Air target intention recognition method, device, equipment and medium

Abstract

This invention provides a method, apparatus, device, and medium for identifying the intent of aerial targets, comprising the following steps: acquiring continuously collected multidimensional feature data of aerial targets to construct a target temporal feature dataset, and preprocessing and labeling each sample in the dataset with target intent tags to construct a training dataset; constructing a domain rule set expressing the logical constraint relationships between target intents using first-order predicate logic rules; constructing a neural symbolic joint reasoning model, and jointly training the neural symbolic joint reasoning model using the training dataset and the domain rule set to obtain a trained neural symbolic joint reasoning model; and predicting the target intent based on the temporal feature data of the aerial target to be predicted using the trained neural symbolic joint reasoning model. This invention achieves an organic unity of data-driven and knowledge-driven approaches by embedding logical rules in a differentiable form into the training objective and reasoning structure of the neural network.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and specifically to a method, apparatus, device, and medium for identifying the intent of aerial targets. Background Technology

[0002] Accurate and timely identification and prediction of the flight intentions of aerial targets are crucial for enhancing airspace situational awareness and dispatch decision-making capabilities. Aerial target intention identification aims to infer the mission and action plans of aerial targets based on their flight status, behavioral patterns, and contextual situation. However, due to the diversity of aerial target types, complex movement patterns, and rapid situational evolution, this task faces multiple challenges, including multi-source information fusion, temporal dependency modeling, and high-level semantic reasoning.

[0003] Currently, publicly available technologies for aerial target intent recognition mainly fall into two categories: One category is based on logical rules and template matching. This method summarizes historical experience and domain knowledge to construct a predefined rule base and behavior templates, and then infers intent based on logical rule chains. While this type of method possesses good interpretability and knowledge controllability, it suffers from limited rule coverage and weak adaptability when facing unpredictable flight behaviors, complex environmental changes, and uncertain information, making it difficult to meet the generalization requirements of highly dynamic situations. The other category is based on deep learning models using neural networks, such as recurrent neural networks (RNNs) and gated recurrent units (GRUs). These models can automatically extract features from multi-dimensional time-series data such as heading, speed, and altitude, and achieve end-to-end intent classification. While this type of method exhibits strong fitting ability at the perception level, its decision-making process is opaque, heavily reliant on high-quality labeled data, and lacks embedding constraints based on physical common sense and rules. Therefore, in practical applications, it is prone to output results that violate dynamic laws, affecting the system's reliability and robustness.

[0004] In summary, existing technologies have not yet effectively addressed the balance between "interpretable reasoning" and "adaptive learning" in dynamic environments for aerial target intent recognition tasks. There is an urgent need to develop a new recognition framework that combines semantic transparency and scene generalization capabilities. Summary of the Invention

[0005] To address the problems of poor adaptability and weak interpretability in existing target intent recognition methods based on a single data source or purely data-driven approaches, this invention provides an aerial target intent recognition method, apparatus, device, and medium, aiming to improve the accuracy, robustness, and decision reliability of intent recognition.

[0006] To achieve the above objectives, the present invention provides the following technical solution:

[0007] On the one hand, a method for identifying the intent of aerial targets is provided, including the following steps:

[0008] Multidimensional feature data of continuously collected aerial targets are acquired to construct a target temporal feature dataset. Each sample in the dataset is preprocessed and labeled with target intent to construct a training dataset.

[0009] Construct a domain rule set that describes the logical constraint relationships between target intentions, expressed by first-order predicate logic rules;

[0010] A neural symbolic joint reasoning model is constructed, and the neural symbolic joint reasoning model is jointly trained using a training dataset and a domain rule set to obtain a trained neural symbolic joint reasoning model.

[0011] The target intent is predicted based on the temporal feature data of the aerial target to be predicted using a pre-trained neural symbolic joint reasoning model.

[0012] On the other hand, an aerial target intent recognition device is provided, comprising:

[0013] The training dataset module is used to acquire multi-dimensional feature data of continuously collected aerial targets to construct a target temporal feature dataset, and to preprocess and label each sample in the dataset with target intent labels to construct the training dataset.

[0014] The domain rule set module is used to construct a domain rule set that describes the logical constraint relationships between target intentions, expressed by first-order predicate logic rules.

[0015] The model training module is used to construct a neural symbolic joint reasoning model. The neural symbolic joint reasoning model is jointly trained using a training dataset and a domain rule set to obtain a trained neural symbolic joint reasoning model.

[0016] The prediction module is used to predict the target intent based on the temporal feature data of the aerial target to be predicted, using a trained neural symbolic joint reasoning model.

[0017] On the other hand, the present invention provides an electronic device including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of the above-described aerial target intent recognition method.

[0018] On the other hand, the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the steps of the above-described method for identifying the intent of an aerial target.

[0019] On the other hand, the present invention provides a computer program product stored on a computer-readable storage medium and including computer instructions that, when executed by a processor, cause an electronic device to implement the steps of the above-described method for identifying the intent of an aerial target.

[0020] Compared with existing technologies, the aerial target intent recognition method provided by this invention achieves an effective unification of model performance and interpretability, data-driven learning and knowledge-driven reasoning through deep coupling of neural and symbolic computation. Specifically, the beneficial effects of this invention are as follows:

[0021] 1) Improve the accuracy and environmental robustness of intent recognition. By introducing logical rules as symbolic constraints, the system effectively corrects erroneous inferences of neural networks under conditions of data noise, scarce samples, or long-tailed distributions, preventing them from outputting absurd results that violate physical common sense or business logic, thus enhancing the reliability of the system in complex real-world scenarios.

[0022] 2) Achieve interpretability and traceability of the decision-making process. This fundamentally improves upon the black-box limitations of pure deep learning models. This framework can clearly indicate the key features that trigger the final decision and the applicable logical rules, meeting the rigid requirements of transparency and accountability in algorithmic decision-making in high-reliability domains.

[0023] 3) Optimize reasoning efficiency and computational resource utilization. By employing a reflexive mechanism, potential inconsistencies in the neural network output are quickly located, focusing the search scope of symbolic reasoning on a limited number of problem points. Regardless of the input data size, the system only needs to perform a single targeted call to the knowledge base to complete the correction, significantly reducing the computational overhead of joint reasoning and enhancing the system's practicality in scenarios with high timeliness requirements. Attached Figure Description

[0024] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0025] Figure 1 This is a flowchart of an embodiment of an aerial target intent recognition method;

[0026] Figure 2 This is a block diagram illustrating the principle of a neural symbolic joint reasoning model in one embodiment;

[0027] Figure 3 This is a block diagram of an electronic device in one embodiment. Detailed Implementation

[0028] The technical solution of the present invention will now be clearly and completely described through specific embodiments. Obviously, the described embodiments are merely some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0029] In another embodiment, reference is made to Figure 1 A method for identifying the intent of an aerial target is provided, comprising the following steps:

[0030] Multidimensional feature data of continuously collected aerial targets are acquired to construct a target temporal feature dataset. Each sample in the dataset is preprocessed and labeled with target intent to construct a training dataset.

[0031] Construct a domain rule set that describes the logical constraint relationships between target intentions, expressed by first-order predicate logic rules;

[0032] A neural symbolic joint reasoning model is constructed, and the neural symbolic joint reasoning model is jointly trained using a training dataset and a domain rule set to obtain a trained neural symbolic joint reasoning model.

[0033] The target intent is predicted based on the temporal feature data of the aerial target to be predicted using a pre-trained neural symbolic joint reasoning model.

[0034] In one embodiment, a method for constructing a training dataset is provided as follows:

[0035] Continuously collect multidimensional feature data of aerial targets to construct a target time-series feature dataset. The multidimensional feature data includes: relative distance and position of aerial targets, altitude, speed, acceleration, heading angle, radar on / off status, and electronic jamming device status.

[0036] Binarize the non-numerical feature data in the target time-series feature dataset and perform time-series normalization on the numerical feature data.

[0037] Based on the preset target intent space, the corresponding intent labels are labeled for each target time-series feature data in the target time-series feature dataset, thus completing the construction of a training dataset with target intent labels.

[0038] In another embodiment, the constructed training dataset comprises two subsets, namely a training dataset with target intent labels. and training dataset without target intent labels Training dataset , , The number of samples in each is respectively , Training dataset Number of samples Training dataset with target intent labels The samples in the data consist of target time-series feature data. and their corresponding target intent tags Composition, training dataset without target intent labels The samples in the sample only contain target time series feature data Specifically, the method for constructing the training dataset is as follows:

[0039] Continuously collect multidimensional feature data of aerial targets to construct a target time-series feature dataset. The multidimensional feature data includes: relative distance and position of aerial targets, altitude, speed, acceleration, heading angle, radar on / off status, and electronic jamming device status.

[0040] Binarize the non-numerical feature data in the target time-series feature dataset and perform time-series normalization on the numerical feature data.

[0041] Based on a predefined target intent space, corresponding intent labels are annotated for a portion of the target temporal feature data in the target temporal feature dataset, thus completing a training dataset with target intent labels. The construction of the target temporal feature dataset involves using the remaining unlabeled target temporal feature data from the target temporal feature dataset to construct a training dataset without target intent labels. .

[0042] In practical applications, target data can be collected through satellites, drones, radar, and various sensing devices. The determination of intent patterns relies on relevant factors, including but not limited to: the aerial target's position, altitude, speed, acceleration, heading angle, radar on / off status, electronic jamming release, and historical spatiotemporal information of various equipment. Based on this collected feature data, including numerical and non-numerical feature data, a multi-dimensional feature space can be constructed. This multi-dimensional feature data can include: the relative distance and position, altitude, speed, acceleration, heading angle, radar on / off status, and electronic jamming device status of the aerial target; the relative distance and position, altitude, speed, acceleration, and heading angle of the aerial target are numerical feature data, while the radar on / off status and electronic jamming device status are non-numerical feature data.

[0043] To overcome the reliability limitations of single-moment identification methods in constantly changing and complex environments, this invention employs a continuous time series modeling approach. The continuous acquisition duration is... m Real-time multidimensional feature data of aerial targets within seconds, from arrive The time series constitutes the target temporal feature dataset ,in l For feature dimension, m In terms of time dimension, n This represents the number of samples.

[0044] Non-numerical feature data in the target temporal feature dataset are binarized and encoded, such as encoding interference status and radar status as 1 (on) and 0 (off), respectively. Numerical feature data undergoes temporal normalization to improve model convergence speed, recognition accuracy, and suppress gradient explosion, thereby enhancing training stability and overall performance.

[0045] Based on typical behavioral patterns of aerial targets, a standardized set of target intent categories is predefined, and domain expert knowledge is transformed into structured intent labels through a discretization coding mechanism, thereby constructing a target intent space. This mapping relationship provides the model with identifiable supervised training objectives, realizing a systematic transformation from expert cognitive experience to machine learning labels. Based on a preset target intent space, the samples in the target temporal feature dataset are labeled with corresponding intent labels (optionally, some samples are labeled with target intent labels while others are left unlabeled; this labeling is used to construct a training dataset with target intent labels). Unlabeled training dataset with unlabeled target intent labels Complete the training dataset. The construction, in which, Represents the first in the target time series feature dataset One sample (i.e., target time-series feature data). express The corresponding target intent category label.

[0046] Construct a domain rule set that describes the logical constraint relationships between target intentions, expressed by first-order predicate logic rules. , Indicates the first This invention constructs a first-order predicate logic rule. The neural symbolic joint reasoning model includes a neural network model f, used to achieve end-to-end mapping from target temporal feature data input to target intent category output. The input to the neural symbolic joint reasoning model is the training dataset. and a finite set of rules consisting of first-order predicate logic rules. The neural symbolic joint reasoning model includes a domain knowledge base KB for storing domain rule sets. The first-order predicate logic rules, expressed by first-order predicate logic rules, describe the logical constraints between target intentions. These rules are expressed as first-order predicate logic rules and can perform symbolic reasoning through a Markov logic network. In the neural symbolic joint reasoning model, the target intention category output by the model must satisfy the logical constraints between target intentions described by first-order predicate logic rules stored in the domain knowledge base KB, to ensure that the reasoning results are consistent with the domain knowledge and avoid output errors caused by logical conflicts. The domain rule set consists of basic rules, default rules, and weighted rules. Variables in the first-order predicate logic rules are instantiated as constants to construct a closed set of atoms, which are then used as nodes to construct a Markov logic network. Rule weights are mapped to edge weights to form a joint probability distribution, thereby constructing a closed Markov logic network. The reasoning method used in this invention is to transform the Markov logic program into a weakly constrained logic program, and then solve it using logic inference engines such as Clingo and DLV.

[0047] The neural symbolic joint reasoning model is jointly trained based on the aforementioned training dataset and domain rule set. The trained model, along with a test set, is then used to identify the target intent of aerial targets. During the neural network output stage, the model synchronously generates a binary reflection vector to identify output positions that potentially conflict with domain knowledge. This vector is then corrected using a symbolic reasoning mechanism, ultimately generating a recognition result that conforms to domain knowledge constraints.

[0048] In one embodiment, reference is made to Figure 2 The neural symbolic joint reasoning model is jointly trained using a training dataset and a domain rule set, specifically including the following steps:

[0049] Samples from the training dataset Transformer encoder input to neural network To obtain a high-dimensional embedding representation This includes feature information of the target time series feature data;

[0050] High-dimensional embedding representation Input to a module consisting of a linear layer and a softmax layer Preliminary target intent recognition results were obtained. That is, to obtain the intuitive output of the model;

[0051] Transformer encoder A reflection layer was then added. High-dimensional embedding representation Simultaneously input into the reflection layer Reflection layer It is a Sigmoid layer that performs binarization by setting a threshold, generating a result similar to... Reflection vectors with the same dimensions Reflective Vectors As a binary classification identifier, it indicates the corresponding output. Does it have any potential logical conflict with the domain rule set? This indicates a potential logical conflict. This indicates that the logical constraint relationship between the target intent and the domain rule set is consistent and there is no conflict.

[0052] Based on reflection vector The initial target intent recognition results are corrected, and the corrected target intent recognition results are output. :

[0053] ;

[0054] in This indicates the preliminary target intent identification results. There is a potential logical conflict with the domain rule set. The abductive reasoning process will be marked as having potential logical conflicts. The blank space is filled abductively by invoking symbolic reasoning based on Markov logic networks, generating a corrected target intent recognition result that is consistent with the logical constraints between target intents in the domain rule set. . This represents the abductive reasoning process, an attention mechanism that can quickly help symbolic reasoning focus on specific areas it identifies incorrectly, ultimately outputting... Ensure consistency between the logical constraints and the target intent in the domain rule set.

[0055] The constructed comprehensive loss function includes: a cross-entropy loss function based on the preliminary target intent recognition results and the true target intent labels; a consistency loss function based on the reflection vector, used to guide the target intent recognition results output by the model to conform to the logical constraint relationship between target intents in the domain rule set; a reflection size loss function based on the number of conflict-marked values ​​in the reflection vector, used to control the sparsity of reflection; and a comprehensive loss function obtained by combining the cross-entropy loss function, the consistency loss function, and the reflection size loss function.

[0056] With the goal of minimizing the comprehensive loss function, the model parameters of the neural symbol joint inference model are iteratively updated until convergence, thus obtaining the trained neural symbol joint inference model.

[0057] if If all elements conform to the constraints in the domain rule set, the consistency measure is 1; otherwise, it is 0. Let the output after error removal be... Specifically, a consistency loss function is constructed based on the reflexive vector, including:

[0058] Define consistency metrics ,in , express Consistency with the domain knowledge base KB all elements If all of them are consistent with the logical constraints between target intents in the domain knowledge base KB, then It is 1 if it is not 0 otherwise; It contains elements that potentially conflict with the domain knowledge base KB. So in Remove elements that have potential logical conflicts with the domain knowledge base (KB). The output after is , express Consistency with the domain knowledge base (KB).

[0059] The higher the value, the stronger the reflection vector. More effective detection The inconsistencies in the training process. The goal of training is to guide the neural network's parameter generation to maximize this value, given... Typically, these are discrete values. The REINFORCE algorithm is used to achieve this goal, optimizing the policy by maximizing a specified reward. A consistent loss function is constructed. ,as follows:

[0060] Constructing a consistency loss function ,as follows:

[0061] ;

[0062] in The model parameters represent the neural symbolic joint reasoning model. Indicates that given input In this context, the neural symbolic joint reasoning model generates the current target intent recognition result. and reflection vector The logarithm of the joint probability, , , This represents the gradient operator.

[0063] The log-likelihood of the strategy represents the confidence level of the model in adopting the "prediction + reflection" strategy. Represents the gradient operator (with respect to parameters) (Calculate the partial derivative), indicating the direction of parameter updates.

[0064] To avoid over-labeling of reflexive vectors, a reflexive size loss is introduced. The reflexive size loss function... ,as follows:

[0065] Reflection on the size loss function ,as follows:

[0066] ;

[0067] in , It is a hyperparameter between 0 and 1. Representing the reflection layer The output value.

[0068] Representing the reflection layer The output value is used to characterize the error probability. A value close to 1 indicates that the model considers there to be an error and that correction is needed. A value close to 0 indicates that the model considers the point correct and retains the intuitive output. This represents the probability of retaining the correct answer, which is also the probability of being correct.

[0069] when When set to a higher value, the reflexive vector tends to retain more intuitive output elements rather than labeling them as errors and delegating them to abductive reasoning. In addition, data-driven supervised training methods, such as the cross-entropy loss function, denoted as... .

[0070] Therefore, by combining all loss functions, the total loss of the model is constructed, which is the comprehensive loss function. ,as follows:

[0071] ;

[0072] The training dataset consists of two subsets: a training dataset with target intent labels and a training dataset with target intent labels. and training dataset without target intent labels Training dataset , , The number of samples in each is respectively , Training dataset Number of samples Training dataset with target intent labels The samples in the data consist of target time-series feature data. and their corresponding target intent tags Composition, training dataset without target intent labels The samples in the sample only contain target time series feature data Using a training dataset with target intent labels The preliminary target intent recognition results of the medium-sized samples are used to construct a cross-entropy loss function with the true target intent labels. , and These are hyperparameters, all of which are non-negative real numbers.

[0073] Whether still The reflection on related losses does not rely on data labeling, but directly uses the domain rule set for training, thus decoupling the optimization goals of intuitive output and reflection mechanism.

[0074] The core of the above embodiments is to construct an end-to-end joint reasoning framework. During the training phase, the joint reasoning framework uses domain knowledge to generate a reflective vector. The reflective vector can automatically identify potential errors in the output of the neural network during the reasoning process and correct them through the abductive mechanism, thereby ensuring that the final output intent recognition result is consistent with the domain knowledge.

[0075] The above embodiments propose an intent recognition framework based on neural symbolic joint reasoning. By embedding logical rules representing the logical constraints between target intents into the training objectives and inference structure of the neural network in a differentiable form, a unified approach combining data-driven and knowledge-driven methods is achieved. This method not only utilizes neural networks to extract high-level abstract features from raw data but also uses logical rules for reflective causal analysis and applies domain knowledge constraints to the network output using a loss function. This effectively solves the dual dilemmas of insufficient flexibility in pure rule-based systems and poor interpretability and low reliability in pure neural network systems, improving the accuracy of intent recognition while ensuring the transparency and compliance of the decision-making process.

[0076] In one embodiment, a multi-source sensing system (including radar, electronic support measures, satellite reconnaissance, etc.) samples seven-dimensional feature data of an aerial target over a continuous 20-second period at 1-second intervals. The numerical features in the seven-dimensional feature data include: relative distance, altitude, speed, acceleration, and heading angle; the non-numerical features include: radar on / off status and electronic jamming status. A target temporal feature dataset is then constructed. ,in n The sample size is specified. Numerical features are normalized using Min-Max based on the physical limits of the equipment to ensure that the features are not distorted. Non-numerical features are binary encoded (1: on, 0: off).

[0077] Based on the rules of engagement and expert knowledge, a standardized label space for the intentions of seven types of aerial targets is established: {0: surprise attack, 1: infiltration, 2: feint attack, 3: electronic jamming, 4: retreat, 5: reconnaissance, 6: cover}.

[0078] The neural symbolic joint reasoning model employs an 8-layer Transformer encoder with 256 hidden layers and 4 attention heads. The output layer generates a 7-dimensional intent probability distribution through linear transformation and softmax. The initial output is defined as follows:

[0079] ;

[0080] Based on the adversarial rules, a domain rule set is established, which describes the logical constraint relationship between target intentions and is expressed by first-order predicate logic rules. The first-order logic rules are converted into ASP (Answer Set Programming) format and stored in the knowledge base. For example: Rule 1: If the target activates fire control radar lock and its heading is towards us, it is definitely not a retreat.

[0081] ;

[0082] In the above formula, is the target variable, representing a specific target entity in the current environment; On is the state constant, representing the "on / locked" state; Towards is the orientation constant, representing the relative orientation "pointing towards us"; Retreat is the intention constant, representing the target's intention to "retreat". An intention predicate, indicating the goal. The intention was to "retreat". As a negation operator, it represents logical "NOT" or "absolutely impossible". Here, it's used to modify intent, indicating the goal. Their intention could never have been to retreat. The sensor state predicate represents the target. The fire control radar is in a locked state. This is a heading predicate. It indicates that the heading of target x is towards us. As an implication operator, it is read as "if...then..." or "defined by..." in logic programming (such as ASP), and the direction is from right to left: that is, "if the condition on the right is true, then the conclusion on the left is true". It is a conjunction operator, representing the logical "AND" relationship.

[0083] Using the cross-entropy loss function as A reflection layer is connected after the Transformer encoder. The output dimension is the same as the number of target intent labels, and the reflection vector is... The model is generated using a Sigmoid activation function, binarized with a threshold of 0.5, and the locations of potential erroneous outputs are marked. Symbolic inference is used to correct the causes of these errors at the marked locations. The Clingo inference engine generates the erroneous outputs. End-to-end joint optimization training is performed based on the aforementioned comprehensive loss function to obtain a trained model. The trained model can then be used to identify the target intent from the temporal feature data of the aerial target to be predicted.

[0084] On the other hand, an aerial target intent recognition device is provided, comprising:

[0085] The training dataset module is used to acquire multi-dimensional feature data of continuously collected aerial targets to construct a target temporal feature dataset, and to preprocess and label each sample in the dataset with target intent labels to construct the training dataset.

[0086] The domain rule set module is used to construct a domain rule set that describes the logical constraint relationships between target intentions, expressed by first-order predicate logic rules.

[0087] The model training module is used to construct a neural symbolic joint reasoning model. The neural symbolic joint reasoning model is jointly trained using a training dataset and a domain rule set to obtain a trained neural symbolic joint reasoning model.

[0088] The prediction module is used to predict the target intent based on the temporal feature data of the aerial target to be predicted, using a trained neural symbolic joint reasoning model.

[0089] Figure 3 The figure shows a block diagram of an electronic device according to an embodiment, such as Figure 3 As shown, the electronic device includes one or more processors and a memory. The memory stores a computer program, and when the processor executes the computer program, it implements the steps of the aerial target intent recognition method provided in any of the above embodiments. The processor may be a central processing unit (CPU) or other processing unit with data processing capabilities and / or instruction execution capabilities, and can control other components in the electronic device to perform desired functions. The memory may include one or more computer program products, which may include various forms of computer-readable storage media, such as volatile memory and / or non-volatile memory. Volatile memory may include, for example, random access memory (RAM) and / or cache memory. Non-volatile memory may include, for example, read-only memory (ROM), hard disk, flash memory, etc.

[0090] In one example, the electronic device may also include input devices and output devices, which are interconnected via a bus system and / or other forms of connection mechanism (not shown).

[0091] Of course, for the sake of simplicity, Figure 3 Only some of the components of the electronic device relevant to this application are shown, omitting components such as buses, input / output interfaces, etc. In addition, the electronic device may include any other suitable components depending on the specific application.

[0092] Embodiments of the present invention may also be computer-readable storage media storing a computer program thereon, which, when executed by a processor, implements the steps of the aerial target intent recognition method provided in any of the above embodiments. The computer-readable storage medium may be any combination of one or more readable media. A readable medium may be a readable signal medium or a readable storage medium. A readable storage medium may include, for example, but not limited to, electrical, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatuses, or devices, or any combination thereof. More specific examples of readable storage media (a non-exhaustive list) include: electrical connections having one or more wires, portable disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fibers, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof.

[0093] The basic principles of this application have been described above with reference to specific embodiments. However, it should be noted that the advantages, benefits, and effects mentioned in this application are merely examples and not limitations, and should not be considered as essential features of each embodiment of this application. Furthermore, the specific details disclosed above are for illustrative and facilitative purposes only, and are not limitations. These details do not restrict the application from being implemented using the specific details described above.

[0094] The block diagrams of devices, apparatuses, devices, and systems involved in this application are merely illustrative examples and are not intended to require or imply that they must be connected, arranged, or configured in the manner shown in the block diagrams. As those skilled in the art will recognize, these devices, apparatuses, devices, and systems can be connected, arranged, and configured in any manner. Words such as “comprising,” “including,” “having,” etc., are open-ended terms meaning “including but not limited to,” and are used interchangeably with them. The terms “or” and “and” as used herein refer to the terms “and / or,” and are used interchangeably with them unless the context clearly indicates otherwise. The term “such as” as used herein refers to the phrase “such as but not limited to,” and is used interchangeably with it.

[0095] It should also be noted that in the apparatus, equipment, and methods of this application, the components or steps can be disassembled and / or recombined. These disassemblies and / or recombinations should be considered as equivalent solutions of this application.

[0096] The above description of the disclosed aspects is provided to enable any person skilled in the art to make or use this application. Various modifications to these aspects will be readily apparent to those skilled in the art, and the general principles defined herein can be applied to other aspects without departing from the scope of this application. Therefore, this application should not be limited to the aspects shown herein, but rather to the widest scope consistent with the principles and novel features disclosed herein.

[0097] The above description has been given for purposes of illustration and description. Furthermore, this description is not intended to limit the embodiments of this application to the forms disclosed herein. Although numerous exemplary aspects and embodiments have been discussed above, those skilled in the art will recognize certain variations, modifications, alterations, additions, and sub-combinations thereof.

Claims

1. A method for recognizing the intent of aerial targets, characterized in that, Includes the following steps: Multidimensional feature data of continuously collected aerial targets are acquired to construct a target temporal feature dataset. Each sample in the dataset is preprocessed and labeled with target intent to construct a training dataset. Construct a domain rule set that describes the logical constraint relationships between target intentions, expressed by first-order predicate logic rules; Constructing a neural symbolic joint reasoning model, and jointly training the neural symbolic joint reasoning model using a training dataset and a domain rule set, including: Samples from the training dataset Transformer encoder input to neural network To obtain a high-dimensional embedding representation This includes feature information of the target time series feature data; High-dimensional embedding representation Input to a module consisting of a linear layer and a softmax layer Preliminary target intent recognition results were obtained. ; High-dimensional embedding representation Simultaneously input into the reflection layer Reflection layer It is a Sigmoid layer that performs binarization by setting a threshold, generating a result similar to... Reflection vectors with the same dimensions ; Based on reflection vector The initial target intent recognition results are corrected, and the corrected target intent recognition results are output. : in This indicates the preliminary target intent identification results. There is a potential logical conflict with the domain rule set. The abductive reasoning process will be marked as having potential logical conflicts. The blank space is filled abductively by invoking symbolic reasoning based on Markov logic networks, generating a corrected target intent recognition result that is consistent with the logical constraints between target intents in the domain rule set. ; The constructed comprehensive loss function includes: a cross-entropy loss function based on the preliminary target intent recognition results and the true target intent labels; a consistency loss function based on the reflection vector, used to guide the target intent recognition results output by the model to conform to the logical constraint relationship between target intents in the domain rule set; a reflection size loss function based on the number of conflict-marked values ​​in the reflection vector, used to control the sparsity of reflection; and a comprehensive loss function obtained by combining the cross-entropy loss function, the consistency loss function, and the reflection size loss function. With the goal of minimizing the comprehensive loss function, the model parameters of the neural symbol joint inference model are iteratively updated until convergence, thus obtaining the trained neural symbol joint inference model. The target intent is predicted based on the temporal feature data of the aerial target to be predicted using a pre-trained neural symbolic joint reasoning model.

2. The method for identifying the intent of an aerial target according to claim 1, characterized in that, Construct the training dataset, including: Continuously collect multidimensional feature data of aerial targets to construct a target time-series feature dataset. The multidimensional feature data includes: relative distance and position of aerial targets, altitude, speed, acceleration, heading angle, radar on / off status, and electronic jamming device status. Binarize the non-numerical feature data in the target time-series feature dataset and perform time-series normalization on the numerical feature data. Based on the preset target intent space, the corresponding intent labels are labeled for each target time-series feature data in the target time-series feature dataset, thus completing the construction of a training dataset with target intent labels.

3. The method for identifying the intent of an aerial target according to claim 1 or 2, characterized in that, A consistency loss function is constructed based on reflexive vectors, including: Define consistency metrics ,in , express Consistency with the domain knowledge base KB all elements If all of them are consistent with the logical constraints between target intents in the domain knowledge base KB, then It is 1 if it is not 0 otherwise; It contains elements that potentially conflict with the domain knowledge base KB. So in Remove elements that have potential logical conflicts with the domain knowledge base (KB). The output after is , express Consistent with the Domain Knowledge Base (KB), the Domain Knowledge Base (KB) is used to store the domain rule set. The first-order predicate logic rules in the text describe the logical constraint relationships between target intentions, expressed by first-order predicate logic rules.

4. The method for identifying the intent of an aerial target according to claim 3, characterized in that, Constructing a consistency loss function ,as follows: in The model parameters represent the neural symbolic joint reasoning model. Indicates that given input In this context, the neural symbolic joint reasoning model generates the current target intent recognition result. and reflection vector The logarithm of the joint probability, , , This represents the gradient operator.

5. The aerial target intent recognition method according to claim 4, characterized in that, Reflection on the size loss function ,as follows: in , It is a hyperparameter between 0 and 1. Representing the reflection layer The output value.

6. The method for identifying the intent of an aerial target according to claim 5, characterized in that, Comprehensive loss function as follows: The training dataset consists of two subsets: a training dataset with target intent labels and a training dataset with target intent labels. and training dataset without target intent labels Training dataset , , The number of samples in each is respectively , Training dataset Number of samples Training dataset with target intent labels The samples in the data consist of target time-series feature data. and their corresponding target intent tags Composition, training dataset without target intent labels The samples in the sample only contain target time series feature data Using a training dataset with target intent labels The preliminary target intent recognition results of the medium-sized samples are used to construct a cross-entropy loss function with the true target intent labels. , and These are hyperparameters, all of which are non-negative real numbers.

7. An aerial target intent recognition device, used to implement the aerial target intent recognition method as described in claim 1, 2, 4, 5, or 6, characterized in that, include: The training dataset module is used to acquire multi-dimensional feature data of continuously collected aerial targets to construct a target temporal feature dataset, and to preprocess and label each sample in the dataset with target intent labels to construct the training dataset. The domain rule set module is used to construct a domain rule set that describes the logical constraint relationships between target intentions, expressed by first-order predicate logic rules. The model training module is used to construct a neural symbolic joint reasoning model. The neural symbolic joint reasoning model is jointly trained using a training dataset and a domain rule set to obtain a trained neural symbolic joint reasoning model. The prediction module is used to predict the target intent based on the temporal feature data of the aerial target to be predicted, using a trained neural symbolic joint reasoning model.

8. An electronic device comprising a memory and a processor, characterized in that: The memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions, which, when executed by the processor, implement the steps of the airborne target intent recognition method as described in claim 1.

9. A computer storage medium storing computer-executable instructions thereon, characterized in that: When the computer-executable instructions are executed by the processor, they implement the steps of the airborne target intent recognition method as described in claim 1.

Images