A coastline inspection scene situation awareness method and system for suspicious target intrusion

Abstract

The application relates to a coastline inspection scene situation awareness method and system for suspicious target intrusion. The method generates a cooperative situation awareness feature through distributed cooperation and fusion, combines an inspection image to realize target identification and threat determination, and realizes positioning and threat quantification of a suspicious target in a complex environment; a node situation value is calculated according to a threat level and a node topological position, and a global situation inconsistency index is calculated in combination with situation description vectors of all nodes to the same target, so that dynamic monitoring of system perception consistency is realized; when the node situation value is lower than a threshold value, the information fusion weight of the node is dynamically adjusted to suppress the influence of unreliable nodes, and meanwhile, the misjudgment sample screened by the global situation inconsistency index is used to update the identification model through incremental learning. Therefore, the monitoring blind area and identification difficulty problem in the long-distance and complex environment of the coastline are effectively solved, and the precision, robustness and self-adaptability of cooperative perception are significantly improved.

Description

Technical Field

[0001] This invention belongs to the field of intelligent monitoring technology, and in particular relates to a situational awareness method and system for coastline patrol scenarios facing suspicious target intrusion. Background Technology

[0002] As a crucial line of defense for border security, the coastline's patrol is of paramount importance in preventing the intrusion of suspicious targets such as smuggling vessels. Traditional coastline patrols rely primarily on fixed radar stations, patrol boats, and manual inspections. However, these methods suffer from significant drawbacks, including incomplete monitoring range, slow response times, and difficulty in achieving continuous, 24 / 7 coverage, especially given the vast coastline, complex and varied terrain, and adverse weather conditions. In particular, suspicious targets such as smuggling vessels are often small, fast, and adept at using complex sea conditions and terrain for cover, making accurate identification and continuous tracking difficult with traditional monitoring methods.

[0003] The development of unmanned systems technologies such as drones and unmanned surface vessels has provided new technological pathways for coastline inspection. However, single-type unmanned systems, limited by their own sensor field of view and endurance, still struggle to independently complete large-scale, continuous situational awareness tasks. Therefore, constructing a collaborative perception network composed of multiple heterogeneous unmanned systems has become a key direction for solving this problem. Existing collaborative perception methods mainly suffer from the following technical bottlenecks: First, in long-distance, poorly communicating coastal environments, there is a lack of information architecture capable of supporting low-latency interaction among distributed heterogeneous unmanned systems, making it difficult to eliminate perception blind spots in remote or inaccessible areas; Second, against the backdrop of dynamically changing and complex sea surfaces, existing target monitoring methods lack sufficient accuracy in identifying suspicious targets of varying shapes and sizes, and struggle to effectively distinguish targets from environmental interference such as waves and sea clutter; Third, traditional situation assessment often adopts a centralized processing model, neglecting the reliability differences among nodes in a distributed system, and the assessment mechanism lags behind real-time situational changes, making it difficult to provide timely feedback and correct system behavior, resulting in insufficient response capability to rapidly intruding targets.

[0004] Therefore, how to achieve efficient collaboration of distributed heterogeneous unmanned systems in complex coastal environments, accurately identify suspicious targets, and evaluate the system's perception status in real time has become a technical challenge that urgently needs to be solved in this field. Summary of the Invention

[0005] Therefore, it is necessary to provide a situational awareness method and system for coastline patrol scenarios facing suspicious target intrusion, in order to address the above-mentioned technical problems.

[0006] Firstly, this application provides a situational awareness method for coastline patrol scenarios involving suspicious target intrusion, including:

[0007] S1. Based on the distributed heterogeneous unmanned system cluster deployed in the coastline inspection area, the time-series perception data streams collected by each unmanned system node are processed in a distributed collaborative manner to obtain the collaborative situational awareness characteristics of each unmanned system node at its respective time step.

[0008] S2. Based on the inspection images and collaborative situational awareness features collected by each unmanned system node, identify targets and determine their threat levels in the inspection images to obtain suspicious targets and their corresponding threat levels. Generate situational description vectors based on the position and motion state of suspicious targets in the inspection images.

[0009] S3. Based on the threat level and the topological location of each unmanned system node in the distributed communication network corresponding to the distributed heterogeneous unmanned system cluster, calculate the node status value of each unmanned system node.

[0010] S4. Calculate the global situation inconsistency index based on the situation description vector of each unmanned system node for the same suspicious target;

[0011] S5. When the node status value is lower than the preset confidence threshold, the information fusion weight of the corresponding unmanned system node is adjusted according to the node status value in the distributed collaborative fusion processing. The misjudged samples corresponding to the low node status value unmanned system nodes that cause status inconsistency are screened according to the global status inconsistency index. The misjudged samples are used to incrementally learn and update the processing model for identification and threat level determination.

[0012] Secondly, this application also provides a situational awareness system for coastline patrol scenarios targeting suspicious target intrusion, used to implement the method described in the first aspect, the system comprising:

[0013] The distributed situation fusion module is used to perform distributed collaborative fusion processing on the time-series perception data streams collected by each unmanned system node based on a distributed heterogeneous unmanned system cluster deployed in the coastline inspection area, so as to obtain the collaborative situation perception characteristics of each unmanned system node at its respective time step.

[0014] The target situation generation module is used to identify targets and determine their threat levels in the inspection images collected by each unmanned system node and the collaborative situational awareness features, to obtain suspicious targets and their corresponding threat levels, and to generate situational description vectors based on the position and motion state of the suspicious targets in the inspection images.

[0015] The node situation quantification module is used to calculate the node situation value of each unmanned system node based on the threat level and the topological location of each unmanned system node in the distributed communication network corresponding to the distributed heterogeneous unmanned system cluster.

[0016] The situation consistency analysis module is used to calculate the global situation inconsistency index based on the situation description vector of each unmanned system node for the same suspicious target.

[0017] The adaptive optimization module is used to adjust the information fusion weight of the corresponding unmanned system node according to the node status value in the distributed collaborative fusion processing when the node status value is lower than the preset confidence threshold. It also filters out the misjudged samples corresponding to the low node status value unmanned system nodes that cause status inconsistency based on the global status inconsistency index, and uses the misjudged samples to incrementally learn and update the processing model for identification and threat level determination.

[0018] Thirdly, this application also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement a situational awareness method for a coastline patrol scenario oriented towards suspicious target intrusion as described in the first aspect.

[0019] Fourthly, this application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements a situational awareness method for a coastline patrol scenario oriented towards suspicious target intrusion as described in the first aspect.

[0020] The aforementioned situational awareness method and system for coastline patrol scenarios targeting suspicious target intrusion firstly integrates the temporal perception data streams collected by various unmanned system nodes through distributed collaborative fusion processing. This allows each node to obtain collaborative situational awareness features containing its own observations and information from neighboring nodes, providing spatiotemporal correlation information from a global perspective for subsequent identification. Based on this, target identification and threat level determination are performed by combining patrol images with the collaborative situational awareness features, generating suspicious targets and their situational description vectors. This achieves accurate location and threat quantification of suspicious targets in complex coastline environments. Furthermore, the situational value of each node is calculated based on the threat level and its topological position in the distributed communication network, quantifying the importance and reliability of each node in the perception task. A global situational inconsistency index is calculated by combining the situational description vectors of all nodes for the same target, enabling dynamic monitoring of the overall system perception consistency. Finally, when a node's situational value falls below a preset threshold, its information fusion weight in collaborative fusion is dynamically adjusted to suppress the negative impact of unreliable nodes. Simultaneously, misjudged samples filtered by the global situational inconsistency index are used to incrementally learn and update the identification model, enabling the system to continuously adapt to environmental changes and correct its own perception biases. Through the above methods, this technical solution achieves a complete closed loop from the collaborative fusion of multi-source heterogeneous information, accurate target identification and threat assessment, quantitative evaluation of node and global situation, to adaptive adjustment of fusion weights and online iterative optimization of the model. It effectively solves the problems of monitoring blind spots, target identification difficulties and slow response mechanisms in traditional perception methods under long-distance coastlines, complex terrain and severe weather conditions. It significantly improves the collaborative perception accuracy, robustness and adaptability of distributed heterogeneous unmanned systems in suspicious target intrusion scenarios. Attached Figure Description

[0021] To more clearly illustrate the technical solutions in the embodiments or related technologies of this application, the accompanying drawings used in the description of the embodiments or related technologies will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0022] Figure 1 A flowchart illustrating a situational awareness method for coastline patrol scenarios targeting suspicious intrusion provided by this invention.

[0023] Figure 2 This is a schematic diagram of the process of generating the collaborative situational awareness features of each unmanned system node at its respective time step in an optional embodiment of the present invention.

[0024] Figure 3 This is a schematic diagram of the structure of a situational awareness system for a coastline patrol scenario oriented towards suspicious target intrusion, provided by the present invention. Detailed Implementation

[0025] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0026] refer to Figure 1 The document presents a flowchart illustrating a situational awareness method for coastline patrol scenarios involving suspicious target intrusion, as provided in this application. The method includes the following steps:

[0027] S1. Based on the distributed heterogeneous unmanned system cluster deployed in the coastline inspection area, the time-series perception data streams collected by each unmanned system node are processed in a distributed collaborative manner to obtain the collaborative situational awareness characteristics of each unmanned system node at its respective time step.

[0028] Specifically, the distributed heterogeneous unmanned system cluster is the hardware foundation of this invention. Its deployment is based on the terrain features, sea conditions, and monitoring blind spots of the coastline inspection area, adopting a heterogeneous networking mode of "fixed nodes + mobile nodes." It consists of three types of nodes: UAV nodes, unmanned surface vessel (USV) nodes, and shore-based fixed monitoring nodes. These three types of nodes form a complementary and collaborative perception network, covering nearshore, offshore, and complex terrain areas. The UAV nodes use multi-rotor UAVs equipped with visible light cameras, infrared thermal imaging sensors, and small millimeter-wave radars, primarily responsible for inspecting areas inaccessible by vehicles, such as nearshore shoals and cliffs. The USV nodes use small, high-speed USVs equipped with marine radar, sonar sensors, and GPS / BeiDou dual-mode positioning modules, responsible for inspecting the sea surface within a designated offshore range. The shore-based fixed monitoring nodes are deployed at key points along the coastline, equipped with high-power millimeter-wave radars and high-definition monitoring cameras, enabling 24-hour continuous monitoring of fixed areas. The three types of nodes are interconnected through a distributed communication network. The communication network adopts a hybrid networking mode of LoRa and 5G. The LoRa module is used for low-bandwidth data transmission in remote areas to ensure low-latency interaction between nodes, while the 5G module is used for high-speed transmission of large-capacity data such as high-definition images and radar data in near-shore areas to ensure data real-time performance.

[0029] Distributed collaborative fusion processing is the core technology of this step. Its core principle lies in breaking the perception limitations of a single node. Through information interaction and fusion between nodes, the accuracy and completeness of situational awareness features are improved. Unlike traditional centralized fusion processing, this solution adopts a two-level distributed fusion architecture of "data layer-feature layer," avoiding the problems of large data transmission volume, high latency, and high risk of single point of failure in centralized processing. In the data layer fusion stage, each unmanned system node first preprocesses its own time-series perception data stream to remove noise interference and complete data alignment. For visible light image data collected by UAV nodes, Gaussian filtering algorithm is used to remove noise caused by sea fog and light changes. At the same time, histogram equalization is used to enhance image contrast, which is convenient for subsequent target identification. For infrared thermal imaging data, median filtering algorithm is used to remove salt-and-pepper noise. Temperature threshold segmentation is also performed to distinguish warm targets such as humans and ships from the sea surface background. For radar data, Kalman filtering algorithm is used to remove sea clutter interference. The state equation is set as a uniform linear motion model, and the observation equation is set as a radar ranging and angle measurement model. Iterative updates are used to achieve denoising and smoothing of radar data. For sonar data collected by unmanned surface vessel nodes, wavelet transform denoising algorithm is used to remove marine environmental noise and ship noise; for GPS positioning data, weighted average algorithm is used to fuse multiple sets of positioning data to reduce positioning error.

[0030] After preprocessing, each node sends the preprocessed data to its neighboring nodes through a distributed communication network. The neighboring nodes then perform data fusion based on data type and reliability. For example, in the fusion of radar data from UAV nodes and shore-based fixed nodes, a weighted fusion algorithm is used. The weights are determined by the distance between the node and the target; the closer the node, the greater its weight. The weight calculation formula is as follows:

[0031]

[0032] in, This represents the radar data fusion weight of the i-th node. Indicates the maximum monitoring distance of the node. This represents the actual distance from the i-th node to the target. This represents the sum of the differences between the maximum monitoring distance of all nodes and the actual distance from the node to the target, where n represents the total number of nodes participating in data fusion. This formula enables complementary correction of data and improves the accuracy of data layer fusion.

[0033] In the feature layer fusion stage, each node extracts features from the data stream after data layer fusion, and then achieves feature fusion through a distributed collaborative algorithm to obtain collaborative situational awareness features. Specifically, for image data, a CNN convolutional neural network is used to extract target features, with the network using the ResNet50 architecture. The input image size is adjusted according to actual needs, and feature vectors are extracted through forward propagation, including features such as the target's contour, texture, and brightness. For radar and sonar data, feature engineering methods are used to extract the target's motion and morphological features to construct feature vectors. For positioning data, the target's position features are extracted to construct feature vectors.

[0034] After feature extraction, a federated learning framework is used to achieve distributed collaborative feature fusion. Each node acts as a client of the federated learning framework, training the feature fusion model locally and sending only the model parameters to the federated server. The server aggregates and updates the parameters of each node. The aggregation uses the FedAvg algorithm, and the aggregation weight is determined based on the amount of data per node; the larger the data volume, the larger the aggregation weight. The formula for calculating the aggregation weight is as follows: .in, This represents the aggregated weights of the model parameters for the j-th client node. This represents the amount of local training data for the j-th client node. This represents the total amount of local training data across all K client nodes, where K represents the total number of client nodes participating in federated learning. The server feeds back the aggregated and updated parameters to each node, enabling collaborative training of the feature fusion model. After multiple rounds of iterative training, each node obtains a unified feature fusion model, inputting its extracted features into the model and outputting collaborative situational awareness features. These features contain multi-dimensional information such as the target's position, motion, and morphology, effectively improving the accuracy of subsequent target recognition. Each time step outputs corresponding collaborative situational awareness features, achieving continuous temporal situational awareness. The time step is consistent with the sensor acquisition frequency of each node.

[0035] S2. Based on the inspection images and collaborative situational awareness features collected by each unmanned system node, identify targets and determine their threat levels in the inspection images to obtain suspicious targets and their corresponding threat levels. Generate situational description vectors based on the position and motion state of the suspicious targets in the inspection images.

[0036] Specifically, the core of this step is threat level determination, and target identification is the foundation of threat level determination. The two work together to achieve accurate identification of suspicious targets and risk quantification.

[0037] During the recognition process, the inspection image is input into the YOLOv8 algorithm, which outputs the preliminary recognition results of the target, including the target category, bounding box coordinates, and confidence score. The confidence score threshold is set according to the actual recognition requirements. Recognition results below the threshold are considered false detections and are discarded.

[0038] The principle of threat level determination is based on multi-dimensional indicators such as target category, motion state, and location information to construct a quantitative threat level evaluation system. This system enables risk classification of suspicious targets, providing a basis for subsequent node situational awareness calculations and system optimization. Threat level determination employs a multi-level fuzzy comprehensive evaluation method. First, the evaluation indicator system is determined, selecting four indicators: target category, motion state, location, and target behavior. The weights of each indicator are determined using the analytic hierarchy process (AHP), with the sum of the weights being 1. The weight allocation satisfies the following conditions: .in, This represents the weight of the target category indicator and is the most crucial evaluation metric. Indicates the weight of motion state indicators. Indicates the weight of the location indicator. This indicates the weight of the target behavior indicator.

[0039] Each evaluation indicator is quantified and assigned a value. The target category indicator is classified and assigned a value based on the risk level of the target type, with a higher value for a higher risk level. The motion status indicator is classified and assigned a value based on the target's speed and direction of movement, with a higher value for faster speed and closer movement to the restricted area. The location indicator is classified and assigned a value based on the distance between the target and the key protection area of ​​the coastline, with a higher value for closer distance. The target behavior indicator is classified and assigned a value based on whether the target exhibits evasion behavior, with a higher value for evasion behavior and abnormal movement trajectory.

[0040] The specific calculation process for threat level determination is as follows: First, calculate the membership degree of each evaluation indicator. Using the triangular membership function, the quantified values ​​of each indicator are mapped to the [0,1] interval. The membership degree calculation formula is: .in, This represents the membership degree corresponding to the quantitative value x of the evaluation index. This indicates the lower limit of a certain level of the evaluation index. This function represents the upper limit of a certain level of the evaluation indicator. It can be used to standardize the quantitative values ​​of each indicator to a uniform range, which is convenient for subsequent comprehensive evaluation.

[0041] Then, the single-indicator evaluation matrix is ​​calculated by multiplying the membership degree of each indicator by its corresponding weight to obtain the single-indicator evaluation result. The formula for calculating the single-indicator evaluation result is as follows: .in, This represents the single-indicator evaluation result of the i-th evaluation indicator under the j-th evaluation level. This represents the membership degree of the i-th evaluation index under the j-th evaluation level. This represents the weight of the i-th evaluation indicator.

[0042] Finally, the threat level quantification value of the target is obtained through fuzzy synthesis. A weighted average synthesis method is used, and the synthesis formula is as follows: Where T represents the threat level quantification value, This represents the single-indicator evaluation result of the i-th evaluation indicator. The comprehensive threat quantification value of the target can be obtained through this formula. Then, the threat level is divided into four levels according to the range of the quantification value. Different ranges of quantification value correspond to different threat levels, namely high risk, medium-high risk, medium risk, and low risk. Different risk levels correspond to different handling requirements.

[0043] After the threat level is determined, a situation description vector is generated. This vector is used to quantify the real-time situation of the suspicious target. Its dimensions are set according to the actual situation description requirements, and specifically include the target center coordinates, target movement speed, movement direction, threat level quantification value, target bounding box size, target identification confidence, and node ID, providing a basis for subsequent global situation consistency judgment.

[0044] S3. Based on the threat level and the topological location of each unmanned system node in the distributed communication network corresponding to the distributed heterogeneous unmanned system cluster, calculate the node status value of each unmanned system node.

[0045] Specifically, the role of node situation values ​​is to quantitatively assess the perception reliability and importance of each unmanned system node, providing a basis for subsequent information fusion weight adjustments and misjudgment sample screening. Its calculation requires comprehensive consideration of two core factors: threat level and node topology location. These two factors are interrelated; the higher the threat level, the higher the node's perception importance to the target; the more critical the node's topology location, the greater its perception reliability weight. First, the threat level is quantified and mapped, using the threat level quantification value obtained in step S2. Normalization is performed; the normalization formula is: .in, This represents the normalized threat level quantization value, with a value range of [0,1]. T represents the original threat level quantization value obtained in step S2. This represents the minimum value of the threat level quantification. This represents the maximum value of the threat level quantification. This value reflects the contribution of the target threat level to the node's situational awareness value. The higher the target threat level, the greater the contribution. The larger the value, the higher the node situation value, because the node perception data corresponding to high-threat targets has a greater impact on the overall situation assessment and needs to be given special attention.

[0046] Secondly, the topological location of nodes is quantitatively evaluated. The topology of the distributed communication network is represented using a graph theory model, denoted as . Where V is the set of nodes, including three types of nodes: UAVs, unmanned surface vessels, and shore-based fixed nodes; and E is the set of communication links between nodes, used to represent the interconnection relationships between nodes. The quantitative indicators for node topological location are betweenness centrality and degree centrality, which respectively reflect the node's pivotal role and connectivity in the network.

[0047] The formula for calculating degree centrality is: .in, Let represent the degree centrality of the i-th node, with a value in the range [0,1]. This represents the number of neighboring nodes of the i-th node, i.e., the number of nodes with which the i-th node has a communication link, where n is the total number of nodes in the network. The larger the value, the stronger the node's connectivity, the more information it can interact with, and the wider the range of its perceived data dissemination.

[0048] The formula for calculating betweenness centrality is: .in, Let represent the betweenness centrality of the i-th node, with values ​​ranging from [0,1]. This represents the number of shortest paths from node s to node t. This represents the number of shortest paths from node s to node t that pass through node i, where s and t are nodes in the network that are different from i. The larger the value, the more critical the node is in the network, serving as an essential pathway for information exchange between other nodes, and the greater the impact of its perceived data on global information fusion.

[0049] To comprehensively reflect the importance of a node's topological location, degree centrality and betweenness centrality are weighted and fused to obtain the topological location weight, calculated as follows: .in, This represents the topological position weight of the i-th node, with a value range of [0,1]. The weighting coefficients represent betweenness centrality. The weight coefficients representing degree centrality are set according to the requirements of the coastline inspection scenario and satisfy the following conditions: .

[0050] The final formula for calculating the nodal situation value is: .in, Let represent the node status value of the i-th node, where α, β, and γ are weighting coefficients set according to the actual inspection scenario requirements, and satisfying the following conditions: α is the contribution weight of threat level, which is the most critical influencing factor; β is the contribution weight of topological location; and γ is the contribution weight of node's own reliability.

[0051] This is a node's own reliability metric, used to quantify the node's sensor performance, battery life, and communication quality. Its calculation formula is: .in, , , The weighting coefficients for each reliability component satisfy the following conditions: , The accuracy of node sensor identification is obtained through statistical analysis of historical data. This represents the percentage of remaining power at the node. For shore-based fixed nodes, which use a continuous power supply mode, this value is 1. The node communication quality is calculated using the packet loss rate. The calculation frequency of the node status value is consistent with the time step to reflect the dynamic status of the node in real time. The preset confidence threshold is set according to the actual scenario. When the node status value is lower than the threshold, it indicates that the node's perception reliability is insufficient and its information fusion weight needs to be adjusted.

[0052] S4. Calculate the global situation inconsistency index based on the situation description vector of each unmanned system node for the same suspicious target.

[0053] Specifically, the core function of the global situation inconsistency index is to quantify the differences in situation assessments of the same suspicious target by multiple unmanned system nodes. The greater the difference, the higher the probability of misjudgment, providing a quantitative basis for subsequent screening of misjudged samples. Its calculation is based on ensuring that multiple nodes are observing the same suspicious target. This is achieved through a target matching algorithm, which associates the situation description vectors generated by different nodes with the same suspicious target. The target matching algorithm uses the target coordinates, velocity, and direction of motion in the situation description vectors as matching features to calculate the feature similarity between two situation description vectors. The similarity threshold is set according to actual matching requirements; when the similarity is not lower than this threshold, the situation description vectors are determined to be the same target.

[0054] After target matching is completed, for the same suspicious target, all associated situation description vectors are collected, denoted as... Where m is the number of nodes observing the target, and m is not less than 2. If only a single node observes, the inconsistency index is not calculated. Let be the situation description vector for the i-th node. , These are the x and y coordinates of the target center, respectively. For the target speed of motion, For the direction of target movement, As a quantification value for threat level, , These are the width and height of the target bounding box, respectively. Confidence level for target identification.

[0055] The global situation inconsistency index is calculated using a vector similarity-based method. First, the situation description vector is normalized to eliminate the influence of differences in magnitude across different dimensions. The normalization range is [0,1]. The normalization formula for the coordinates is... , ,in , These are the normalized x and y coordinates of the target center. , These represent the minimum and maximum values ​​of the abscissa of the coastline inspection area, respectively. , These represent the minimum and maximum values ​​of the ordinate of the coastline inspection area, respectively.

[0056] The normalized formula for velocity is: ,in The normalized target velocity, Let be the maximum monitoring speed of the node. The normalized formula for the direction of motion is: ,in This represents the normalized target motion direction. The normalization formula for the bounding box dimensions is: , ,in , These are the normalized target bounding box width and height, respectively. , These represent the width and height of the inspection image, respectively. The threat level quantification value and identification confidence level are already in the [0,1] range and do not require normalization. The normalized situation description vector is denoted as... .

[0057] Then, the cosine similarity between any two normalized situation description vectors is calculated. Cosine similarity measures the directional consistency between the two vectors; a similarity closer to 1 indicates a more consistent situation assessment between the two nodes, while a similarity closer to 0 indicates a greater difference in assessment. The formula for calculating cosine similarity is: .in, Let represent the cosine similarity between the normalized situation description vectors of the i-th node and the j-th node. The dot product of two vectors. , Let these be the magnitudes of the two vectors, and the formula for calculating the magnitude of a vector is: ,in This represents the k-th dimension value of the normalized situation description vector for the i-th node.

[0058] Calculate the cosine similarity between all pairs of nodes to obtain the similarity matrix. Where i and j both range from 1 to m, and This means that the similarity between nodes themselves is 1. The formula for calculating the global situation inconsistency index is: Where I represents the global situation inconsistency index, with a value range of [0,1], and m is the number of nodes observing the same target. I represents the sum of the cosine similarities of all pairs of nodes. The core logic of this formula is to calculate the average of all pairwise similarities and subtract this average from 1 to obtain the inconsistency index. The closer I is to 1, the greater the difference in the situation assessment of multiple nodes and the more inconsistent the global situation. The closer I is to 0, the more consistent the situation assessment of multiple nodes and the more stable the global situation.

[0059] To ensure the reliability of the calculation results, an inconsistency threshold needs to be set. Based on the actual scenario, when I is greater than When I is greater than a certain value, it indicates a high degree of inconsistency in the overall situation, suggesting the possibility of node misjudgment, requiring further screening of misjudged samples; when I is not greater than a certain value... When the overall situation is consistent, there is no need to screen for misjudged samples. Furthermore, to avoid the impact of single-dimensional differences on inconsistency indicators, the dimensions of the situation description vector can be weighted during the calculation process. The weights of each dimension are set according to the importance of the situation assessment, with higher weights for core dimensions and lower weights for secondary dimensions. This weighting can further improve the accuracy of inconsistency indicators and better reflect the core differences in the situation assessment at each node.

[0060] S5. When the node status value is lower than the preset confidence threshold, the information fusion weight of the corresponding unmanned system node is adjusted according to the node status value in the distributed collaborative fusion processing. The misjudged samples corresponding to the low node status value unmanned system nodes that cause status inconsistency are screened according to the global status inconsistency index. The misjudged samples are used to incrementally learn and update the processing model for identification and threat level determination.

[0061] Specifically, this step improves the accuracy of information fusion by adjusting weights and updates the model through incremental learning to enhance the accuracy of target identification and threat level determination, thus addressing the misjudgment problem caused by environmental changes and node performance degradation during long-term operation. First, the information fusion weights are adjusted. These weights, used in distributed collaborative fusion processing, determine the contribution of each node's perceived data and features. Higher weights indicate a greater impact of node data on collaborative situational awareness features. Traditional methods use fixed weights, which cannot adapt to dynamic changes in node situational awareness. This solution dynamically adjusts weights based on node situational values, ensuring higher weights for high-reliability nodes and lower weights for low-reliability nodes, thereby improving the accuracy of the fusion results.

[0062] The adjustment rules for information fusion weights are as follows: First, calculate the node status values ​​of all nodes. And filter out The low-reliability nodes are denoted as set L, where A preset credibility threshold is set according to the actual scenario. For nodes in set L, their information fusion weights are adjusted based on the node's status value, using the following formula: .in, This represents the adjusted information fusion weight for the i-th low-reliability node. This represents the node status value of the i-th low-reliability node, where n is the total number of nodes in the network. This represents the sum of the state values ​​of all nodes. This formula ensures that the sum of the fusion weights of all nodes is 1, and the higher the state value of a node, the greater its weight, and the lower the state value of a node, the smaller its weight.

[0063] for For high-reliability nodes, their weights retain their initial weights. The initial weights are set according to the node type; different node types have their initial weights determined based on factors such as their sensing capabilities and coverage. If the high-reliability node's... If a node's reliability is significantly higher than other nodes, its weight can be appropriately increased. The increase can be set according to actual needs to ensure the dominant role of high-reliability nodes. The frequency of weight adjustment should be consistent with the calculation frequency of node situation values ​​to adapt to the dynamic changes in node situation in real time. By adjusting the weight, the impact of low-reliability nodes on the fusion results can be reduced, and the accuracy of collaborative situational awareness features can be improved.

[0064] Next, the misclassified samples are screened. The screening logic is as follows: nodes with high global situation inconsistency and low node situation values ​​are likely to be misclassified samples in terms of their generated situation description vectors and corresponding target identification and threat level determination results, and these need to be screened out for model updates. The specific screening steps are as follows: First, determine the global situation inconsistency index I greater than 1. The first step is to identify suspicious targets, which may be subject to node misclassification; the second step is to collect the situation description vectors and situation values ​​of all nodes corresponding to the target; the third step is to filter out... The fourth step involves extracting the target identification results, threat level judgment results, original inspection images, and collaborative situational awareness features corresponding to these low-reliability nodes as misjudged samples. These misjudged samples must be correctly labeled with the target category and threat level. Correct labeling is determined through manual review or the judgment results of high-reliability nodes. To ensure the validity of misjudged samples, a sample screening threshold must be set. Only when the judgment results of low-reliability nodes differ significantly from those of most high-reliability nodes should they be considered misjudged samples, thus avoiding the selection of invalid samples.

[0065] Finally, the misclassified samples are used to incrementally learn and update the identification and threat level determination models. The core advantage of incremental learning is that it eliminates the need to retrain the entire model; it only uses new misclassified samples to fine-tune the model, saving computing power and time, while adapting to the dynamic changes in the coastline scene. The incremental learning and update processes of the identification model and the threat level determination model are carried out separately. First, the identification model is updated: the original inspection images of the misclassified samples and the correct target category labels are input into the model, and the mini-batch gradient descent method is used for fine-tuning. The feature extraction layer of the model is frozen, and only the classification layer and regression layer are trained to avoid model overfitting. The learning rate, batch size, and number of iterations are set according to the fine-tuning requirements. After each iteration, the identification accuracy of the model is calculated, and the iteration stops when the accuracy no longer improves.

[0066] The threat level determination model was then updated: the collaborative situational awareness features of misjudged samples and the correct threat level labels were input into the model, the weights of the evaluation indicators were adjusted, and the weight adjustments were recalculated using the analytic hierarchy process (AHP). Simultaneously, the parameters of the membership function were optimized to better adapt the model to new target characteristics and threat scenarios. The frequency of incremental learning updates was determined based on the number of misjudged samples. When the number of misjudged samples reached a preset threshold, a model update was performed. After the update, the updated model parameters were synchronized to all unmanned system nodes to ensure that the processing models of each node remained consistent.

[0067] Meanwhile, to avoid model degradation, after each incremental learning iteration, the model is validated using historical normal samples. The validation accuracy must not be lower than a preset validation threshold. If the validation accuracy falls below this threshold, the learning rate and number of iterations are adjusted, and fine-tuning is performed until the validation requirements are met. Through dynamic adjustment of information fusion weights and incremental learning updates of the model, it can continuously adapt to changes in complex coastal environments, improve the accuracy of suspicious target identification and threat level determination, ensure the real-time performance and reliability of situational awareness, and effectively solve the problems of perception blind spots, slow response, and high false positive rates in traditional methods.

[0068] The aforementioned situational awareness method for coastline patrol scenarios targeting suspicious target intrusion firstly integrates the temporal perception data streams collected by each unmanned system node through distributed collaborative fusion processing. This allows each node to obtain collaborative situational awareness features containing its own observations and information from neighboring nodes, providing spatiotemporal correlation information from a global perspective for subsequent identification. Based on this, target identification and threat level determination are performed by combining patrol images with the collaborative situational awareness features, generating suspicious targets and their situational description vectors. This achieves accurate location and threat quantification of suspicious targets in complex coastline environments. Furthermore, the situational value of each node is calculated based on the threat level and its topological position in the distributed communication network, quantifying the importance and reliability of each node in the perception task. A global situational inconsistency index is calculated by combining the situational description vectors of all nodes for the same target, enabling dynamic monitoring of the overall system perception consistency. Finally, when a node's situational value falls below a preset threshold, its information fusion weight in collaborative fusion is dynamically adjusted to suppress the negative impact of unreliable nodes. Simultaneously, misjudged samples selected using the global situational inconsistency index are used to incrementally learn and update the identification model, enabling the system to continuously adapt to environmental changes and correct its own perception biases. Through the above methods, this technical solution achieves a complete closed loop from the collaborative fusion of multi-source heterogeneous information, accurate target identification and threat assessment, quantitative evaluation of node and global situation, to adaptive adjustment of fusion weights and online iterative optimization of the model. It effectively solves the problems of monitoring blind spots, target identification difficulties and slow response mechanisms in traditional perception methods under long-distance coastlines, complex terrain and severe weather conditions. It significantly improves the collaborative perception accuracy, robustness and adaptability of distributed heterogeneous unmanned systems in suspicious target intrusion scenarios.

[0069] refer to Figure 2 In one optional embodiment, based on a distributed heterogeneous unmanned system cluster deployed within the coastline inspection area, distributed collaborative fusion processing is performed on the time-series perception data streams collected by each unmanned system node to obtain the collaborative situational awareness features of each unmanned system node at its respective time step, including the following steps:

[0070] S11. At each unmanned system node, the sensor data sequence collected within a preset time window is input into the bidirectional gating loop unit. The forward hidden state is extracted through the forward gating loop unit, and the backward hidden state is extracted through the backward gating loop unit. The forward hidden state and the backward hidden state are spliced ​​together to obtain the local spatiotemporal dynamic characteristics of each unmanned system node.

[0071] Specifically, the window length is reasonably set based on the target's movement speed and sensor response speed in the actual inspection scenario, so that the sensor data sequence within the window can reflect the target's state changes over a continuous period of time. The sensor data sequence collected by each unmanned system node contains multi-dimensional perception information, covering various types of data such as images, radar, sonar, and positioning. Before being input into the bidirectional gating loop unit, the data needs to be standardized to eliminate the differences in magnitude between different types of sensor data.

[0072] The bidirectional gated loop unit consists of a forward gated loop unit and a backward gated loop unit, which operate in parallel and in opposite directions. The forward gated loop unit processes the sensor data sequence in chronological order from the start to the end of a preset time window, extracting the forward spatiotemporal features of the target, i.e., the state change pattern of the target from the past to the present. The backward gated loop unit processes the sensor data sequence in reverse chronological order from the end to the start of the preset time window, extracting the backward spatiotemporal features of the target, i.e., the state evolution trend of the target from the present to the future. Through the synergistic effect of input gates, forget gates, and output gates, the gated loop unit effectively alleviates the gradient vanishing and gradient exploding problems in time-series data processing, accurately captures the long-short-term dependencies in the sensor data sequence, and more comprehensively mines the dynamic features of the target.

[0073] The forward and backward hidden states correspond to the outputs of the forward and backward gated recurrent units, respectively. Both are fixed-dimensional feature vectors containing dynamic feature information of the target in different time dimensions. By concatenating these two feature vectors column-wise, the local spatiotemporal dynamic features of the node are obtained. These features include both the forward and backward temporal characteristics of the target, comprehensively reflecting the spatiotemporal changes of the target within a preset time window.

[0074] S12. Through a distributed communication network, the local spatiotemporal dynamic characteristics of each unmanned system node are broadcast to neighboring nodes within the communication range.

[0075] Specifically, each unmanned system node has a preset communication range threshold, which is set according to the node type, sensor performance and coastline terrain conditions. Other unmanned system nodes within the communication range are the neighboring nodes of that node, and the number and distribution of neighboring nodes change dynamically according to the node deployment location and communication environment.

[0076] The broadcast of local spatiotemporal dynamic features adopts a combination of point-to-point transmission and multicast. After generating local spatiotemporal dynamic features, each node immediately encapsulates them into a standard data frame through the communication module. The data frame contains key information such as node ID, time step identifier, and local spatiotemporal dynamic feature data, ensuring that neighboring nodes can accurately identify the source of the features and the corresponding time step.

[0077] S13. For each unmanned system node, generate the corresponding query vector, key vector, and value vector based on its own local spatiotemporal dynamic characteristics, and generate the corresponding key vector and value vector based on the local spatiotemporal dynamic characteristics of all neighboring nodes.

[0078] Specifically, the generation of query vectors, key vectors, and value vectors is achieved through linear transformations. Each node has three preset independent linear transformation matrices, which are used to map local spatiotemporal dynamic features to query vectors, key vectors, and value vectors, respectively.

[0079] For the local spatiotemporal dynamic features of a node, three independent linear transformation matrices are used for mapping to generate the corresponding query vector, key vector, and value vector. These three vectors maintain consistent dimensions, matching the dimensionality requirements for subsequent attention score calculations. For the local spatiotemporal dynamic features of each neighboring node, only two independent linear transformation matrices are used for mapping to generate the corresponding key vector and value vector. No query vector is generated because the query vector is only used for calculating the node's attention to other nodes (its own and neighboring nodes), while the neighboring node's query vector is only used for its own attention calculation and is unrelated to the current node.

[0080] S14. For each unmanned system node, calculate the attention score of the unmanned system node to each computed node based on its own query vector and the key vector of each computed node; where the computed nodes include the unmanned system node itself and all its neighboring nodes; the formula for calculating the attention score is:

[0081]

[0082] in, Indicates at time step node For nodes Attention score Represents a node At time step The query vector, Represents a node At time step The key vector, The dimension of the key vector. Represents a node of The node index of each neighboring node. .

[0083] Specifically, the attention score quantifies the attention the current node pays to each feature of the computed node. The higher the attention, the higher the corresponding attention score, and the greater the weight of the feature of the computed node in the subsequent fusion process. Conversely, the lower the attention score, the smaller the weight. By calculating the attention score, adaptive weighting of features of different nodes can be achieved, highlighting the effective features of key nodes and suppressing ineffective or interfering features.

[0084] In the formula for calculating attention score, This represents the attention score of node i to node j at time step t, with a value range of [0,1]. The higher the score, the higher the attention of node i to the features of node j, and the greater the contribution of its features to the fusion process. The query vector for node i at time step t is generated by linear transformation of the local spatiotemporal dynamic features of node i itself, and is used to capture the feature requirements and focus of node i. The key vector of node j at time step t is generated by linear transformation of the local spatiotemporal dynamic features of node j. It is used to characterize the feature attributes of node j and facilitates similarity matching with the query vector of node i. The dimension of the key vector is used to normalize the dot product of the query vector and the key vector, avoiding excessively large dot product results that could saturate the exponential function and improving the stability and accuracy of attention score calculation. This represents the node index of node i's K neighboring nodes, where K is the number of neighboring nodes of node i, and its value changes dynamically according to the node's deployment location and communication range.

[0085] S15. For each unmanned system node, the value vectors of itself and each neighbor node are weighted and summed according to the attention score to obtain the collaborative situational awareness feature that integrates neighbor information; wherein, the calculation formula of the collaborative situational awareness feature is:

[0086]

[0087] in, Represents a node At time step The characteristics of collaborative situational awareness. Represents a node For nodes Attention score Represents a node At time step The value vector.

[0088] Specifically, the value vector is the core representation of the local spatiotemporal dynamic features of the computed node. It contains the spatiotemporal dynamic information of the target captured by the computed node. By weighting and summing the value vector through attention scores, it is possible to achieve adaptive fusion of the features of the node itself and its neighboring nodes, highlighting the effective features of nodes with high attention. The fused collaborative situational awareness features can integrate the perception information of multiple nodes, make up for the perception limitations of a single node, and improve the completeness and accuracy of the features.

[0089] In the calculation formula of collaborative situational awareness characteristics, It represents the collaborative situational awareness characteristics of node i at time step t. It is a fixed-dimensional feature vector that integrates the local spatiotemporal dynamic characteristics of node i itself and all neighboring nodes, and can comprehensively reflect the spatiotemporal state of the target at time step t. The attention score of node i to node j is calculated by step S14 and is used to determine the weight of the value vector of node j in the fusion process. The higher the attention score, the greater the corresponding weight and the greater the contribution of the value vector to the collaborative situational awareness features. This represents the value vector of node j at time step t, which is generated by linear transformation of the local spatiotemporal dynamic features of node j and contains the core spatiotemporal dynamic features of the target captured by node j.

[0090] In one optional embodiment, based on the inspection images and collaborative situational awareness features collected by each unmanned system node, targets in the inspection images are identified and their threat levels are determined to obtain suspicious targets and their corresponding threat levels, including the following steps:

[0091] S21. Input the inspection images collected by each unmanned system node into the semantic segmentation network, and perform pixel-level classification on the inspection images through the semantic segmentation network to obtain a segmentation mask containing the category of suspicious targets.

[0092] Specifically, the inspection images contain various pixel information, including the sea surface background, waves, sea clutter, and various targets. The semantic segmentation network plays a crucial role in achieving pixel-level category differentiation and accurately locating pixel regions in the image that may belong to suspicious targets, laying the foundation for subsequent target extraction. The selected semantic segmentation network is suitable for coastline inspection scenarios, possesses strong anti-interference capabilities, and can effectively distinguish suspicious targets from interference factors such as the sea surface background and waves. The U-Net improved architecture is preferentially adopted. This architecture achieves multi-scale feature fusion through an encoder-decoder structure, which can accurately capture pixel features of suspicious targets of different sizes, and is particularly suitable for segmentation tasks of small suspicious targets.

[0093] The input to the semantic segmentation network is the raw inspection image collected by the unmanned system nodes. The input image needs to be preprocessed, including normalization and size unification. The normalization formula is as follows: ,in This represents the normalized pixel value. Represents the original pixel value. This represents the minimum pixel value in the image. This represents the maximum pixel value in the image, thus eliminating differences in pixel values ​​under different lighting and sea conditions and ensuring the consistency of network input. During the network training phase, a labeled dataset from a coastline inspection scenario is used. The labeled data includes pixel-level annotations of various suspicious and non-suspicious targets. The network parameters are optimized using the cross-entropy loss function, the formula of which is: ,in Indicates the loss value. Indicates the total number of pixels. This represents the true label of the i-th pixel (0 indicates a non-suspicious target, and 1 indicates a suspicious target). This indicates the probability that the network predicts the i-th pixel as a suspicious target, thus improving the accuracy of pixel classification.

[0094] During segmentation, the semantic segmentation network extracts multi-scale features from the inspected image through the encoder. Shallow features capture detailed information such as the edges and textures of the target, while deep features capture the global contour and semantic information of the target. The decoder fuses the deep and shallow features through upsampling, gradually restoring the image resolution, and finally outputs a segmentation mask with the same size as the input image. The segmentation mask is a binary or multi-valued image, where different pixel values ​​correspond to different category labels, clearly marking pixel regions belonging to the suspicious target category and pixel regions belonging to the non-suspicious target category, achieving pixel-level localization of suspicious targets.

[0095] S22. Extract the pixel regions marked as suspicious target categories from the segmentation mask, and crop the corresponding image sub-regions in the inspection image according to the pixel regions; perform size adjustment and flattening processing on the image sub-regions to obtain the suspicious target image block vector.

[0096] Specifically, the pixel regions marked as suspicious target categories in the segmentation mask are usually irregular in shape. First, the minimum bounding rectangle of the pixel region is extracted using a connected component analysis algorithm to determine the coordinate range of the region in the inspection image. ,in The coordinates of the top left corner of the rectangular area. The coordinates are the bottom right corner coordinates of the rectangular area, in pixels.

[0097] Based on the determined coordinate range, corresponding image sub-regions are cropped from the original inspection image. These sub-regions contain only suspicious targets and a small amount of background information, effectively reducing the impact of background interference on subsequent feature extraction. Since the cropped image sub-regions vary in size, resizing is required. A bilinear interpolation algorithm is used to adjust all image sub-regions to a uniform size. ,in To adjust the width of the image, The height and size of the adjusted image are set according to the input requirements of the subsequent stacked sparse autoencoder.

[0098] After resizing, the image sub-regions are flattened, which converts the two-dimensional image pixel matrix into a one-dimensional vector. The flattening order is row-major, and the dimension of the flattened vector is [missing information]. That is, the image patch vector of the suspicious target. ,in This represents a vector of image blocks representing a suspicious target. This represents the dimensional space of a vector, which contains pixel feature information of the suspicious target, providing a foundation for subsequent deep visual feature extraction.

[0099] S23. Input the image block vector of the suspicious target into the stacked sparse autoencoder, extract deep features through the stacked sparse autoencoder to obtain the deep visual features of the suspicious target; calculate the spatial position parameters and motion state parameters of the suspicious target in the preset coordinate system according to the pixel position of the suspicious target in the inspection image to obtain the motion features.

[0100] Specifically, the stacked sparse autoencoder consists of multiple stacked sparse autoencoders. Each sparse autoencoder contains an input layer, a hidden layer, and an output layer. Its function is to perform unsupervised feature learning on the image patch vectors of suspicious targets, mine deep abstract features from the image patch vectors, remove redundant information, and improve the discriminative ability of the features. Suspicious target image patch vectors After inputting a stacked sparse autoencoder, primary visual features are extracted through the lower-level autoencoder. The mapping relationship of the lower-level autoencoder is as follows: ,in Represents the primary visual feature vector. This represents the activation function. This represents the weight matrix of the underlying autoencoder. This represents the bias vector of the lower-level autoencoder; the upper-level autoencoder then further processes and fuses the primary features, and the mapping relationship of the upper-level autoencoder is as follows: ,in This represents the output features of the k-th layer autoencoder. This represents the weight matrix of the k-th layer autoencoder. This represents the bias vector of the k-th layer autoencoder, ultimately yielding more representative deep visual features.

[0101] During training, the activation states of neurons in the hidden layer are constrained by a sparse constraint regularization term, which is: ,in Represents the regularization coefficient. Indicates the number of neurons in the hidden layer. This represents the preset sparsity parameter. This represents the average activation probability of the j-th hidden layer neuron. KL divergence is used to measure... The differences in these features cause most neurons to be in an inhibited state, with only a few neurons activated. This ensures that the extracted deep visual features are sparse and representative, effectively distinguishing different types of targets. The output of the stacked sparse autoencoder is the deep visual feature of the suspicious target. ,in The dimension representing deep visual features is a fixed-dimensional feature vector that contains the core visual feature information of suspicious targets and can characterize key attributes such as the shape and outline of suspicious targets.

[0102] The calculation of spatial position parameters and motion state parameters is based on a preset coordinate system, which adopts the UTM coordinate system. This coordinate system has a fixed point on the coastline as the origin, with the positive x-axis pointing east, the positive y-axis pointing north, and the positive z-axis pointing perpendicular to the sea level. First, the pixel position of the suspected target in the inspection image is used... Combined with the positioning information of unmanned system nodes Sensor parameters (camera focal length) Shooting angle (This involves) transforming pixel coordinates into three-dimensional spatial coordinates in a preset coordinate system through perspective projection transformation. The formula for perspective projection transformation is:

[0103]

[0104] in, The pixel coordinates of the camera's optical center on the image are used to obtain the spatial location parameters of the suspected target. ,in The altitude or water depth corresponding to the target.

[0105] Motion state parameters are calculated from the spatial position parameters of consecutive time steps, selecting two adjacent time steps. and Spatial location parameters and Calculate the speed of movement of the suspicious target and direction of motion The formula for velocity is: .in, This represents the time interval between two time steps; the formula for the direction of motion is... , representing the angle between the line connecting two adjacent spatial positions and the positive x-axis of the coordinate system. Integrating the spatial position parameters with the motion state parameters yields the motion characteristics. , which is a fixed-dimensional vector, can reflect the real-time movement status of suspicious targets and provide movement dimension support for subsequent threat level determination.

[0106] S24. The collaborative situational awareness features, the deep visual features of suspicious targets, and the motion features are concatenated into vectors to obtain a fused feature vector.

[0107] Specifically, collaborative situational awareness characteristics The collaborative fusion processing from a distributed heterogeneous unmanned system cluster includes multi-dimensional target information perceived by multiple nodes, which can overcome the limitations of single-node perception; deep visual features of suspicious targets. Derived from a stacked sparse autoencoder, focusing on the visual morphological features and motion features of the target. Reflecting the spatial motion state of the target, these three features characterize the attributes of the suspicious target from different dimensions. Using any one feature alone cannot fully and accurately identify the suspicious target and determine the threat level. By using vector concatenation to achieve multi-feature fusion, the advantages of features from various dimensions can be integrated to improve the discriminative ability of the features.

[0108] Vector concatenation employs a column-wise concatenation method, sequentially merging the collaborative situational awareness feature vector, the deep visual feature vector of suspicious targets, and the motion feature vector, thus fusing the feature vectors. ,in , , These represent the transposes of the three feature vectors, and the dimension of the fused feature vector. ,in Dimensions representing the characteristics of collaborative situational awareness. Dimensions representing deep visual features This represents the dimension of the motion features. During the concatenation process, ensure that the time steps of the three feature vectors are consistent, i.e., they all correspond to the same time step. To identify suspicious target features, the feature vector is normalized to avoid biases in feature fusion due to time step mismatch. After concatenation, the fused feature vector is normalized using the following formula: ,in This represents the normalized fused feature vector. This represents the minimum value of each dimension of the fused feature vector. This represents the maximum value of each dimension of the fused feature vector. The values ​​of each dimension of the vector are mapped to the interval [0,1] to eliminate the difference in magnitude between features of different dimensions and ensure the computational stability of the subsequent multilayer perceptron classifier.

[0109] S25. Input the fused feature vector into the multilayer perceptron classifier. The fused feature vector is nonlinearly transformed through the fully connected layer of the multilayer perceptron classifier to obtain the hidden layer features. The hidden layer features are then mapped to threat probability and non-threat probability through the output layer of the multilayer perceptron classifier.

[0110] Specifically, a multilayer perceptron classifier consists of an input layer, several hidden layers, and an output layer. The dimension of the input layer is the same as the dimension of the fused feature vector. Consistent, used to receive fused feature vectors The hidden layer uses the ReLU activation function to achieve a non-linear transformation of the fused feature vector. The output formula of the k-th hidden layer is as follows: ,in This represents the output feature of the k-th hidden layer. This represents the weight matrix of the k-th hidden layer. This represents the bias vector of the k-th hidden layer. The number of neurons in each hidden layer is reasonably set according to the feature dimension and classification requirements. By gradually reducing the feature dimension, more discriminative hidden layer features are extracted. The output layer uses the Softmax activation function to map the hidden layer features to a probability distribution. The output formula is:

[0111]

[0112] in, This represents the output probability vector. The threat probability corresponds to the first component of the output vector, representing the probability that the target is a threat target; This is the non-threat probability, corresponding to the second component of the output vector, representing the probability that the target is a non-threat target; This represents the weight matrix of the output layer, with dimension 1. , This represents the output feature dimension of the last hidden layer, used to map hidden layer features to the output dimension; This represents the bias vector of the output layer, with a dimension of 2×1, and is used to adjust the baseline value of the output probability; Represents the output layer weight matrix The first row corresponds to the mapping weights of the threat probability; Represents the output layer weight matrix The second row corresponds to the mapping weights for the non-threat probability; Represents the output layer bias vector The first element corresponds to the bias adjustment of the threat probability; Represents the output layer bias vector The second element corresponds to the bias adjustment of the non-threat probability; This represents the output hidden layer feature of the last hidden layer (the Lth layer), which is the input feature of the output layer; This indicates the total number of hidden layers.

[0113] The training process of the multilayer perceptron classifier is supervised by a labeled dataset, which includes fused feature vectors and their corresponding labels. (Threat target correspondence) Non-threat target correspondence The error between the model's predicted value and the true label is calculated using the cross-entropy loss function. The formula for the loss function is as follows: ,in This represents the j-th component of the real label. The j-th component represents the predicted probability of the model. The model parameters are iteratively optimized using gradient descent, and the gradient update formula is: , ,in Indicates the learning rate. Represents the loss function with respect to the weight matrix. and bias vector The gradient is calculated until the model converges, ensuring that the model can accurately map the relationship between fused features and threat and non-threat probabilities.

[0114] S26. When the threat probability is greater than the preset threat threshold, the target in the inspection image is determined to be a suspicious target, and the corresponding threat level is assigned to the suspicious target based on the difference between the threat probability and the preset threat threshold.

[0115] Specifically, preset threat thresholds This setting, based on the security requirements of coastline patrol scenarios, is used to distinguish between threatening and non-threatening targets; its value range is [value range missing]. During the determination process, the threat probability output by the multilayer perceptron classifier is used. Compared with preset threat threshold If a comparison is made, If the target is identified as suspicious, further threat level classification is required; if If the target is not identified as a threat, it is determined to be non-threatening and no further threat level assessment is required. Threat levels are determined based on the difference between the threat probability and a preset threat threshold. Based on the core evidence, among which This indicates the degree to which the threat probability exceeds a threshold. The larger the difference, the higher the threat level of the target and the higher the corresponding threat level. By setting multiple difference thresholds, different threat levels can be divided, and each difference range corresponds to a specific threat level.

[0116] Specifically, two difference thresholds are set. and (satisfy ), the difference Divided into three intervals: when When, it corresponds to the lowest threat level; when At this time, it corresponds to a medium threat level; when At that time, it corresponds to the highest threat level. , The preset difference threshold is set according to actual security needs and experimental data. This method enables the quantitative classification of the threat level of suspicious targets.

[0117] In one optional embodiment, based on the threat level and the topological location of each unmanned system node in the distributed communication network corresponding to the distributed heterogeneous unmanned system cluster, the node status value of each unmanned system node is calculated, including the following steps:

[0118] S31. Obtain the suspicious targets observed by each unmanned system node at the current time step, query the preset threat weight mapping table according to the threat level of each suspicious target, and obtain the corresponding threat weight; sum up the corresponding threat weights of all suspicious targets to obtain the node importance.

[0119] Specifically, the current time step In this process, each unmanned system node, through its own perception module and distributed collaborative fusion processing, identifies all suspicious targets it observes and records them as nodes. At time step The observed set of suspicious targets is ,in For nodes At time step The number of suspicious targets observed, Represents a node At time step The observed first A suspicious target.

[0120] The pre-defined threat weight mapping table, based on the security requirements of coastline patrol scenarios and calibrated through experimental data and expert experience, establishes a one-to-one correspondence between the threat level of suspicious targets and their corresponding threat weights. The higher the threat level, the greater the corresponding threat weight, indicating a higher degree of threat posed by the suspicious target to coastline security and a higher importance for the node observing that target. Threat levels are categorized into... Levels (e.g., low, medium, high, very high) have corresponding threat weights of [missing information]. And satisfy Threat weight mapping table is denoted as ,in Indicates the first Threat level.

[0121] For nodes Each suspicious target observed According to its corresponding threat level Query the threat weight mapping table The threat weight corresponding to the suspicious target is obtained. Node importance The calculation formula is obtained by summing the threat weights of all suspicious targets. ,in Represents a node At time step The importance of nodes, Represents a node At time step The observed first Threat weight of each suspicious target, For nodes At time step The number of suspicious targets observed. If the node If no suspicious targets are observed at the current time step, then the node importance level is... This indicates that the node currently contributes little to the safety and protection of the coastline.

[0122] S32. Based on the link quality between all nodes in the distributed communication network at the current time step, calculate the node centrality of each unmanned system node in the distributed communication network.

[0123] Specifically, the topology of distributed communication networks is modeled using graph theory. It means that among them For the set of all unmanned system nodes, For time step A set of communication links between downstream nodes; the existence of a link is determined by its quality. Link quality. Used for quantization nodes With nodes The reliability of the communication link between them is calculated by weighting three core indicators: packet loss rate, transmission delay, and signal strength. The calculation formula is as follows: .in, Indicates time step Next node With nodes The link quality between them, with a value range of [0,1]; Indicates time step Next node With nodes Packet loss rate in communication between them; Indicates time step Next node With nodes Communication transmission delay between them; This indicates the preset maximum allowable transmission delay; Indicates time step Next node With nodes The strength of the communication signals between them; Indicates the preset maximum signal strength; The weighting coefficients for each link quality indicator satisfy... It is set according to the performance requirements of the communication network.

[0124] Set link quality threshold ,when At that time, determine the node With nodes There is a valid communication link between them, i.e. ;when At that time, determine the node With nodes There is no effective communication link between them, that is Node centrality Used for quantization nodes In distributed communication networks, topological importance is calculated using weighted degree centrality, taking into account the impact of link quality. The formula is as follows: .in, Represents a node At time step The node centrality of has a value range of [0,1]. Represents a node At time step The set of neighboring nodes, that is, the set of nodes with which the node is located. The set of nodes with valid communication links; Represents a node with neighboring nodes Link quality between; denominator This represents the maximum sum of the neighbor link quality of all nodes in the network. It is used to normalize the node centrality to the interval [0,1]. The higher the node centrality, the more critical the node's topological position in the network, the stronger its information interaction capability, and the greater its contribution to global collaborative perception.

[0125] S33. Collect the operational status data of each unmanned system node, normalize the operational status data to obtain normalized operational status data; perform weighted fusion of the normalized operational status data to obtain the perception health factor of each unmanned system node; wherein, the operational status data includes sensor status data, computing load data and communication packet loss rate data.

[0126] Specifically, the operational status data is collected using a real-time sampling method, with the sampling frequency and time step... Maintain consistency. Sensor status data. Used for quantization nodes The reliability of the sensor is characterized by its recognition accuracy, which is calculated by comparing historical sensing data of the node with real labeled data. The value range is [0,1]. The closer the value is to 1, the better the sensor is working and the higher the accuracy of the sensed data.

[0127] Calculate load data Used for quantization nodes The computational resource usage is represented by CPU utilization, i.e. The value range is [0,1]. The closer a value is to 0, the lower the node's computational load and the stronger its data processing capability; the closer a value is to 1, the higher the node's computational load and the more likely it is to experience data processing delays, stuttering, and other problems.

[0128] Communication packet loss rate data Used for quantization nodes Communication reliability, i.e. The value range is [0,1]. The closer the value is to 0, the better the node's communication status and the higher the integrity of data transmission.

[0129] Because the magnitude and meaning of the three types of operational status data differ, normalization processing is required to eliminate the influence of dimensionality and ensure the rationality of subsequent weighted fusion. Sensor status data The data is already in the [0,1] interval and requires no additional normalization; the load data is calculated using inverse normalization, as shown in the formula. ,in This represents the normalized computational load data, with values ​​ranging from [0,1]. The closer the value is to 1, the better the node's computational state; the communication packet loss rate data uses inverse normalization, and the formula is... ,in This represents the normalized packet loss rate data, with values ​​ranging from [0,1]. The closer it is to 1, the better the node's communication status.

[0130] Perceived health factors The data is obtained by weighted fusion of three types of normalized operational status data, and the fusion formula is as follows: ,in Represents a node At time step The perceived health factor has a value range of [0,1]. The closer it is to 1, the better the node's operating status and the higher the perception reliability; The weighting coefficients for the three types of normalized operating state data satisfy the following conditions: ,in It has the highest weight because the sensor status directly determines the accuracy of the sensed data and is the core factor affecting the reliability of node sensing. and Configure the nodes reasonably based on their computing and communication requirements.

[0131] S34. Calculate the node status value based on node importance, node centrality, and perception health factor; the formula for calculating the node status value is as follows:

[0132]

[0133] in, Represents a node At time step The node status value, Represents a node At time step Perceived health factors Represents a node At time step The importance of nodes, Represents a node At time step The node centrality, The first weighting coefficient represents the importance of a node. The second weighting coefficient represents the node centrality.

[0134] Specifically, node status values It is a comprehensive quantitative result of node importance, node centrality, and perceived health factor, with a value range of [0,1]. The closer a node is to 1, the better its overall performance at the current time step, the higher its perceived reliability and importance, and the higher the weight of its perceived data and features in the distributed collaborative fusion processing should be; conversely, The closer the value is to 0, the worse the overall performance of the node and the lower the reliability of its perception. Therefore, its information fusion weight should be reduced to avoid interfering with the overall situational awareness.

[0135] Represents a node At time step The perception health factor, calculated by step S33, is used to characterize the node's own operating status and perception reliability. As a basic correction factor for the node's status value, it ensures that only nodes with good operating status can fully reflect their importance and centrality. The first weighting coefficient represents the importance of a node. The second weighting coefficient represents the node centrality; both are preset constants that satisfy... , and ,in Because the importance of a node is directly related to the threat of a suspected target, it has a greater impact on the security of the coastline and is the core influencing factor of the node situation value. Represents a node At time step The importance of a node is calculated by step S31 and characterizes the threat correlation of a node observing suspicious targets. Represents a node At time step The node centrality, calculated by step S32, characterizes the topological importance of a node in a distributed communication network.

[0136] During the calculation, the weighted sum of node importance and node centrality is first calculated. This weighted sum represents the "potential importance" of a node, that is, its importance without considering the node's own operational status; subsequently, this weighted sum is compared with the perceived health factor. Multiplication corrects the potential importance of nodes, ensuring that the node status value truly reflects the overall performance of the node. It considers both the threat correlation and topological location of the node, as well as the operational reliability of the node itself, providing accurate quantitative basis for subsequent system optimization.

[0137] In one optional embodiment, adjusting the information fusion weights of corresponding unmanned system nodes based on node status values ​​in the distributed collaborative fusion processing includes the following steps:

[0138] S41. Obtain the node status values ​​of all unmanned system nodes at the current time step, input the status values ​​of each node into the sigmoid function, and obtain the confidence scaling factor corresponding to each unmanned system node.

[0139] S42. Adjust the attention score according to the credibility scaling factor to obtain the adjusted attention score; where the expression for the adjusted attention score is:

[0140]

[0141] in, Represents a node At time step For nodes Attention score Represents a node At time step The query vector, Represents a node At time step The key vector, The dimension of the key vector. Indicates a node Node status values The sigmoid function that performs the transformation. Represents a node of The node index of each neighboring node. .

[0142] Specifically, the current time step The following step, S34, yields the set of node status values ​​for all unmanned system nodes. ,in This represents the total number of nodes in a distributed heterogeneous unmanned system cluster. Represents a node At time step The node status value ranges from [0,1]. A higher value indicates a stronger node status. The better the overall performance, the higher the perceived reliability and importance.

[0143] The credibility scaling factor is used to map the node situation value to a coefficient suitable for correcting the attention score, so as to achieve precise control of the information fusion weight by the node situation value. It is obtained by transforming the node situation value through the sigmoid function. The sigmoid function has non-linear mapping capability, which can normalize the node situation value to the (0,1) interval, while amplifying the difference between nodes with medium and high situation values ​​and nodes with low situation values, so as to ensure that the corrected attention score can effectively distinguish nodes with different reliability.

[0144] node At time step The confidence scaling factor is the sigmoid function applied to the node state value. The transformation result is denoted as The expression for the sigmoid function is: ,in This is the gain coefficient, used to adjust the slope of the sigmoid function and amplify the differences in nodal situation values. , The larger the value, the greater the slope of the function, and the higher the distinction between medium-high situation values ​​and low situation values. This is a threshold parameter used to adjust the function's offset, ensuring the function output matches the distribution of node situation values. We can take 0.5 to ensure that when the node status value is 0.5, the confidence scaling factor is 0.5, thus achieving neutral correction.

[0145] When the node status value hour, A credibility scaling factor greater than 0.5 will positively enhance the attention score. The weight of the perceived features is increased during the fusion process; when the node's situation value hour, A credibility scaling factor less than 0.5 will have a negative inhibitory effect on the attention score, affecting the node. The weight of the perceptual features is reduced during the fusion process; when hour, It does not modify the attention score, keeping the original attention score unchanged, thus ensuring the rationality and relevance of the modification logic.

[0146] This correction method ensures that the attention score depends not only on the feature similarity between nodes but also on the node's overall performance (node ​​status value), enabling dynamic adjustment of information fusion weights. When a node... When the node status value is high, its confidence scaling factor Larger, corrected attention score Corresponding improvements, nodes The weight of perception features in collaborative situational awareness feature fusion increases; when nodes When the node's state value is low, its confidence scaling factor is small, and the corrected attention score decreases accordingly. The weights of perceived features are suppressed, thereby improving the accuracy of distributed collaborative fusion processing and avoiding the impact of interference from low-reliability nodes on global situational awareness.

[0147] The aforementioned situational awareness method for coastline patrol scenarios targeting suspicious target intrusion firstly integrates the temporal perception data streams collected by each unmanned system node through distributed collaborative fusion processing. This allows each node to obtain collaborative situational awareness features containing its own observations and information from neighboring nodes, providing spatiotemporal correlation information from a global perspective for subsequent identification. Based on this, target identification and threat level determination are performed by combining patrol images with the collaborative situational awareness features, generating suspicious targets and their situational description vectors. This achieves accurate location and threat quantification of suspicious targets in complex coastline environments. Furthermore, the situational value of each node is calculated based on the threat level and its topological position in the distributed communication network, quantifying the importance and reliability of each node in the perception task. A global situational inconsistency index is calculated by combining the situational description vectors of all nodes for the same target, enabling dynamic monitoring of the overall system perception consistency. Finally, when a node's situational value falls below a preset threshold, its information fusion weight in collaborative fusion is dynamically adjusted to suppress the negative impact of unreliable nodes. Simultaneously, misjudged samples selected using the global situational inconsistency index are used to incrementally learn and update the identification model, enabling the system to continuously adapt to environmental changes and correct its own perception biases. Through the above methods, this technical solution achieves a complete closed loop from the collaborative fusion of multi-source heterogeneous information, accurate target identification and threat assessment, quantitative evaluation of node and global situation, to adaptive adjustment of fusion weights and online iterative optimization of the model. It effectively solves the problems of monitoring blind spots, target identification difficulties and slow response mechanisms in traditional perception methods under long-distance coastlines, complex terrain and severe weather conditions. It significantly improves the collaborative perception accuracy, robustness and adaptability of distributed heterogeneous unmanned systems in suspicious target intrusion scenarios.

[0148] It should be understood that although the steps in the flowcharts of the embodiments described above are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the embodiments described above may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.

[0149] Based on the same inventive concept, this application also provides a system for implementing the aforementioned situational awareness method for coastline patrol scenarios involving suspicious target intrusion. The solution provided by this system is similar to the implementation scheme described in the above method. Therefore, the specific limitations of one or more embodiments of the situational awareness system for coastline patrol scenarios involving suspicious target intrusion provided below can be found in the limitations of the situational awareness method for coastline patrol scenarios involving suspicious target intrusion described above, and will not be repeated here.

[0150] In one exemplary embodiment, such as Figure 3 As shown, a situational awareness system 30 for coastline patrol scenarios targeting suspicious intrusion is provided to implement the methods in the above-described embodiments. The system includes:

[0151] The distributed situation fusion module 31 is used to perform distributed collaborative fusion processing on the time-series perception data streams collected by each unmanned system node based on the distributed heterogeneous unmanned system cluster deployed in the coastline inspection area, so as to obtain the collaborative situation perception characteristics of each unmanned system node at its respective time step.

[0152] The target situation generation module 32 is used to identify targets and determine their threat levels in the inspection images collected by each unmanned system node and the collaborative situational awareness features, to obtain suspicious targets and their corresponding threat levels, and to generate situational description vectors based on the position and motion state of the suspicious targets in the inspection images.

[0153] The node situation quantification module 33 is used to calculate the node situation value of each unmanned system node based on the threat level and the topological position of each unmanned system node in the distributed communication network corresponding to the distributed heterogeneous unmanned system cluster.

[0154] The situation consistency analysis module 34 is used to calculate the global situation inconsistency index based on the situation description vector of each unmanned system node for the same suspicious target.

[0155] The adaptive optimization module 35 is used to adjust the information fusion weight of the corresponding unmanned system node according to the node status value in the distributed collaborative fusion processing when the node status value is lower than the preset confidence threshold. It also filters out the misjudged samples corresponding to the low node status value unmanned system nodes that cause status inconsistency based on the global status inconsistency index, and uses the misjudged samples to incrementally learn and update the processing model for identification and threat level determination.

[0156] Embodiments of this application also provide a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps in the aforementioned method embodiments.

[0157] Embodiments of this application also provide a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps in the above-described method embodiments.

[0158] For the device embodiments, since they basically correspond to the method embodiments, the relevant parts can be referred to in the description of the method embodiments. The device embodiments described above are merely illustrative. The components described as separate parts may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this disclosure according to actual needs. Those skilled in the art can understand and implement this without creative effort.

[0159] The above-described embodiments are merely illustrative of several implementation methods of the embodiments of this application, and their descriptions are relatively specific and detailed. However, they should not be construed as limiting the scope of the patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the embodiments of this application, and these modifications and improvements all fall within the protection scope of the embodiments of this application.

Claims

1. A method for situation awareness of a coastline patrol scenario facing suspicious target intrusion, characterized in that, The method includes: S1. Based on the distributed heterogeneous unmanned system cluster deployed in the coastline inspection area, the time-series perception data streams collected by each unmanned system node are processed in a distributed collaborative manner to obtain the collaborative situational awareness features of each unmanned system node at its respective time step. S2. Based on the inspection images collected by each of the unmanned system nodes and the collaborative situational awareness features, identify and determine the threat level of targets in the inspection images to obtain suspicious targets and their corresponding threat levels, and generate a situational description vector based on the position and motion state of the suspicious targets in the inspection images. S3. Based on the threat level and the topological location of each unmanned system node in the distributed communication network corresponding to the distributed heterogeneous unmanned system cluster, calculate the node status value of each unmanned system node. S4. Calculate the global situation inconsistency index based on the situation description vector of each unmanned system node for the same suspected target; S5. When the node situation value is lower than the preset confidence threshold, the information fusion weight of the corresponding unmanned system node is adjusted according to the node situation value in the distributed collaborative fusion processing, and the misjudged samples corresponding to the low node situation value unmanned system nodes that cause situation inconsistency are screened according to the global situation inconsistency index. The misjudged samples are used to incrementally learn and update the processing model for the identification and threat level determination.

2. The method of claim 1, wherein, The distributed heterogeneous unmanned system cluster deployed within the coastline inspection area performs distributed collaborative fusion processing on the time-series sensing data streams collected by each unmanned system node to obtain the collaborative situational awareness features of each unmanned system node at its respective time step, including: S11. At each unmanned system node, the sensor data sequence collected within a preset time window is input into a bidirectional gating loop unit. The forward hidden state is extracted through the forward gating loop unit, and the backward hidden state is extracted through the backward gating loop unit. The forward hidden state and the backward hidden state are spliced ​​together to obtain the local spatiotemporal dynamic features of each unmanned system node. S12. Through the distributed communication network, the local spatiotemporal dynamic characteristics corresponding to each of the unmanned system nodes are broadcast to neighboring nodes within the communication range; S13. For each unmanned system node, generate a corresponding query vector, key vector, and value vector based on its own local spatiotemporal dynamic characteristics, and generate a corresponding key vector and value vector based on the local spatiotemporal dynamic characteristics of all neighboring nodes. S14. For each of the unmanned system nodes, calculate the attention score of the unmanned system node to each computed node based on its own query vector and the key vector of each computed node; wherein, the computed nodes include the unmanned system node itself and all its neighboring nodes; wherein, the formula for calculating the attention score is: wherein, denotes the attention score of node at time step , denotes the attention score of node , denotes the query vector of node at time step , denotes the key vector of node at time step , denotes the dimension of the key vector, denotes the node index of the neighbor nodes of node ; S15. For each of the unmanned system nodes, the value vectors of itself and each neighbor node are weighted and summed according to the attention score to obtain the cooperative situational awareness feature that integrates neighbor information; wherein, the calculation formula of the cooperative situational awareness feature is: wherein, representing a node at a time step the co-situational awareness feature, representing a node the attention score for a node representing a node representing a node a value vector at a time step .

3. The method according to claim 1, characterized in that, The step of identifying and determining the threat level of targets in the inspection images based on the inspection images collected by each of the unmanned system nodes and the collaborative situational awareness features, to obtain suspicious targets and their corresponding threat levels, includes: S21. Input the inspection images collected by each unmanned system node into the semantic segmentation network, and perform pixel-level classification on the inspection images through the semantic segmentation network to obtain a segmentation mask containing the category of suspicious targets. S22. Extract pixel regions marked as suspicious target categories from the segmentation mask, and crop corresponding image sub-regions in the inspection image according to the pixel regions; perform size adjustment and flattening processing on the image sub-regions to obtain suspicious target image block vectors; S23. Input the image block vector of the suspected target into a stacked sparse autoencoder, extract deep features through the stacked sparse autoencoder to obtain the deep visual features of the suspected target; calculate the spatial position parameters and motion state parameters of the suspected target in the preset coordinate system according to the pixel position of the suspected target in the inspection image to obtain the motion features. S24. The collaborative situational awareness features, the deep visual features of the suspicious target, and the motion features are concatenated to obtain a fused feature vector; S25. Input the fused feature vector into a multilayer perceptron classifier, perform a nonlinear transformation on the fused feature vector through the fully connected layer of the multilayer perceptron classifier to obtain hidden layer features, and map the hidden layer features into threat probability and non-threat probability through the output layer of the multilayer perceptron classifier. S26. When the threat probability is greater than the preset threat threshold, the target in the inspection image is determined to be the suspicious target, and the threat level corresponding to the suspicious target is assigned according to the difference between the threat probability and the preset threat threshold.

4. The method according to claim 2, characterized in that, The step of calculating the node status value of each unmanned system node based on the threat level and the topological location of each unmanned system node in the distributed communication network corresponding to the distributed heterogeneous unmanned system cluster includes: S31. Obtain the suspicious targets observed by each unmanned system node at the current time step, query the preset threat weight mapping table according to the threat level of each suspicious target, and obtain the corresponding threat weight; sum up the corresponding threat weights of all the suspicious targets to obtain the node importance. S32. Calculate the node centrality of each unmanned system node in the distributed communication network based on the link quality between all nodes in the distributed communication network at the current time step. S33. Collect the operating status data of each of the unmanned system nodes, normalize the operating status data to obtain normalized operating status data; perform weighted fusion on the normalized operating status data to obtain the perception health factor of each of the unmanned system nodes; wherein, the operating status data includes sensor status data, computing load data and communication packet loss rate data. S34. Calculate the node status value based on the node importance, node centrality, and perceived health factor; wherein the formula for calculating the node status value is: in, Represents a node At time step The node situation value, Represents a node At time step The perceived health factor, Represents a node At time step The importance of the nodes mentioned above, Represents a node At time step The node centrality, The first weighting coefficient represents the importance of the node. The second weighting coefficient represents the node centrality.

5. The method according to claim 4, characterized in that, The step of adjusting the information fusion weight of the corresponding unmanned system node based on the node status value in the distributed collaborative fusion processing includes: S41. Obtain the node status values ​​of all the unmanned system nodes at the current time step, input each node status value into the sigmoid function, and obtain the confidence scaling factor corresponding to each unmanned system node. S42. The attention score is corrected according to the credibility scaling factor to obtain the corrected attention score; wherein, the expression for the corrected attention score is: in, Represents a node At time step For nodes The attention score, Represents a node At time step The query vector, Represents a node At time step The key vector, The dimension of the key vector. Indicates a node Node status values The sigmoid function that performs the transformation. Represents a node of The node index of each neighboring node. .

6. A situational awareness system for coastline patrol scenarios targeting suspicious target intrusion, used to implement the method according to any one of claims 1 to 5, characterized in that, The system includes: The distributed situation fusion module is used to perform distributed collaborative fusion processing on the time-series perception data streams collected by each unmanned system node based on a distributed heterogeneous unmanned system cluster deployed in the coastline inspection area, so as to obtain the collaborative situation perception features of each unmanned system node at its respective time step. The target situation generation module is used to identify and determine the threat level of targets in the inspection images based on the inspection images collected by each of the unmanned system nodes and the collaborative situational awareness features, to obtain suspicious targets and their corresponding threat levels, and to generate a situational description vector based on the position and motion state of the suspicious targets in the inspection images. The node situation quantification module is used to calculate the node situation value of each unmanned system node based on the threat level and the topological position of each unmanned system node in the distributed communication network corresponding to the distributed heterogeneous unmanned system cluster. The situation consistency analysis module is used to calculate a global situation inconsistency index based on the situation description vector of each unmanned system node for the same suspected target. An adaptive optimization module is used to adjust the information fusion weight of the corresponding unmanned system node according to the node status value in the distributed collaborative fusion processing when the node status value is lower than a preset confidence threshold, and to filter out the misjudged samples corresponding to the unmanned system nodes with low node status values ​​that cause status inconsistency according to the global status inconsistency index, and to use the misjudged samples to incrementally learn and update the processing model for identification and threat level determination.

7. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the method of any one of claims 1 to 5.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the method of any one of claims 1 to 5.

Images