Unmanned aerial vehicle defense decision support system and method based on machine learning
Through improved convolution feature extraction and timing adaptive recursive feature extraction algorithms, combined with threat detection model, the limitations of modal interaction information being ignored and single modal data processing in traditional methods are solved, and efficient threat detection and defense strategy generation is achieved.
Patent Information
- Application Number
- CN202510123888.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-01-26
- Publication Date
- 2025-05-16
- Estimated Expiration
- Not applicable · inactive patent
AI Technical Summary
Traditional multimodal data processing methods ignore the interactive information between modes and cannot fully utilize the complementarity between modes. In addition, traditional threat detection relies on single mode data, cannot comprehensively consider the interaction between modes, and is difficult to deal with complex modes and environmental changes.
The improved convolution feature extraction algorithm and timing adaptive recursive feature extraction algorithm are used to extract convolutional features and recursive features of each modality through convolution operations and recursive neural networks, and weighted fusion is performed to generate derivative features. The threat detection model is then used to map the derivative features to the threat probability space to generate the best defense strategy.
It improves the quality and performance of feature extraction, enhances the ability to identify complex patterns, improves the accuracy and robustness of threat detection, and achieves a balance between threat probability and defense cost.
Smart Images
Figure CN120010508A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of machine learning, and in particular to a drone defense decision support system and method based on machine learning. Background Art
[0002] As a low-cost, high-efficiency, highly maneuverable and flexible aerial platform, drones have been widely used in intelligence reconnaissance, target strikes, logistics and transportation, agricultural monitoring and other fields. However, the popularity of drones has also brought new security risks, especially threats to key infrastructure, military facilities, important political activities and other targets. The risk of hostile forces conducting malicious attacks or espionage through drones has greatly increased. Therefore, how to effectively defend against drone attacks has become an important issue in the field of modern security protection.
[0003] In addition, drone defense not only needs to deal with the direct threats posed by drones, but also needs to effectively respond to the diverse tactics, complex environmental changes, and evolving attack patterns of drones. Against this background, a drone defense decision support system based on machine learning has emerged. It can analyze the behavior patterns of drones, combine real-time sensor data and historical attack data, and make intelligent defense decisions, thereby maximizing defense efficiency and reducing errors and lags in human decision-making. Summary of the invention
[0004] The present invention provides a drone defense decision support system and method based on machine learning to solve the problems that in traditional multimodal data processing methods, only the features of a single modality are usually processed, and the interactive information between the modalities is ignored, resulting in the inability to fully utilize the complementarity between the modalities, limiting the quality and performance of feature extraction; the correlation between the historical information and the current moment in the time series data is not evenly distributed, and the traditional recursive neural network processing method ignores the differences in information at different moments; and traditional threat detection relies on single modality data, cannot comprehensively consider the interaction between the modalities, and is difficult to cope with complex patterns and environmental changes.
[0005] The UAV defense decision support system and method based on machine learning of the present invention specifically include the following technical solutions:
[0006] The drone defense decision support method based on machine learning includes the following steps:
[0007] S1: Capture and preprocess the multimodal data of the drone and the interference to obtain the preprocessed multimodal data; based on the preprocessed multimodal data, use the improved convolution feature extraction algorithm to extract the convolution features of each mode;
[0008] S2: Based on the convolutional features of each modality, the recursive features of each modality are calculated using the time-series adaptive recursive feature extraction algorithm;
[0009] S3: Generate derivative features of each modality by weighted fusion of convolutional features and recursive features of each modality; Based on the derivative features of each modality, use the threat detection model to obtain the threat probability of each modality;
[0010] S4: Based on the threat probability of each mode, use the defense strategy generation algorithm to generate the best defense strategy; implement defense operations according to the best defense strategy.
[0011] Preferably, the S1 specifically includes:
[0012] The improved convolutional feature extraction algorithm performs weighted processing on the preprocessed multimodal data through convolution operations, further introduces the interaction features between modalities, and uses the interaction weights to adjust the interaction between different modalities. At the same time, a bias term is introduced, and finally a nonlinear transformation is performed through the LeakyReLU activation function.
[0013] Preferably, the S2 specifically includes:
[0014] The recursive features are calculated using the time-adaptive recursive feature extraction algorithm through the recursive hidden state of the previous moment and the convolutional features of the current moment.
[0015] Preferably, the S2 specifically includes:
[0016] In the implementation process of the time-series adaptive recursive feature extraction algorithm, the time-series weighting mechanism and the local time window mechanism are introduced. The hidden state is updated based on the historical hidden state, the convolutional features of the current moment and the time weighting coefficient, and nonlinear mapping is performed through a nonlinear activation function to obtain a new hidden state.
[0017] Preferably, the S2 specifically includes:
[0018] The calculation formula of hidden state is:
[0019]
[0020] Among them, h t，i is the hidden state of the i-th mode at time t, representing the recursive feature of the i-th mode at time t; tanh represents the hyperbolic tangent function, which is used as a nonlinear activation function; is the summation operation from time t-Δt to t-1, which means weighted summation of all historical hidden states within the Δt time steps before time t; Δt represents the size of the historical data window; the historical data is the hidden state of the i-th mode from time t-Δt to t-1; α k，iIt represents the time weighting coefficient of the i-th mode at time k, and the calculation formula is: ζ represents the adjustment parameter; h k，i represents the hidden state of the i-th mode at time k; β t，i Represents the influence weight of different modal convolution features at time t; y t，i represents the convolution feature of the i-th mode at time t; Represents the bias term of the recurrent neural network.
[0021] Preferably, the S3 specifically includes:
[0022] Using the threat detection model, the derived features of each modality are mapped to the threat probability space to obtain the threat probability of each modality.
[0023] Preferably, the S3 specifically includes:
[0024] The threat detection model establishes a nonlinear relationship between the derived features of each modality and the threat probability of each modality by combining exponential transformation and hyperbolic tangent function; and introduces the Sigmoid function for normalization in the mapping process.
[0025] Preferably, the S4 specifically includes:
[0026] The defense strategy generation algorithm calculates the total threat probability, evaluates the overall threat level, and weighs the threat probability and defense cost, so that the threat is effectively suppressed while the cost of the defense strategy is minimized.
[0027] Preferably, the S4 specifically includes:
[0028] The formula for selecting the best defense strategy is:
[0029]
[0030] Among them, d t represents the best defense strategy at time t; Denotes the choice of defense strategy so that the objective function The value of P is maximized; t，i represents the threat probability of the i-th mode at time t; λ1 represents the threat weight coefficient; N represents the total number of modes; is the total threat probability; λ2 represents the weight coefficient of the defense cost; C(d) represents the cost function of the defense strategy.
[0031] The drone defense decision support system based on machine learning includes the following parts:
[0032] Data acquisition module, data preprocessing module, convolution feature extraction module, recursive feature extraction module, derivative feature generation module, threat detection module, defense strategy generation module;
[0033] Data acquisition module: captures multimodal data of drones and interference objects through different sensors, and outputs the multimodal data to the data preprocessing module;
[0034] Data preprocessing module: preprocesses the multimodal data to obtain preprocessed multimodal data, and outputs the preprocessed multimodal data to the convolution feature extraction module;
[0035] Convolution feature extraction module: Use the improved convolution feature extraction algorithm to perform convolution operation on the preprocessed multimodal data, extract the convolution features of each modality, and output the convolution features of each modality to the recursive feature extraction module and the derived feature generation module;
[0036] Recursive feature extraction module: Based on the convolution features of each mode, the time-series adaptive recursive feature extraction algorithm is used to calculate the recursive features of each mode, and the recursive features of each mode are output to the derived feature generation module;
[0037] Derived feature generation module: performs weighted fusion on the convolutional features and recursive features of each modality to generate derived features of each modality, and outputs the derived features of each modality to the threat detection module;
[0038] Threat detection module: Use the threat detection model to map the derived features of each modality to the threat probability space, obtain the threat probability of each modality, and output the threat probability of each modality to the defense strategy generation module;
[0039] Defense strategy generation module: Based on the threat probability of each mode, the defense strategy generation algorithm is used to dynamically generate the best defense strategy; according to the best defense strategy, defense operations are implemented.
[0040] The beneficial effects of the technical solution of the present invention are:
[0041] 1. The improved convolution feature extraction algorithm captures the complex correlation between different modalities by introducing the interaction features between modalities based on the traditional convolution operation. The LeakyReLU activation function is used to avoid the "dead neuron" problem in ReLU, improve the nonlinear expression ability, enhance the expression ability of features, perform weighted summation of multimodal data through convolution weights, and introduce interaction weights to calculate the interaction features between modalities, further improving the adaptability of heterogeneous data sources.
[0042] 2. The time-adaptive recursive feature extraction algorithm obtains the recursive features by calculating the recursive hidden state of the previous moment and the convolution feature of the current moment, and dynamically adjusts the influence of the historical hidden state by introducing a time-series weighting mechanism. At the same time, in order to avoid the redundancy and interference caused by global information transmission, a local time window mechanism is introduced to improve the efficiency and accuracy of information extraction.
[0043] 3. Efficiently integrate convolutional features with recursive features to improve the accuracy and robustness of threat detection.
[0044] 4. The defense strategy generation algorithm evaluates the current threat level based on the sum of threat probabilities, and generates the optimal defense strategy by weighing the threat probability and defense cost, achieving a balance between threat probability and defense cost. BRIEF DESCRIPTION OF THE DRAWINGS
[0045] Figure 1 This is a structural diagram of the UAV defense decision support system based on machine learning according to the present invention;
[0046] Figure 2 This is a flow chart of the UAV defense decision support method based on machine learning described in the present invention. DETAILED DESCRIPTION
[0047] In order to further explain the technical means and effects adopted by the present invention to achieve the predetermined invention purpose, the technical scheme in the embodiment of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiment of the present invention. Obviously, the described embodiment is only a part of the embodiment of the present invention, not all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.
[0048] Unless defined otherwise, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention belongs.
[0049] The specific scheme of the UAV defense decision support system and method based on machine learning provided by the present invention is described in detail below with reference to the accompanying drawings.
[0050] See attached Figure 1 , which shows a structure diagram of a drone defense decision support system based on machine learning provided by an embodiment of the present invention, the system includes the following parts:
[0051] Data acquisition module, data preprocessing module, convolution feature extraction module, recursive feature extraction module, derivative feature generation module, threat detection module, defense strategy generation module;
[0052] Data acquisition module: captures multimodal data of drones and interference objects through different sensors, and outputs the multimodal data to the data preprocessing module;
[0053] Data preprocessing module: preprocesses the multimodal data from the data acquisition module to obtain preprocessed multimodal data, and outputs the preprocessed multimodal data to the convolution feature extraction module;
[0054] Convolution feature extraction module: uses an improved convolution feature extraction algorithm to perform convolution operation on the preprocessed multimodal data from the data preprocessing module, extracts the convolution features of each modality, and outputs the convolution features of each modality to the recursive feature extraction module and the derivative feature generation module;
[0055] Recursive feature extraction module: Based on the convolution features of each mode of the convolution feature extraction module, the recursive features of each mode are calculated using the time-series adaptive recursive feature extraction algorithm, and the recursive features of each mode are output to the derived feature generation module;
[0056] Derived feature generation module: performs weighted fusion on the convolutional features of each modality from the convolutional feature extraction module and the recursive features of each modality from the recursive feature extraction module to generate derived features of each modality, and outputs the derived features of each modality to the threat detection module;
[0057] Threat detection module: Use the threat detection model to map the derived features of each modality of the derived feature generation module to the threat probability space, obtain the threat probability of each modality, and output the threat probability of each modality to the defense strategy generation module;
[0058] Defense strategy generation module: Based on the threat probability of each mode of the threat detection module, the defense strategy generation algorithm is used to dynamically generate the best defense strategy, and specific defense operations are implemented according to the best defense strategy.
[0059] See attached Figure 2 , which shows a flow chart of a drone defense decision support method based on machine learning provided by an embodiment of the present invention, the method comprising the following steps:
[0060] S1. Capture and preprocess the multimodal data of the UAV and the interference to obtain the preprocessed multimodal data; based on the preprocessed multimodal data, use the improved convolution feature extraction algorithm to extract the convolution features of each mode;
[0061] The multimodal data of the drone and the interference objects are captured by different sensors (such as radar, vision, infrared), such as flight parameters (such as speed, heading, altitude), environmental data (such as temperature, humidity, air pressure), and the multimodal data are preprocessed to obtain preprocessed multimodal data to ensure that different modal data are processed under a unified standard, laying the foundation for subsequent steps. The preprocessing methods include data synchronization, data cleaning and data standardization, which are technical means well known to those skilled in the art and will not be elaborated here.
[0062] An improved convolutional feature extraction algorithm is used to extract convolutional features from the preprocessed multimodal data through convolution operations, while introducing interactive information between modalities, thereby improving the ability to capture complex associations.
[0063] The core of the improved convolutional feature extraction algorithm is to capture the correlation between different modalities by introducing interactive features between modalities, and at the same time adopt the LeakyReLU activation function to increase the nonlinear expression ability, so that the improved convolutional feature extraction algorithm can effectively improve the quality and performance of feature extraction when processing multimodal data, especially when facing a variety of heterogeneous data sources, and can give full play to the complementarity of each modality.
[0064] The preprocessed multimodal data is weighted through a convolution operation. Specifically, for each modality, the improved convolution feature extraction algorithm uses a convolution weight to perform a weighted summation on the input preprocessed multimodal data.
[0065] Since only the feature information of a single modality is usually considered in traditional convolution operations, the interaction features between modalities are introduced to effectively capture the complex correlations between modalities by calculating the interactions between different modalities. Specifically, each pair of modal data is firstly subjected to element-level multiplication operations to calculate the interaction features between modalities, and then the interaction weights are further used to adjust the interactions between different modalities to accurately mine the connections between different modalities, thereby providing feature representations that are richer than single modal information.
[0066] Based on the weighted summation of the preprocessed multimodal data and the weighted summation of the interaction features between the modalities, a bias term is added, and then a nonlinear transformation is performed through the LeakyReLU activation function. The LeakyReLU activation function can effectively avoid the common "dead neuron" problem in the ReLU function, while maintaining certain nonlinear characteristics, enhancing the expression ability of the convolution feature, and avoiding the performance bottleneck caused by the overly linear algorithm structure;
[0067] The convolution feature extraction formula for each modality is as follows:
[0068]
[0069] Among them, y t，i Represents the convolution feature of the i-th mode at time t; LeakyReLU represents a nonlinear activation function, which is used to introduce nonlinear transformation and avoid the "dead neuron" problem caused by ReLU when it is negative, so that the neural network maintains a certain gradient flow during back propagation; N represents the total number of modes; represents the i-th modal data after preprocessing; represents the jth modal data after preprocessing; w i，jrepresents the convolution weight between the i-th mode and the j-th mode, which determines the combination mode and strength between different modal data and is used to control the contribution of different modal data in the convolution operation; a i，j The interaction weight represents the interaction information between the i-th mode and the j-th mode. Unlike the convolution weight, the interaction weight focuses on the interaction between different modes and determines the strength of the interaction information between the modes. ⊙ represents element-by-element multiplication. represents the element-by-element multiplication of the preprocessed i-th modal data and the preprocessed, ,th modal data at time t, which helps to capture the correlation between different modalities; b i Represents the bias term of the convolutional neural network, which is used to adjust the output of the convolution operation, allowing the convolution operation to maintain a certain output value without data input;
[0070] The improved convolutional feature extraction algorithm not only retains the advantages of traditional convolutional feature extraction, but also improves the ability to recognize complex patterns through the introduction of interactive features between modalities and nonlinear activation functions, especially when processing multimodal data.
[0071] S2, based on the convolutional features of each modality, the recursive features of each modality are calculated using a time-series adaptive recursive feature extraction algorithm;
[0072] The time-series adaptive recursive feature extraction algorithm is used to calculate the recursive features of each modality through the recursive hidden state of the previous moment and the convolution feature of the current moment;
[0073] In traditional recurrent neural networks, the update of hidden states is usually performed through a weighted summation method, in which the influence of historical hidden states is treated equally. However, the correlation between historical hidden states and the current moment in time series data is not evenly distributed, so a time series weighting mechanism is introduced to dynamically adjust the influence of historical hidden states. Specifically, the time difference between the historical moment and the current moment is calculated, and combined with the adjustment parameters, the influence of historical hidden states can be adaptively changed according to time, thereby effectively highlighting the influence of the current moment; at the same time, in order to reduce the redundancy and interference caused by global information transmission, a local time window mechanism is introduced. In each time step, only the historical hidden states of the previous several time points are considered, instead of transferring the historical hidden states of the entire time series to the current moment. This can avoid transferring irrelevant noise brought by distant time step data to the current moment without losing time series information, thereby ensuring the efficiency and accuracy of the recursive feature extraction process;
[0074] In each time step, the time-series adaptive recursive feature extraction algorithm updates the hidden state based on the historical hidden state, the convolutional features at the current moment, and the calculated time weighting coefficient, and obtains a new hidden state through nonlinear mapping through a nonlinear activation function such as the hyperbolic tangent function;
[0075] The calculation formula of hidden state is:
[0076]
[0077] Among them, h t，i represents the hidden state of the i-th mode at time t, i.e., the recursive feature; tanh represents the hyperbolic tangent function, which, as a nonlinear activation function, provides a nonlinear mapping that maps the input to the interval [-1, 1]; represents the summation operation from time t-Δt to t-1, that is, the weighted summation of all historical hidden states within the Δt time steps before time t; Δt represents the size of the historical data window, which can be set according to the specific implementation scenario and is not limited here; the historical data is the hidden state of the i-th mode from time t-Δt to t-1; α k，i It represents the time weighting coefficient of the i-th mode at time k. It dynamically adjusts the influence of each historical data according to the time difference between the historical moment and the current moment. The calculation formula is: ζ represents the adjustment parameter, which is used to adaptively control the influence range of historical data at each time step on the current moment; h k，i represents the hidden state of the i-th mode at time k; β t，i Represents the influence weight of different modal convolution features at time t; y t，i represents the convolution feature of the i-th mode at time t; Represents the bias term of the recurrent neural network, which is used to adjust the linear combination in the hidden state calculation and is flexible.
[0078] S3, by weighted fusion of the convolutional features and recursive features of each modality, the derivative features of each modality are generated; based on the derivative features of each modality, the threat probability of each modality is obtained using the threat detection model;
[0079] By weighted fusion of the convolutional features and recursive features of each modality, the derived features of each modality are generated. The calculation formula is as follows:
[0080]
[0081] in, represents the derivative features of the ith mode at time t; y t，i represents the convolution feature of the i-th mode at time t; h t，i represents the hidden state of the i-th mode at time t, i.e., the recursive feature; ⊙ represents element-wise multiplication; γi Represents the weighting coefficient, which is used to adjust the interaction sensitivity between convolutional features and recursive features;
[0082] By weighted fusion of convolutional features and recursive features, the accuracy and robustness of overall threat detection can be improved.
[0083] Use the threat detection model to map the derived features of each modality to the threat probability space to obtain the threat probability of each modality;
[0084] The threat detection model establishes a complex nonlinear relationship between the derived features of each modality and the actual threat probability of each modality by combining exponential transformation and hyperbolic tangent function;
[0085] To ensure that the threat detection model outputs a valid probability value, the entire mapping process is normalized by a Sigmoid function. The Sigmoid function compresses all outputs between 0 and 1 to ensure that the output can be interpreted as the probability of a threat occurring.
[0086] The calculation formula of threat probability of each mode is:
[0087]
[0088] Among them, P t，i represents the threat probability of the i-th mode at time t; Sigmoid represents the activation function, which is used to compress the output result to between 0 and 1 to ensure a valid probability value; δ represents the mapping weight coefficient, which is used to control the contribution of the derived feature to the threat probability; c represents the exponential coefficient of the nonlinear transformation, which is used to control the degree of exponential transformation of the derived feature; tanh represents the hyperbolic tangent function, which is used to provide nonlinear transformation of the derived feature, which can strengthen the influence of extreme values and help the threat detection model learn deep nonlinear relationships; Represents the bias term, which is used to adjust the output of the threat detection model.
[0089] S4. Based on the threat probability of each mode, use the defense strategy generation algorithm to generate the best defense strategy; implement defense operations according to the best defense strategy;
[0090] Based on the threat probability of each modality, the defense strategy generation algorithm is used to dynamically generate the best defense strategy to achieve a balance between threat and cost;
[0091] The defense strategy generation algorithm calculates the total threat probability, evaluates the overall threat level, and weighs the threat probability and defense cost, so that the threat is effectively suppressed while the cost of the defense strategy is minimized;
[0092] The formula for selecting the best defense strategy is:
[0093]
[0094] Among them, d t represents the best defense strategy at time t; Denotes the choice of defense strategy so that the objective function The value of P is maximized; t，i represents the threat probability of the i-th mode at time t; λ1 represents the threat weight coefficient, which is used to control the total threat probability Impact on defense strategies; is the total threat probability, which represents the sum of the threat probabilities calculated for all modes and is used to evaluate the overall threat level; λ2 represents the weight coefficient of the defense cost, which is used to control the importance of the defense cost in the defense strategy; C(d) represents the cost function of the defense strategy, which reflects the resource consumption required for selecting the defense strategy, such as materials and economy, ensuring that the defense strategy not only takes into account the threat defense effect, but also the cost of implementing the defense strategy. The calculation formula is:
[0095]
[0096] Among them, R u (d) represents the consumption of the d-th defense strategy on the u-th resource, such as energy and time; U represents the number of resource types; E v (d) represents the vth economic cost of the dth defense strategy, such as equipment usage fees; V represents the number of economic cost items.
[0097] Based on the best defense strategy, implement specific defense operations, such as:
[0098] Electromagnetic interference: Interfere with the communication and navigation systems of enemy drones by emitting electromagnetic waves, causing them to lose control or deviate from their targets;
[0099] Flight path adjustment: adjust the flight path of your own or friendly drones to avoid the attack range or threat area of enemy drones;
[0100] Target interception: Use interception weapons (such as missiles, lasers) to directly attack enemy drones and eliminate the threat.
[0101] In summary, the UAV defense decision support system and method based on machine learning has been completed.
[0102] The order of the embodiments of the invention is for description only and does not represent the advantages and disadvantages of the embodiments. The processes depicted in the drawings do not necessarily require the specific order or continuous order shown to achieve the desired results. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.
[0103] The various embodiments in this specification are described in a progressive manner, and the same or similar parts between the various embodiments can be referenced to each other, and each embodiment focuses on the differences from other embodiments.
[0104] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that the technical solutions described in the above embodiments may still be modified, or some of the technical features may be replaced by equivalents. Such modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention, and should be included in the protection scope of the present invention.
Claims
1. A drone defense decision support method based on machine learning, characterized in that: The following steps are involved: S1: Capture and preprocess the multimodal data of the UAV and the interference objects to obtain the preprocessed multimodal data; Based on the preprocessed multimodal data, the convolution features of each modality are extracted using an improved convolution feature extraction algorithm; S2: Based on the convolutional features of each modality, the recursive features of each modality are calculated using the time-series adaptive recursive feature extraction algorithm; S3: Generate derivative features of each modality by weighted fusion of convolutional features and recursive features of each modality; Based on the derivative features of each modality, use the threat detection model to obtain the threat probability of each modality; S4: Based on the threat probability of each mode, use the defense strategy generation algorithm to generate the best defense strategy; implement defense operations according to the best defense strategy.
2. The method for drone defense decision support based on machine learning according to claim 1 is characterized in that: The S1 specifically includes: The improved convolutional feature extraction algorithm performs weighted processing on the preprocessed multimodal data through convolution operations, further introduces the interaction features between modalities, and uses the interaction weights to adjust the interaction between different modalities. At the same time, a bias term is introduced, and finally a nonlinear transformation is performed through the LeakyReLU activation function.
3. The method for drone defense decision support based on machine learning according to claim 1, characterized in that: The S2 specifically includes: The recursive features are calculated using the time-adaptive recursive feature extraction algorithm through the recursive hidden state of the previous moment and the convolutional features of the current moment.
4. The method for drone defense decision support based on machine learning according to claim 3 is characterized in that: The S2 specifically includes: In the implementation process of the time-series adaptive recursive feature extraction algorithm, the time-series weighting mechanism and the local time window mechanism are introduced. The hidden state is updated based on the historical hidden state, the convolutional features of the current moment and the time weighting coefficient, and nonlinear mapping is performed through a nonlinear activation function to obtain a new hidden state.
5. The method for drone defense decision support based on machine learning according to claim 4 is characterized in that: The S2 specifically includes: The calculation formula of hidden state is: Among them, h t，i is the hidden state of the i-th mode at time t, representing the recursive feature of the i-th mode at time t; tanh represents the hyperbolic tangent function, which is used as a nonlinear activation function; is the summation operation from time t-Δt to t-1, which means weighted summation of all historical hidden states within the Δt time steps before time t; Δt represents the size of the historical data window; the historical data is the hidden state of the i-th mode from time t-Δt to t-1; α k，i It represents the time weighting coefficient of the i-th mode at time k, and the calculation formula is: ζ represents the adjustment parameter; h k，i represents the hidden state of the i-th mode at time k; β t，i Represents the influence weight of different modal convolution features at time t; y t，i represents the convolution feature of the i-th mode at time t; Represents the bias term of the recurrent neural network.
6. The method for drone defense decision support based on machine learning according to claim 1, characterized in that: The S3 specifically includes: Using the threat detection model, the derived features of each modality are mapped to the threat probability space to obtain the threat probability of each modality.
7. The method for drone defense decision support based on machine learning according to claim 6, characterized in that: The S3 specifically includes: The threat detection model establishes a nonlinear relationship between the derived features of each modality and the threat probability of each modality by combining exponential transformation and hyperbolic tangent function; and introduces the Sigmoid function for normalization in the mapping process.
8. The method for drone defense decision support based on machine learning according to claim 1, characterized in that: The S4 specifically includes: The defense strategy generation algorithm calculates the total threat probability, evaluates the overall threat level, and weighs the threat probability and defense cost, so that the threat is effectively suppressed while the cost of the defense strategy is minimized.
9. The method for drone defense decision support based on machine learning according to claim 8, characterized in that: The S4 specifically includes: The formula for selecting the best defense strategy is: Among them, d t represents the best defense strategy at time t; Denotes the choice of defense strategy so that the objective function The value of P is maximized; t，i represents the threat probability of the i-th mode at time t; λ1 represents the threat weight coefficient; N represents the total number of modes; is the total threat probability; λ2 represents the weight coefficient of the defense cost; C(d) represents the cost function of the defense strategy.
10. A drone defense decision support system based on machine learning, applied to the drone defense decision support method based on machine learning as claimed in claim 1, characterized in that: Includes the following parts: Data acquisition module, data preprocessing module, convolution feature extraction module, recursive feature extraction module, derivative feature generation module, threat detection module, defense strategy generation module; Data acquisition module: captures multimodal data of drones and interference objects through different sensors, and outputs the multimodal data to the data preprocessing module; Data preprocessing module: preprocesses the multimodal data to obtain preprocessed multimodal data, and outputs the preprocessed multimodal data to the convolution feature extraction module; Convolution feature extraction module: Use the improved convolution feature extraction algorithm to perform convolution operation on the preprocessed multimodal data, extract the convolution features of each modality, and output the convolution features of each modality to the recursive feature extraction module and the derived feature generation module; Recursive feature extraction module: Based on the convolution features of each mode, the time-series adaptive recursive feature extraction algorithm is used to calculate the recursive features of each mode, and the recursive features of each mode are output to the derived feature generation module; Derived feature generation module: performs weighted fusion on the convolutional features and recursive features of each modality to generate derived features of each modality, and outputs the derived features of each modality to the threat detection module; Threat detection module: Use the threat detection model to map the derived features of each modality to the threat probability space, obtain the threat probability of each modality, and output the threat probability of each modality to the defense strategy generation module; Defense strategy generation module: Based on the threat probability of each mode, the defense strategy generation algorithm is used to dynamically generate the best defense strategy; according to the best defense strategy, defense operations are implemented.