Adversary intent pattern recognition method based on multi-modal feature fusion
By using a multimodal feature fusion method for wargaming simulation systems, the problem of feature fragmentation caused by single-modal data in existing technologies is solved, achieving high accuracy and robust recognition of the opponent's intentions and improving the decision-making capabilities of wargaming simulation systems.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- CHINESE PEOPLES LIBERATION ARMY INFORMATION SUPPORT CORPS ENGINEERING UNIVERSITY
- Filing Date
- 2026-02-14
- Publication Date
- 2026-05-29
Smart Images
Figure CN122114168A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of intent recognition technology, and in particular to a method for opponent intent pattern recognition based on multimodal feature fusion. Background Technology
[0002] Existing methods for identifying adversary intentions have fundamental limitations in wargaming simulation systems. Specifically, they rely too heavily on single-modal data, such as analyzing only action state sequences or processing only static geographical location information, resulting in fragmented and incomplete extracted situational features.
[0003] Furthermore, due to the lack of a multimodal collaboration mechanism, it is impossible to effectively integrate the temporal evolution of action states with the dynamic correlation characteristics of the spatial environment, resulting in a significant increase in intent recognition bias.
[0004] Meanwhile, traditional models are not adaptable to local observation information under the "fog of war" mechanism, and are not robust in complex adversarial environments with high information loss and strong noise interference, making it difficult to support the need for accurate decision-making under high uncertainty conditions.
[0005] In summary, existing technologies suffer from technical problems such as feature fragmentation due to reliance on single-modal data, resulting in insufficient accuracy in intent recognition and poor adaptability to local observation information. Summary of the Invention
[0006] To address the aforementioned deficiencies or improvement needs of existing technologies, this invention provides a method for opponent intent pattern recognition based on multimodal feature fusion. This method addresses the technical problem that existing technologies suffer from insufficient accuracy in intent recognition and poor adaptability to local observation information due to feature fragmentation caused by reliance on single-modal data.
[0007] To achieve the above objectives, this invention provides a method for opponent intent pattern recognition based on multimodal feature fusion, the method comprising:
[0008] Historical game data is retrieved from the adversarial simulation system and segmented into time slices to obtain multiple opponent time-series slices, each bearing a preset intent label. Opponent situation information features are extracted from these multiple opponent time-series slices to generate multiple sample multimodal situation information features. These sample multimodal situation information features are used as training data, and the corresponding preset intent labels are used as supervision signals to iteratively optimize the parameters of the opponent intent recognition model until the model output meets a preset convergence condition. During real-time gameplay, locally observed game information within a preset time window is input into the opponent intent recognition model for intent inference, outputting the current opponent intent probability distribution.
[0009] In one implementation, historical game data is retrieved from the adversarial simulation system, time-sliced to obtain multiple opponent time-series slices, and the following processing is also performed:
[0010] The opponent's situation information is extracted from the global confrontation record of game P in the historical game data to obtain the opponent's situation time series set of game P; the opponent's situation time series set of game P is divided into time slices using a preset time window length to obtain multiple opponent time slices, wherein each opponent time slice includes multiple time-series action state data and multiple time-series geographical location data of multiple opponent combat units.
[0011] In one implementation, the following processing is also performed:
[0012] A pre-constructed action state feature extraction module is provided, comprising a cascaded first input layer, a masking layer, a first LSTM layer, and a first dropout layer. The first LSTM layer has a 2-layer hidden structure with a hidden layer dimension of 256, and the first dropout layer has a dropout rate of 0.4. A pre-constructed geolocation feature extraction module is also provided, comprising a cascaded second input layer, a flattening layer, a sine-cosine position encoding layer, a Transformer encoder layer, a second LSTM layer, a global average pooling layer, and a second dropout layer. The system is configured with 3 layers, 8 multi-head attention heads, a feedforward neural network with a hidden layer dimension of 512 and a ReLU activation function. The second LSTM layer is a 2-layer hidden structure with a hidden layer dimension of 256, and the dropout rate of the second Dropout layer is 0.4. A pre-built feature fusion and classification module is constructed, which includes a concatenation layer, a fully connected layer with a ReLU activation function, a classification layer, and an output layer with a Softmax activation function. The first Dropout layer of the action state feature extraction module and the second Dropout layer of the geographic location feature extraction module are connected to the concatenation layer of the feature fusion and classification module to complete the framework construction of the opponent intent recognition model.
[0013] In one implementation, the multimodal situational information features of the multiple samples are used as training data, and the corresponding multiple preset intent labels are used as supervision signals to perform iterative optimization training of the opponent intent recognition model until the model output meets the preset convergence condition. Then, the following processing is performed:
[0014] The training dataset, consisting of the multimodal situational information features of multiple samples and multiple preset intent labels, is separated into a training set, a validation set, and a test set. The Adam optimizer is used, and optimizer parameters are set, including the learning rate and batch size. Using the optimizer parameters as constraints, training data is retrieved from the training set to iterate the opponent intent recognition model for a single round, and the intent recognition accuracy on the validation set is calculated. When the improvement in intent recognition accuracy of the opponent intent recognition model is less than 0.1% after M consecutive iterations, the output of the opponent intent recognition model is determined to meet the preset convergence condition. The recognition accuracy of the opponent intent recognition model is verified using the test set.
[0015] In one implementation, during real-time game play, locally observed game information within a preset time window is input into the opponent's intent recognition model for intent reasoning, outputting the current opponent's intent probability distribution, and further performing the following processing:
[0016] The local observation game information is segmented using the specified time window length to obtain multiple real-time action state data and multiple real-time geographical location data for multiple opponent combat units; the multiple real-time action state data are integrated using the specified time window length to obtain real-time action state information; the real-time action state information is input into the action state feature extraction module for action state temporal feature extraction, and a partial action state feature is output; the multiple real-time geographical location data are integrated using the specified time window length to obtain real-time geographical location information; the real-time geographical location information is input into the geographical location feature extraction module for spatiotemporal fusion feature extraction, and a partial geographical location feature is output; the partial action state feature and partial geographical location feature are input into the feature fusion and classification module for intent classification under feature fusion, and the current opponent intent probability distribution is output.
[0017] In one implementation, the real-time action state information is input into the action state feature extraction module for action state temporal feature extraction, and partial action state features are output. The following processing is also performed:
[0018] After the real-time action state information is input into the Masking layer through the first input layer to mask invalid data, the action state temporal feature extraction based on the Long Short-Term Memory network is performed through the first LSTM layer to obtain a partial action state feature vector; the partial action state feature vector is then subjected to random deactivation regularization processing with a dropout rate of 0.4 through the first Dropout layer to output the partial action state feature.
[0019] In one implementation, the real-time geographic location information is input into the geographic location feature extraction module for spatiotemporal fusion feature extraction, outputting partial geographic location features, and the following processing is also performed:
[0020] The real-time geographic location information is input into the flattening layer through the second input layer for two-dimensional spatial feature serialization. Then, the location encoding is added through the sine-cosine location encoding layer to obtain location-encoded sequence data. The location-encoded sequence data is then modeled using a multi-head attention mechanism for spatial global association through the Transformer encoder layer to output a spatial feature vector. The spatial feature vector is then input into the second LSTM layer for spatiotemporal fusion feature extraction to obtain a spatiotemporal feature vector. The spatiotemporal feature vector is then compressed using the global average pooling layer and then subjected to random deactivation regularization with a dropout rate of 0.4 through the second Dropout layer to output the partial geographic location features.
[0021] In one implementation, the partial action state features and partial geographic location features are input into the feature fusion and classification module for intent classification based on feature fusion, and the current opponent intent probability distribution is output. The following processing is also performed:
[0022] The partial action state features and partial geographical location features are input into the splicing layer for feature splicing to obtain the opponent's situation information feature vector. After the opponent's situation information feature vector is nonlinearly transformed by the fully connected layer, the intent probability distribution is calculated by the Softmax activation function of the classification layer, and the current opponent's intent probability distribution is output by the output layer.
[0023] In one implementation, adversary situation information feature extraction is performed on the multiple adversary time-series slices to generate multiple sample multimodal situation information features, and the following processing is also performed:
[0024] Perform bimodal feature extraction on the time slices of the multiple opponents to obtain multiple sample action state features and multiple sample geographical location features; fuse the multiple sample action state features and multiple sample geographical location features to generate the multiple sample multimodal situational information features.
[0025] In one implementation, bimodal feature extraction is performed on the multiple opponent time slices to obtain multiple sample action state features and multiple sample geographical location features, and the following processing is also performed:
[0026] The first action state data of multiple adversary combat units in the first adversary time slice are integrated using the time window length to obtain the first action state information, and then the first sample action state features are extracted. The first geographical location data of multiple adversary combat units in the first adversary time slice are integrated using the time window length to obtain the first geographical location information, and then the first sample geographical location features are extracted. In this way, bimodal feature extraction is performed on the multiple adversary time slices to obtain the multiple sample action state features and multiple sample geographical location features.
[0027] One or more technical solutions provided in this invention have at least the following technical effects or advantages:
[0028] The method provided in this invention retrieves historical game data from an adversarial simulation system, segments it into time slices to obtain multiple opponent time-series slices, each bearing a preset intent label. It then extracts opponent situation information features from these slices, generating multiple sample multimodal situation information features. These features are used as training data, and the corresponding preset intent labels are used as monitoring signals to iteratively optimize the parameters of an opponent intent recognition model until the model output meets a preset convergence condition. During real-time gameplay, locally observed game information within a preset time window is input into the opponent intent recognition model for intent inference, outputting the current opponent intent probability distribution. This method effectively integrates multimodal time-series data, reduces information uncertainty during gameplay, improves the accuracy and robustness of intent recognition, and helps the player enhance their game adversarial effectiveness and survivability. Attached Figure Description
[0029] 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.
[0030] Figure 1 A schematic diagram of the adversary intent pattern recognition method based on multimodal feature fusion provided by the present invention is shown.
[0031] Figure 2 The diagram illustrates the process of inferring and outputting the probability distribution of the current opponent's intent in the opponent intent pattern recognition method based on multimodal feature fusion provided by the present invention. Detailed Implementation
[0032] This invention provides a method for opponent intent pattern recognition based on multimodal feature fusion, which solves the technical problem that existing technologies rely on single-modal data, resulting in feature fragmentation, insufficient accuracy of intent recognition, and poor adaptability to local observation information.
[0033] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. Furthermore, the technical features involved in the various embodiments of this invention described below can be combined with each other as long as they do not conflict with each other.
[0034] Unless otherwise expressly stated, throughout the specification and claims, the term "comprising" or its variations such as "including" or "comprises" shall be understood to include the stated elements or components without excluding other elements or other components.
[0035] The flowchart of the adversary intent pattern recognition method based on multimodal feature fusion provided in this embodiment of the invention is shown below. Figure 1 The method includes:
[0036] A100: Retrieve historical game data from the adversarial simulation system, perform time-slice segmentation, and obtain multiple opponent time-series slices, in which the opponent time-series slices are marked with preset intent labels.
[0037] In one implementation, historical game data is retrieved from the adversarial simulation system, and time-sliced to obtain multiple opponent time-series slices. Step A100 of the method includes:
[0038] The opponent's situation information is extracted from the global confrontation record of game P in the historical game data to obtain the opponent's situation time series set of game P; the opponent's situation time series set of game P is divided into time slices using a preset time window length to obtain multiple opponent time slices, wherein each opponent time slice includes multiple time-series action state data and multiple time-series geographical location data of multiple opponent combat units.
[0039] Specifically, this embodiment obtains historical game data from the simulation records of the wargaming simulation system, which includes behavioral and situational data of the combat units of both sides at continuous time points in multiple confrontations.
[0040] Behavioral data covers the types of actions of combat units, such as movement, stopping, firing, and aiming; situational data includes terrain and environmental information within the observation range of combat units, as well as the health, position, and dynamic movement information of other combat units.
[0041] These data are generated through the fog of war mechanism unique to wargaming systems, simulating the environmental constraints of real combat where one's own side can only observe local information.
[0042] This embodiment first extracts the opponent's situation information from the global confrontation record of game P in the historical game data. Here, each confrontation record is organized into an opponent situation time series set in chronological order, which fully records the changes in the action status and geographical location evolution of all combat units in that game at each time node. For example, the action status of a combat unit at time node t can be reflected as moving towards a specific coordinate or firing at a target, and its geographical location can be reflected as the terrain visibility relationship and elevation gradient distribution centered on the unit.
[0043] Each game's time series is divided into equal-length segments using a preset time window length T, generating multiple opponent time series slices. The preferred time window length T is 50 time steps. Each slice covers 50 consecutive time steps and includes the temporal action status data and temporal geographical location data of all opponent combat units within the window.
[0044] Temporal action status data refers to the action sequence of a combat unit within 50 steps, with a dimension of 50×16×8. For example, the 5th unit has a mobility speed of 5.2 and a health value of 80% in step 30. Temporal geolocation data refers to the spatial information sequence of a combat unit within 50 steps, with a dimension of 50×16×(11×11×8). For example, the elevation gradient distribution of the 3rd unit in the 11×11 grid around its perimeter in step 40.
[0045] Each slice is bound to a preset intent label at the end of its window. This label is derived from the opponent's actual decision intent recorded by the simulation system. For example, the intent in step 50 is labeled as a detour.
[0046] This embodiment uses a fixed time window to discretize the continuous game, adapting to the input requirements of the action state feature extraction module and the geographical location feature extraction module in the subsequent opponent intention recognition model. The slice contains dual-modal data of action state and geographical location, providing a foundation for multimodal feature fusion. Intention label binding enables the subsequent opponent intention recognition model to learn the mapping rules from situation features to intentions, forming the data foundation for training the opponent intention recognition model.
[0047] A200: Perform opponent situation information feature extraction on the multiple opponent time-series slices to generate multiple sample multimodal situation information features.
[0048] In one implementation, opponent situation information feature extraction is performed on the multiple opponent time-series slices to generate multiple sample multimodal situation information features. The method step A200 includes:
[0049] A210: Perform bimodal feature extraction on the time slices of the multiple opponents to obtain multiple sample action state features and multiple sample geographical location features; A220: Fuse the multiple sample action state features and multiple sample geographical location features to generate the multiple sample multimodal situational information features.
[0050] In one implementation, bimodal feature extraction is performed on the multiple opponent time slices to obtain multiple sample action state features and multiple sample geographical location features. Method step A210 includes:
[0051] A211: Using the time window length, integrate multiple first action state data of the multiple adversary combat units in the first adversary time slice to obtain first action state information, and then extract first sample action state features; A212: Using the time window length, integrate multiple first geographical location data of the multiple adversary combat units in the first adversary time slice to obtain first geographical location information, and then extract first sample geographical location features; and so on, perform bimodal feature extraction on the multiple adversary time slices to obtain multiple sample action state features and multiple sample geographical location features.
[0052] This embodiment performs bimodal feature extraction on multiple opponent time-series slices to generate sample action state features and sample geographical location features respectively. In specific implementation, the action state data and geographical location data of each opponent time-series slice are processed independently for feature extraction.
[0053] Specifically, a fixed time window length is used to integrate the action status data of all opponent combat units in the first opponent time-series slice. The specific operation is as follows: the action status data of 16 combat units in the slice are arranged in chronological order over 50 time steps to form a data structure with dimensions of 50×16×8, where 8 represents the action feature dimension of each unit, including type, behavior, health, mobility speed, etc. After integration, the first action status information is obtained, and the first sample action status features are extracted.
[0054] A fixed time window of 50 is used to integrate the geographic location data of all adversary combat units in the first adversary time-series slice. Specifically, the geographic location data of 16 combat units within the slice are arranged chronologically over 50 time steps. Each unit corresponds to 11×11 grid spatial data, containing 8 feature channels such as elevation and vegetation. After integration, a data structure with dimensions of 50×16×(11×11×8) is formed. After obtaining the first geographic location information, the geographic location features of the first sample are extracted.
[0055] Repeat steps A211 and A212 for all remaining opponent time slices to obtain the multiple sample action state features and multiple sample geographical location features. Fuse the multiple sample action state features and multiple sample geographical location features to generate the multiple sample multimodal situation information features.
[0056] A300: Using the multimodal situational information features of the multiple samples as training data and the corresponding multiple preset intent labels as supervision signals, the parameters of the adversary intent recognition model are iteratively optimized and trained until the model output meets the preset convergence condition.
[0057] In one implementation, the method further includes:
[0058] A pre-built action state feature extraction module is provided, wherein the action state feature extraction module includes a cascaded first input layer, a Masking layer, a first LSTM layer and a first Dropout layer, wherein the first LSTM layer is a 2-layer hidden structure with a hidden layer dimension of 256, and the dropout rate of the first Dropout layer is 0.4.
[0059] A pre-built geographic location feature extraction module is provided, wherein the action state feature extraction module includes a cascaded second input layer, a flattening layer, a sine-cosine position encoding layer, a Transformer encoder layer, a second LSTM layer, a global average pooling layer, and a second Dropout layer. The Transformer encoder layer is set to 3 layers, the multi-head attention head is 8, the hidden layer dimension of the feedforward neural network is 512, the activation function is ReLU, the second LSTM layer is a 2-layer hidden structure with a hidden layer dimension of 256, and the dropout rate of the second Dropout layer is 0.4.
[0060] A pre-built feature fusion and classification module is provided, which includes a splicing layer, a fully connected layer with ReLU activation function, a classification layer, and an output layer with Softmax activation function.
[0061] The first Dropout layer of the action state feature extraction module and the second Dropout layer of the geographic location feature extraction module are connected to the splicing layer of the feature fusion and classification module to complete the framework construction of the opponent intent recognition model.
[0062] In one implementation, the multimodal situational information features of the multiple samples are used as training data, and the corresponding multiple preset intent labels are used as supervision signals to perform iterative optimization training of the opponent intent recognition model until the model output satisfies the preset convergence condition. The method step A300 includes:
[0063] A310: Separate the training dataset, consisting of the multimodal situational information features of the multiple samples and multiple preset intent labels, into a training set, a validation set, and a test set; A320: Use the Adam optimizer and set the optimizer parameters, including the learning rate and batch size; A330: Using the optimizer parameters as constraints, retrieve training data from the training set and iterate the opponent intent recognition model for a single round, then calculate the intent recognition accuracy on the validation set; A340: When the improvement in intent recognition accuracy of the opponent intent recognition model is less than 0.1% after M consecutive iterations, determine that the output of the opponent intent recognition model meets the preset convergence condition; A350: Use the test set to verify the recognition accuracy of the opponent intent recognition model.
[0064] Specifically, the specific model structure of the pre-constructed adversary intent recognition model in this embodiment is as follows:
[0065] A pre-built action state feature extraction module is constructed, consisting of four cascaded components: a first input layer, a Masking layer, a first LSTM layer, and a first Dropout layer. The first LSTM layer employs a two-layer hidden structure, with each hidden unit having a dimension of 256, used to capture the temporal evolution of action states. The first Dropout layer has a dropout rate of 0.4, using random deactivation regularization to prevent model overfitting. The function of the action state feature extraction module is to process action state information within a time window and output a feature vector with temporal representation capabilities.
[0066] A pre-built geolocation feature extraction module consists of seven cascaded components: a second input layer, a flattening layer, a sine-cosine position encoding layer, a Transformer encoder layer, a second LSTM layer, a global average pooling layer, and a second Dropout layer. The Transformer encoder layer is a three-layer structure with an 8-head attention mechanism. The feedforward neural network has a hidden layer dimension of 512 and uses the ReLU activation function to capture the global correlation of spatial features. The second LSTM layer also uses a two-layer hidden structure, with each hidden unit having a dimension of 256, to fuse spatiotemporal dynamic features. The global average pooling layer compresses the features into a low-dimensional vector, and the second Dropout layer has a dropout rate of 0.4 to prevent overfitting. The geolocation feature extraction module processes spatial information within a time window and outputs a feature vector that fuses spatiotemporal characteristics.
[0067] A pre-built feature fusion and classification module is constructed, consisting of four core components: a concatenation layer, a fully connected layer with ReLU activation, a classification layer, and an output layer with Softmax activation. The concatenation layer connects action state features and geographic location features along the feature dimension to form a comprehensive situational feature vector. The fully connected layer with ReLU activation performs a non-linear transformation on the fused features to extract higher-order abstract features. The classification layer calculates the probability distribution of intent categories using the Softmax activation function, ultimately outputting the prediction result of the opponent's intent.
[0068] The final output layer (first Dropout layer) of the action state feature extraction module and the final output layer (second Dropout layer) of the geolocation feature extraction module are connected to the concatenation layer of the feature fusion and classification module to complete the overall framework construction of the opponent intent recognition model. This connection method ensures that the bimodal features are deeply fused while retaining their respective temporal and spatial characteristics, providing complete feature input for intent classification.
[0069] The training process of the adversary intent recognition model is as follows:
[0070] The training dataset is constructed by combining multimodal situational information features from multiple samples with corresponding predefined intent labels. This dataset is then divided into training, validation, and test sets in a 6:2:2 ratio. The training set is used for updating model parameters, the validation set is used to monitor the training process and adjust hyperparameters, and the test set is used for final model performance evaluation.
[0071] The Adam optimizer was used for model training, with the optimizer parameters set to a learning rate of 0.001 and a batch size of 32. The learning rate controls the step size of the update parameter, and the batch size determines the number of samples used in each iteration. These two parameters directly affect the model's convergence speed and stability.
[0072] With the optimizer parameters as constraints, sample data in batches of 32 are retrieved from the training set and input into the hand intent recognition model for forward and backward propagation to update the model parameters. After each round of full training set iteration, the intent recognition accuracy on the validation set is calculated, which reflects the model's generalization ability on unseen data.
[0073] If the improvement in validation set accuracy is less than 0.1% after 10 consecutive iterations (M is preferably 10), the model output is considered to have met the preset convergence condition. For example, if the accuracy increases from 75.2% to 75.25% between the 20th and 30th iterations, an increase of 0.05% which is less than the threshold, training is stopped. This dynamic early stopping mechanism avoids overfitting and ensures the model reaches its optimal state. Finally, the trained model is input using the test set to calculate the intent recognition accuracy for final performance verification.
[0074] A400: During real-time game play, the game information observed locally within a preset time window is input into the opponent's intent recognition model for intent reasoning, and the current opponent's intent probability distribution is output.
[0075] In one implementation, see Figure 2 Using the multimodal situational information features of the multiple samples as training data, during real-time game play, locally observed game information of a preset time window length is input into the opponent's intent recognition model for intent inference, and the current opponent's intent probability distribution is output. The method step A400 includes:
[0076] A410: Segment the local observation game information using the time window length to obtain multiple real-time action state data and multiple real-time geographical location data of the multiple opponent combat units; A420: Integrate the multiple real-time action state data using the time window length to obtain real-time action state information; A430: Input the real-time action state information into the action state feature extraction module for action state temporal feature extraction, and output partial action state features; A440: Integrate the multiple real-time geographical location data using the time window length to obtain real-time geographical location information; A450: Input the real-time geographical location information into the geographical location feature extraction module for spatiotemporal fusion feature extraction, and output partial geographical location features; A460: Input the partial action state features and partial geographical location features into the feature fusion and classification module for intent classification under feature fusion, and output the current opponent intent probability distribution.
[0077] In one implementation, the real-time action state information is input into the action state feature extraction module for action state temporal feature extraction, and partial action state features are output. Method step A430 includes:
[0078] After the real-time action state information is input into the Masking layer through the first input layer to mask invalid data, the action state temporal feature extraction based on the Long Short-Term Memory network is performed through the first LSTM layer to obtain a partial action state feature vector; the partial action state feature vector is then subjected to random deactivation regularization processing with a dropout rate of 0.4 through the first Dropout layer to output the partial action state feature.
[0079] In one implementation, the multimodal situational information features of the multiple samples are used as training data, and the real-time geographic location information is input into the geographic location feature extraction module for spatiotemporal fusion feature extraction, outputting partial geographic location features. The method step A450 includes:
[0080] The real-time geographic location information is input into the flattening layer through the second input layer for two-dimensional spatial feature serialization. Then, the location encoding is added through the sine-cosine location encoding layer to obtain location-encoded sequence data. The location-encoded sequence data is then modeled using a multi-head attention mechanism for spatial global association through the Transformer encoder layer to output a spatial feature vector. The spatial feature vector is then input into the second LSTM layer for spatiotemporal fusion feature extraction to obtain a spatiotemporal feature vector. The spatiotemporal feature vector is then compressed using the global average pooling layer and then subjected to random deactivation regularization with a dropout rate of 0.4 through the second Dropout layer to output the partial geographic location features.
[0081] In one implementation, the partial action state features and partial geographic location features are input into the feature fusion and classification module for intent classification under feature fusion, and the current opponent intent probability distribution is output. Method step A460 includes:
[0082] The partial action state features and partial geographical location features are input into the splicing layer for feature splicing to obtain the opponent's situation information feature vector. After the opponent's situation information feature vector is nonlinearly transformed by the fully connected layer, the intent probability distribution is calculated by the Softmax activation function of the classification layer, and the current opponent's intent probability distribution is output by the output layer.
[0083] Specifically, the local observation game information is structured and segmented using the time window length to obtain multiple real-time action status data and multiple real-time geographical location data of the multiple opposing combat units.
[0084] Real-time action status data includes the action sequences of 16 combat units (i.e., multiple opposing combat units) over the last 50 time steps, such as movement direction and changes in health points; real-time geolocation data includes the spatial information sequences of the same 16 units (i.e., multiple opposing combat units) within the same time window, such as the terrain elevation distribution within an 11×11 grid around each unit.
[0085] Using a time window length of 50 as a baseline, the segmented real-time action state data are arranged according to combat units and time steps, forming real-time action state information with dimensions of 50×16×8. Here, 8 represents the action feature dimension of each unit, including type, behavior, percentage of health, and movement speed. This information serves as the input to the action state feature extraction module.
[0086] When the real-time motion state information is input into the motion state feature extraction module for motion state temporal feature extraction and partial motion state features are output, the data processing procedure within the module is as follows:
[0087] First, real-time action state information with dimensions of 50×16×8 is input into the Masking layer through the first input layer. This layer automatically identifies and filters out invalid padding data, such as placeholder data added when the time step is insufficient in the early stage of the game, using a mask value of 999.0, to ensure that the model only processes valid action state sequences.
[0088] The masked valid data is input into the first LSTM layer, which adopts a two-layer hidden structure. Each layer contains 256 hidden units. The temporal dependencies of action states within 50 time steps are captured through the gating mechanism of forget gate, input gate, and output gate. For example, the behavior transition mode of a combat unit from moving to firing. The output is a temporal feature vector with a dimension of 256.
[0089] The feature vector is then fed into the first Dropout layer, which randomly masks some neuron outputs at a fixed dropout rate of 0.4. For example, 40% of neuron activation values are randomly dropped during the training phase. This forces the model to disperse its feature representation capabilities and prevents overfitting caused by excessive sensitivity to specific action patterns.
[0090] The final output is a 256-dimensional real-time action state feature vector, which preserves the temporal evolution of the action state and serves as the basic input for subsequent feature fusion. The processing flow of the action state feature extraction module strictly reuses the parameters of the training phase module to ensure consistency between the real-time inference and training feature spaces.
[0091] Using a time window length of 50 as a baseline, the segmented real-time geographic location data are arranged according to combat units and time steps, forming real-time geographic location information with dimensions of 50×16×(11×11×8). Here, 8 represents the number of geographic location feature channels, including vegetation, roads, elevation, etc. This information serves as the input to the geographic location feature extraction module.
[0092] When the real-time geographic location information is input into the geographic location feature extraction module for spatiotemporal fusion feature extraction and partial geographic location features are output, the data processing procedure within the module is as follows:
[0093] First, the real-time geographic location information with dimensions of 50×16×(11×11×8) is input into the flattening layer through the second input layer. This layer flattens the 11×11 spatial grid of each combat unit at a single time step into 121×8 one-dimensional sequence data. For example, the 11×11×8 three-dimensional tensor of the third unit at the 40th step is converted into 121 spatial location points, each point containing 8-dimensional feature vectors, which are serialized data.
[0094] Next, a sine-cosine positional coding layer adds positional coding to the flattened sequence. The sine function is used to generate coding vectors for even positions and the cosine function for odd positions. The coding dimension is consistent with the feature dimension, which is 8-dimensional, to ensure that the absolute and relative positional relationships of each spatial position are preserved. For example, the center point and the edge point of the grid form different coding patterns.
[0095] The position-encoded sequence data is then input into the Transformer encoder layer, which contains three cascaded encoder blocks. Each block performs a multi-head attention mechanism and a feedforward neural network operation: the multi-head attention mechanism splits the sequence data into eight independent subspaces and computes the global association weights between 121 location points in parallel, such as identifying spatial dependencies in areas of abrupt changes in terrain elevation; the feedforward neural network uses a 512-dimensional hidden layer and a ReLU activation function to perform a nonlinear transformation on the attention output, enhancing the expressive power of the features. The spatial feature vector output by the Transformer is input into the second LSTM layer, which is a two-layer hidden structure with 256 hidden units per layer. It fuses the dynamic evolution patterns over 50 time steps through a time-gating mechanism, such as the spatial trajectory changes of combat units moving from flat terrain to high ground, and outputs a spatiotemporal fusion feature vector.
[0096] The vector is compressed along the time dimension by a global average pooling layer, and the average of the features at 50 time steps is taken to reduce the dimensionality to a low-dimensional vector of 256 dimensions to retain key information.
[0097] Finally, a second Dropout layer randomly masks some neuron outputs at a fixed dropout rate of 0.4. For example, a random inactivation rate of 40% is maintained during the real-time inference stage to ensure consistency with the regularization effect during the training stage. The final output is a 256-dimensional real-time geographic location feature vector.
[0098] The data processing flow of the geographic location feature extraction module strictly reuses the module parameters from the training phase to achieve a deep transformation from raw spatial data to temporal-spatial fusion features.
[0099] The partial action state features and partial geographical location features are input into the feature fusion and classification module for intent classification based on feature fusion, and the current opponent intent probability distribution is output. The specific data processing process of this module is as follows:
[0100] First, the 256-dimensional real-time action state feature vector output by the action state feature extraction module and the 256-dimensional real-time geographic location feature vector output by the geographic location feature extraction module are input into the splicing layer. This splicing layer performs a head-to-tail connection operation on the feature dimensions to form a 512-dimensional comprehensive opponent situation information feature vector, thereby achieving deep fusion of action state temporal features and geographic location spatiotemporal features.
[0101] The 512-dimensional feature vector is then input into a fully connected layer containing 1024 neurons and employing the ReLU activation function. High-order abstract features are extracted through nonlinear transformation, such as identifying spatial correlation patterns between sudden changes in action state and terrain obstacles.
[0102] The transformed feature vector is input into the classification layer, which uses the Softmax activation function to calculate the probability distribution of the intent category: based on the weight parameters of the three types of intents (attack, defense, and flanking) learned during the training phase, the input features are linearly weighted and summed, and then exponential normalization is performed to output a three-dimensional probability vector.
[0103] Finally, the output layer generates the probability distribution of the current adversary's intentions, such as an offensive intention probability of 0.65, a defensive intention probability of 0.30, and a flanking intention probability of 0.05. This distribution reflects the model's quantitative prediction of the adversary's decision-making intentions in real-time situations.
[0104] The data processing flow of the feature fusion and classification module strictly reuses the module structure and parameters from the training phase to ensure consistency between the real-time inference and training feature spaces.
[0105] This embodiment achieves the following technical effects: significantly improves the accuracy and robustness of adversary intent recognition, effectively reduces information uncertainty in the game process; enhances the adaptability of the adversary intent recognition model to local observation information and improves its generalization performance in complex adversarial environments; optimizes the integrity of situational features through deep fusion of multimodal features and reduces recognition bias; provides reliable decision support for resource allocation and strategy formulation, and comprehensively improves the combat effectiveness and battlefield survivability of our own side.
[0106] The above description is merely a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.
Claims
1. A method for opponent intent pattern recognition based on multimodal feature fusion, characterized in that, include: Historical game data is retrieved from the adversarial simulation system and segmented into time slices to obtain multiple opponent time slices, in which the opponent time slices are marked with preset intention labels. Perform opponent situation information feature extraction on the multiple opponent time-series slices to generate multiple sample multimodal situation information features; The multimodal situational information features of the multiple samples are used as training data, and the corresponding multiple preset intent labels are used as supervision signals to perform iterative optimization training of the opponent intent recognition model until the model output meets the preset convergence condition. During real-time game play, the game information observed locally within a preset time window is input into the opponent's intent recognition model for intent reasoning, and the current opponent's intent probability distribution is output.
2. The adversary intent pattern recognition method based on multimodal feature fusion as described in claim 1, characterized in that, Historical game data was retrieved from the adversarial simulation system and segmented into time slices to obtain multiple opponent time-series slices, including: The opponent's situation information is extracted from the global confrontation record of game P in the historical game data to obtain the time series set of the opponent's situation in game P. The time sequence set of the opponent situation in Bureau P is divided into time slices using a preset time window length to obtain multiple opponent time slices. Each opponent time slice includes multiple time sequence action status data and multiple time sequence geographical location data of multiple opponent combat units.
3. The adversary intent pattern recognition method based on multimodal feature fusion as described in claim 2, characterized in that, Also includes: A pre-built action state feature extraction module is provided, wherein the action state feature extraction module includes a cascaded first input layer, a Masking layer, a first LSTM layer and a first Dropout layer, wherein the first LSTM layer is a 2-layer hidden structure with a hidden layer dimension of 256 and the dropout rate of the first Dropout layer is 0.
4. A pre-built geographic location feature extraction module is provided, wherein the action state feature extraction module includes a cascaded second input layer, a flattening layer, a sine-cosine position encoding layer, a Transformer encoder layer, a second LSTM layer, a global average pooling layer, and a second Dropout layer. The Transformer encoder layer is set to 3 layers, the number of multi-head attention heads is 8, the hidden layer dimension of the feedforward neural network is 512, the activation function is ReLU, the second LSTM layer is a 2-layer hidden structure with a hidden layer dimension of 256, and the dropout rate of the second Dropout layer is 0.
4. A pre-built feature fusion and classification module is provided, wherein the feature fusion and classification module includes a concatenation layer, a fully connected layer with ReLU activation function, a classification layer, and an output layer with Softmax activation function; The first Dropout layer of the action state feature extraction module and the second Dropout layer of the geographic location feature extraction module are connected to the splicing layer of the feature fusion and classification module to complete the framework construction of the opponent intent recognition model.
4. The adversary intent pattern recognition method based on multimodal feature fusion as described in claim 3, characterized in that, Using the multimodal situational information features of the multiple samples as training data and the corresponding multiple preset intent labels as supervision signals, the parameters of the adversary intent recognition model are iteratively optimized and trained until the model output satisfies the preset convergence conditions, including: The training dataset, consisting of the multimodal situational information features of the multiple samples and the multiple preset intent labels, is separated into a training set, a validation set, and a test set. The Adam optimizer is used, and optimizer parameters are set, including the learning rate and batch size. Using the optimizer parameters as constraints, the opponent intent recognition model is iterated once by retrieving training data from the training set, and then the intent recognition accuracy on the validation set is calculated. When the improvement in the accuracy of the adversary intent recognition model is less than 0.1% after M consecutive iterations, it is determined that the output of the adversary intent recognition model meets the preset convergence condition. The accuracy of the adversary intent recognition model was verified using the test set.
5. The adversary intent pattern recognition method based on multimodal feature fusion as described in claim 3, characterized in that, During real-time game play, locally observed game information within a preset time window is input into the opponent's intent recognition model for intent inference, and the current opponent's intent probability distribution is output, including: The local observation game information is segmented using the time window length to obtain multiple real-time action status data and multiple real-time geographical location data of the multiple opposing combat units; The multiple real-time action status data are integrated using the time window length to obtain real-time action status information; The real-time action state information is input into the action state feature extraction module to extract action state temporal features and output partial action state features. The multiple real-time geographic location data are integrated using the specified time window length to obtain real-time geographic location information; The real-time geographic location information is input into the geographic location feature extraction module for spatiotemporal fusion feature extraction, and a partial geographic location feature is output. The partial action state features and partial geographical location features are input into the feature fusion and classification module to perform intent classification under feature fusion, and the current opponent intent probability distribution is output.
6. The adversary intent pattern recognition method based on multimodal feature fusion as described in claim 5, characterized in that, The real-time action state information is input into the action state feature extraction module for action state temporal feature extraction, and partial action state features are output, including: After the real-time action state information is input into the Masking layer through the first input layer to filter out invalid data, the action state temporal feature extraction based on the long short-term memory network is performed through the first LSTM layer to obtain a partial action state feature vector. The partial action state feature vector is subjected to random deactivation regularization with a dropout rate of 0.4 through the first Dropout layer, and the partial action state feature is output.
7. The adversary intent pattern recognition method based on multimodal feature fusion as described in claim 5, characterized in that, The real-time geographic location information is input into the geographic location feature extraction module for spatiotemporal fusion feature extraction, and partial geographic location features are output, including: After the real-time geographic location information is input into the flattening layer through the second input layer for two-dimensional spatial feature serialization, the location encoding is added through the sine-cosine location encoding layer to obtain location-encoded sequence data. The spatial global correlation modeling of the position-encoded sequence data is performed on the Transformer encoder layer based on a multi-head attention mechanism, and a spatial feature vector is output. The spatial feature vector is input into the second LSTM layer for spatiotemporal fusion feature extraction to obtain the spatiotemporal feature vector; After the spatiotemporal feature vector is compressed using the global average pooling layer, it is then subjected to random deactivation regularization with a dropout rate of 0.4 by the second Dropout layer to output the partial geographic location feature.
8. The adversary intent pattern recognition method based on multimodal feature fusion as described in claim 5, characterized in that, The partial action state features and partial geographical location features are input into the feature fusion and classification module for intent classification under feature fusion, and the current opponent intent probability distribution is output, including: The partial action state features and partial geographical location features are input into the splicing layer for feature splicing to obtain the opponent's situation information feature vector; After the opponent's situation information feature vector is nonlinearly transformed by the fully connected layer, the intent probability distribution is calculated by the Softmax activation function of the classification layer, and the current opponent intent probability distribution is output by the output layer.
9. The adversary intent pattern recognition method based on multimodal feature fusion as described in claim 2, characterized in that, Adversary situation information features are extracted from the multiple adversary time-series slices to generate multiple sample multimodal situation information features, including: Bimodal feature extraction is performed on the time slices of the multiple opponents to obtain multiple sample action state features and multiple sample geographical location features; By integrating the action state features and geographical location features of the multiple samples, multimodal situational information features of the multiple samples are generated.
10. The adversary intent pattern recognition method based on multimodal feature fusion as described in claim 9, characterized in that, Bimodal feature extraction is performed on the time slices of the multiple opponents to obtain multiple sample action state features and multiple sample geographical location features, including: After integrating multiple first action state data of multiple adversary combat units in the first adversary time slice using the time window length to obtain first action state information, the first sample action state features are extracted. After integrating multiple first geographical location data of the multiple adversary combat units in the first adversary time slice using the time window length to obtain the first geographical location information, the geographical location features of the first sample are extracted. Similarly, bimodal feature extraction is performed on the time slices of the multiple opponents to obtain the action state features and geographical location features of the multiple samples.