Unmanned aerial vehicle multi-modal interference autonomous identification method based on hierarchical cognitive architecture

Abstract

This invention relates to the field of UAV communication technology, and in particular to an autonomous identification method for multimodal interference in UAVs based on a hierarchical cognitive architecture. By introducing homomorphic encryption and secure multi-party computation, it achieves encrypted domain processing and fusion of multimodal data throughout its entire lifecycle, avoiding the risk of leakage of original information during transmission and computation, and constructing a data security barrier in high-adversarial environments. Furthermore, at the cognitive layer, it integrates an attention mechanism and a causal reasoning engine, elevating multimodal features to a deep semantic understanding level of intent and evolutionary patterns, generating a spatiotemporal interference cognitive situation map, breaking through the limitation of traditional methods that can only identify interference types. Finally, the decision layer adopts hierarchical reinforcement learning to generate the optimal anti-interference strategy in real time based on the dynamic situation, and optimizes the perception and cognition modules through feedback of evaluation results, forming a closed-loop self-evolutionary mechanism.

Description

Technical Field

[0001] This invention relates to the field of unmanned aerial vehicle (UAV) communication technology, and in particular to an autonomous identification method for multimodal interference of UAVs based on a hierarchical cognitive architecture. Background Technology

[0002] With the widespread application of drone technology in numerous fields, the electromagnetic environment they face has evolved from a relatively simple and static scenario to an increasingly complex and dynamic battlefield. Various malicious interference signals, such as deceptive interference that can trick receivers, suppressive interference that can suppress and block communication links with high power, and targeted interference that targets specific frequency bands, pose a serious threat to the reliability of drone communication links, the accuracy of navigation and positioning, and the normal operation of mission payloads.

[0003] However, existing methods cannot achieve a secure closed-loop intelligent cognition and decision-making mechanism, making it difficult to optimize multimodal interference information through deep semantic understanding and dynamic adaptive anti-interference strategies while ensuring data security. Summary of the Invention

[0004] The purpose of this invention is to provide an autonomous identification method for multimodal interference of unmanned aerial vehicles based on a hierarchical cognitive architecture. This method solves the technical problems in the prior art, which cannot achieve a safe closed-loop intelligent cognitive and decision-making mechanism and is difficult to optimize multimodal interference information through deep semantic understanding and dynamic adaptive anti-interference strategies while ensuring data security.

[0005] To achieve the above objectives, this invention provides an autonomous identification method for multimodal interference of unmanned aerial vehicles (UAVs) based on a hierarchical cognitive architecture, comprising: Construct a hierarchical cognitive architecture, which includes a perception layer, a cognition layer, and a decision-making layer; Multimodal signals are acquired through encrypted transmission, and multimodal signal preprocessing and feature extraction are performed at the perception layer. By introducing a dynamic attention mechanism based on electromagnetic situational awareness, sensor data of different modalities are adaptively weighted and fused to generate a multi-dimensional feature vector set. At the cognitive layer, autonomous identification of interference patterns is performed. A multi-scale feature fusion network based on an attention mechanism is used to perform deep semantic encoding on the multi-dimensional feature vector set. Combined with the interference prior knowledge in the dynamic semantic knowledge graph, the interference type, intensity and intent are inferred through a causal reasoning engine to generate an interference cognitive situation map containing spatiotemporal evolution information. At the decision-making level, anti-interference strategies are dynamically generated. Based on the interference cognitive situation map, the optimal anti-interference decision sequence is generated through a hierarchical reinforcement learning-driven policy network. The effectiveness of the strategy execution is evaluated in real time, and the evaluation results are fed back to the perception and cognition layers to form a closed-loop feedback optimization mechanism. The specific method for obtaining multimodal signals through encrypted transmission is as follows: After acquiring multimodal sensor data, the perception layer uses a distributed encryption protocol to segment and encrypt the original data, generating encrypted data blocks with timestamps. The encrypted data blocks are then input into a feature extraction network based on homomorphic encryption to perform preliminary feature extraction without decryption, generating encrypted intermediate feature representations. These encrypted intermediate feature representations are then transmitted to the cognitive layer via a secure channel. The cognitive layer uses a secure multi-party computation protocol to collaboratively compute the encrypted intermediate feature representations of each modality, achieving cross-modal fusion and semantic understanding in the ciphertext domain.

[0006] Within the constructed hierarchical cognitive architecture, The perception layer is equipped with a multi-source heterogeneous sensor array for real-time acquisition of multimodal environmental data from the UAV; The cognitive layer deploys a dynamic semantic knowledge graph and a causal reasoning engine for deep semantic understanding and interference intent inference of multimodal data; The decision-making level establishes an adaptive adversarial strategy generation mechanism to dynamically generate and optimize anti-interference decisions.

[0007] The dynamic attention mechanism includes a spectrum situational awareness module and an adaptive weighted fusion module. The spectrum situation awareness module is used to analyze the full-band energy distribution of the radio frequency front end in real time, identify abnormal energy peaks and frequency band blocking regions, and generate an electromagnetic situation heat map. The adaptive weighted fusion module dynamically assigns attention weights to each modality of data based on the electromagnetic situation heatmap. When broadband blocking interference is detected in the satellite navigation signal frequency band, the initial weight of the navigation mode is reduced, while the weight of the inertial measurement unit and the visual image mode is increased. When tracking interference targeting the communication link is detected, the weights of frequency hopping communication mode and radar echo mode are increased.

[0008] Among them, the dynamic semantic knowledge graph of the cognitive layer is constructed and updated using graph neural networks, including the interference entity layer, the relation layer, and the spatiotemporal evolution layer; The interference entity layer defines the interference source ontology, interference style ontology, and platform characteristic ontology, with each node embedding a multi-dimensional attribute vector. The relation layer defines causal, temporal, and spatial relationships between entities, with causal relationship edges learned from historical data through Granger causality tests. The spatiotemporal evolution layer predicts the future evolution trend of knowledge graphs through long short-term memory networks, enabling forward-looking cognition of disturbance situations.

[0009] Among them, the causal reasoning engine adopts an architecture that combines structural causal modeling with counterfactual reasoning; The causal reasoning engine includes a causal discovery module for automatically learning the directed acyclic graph structure of the disturbance generation mechanism from a multi-dimensional feature vector set; The causal reasoning engine also includes a counterfactual prediction mechanism. Based on the generated directed acyclic graph structure, it simulates the system response under different anti-interference strategies through intervention operations, evaluates the effectiveness of potential intervention measures, and provides causal explanatory support for decision-making.

[0010] Among them, the multi-scale feature fusion network based on the attention mechanism includes a cross-modal feature alignment sub-network and an adaptive attention fusion sub-network; The cross-modal feature alignment sub-network aligns the distribution of features from different modalities to generate modality-independent features in a unified representation space. The cross-modal feature alignment sub-network uses homomorphic encryption technology to perform normalization and alignment operations on feature distribution in the ciphertext domain. An adaptive attention fusion subnetwork is designed with an intermodal attention gating mechanism to dynamically calculate the contribution weight of each modality feature to the current interference recognition task, thereby achieving adaptive selection and weighted fusion of features. The attention gating mechanism is based on a secure multi-party computation protocol and completes weight calculation and feature fusion without exposing the original features of each modality.

[0011] Among them, the hierarchical reinforcement learning-driven policy network of the decision layer includes a high-level policy network and a low-level policy network; Based on the interference cognitive situation map, the high-level strategy network outputs discrete anti-interference mode selection decisions, including spectrum avoidance mode, power adaptive mode, waveform agility mode or trajectory replanning mode. The low-level policy network continuously adjusts specific anti-jamming parameters, including frequency hopping rate, transmit power level, or beam pointing angle, for the selected high-level mode. The policy network introduces a curiosity-driven intrinsic reward mechanism, which estimates the uncertainty of state transitions by predicting dynamic models and provides additional rewards for exploring unknown disturbance patterns, thereby enhancing the adaptability to new disturbances.

[0012] This invention presents an autonomous identification method for multimodal interference in unmanned aerial vehicles (UAVs) based on a hierarchical cognitive architecture. By introducing homomorphic encryption and secure multi-party computation, it achieves encrypted domain processing and fusion of multimodal data throughout its entire lifecycle, avoiding the risk of leakage of original information during transmission and computation, and constructing a data security barrier in high-adversarial environments. Furthermore, at the cognitive layer, it integrates an attention mechanism and a causal reasoning engine, elevating multimodal features to a deep semantic understanding level of intent and evolutionary patterns, generating a spatiotemporal interference cognitive situation map, breaking through the limitation of traditional methods that can only identify interference types. Finally, the decision layer adopts hierarchical reinforcement learning to generate the optimal anti-interference strategy in real time based on the dynamic situation, and optimizes the perception and cognition modules through feedback of evaluation results, forming a closed-loop self-evolutionary mechanism. This approach solves the technical problem in existing technologies where intelligent cognitive and decision-making mechanisms cannot achieve a secure closed loop, making it difficult to optimize multimodal interference information for deep semantic understanding and dynamic adaptive anti-interference strategies while ensuring data security. Attached Figure Description

[0013] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below.

[0014] Figure 1 This is a flowchart of the autonomous identification method for multimodal interference of unmanned aerial vehicles based on a hierarchical cognitive architecture according to the present invention. Detailed Implementation

[0015] The embodiments of the present invention are described in detail below. Examples of the embodiments are shown in the accompanying drawings. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain the present invention, but should not be construed as limiting the present invention.

[0016] Please see Figure 1 , Figure 1 This is a flowchart of the autonomous identification method for multimodal interference of unmanned aerial vehicles based on a hierarchical cognitive architecture according to the present invention.

[0017] This invention provides an autonomous identification method for multimodal interference in unmanned aerial vehicles (UAVs) based on a hierarchical cognitive architecture, comprising: S1. Construct a hierarchical cognitive architecture, which includes a perception layer, a cognitive layer, and a decision-making layer. In this specific implementation, the constructed hierarchical cognitive architecture, The perception layer is equipped with a multi-source heterogeneous sensor array for real-time acquisition of multimodal environmental data from the UAV; The cognitive layer deploys a dynamic semantic knowledge graph and a causal reasoning engine for deep semantic understanding and interference intent inference of multimodal data; The decision-making level establishes an adaptive adversarial strategy generation mechanism to dynamically generate and optimize anti-interference decisions.

[0018] S2. Multimodal signals are acquired through encrypted transmission, and multimodal signal preprocessing and feature extraction are performed at the perception layer. By introducing a dynamic attention mechanism based on electromagnetic situational awareness, sensor data of different modalities are adaptively weighted and fused to generate a multi-dimensional feature vector set. In this specific implementation, the dynamic attention mechanism includes a spectrum situational awareness module and an adaptive weighted fusion module; The spectrum situation awareness module is used to analyze the full-band energy distribution of the radio frequency front end in real time, identify abnormal energy peaks and frequency band blocking regions, and generate an electromagnetic situation heat map. The adaptive weighted fusion module dynamically assigns attention weights to each modality of data based on the electromagnetic situation heatmap. When broadband blocking interference is detected in the satellite navigation signal frequency band, the initial weight of the navigation mode is reduced, while the weight of the inertial measurement unit and the visual image mode is increased. When tracking interference targeting the communication link is detected, the weights of frequency hopping communication mode and radar echo mode are increased.

[0019] S3. Perform autonomous identification of interference patterns at the cognitive layer. Use a multi-scale feature fusion network based on attention mechanism to perform deep semantic encoding on the multi-dimensional feature vector set. Combine the interference prior knowledge in the dynamic semantic knowledge graph and infer the interference type, intensity and intent through the causal reasoning engine to generate an interference cognitive situation map containing spatiotemporal evolution information. In this specific implementation, the dynamic semantic knowledge graph of the cognitive layer is constructed and updated using a graph neural network, which includes an interfering entity layer, a relation layer, and a spatiotemporal evolution layer. The interference entity layer defines the interference source ontology, interference style ontology, and platform characteristic ontology, with each node embedding a multi-dimensional attribute vector. The relation layer defines causal, temporal, and spatial relationships between entities, with causal relationship edges learned from historical data through Granger causality tests. The spatiotemporal evolution layer predicts the future evolution trend of knowledge graphs through long short-term memory networks, enabling forward-looking cognition of disturbance situations.

[0020] The causal reasoning engine adopts an architecture that combines structural causal models with counterfactual reasoning; The causal reasoning engine includes a causal discovery module for automatically learning the directed acyclic graph structure of the disturbance generation mechanism from a multi-dimensional feature vector set; The causal reasoning engine also includes a counterfactual prediction mechanism. Based on the generated directed acyclic graph structure, it simulates the system response under different anti-interference strategies through intervention operations, evaluates the effectiveness of potential intervention measures, and provides causal explanatory support for decision-making.

[0021] The attention-based multi-scale feature fusion network includes a cross-modal feature alignment sub-network and an adaptive attention fusion sub-network; The cross-modal feature alignment sub-network aligns the distribution of features from different modalities to generate modality-independent features in a unified representation space. The cross-modal feature alignment sub-network uses homomorphic encryption technology to perform normalization and alignment operations on feature distribution in the ciphertext domain. An adaptive attention fusion subnetwork is designed with an intermodal attention gating mechanism to dynamically calculate the contribution weight of each modality feature to the current interference recognition task, thereby achieving adaptive selection and weighted fusion of features. The attention gating mechanism is based on a secure multi-party computation protocol and completes weight calculation and feature fusion without exposing the original features of each modality.

[0022] S4. Dynamically generate anti-interference strategies at the decision-making level. Based on the interference cognitive situation map, generate the optimal anti-interference decision sequence through a hierarchical reinforcement learning-driven policy network, evaluate the strategy execution effect in real time, and feed the evaluation results back to the perception layer and cognition layer to form a closed-loop feedback optimization mechanism. In this specific implementation, the hierarchical reinforcement learning-driven policy network of the decision layer includes a high-level policy network and a low-level policy network. Based on the interference cognitive situation map, the high-level strategy network outputs discrete anti-interference mode selection decisions, including spectrum avoidance mode, power adaptive mode, waveform agility mode or trajectory replanning mode. The low-level policy network continuously adjusts specific anti-jamming parameters, including frequency hopping rate, transmit power level, or beam pointing angle, for the selected high-level mode. The policy network introduces a curiosity-driven intrinsic reward mechanism, which estimates the uncertainty of state transitions by predicting dynamic models and provides additional rewards for exploring unknown disturbance patterns, thereby enhancing the adaptability to new disturbances.

[0023] The specific method for obtaining multimodal signals through encrypted transmission is as follows: After acquiring multimodal sensor data, the perception layer uses a distributed encryption protocol to segment and encrypt the original data, generating encrypted data blocks with timestamps. The encrypted data blocks are then input into a feature extraction network based on homomorphic encryption to perform preliminary feature extraction without decryption, generating encrypted intermediate feature representations. These encrypted intermediate feature representations are then transmitted to the cognitive layer via a secure channel. The cognitive layer uses a secure multi-party computation protocol to collaboratively compute the encrypted intermediate feature representations of each modality, achieving cross-modal fusion and semantic understanding in the ciphertext domain.

[0024] The UAV multimodal interference autonomous identification method based on hierarchical cognitive architecture also includes an online learning mechanism, which uses real-time collected data and identification results to update the dynamic semantic knowledge graph and anti-interference strategy library; When the matching degree between the generated interference cognitive situation map and all known interference templates in the knowledge base is lower than the preset threshold, it is determined to be a new type of unknown interference, triggering an online update of the knowledge base. The cognitive situation map corresponding to the new interference is then abstracted and stored as a new cognitive template in the knowledge base, and the state space of the hierarchical reinforcement learning network is updated synchronously.

[0025] When performing multimodal signal preprocessing and feature extraction at the perception layer, the process also includes: cleaning noisy signals using a signal denoising module based on a deep residual network, and simultaneously using a missing data completion technique based on a generative adversarial network to reconstruct data packets lost due to interference, ensuring the integrity and robustness of the input features.

[0026] Furthermore, it also includes setting up an interference trend prediction mechanism at the cognitive layer: the interference cognitive situation map containing spatiotemporal evolution information is input into a spatiotemporal sequence prediction model based on a convolutional long short-term memory network, and the interference situation evolution trajectory within a preset time domain is output, providing forward-looking decision support for the decision-making layer.

[0027] This embodiment employs a hierarchical cognitive architecture-based autonomous identification method for UAV multimodal interference. By introducing homomorphic encryption and secure multi-party computation, it achieves encrypted domain processing and fusion of multimodal data throughout its entire lifecycle, avoiding the risk of leakage of original information during transmission and computation, and constructing a data security barrier in high-adversarial environments. Furthermore, at the cognitive layer, it integrates an attention mechanism and a causal reasoning engine, elevating multimodal features to a deep semantic understanding level of intent and evolutionary patterns, generating a spatiotemporal interference cognitive situation map, breaking through the limitation of traditional methods that can only identify interference types. Finally, the decision layer adopts hierarchical reinforcement learning to generate the optimal anti-interference strategy in real time based on the dynamic situation, and optimizes the perception and cognition modules through evaluation results, forming a closed-loop self-evolutionary mechanism. This approach solves the technical problem in existing technologies where intelligent cognitive and decision-making mechanisms cannot achieve a secure closed loop, making it difficult to optimize multimodal interference information through deep semantic understanding and dynamic adaptive anti-interference strategies while ensuring data security.

[0028] The above-disclosed embodiments are merely one or more preferred embodiments of this application and should not be construed as limiting the scope of this application. Those skilled in the art can understand that all or part of the processes for implementing the above embodiments and equivalent changes made in accordance with the claims of this application still fall within the scope of this application.

Claims

1. A method for autonomous identification of multimodal interference in unmanned aerial vehicles (UAVs) based on a hierarchical cognitive architecture, characterized in that, include: Construct a hierarchical cognitive architecture, which includes a perception layer, a cognition layer, and a decision-making layer; Multimodal signals are acquired through encrypted transmission, and multimodal signal preprocessing and feature extraction are performed at the perception layer. By introducing a dynamic attention mechanism based on electromagnetic situational awareness, sensor data of different modalities are adaptively weighted and fused to generate a multi-dimensional feature vector set. At the cognitive layer, autonomous identification of interference patterns is performed. A multi-scale feature fusion network based on an attention mechanism is used to perform deep semantic encoding on the multi-dimensional feature vector set. Combined with the interference prior knowledge in the dynamic semantic knowledge graph, the interference type, intensity and intent are inferred through a causal reasoning engine to generate an interference cognitive situation map containing spatiotemporal evolution information. At the decision-making level, anti-interference strategies are dynamically generated. Based on the interference cognitive situation map, the optimal anti-interference decision sequence is generated through a hierarchical reinforcement learning-driven policy network. The effectiveness of the strategy execution is evaluated in real time, and the evaluation results are fed back to the perception and cognition layers to form a closed-loop feedback optimization mechanism. The specific method for obtaining multimodal signals through encrypted transmission is as follows: After acquiring multimodal sensor data, the perception layer uses a distributed encryption protocol to segment and encrypt the original data, generating encrypted data blocks with timestamps. The encrypted data blocks are then input into a feature extraction network based on homomorphic encryption to perform preliminary feature extraction without decryption, generating encrypted intermediate feature representations. These encrypted intermediate feature representations are then transmitted to the cognitive layer via a secure channel. The cognitive layer uses a secure multi-party computation protocol to collaboratively compute the encrypted intermediate feature representations of each modality, achieving cross-modal fusion and semantic understanding in the ciphertext domain.

2. The autonomous identification method for multimodal interference of unmanned aerial vehicles based on a hierarchical cognitive architecture as described in claim 1, characterized in that, In the constructed hierarchical cognitive architecture, The perception layer is equipped with a multi-source heterogeneous sensor array for real-time acquisition of multimodal environmental data from the UAV; The cognitive layer deploys a dynamic semantic knowledge graph and a causal reasoning engine for deep semantic understanding and interference intent inference of multimodal data; The decision-making level establishes an adaptive adversarial strategy generation mechanism to dynamically generate and optimize anti-interference decisions.

3. The autonomous identification method for multimodal interference of unmanned aerial vehicles based on a hierarchical cognitive architecture as described in claim 2, characterized in that, The dynamic attention mechanism includes a spectrum situational awareness module and an adaptive weighted fusion module; The spectrum situation awareness module is used to analyze the full-band energy distribution of the radio frequency front end in real time, identify abnormal energy peaks and frequency band blocking regions, and generate an electromagnetic situation heat map. The adaptive weighted fusion module dynamically assigns attention weights to each modality of data based on the electromagnetic situation heatmap. When broadband blocking interference is detected in the satellite navigation signal frequency band, the initial weight of the navigation mode is reduced, while the weight of the inertial measurement unit and the visual image mode is increased. When tracking interference targeting the communication link is detected, the weights of frequency hopping communication mode and radar echo mode are increased.

4. The autonomous identification method for multimodal interference of unmanned aerial vehicles based on a hierarchical cognitive architecture as described in claim 3, characterized in that, The dynamic semantic knowledge graph of the cognitive layer is constructed and updated using graph neural networks, including an interfering entity layer, a relation layer, and a spatiotemporal evolution layer; The interference entity layer defines the interference source ontology, interference style ontology, and platform characteristic ontology, with each node embedding a multi-dimensional attribute vector. The relation layer defines causal, temporal, and spatial relationships between entities, with causal relationship edges learned from historical data through Granger causality tests. The spatiotemporal evolution layer predicts the future evolution trend of knowledge graphs through long short-term memory networks, enabling forward-looking cognition of disturbance situations.

5. The autonomous identification method for multimodal interference of unmanned aerial vehicles based on a hierarchical cognitive architecture as described in claim 4, characterized in that, The causal reasoning engine adopts an architecture that combines structural causal models with counterfactual reasoning; The causal reasoning engine includes a causal discovery module for automatically learning the directed acyclic graph structure of the disturbance generation mechanism from a multi-dimensional feature vector set; The causal reasoning engine also includes a counterfactual prediction mechanism. Based on the generated directed acyclic graph structure, it simulates the system response under different anti-interference strategies through intervention operations, evaluates the effectiveness of potential intervention measures, and provides causal explanatory support for decision-making.

6. The autonomous identification method for multimodal interference of unmanned aerial vehicles based on a hierarchical cognitive architecture as described in claim 5, characterized in that, The attention-based multi-scale feature fusion network includes a cross-modal feature alignment sub-network and an adaptive attention fusion sub-network; The cross-modal feature alignment sub-network aligns the distribution of features from different modalities, generating modality-independent features in a unified representation space; The cross-modal feature alignment sub-network employs homomorphic encryption technology to complete the normalization and alignment operations of feature distribution in the ciphertext domain; An adaptive attention fusion subnetwork is designed with an intermodal attention gating mechanism to dynamically calculate the contribution weight of each modality feature to the current interference recognition task, thereby achieving adaptive selection and weighted fusion of features. The attention gating mechanism is based on a secure multi-party computation protocol and completes weight calculation and feature fusion without exposing the original features of each modality.

7. The autonomous identification method for multimodal interference of unmanned aerial vehicles based on a hierarchical cognitive architecture as described in claim 6, characterized in that, The hierarchical reinforcement learning-driven policy network of the decision layer consists of a high-level policy network and a low-level policy network. Based on the interference cognitive situation map, the high-level strategy network outputs discrete anti-interference mode selection decisions, including spectrum avoidance mode, power adaptive mode, waveform agility mode or trajectory replanning mode. The low-level policy network continuously adjusts specific anti-jamming parameters, including frequency hopping rate, transmit power level, or beam pointing angle, for the selected high-level mode. The policy network introduces a curiosity-driven intrinsic reward mechanism, which estimates the uncertainty of state transitions by predicting dynamic models and provides additional rewards for exploring unknown disturbance patterns, thereby enhancing the adaptability to new disturbances.

Images