A collaborative perception method based on region hierarchical clustering

Abstract

This invention discloses a collaborative sensing method based on regional hierarchical clustering, belonging to the field of spectrum sensing technology. The method includes: each satellite sending data packets to a fusion center, including time-frequency maps, monitoring time, and satellite IDs; the fusion center mapping the time-frequency maps to corresponding monitoring areas based on the monitoring time and satellite IDs; filtering overlapping areas from multiple monitoring areas through regional hierarchical clustering, and integrating the time-frequency maps collected during monitoring of the overlapping areas according to the mapping relationship; arranging the integrated time-frequency maps to form a dataset containing the correspondence between overlapping areas, monitoring time, and time-frequency map groups; and the fusion center inputting the time-frequency map groups from the dataset one by one into a pre-trained neural network model to determine the spectrum occupancy status of the monitoring area at different time periods. This invention achieves high-precision, high-reliability, and near-real-time discrimination of spectrum occupancy status over a large area, providing technical support for wide-area electromagnetic spectrum monitoring.

Description

Technical Field

[0001] This invention belongs to the field of spectrum sensing technology, and particularly relates to a collaborative sensing method based on regional hierarchical clustering. Background Technology

[0002] With the surge in the number of wireless communication devices, electromagnetic spectrum resources are becoming increasingly scarce, making efficient and accurate spectrum monitoring a key support for spectrum monitoring and security control. Low-Earth orbit satellite constellations, with their wide-area coverage, rapid revisit capabilities, and flexible deployment, are widely used for spectrum situational awareness in ground-based monitoring blind spots.

[0003] In existing technologies, some spectrum sensing methods utilize the simultaneous coverage of the same area by multiple low-Earth orbit (LEO) satellites (i.e., multi-satellite co-view) for collaborative sensing, improving detection performance by fusing multi-source observation data. However, this type of method has significant limitations in practical wide-area monitoring scenarios: due to the low orbital altitude and high speed of LEO satellites, wide areas are usually not simultaneously covered by multiple satellites, making it difficult to meet the co-view condition. Forcibly fusing monitoring data from different time periods and coverage areas directly introduces a large amount of irrelevant interference due to the lack of spatial and temporal correlation between the data, severely reducing sensing accuracy.

[0004] Furthermore, most existing methods assume that all observation data originate from the same target area, failing to consider interference signals generated by external frequency-using equipment in non-common-view areas. These interference signals mix with the target signal at the feature level, and if not distinguished, will lead to model misjudgment.

[0005] Therefore, there is an urgent need for a new collaborative spectrum sensing scheme that can effectively process wide-area, non-common-view, and temporally heterogeneous monitoring data and has strong interference suppression capabilities. Summary of the Invention

[0006] To address the aforementioned technical problems, this invention proposes a collaborative sensing method based on regional hierarchical clustering. This method achieves intelligent discrimination of spectrum occupancy status based on regional hierarchical clustering, spatiotemporal data organization, and deep learning feature fusion.

[0007] The collaborative sensing method based on regional hierarchical clustering proposed in this invention is applied to a multi-satellite collaborative monitoring scenario. The method specifically includes: Step S1: After completing monitoring, each satellite sends a data packet to the fusion center. The data packet includes the time-frequency map collected during the monitoring process, the monitoring time, and the satellite number. Step S2: The fusion center receives the data packet and parses it, mapping the time-frequency map to the corresponding monitoring area according to the monitoring time and satellite number; Step S3: Select overlapping regions from multiple monitoring regions through hierarchical clustering, and integrate all time-frequency maps collected when monitoring overlapping regions according to the mapping relationship; Step S4: Arrange the integrated time-frequency maps in chronological order to form a dataset containing the correspondence between overlapping areas, monitoring times, and time-frequency map groups; Step S5: The fusion center inputs the time-frequency graphs from the dataset into the pre-trained neural network model one by one to determine the spectrum occupancy status of the monitoring area at different time periods.

[0008] In step S2, the time-frequency map is mapped to the corresponding monitoring area according to the monitoring time and satellite number, wherein: With A i,t Let represent the set of monitoring areas covered by satellite i during monitoring time t. Then: A i,t = {j | d(jcenter, S i,t )≤R} Where j represents the grid number after the monitoring area is divided, jcenter represents the coordinates of the center point of the j-th grid, and S i,t Let d(jcenter, S) represent the coordinates of satellite i at monitoring time t. i,t ) indicates jcenter and S i,t The Euclidean distance between them, where R represents the monitoring radius.

[0009] In step S2, the time-frequency map is mapped to the corresponding monitoring area according to the monitoring time and satellite number, wherein: The mapping relationship is represented as follows: A i,t D i,t Among them, D i,t This represents the time-frequency graph collected by satellite i at monitoring time t.

[0010] In a multi-satellite collaborative monitoring scenario: if the moving speed of the frequency-using device is equivalent to being stationary relative to the satellite, then the time-frequency map without the aforementioned mapping relationship is unrelated to the monitoring area.

[0011] In a multi-satellite collaborative monitoring scenario: the correlation between different time-frequency maps is equivalent to the correlation between monitoring areas that have a mapping relationship with the time-frequency maps, and the correlation between different monitoring areas is judged based on set similarity.

[0012] In step S3, the hierarchical clustering of regions includes: In the first-level clustering, the intersection of any two sets of monitoring regions is taken. If the intersection is not empty, a new set of regions B1 is formed, and the element positions of the data in the intersection set in the set of monitoring regions are stored in set b1. The above process is executed n-1 times to obtain n-1 sets of regions B1 and corresponding set of element positions b1; where n represents the number of sets of monitoring regions. In the second-level clustering, for any two sets of regions, the intersection is taken. If the intersection is not empty, a new set of regions B2 is formed, and the element positions of the data in the intersection are stored in set b2. The above process is performed n-2 times to obtain n-2 sets of regions B2 and corresponding set of element positions b2. Repeat the above hierarchical clustering until all clustering is completed at level L, obtain the region set BL and the corresponding element location set bL, and determine the overlapping region. Based on the mapping relationship, integrate all time-frequency maps collected when monitoring the overlapping region.

[0013] In step S3, when determining the overlapping region based on the region set BL and the corresponding element position set bL, constraints are set, including: the area of ​​the overlapping region is not less than the area threshold; and the number of times the overlapping region is detected at any monitoring time is not less than the monitoring number threshold.

[0014] In the method, the pre-trained neural network model includes convolutional layers, attention layers, and CNN network layers. The convolutional layers are used to extract features from the input time-frequency graph group, the attention layers are used to assign attention weights to the extracted features to focus on the target features, and the CNN network layers are used to perform spectral occupancy status determination based on the signal patterns contained in the target features.

[0015] In the method, the convolutional layer employs a multi-scale convolutional neural network, which extracts features from the input time-frequency map group at different time and frequency resolutions, and the feature extraction at different time and frequency resolutions is performed in the same representation space.

[0016] In the method, the attention layer adopts a multi-head attention mechanism, which assigns attention weights so that each attention head focuses on the target feature, which refers to features that have common characteristics.

[0017] This invention addresses the technical challenges of wide-area scenarios, such as dispersed regions, asynchronous coverage, weak data correlation, and severe interference, by introducing a time dimension, constructing a regional correlation clustering mechanism, and combining deep learning common feature extraction technology. It achieves high-precision, high-reliability, and near-real-time discrimination of spectrum occupancy status over a wide area, providing technical support for wide-area electromagnetic spectrum regulation. Attached Figure Description

[0018] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0019] Figure 1 This is a flowchart illustrating a collaborative sensing method based on region hierarchical clustering according to an embodiment of the present invention. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] This invention discloses a collaborative sensing method based on regional hierarchical clustering, which is applied to multi-satellite collaborative monitoring scenarios, such as... Figure 1 As shown, the method specifically includes: Step S1: After completing monitoring, each satellite sends a data packet to the fusion center. The data packet includes the time-frequency map collected during the monitoring process, the monitoring time, and the satellite number. Step S2: The fusion center receives the data packet and parses it, mapping the time-frequency map to the corresponding monitoring area according to the monitoring time and satellite number; Step S3: Select overlapping regions from multiple monitoring regions through hierarchical clustering, and integrate all time-frequency maps collected when monitoring overlapping regions according to the mapping relationship; Step S4: Arrange the integrated time-frequency maps in chronological order to form a dataset containing the correspondence between overlapping areas, monitoring times, and time-frequency map groups; Step S5: The fusion center inputs the time-frequency graphs from the dataset into the pre-trained neural network model one by one to determine the spectrum occupancy status of the monitoring area at different time periods.

[0022] In one embodiment, in a scenario of collaborative monitoring by multiple satellites, each satellite, after completing its observations, sends the collected time-frequency map data to the fusion center, along with the precise monitoring time and corresponding satellite number. Upon receiving this information, the fusion center first maps the time-frequency maps to the corresponding monitoring areas based on the satellite number and monitoring time. Next, using a hierarchical clustering method, overlapping areas that meet the requirements are selected from numerous monitoring ranges, and all time-frequency maps obtained from monitoring these overlapping areas are integrated. To more clearly observe dynamic changes, these time-frequency maps are arranged in chronological order, forming a complete dataset that reflects the correspondence between "region, time, and time-frequency map group." Subsequently, the fusion center inputs these time-frequency map groups one by one into a neural network model, thereby accurately determining the occupancy status of each region at different time periods.

[0023] In this overall process, data preprocessing and feature extraction play crucial roles. By rationally dividing the monitoring area, interfering information irrelevant to the target area can be eliminated before data fusion, thus ensuring high data relevance. Simultaneously, during the feature extraction stage, technologies such as deep learning networks effectively preserve common features among multiple satellite monitoring data while discarding non-common features that only exist in a single satellite or non-target areas. In this way, the features obtained by the fusion center can more accurately reflect the true state of the target area, further improving the reliability and accuracy of the final judgment. Through the above processing flow, the fusion center can comprehensively integrate multi-satellite monitoring data, fully utilizing information from both time and region dimensions to obtain information on region-time-spectrum occupancy status.

[0024] In step S2, the time-frequency map is mapped to the corresponding monitoring area according to the monitoring time and satellite number, wherein: With A i,t Let represent the set of monitoring areas covered by satellite i during monitoring time t. Then: A i,t = {j | d(jcenter, S i,t )≤R} Where j represents the grid number after the monitoring area is divided, jcenter represents the coordinates of the center point of the j-th grid, and S i,t Let d(jcenter, S) represent the coordinates of satellite i at monitoring time t. i,t ) indicates jcenter and S i,t The Euclidean distance between them, where R represents the monitoring radius.

[0025] In step S2, the time-frequency map is mapped to the corresponding monitoring area according to the monitoring time and satellite number, wherein: The mapping relationship is represented as follows: Ai,t D i,t Where Di,t represents the time-frequency map collected by satellite i at monitoring time t.

[0026] In one embodiment, this mapping relationship allows for a one-to-one correspondence between monitored data and regions. In this scenario, the moving speed of the frequency-using equipment in the monitoring area is equivalent to being stationary relative to the satellite; therefore, the time-frequency maps of completely different monitoring areas are entirely unrelated. Completely unrelated data has no value for fusion; it is necessary to find highly correlated time-frequency maps for data fusion judgment. The mapping relationship shows that the correlation of a time-frequency map is equivalent to the correlation of its set of monitoring areas. The correlation between sets of monitoring areas can be judged using set similarity, i.e., calculating the intersection of different sets of monitoring areas. If the values ​​of elements within the intersection are greater than a threshold, they can be considered similar, and the two are correlated.

[0027] In a multi-satellite collaborative monitoring scenario: if the moving speed of the frequency-using device is equivalent to being stationary relative to the satellite, then the time-frequency map and the monitoring area that do not have the aforementioned mapping relationship are unrelated; the correlation between different time-frequency maps is equivalent to the correlation between monitoring areas that have a mapping relationship with the time-frequency map, and the correlation between different monitoring areas is judged based on set similarity.

[0028] In step S3, the hierarchical clustering of regions includes: In the first-level clustering, the intersection of any two sets of monitoring regions is taken. If the intersection is not empty, a new set of regions B1 is formed, and the element positions of the data in the intersection set in the set of monitoring regions are stored in set b1. The above process is executed n-1 times to obtain n-1 sets of regions B1 and corresponding set of element positions b1; where n represents the number of sets of monitoring regions. In the second-level clustering, for any two sets of regions, the intersection is taken. If the intersection is not empty, a new set of regions B2 is formed, and the element positions of the data in the intersection are stored in set b2. The above process is performed n-2 times to obtain n-2 sets of regions B2 and corresponding set of element positions b2. Repeat the above hierarchical clustering until all clustering is completed at level L, obtain the region set BL and the corresponding element location set bL, and determine the overlapping region. Based on the mapping relationship, integrate all time-frequency maps collected when monitoring the overlapping region.

[0029] In one embodiment, the first-level clustering involves grouping two highly correlated regions together and taking their intersection to form a new set of regions. Simultaneously, the element positions are combined and stored as a union in another set. This signifies that the data corresponding to each element contains data related to the region's monitoring. In subsequent clustering decisions, considering the data volume, if two regions did not intersect in the previous round, they will certainly not intersect in the next round either. In this case, it is only necessary to determine whether the intersection of the sets is empty. This process is then continued sequentially.

[0030] In step S3, when determining the overlapping region based on the region set BL and the corresponding element position set bL, constraints are set, including: the area of ​​the overlapping region is not less than the area threshold; and the number of times the overlapping region is detected at any monitoring time is not less than the monitoring number threshold.

[0031] In one embodiment, the final judgment condition includes two conditions: first, the area of ​​the control region should not be too small, as a small region will cause a large number of signals to be outside the region, resulting in significant interference; second, the region must be monitored at least k times within each defined time period. At least k time-frequency maps are available for the fusion center to make fusion decisions at each time point. The size of k depends on the convergence of the clustering and the performance of the subsequent neural network; too large a value will prevent the clustering method from converging, while too small a value will result in insufficient input data for the neural network and poor judgment results.

[0032] In one embodiment, multi-round hierarchical clustering is performed based on the coverage grid set of all satellites at different times: First, the sets of any two coverage areas with an intersection grid number greater than a preset area threshold (e.g., one-quarter of the single coverage area of ​​a satellite) are merged, and their intersection is taken as a new cluster area, and all "satellite-time" combinations participating in the clustering are recorded; then, the intersection judgment and merging of the generated cluster areas are continued at a higher level; finally, the clustering results that meet two conditions are retained: (1) the number of grids contained in the cluster area is not less than a preset minimum area threshold; (2) within each preset monitoring period (e.g., 10 seconds), the area is covered by no less than a preset number of satellites (e.g., 5). This process effectively filters out effective monitoring units with strong spatiotemporal correlation.

[0033] In the method, the pre-trained neural network model includes convolutional layers, attention layers, and CNN network layers. The convolutional layers are used to extract features from the input time-frequency graph group, the attention layers are used to assign attention weights to the extracted features to focus on the target features, and the CNN network layers are used to perform spectral occupancy status determination based on the signal patterns contained in the target features.

[0034] In the method, the convolutional layer employs a multi-scale convolutional neural network, which extracts features from the input time-frequency map group at different time and frequency resolutions, and the feature extraction at different time and frequency resolutions is performed in the same representation space.

[0035] In the method, the attention layer adopts a multi-head attention mechanism, which assigns attention weights so that each attention head focuses on the target feature, which refers to features that have common characteristics.

[0036] In one embodiment, the neural network needs to extract common features from the input time-frequency maps to eliminate interference from signals in non-common coverage areas. First, a multi-scale convolutional neural network (CNN) is used to extract features from time-frequency maps from multiple satellites. All satellite time-frequency maps are input into a shared CNN to ensure their features are processed in the same representation space. The shared convolutional network design helps reduce the number of network parameters and ensures the network can uniformly learn common features from signals from different receivers. The multi-scale convolution design enables the network to extract detailed features at different time and frequency resolutions to adapt to signal variations at different scales, further improving the accuracy of feature extraction. In time-frequency maps, signals may exhibit different characteristics at different scales. Using multi-scale convolution ensures that more global information is captured in the lower frequency range, while detailed features are captured in the higher frequency range, thus adapting to variations in various signals.

[0037] In one embodiment, after feature extraction, the resulting representation vectors are concatenated and fed into a multi-head attention mechanism for further processing. The multi-head attention mechanism helps the network focus on common signal features while ignoring the influence of interfering signals by independently learning the relationships between receivers in different subspaces. Each head can independently focus on different signal regions when processing the signal, providing diversity and flexibility for fusing features from different receivers. In cooperative spectrum sensing tasks, receivers may receive interference from other signals. The attention mechanism automatically adjusts the weights of signals from different receivers, ensuring that important signals dominate the final representation and reducing the impact of interference. This mechanism effectively ensures that the importance of signals is correctly identified and has strong robustness to noise and interference.

[0038] In one embodiment, the multi-head attention mechanism reweights each input vector according to its relationship with other inputs, focusing on more relevant features. In this way, the interrelationships between different inputs are fused to perform data fusion, which has strong expressive power in feature fusion and modeling.

[0039] In one embodiment, the features processed by the attention mechanism are fed into a CNN containing two convolutional blocks for classification. This CNN extracts high-level features of the signal through multiple convolutional layers, enabling more effective integration of multi-level signal features and thus improving classification accuracy. By feeding the features extracted by the attention mechanism into this convolutional network, complex signal patterns can be better captured, making it suitable for judging complex signals in time-frequency graphs. Ultimately, the network can accurately determine the spectral occupancy state based on strong discriminative capabilities, thereby achieving higher accuracy.

[0040] In one embodiment, within each monitoring cycle, hierarchical clustering of regions is first performed to generate effective monitoring units; then, a neural network is invoked for each unit to determine its spectral status; finally, the determination results of all regions, time periods, and frequency bands are integrated to generate a wide-area three-dimensional spectral situation map (region-time-frequency band), which is output to the spectrum monitoring platform for visualization and decision support. This module realizes an end-to-end intelligent sensing closed loop from raw observation data to the overall spectral situation.

[0041] This invention constructs a collaborative perception framework for clustering-fusion-discrimination in wide-area non-common-view scenarios. By organically combining hierarchical regional clustering with deep learning features, it effectively solves the problems of weak data correlation and poor interference suppression capability of traditional methods under wide-area dynamic coverage. Specifically, this invention introduces set similarity and hierarchical clustering concepts into the satellite spectrum monitoring data preprocessing workflow for the first time. Utilizing the intersection of coverage areas as a correlation metric, it automatically aggregates multi-temporal and multi-satellite observation data from the same geographical target, avoiding performance degradation caused by cross-regional data mixing. Furthermore, by constructing a regional-temporal-multi-satellite time-frequency map group, it ensures strong spatiotemporal consistency of the neural network input. Moreover, the designed shared multi-scale convolutional and multi-head attention network structure can automatically identify and enhance the real signal characteristics jointly observed by multiple satellites, while effectively filtering external interference that only appears in the field of view of a single satellite, significantly improving the model's robustness in complex electromagnetic environments. The final output regional-temporal-frequency three-dimensional spectrum occupancy model comprehensively depicts the dynamic evolution of the wide-area electromagnetic situation, providing a data foundation for refined spectrum resource management. Simultaneously, the overall architecture balances lightweight on-board design with centralized ground processing, exhibiting good engineering deployability.

[0042] This invention significantly improves wide-area spectrum sensing performance through a collaborative design of region clustering and deep learning. Simulation experiments show that even under harsh conditions with a signal-to-noise ratio as low as -4 dB, the average detection probability still exceeds 85%, and the false alarm rate is less than 12%. Compared with traditional methods that directly fuse all observation data, this invention improves the detection probability by 18% to 25% at the same signal-to-noise ratio. The hierarchical region clustering mechanism effectively suppresses interference from non-common-view areas, enabling the model to maintain stable and reliable discrimination capabilities even in extreme scenarios where the number of interfering devices exceeds the target device. Through system simulation and optimization of parameters such as grid size, sensing interval, region area threshold, and time-period coverage threshold, this invention achieves a good balance between computational efficiency and sensing accuracy. It avoids the problem of excessive interference due to excessively small grid size and prevents data sparsity due to insufficient coverage. The overall method supports dynamic satellite trajectories and random user distribution, possesses strong generalization ability and practical deployment feasibility, and can be widely applied to important scenarios such as spectrum monitoring, emergency communication support, and electromagnetic environment mapping.

[0043] To verify the effectiveness of this invention, a typical wide-area monitoring scenario was constructed: the monitoring area was set at 1000 km × 1000 km, divided into a 20 km × 20 km geographic grid; eight low-Earth orbit satellites were deployed at an altitude of 500 km, with a monitoring half-angle of 30 degrees, acquiring broadband spectrum data once per second, for a total monitoring duration of 60 seconds; 30 frequency-using devices were randomly deployed on the ground, divided into 6 groups, with each group communicating using a designated frequency band for 10 consecutive seconds, and the modulation method randomly employing BPSK, QPSK, 16QAM, or GMSK. Each satellite generated a time-frequency map from the acquired signals and uploaded it to the fusion center. The fusion center first determined the grid set covered by each observation based on the coordinates of the nadir points, and then performed hierarchical clustering of the region, setting the region area threshold to one-quarter of the coverage area of ​​a single satellite and the time period coverage number threshold to 5, ultimately obtaining 27 effective region-time period monitoring units, each unit containing an average of 31 time-frequency maps from different satellites. These data are input into a pre-trained multi-scale convolutional and multi-head attention neural network for spectrum occupancy status discrimination. The system outputs the overall spectrum situation and calculates four comprehensive performance indicators: the detection probability of the best region in time period 1, the average detection probability of all regions in time period 1, the average detection probability of the best region in time period, and the average detection probability of all regions in time period. Experimental results show that, under the condition that the signal-to-noise ratio is not lower than -4 dB, all four indicators exceed 82%, which fully verifies that the present invention can still achieve high-precision and high-robust spectrum sensing capabilities in wide-area, dynamic, and strong interference environments, and has significant practical value and promotion prospects.

[0044] Please note that the technical features of the above embodiments can be combined arbitrarily. For the sake of brevity, not all possible combinations of the technical features in the above embodiments have been described. However, as long as the combination of these technical features does not contradict each other, it should be considered within the scope of this specification. The above embodiments only illustrate several implementation methods of the present invention, and their descriptions are relatively specific and detailed, but they should not be construed as limiting the scope of the invention patent. It should be pointed out that for those skilled in the art, several modifications and improvements can be made without departing from the concept of the present invention, and these all fall within the protection scope of the present invention. Therefore, the protection scope of this invention patent should be determined by the appended claims.

Claims

1. A collaborative sensing method based on region hierarchical clustering, characterized in that, The method is applied to a multi-satellite collaborative monitoring scenario, and the method specifically includes: Step S1: After completing monitoring, each satellite sends a data packet to the fusion center. The data packet includes the time-frequency map collected during the monitoring process, the monitoring time, and the satellite number. Step S2: The fusion center receives the data packet and parses it, mapping the time-frequency map to the corresponding monitoring area according to the monitoring time and satellite number; Step S3: Select overlapping regions from multiple monitoring regions through hierarchical clustering, and integrate all time-frequency maps collected when monitoring overlapping regions according to the mapping relationship; Step S4: Arrange the integrated time-frequency maps in chronological order to form a dataset containing the correspondence between overlapping areas, monitoring times, and time-frequency map groups; Step S5: The fusion center inputs the time-frequency graphs from the dataset into the pre-trained neural network model one by one to determine the spectrum occupancy status of the monitoring area at different time periods.

2. The collaborative sensing method based on region hierarchical clustering according to claim 1, characterized in that, In step S2, the time-frequency map is mapped to the corresponding monitoring area according to the monitoring time and satellite number, wherein: With A i,t Let represent the set of monitoring areas covered by satellite i during monitoring time t. Then: A i,t = {j | d(jcenter, S i,t )≤R} Where j represents the grid number after the monitoring area is divided, jcenter represents the coordinates of the center point of the j-th grid, and S i,t Let d(jcenter, S) represent the coordinates of satellite i at monitoring time t. i,t ) indicates jcenter and S i,t The Euclidean distance between them, where R represents the monitoring radius.

3. The collaborative sensing method based on region hierarchical clustering according to claim 2, characterized in that, In step S2, the time-frequency map is mapped to the corresponding monitoring area according to the monitoring time and satellite number, wherein: The mapping relationship is represented as follows: A i,t D i,t Among them, D i,t This represents the time-frequency graph collected by satellite i at monitoring time t.

4. The collaborative sensing method based on region hierarchical clustering according to claim 3, characterized in that, In a multi-satellite collaborative monitoring scenario: if the moving speed of the frequency-using device is equivalent to being stationary relative to the satellite, then the time-frequency map without the aforementioned mapping relationship is unrelated to the monitoring area.

5. The collaborative sensing method based on region hierarchical clustering according to claim 4, characterized in that, In a multi-satellite collaborative monitoring scenario: the correlation between different time-frequency maps is equivalent to the correlation between monitoring areas that have a mapping relationship with the time-frequency maps, and the correlation between different monitoring areas is judged based on set similarity.

6. The collaborative sensing method based on region hierarchical clustering according to claim 5, characterized in that, In step S3, the hierarchical clustering of regions includes: In the first-level clustering, the intersection of any two sets of monitoring regions is taken. If the intersection is not empty, a new set of regions B1 is formed, and the element positions of the data in the intersection set in the set of monitoring regions are stored in set b1. The above process is executed n-1 times to obtain n-1 sets of regions B1 and corresponding set of element positions b1; where n represents the number of sets of monitoring regions. In the second-level clustering, for any two sets of regions, the intersection is taken. If the intersection is not empty, a new set of regions B2 is formed, and the element positions of the data in the intersection are stored in set b2. The above process is performed n-2 times to obtain n-2 sets of regions B2 and corresponding set of element positions b2. Repeat the above hierarchical clustering until all clustering is completed at level L, obtain the region set BL and the corresponding element location set bL, and determine the overlapping region. Based on the mapping relationship, integrate all time-frequency maps collected when monitoring the overlapping region.

7. The collaborative sensing method based on region hierarchical clustering according to claim 6, characterized in that, In step S3, when determining the overlapping region based on the region set BL and the corresponding element position set bL, constraints are set, including: the area of ​​the overlapping region is not less than the area threshold; and the number of times the overlapping region is detected at any monitoring time is not less than the monitoring number threshold.

8. The collaborative sensing method based on region hierarchical clustering according to claim 7, characterized in that, In the method, the pre-trained neural network model includes convolutional layers, attention layers, and CNN network layers. The convolutional layers are used to extract features from the input time-frequency graph group, the attention layers are used to assign attention weights to the extracted features to focus on the target features, and the CNN network layers are used to perform spectral occupancy status determination based on the signal patterns contained in the target features.

9. A collaborative sensing method based on region hierarchical clustering according to claim 8, characterized in that, In the method, the convolutional layer employs a multi-scale convolutional neural network, which extracts features from the input time-frequency map group at different time and frequency resolutions, and the feature extraction at different time and frequency resolutions is performed in the same representation space.

10. A collaborative sensing method based on region hierarchical clustering according to claim 9, characterized in that, In the method, the attention layer adopts a multi-head attention mechanism, which assigns attention weights so that each attention head focuses on the target feature, which refers to features that have common characteristics.

Images