * X + b (5)

[0043] y = sigmoid(Z) (6)

[0044] where W and b are the parameters to be learned in the neural network model.

[0045] According to the calculated gradient, use the optimizer SGD to update the weights and biases of the neural network. The weight update formula 7 of SGD is:

[0046]

[0047] where X is the weight, α is the learning rate, is the gradient of the weight.

[0048] Step 4.2, fix the neural network module, and optimize the belief rule base module of the model using the differential evolution algorithm. The differential evolution algorithm first needs to initialize a population, and each individual in the population represents a parameter setting of the belief rule base. The parameters include the antecedent attribute weights of the rules, the consequent confidence degrees, etc. For each individual, a set of initial values that meet the parameter value ranges are randomly generated.

[0049] The output of the belief rule base is the confidence degree vector of each category, expressed as β = (β1, β2, …, β m ), where

[0050] The goal of parameter optimization is to minimize the cross-entropy between (y n , β n ) to make the distributions of y n and β n closer to being consistent. The objective function is shown in Formula 8.

[0051]

[0052] In addition, the parameter optimization process of the BRB is also restricted by the following constraints:

[0053] 1) The confidence degree of the consequent of each rule is between 0 and 1.

[0054] 0 ≤ β nk ≤ 1, n = 1, 2, ..., N

[0055] 2) Each rule is complete, that is, the sum of the consequent confidence degrees of the k-th rule is 1.

[0056]

[0057] 3) The weight of each rule should be between 0 and 1

[0058] 0 ≤ θ k ≤ 1, k = 1, 2, …, L

[0059] The weight of each feature should be between 0 and 1

[0060] 0 ≤ δ i ≤ 1, i = 1, 2, …, T

[0061] Optimize the above parameters of the belief rule base through the differential evolution algorithm. When the fitness value of the optimal individual no longer has a significant improvement in several consecutive generations, it is considered that the optimization of the belief rule base has converged. At this time, the belief rule base reaches a relatively optimal state under the current fixed neural network.

[0062] Step 4.3: Continuously repeat Step 4.1 and Step 4.2, that is, alternately fix the belief rule base to optimize the neural network, and then fix the neural network to optimize the belief rule base.

[0063] Compared with the existing technical solutions, the beneficial effects of the present invention are as follows:

[0064] 1. The present invention makes full use of the role of expert knowledge in the process of intention recognition, improving the accuracy of the model.

[0065] 2. Good intention recognition accuracy can still be obtained when the amount of data is not large.

[0066] 3. The natural interpretability of the belief rule base breaks through the limitation of the black box of the deep learning model, enabling researchers to see the internal structure of the model. Brief Description of the Drawings

[0067] Figure 1 is a flowchart of a method for target intention recognition based on the joint drive of knowledge and data according to the present invention;

[0068] Figure 2 is the framework of the BRB-NN model;

[0069] Figure 3 is an example diagram of the collaborative optimization algorithm;

[0070] Figure 4(a) is a comparison of different learning rates (loss values);

[0071] Figure 4(b) is the comparison of different learning rates (accuracy);

[0072] Figure 4(c) is the comparison of different batch sizes (loss value);

[0073] Figure 4(d) is the comparison of different batch sizes (accuracy);

[0074] Figure 4(e) is the comparison of different neural network structures (loss value);

[0075] Figure 4(f) is the comparison of different neural network structures (accuracy);

[0076] Figure 5 It is a comparison diagram of the predicted value and the true value of the BRB-NN model on the test set;

[0077] Figure 6(a) is the weights of different rules;

[0078] Figure 6(b) is the weights of each feature in different rules. Detailed implementation manners

[0079] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0080] A target intention recognition method based on the joint drive of knowledge and data includes the following steps:

[0081] Step 1, collect the flight data of aircraft with different intentions and relevant variable data in the simulation system, and perform data cleaning and reduction;

[0082] Each data sample represents the flight behavior data of an enemy aircraft and is a multi-feature time series data. To ensure the consistency of the input model data, each sample data needs to be adjusted to have the same number of frames and the same number of features. In this patent, all sample data are uniformly adjusted into 12-frame time series data.

[0083] Since the dimensions of different features are different, it is also necessary to normalize the data of all features and compress the values between 0 and 1. The normalization formula is as follows:

[0084]

[0085] Among them, Xi refers to the sample in the random variable X, which refers to the behavior data of a single aircraft, X min 、X maxRefers to the minimum and maximum values in the data column;

[0086] Step 2: Obtain expert knowledge through the Delphi method and construct a belief rule base. The selected experts should have profound professional knowledge, rich practical experience, and excellent analytical and judgment abilities in the field of target intention recognition. The number of experts is controlled between 3 and 7, which can not only ensure the full exchange of diverse views but also avoid the problems of high communication costs and difficult integration of opinions caused by too many people. Provide detailed background information and clear problem guidance to the experts so that they can sort out the knowledge in the field of target intention recognition from different perspectives based on their professional knowledge and experience. When organizing, convert the expert knowledge into a rule form composed of elements such as premise attributes, result attributes, and corresponding confidence levels according to the specific format specifications of the belief rule base to ensure the systematicness and operability of the knowledge. Anonymously distribute the belief rule bases sorted out by each expert to other experts for evaluation. When evaluating, the experts need to elaborate in detail the reasons for approval or disapproval and put forward specific opinions on the rationality, integrity, applicability, etc. of the rules. After receiving the feedback, the experts re-examine their own belief rule bases in combination with the opinions of others and revise and supplement the controversial or imperfect parts. Set clear consistency judgment criteria, such as the confidence judgment deviation of key rules within a certain range and no obvious differences in the core content of the rules. In each round of the repetition process, pay attention to the effective transmission and feedback of information, continuously narrow the opinion gap among the experts until the pre-set consistency standard is reached, so as to sort out the final belief rule base jointly recognized by the experts in the field of target intention recognition.

[0087] Table 1 Sample of Belief Rule Base

[0088]

[0089] Table 1 is a sample of a belief rule base. Symbols such as β, D, and V in the rule antecedent represent different features. Each row in the rule base represents a knowledge-based judgment. "L", "ML", "M", "MH", and "H" in the rule antecedent represent low, medium-low, medium, medium-high, and high respectively, while "any" represents any value. The first rule means that if the azimuth angle (β) value of an air target is low, the distance (D) is far, the speed (V) is medium-fast, and the flight direction (θ) angle is medium, then the intention of the target is reconnaissance.

[0090] Step 3: Input the data into the belief rule base module to obtain the embedded representation x of the expert knowledge e , concatenate the embedded representation of the expert knowledge and the data to form <x, x e >, and input them into the neural network module together.

[0091] Step 3.1: First, for each rule in the belief rule base, the premise conditions of the rule usually contain multiple attributes. Compare the feature values of the data with the antecedent attributes of the rule one by one. For example, if the antecedent attribute of the rule is a numerical range, calculate the membership degree of the data feature value within this numerical range. The membership degree function in fuzzy mathematics can be used, such as triangular membership degree function, Gaussian membership degree function, trapezoidal membership degree function, etc. (this patent selects the triangular membership degree function) to quantify the matching degree α between the data and the antecedent attribute of the rule;

[0092] Step 3.2: Match the fuzzified data with each rule and calculate the activation weight of each rule. The following formula determines the calculation method of the activation weight w k of the k-th rule, where similarly, δ i represents the importance degree of the i-th antecedent attribute relative to other attributes, and θ k is the rule weight, representing the importance of the k-th rule relative to other rules. represents the matching degree between the input data and the j-th reference value of the i-th conditional attribute in the k-th rule. Formula 1 converts the matching degree into the activation weight of each rule in the rule base;

[0093]

[0094] After obtaining the activation weight of each rule, use the Dempster combination method to aggregate the consequent information of each rule, thereby obtaining the inference result. The formula of the Dempster combination method is as follows:

[0095]

[0096]

[0097] The above formula shows the inference process of the belief rule base, where β j represents the confidence degree of the j-th result. The closer it is to 1, the higher the confidence degree of the j-th result; the closer it is to 0, the lower the confidence degree of the j-th result. β j is the embedded representation x e of knowledge.

[0098] Step 3.3: The knowledge embedding vector x e is already a numerical vector. Concatenate the processed original data vector x and the knowledge embedding vector x e in a certain order to form a new vector <x, x e >. For example, if the dimension of the original data vector x is n and the dimension of the knowledge embedding vector x e is m, then the dimension of the concatenated vector is n + m.

[0099] Finally, the concatenated vector <x, x e > is used as the input to the neural network module, providing a data basis for subsequent model training and intent recognition tasks. The neural network can learn the complex relationship between data features and intents from the input data, thereby achieving accurate intent recognition. The average envelope line is obtained by averaging the two extreme value curves;

[0100] Step 4: Iteratively optimize the parameters of the belief rule base and the neural network through a collaborative optimization algorithm until the model converges. Collaborative optimization aims to give full play to the respective advantages of the belief rule base and the neural network. By alternately optimizing the parameters of the two, the overall performance of the model is optimized. In each iteration, targeted adjustments are made to the belief rule base and the neural network, gradually reducing the loss value of the model until the model converges.

[0101] Step 4.1: Fix the parameters of the belief rule base module and optimize the neural network module of the model using the backpropagation algorithm. Before starting the optimization, ensure that the weights and biases of the neural network have been randomly initialized. Input the training data into the neural network and perform forward propagation calculations according to the network structure. Each layer of neurons performs a weighted sum based on the input data and the weights and biases of the current layer, and obtains the output through an activation function. These outputs will be used as the input for the next layer until the prediction result is output by the last layer. The operation logic of each layer is as follows:

[0102] Z = W * X + b

[0103] y = sigmoid(Z)

[0104] where W and b are the parameters to be learned in the neural network model. The parameter optimization of the neural network is usually achieved through the backpropagation mechanism. The backpropagation mechanism is based on the calculated loss value. Starting from the output layer, it uses the chain rule to backpropagate and calculate the gradients of the loss value with respect to the weights and biases of each layer. During the backpropagation process, each layer calculates the gradients of the weights and biases of that layer according to the gradient information passed from the next layer, combined with its own input and the derivative of the activation function. According to the calculated gradients, the optimizer SGD is used to update the weights and biases of the neural network. The weight update formula of SGD is:

[0105]

[0106] where W is the weight and α is the learning rate, is the gradient of the weight. The update formula for the bias is similar. The learning rate determines the step size of each parameter update. It is a hyperparameter that needs to be adjusted according to experiments. When the decrease in the loss value is less than a certain threshold (such as 0.0001) in several consecutive iterations, it is considered that the neural network has converged. At this time, the neural network reaches a relatively optimal state under the current fixed confidence rule base parameters;

[0107] Step 4.2, fix the neural network module and optimize the confidence rule base module of the model using the differential evolution algorithm. The differential evolution algorithm first needs to initialize a population, and each individual in the population represents a parameter setting of the confidence rule base. The parameters include the weights of the antecedent attributes of the rules, the consequent confidence degrees, etc. For each individual, a set of initial values that meet the parameter value ranges are randomly generated.

[0108] The output of the confidence rule base is the confidence degree vector of each category, denoted as β = (β1, β2,..., β m ), where Suppose there are N feature-label pairs (X n , y n ). After each feature-label pair enters the BRB, an inference result can be generated, that is, the knowledge embedding information x e . Then, inputting <x, x e > into the neural network model can obtain the intention distribution β n predicted by the model. The goal of parameter optimization is to minimize the cross-entropy between (y n , β n ), so that the distributions of y n and β n are closer to being consistent. The objective function is shown in Equation 5.

[0109]

[0110] In addition, the parameter optimization process of the BRB is also restricted by the following constraints:

[0111] 1) The confidence degree of the consequent of each rule is between 0 and 1.

[0112] 0 ≤ β nk ≤ 1, n = 1, 2,..., N

[0113] 2) Each rule is complete, that is, the sum of the consequent confidence degrees of the k-th rule is 1.

[0114]

[0115] 3) The weight of each rule should be between 0 and 1

[0116] 0 ≤ θ k ≤ 1, k = 1, 2,..., L

[0117] The weight of each feature should be between 0 and 1

[0118] 0 ≤ δ i ≤ 1, i = 1, 2, ..., T

[0119] Optimize the above parameters of the belief rule base through the differential evolution algorithm. When the fitness value of the optimal individual no longer has a significant improvement (such as the improvement amplitude is less than a certain threshold) in several consecutive generations, it is considered that the optimization of the belief rule base has converged. At this time, the belief rule base reaches a relatively optimal state under the current fixed neural network.

[0120] Step 4.3: Continuously repeat Step 4.1 and Step 4.2, that is, alternately fix the belief rule base to optimize the neural network, and then fix the neural network to optimize the belief rule base. In each iteration process, calculate the overall loss value of the model and monitor its change. When the decrease amplitude of the overall loss value of the model is less than a pre-set threshold (such as 0.001) in several consecutive iterations, it is considered that the model has reached the convergence state. At this time, the parameters of the belief rule base and the neural network reach a relatively optimal combination, making the model perform best in processing the target intention recognition task.

[0121] The present invention is proved to be feasible by comparing with the existing prediction model method through experiments.

[0122] Enlightened by the above ideal embodiments according to the present invention, through the above description, those skilled in the art can make various changes and modifications completely within the scope of not departing from the technical idea of the present invention. The technical scope of the present invention is not limited to the content in the specification, and its technical scope must be determined according to the scope of the claims.

Claims

1. A method for recognizing the intention of aerial targets driven jointly by data and knowledge, characterized in that, The steps are as follows: Step 1: Collect the behavioral variable data of airborne flying targets and perform data cleaning to obtain x; Step 2: Obtain expert knowledge through the Delphi method and construct a belief rule base; Step 3: Input the data into the belief rule base module to obtain the embedded representation x of the expert knowledge e , concatenate the embedded representation of the expert knowledge and the data to form <x, x e >, and input them into the neural network module together; specifically as follows: Step 3.1: For each rule in the belief rule base, the premise conditions of the rule usually contain no less than 2 attributes; compare the eigenvalue of the data with the antecedent attributes of the rule one by one; Step 3.2: Match the fuzzified data with each rule and calculate the activation weight of each rule; Determine the activation weight w of the k-th rule k The calculation method of δ i represents the importance degree of the i-th premise attribute relative to other attributes, and θ k is the rule weight, representing the importance of the k-th rule relative to other rules; represents the matching degree between the input data and the j-th reference value of the i-th conditional attribute in the k-th rule; Formulas 1 and 2 convert the matching degree into the activation weight of each rule in the rule base; After obtaining the activation weight of each rule, use the Dempster combination method to aggregate the consequent information of each rule to obtain the inference result; the formulas of the Dempster combination method are shown in Equations 3 and 4: where β j represents the confidence of the j-th type of result. The closer it is to 1, the higher the confidence of the j-th result; the closer it is to 0, the lower the confidence of the j-th result. β j is the embedded representation x of the knowledge e ; Step 3.3, knowledge embedding vector x e is already a numerical vector; concatenate the processed original data vector x and the knowledge embedding vector x e in a certain order to form a new vector <x, x e >; Finally, the concatenated vector <x, x e > is used as the input of the neural network module to provide a data basis for subsequent model training and intent recognition tasks; Step 4: Iteratively optimize the parameters of the belief rule base and the neural network in a collaborative optimization manner until the model converges.

2. The method for identifying the intention of an aerial target jointly driven by data and knowledge as claimed in claim 1, wherein The specific content of the above-mentioned Step 1 is as follows: Step 1.1: Select appropriate digital domain experts, and the number is controlled between 3 and 7; Step 1.2: Let each expert independently give their own knowledge in the field of target intention recognition and organize it in the form of a belief rule base; Step 1.3: Let each expert evaluate the knowledge given by other experts and adjust their own belief rule base according to the belief rule base given by other experts; Step 1.4: Repeat the above operations repeatedly until the opinions of the experts reach an agreement, and sort out the final and consistent belief rule base.

3. The method for identifying the intention of an aerial target driven jointly by data and knowledge as claimed in claim 1, wherein The specific content of the above-mentioned Step 2 is as follows: Step 2.1: Calculate the matching degree between the data and each rule in the belief rule base module, and then calculate the activation degree of the rule; Step 2.2, calculate the intention distribution of each piece of data through the activation degree and the consequent part of the confidence rule base, and represent this distribution as a vector x e ; Step 2.3, concatenate the original data and the knowledge embedding vectors to form <x, x e >, and input them into the neural network module together.

4. The method for identifying the intention of an aerial target driven jointly by data and knowledge according to claim 1, wherein, The specific content of the above-mentioned Step 4 is as follows: Step 4.1: Fix the parameters of the belief rule base module and optimize the neural network module of the model using the backpropagation algorithm; before starting the optimization, ensure that the weights and biases of the neural network have been randomly initialized; input the training data into the neural network and perform forward propagation calculations according to the structure of the network; each layer of neurons performs weighted summation according to the input data and the weights and biases of the current layer, and obtains the output through the activation function, and the output will be used as the input of the next layer until the prediction result is output by the last layer; the operation logic of each layer is shown in Equations 5 and 6: Z = W * X + b (5) y = sigmoid(Z) (6) Where W and b are the parameters to be learned in the neural network model; According to the calculated gradient, use the optimizer SGD to update the weights and biases of the neural network; the weight update formula 7 of SGD is: where W is the weight and α is the learning rate, is the gradient of the weight; Step 4.2, fix the neural network module and optimize the belief rule base module of the model using the differential evolution algorithm; the output of the belief rule base is the confidence vector for each category, denoted as β = (β1, β2,..., β m ), where the objective of parameter optimization is to minimize the cross-entropy between (y n , β n ), making the distributions of y n and β n closer to being consistent, and the objective function is shown in Equation 8; Step 4.3: Continuously repeat Step 4.1 and Step 4.2, that is, alternately fix the belief rule base to optimize the neural network, and then fix the neural network to optimize the belief rule base.

5. The method for identifying the intention of an aerial target driven jointly by data and knowledge as claimed in claim 4, wherein In the above-mentioned Step 4.2, the parameter optimization process is also restricted by the following constraints: 1) The confidence of the consequent of each rule ranges from 0 to 1; 0 ≤ β nk ≤ 1, n = 1, 2, …, N 2) Each rule is complete, that is, the sum of the consequent confidences of the k-th rule is 1. 3) The weight of each rule should be between 0 and 1 0 ≤ θ k ≤ 1, k = 1, 2, ..., L. The weight of each feature should be between 0 and 1. 0 ≤ δ i ≤ 1, i = 1, 2, ..., T.

Images