An Unmanned Cluster Cooperative Sensing Method and System Based on Unsupervised Hashing Learning

Abstract

The present invention belongs to the technical field of unmanned cluster control, and particularly relates to an unmanned cluster collaborative perception method and system based on unsupervised hashing learning, including: acquiring data in the unmanned cluster system architecture, applying an unsupervised hashing learning algorithm to perform hashing dimensionality reduction processing on the data, and generating low-dimensional hash codes capable of characterizing the similarity between data; applying a distributed approximate clustering algorithm to randomly sample and divide the low-dimensional hash codes, and forming fine approximate clustering centers and clustering clusters through a clustering ensemble method; using an improved DBSCAN algorithm for each cluster to form data block grouping based on the nearest neighbor; and supporting collaborative perception among unmanned cluster nodes by using the low-dimensional hash codes and / or clustering or grouping. The present invention has outstanding advantages in terms of the high efficiency and reliability of data transmission, greatly improves the overall task execution efficiency, and significantly improves the energy consumption.

Description

Technical Field

[0001] The present invention belongs to the technical field of unmanned cluster control, and particularly relates to a method and system for collaborative perception of unmanned clusters based on unsupervised hashing learning. Background Art

[0002] An unmanned cluster is usually composed of a variety of relatively simple unmanned platforms, which has the characteristic of large scale. Each platform individual realizes global collaborative behavior through local cooperation. However, with the increase in the number of unmanned platforms and the improvement of the requirements for data transmission security, how to achieve high-speed and secure data transmission in complex situations is a problem that needs to be considered at present. For the hierarchical unmanned cluster scenario combining distribution and concentration, it faces multiple challenges in various aspects of the "perception - judgment - decision - action" loop in a complex environment. The key to using unmanned cluster devices for efficient collaborative perception of the environmental situation lies in how to efficiently compress and map various types of modal perception data, so as to achieve high-speed and secure transmission.

[0003] When unmanned cluster systems (such as unmanned aerial vehicle swarms, unmanned vehicle swarms) perform complex tasks such as reconnaissance, mapping, search and rescue, logistics, and collaborative combat, they will generate and interact with a large amount of multi-modal perception data (such as images, videos, point clouds, sensor readings, etc.). How to efficiently store, transmit, and retrieve these data, and on this basis, conduct effective collaborative perception and decision-making, is the key challenge faced by current unmanned cluster technologies. Unmanned cluster collaborative perception mainly includes system positioning technology, collaborative target tracking technology, and system target recognition technology, etc. Only by organizing various technical elements and unsupervised hashing learning in a reasonable way can the greater efficiency of the cluster be exerted. Therefore, the architecture is an important method for constructing an unmanned cluster system, which can significantly improve the task execution efficiency of the cluster.

[0004] However, the existing technologies have the following defects: low data transmission efficiency: the large amount of multi-modal data generated by unmanned clusters leads to large storage and transmission overheads and high response delays; low retrieval efficiency: it is very difficult to quickly find the required information from unstructured massive data; insufficient system reliability: in a decentralized architecture and a dynamic environment, node failures or data losses easily lead to system instability, and there is a lack of an effective fault self-healing mechanism; difficult heterogeneous data fusion: multi-modal data is difficult to be uniformly processed, and cross-modal relevance and spatio-temporal consistency are difficult to retain; poor adaptability to resource constraints: high-precision model calculation complexity is high, and it is difficult to be deployed on resource-constrained nodes; insufficient adaptability to dynamic environments: high task dynamics and frequent changes in node states lead to uneven load and large energy consumption.

[0005] Since unsupervised hashing learning algorithms can meet the special requirements of large-scale data retrieval for data storage space and retrieval accuracy, but there is currently a lack of practical applications of hashing learning in the field of unmanned cluster collaborative perception. Traditional hashing algorithms are mostly supervised or static methods, which are difficult to preserve the similarity of complex data; clustering algorithms are difficult to process massive data; consistent hashing lacks a dynamic adjustment mechanism. These deficiencies limit the collaborative perception ability of unmanned clusters in dynamic environments. There is an urgent need for a method for processing and managing unmanned cluster data that can efficiently process massive heterogeneous data, adapt to dynamic environments, and support intelligent collaborative perception. Summary of the Invention

[0006] Object of the Invention: The object of the present invention is to provide an unmanned cluster collaborative perception method and system based on unsupervised hashing learning, which has outstanding advantages in the efficiency and reliability of data transmission, greatly improves the efficiency of overall task execution, and has obvious improvement in energy consumption.

[0007] Technical Solution: The unmanned cluster collaborative perception method based on unsupervised hashing learning according to the present invention includes the following steps:

[0008] (1) Apply an unsupervised similarity adaptive hashing algorithm to perform hashing dimensionality reduction processing on data blocks in the unmanned cluster system to generate low-dimensional hash codes that can preserve the similarity between data blocks;

[0009] (2) Apply a distributed approximate clustering algorithm to the low-dimensional hash codes to obtain the clustering results of the low-dimensional hash codes, where the clustering results include approximate clustering centers and clustering clusters;

[0010] (3) Apply an improved DBSCAN algorithm to group the clustering clusters to obtain data block groups, control the number of data blocks in each group not to exceed a preset threshold, and apply a data compression algorithm to the data block groups to generate compressed groups;

[0011] (4) Use one or a combination of the low-dimensional hash codes, approximate clustering centers, clustering clusters, and compressed groups to perform at least one of the following operations in the unmanned cluster system to support collaborative perception:

[0012] (401) Transmit the formed approximate clustering centers or compressed groups among the nodes of the unmanned cluster system to share perception information among the nodes;

[0013] (402) In the unmanned cluster system, transmit the generated low-dimensional hash codes among the nodes, and the receiving node performs nearest neighbor search based on the received hash codes through the Hamming distance calculation method to achieve the association or identification of the targets observed by other nodes;

[0014] (403) Use the low-dimensional hash code as the representation of the node perception state or environmental features, perform state estimation in a distributed computing manner, and achieve collaborative decision-making based on the estimation results and task requirements.

[0015] Further improve the above technical solution. The unsupervised similarity adaptive hashing algorithm in step (1) includes:

[0016] Use the Euclidean loss function to measure the difference between the output of the hash function and the target hash code learned in the previous iteration, and optimize the parameters of the hash function using backpropagation and stochastic gradient descent;

[0017] Based on the feature representation corresponding to the currently trained hash function, update the similarity between data blocks, and generate an updated similarity graph representing the similarity structure;

[0018] Based on the updated similarity graph, construct a graph Laplacian matrix, and use the graph hashing optimization method to generate an updated target hash code as the target for the next iteration, where the graph Laplacian matrix is calculated using the anchor graph method;

[0019] After the iterative optimization loop ends, use the optimized learnable hash function to process the data blocks to generate the low-dimensional hash code.

[0020] Further, the use of the Euclidean loss function to measure the difference between the output of the hash function and the hash code learned in the previous iteration is expressed as follows: ,

[0021] where F is the hash function, X is the input data block, is the set of learnable hash function parameters, is the weight parameter, is the target hash code;

[0022] The updated similarity graph is expressed as follows:

[0023] ;

[0024] where, is the feature representation model, is the bandwidth parameter, represents the similarity between data pairs in the input feature space, are the fixed hash function parameters after training;

[0025] The graph hashing optimization method is expressed as follows:

[0026] ;

[0027] where, is the trace operation, representing the similarity loss between the hash matrix B and the graph Laplacian matrix L. The size of the hash matrix B is , r is the length of the hash code, and n is the number of data blocks; is the graph Laplacian matrix, represents the identity matrix, and represent the regularization parameters, is a vector of all 1s with length n, used for balance regularization; is the constraint condition to ensure that the elements of the hash matrix are binary values;

[0028] The graph Laplacian matrix is calculated using the anchor graph method, and the formula is as follows:

[0029] ;

[0030] Among them, is the identity matrix of, is the anchor graph matrix, calculated from the similarity between n data blocks and anchors, is a diagonal matrix, , is a vector of all 1s with length n, is the inverse matrix of, is the transpose matrix of Z.

[0031] Furthermore, the low-dimensional hash codes are randomly divided into subsets of random sample data blocks;

[0032] On each parallel node, a clustering algorithm including K-means is applied to the subset of random sample data blocks to generate parallel node clustering centers;

[0033] The parallel node clustering centers are transmitted to the master node and integrated by the ACEM-MP method to form approximate clustering centers;

[0034] The approximate clustering centers are broadcast to the parallel nodes, and all subsets of random sample data blocks are clustered according to the approximate clustering centers to generate clustering clusters.

[0035] Furthermore, the distributed approximate clustering algorithm used in step (2) includes seven layers of steps, and the first layer to the seventh layer perform the following operations in sequence:

[0036] The first layer: Obtain and store the low-dimensional hash codes obtained after hash dimensionality reduction;

[0037] The second layer: Use the RSP algorithm on the low-dimensional hash codes to form subsets of random sample data blocks;

[0038] The third layer: randomly select a subset of random sample data blocks and load them onto parallel nodes;

[0039] The fourth layer: use the I-niceDP and K-means algorithms on each parallel node to generate the parallel node clustering centers for each subset of random sample data blocks;

[0040] The fifth layer: transfer all the parallel node clustering centers to the master node, and use the ACEM-MP method to integrate the set of clustering centers formed by the parallel node clustering centers as the approximate clustering centers;

[0041] The sixth layer: broadcast the approximate clustering centers to the parallel nodes;

[0042] The seventh layer: use the K-meansOne algorithm to cluster all subsets of random sample data blocks according to the approximate clustering centers to generate clustering clusters.

[0043] Further, the density-based clustering algorithm of the nearest neighbor in step (3) includes the following steps: set the neighborhood radius Eps, the minimum number of points MinPts within the neighborhood radius, and the maximum number of points MaxPts; use the nearest neighbor search method, start from any point, mark it as "visited" and check whether it is a core point; iteratively add the objects in the candidate set that do not belong to other clustering clusters to the current cluster until the candidate set is empty or the number of data in the current cluster reaches MaxPts; randomly select the next unvisited object from the remaining objects and repeat the above process until all objects are visited.

[0044] Further, the method is executed by a hash learning module that interacts with the database management system of the unmanned cluster system. The method also includes storing the processed low-dimensional hash codes, the clustering results, and the compressed groups in the database management system, and using the hash codes to accelerate the search or storage operations of data related to system design or analysis.

[0045] Further, the method is executed by the hash learning algorithm layer in the task hierarchical architecture of the unmanned cluster system. The hash learning algorithm layer is located between the cluster algorithm layer and the data layer. The method also includes using the processed low-dimensional hash codes, the clustering results, and the compressed groups as the medium for data processing and transmission between the cluster algorithm layer and the data layer to improve the processing efficiency of inter-layer data interaction.

[0046] Further, construct a hash index using the low-dimensional hash codes stored in the database management system, and perform content-based fast data retrieval based on the hash index.

[0047] A system for implementing the above-mentioned unsupervised hashing learning-based collaborative perception method for unmanned clusters, comprising:

[0048] A data acquisition module, configured to acquire the perception data collected by the unmanned cluster system and represent the perception data as data blocks;

[0049] An unsupervised hashing processing module, configured to apply an unsupervised similarity adaptive hashing algorithm to perform hashing dimensionality reduction processing on the data blocks to generate low-dimensional hash codes capable of characterizing the similarity between the data blocks;

[0050] A clustering module, configured to apply a distributed approximate clustering algorithm to process the low-dimensional hash codes to obtain a clustering result according to the similarity of the low-dimensional hash codes, where the clustering result includes approximate clustering centers and clustering clusters;

[0051] A grouping processing unit, configured to apply an improved DBSCAN algorithm to the clustering clusters in the clustering result to obtain data block groupings, and control the number of data blocks in each data block grouping not to exceed a preset threshold;

[0052] A compression processing unit, configured to apply the LZ4 compression algorithm to each data block grouping to obtain compressed groupings;

[0053] A collaborative perception support module, configured to use one or a combination of the low-dimensional hash codes, approximate clustering centers, clustering clusters, and compressed groupings to perform at least one of the following operations in the unmanned cluster system to support collaborative perception: transmit the formed approximate clustering centers or compressed groupings among the nodes of the unmanned cluster system to share perception information among the nodes; within the unmanned cluster system, transmit the generated low-dimensional hash codes among the nodes, and the receiving nodes perform nearest neighbor search based on the received hash codes through the Hamming distance calculation method to achieve the association or identification of the targets observed by other nodes; use the low-dimensional hash codes as representations of the node perception states or environmental characteristics, perform state estimation in a distributed computing manner, and implement collaborative decision-making based on the estimation results and task requirements.

[0054] Advantageous effects: Compared with the prior art, the advantages of the present invention are as follows:

[0055] (1) Significantly improve data processing and transmission efficiency: By adopting unsupervised hashing learning, the high-dimensional original perception data blocks are mapped into low-dimensional hash codes, greatly compressing the data representation and reducing the computational overhead of subsequent processing. Combining distributed approximate clustering and LZ4 compression can efficiently process massive data and significantly reduce the bandwidth and time required for data transmission among the unmanned cluster nodes.

[0056] (2) Reduce storage space requirements: Hash dimensionality reduction and subsequent similarity-based grouped compression significantly reduce the space required to store unmanned cluster perception data, which is particularly important for resource-constrained individual unmanned platforms.

[0057] (3) Improve the speed of similarity retrieval and matching: The generated low-dimensional hash codes preserve the similarity between the original data blocks. By calculating the Hamming distance between the hash codes (which is extremely fast to compute), similarity retrieval can be performed quickly, such as for fast target association, identification, or duplicate data detection, improving the efficiency of information matching in collaborative perception.

[0058] (4) Facilitate more effective collaborative perception behavior: Using the hash codes or compressed data groups for inter-node communication, or using the hash codes as a compact representation of the state, can directly support more advanced collaborative perception functions, such as shared situation map construction, distributed state estimation, and collaborative decision-making.

[0059] (5) System integration optimization: By integrating the hash learning ability on top of the database management system or as an independent layer in the task architecture, this efficient data processing method can be systematically applied to all aspects of the unmanned cluster (requirements analysis, design, positioning, tracking, identification), improving the performance of the overall architecture.

[0060] In summary, the present invention deeply integrates a variety of unsupervised hash learning techniques and their variants, and applies them to the specific challenges of unmanned cluster collaborative perception, achieving a significant improvement in the efficiency of data processing, transmission, and storage, enhancing the scalability, robustness, adaptability, and intelligence level of the system, and ultimately improving the overall collaborative perception effectiveness of the unmanned cluster. BRIEF DESCRIPTION OF THE DRAWINGS

[0061] Figure 1 is a flowchart of the method for collaborative perception of unmanned clusters based on unsupervised learning provided by the present invention.

[0062] Figure 2 is a schematic diagram of the application of the method provided by the present invention in the design of an unmanned cluster system.

[0063] Figure 3 is a schematic diagram of the application of the method provided by the present invention in the task architecture of an unmanned cluster. DETAILED DESCRIPTION OF THE INVENTION

[0064] The technical solution of the present invention will be described in detail below with reference to the drawings, but the protection scope of the present invention is not limited to the described embodiments.

[0065] Embodiment 1: As Figure 1 shown, the method for collaborative perception of unmanned clusters based on unsupervised learning provided in this embodiment includes the following steps:

[0066] (1) Obtain data blocks in the unmanned clustering system and use the unsupervised Similarity-Adaptive Hashing (SAH) algorithm to perform hash dimensionality reduction on massive fixed-size data blocks. The low-dimensional binary hash code after dimensionality reduction not only retains the byte similarity between data blocks, but also accelerates the similarity calculation and processing overhead in the low-dimensional Hamming space, thereby significantly reducing the data storage space and improving the efficiency of data block clustering.

[0067] (2) Perform random sample division on the massive low-dimensional hash codes after dimensionality reduction to generate random sample data block subsets, maintain statistical consistency, select random sample data block subsets to execute the distributed approximate clustering algorithm for collaborative sensing data blocks, cluster the random sample data block subsets in a distributed manner, and transmit them to the master node to form approximate clustering centers and clustering clusters, thereby efficiently completing the approximate clustering of massive hash codes;

[0068] (3) The improved DBSCAN (Density-Based Spatial Clustering of Applications with Noise) algorithm is used for clustering, and finally the data block grouping based on the nearest neighbor is formed to ensure that the number of data blocks in the group does not exceed the threshold value. Since the data blocks in the group generated by the DBSCAN algorithm are highly similar, the LZ4 compression algorithm is used to take advantage of its fast compression and decompression characteristics and good compression rate for similar data to further reduce the storage space;

[0069] (4) Based on the results generated in steps (1) to (3), the UAV swarm system is optimized and its indicators are evaluated.

[0070] Specifically, in step (1), the unsupervised similarity-adaptive hashing algorithm aims to map the input data (X) to a low-dimensional hash code. First, a learnable hash function is used to generate a similarity-preserving hash code. The hash code is used to generate a similarity graph and an unsupervised objective function is constructed. The hash function parameters are updated through iterative optimization of the objective function to train the model to generate a similarity-adaptive hash code. This process aims to significantly reduce the amount of data while retaining the similarity information between the original data blocks as much as possible so as to quickly perform similarity comparison and subsequent processing. The specific implementation process is as follows:

[0071] 1.1 Initialize the hash function

[0072] The present invention uses traditional hashing methods (such as MinHash or SimHash) as the initialization model of the learnable hashing function. First, it extracts the features of the data block, assigns corresponding weights to each feature, generates the corresponding binary hash code for each feature, then performs a weighted sum calculation, and finally binarizes each dimension of the weighted sum to complete the initialization.

[0073] 1.2 Conduct training of the learnable hashing function

[0074] The Euclidean loss function is used to measure the difference between the output of the hashing function and the hash code learned in the previous iteration, expressed as follows:

[0075] ;

[0076] where, is the learnable hashing function, which obtains the feature representation by passing the input data X through F, compresses the feature representation to the range; represents the set of parameters of the learnable hashing function, represents the weight parameter, represents the target hash code.

[0077] 1.3 Parameter optimization

[0078] Using the simulated data set generated by the unmanned cluster, the weight parameter is fine-tuned and optimized through standard backpropagation and stochastic gradient descent (SGD), so that the hashing function adapts to the byte similarity characteristics of the data block.

[0079] 1.4 Data similarity update

[0080] After training, the optimized hashing function is obtained. The trained hashing function is used to map the input data into a binary hash code, and the hash code quality is further improved through similarity graph update and hash code optimization. The similarity graph A is updated through the following formula:

[0081] ;

[0082] where, is the feature representation model, is the bandwidth parameter, is an element of the similarity graph, representing the and similarity between.

[0083] The similarity graph update serves as a bridge between the hash model and binary codes, making them more compatible and capturing the similar structures of the output hash representation more effectively. The updated similarity graph participates in the next iterative hash code optimization step, where the general optimization idea of graph hashing is adopted:

[0084] ;

[0085] Among them, is the trace operation, representing the similarity loss between the hash matrix B and the graph Laplacian matrix L. The size of the hash matrix B is , r is the length of the hash code, and n is the number of data blocks; is the graph Laplacian matrix, represents the identity matrix, and represent the regularization parameters, is a vector of all 1s with length n, used for balance regularization; is the constraint condition to ensure that the elements of the hash matrix are binary values. The latter two terms in the formula control the irrelevance and balance of the hash code respectively.

[0086] To reduce the computational complexity, the anchor graph scheme is adopted to calculate the graph Laplacian , and the anchor graph constructs the similarity matrix between all data and the anchor points by selecting representative anchor points, significantly reducing the computational overhead. The specific formula is as follows:

[0087] ;

[0088] Among them, is the identity matrix, is the anchor graph matrix, .

[0089] Specifically, in step (2), the implementation process of the distributed approximate clustering algorithm for a large number of data blocks is as follows:

[0090] To improve the computational efficiency and scalability for perceptual data, ACEM (Approximate Clustering Ensemble Method) uses a distributed computing framework to calculate the approximate clustering results of large datasets. The computing framework consists of seven layers from input to output. The first layer is the dataset as the input of ACEM, and the last layer is the output result of ACEM. The specific operation steps between each layer are as follows:

[0091] The first layer: Store all the input hash codes;

[0092] Second layer: Use the RSP (Random Subspace Projection) algorithm on the large dataset to form multiple random sample data blocks, where each random sample data block is a subsample. The RSP algorithm is a data dimensionality reduction and sampling technique that forms multiple random sample data blocks by randomly selecting feature subspaces and projecting the data, and each data block maintains the statistical characteristics of the original data distribution;

[0093] Third layer: Randomly select a subset of the random sample data blocks and load them onto each parallel processing node;

[0094] Fourth layer: Use the I-niceDP and K-means algorithms on each parallel node to generate the cluster center groups for each random sample data block , the I-niceDP (Improved Nearest-neighbor-based Initialization and Cluster Estimation with Density Peaks) algorithm is an improved density peak clustering initialization method that identifies points with high local density and far from high-density points as candidate cluster centers, overcoming the drawback of the standard K-means being sensitive to the initial points; K-means then uses these high-quality candidate cluster centers and, through fast iterative optimization, precisely partitions all points in the data block into each cluster and further fine-tunes the positions of the center points to obtain the final local cluster centers Cᵢ, and the cluster center groups are integrated into a new cluster center group ;

[0095] Fifth layer: Transmit all the cluster centers to the master node, and use the clustering ensemble method ACEM-MP (Approximate Clustering Ensemble Method - MapReduce) to integrate the cluster centers of the parallel nodes to form the final set of cluster centers as the approximate cluster centers of the large dataset;

[0096] Sixth layer: Broadcast the approximate cluster centers to the parallel nodes;

[0097] Seventh layer: After each parallel node receives the global cluster centers, use the K-meansOne algorithm (K-meansOne is a single-iteration variant of the K-means algorithm for quickly assigning data points to the nearest cluster center) to cluster all the random sample data blocks according to the approximate cluster centers and generate the final clustering clusters .

[0098] Specifically, for each clustering cluster obtained in step (2), further refine the grouping, use DBSCAN to identify sub-regions with higher density, and compress the original data blocks corresponding to these finally determined groupings to reduce storage space. The process is as follows:

[0099] In the first step, set the neighborhood radius Eps, the minimum number of points MinPts within the neighborhood radius to become a core object, and the maximum number of points MaxPts that each cluster can contain at most;

[0100] In the second step, select an unvisited hash code x from the current cluster, mark it as "visited", and check whether it is a core point, that is, there are at least MinPts objects in the Eps neighborhood of x. If it is not a core point, mark it as a noise point, otherwise create a new cluster for x , and put all the objects in the Eps neighborhood of x into the candidate set ;

[0101] In the third step, iteratively add the objects in that do not belong to other clusters to . During this process, for the object x1 marked as "unvisited" in , mark it as "visited", and check its Eps neighborhood. If x1 is also a core object, the objects in the Eps neighborhood of x1 are added to , and continue to add objects to until is empty or the number of data in reaches MaxPts. At this time, the cluster is completely generated;

[0102] In the fourth step, randomly select the next unvisited object from the remaining objects, and repeat the process of the third step until all objects are visited.

[0103] An original clustering cluster is decomposed by DBSCAN into one or more smaller and tighter groupings. The hash codes within each grouping are very similar to each other, and the grouping size is limited by MaxPts.

[0104] Traverse each grouping: Perform the following operations on each grouping generated by DBSCAN in the previous step.

[0105] Data retrieval: Find the original data blocks corresponding to all the hash codes in the grouping.

[0106] Compression: Take the set of these retrieved original data blocks as a unit and compress it using the LZ4 compression algorithm. The LZ4 compression algorithm is a high-speed lossless compression algorithm, especially suitable for compressing similar data blocks, and is used to reduce the final data storage space.

[0107] Output: A series of grouped files of data blocks compressed by LZ4. Each compressed unit corresponds to a tight group identified by DBSCAN and contains the compressed versions of all the original data blocks within the group, greatly reducing the space required for final storage.

[0108] Specifically, step (4) uses one or a combination of low-dimensional hash codes, approximate cluster centers, cluster clustering, and compressed groups to perform at least one of the following operations in the unmanned cluster system to support collaborative perception:

[0109] (401) Transmit the formed approximate cluster center or compressed group among the nodes of the unmanned cluster system to share perception information among the nodes;

[0110] (402) Within the unmanned cluster system, transmit the generated low-dimensional hash code among the nodes. The receiving node performs a nearest neighbor search based on the received hash code through the Hamming distance calculation method to achieve the association or identification of the targets observed by other nodes;

[0111] (403) Use the low-dimensional hash code as the representation of the node's perception state or environmental characteristics, perform state estimation in a distributed computing manner, and based on the estimation results and task requirements, achieve collaborative decision-making.

[0112] Embodiment 2: As Figure 2 shown, it is a schematic diagram of the design and integration of an unmanned cluster system. The top-level module is divided into a task requirement analysis module, a task system architecture model development module, and a task system architecture model generation module. Among them, the task requirement module is responsible for generating requirement analysis, function analysis, comprehensive design, and module design; the task system architecture model generation module is responsible for generating relevant data for system function verification and integration testing; the information generated by the task requirement analysis module and the task architecture system model generation module undergoes iteration to generate the overall design plan and core capability indicators, etc.

[0113] Among them, the data of task requirement analysis needs to interact with the database management system. A hash learning module is added on top of the database management system in the unmanned cluster system. When the data of task requirement analysis interacts with the database management system, first, the data is processed by an unsupervised similarity adaptive hashing algorithm for hashing dimensionality reduction to generate low-dimensional hash codes, thereby reducing the data volume and improving the transmission efficiency. Then, the distributed approximate clustering algorithm is used to process the low-dimensional hash codes to generate clustering centers and clustering results. The improved DBSCAN algorithm is used for the clustering results to form data block grouping based on the nearest neighbors, and the LZ4 compression algorithm is executed to obtain compressed grouping, further reducing the data storage space. As a module connecting the top layer and the bottom layer, the hash learning module stores the processed hash codes, clustering results, and compressed grouping into the database management system during this process. During data retrieval, the relevant data can be quickly located through the hash codes, which functions to improve the data transmission efficiency and facilitate efficient management and retrieval.

[0114] Embodiment 3: As Figure 3 shown, the tasks of the unmanned cluster system are divided into a task layer, a phase layer, a tactical layer, an operation layer, and a cluster algorithm layer. Each layer is responsible for the corresponding work. A hash learning algorithm layer is established above the data layer in the hierarchical architecture of the unmanned cluster task. The data transmitted from the cluster algorithm layer first generates low-dimensional hash codes through an unsupervised similarity adaptive hashing algorithm. This step reduces the dimensionality of the original high-dimensional data. Then, the distributed approximate clustering algorithm is used to perform clustering operations on the hash codes to form clustering results. Subsequently, the improved DBSCAN algorithm is used to process the clustering results to obtain data block grouping and compression. During this process, the hash learning algorithm layer transmits the processed data to the data layer for storage. When reading data, fast similarity search can be performed in the Hamming space using the hash codes, and the relevant data can be quickly found, improving the data retrieval efficiency.

[0115] As described above, although the present invention has been shown and described with reference to specific preferred embodiments, it should not be construed as a limitation of the present invention itself. Various changes can be made in its form and details without departing from the spirit and scope of the present invention defined by the appended claims.

Claims

1. An unsupervised hash learning-based collaborative perception method for unmanned clusters, characterized in that It includes the following steps: (1) Apply an unsupervised similarity adaptive hashing algorithm to perform hashing dimensionality reduction on data blocks in the unmanned cluster system, generating low-dimensional hash codes that can retain the similarity between data blocks; (2) Apply a distributed approximate clustering algorithm to the low-dimensional hash codes to obtain the clustering results of the low-dimensional hash codes, where the clustering results include approximate clustering centers and clustering clusters; (3) Apply an improved DBSCAN algorithm to group the clustering clusters to obtain data block groups, control the number of data blocks in each group not to exceed a preset threshold, and apply a data compression algorithm to the data block groups to generate compressed groups; (4) Use one or a combination of the low-dimensional hash codes, approximate clustering centers, clustering clusters, and compressed groups to perform at least one of the following operations in the unmanned cluster system to support collaborative perception: (401) Transmit the formed approximate clustering centers or compressed groups among the nodes of the unmanned cluster system to share perception information among the nodes; (402) In the unmanned cluster system, transmit the generated low-dimensional hash codes among the nodes, and the receiving node performs nearest neighbor search based on the received hash codes through the Hamming distance calculation method to achieve the association or identification of the targets observed by other nodes; (403) Use the low-dimensional hash codes as the representation of the node perception state or environmental characteristics, perform state estimation in a distributed computing manner, and achieve collaborative decision-making based on the estimation results and task requirements.

2. The unsupervised hash learning-based unmanned cluster collaborative perception method according to claim 1, wherein The unsupervised similarity adaptive hashing algorithm in step (1) includes: Generate hash codes through a learnable hash function; Use the Euclidean loss function to measure the difference between the output of the hash function and the target hash codes learned in the previous iteration, and optimize the parameters of the hash function using backpropagation and stochastic gradient descent; Based on the feature representation corresponding to the currently trained hash function, update the similarity between data blocks, generating an updated similarity graph representing the similarity structure; Based on the updated similarity graph, construct a graph Laplacian matrix, and use a graph hashing optimization method to generate updated target hash codes as the target for the next iteration, where the graph Laplacian matrix is calculated using the anchor graph method; After the iterative optimization loop ends, use the optimized learnable hash function to process the data blocks to generate the low-dimensional hash codes.

3. The method for collaborative perception of unmanned clusters based on unsupervised hashing learning according to claim 2, wherein The difference between the output of the hash function and the hash code learned in the previous iteration, measured using the Euclidean loss function, is expressed as follows: , Among them, F is a hash function, X is the input data block, is a set of learnable hash function parameters, is the weight parameter, is the target hash code; The updated similarity graph is represented as follows: ； Among them, is a feature representation model, is a bandwidth parameter, represents the similarity between data pairs in the input feature space, is the hash function parameter fixed after training; The graph hashing optimization method is represented as follows: ； Among them, is the trace operation, representing the similarity loss between the hash matrix B and the graph Laplacian matrix L. represents the identity matrix. and represent the regularization parameter. is a vector of all 1s with length n. is the constraint condition to ensure that the elements of the hash matrix are binary values, r is the length of the hash code, and n is the number of data blocks. Use the anchor graph method to calculate the graph Laplacian matrix, and the formula is as follows: ； Among them, is the identity matrix, is the anchor graph matrix, which is calculated from the similarity between data blocks and anchors, is a diagonal matrix, , is a vector of all 1s with length n, is the inverse matrix of, is the transpose matrix of Z.

4. The method for unmanned cluster collaborative perception based on unsupervised hash learning according to claim 1, wherein Randomly divide the low-dimensional hash codes into subsets of random sample data blocks; Apply a clustering algorithm including K-means to the subsets of random sample data blocks on each parallel node to generate parallel node clustering centers; Transmit the parallel node clustering centers to the master node, and integrate them through the ACEM-MP method to form approximate clustering centers; Broadcast the approximate clustering centers to the parallel nodes, and cluster all subsets of random sample data blocks according to the approximate clustering centers to generate clustering clusters.

5. The method for collaborative perception of unmanned clusters based on unsupervised hashing learning according to claim 4, wherein The distributed approximate clustering algorithm used in step (2) includes seven layers of steps, and the operations from the first layer to the seventh layer are performed in sequence as follows: The first layer: Obtain and store the low-dimensional hash codes obtained after hash dimensionality reduction processing; The second layer: Use the RSP algorithm on the low-dimensional hash codes to form a subset of random sample data blocks; The third layer: Randomly select a subset of random sample data blocks and load them onto parallel nodes; The fourth layer: Use the I-niceDP and K-means algorithms on each parallel node to generate the parallel node clustering centers of each subset of random sample data blocks; The fifth layer: Transmit all the parallel node clustering centers to the master node, and use the ACEM-MP method to integrate the parallel node clustering centers to form a set of clustering centers as the approximate clustering centers; The sixth layer: Broadcast the approximate clustering centers to the parallel nodes; The seventh layer: Use the K-meansOne algorithm to cluster all subsets of random sample data blocks according to the approximate clustering centers to generate clustering clusters.

6. The method for unsupervised hash learning-based unmanned cluster collaborative perception according to claim 1 or 4, characterized in that The density-based clustering algorithm of the nearest neighbor in step (3) includes the following steps: Set the neighborhood radius Eps, the minimum number of points MinPts within the neighborhood radius, and the maximum number of points MaxPts that can be included; Use the nearest neighbor search method, starting from any point, mark it as "visited" and check if it is a core point; Iteratively add the objects in the candidate set that do not belong to other clustering clusters to the current cluster until the candidate set is empty or the number of data in the current cluster reaches MaxPts; Randomly select the next unvisited object from the remaining objects and repeat the above process until all objects are visited.

7. The unsupervised hash learning-based unmanned cluster collaborative perception method according to claim 1, wherein The method is executed by a hash learning module that interacts with the database management system of the unmanned cluster system. The method further includes storing the processed low-dimensional hash codes, the clustering results, and the compressed packets into the database management system, and using the hash codes to accelerate the search or storage operations related to system design or analysis data.

8. The unsupervised hash learning-based collaborative perception method for unmanned clusters according to claim 1, characterized in that, The method is executed by the hash learning algorithm layer in the task hierarchical architecture of the unmanned cluster system. The hash learning algorithm layer is located between the cluster algorithm layer and the data layer. The method further includes using the processed low-dimensional hash codes, the clustering results, and the compressed packets as the medium for data processing and transmission between the cluster algorithm layer and the data layer to improve the processing efficiency of inter-layer data interaction.

9. The unsupervised hash learning-based unmanned cluster collaborative perception method according to claim 7, characterized in that, Construct a hash index using the low-dimensional hash codes stored in the database management system, and perform content-based fast data retrieval based on the hash index.

10. A system for implementing the unsupervised hash learning-based unmanned cluster collaborative perception method according to claim 1, characterized in that, Including: A data acquisition module configured to acquire the perception data collected by the unmanned cluster system and represent the perception data as data blocks; An unsupervised hash processing module configured to apply an unsupervised similarity adaptive hash algorithm to perform hash dimensionality reduction processing on the data blocks to generate low-dimensional hash codes that can characterize the similarity between the data blocks; A clustering module configured to apply a distributed approximate clustering algorithm to process the low-dimensional hash codes to obtain a clustering result according to the similarity of the low-dimensional hash codes. The clustering result includes approximate clustering centers and clustering clusters; A grouping processing unit, configured to apply an improved DBSCAN algorithm to group the clustering clusters in the clustering result to obtain data block groups, and control the number of data blocks in each data block group not to exceed a preset threshold; A compression processing unit, configured to apply the LZ4 compression algorithm to each data block group to obtain compressed groups; A collaborative perception support module, configured to use one or a combination of the low-dimensional hash codes, approximate clustering centers, clustering clusters, and compressed groups to perform at least one of the following operations in an unmanned cluster system to support collaborative perception: Transmit the formed approximate clustering center or compressed group among the nodes of the unmanned cluster system to share perception information among the nodes; within the unmanned cluster system, transmit the generated low-dimensional hash codes among the nodes, and the receiving node performs a nearest neighbor search based on the received hash codes through the Hamming distance calculation method to achieve the association or identification of the targets observed by other nodes; use the low-dimensional hash code as the representation of the node perception state or environmental characteristics, perform state estimation in a distributed computing manner, and implement collaborative decision-making based on the estimation results and task requirements.

Images