Military target behavior analysis agent test method and equipment

Abstract

The invention belongs to the technical field of artificial intelligence analysis, and particularly relates to a military target behavior analysis agent test method and equipment. The method comprises the following steps: step 1, acquiring multi-source data of a military target, and performing data preprocessing on each category of data in the multi-source data; 2, extracting tensor features of the input tensor; defining a multi-layer intention hierarchical structure based on the tensor features; 3, defining a likelihood function of each intention layer; modeling a prior probability using a recurrent neural network; calculating a posterior probability by using a Bayesian rule; when the number of times of iterative execution reaches the set number of times, executing the step 4: performing intention reasoning by using approximate posterior distribution, and calculating the uncertainty of intention reasoning. According to the method, the precision, the real-time performance, the interpretability and the reliability of military target behavior analysis are remarkably improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of artificial intelligence analysis, and specifically relates to a method and equipment for testing a militarized target behavior analysis intelligent agent. Background Art

[0002] In modern military operations, the application of intelligent and automated technologies is becoming increasingly widespread, especially in the field of militarized target behavior analysis. By accurately predicting and analyzing enemy target behavior, military commanders can better make strategic decisions, optimize resource allocation, and improve combat effectiveness. However, the existing militarized target behavior analysis technology still has many problems and challenges in practical applications. At present, militarized target behavior analysis technology mainly relies on data-driven methods, including machine learning and deep learning technologies. These technologies build predictive models by training a large amount of historical data to identify and infer the behavioral intentions of enemy targets. For example, using supervised learning methods, researchers can train classifiers to identify specific target behavior patterns; using unsupervised learning methods, hidden behavior patterns and associations can be discovered; using deep learning methods, especially convolutional neural networks (CNNs) and recurrent neural networks (RNNs), complex spatiotemporal data can be processed and analyzed to make more refined target behavior predictions.

[0003] Militarized target behavior analysis relies on a large amount of high-quality training data. However, in actual operations, there are many difficulties in obtaining and processing this data. Military intelligence data often contains noise, incompleteness, and inconsistency, which affects the training effect and prediction accuracy of the model. In addition, the data formats and types of different data sources vary greatly, and how to effectively integrate multi-source data is also a major challenge. Existing machine learning and deep learning models are prone to overfitting during the training process, that is, the model performs well on the training data, but has poor generalization ability in practical applications. This is because the model remembers specific data patterns during the training process and fails to effectively learn generally applicable features. This problem is particularly prominent in militarized target behavior analysis, because the actual combat environment is complex and changeable, and training data often cannot cover all possible situations. Militarized target behavior analysis requires real-time processing and analysis of a large amount of data to provide accurate prediction results. However, existing deep learning models usually have high computational complexity and slow processing speed, and it is difficult to meet real-time requirements. This may cause delays and information lags in actual combat, affecting the timeliness and accuracy of decision-making. Although deep learning models, especially neural network models, perform well in complex pattern recognition and prediction, their internal structure is complex and lacks interpretability. This makes it difficult for decision makers to understand the model's prediction results and inference process, thereby reducing their trust and reliance on the model. In military decision-making, the interpretability of the model is crucial because decision makers need to make judgments based on clear and reliable information. Existing technologies often ignore the assessment of the uncertainty of prediction results when predicting target behavior. This results in decision makers being unable to effectively assess the reliability and risk of prediction results when using them, which may lead to wrong decisions. In actual combat, uncertainty assessment can provide important information about the confidence of prediction results, helping decision makers make more cautious and reasonable decisions in uncertain environments. Summary of the invention

[0004] In view of this, the main purpose of the present invention is to provide a militarized target behavior analysis intelligent agent testing method and equipment, which significantly improves the accuracy, real-time performance, interpretability and reliability of militarized target behavior analysis.

[0005] The technical solution adopted by the present invention is as follows: A militarized target behavior analysis agent testing method, the method comprising: Step 1: Obtain multi-source data of militarized targets, pre-process the data of each category in the multi-source data, align the timestamps and register the space, and then fuse them into an input tensor; Step 2: Extract tensor features of the input tensor; based on the tensor features, define a multi-layer intent hierarchy, where each intent layer contains multiple possible intents, and use conditional probabilities to represent the dependencies between intent layers; Step 3: Define the likelihood function of each intent layer; use a recursive neural network to model the prior probability; use Bayes' rule to calculate the posterior probability; iteratively perform the following process to complete the training: use variational inference to define an approximate posterior distribution, and then minimize the KL divergence; use importance sampling to approximate the expectation, and use stochastic gradient descent to optimize the parameters of variational inference; when the number of iterations reaches the set number, execute step 4: Step 4: Use the approximate posterior distribution to perform intent inference and calculate the uncertainty of intent inference.

[0006] Furthermore, the multi-layer intention hierarchy in step 2 includes at least three intention layers, namely: a strategic intention layer, a tactical intention layer and a specific action intention layer; the strategic intention layer includes at least the following intentions: attack, defense, retreat, reorganization and deception; the tactical intention layer includes at least the following intentions: encirclement, breakthrough, reconnaissance, blocking, detour and feint; the specific action intention layer includes at least the following intentions: deploying weapon systems, establishing fortifications, carrying out air strikes, conducting electronic interference, launching ground attacks, conducting special operations, and establishing supply lines.

[0007] Furthermore, the multi-source data includes at least the following categories of data: position data, speed data, acceleration data, direction data, environmental data, altitude data, longitude data and latitude data; the process of performing data preprocessing on each category of the multi-source data in step 1 includes: performing data noise reduction and data normalization processing on each category of the data; in step 1, after performing timestamp alignment and spatial registration on the preprocessed data of each category, performing data resampling and interpolation processing and fusing them into an input tensor; in step 2, using a deep neural network to extract tensor features of the input tensor .

[0008] Furthermore, in step 2, define Layer Intent Structure, Layer intention layer possible intents; initialize the intent vector for each layer: ; in, , is the index number; Indicates The intention layer intention; each is a discrete variable and can be possible values; Tensor features The nonlinear transformation is performed through the following formula to obtain the feature , to accommodate different levels of intent reasoning: ; in, represents element-wise multiplication, and They are the preset feature weight matrix and feature bias matrix respectively; Use conditional probability to represent the dependencies between intent layers: ; in, Indicates the highest level of intention. Indicates the lowest level of intention; For a given feature The joint probability distribution of the entire intent hierarchy when ; For a given When the intention and features of the layers are combined, the conditional probability of the lowest level intention; For a given When layering intent and features, Conditional probability of layer intention; The conditional probability of the top-level intent given the features.

[0009] Further, in step 3, the following formula is used to define Likelihood function of the layer intention layer , which contains a multi-layer perceptron of a main branch and a multi-layer sensor of a parallel branch. The main branch uses the tanh activation function, and the parallel branch uses the LeakyReLU activation function, which is expressed by the following formula: ; in, From the 1st layer to the The intentional sequence of layers; is the embedding representation of the low-level intent, which is a function that maps a discrete sequence of intents to a continuous vector space. Let the dimension after embedding be ;set up The latitude is , is the concatenation of feature and low-level intent embedding, with dimension ; For the The weight matrix of the layer intention layer has the dimension ,in It is Dimensions of the layer intention layer; For the The bias vector of the layer intent layer, dimension is ; For the The weight matrix of the layer intention layer has the dimension ,in is the dimension of the output; For the The bias vector of the layer intent layer, dimension is ; For the parallel branch The weight matrix of the layer intention layer has the dimension ; For the parallel branch The bias vector of the layer intent layer, dimension is ; For the parallel branch The weight matrix of the layer intention layer has the dimension ; For the parallel branch The bias vector of the layer intent layer, dimension is ; is the rectified linear unit activation function with leakage, is a function variable; is the Softmax function, which is used to convert the output into a probability distribution; is the Sigmoid activation function, which is used to compress the output to the (0,1) interval.

[0010] Furthermore, a recurrent neural network is used to model the first Prior probability of the intention layer , each intent layer is a hidden layer of the recurrent neural network: ; ; in, represents a gated recurrent unit; The recurrent neural network corresponds to Hidden layers of the layer intention layer; The recurrent neural network corresponds to Hidden layers of the layer intention layer; is the weight matrix of the recurrent neural network; is the bias matrix of the recurrent neural network.

[0011] Further, the Bayesian rule is used to calculate the Posterior probability of the layer intention : .

[0012] in, is the subscript index; using variational inference, an approximate posterior distribution is defined using the following formula: ; in, is the parameterized approximate posterior distribution; are the parameters of variational inference; is the weight matrix for variational inference; is the bias vector for variational inference.

[0013] Furthermore, the KL divergence is minimized by the following formula: ; in, is the loss function; represents the KL divergence; importance sampling is used to approximate the expectation through the following formula: +const; in, is the sampling times; is the kth intention layer obtained by the sth sampling: const is a constant term; the parameters of variational inference are optimized using stochastic gradient descent through the following formula: ; in, For the The parameters of variational inference after iterative optimization; Parameters of variational inference after optimization at the t+1th iteration; is the learning rate; is the gradient of the loss function with respect to the parameter θ.

[0014] Furthermore, in step 4, the approximate posterior distribution is used to perform intent inference through the following formula: ; in, is the result of intention reasoning; then use the following formula to calculate the uncertainty of intention reasoning: ; in, Reasoning about uncertainty in intention.

[0015] A militarized target behavior analysis intelligent agent testing device includes: a data acquisition part, which is used to acquire multi-source data of militarized targets, perform data preprocessing on each category of data in the multi-source data, perform timestamp alignment and spatial registration, and then fuse them into an input tensor; an intention model construction part, which is used to extract tensor features of the input tensor; based on the tensor features, a multi-layer intention hierarchy is defined, each intention layer contains multiple possible intentions, and conditional probability is used to represent the dependency relationship between the intention layers; an intention reasoning model construction part, which is used to define the likelihood function of each intention layer; a recursive neural network is used to model the prior probability; a posterior probability is calculated using the Bayesian rule; the following process is iteratively executed to complete the training: using variational inference, defining an approximate posterior distribution, and then minimizing the KL divergence; using importance sampling to approximate expectations, and using stochastic gradient descent to optimize the parameters of variational inference; when the number of iterations reaches the set number, executing step 4: the intention reasoning part, which is used to perform intention reasoning using the approximate posterior distribution, and calculate the uncertainty of intention reasoning By adopting the above technical scheme, the present invention produces the following beneficial effects: The present invention effectively improves the generalization ability of the model through the combination of a multi-layer intention hierarchy and a recursive neural network (RNN). Each intention layer models the prior probability through GRU (gated recurrent unit), so that the model can handle long-term dependencies and capture the complex patterns of target behaviors. Through variational inference, an approximate posterior distribution is defined to approximate the true posterior distribution. By minimizing the KL divergence, the parameters of variational inference are optimized so that the approximate posterior distribution is as close to the true posterior distribution as possible. This method can effectively avoid the problem of model overfitting, improve the generalization ability in different environments, and make the model more robust in practical applications. The present invention has also done a lot of work in enhancing the interpretability of the model. Through a multi-layer intention hierarchy, the target behavior is decomposed into strategic intention, tactical intention and specific action intention, and the intention and calculation process of each layer have clear physical meaning and explanation. This design makes the reasoning process of the model more transparent, which is convenient for decision makers to understand and trust. For each layer of intent, entropy is calculated to evaluate uncertainty, allowing decision makers to more clearly understand the confidence and reliability of model predictions and make more informed decisions in actual operations. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 A flowchart of a militarized target behavior analysis agent testing method provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0017] All features disclosed in this specification, or steps in all methods or processes disclosed, except mutually exclusive features and / or steps, can be combined in any manner.

[0018] Any feature disclosed in this specification (including any additional claims and abstract), unless otherwise stated, may be replaced by other alternative features that are equivalent or have similar purposes. That is, unless otherwise stated, each feature is only an example of a series of equivalent or similar features.

[0019] Example 1: Reference Figure 1 , a militarized target behavior analysis agent testing method, the method comprising: Step 1: Obtain multi-source data of militarized targets, pre-process the data of each category in the multi-source data, align the timestamps and register the space, and then fuse them into an input tensor; Multi-source data usually includes data from different sensors, such as image data, radar data, communication records, etc., which provide rich information about the target in different time and space dimensions. However, due to the diversity and complexity of data sources, directly using raw data for analysis often leads to accuracy and reliability problems. Therefore, data preprocessing is a key step. First, each data is standardized to ensure that different data types have the same dimension and scale. This process usually involves operations such as noise removal, missing value filling, and data smoothing, aiming to improve the quality and consistency of the data. After standardization, timestamp alignment is also required. Multi-source data is often collected asynchronously, so the data needs to be time-aligned to ensure that it is analyzed at the same time point. This step is achieved through methods such as interpolation or resampling, mapping data at different time points to a unified time axis. Next is spatial registration, which refers to spatially aligning data collected by different sensors. For example, image data and radar data may come from different spatial coordinate systems. Through registration technology, they can be converted to the same coordinate system, which is convenient for fusion and analysis. Spatial registration usually uses geometric transformation and image registration algorithms, such as feature point-based registration methods or density-based registration methods. The purpose of this process is to ensure the spatial consistency of different data sources and eliminate errors caused by different sensor positions and viewing angles. After completing timestamp alignment and spatial registration, all preprocessed data are fused into a multidimensional input tensor. The input tensor is a high-dimensional data structure that contains information in multiple dimensions such as time, space, and data type. By fusing multi-source data into an input tensor, the spatiotemporal correlation between the data sources can be fully utilized to provide a complete and detailed description of the target behavior. For example, image data can provide visual information of the target, radar data can provide distance and speed information of the target, and communication records can provide interaction and behavior information of the target. This information is fused together in the form of an input tensor, making subsequent feature extraction and behavior analysis more accurate and reliable. The construction of the input tensor is not just a simple superposition of data, but is based on a deep understanding of data features and correlations. In the form of a high-dimensional tensor, the dynamic changes of target behavior in time and space, as well as the complex relationship between different data sources, can be captured. The input tensor provides a solid foundation for tensor feature extraction and the definition of multi-layer intent hierarchy in subsequent steps. Through this data preprocessing and fusion method, the present invention can better process and analyze complex militarized target behavior data and improve the accuracy and reliability of intelligent agent testing. In general, step 1 plays a vital role in the present invention. Through comprehensive preprocessing and fusion technology, a complete and accurate input data foundation is constructed, making subsequent behavior analysis and intent reasoning more efficient and accurate.

[0020] Step 2: Extract tensor features of the input tensor; based on the tensor features, define a multi-layer intent hierarchy, where each intent layer contains multiple possible intents, and use conditional probabilities to represent the dependencies between intent layers; First, the input tensor is feature extracted by deep learning methods. In this step, the input tensor is a high-dimensional data structure that contains rich information of time, space and multi-source data. Through deep neural networks, especially convolutional neural networks (CNNs), important features in the tensor can be effectively extracted. Through its hierarchical structure and convolution operations, convolutional neural networks can automatically learn and capture spatial and temporal patterns in data, generate a low-dimensional but rich feature vector, and these feature vectors become the basis for subsequent intent analysis. Next, based on the extracted tensor features, a multi-layer intent hierarchy is defined. The multi-layer intent hierarchy refers to modeling the target's behavioral intention in a hierarchical manner, with each layer representing a different level of intent, from low-level specific behavior to high-level strategic intent. This hierarchy can better capture the complexity and diversity of target behavior. For example, in a militarized scenario, low-level intent may involve basic actions of the target, such as moving or attacking, while high-level intent may involve more complex tactical or strategic goals. Through this hierarchical intent modeling, the target's behavior can be more comprehensively understood and predicted. In order to model the dependencies between the intents of each layer in the multi-layer intent hierarchy, the conditional probability method is used to represent this relationship. Conditional probability can describe the probability distribution of an intent layer under the condition of the upper layer intent, so that the complex intent hierarchy relationship can be clearly expressed by mathematical models. The probability distribution of each layer of intent depends not only on the tensor features of the current layer, but also on the intent state of the previous layer, thus forming a recursive dependency relationship. This method can effectively capture the dynamic changes and hierarchical dependencies of the target behavior, making the model's inference of behavioral intention more accurate. In the multi-layer intent hierarchy, the intent of each layer can be regarded as a hidden variable. By defining the likelihood function of each layer of intent, the distribution of observed data under different intents can be described. The definition of the likelihood function is based on the relationship between tensor features and intent. By maximizing the likelihood function, the most likely intent explanation can be found. On this basis, a recurrent neural network (RNN) is used to model the prior probability of each layer of intent. RNN has the ability to process time series data and can capture the dynamic changes of intent over time. The prior modeling of intent by RNN can better reflect the temporal pattern of the target behavior. In order to infer the intent of the target, the Bayesian inference method is used to calculate the posterior probability. Bayesian inference can combine observed data and prior knowledge to give a comprehensive probability distribution of intent. In actual calculations, since the posterior distribution is often complex and difficult to solve, the posterior distribution can be approximated by variational inference. Variational inference finds an effective approximate solution by defining an approximate distribution and minimizing its KL divergence with the true posterior distribution. In the optimization process, importance sampling is used to approximate expectations, and the stochastic gradient descent (SGD) algorithm is combined to optimize the parameters of variational inference, so that the model can gradually approach the optimal solution during training.

[0021] Step 3: Define the likelihood function of each intent layer; use a recursive neural network to model the prior probability; use Bayes' rule to calculate the posterior probability; iteratively perform the following process to complete the training: use variational inference to define an approximate posterior distribution, and then minimize the KL divergence; use importance sampling to approximate the expectation, and use stochastic gradient descent to optimize the parameters of variational inference; when the number of iterations reaches the set number, execute step 4: The core of step 3 is to model and infer the multi-layer intent hierarchy through recurrent neural network (RNN) and Bayesian inference. This process involves defining the likelihood function of each intent layer, using RNN to model the prior probability, and calculating the posterior probability through Bayesian rule. Finally, optimization and approximate calculation are performed through variational inference and importance sampling to complete the training and inference of the target behavior intention. First, in the multi-layer intent hierarchy, the intent of each layer can be regarded as a hidden variable. By defining the likelihood function of each intent layer, the distribution of observed data under different intents can be described. The likelihood function refers to the probability distribution of observed data under a given intent. By maximizing the likelihood function, the most likely intent explanation can be found. This process is based on the relationship between tensor features and intent. Through the definition of the likelihood function, the observed data can be linked to the intent, thus providing a basis for the subsequent inference process. Next, a recurrent neural network (RNN) is used to model the prior probability of each layer of intent. RNN has the ability to process time series data and can capture the dynamic changes of intent over time. Specifically, RNN can memorize and process long-range dependencies in sequence data through its internal loop structure, which is very important for modeling the temporal dynamic changes of target behavior. In the modeling process of RNN, the intention state at each moment depends not only on the current observation data, but also on the intention state at the previous moment, thus forming a recursive dependency. By modeling the intention a priori through RNN, the temporal pattern of the target behavior can be better reflected, so that the model has higher accuracy and robustness when processing complex temporal data. In order to infer the intention of the target, the Bayesian inference method is used to calculate the posterior probability. Bayesian inference can combine the observation data and prior knowledge to give a comprehensive intention probability distribution. Specifically, Bayesian inference combines the prior probability and the likelihood function through the Bayesian formula to calculate the posterior probability. The posterior probability represents the probability distribution that the target has a certain intention given the observation data. Since the posterior distribution is often complex and difficult to solve, direct calculation is very difficult. Therefore, in practical applications, the approximate inference method is often used to simplify the calculation process. In the present invention, the variational inference method is used to approximate the posterior distribution. Variational inference is an optimization technique that finds an effective approximate solution by defining an approximate distribution and minimizing the KL divergence between it and the true posterior distribution. Specifically, variational inference transforms complex integral calculations into optimization problems to find a parameterized approximate distribution that is as close to the true posterior distribution as possible. During the optimization process, importance sampling is used to approximate expectations. Importance sampling is a Monte Carlo method that estimates the expected value by reweighting the samples. This method can achieve higher estimation accuracy with a smaller number of samples and improves computational efficiency. In addition, the stochastic gradient descent (SGD) algorithm is used to optimize the parameters of variational inference.SGD is an efficient optimization algorithm that uses the gradient information of the current parameters in each iteration to gradually adjust the parameters so that the objective function gradually converges to the optimal solution. In the present invention, SGD is used to minimize the KL divergence in variational inference so that the approximate distribution can better approximate the true posterior distribution. Through this optimization process, the model can gradually approach the optimal solution during training, thereby improving the accuracy and robustness of intent inference.

[0022] When the number of iterations reaches the set number, the training process is completed and the model has a high inference ability. On this basis, the approximate posterior distribution is used for intent inference. The approximate posterior distribution represents the probability distribution of the target intent given the observed data. Through this probability distribution, the most likely intent state of the target can be inferred and the uncertainty of intent inference can be calculated. Uncertainty represents the confidence of the model's inference of the target intent given the observed data. Specifically, uncertainty can be measured by the entropy of the posterior distribution. The higher the entropy value, the greater the uncertainty of the model's intent inference; conversely, the lower the entropy value, the higher the confidence of the model in the intent inference.

[0023] Step 4: Use the approximate posterior distribution to perform intent inference and calculate the uncertainty of intent inference.

[0024] The approximate posterior distribution represents the probability distribution of the target having a certain intention given the observed data. By maximizing this probability distribution, the target's most likely behavioral intention can be inferred. This reasoning process is based on the result of Bayesian inference. It combines the prior probability and the likelihood of the observed data, so that the reasoning result not only considers the influence of historical data, but also combines the currently observed data, so that it has high reliability and accuracy. In actual operation, when using the approximate posterior distribution for intention reasoning, all possible intention states are traversed, the posterior probability of each state is calculated, and the state with the largest probability is selected as the predicted intention of the target. Although this process seems simple, it actually requires efficient calculation methods and powerful computing power because it involves high-dimensional data and complex probability calculations. By using variational inference and recurrent neural networks, the calculation process can be effectively simplified and the calculation efficiency can be improved, thereby realizing real-time intention reasoning. Calculating the uncertainty of intention reasoning is another key step. Uncertainty assessment is very important in practical applications. It can tell the model how confident it is in the current inference result. When the uncertainty of the model is high, it means that the current observation data is not enough to support a clear inference. At this time, additional observation data or further analysis may be required to improve the reliability of the inference. On the contrary, when the uncertainty of the model is low, it means that the model has a high confidence in the current inference result, and corresponding decisions can be made according to the inference result. The calculation of uncertainty is usually based on the entropy of the posterior distribution. Entropy is an indicator of the uncertainty of probability distribution. The higher the entropy value, the more dispersed the distribution and the greater the uncertainty; the lower the entropy value, the more concentrated the distribution and the smaller the uncertainty. In this method, by calculating the entropy of the posterior distribution, the confidence of the model in each intention inference result can be quantified. This process can not only help understand the behavior of the model, but also provide important information support for actual decision-making. Combined with the above steps, the method of the present invention can achieve efficient and accurate intention reasoning and uncertainty assessment when dealing with militarized target behavior analysis. A powerful analysis model is constructed by preprocessing and fusion of multi-source data, definition of multi-layer intention hierarchy, combination of recursive neural network and Bayesian inference, and optimization of variational inference and importance sampling. This model can not only process complex time series data and multi-dimensional information, but also provide reasonable evaluation in the face of uncertainty, thereby improving the reliability and practicality of the analysis results.

[0025] Embodiment 2: The multi-layer intention hierarchy in step 2 includes at least three intention layers, namely: a strategic intention layer, a tactical intention layer and a specific action intention layer; the strategic intention layer includes at least the following intentions: attack, defense, retreat, reorganization and deception; the tactical intention layer includes at least the following intentions: encirclement, breakthrough, reconnaissance, blocking, detour and feint; the specific action intention layer includes at least the following intentions: deploying weapon systems, establishing fortifications, carrying out air strikes, conducting electronic interference, launching ground attacks, conducting special operations, and establishing supply lines.

[0026] Specifically, the strategic intention layer is the highest level of intention layer, which is mainly used to describe the macro-strategic decision-making of militarized targets. The intentions at this level include attack, defense, retreat, reorganization and deception. The offensive intention indicates that the main purpose of the target is to launch an active attack to suppress the enemy's strength. The defensive intention is to protect one's own forces and resist the enemy's attack. The retreat intention indicates that the target chooses to retreat in the current situation to avoid unnecessary losses or make strategic adjustments. The reorganization intention indicates that the target is reorganizing its own forces, which may be to fight more effectively or adjust strategic deployment. The deception intention is to achieve strategic advantages by disguising or confusing the enemy and making it misjudge its true intentions. The setting of these strategic intentions provides a macro-direction and framework for the overall military operation. Secondly, the tactical intention layer is the middle level of intention layer, which is mainly used to describe the tactical actions of militarized targets in specific battles. The intentions at this level include encirclement, breakthrough, reconnaissance, blocking, detour and feint attack. The encirclement intention indicates that the target attempts to encircle the enemy so that it cannot escape or launch an effective counterattack. The intention of breakthrough is to carry out a strong assault at the weak point of the enemy's defense line to break the enemy's defense. The intention of reconnaissance means that the target collects intelligence to understand the enemy's movements and deployment. The intention of blocking is to set up defense lines at key positions to prevent the enemy's attack or advancement. The intention of detour is to bypass the enemy's main force and attack its flank or rear to achieve the effect of surprise. The intention of feint attack is to attract the enemy's attention by pretending to attack, so that it misjudges its true actions. The setting of these tactical intentions provides tactical guidance for specific actions in actual combat. Finally, the specific action intention layer is the lowest level of intention, which is mainly used to describe the specific operations of militarized targets in specific execution. The intentions at this level include deploying weapon systems, building fortifications, carrying out air strikes, conducting electronic interference, launching ground attacks, carrying out special operations, and establishing supply lines. The intention of deploying weapon systems means that the target places weapons at specific locations to enhance firepower coverage. The intention of building fortifications is to improve one's own defense capabilities by building defense facilities. The intention to carry out air strikes means that the target uses air power to strike the enemy. The intention to carry out electronic jamming means that the enemy's communications and radar systems are interfered with by electronic warfare means. The intention to launch a ground offensive means that the target uses ground forces to attack the enemy's positions. The intention to carry out special operations means that the target uses special forces to carry out high-precision and high-risk military operations. The intention to establish supply lines means that the target is ensuring logistical supplies to support front-line operations. The setting of these specific action intentions provides detailed guidance for each step of the actual operation.

[0027] For each layer, a probability distribution is output, indicating the probability of each possible intent. For example: (Layer 1); (Layer 2); (Layer 3).

[0028] The most likely intent combination: By selecting the intent with the highest probability in each layer, the most likely intent combination can be obtained: ; For example: (attack, break through, launch a ground attack).

[0029] Confidence of intent: For each layer of intent, its entropy can be calculated to represent the uncertainty of reasoning: ; Lower entropy values ​​indicate higher confidence.

[0030] Temporal Intention Reasoning: Considering the dynamic nature of military operations, intention reasoning can be extended in the temporal dimension: ; This allows tracking changes in intent over time.

[0031] Conditional intent prediction: Based on the currently inferred intent, possible future intent changes can be predicted: .

[0032] This multi-level, dynamic intention reasoning result provides rich information for military decision-making, including not only the most likely intention combination at the moment, but also the probability distribution of various possibilities, as well as the trend of intention change over time. This information can be used to formulate corresponding strategies, tactics and specific action plans.

[0033] Embodiment 3: The multi-source data includes at least the following categories of data: position data, speed data, acceleration data, direction data, environmental data, altitude data, longitude data and latitude data; Step 1 The process of preprocessing the data of each category in the multi-source data includes: performing data noise reduction and data normalization processing on the data of each category; In step 1, after the preprocessed data of each category is timestamp aligned and spatially registered, the data is resampled and interpolated to be fused into an input tensor; In step 2, the tensor features of the input tensor are extracted using a deep neural network .

[0034] Specifically, multi-source data comes from a wide range of sources, and each category of data has different physical quantities and units, so it is very necessary to preprocess these data. Data denoising is the first step in preprocessing. It removes noise from the data through filters or other denoising techniques to retain useful signals. Common denoising methods include low-pass filtering, high-pass filtering, and band-pass filtering, which can effectively remove high-frequency noise or low-frequency interference, thereby improving the signal-to-noise ratio of the data. Data normalization is the second step in preprocessing. Through normalization, data of different categories are converted to the same scale, making subsequent processing and analysis more convenient. Normalization methods include minimum-maximum normalization and z-score standardization, which can narrow the value range of the data to a standard interval, such as [0,1] or [-1,1], thereby eliminating the dimensional differences between different data categories. The preprocessed data needs to be timestamp aligned and spatially registered to ensure the consistency of different data sources at the same time point and spatial position. Timestamp alignment refers to aligning the data of each data source by timestamp to ensure that all categories of data can be obtained at the same time point. This step is usually implemented by interpolation or resampling. Interpolation methods include linear interpolation, spline interpolation, etc., and resampling methods include downsampling and upsampling, etc. These methods can map data of different time resolutions to a unified time axis. Spatial registration refers to aligning data from different data sources according to spatial positions to ensure that all categories of data can be obtained at the same spatial position. This step is usually implemented through geometric transformation and coordinate transformation, such as aligning image data and lidar data using camera calibration technology, and aligning geographic data using GPS coordinate conversion. After completing timestamp alignment and spatial registration, the data is resampled and interpolated to merge it into an input tensor. The input tensor is a high-dimensional data structure that contains multidimensional information of time, space, and multi-source data. Resampling and interpolation processing can further improve the consistency and integrity of the data. Through these processes, the data is fused into a unified high-dimensional representation, making subsequent feature extraction and behavior analysis more efficient and accurate. In step 2, a deep neural network is used to extract the tensor feature H of the input tensor. Deep neural networks, especially convolutional neural networks (CNNs), have strong capabilities in extracting high-dimensional data features. Through its hierarchical structure and convolution operations, CNN can automatically learn and capture spatial and temporal patterns in data, generating a low-dimensional but information-rich feature vector. Specifically, the input tensor is passed through multiple convolutional layers, pooling layers, and fully connected layers to gradually extract features at different levels. These features not only contain the spatiotemporal information of the original data, but also contain deep semantic information. The convolutional layer can effectively extract spatial features through the local receptive field and weight sharing mechanism; the pooling layer can reduce the dimension of the data and prevent overfitting through downsampling operations; the fully connected layer can integrate different features through linear transformation to generate the final tensor features.

[0035] Example 4: In step 2, define Layer Intent Structure, Layer intention layer possible intents; initialize the intent vector for each layer: ; in, , is the index number; Indicates The intention layer intention; each is a discrete variable and can be possible values; Tensor features The nonlinear transformation is performed through the following formula to obtain the feature , to accommodate different levels of intent reasoning: ; in, represents element-wise multiplication, and They are the preset feature weight matrix and feature bias matrix respectively; Use conditional probability to represent the dependencies between intent layers: ; in, Indicates the highest level of intention. Indicates the lowest level of intention; For a given feature The joint probability distribution of the entire intent hierarchy when ; For a given When the intention and features of the layers are combined, the conditional probability of the lowest level intention; For a given When layering intent and features, Conditional probability of layer intention; The conditional probability of the top-level intent given the features.

[0036] Specifically, in this multi-layer intent hierarchy, each intent layer is defined as a A vector of possible intentions . This definition allows each intent layer to express the specific behavioral intentions of the target at different levels. For example, the highest-level intent can represent macro-strategic decisions, such as attack, defense, retreat, etc.; the middle-level intent can represent specific tactical actions, such as encirclement, breakthrough, reconnaissance, etc.; and the lowest-level intent can be refined to specific operations, such as deploying weapon systems, building fortifications, etc. Such a hierarchical design can not only capture the diversity and complexity of target behaviors, but also provide a clearer hierarchical structure during analysis, making the relationship between each intent layer clearer. In order to adapt to the intention reasoning at different levels, the preprocessed tensor features are Perform nonlinear transformation to obtain features This process is implemented through a transformation formula with a nonlinear activation function, which can deeply process the original features to make them better suited for subsequent intent reasoning. Through this transformation, the model can capture more complex relationships and patterns between features, thereby providing richer information support in the reasoning process. Nonlinear transformation plays a vital role in this process, making the features It not only contains the basic information of the original data, but also reflects the abstract features of a higher level. In the process of establishing a multi-layer intention hierarchy, the key is to use conditional probability to represent the dependency between the layers of intention. By defining the conditional probability of each layer of intention, the joint probability distribution of the intentions of each layer can be derived layer by layer. This conditional probability representation method can effectively capture the dependency between intentions at different levels, so that the intentions of each layer not only depend on the characteristics of the current layer, but also are affected by the intentions of the previous layer. For example, tactical intentions not only depend on the current observation data, but also are guided by strategic intentions; specific action intentions are affected by the current tactical intentions and are constrained by the overall strategy. Through layer-by-layer reasoning and calculation, the probability distribution of each layer of intention under the given characteristics and the previous layer of intentions can be accurately determined, thereby realizing a comprehensive analysis and inference of the target behavior. The combination of this multi-layer intention hierarchy and conditional probability makes the method of the present invention have high accuracy and robustness when dealing with complex militarized target behaviors. In practical applications, through layer-by-layer derivation and reasoning, the probability distribution of each layer of intentions can be updated and adjusted in real time, thereby dynamically reflecting the behavioral changes of the target. Whether at the strategic level or at the tactical and operational level, it can provide accurate intention analysis and prediction, which not only improves the intelligent agent's ability to understand the target's behavior, but also provides strong technical support for military decision-making.

[0037] Embodiment 5: In step 3, the following formula is used to define Likelihood function of the layer intention layer , which contains a multi-layer perceptron of a main branch and a multi-layer sensor of a parallel branch. The main branch uses the tanh activation function, and the parallel branch uses the LeakyReLU activation function, which is expressed by the following formula: ; in, From the 1st layer to the The intentional sequence of layers; is the embedding representation of the low-level intent, which is a function that maps a discrete sequence of intents to a continuous vector space. Let the dimension after embedding be ;set up The latitude is , is the concatenation of feature and low-level intent embedding, with dimension ; For the The weight matrix of the layer intention layer has the dimension ,in It is Dimensions of the layer intention layer; For the The bias vector of the layer intent layer, dimension is ; For the The weight matrix of the layer intention layer has the dimension ,in is the dimension of the output; For the The bias vector of the layer intent layer, dimension is ; For the parallel branch The weight matrix of the layer intention layer has the dimension ; For the parallel branch The bias vector of the layer intent layer, dimension is ; For the parallel branch The weight matrix of the layer intention layer has the dimension ; For the parallel branch The bias vector of the layer intent layer, dimension is ; is the rectified linear unit activation function with leakage, is a function variable; is the Softmax function, which is used to convert the output into a probability distribution; is the Sigmoid activation function, which is used to compress the output to the (0,1) interval.

[0038] Specifically, the likelihood function The calculation is done by the main branch and the parallel branch. The core of the main branch is the tanh activation function, which can introduce nonlinear factors to make the feature transformation more complex and diverse. The output of the tanh function is This makes it work better when processing signed data. The parallel branch uses the LeakyReLU activation function, which is a leaky rectified linear unit activation function that can retain a certain gradient when the input is negative, avoiding the problem of the traditional ReLU gradient being zero in the negative region. The form of LeakyReLU is , through this activation function, the model can better handle sparse and negative features. In the specific implementation, the input feature The embedding representation of the lower-level intent is concatenated to form a new feature vector This operation integrates information from different levels, allowing the model to simultaneously consider the influence of the current feature and the previous layer's intent. After this concatenation operation, the weight matrix and the bias vector Perform linear transformation to generate an intermediate feature vector. The intermediate features in the main branch are processed by the tanh activation function to output a feature vector after nonlinear transformation; the intermediate features in the parallel branch are processed by the LeakyReLU activation function to output another feature vector after nonlinear transformation. The outputs of these two branches are normalized by the softmax function and the sigmoid function respectively. The softmax function is used to convert the output into a probability distribution, while the sigmoid function is used to compress the output to interval. The final result is the element-by-element product of the two. This design makes the model more flexible and efficient in processing and fusing multiple features. By combining different activation functions and normalization methods, the model can capture both linear and nonlinear feature relationships at the same time, thereby making more accurate inferences in complex environments. In the analysis of militarized target behavior, this combination of multi-layer intention structure and conditional probability enables the calculation of each layer of intention to take into account the intention of the previous layer and the current features. For example, in militarized scenarios, low-level intentions such as specific action intentions (such as deploying weapon systems, building fortifications, etc.) are affected by mid-level intentions (such as tactical intentions such as encirclement and breakthrough) and high-level intentions (such as strategic intentions such as attack and defense). Through this layer-by-layer processing method, the model can more accurately infer the actual intention of the target. In addition, the use of embedding representation to map discrete intention sequences to continuous vector space enables the model to better handle discrete variables. Dimensions of embedding representation and Features Dimensions The combined feature vector is transformed linearly by the weight matrix and the bias vector, and then processed by the nonlinear activation function to generate intermediate features. These intermediate features are further processed to obtain the final output. Through this multi-level and multi-branch structural design, the model can perform well in the analysis of complex militarized target behavior. In step 3, the likelihood function of each layer of intent not only considers the features of the current layer, but also considers the , and also combines the first to the Intent sequence of layers . By embedding , these low-level intentions are mapped into a continuous vector space so that they can be concatenated and processed with the current features. and Linear transformation for the main branch and parallel branch, bias vector and The output of the main branch is processed by the tanh activation function and converted to a probability distribution by the softmax function; the output of the parallel branch is processed by the LeakyReLU activation function and compressed to This combination of multi-layer structure and activation function allows the calculation of each layer of intent to take into account the intent of the previous layer and the current features. This hierarchical processing method can capture the complex dependencies in the target behavior. For example, in a militarized scenario, low-level intent such as specific action intentions will be affected by mid-level and high-level intentions. Through this layer-by-layer processing method, the model can more accurately infer the actual intent of the target.

[0039] Example 6: Using the following formula, a recursive neural network is used to model the Prior probability of the intention layer , each intent layer is a hidden layer of the recurrent neural network: ; ; in, represents a gated recurrent unit; The recurrent neural network corresponds to Hidden layers of the layer intention layer; The recurrent neural network corresponds to Hidden layers of the layer intention layer; is the weight matrix of the recurrent neural network; is the bias matrix of the recurrent neural network.

[0040] Specifically, the core of the recurrent neural network lies in its ability to process sequence data. By combining the hidden state of the previous moment with the current input and updating the current hidden state, the RNN can effectively capture the dependencies in the time series. In the present invention, each intent layer corresponds to a hidden layer of the recurrent neural network. , updated through the gated recurrent unit (GRU). GRU solves the gradient vanishing and gradient exploding problems of traditional RNN when processing long sequence data by introducing update gates and reset gates. In the operation of GRU, the update gate determines how to combine the hidden state of the previous moment with the current input to generate a new hidden state; the reset gate controls how to forget part of the hidden state of the previous moment and focus on the impact of the current input. Through this mechanism, GRU can dynamically adjust the degree of dependence on historical information and current information to capture information over a longer time range. For the prior probability modeling of the intent layer in the present invention, the intent embedding representation of each layer and the hidden state of the previous layer Together as the input of GRU, update the hidden state of the current layer This processing method can transfer and integrate the intention information of the previous layer layer by layer to form a comprehensive judgment of the intention of the current layer. In order to convert the hidden state into the intention probability distribution, the present invention uses the softmax function. By performing linear transformation and combining it with the softmax activation function, we can get the conditional probability distribution of the current layer intent. . This method can effectively map continuous hidden states to discrete intent spaces and generate the probabilities of each intent. In specific applications, this recursive modeling method has significant advantages. First, through the gating mechanism of GRU, the model can flexibly handle long-term dependencies and ensure accurate capture of target behaviors. Second, through layer-by-layer recursive modeling, the model can effectively integrate the complex relationships of multiple layers of intent, thereby improving the accuracy of inference. In militarized target behavior analysis, this method can accurately model complex multi-layer intent relationships, such as intent associations at the strategic, tactical, and specific action levels, enabling intelligent agents to more accurately predict and analyze the behavior of targets. In addition, using embedded representations to map discrete intent sequences to continuous vector spaces enables the model to handle discrete variables. The dimensions and features of the embedded representation are combined to form new feature vectors, which are linearly transformed by weight matrices and bias vectors, and then processed by GRU to generate new hidden states. These hidden states are then converted into probability distributions through the softmax function to generate the prior probabilities of intent at each layer.

[0041] Example 7: Using the following formula, the Bayesian rule is used to calculate Posterior probability of the layer intention : .

[0042] in, is the subscript index; using variational inference, an approximate posterior distribution is defined using the following formula: ; in, is the parameterized approximate posterior distribution; are the parameters of variational inference; is the weight matrix for variational inference; is the bias vector for variational inference.

[0043] Specifically, the core idea of ​​Bayes' rule is to combine the currently observed data (i.e., features ) and previous knowledge (i.e., previous intention ), to update the probability distribution of the target intent. Specifically, the likelihood function Describes the given intention Under the condition of its previous intention, the current feature is observed The prior probability It reflects the fact that before observing the current feature, By combining these two, we can get a comprehensive posterior probability , which is used to more accurately infer the actual intention of the target. In order to effectively calculate the posterior probability, the present invention introduces variational inference. Variational inference defines a parameterized approximate posterior distribution , to approximate the true posterior distribution. This approximate posterior distribution is formed by concatenating the features and the embedded representation of the previous layer's intent, undergoing a weighted linear transformation, and then converting it into a probability distribution through the softmax function. Through this parameterized method, variational inference can transform the calculation problem of the posterior probability into a parameter optimization problem, thereby simplifying the calculation process. In the specific implementation of variational inference, the goal is to optimize the parameters by minimizing the KL divergence between the approximate posterior distribution and the true posterior distribution. . KL divergence is a measure of the difference between two probability distributions. By minimizing KL divergence, the approximate posterior distribution can be made as close to the true posterior distribution as possible. In the optimization process, the parameters of variational inference are iteratively updated using optimization algorithms such as stochastic gradient descent (SGD) to gradually approach the optimal solution. The advantage of this method is that it can efficiently handle complex probability distribution calculation problems, especially when involving large amounts of data and high-dimensional space. By introducing parameterized approximate posterior distributions, the present invention effectively reduces the computational complexity, making the model more efficient and accurate when processing complex multi-layer intention inference. In militarized target behavior analysis, this method is particularly suitable for processing large amounts of complex data and highly uncertain situations. For example, in a complex military scenario, it may be necessary to analyze the intentions of the target at different levels, including strategic-level intentions such as attack and defense, and tactical-level intentions such as encirclement and assault. Through the method of the present invention, the observed feature data and the intention information of the previous level can be combined to calculate the posterior probability of the target intention at each level, thereby providing a comprehensive and detailed intention analysis. In addition, the discrete intention sequence is mapped to a continuous vector space using an embedded representation, so that the model can handle discrete variables. The dimensions and features of the embedded representation are combined to form a new feature vector, which is transformed linearly by the weight matrix and bias vector, and then processed by the nonlinear activation function to generate a new feature representation. These feature representations are then processed by variational inference to generate an approximate posterior probability distribution of intent at each layer. By calculating and updating layer by layer, this method can capture the complex dependencies between intents at different levels, thereby improving the accuracy of intent inference.

[0044] Example 8: Minimize the KL divergence by the following formula: ; in, is the loss function; represents the KL divergence; importance sampling is used to approximate the expectation through the following formula: +const; in, is the sampling times; is the kth intention layer obtained by the sth sampling: const is a constant term; the parameters of variational inference are optimized using stochastic gradient descent through the following formula: ; in, For the The parameters of variational inference after iterative optimization; Parameters of variational inference after optimization at the t+1th iteration; is the learning rate; is the gradient of the loss function with respect to the parameter θ.

[0045] Specifically, in the present invention, the loss function Used to represent KL divergence. By minimizing this function, the parameters of variational inference can be optimized. The form of the loss function integrates the log-likelihood function, the log of the prior distribution, and the log of the approximate posterior distribution. Specifically, this process involves calculating the log of the likelihood function and the prior distribution and subtracting the log of the approximate posterior distribution to obtain the loss function During the calculation, the evidence item It is usually regarded as a constant term and can be ignored to simplify the calculation. In order to approximate the expected value, the present invention uses the importance sampling method. Importance sampling is a Monte Carlo method that estimates the expected value by sampling the distribution multiple times. In this method, by sampling the samples multiple times, an approximation of the expected value can be obtained, thereby effectively estimating the expected value of the loss function. Specifically, the approximate representation of the loss function can be obtained by weighted summing the intentions obtained by sampling, thereby estimating the expected value. This approximate calculation method can greatly reduce the computational complexity while maintaining a high accuracy. During the optimization process, the present invention uses stochastic gradient descent (SGD) to adjust the parameters of variational inference SGD is an iterative optimization algorithm that calculates the loss function About parameters Gradient , gradually update the parameters to minimize the loss function. At each iteration, by adjusting the parameters By making small adjustments, we can gradually approach the optimal solution. The update formula is ,in For the The parameters after iterations, For the The parameters after iterations, is the learning rate. Through this iterative optimization method, the optimal parameters can be gradually approached so that the approximate posterior distribution is as close to the true posterior distribution as possible. This optimization process is of great significance in the analysis of militarized target behavior. In complex military scenarios, the behavior of the target is usually highly uncertain and has complex dependencies. By minimizing the KL divergence, the model can capture these complex relationships more accurately, thereby improving the prediction and analysis capabilities of the target behavior. For example, when analyzing the strategic intention of the target, the current observed data and previous knowledge can be combined to calculate the probability distribution of the target's intention in different situations through the Bayesian inference method. Then, by minimizing the KL divergence, these probability distributions are further optimized to make them more in line with the actual situation. In practical applications, the combination of importance sampling and stochastic gradient descent enables the model to maintain high efficiency and accuracy when processing large-scale data. Importance sampling can effectively estimate the expected value by sampling the distribution multiple times, thereby avoiding the problem of directly calculating complex integrals. Stochastic gradient descent can find the optimal solution in a relatively short time by gradually optimizing the parameters. This method is particularly suitable for processing high-dimensional space and large-scale data scenarios, so that the present invention has a wide range of applicability in practical applications. In order to achieve this optimization process, it is first necessary to model the prior probability of each layer of intent. These prior probabilities are represented by a recursive neural network (RNN), and the prior probability of each layer of intent depends on the hidden state of the previous layer. Through this recursive modeling method, the complex dependencies between the intents of each layer can be captured, thereby providing accurate prior information for subsequent variational inference. When calculating the likelihood function, the likelihood value corresponding to each layer of intent can be obtained by combining the current features and the low-level intent embedding representation. These likelihood values ​​are then combined with the prior probabilities to form a complete posterior probability calculation framework. In practical applications, through this optimization method, the present invention can efficiently handle complex multi-layer intent inference problems. For example, in a complex military scenario, it may be necessary to analyze the intent of the target at different levels, including strategic-level intent such as attack and defense, and tactical-level intent such as encirclement and assault. Through the method of the present invention, the observed feature data and the intent information of the previous level can be combined to calculate the posterior probability of the target intent at each level, thereby providing a comprehensive and detailed intent analysis. This analysis not only considers the current observation data, but also combines historical data and the complex dependencies between intents at each layer, making intent inference more accurate.

[0046] Example 9: In step 4, the approximate posterior distribution is used to perform intent inference through the following formula: ; in, is the result of intention reasoning; then use the following formula to calculate the uncertainty of intention reasoning: ; in, Reasoning about uncertainty in intention.

[0047] Specifically, the core of intention reasoning is to use the approximate posterior distribution To determine the most likely intention. Specifically, by calculating the probability of each possible intention and selecting the intention with the highest probability, the result of intention reasoning can be obtained. This process ensures that the inferred intention is the best explanation based on the current observation data and the previous intention sequence, providing a most likely target behavior prediction. This method utilizes the efficient computing characteristics of the approximate posterior distribution, so that the model can quickly and accurately perform intention reasoning in a complex and changing environment. In addition to inferring the most likely intention, the present invention also pays special attention to the uncertainty assessment of the inference results. By calculating the entropy of the approximate posterior distribution to measure uncertainty, the reliability of the prediction results can be quantified. Entropy is an indicator that measures the degree of dispersion of probability distribution. The higher the entropy value, the more dispersed the distribution, and the greater the uncertainty of the inference result; the lower the entropy value, the more concentrated the distribution, and the higher the confidence of the inference result. This method can effectively quantify the uncertainty of the prediction results, so that decision makers can more clearly understand the reliability of the model inference. In the analysis of militarized target behavior, this uncertainty assessment is particularly important. Decision makers not only need to know the most likely target intention, but also need to understand the credibility of the inference result. For example, in a complex tactical scenario, if the uncertainty of the inference result is high, then the decision maker may need more information or further analysis to confirm the target's intention; if the uncertainty is low, then the decision maker can act more confidently based on the inference result. The specific implementation of this method includes the following steps. First, the probability of each possible intention is calculated using the approximate posterior distribution, and the intention with the highest probability is selected as the inference result. This process is achieved by maximizing the approximate posterior distribution, that is, finding the intention that makes The best intentions . Then, the uncertainty of the inference result is evaluated by calculating the entropy of the approximate posterior distribution. The calculation of entropy involves summing and taking the logarithm of the probability distribution of all possible intentions, thereby obtaining a quantitative uncertainty measure. In this way, the present invention can provide not only accurate intention inference results, but also provide a clear confidence assessment for these results. In practical applications, this method can help military commanders rely more confidently on the inference results of the model when formulating strategies. For example, when faced with complex enemy actions, commanders can use the enemy intentions inferred by the model and the corresponding uncertainty assessment to respond quickly and formulate corresponding tactics.

[0048] Embodiment 10: A militarized target behavior analysis intelligent agent testing device, characterized in that it includes: a data acquisition part, used to acquire multi-source data of militarized targets, perform data preprocessing on each category of data in the multi-source data, perform timestamp alignment and spatial registration, and then fuse them into an input tensor; an intention model construction part, used to extract tensor features of the input tensor; based on the tensor features, define a multi-layer intention hierarchy, each intention layer contains multiple possible intentions, and use conditional probability to represent the dependency relationship between the intention layers; an intention reasoning model construction part, used to define the likelihood function of each intention layer; use a recursive neural network to model prior probabilities; use Bayes' rule to calculate posterior probabilities; iteratively execute the following process to complete training: use variational inference to define an approximate posterior distribution, and then minimize the KL divergence; use importance sampling to approximate expectations, and use stochastic gradient descent to optimize the parameters of variational inference; when the number of iterative executions reaches the set number, execute step 4: the intention reasoning part, used to perform intention reasoning using the approximate posterior distribution, and calculate the uncertainty of intention reasoning.

[0049] Although the specific embodiments of the present invention are described above, it should be understood by those skilled in the art that these specific embodiments are only illustrative, and those skilled in the art may omit, replace, and change the details of the above methods and systems in various ways without departing from the principles and essence of the present invention. For example, merging the above method steps so as to perform substantially the same functions in substantially the same manner to achieve substantially the same results is within the scope of the present invention. Therefore, the scope of the present invention is limited only by the appended claims.

Claims

1. A militarized target behavior analysis agent testing method, characterized in that: The method comprises: Step 1: Obtain multi-source data of militarized targets, pre-process the data of each category in the multi-source data, align the timestamps and register the space, and then fuse them into an input tensor; Step 2: Extract tensor features of the input tensor; based on the tensor features, define a multi-layer intent hierarchy, where each intent layer contains multiple possible intents, and use conditional probabilities to represent the dependencies between intent layers; Step 3: Define the likelihood function of each intent layer; use a recursive neural network to model the prior probability; use Bayes' rule to calculate the posterior probability; iteratively perform the following process to complete the training: use variational inference to define an approximate posterior distribution, and then minimize the KL divergence; use importance sampling to approximate the expectation, and use stochastic gradient descent to optimize the parameters of variational inference; when the number of iterations reaches the set number, execute step 4: Step 4: Use the approximate posterior distribution to perform intent inference and calculate the uncertainty of intent inference.

2. The militarized target behavior analysis agent testing method according to claim 1, characterized in that: The multi-layered intention hierarchy in step 2 includes at least three intention layers, namely: strategic intention layer, tactical intention layer and specific action intention layer; the strategic intention layer includes at least the following intentions: attack, defense, retreat, reorganization and deception; the tactical intention layer includes at least the following intentions: encirclement, breakthrough, reconnaissance, blocking, detour and feint; the specific action intention layer includes at least the following intentions: deploying weapon systems, building fortifications, carrying out air strikes, conducting electronic interference, launching ground attacks, conducting special operations, and establishing supply lines.

3. The militarized target behavior analysis agent testing method according to claim 2, characterized in that: The multi-source data includes at least the following categories of data: position data, speed data, acceleration data, direction data, environment data, altitude data, longitude data and latitude data; Step 1: The process of preprocessing the data of each category in the multi-source data includes: performing data noise reduction and data normalization on the data of each category; In step 1, the preprocessed data of each category is timestamped and spatially registered, and then the data is resampled and interpolated to form an input tensor. In step 2, a deep neural network is used to extract the tensor features of the input tensor. .

4. The militarized target behavior analysis agent testing method according to claim 3, characterized in that: In step 2, define Layer Intent Structure, Layer intention layer possible intents; initialize the intent vector for each layer: ； in, , is the index number; Indicates The intention layer intention; each is a discrete variable and can be possible values; Tensor features The nonlinear transformation is performed through the following formula to obtain the feature , to accommodate different levels of intent reasoning: ； in, represents element-wise multiplication, and They are the preset feature weight matrix and feature bias matrix respectively; Use conditional probability to represent the dependencies between intent layers: ； in, Indicates the highest level of intention. Indicates the lowest level of intention; For a given feature The joint probability distribution of the entire intent hierarchy when ; For a given When the intention and features of the layers are combined, the conditional probability of the lowest level intention; For a given When layering intent and features, Conditional probability of layer intention; The conditional probability of the top-level intent given the features.

5. The militarized target behavior analysis agent testing method according to claim 4, characterized in that: In step 3, the following formula is used to define Likelihood function of the layer intention layer , which contains a multi-layer perceptron of a main branch and a multi-layer sensor of a parallel branch. The main branch uses the tanh activation function, and the parallel branch uses the LeakyReLU activation function, which is expressed by the following formula: ； in, From the 1st layer to the The intentional sequence of layers; is the embedding representation of the low-level intent, which is a function that maps a discrete sequence of intents to a continuous vector space. Let the dimension after embedding be ;set up The latitude is , is the concatenation of feature and low-level intent embedding, with dimension ; For the The weight matrix of the layer intention layer has the dimension ,in It is Dimensions of the layer intention layer; For the The bias vector of the layer intent layer, dimension is ; For the The weight matrix of the layer intention layer has the dimension ,in is the dimension of the output; For the The bias vector of the layer intent layer, dimension is ; For the parallel branch The weight matrix of the layer intention layer has the dimension ; For the parallel branch The bias vector of the layer intent layer, dimension is ; For the parallel branch The weight matrix of the layer intention layer has the dimension ; For the parallel branch The bias vector of the layer intent layer, dimension is ; is the rectified linear unit activation function with leakage, is a function variable; is the Softmax function, which is used to convert the output into a probability distribution; is the Sigmoid activation function, which is used to compress the output to the (0,1) interval.

6. The militarized target behavior analysis agent testing method according to claim 5, characterized in that: Using the following formula, we use a recurrent neural network to model the Prior probability of the intention layer , each intent layer is a hidden layer of the recurrent neural network: ； ； in, represents a gated recurrent unit; The recurrent neural network corresponds to Hidden layers of the layer intention layer; The recurrent neural network corresponds to Hidden layers of the layer intention layer; is the weight matrix of the recurrent neural network; is the bias matrix of the recurrent neural network.

7. The militarized target behavior analysis agent testing method according to claim 6, characterized in that: The Bayesian rule is used to calculate the Posterior probability of the layer intention : 。 in, is the subscript index; using variational inference, an approximate posterior distribution is defined using the following formula: ； in, is the parameterized approximate posterior distribution; are the parameters of variational inference; is the weight matrix for variational inference; is the bias vector for variational inference.

8. The militarized target behavior analysis agent testing method according to claim 7, characterized in that: Minimize the KL divergence by the following formula: ； in, is the loss function; represents the KL divergence; importance sampling is used to approximate the expectation through the following formula: +const; in, is the sampling times; is the kth intention layer obtained by the sth sampling: const is a constant term; the parameters of variational inference are optimized using stochastic gradient descent through the following formula: ； in, For the The parameters of variational inference after iterative optimization; Parameters of variational inference after optimization at the t+1th iteration; is the learning rate; is the gradient of the loss function with respect to the parameter θ.

9. The militarized target behavior analysis agent testing method according to claim 8, characterized in that: In step 4, the approximate posterior distribution is used for intent inference using the following formula: ； in, is the result of intention reasoning; then use the following formula to calculate the uncertainty of intention reasoning: ； in, Reasoning about uncertainty in intention.

10. A militarized target behavior analysis agent testing device, characterized in that: include: The data acquisition part is used to obtain multi-source data of militarized targets, perform data preprocessing on each category of the multi-source data, perform timestamp alignment and spatial registration, and then fuse them into an input tensor; Intent The model building part is used to extract tensor features of the input tensor. Based on the tensor features, a multi-layer intent hierarchy is defined. Each intent layer contains multiple possible intents, and conditional probabilities are used to represent the dependencies between the intent layers. The intention inference model building part is used to define the likelihood function of each intent layer; a recurrent neural network is used to model the prior probability; Calculate posterior probabilities using Bayes’ rule; The following process is iterated to complete the training: using variational inference, an approximate posterior distribution is defined, and then the KL divergence is minimized; importance sampling is used to approximate the expectation, and stochastic gradient descent is used to optimize the parameters of variational inference; when the number of iterations reaches the set number, step 4 is executed: the intention reasoning part, which is used to perform intention reasoning using the approximate posterior distribution and calculate the uncertainty of intention reasoning.

Images