An adaptive radar signal sorting method and system based on an RFSoC heterogeneous platform

Abstract

The application provides a self-adaptive radar signal sorting method and system based on an RFSoC heterogeneous platform. The method works in cooperation with a PS end and a PL end. First, a normalization constant is calculated based on input PDW data, and parallel normalization is performed. Subsequently, a three-part search is used to determine the most dominant energy field radiation factor online, false alarm pulses are removed, and an index mapping is established. The effective pulses are re-normalized, the pulse pair Euclidean distance matrix is calculated, and the nearest neighbor sorting is performed, and then the local density neighborhood scale is determined. The PS end calculates the local density and the relative distance, generates the decision graph basic data, and automatically selects the clustering center through a second-order polynomial regression. Based on the clustering center, pulse assignment and cluster merging are performed, and finally the sorting result of the original batch is output through the index mapping. The application has the characteristics of heterogeneous architecture cooperative design and algorithm hardware friendliness, realizes self-adaptive sorting, and maintains the operation precision.

Description

Technical Field

[0001] This application belongs to the fields of electronic countermeasures and signal processing technology and radar signal sorting technology, specifically relating to an adaptive radar signal sorting method and system based on the RFSoC heterogeneous platform. Background Technology

[0002] The main task of radar signal sorting technology is to accurately separate pulse sequences belonging to the same radar radiation source from a mixed pulse stream received by the receiver, providing reliable basic data for subsequent radar signal identification and radiation source localization. With the continuous evolution of new radar technologies, the continuous increase in the number of electromagnetic devices, and the gradual increase in the complexity of the electromagnetic environment, the performance of the signal sorting stage has become one of the key factors affecting the overall effectiveness of electronic reconnaissance systems.

[0003] In the development of radar signal sorting technology, based on the different characteristic parameters utilized, mainstream methods can be broadly categorized into several types: sorting methods based on Pulse Repetition Interval (PRI), sorting methods based on Pulse Description Word (PDW), and sorting methods based on intra-pulse modulation characteristics. Among these, PDW-based sorting methods use the Pulse Description Word as input data. The PDW is a comprehensive description of the multidimensional characteristic parameters of a radar pulse, typically including parameters such as Radio Frequency (RF), Pulse Width (PW), Direction of Arrival (DOA), and Pulse Amplitude (PA). This type of method performs cluster analysis on the pulse data within a multidimensional feature space, grouping pulses belonging to the same radiation source into the same cluster, thereby deinterlacing the mixed pulse stream.

[0004] Among the PDW parameters, parameters such as DOA, RF, and PW exhibit good stability for the same radiation source within a short observation window. This allows PDW-based sorting methods to make joint decisions in a multi-dimensional feature space, maintaining high sorting accuracy in complex electromagnetic environments such as overlapping signals from multiple radiation sources and local parameter aliasing. Furthermore, since this type of method does not rely on the temporal modulation pattern of the pulse sequence, it exhibits good robustness to non-stationary repetition frequency modulation schemes such as agile repetition frequency, uneven repetition frequency, and jittery repetition frequency. In addition, PDW-based sorting methods can classify pulses without relying on prior parameter templates of the radiation source, facilitating the analysis and processing of intercepted pulse streams through unsupervised clustering, and demonstrating good adaptability to scenarios without prior data.

[0005] However, PDW-based sorting methods still face significant hardware implementation challenges when migrated to engineering applications. On one hand, existing research largely focuses on software simulation and verification on general-purpose computing platforms, with relatively little research on deployment on edge heterogeneous hardware platforms. The limited hardware implementation work already undertaken is mostly based on point-like deployments on general-purpose Field Programmable Gate Arrays (FPGAs) or Digital Signal Processors (DSPs), lacking targeted design for the reasonable allocation of different hardware resources across various computational stages of the algorithm, resulting in processing delays that cannot meet real-time sorting requirements. On the other hand, a significant portion of modern PDW sorting algorithms are adaptive methods, automatically determining the key parameters required for algorithm operation based on the statistical characteristics of the input pulse flow. However, during hardware deployment, limited by the programmable logic circuit's support for control flow branches and floating-point operations, existing solutions often pre-fix adaptive parameters as constants, causing the hardware version to lose the adaptive advantages of the software version, necessitating offline parameter calibration for different electromagnetic environments. In addition, the PDW sorting algorithm usually contains a large number of division, exponentiation and floating-point accumulation operations. When these operations are directly implemented at the programmable logic end, they face problems such as large resource consumption, difficulty in timing convergence and precision degradation, which further exacerbates the difficulty of hardware implementation.

[0006] In summary, existing PDW-based radar signal sorting methods are relatively mature at the algorithm level, but their engineering migration to edge heterogeneous hardware platforms still faces many challenges. How to ensure that the PDW sorting algorithm, under the resource and timing constraints of heterogeneous hardware platforms, can meet real-time processing requirements in hardware implementation while retaining its original adaptive characteristics and without sacrificing algorithm accuracy is a key technical problem that needs to be solved in this field. Summary of the Invention

[0007] The purpose of this application is to overcome the technical defects of existing PDW-based radar signal sorting algorithms when deployed on edge heterogeneous hardware platforms, such as improper matching of computational tasks and hardware resources at each stage of the algorithm, difficulty in preserving adaptive characteristics, and degradation of computational accuracy.

[0008] To achieve the above objectives, this application proposes an adaptive radar signal sorting method based on an RFSoC heterogeneous platform, comprising: Step 1: The PS side calculates the normalization constants for each feature dimension based on the input PDW data, and the PL side performs parallel normalization processing on the PDW data based on the normalization constants; Step 2: The PS and PL ends work together to perform a three-way search for the potential energy field radiation factor: In each iteration, the PS end sends the current candidate radiation factor to the PL end. The PL end calculates the potential energy value based on the pulse pair distance and Gaussian kernel lookup table calculated online and then sends it back to the PS end. The PS end calculates the potential energy entropy based on floating-point operations and shrinks the search interval accordingly. After the iteration ends, the PL end calculates the potential energy value of each pulse based on the optimal radiation factor. Step 3: The PS terminal removes false alarm pulses based on potential energy to obtain the effective pulse set after removing false alarms, and establishes a mapping between each pulse index in the effective pulse set and the corresponding pulse index in the original set; Step 4: The PS end recalculates the normalization constants of each feature dimension for the effective pulse set, and the PL end performs normalization processing on the effective pulse set based on the normalization constants; Step 5: The PL terminal calculates the pulse pair Euclidean distance matrix of the effective pulse set and writes it into the PS terminal memory; Step 6: The PL end performs nearest neighbor sorting on the distance matrix and sends it back to the PS end. The PS end performs iterative termination determination of natural nearest neighbor search based on the sorting results to determine the size of the local density neighborhood of this batch of data. Step 7: The PS terminal calculates the local density of each pulse based on floating-point operations, obtains the density sorting index table in descending order of density, and sends it to the PL terminal; the PL terminal calculates the relative distance of each pulse based on the density sorting index table and the distance matrix in the PS terminal's memory, and sends it back to the PS terminal. Step 8: The PS and PL ends work together to normalize the local density vector and relative distance vector to obtain the basic data of the decision graph; Step 9: Based on the decision graph data, the PS terminal automatically selects the cluster centers of this batch of data through second-order polynomial regression and confidence interval determination; Step 10: The PS terminal sequentially performs core pulse allocation, unallocated pulse voting allocation, and remaining pulse rollback allocation based on the selected cluster centers to obtain the initial cluster partitioning results for this batch of data; Step 11: The PS end adaptively calculates the merging threshold based on the distribution of natural nearest neighbor numbers within each cluster, performs cluster merging on the initial cluster division results, and outputs the final sorting results after backfilling the final cluster division results to the original batch pulse index position based on the index mapping table.

[0009] As an improvement to the above method, step 2 includes: The PS terminal determines the potential energy entropy within a preset search interval using a ternary search method. Optimal radiation factor that achieves minimum value The collaborative timing for each iteration is as follows: the PS end determines the timing based on the two triangulation points of the current iteration. , Calculate separately and The value is written into the configuration register of the PL potential energy field calculation unit after being processed by fixed-point conversion; The PL distance matrix calculation unit calculates the squared distance of each pulse pair online based on normalized coordinates. The Gaussian kernel lookup table (LUT) in the PL potential field calculation unit is based on an index. Approximating the output Gaussian kernel value The potential energy value is obtained by summing the Gaussian kernel values ​​corresponding to all pulse pairs for each pulse. This forms a potential energy array that is transmitted back to the PS terminal; among which, Radiation factor; The PS side calculates potential energy entropy based on floating-point operations. and Based on the relationship between the two, the search interval is narrowed to three parts to enter the next iteration. The iteration terminates when the length of the search interval is less than the preset threshold. After the calculation is completed, the PS end will Write to the PL terminal, and the PL terminal is based on Complete the potential energy values ​​of each pulse The calculation is then sent back to the PS end; Among them, potential energy entropy The expression is: ; in, For the first Each PDW data point in the radiation factor The potential energy value below; Normalization factor; N This represents the number of data entries in the PDW.

[0010] As an improvement to the above method, step 5 includes: The PL distance matrix calculation unit receives the normalized effective pulse set PDW data and stores the normalized coordinates into two types of internal BRAMs: sequential storage BRAM for reading the pulse coordinates corresponding to the current row, and multi-interleaved BRAM for reading the coordinates of multiple comparison pulses in parallel each cycle; the parallel expansion structure of the distance matrix calculation unit is: the three dimensions of DOA, CF, and PW adopt a three-stage pipeline structure of subtraction-square-accumulation, and multiple parallel processing units simultaneously calculate the squared Euclidean distance between the current pulse and multiple comparison pulses in the three-dimensional feature space, and output a transmission beat containing multiple sets of squared distance values ​​each cycle; After each row of calculation is completed, the PL terminal temporarily stores the result of that row in the internal row buffer BRAM, and then writes it into the pre-allocated contiguous physical address space in the PS terminal memory in a burst write mode; the distance matrix is ​​stored contiguously row by row in a fixed-point format.

[0011] As an improvement to the above method, step 6 includes: The nearest neighbor sorting unit (PL) reads the distance matrix from the PS-side memory in row-by-row bursts. Internally, the PL maintains an ordered register array with a depth equal to the preset maximum neighborhood size plus one, sorted in ascending order of distance values. Upon receiving a transmission beat, a multi-channel parallel comparator simultaneously determines the relationship between the distance values ​​of each path and the maximum value at the tail of the ordered array, marking all paths with values ​​smaller than the tail value as awaiting insertion. The distance value and column index of the path to be inserted are then used in a parallel shift insertion network to insert it one by one into the correct position in the ordered array. After processing each row, the first row in the ordered array is... The nearest neighbor entries are packaged and sent back to the working memory of the PS; among them... This indicates the preset maximum neighborhood size.

[0012] The PS (Power Supply) performs a natural nearest neighbor search based on nearest neighbor information to adaptively determine the local density neighborhood size of this batch of data. The iteration starts with a neighborhood size of 1 and increments incrementally. In each round, a K-nearest neighbor list and an inverse K-nearest neighbor list are constructed based on the current neighborhood size, and the intersection of the two is taken to obtain the natural nearest neighbor list. After convergence according to the preset iteration termination criterion, the result is output. And the corresponding natural nearest neighbor table.

[0013] As an improvement to the above method, step 7 includes: The PS (Power Sequencer) calculates the local density value of each pulse using floating-point arithmetic, forming a density array. This density is determined by the pulse's density in its local density neighborhood. The value consists of two parts: the exponential function value of the average distance within the nearest neighbor range and the number of natural nearest neighbors of the pulse. The PS end sorts the obtained density array in descending order to obtain a density descending index table, and sends the index table to the PL relative distance calculation unit; The PL relative distance calculation unit maintains a high-density pulse bitmap internally, traversing each pulse in descending density order: when traversing to the current pulse, the corresponding position of the previous pulse in the bitmap is first set, and then the distance matrix row corresponding to the pulse is read from the PS-end memory using the current pulse index as the base address; for each received transmission beat, the effective distance value corresponding to the traversed high-density pulse is filtered out through the bitmap mask, and then the minimum distance within the transmission beat is calculated and merged with the minimum value already obtained in the current row; after the row is processed, the relative distance value of the pulse and the corresponding nearest high-density neighbor index are output to form a relative distance array and a nearest high-density neighbor index array; the final result is sent back to the PS end; The PS end provides the maximum value among all relative distances for the pulse complement with the highest density.

[0014] As an improvement to the above method, step 9 includes: Based on the decision graph data, the PS terminal automatically selects the cluster centers for this batch of data through second-order polynomial regression and confidence interval determination. This selection process includes two independent decision paths: the first path uses normalized density as the independent variable and normalized relative distance as the dependent variable to perform second-order polynomial fitting, and for each pulse, it uses Student-... The distribution is judged based on the 95% prediction interval. Pulses whose relative distance exceeds the upper limit of the prediction interval and whose density is greater than the preset ratio threshold are added to the initial candidate center set. The second path uses the descending sequence of normalized relative distance to normalized density ratio as the object to perform second-order polynomial fitting. Pulses that deviate from the prediction interval are added to the candidate exclusion set, and the pulse with the lowest density is appended to the end of the set. Finally, the cluster center set is obtained by performing a difference operation on the initial candidate center set and the candidate exclusion set.

[0015] As an improvement to the above method, step 10 includes: The PS side performs a three-step cluster allocation based on the final set of cluster centers: First, using each cluster center in the final set as a seed point, a breadth-first search is performed along the natural nearest neighbor list to allocate core pulses, identifying clusters that have natural neighbor relationships with the seed point and whose natural nearest neighbors satisfy a certain condition. Local density neighborhood size The first step is to assign the pulses to the corresponding clusters; the second step is to count the pulses that are still unassigned. The number of pulses in each assigned cluster in the nearest neighbor is iteratively allocated according to the principle of the maximum number of votes until there are no more pulses to be allocated; in the third step, the remaining unassigned pulses are traversed in descending order of density, and each unassigned pulse is assigned to the cluster to which its nearest high-density neighbor belongs; after the three-step allocation is completed, the initial cluster label vector of this batch of data is obtained.

[0016] As an improvement to the above method, step 11 includes: For each cluster, the PS terminal calculates the distribution of the number of natural nearest neighbors of all pulses within that cluster and then calculates the mean of this distribution. with standard deviation The number of natural nearest neighbors is less than The pulses marked are the boundary pulses of the cluster, and the rest are marked as the core pulses of the cluster; For each pair of clusters on the PS side ,statistics Among the natural nearest neighbors of cluster core pulses, those belonging to The number of pulses in the cluster, as Cluster to The number of bridges between clusters; the smaller of the two clusters is multiplied by a preset ratio and then rounded down to the nearest integer as the merging threshold. If the number of bidirectional bridges between the two clusters exceeds the merging threshold, the cluster pair is marked as a mergeable pair; the PS performs transitive merging on all mergeable pairs, unifying the clusters involved in all mergeable pairs to the same cluster number, and obtaining the final cluster partitioning result of this batch of data. The PS terminal uses the index mapping table to backfill the final cluster division result to the original batch pulse index position: the cluster label corresponding to the position of the false alarm pulse that was removed in the original batch is set to zero, and the cluster number of each pulse in the effective pulse set is written into the original index position pointed to by the index mapping table; finally, the radiation source cluster label vector of each pulse in the original batch is output.

[0017] This application also provides an adaptive radar signal sorting system based on an RFSoC heterogeneous platform, implemented using the above method, the system comprising: The first normalization module is used to calculate the normalization constants of each feature dimension based on the input PDW data, and to perform parallel normalization processing on the PDW data based on the normalization constants. The pulse potential energy value calculation module is used to perform a three-way search of the potential energy field radiation factor and complete the calculation of the potential energy value of each pulse. A pulse index mapping table module is established to remove false alarm pulses based on potential energy, obtain the effective pulse set after removing false alarms, and establish a mapping between each pulse index in the effective pulse set and the corresponding pulse index in the original set. The second normalization module is used to recalculate the normalization constants of each feature dimension of the effective pulse set, and perform normalization processing on the effective pulse set based on the normalization constants; The distance matrix calculation module is used to calculate the Euclidean distance matrix between pulse pairs of the effective pulse set; The local density neighborhood size calculation module is used to perform nearest neighbor sorting on the distance matrix, perform iterative termination determination of natural nearest neighbor search based on the sorting results, and determine the local density neighborhood size of this batch of data. The relative distance calculation module is used to calculate the local density of each pulse based on floating-point operations, obtain a density sorting index table in descending order of density, and calculate the relative distance of each pulse based on the density sorting index table and the distance matrix. The decision graph basic data calculation module is used to complete the normalization processing of local density vectors and relative distance vectors to obtain the decision graph basic data; The cluster center calculation module is used to automatically select the cluster centers of the current batch of data based on the decision graph base data, through second-order polynomial regression and confidence interval determination. The initial cluster partitioning module is used to sequentially perform core pulse allocation, unallocated pulse voting allocation, and remaining pulse back-off allocation based on the selected cluster centers to obtain the initial cluster partitioning results for this batch of data; The cluster merging module is used to adaptively calculate the merging threshold based on the distribution of the number of natural nearest neighbors within each cluster, perform cluster merging on the initial cluster division results, and output the final sorting results after backfilling the final cluster division results to the original batch pulse index positions based on the index mapping table.

[0018] Compared with existing technologies, the advantages of this application are: 1. Heterogeneous Architecture Collaborative Design, Preservation of Adaptive Characteristics. Based on the differences in computational characteristics across different stages of the PDW sorting algorithm, this invention divides the algorithm into two categories: data-parallel rule computation and control-intensive adaptive solution. These tasks are respectively carried out on the PL acceleration core and the PS adaptive layer. This aligns the algorithm's computational characteristics with the two types of computational resources on the RFSoC heterogeneous platform at the underlying execution mode. The PL acceleration core fully leverages the throughput advantage of programmable logic in parallel rule computation, while the PS adaptive layer fully leverages the flexibility of the processor system in control flow and floating-point operations. This significantly reduces the overall processing latency of the algorithm compared to software implementations on general-purpose computing platforms. Furthermore, the PS adaptive layer and the PL acceleration core work in a batch-by-batch collaborative closed-loop manner, ensuring that various adaptive parameters in the algorithm are independently calculated on each batch of PDW data, independent of historical batch results. This fully preserves the adaptive characteristics at the algorithm level in the hardware implementation, eliminating the need for offline parameter calibration for different electromagnetic environments, and providing excellent engineering applicability and cross-scenario versatility.

[0019] 2. Hardware-friendly algorithm rewrite, effectively preserving computational accuracy. This invention addresses the characteristics of the fixed pipelined execution environment at the PL end by rewriting the PDW sorting algorithm's computation process in a hardware-friendly manner: all division operations in the normalization, potential field calculation, and local density calculation stages are moved to the PS adaptive layer to pre-calculate the reciprocal, allowing the PL acceleration core to retain only multiplication and addition operations; the Gaussian kernel operation in the potential field calculation is implemented using a lookup table approximation method, avoiding the resource overhead and timing pressure faced by directly implementing exponential functions in programmable logic; the distance matrix calculation is implemented using a parallel computation array deployed along the PDW dimension, ensuring that large-scale parallel computation is time-closed within the programmable logic resource budget of the target RFSoC. During this hardware-friendly rewrite, precision-sensitive floating-point operations and control flow branches are retained for execution in the PS adaptive layer, effectively preserving the algorithm accuracy of the hardware implementation compared to the software version.

[0020] 3. Dedicated hardware platform, freeing it from the constraints of commercial evaluation boards. This invention relies on the dedicated RFSoC hardware platform disclosed in Chinese Invention Patent No. CN122068983A, "An Electronic Reconnaissance Software Radio Platform Based on RFSoC and Its Usage Method." This hardware platform features customized power topology, clock topology, peripheral interfaces, and board-level integration design for signal sorting applications. This allows the invention to break free from the inherent constraints of commercial evaluation boards in terms of interface form and peripheral resources, providing the hardware foundation required for practical engineering applications. Furthermore, the method provided by this invention does not depend on specific RFSoC devices in its algorithm-hardware mapping relationship, exhibiting good portability and can be ported to other hardware platforms that conform to the aforementioned heterogeneous architecture characteristics. Attached Figure Description

[0021] Figure 1 The diagram shows the overall collaborative architecture of the PL acceleration core and the PS adaptive layer. Figure 2 The diagram shown is a schematic of the hardware platform architecture. Figure 3 The diagram shows the actual resources occupied by the programmable logic. Detailed Implementation

[0022] The technical solution of this application will be described in detail below with reference to the accompanying drawings.

[0023] This application provides an adaptive radar signal sorting method and system based on an RFSoC heterogeneous platform. This method is based on the collaborative design and execution of the programmable logic (PL) end and the processor system end (PS) end of the RFSoC chip. The RFSoC chip structure is described in Chinese Invention Patent Publication No. CN122068983A, entitled "An Electronic Reconnaissance Software Radio Platform and its Usage Method Based on RFSoC". The PL end undertakes the parallel data rule calculation task, serving as the hardware acceleration carrier for the large data scale and computationally structured rule-based stages in the algorithm; the PS end undertakes the control-intensive adaptive parameter calculation task, serving as the carrier for the complex control flow stages in the algorithm that require floating-point operations and mathematical library support; the two ends interact via an AXI bus.

[0024] like Figure 1 As shown, the method includes the following steps: Step 1) The PS side calculates the normalization constants of each feature dimension based on the input PDW data, and the PL side performs parallel normalization processing on the PDW data based on the normalization constants.

[0025] Step 2) The PS and PL ends work together to perform a three-way search for the potential energy field radiation factor: In each iteration, the PS end sends the current candidate radiation factor to the PL end. The PL end calculates the potential energy value based on the online calculated pulse pair distance and Gaussian kernel lookup table and then sends it back to the PS end. The PS end calculates the potential energy entropy based on floating-point operations and shrinks the search interval accordingly. After the iteration ends, the PL end calculates the potential energy value of each pulse based on the optimal radiation factor.

[0026] Step 3) The PS terminal removes false alarm pulses based on potential energy to obtain the effective pulse set after removing false alarms, and establishes a mapping between each pulse index in the effective pulse set and the corresponding pulse index in the original set.

[0027] Step 4) The PS end recalculates the normalization constants of each feature dimension of the effective pulse set, and the PL end performs renormalization processing on the effective pulse set based on the normalization constants.

[0028] Step 5) The PL terminal calculates the pulse pair Euclidean distance matrix of the effective pulse set and writes it to the PS terminal memory. This memory can be DDR4 or other high-speed memory.

[0029] Step 6) The PL end performs nearest neighbor sorting on the distance matrix and sends it back to the PS end. The PS end performs iterative termination determination of natural nearest neighbor search based on the sorting results to determine the local density neighborhood size of this batch of data.

[0030] Step 7) The PS terminal calculates the local density of each pulse based on floating-point operations, obtains the density sorting index table in descending order of density, and sends it to the PL terminal; the PL terminal calculates the relative distance of each pulse based on the density sorting index table and the distance matrix in the PS terminal's memory, and sends it back to the PS terminal.

[0031] Step 8) The PS and PL ends work together to complete the normalization of the local density vector and the relative distance vector to obtain the basic data of the decision graph.

[0032] Step 9) Based on the decision graph data, the PS terminal automatically selects the cluster centers of this batch of data through second-order polynomial regression and confidence interval determination.

[0033] Step 10) The PS terminal sequentially performs core pulse allocation, unallocated pulse voting allocation, and remaining pulse rollback allocation based on the selected cluster centers to obtain the initial cluster partitioning results of this batch of data.

[0034] Step 11) The PS end adaptively calculates the merging threshold based on the distribution of the number of natural nearest neighbors within each cluster, performs cluster merging on the initial cluster division results, and outputs the final sorting results after backfilling the final cluster division results to the original batch pulse index position based on the index mapping table established in Step 3).

[0035] Example 1 I. Hardware Platform Structure like Figure 2 As shown, the RFSoC heterogeneous hardware platform used in this embodiment includes an RFSoC main chip, a PS-side DDR4 storage module, a PDW data input interface, a power management module, a clock distribution module, and an external communication interface module.

[0036] The RFSoC main chip internally comprises two independent computing domains: a processor system and programmable logic (PL). The processor system houses a multi-core ARM processor and on-chip cache, running an embedded operating system, serving as the physical carrier for the PS (Power Switch). The PPL, based on its implemented functions, is logically divided into a normalization module, a distance matrix calculation unit, a potential energy field calculation unit, a nearest neighbor sorting unit, and a relative distance calculation unit, serving as the physical carrier for the PL. The processor system and PPL are interconnected via an AXI bus, providing three types of interface channels: AXI-Lite, AXI-Stream, and AXI4-Memory Mapped. These are used to transmit adaptive parameters and control configuration signals, PDW (Power Controller Wave) data streams and intermediate calculation results, and batch data such as distance matrices, respectively.

[0037] The PS-side DDR4 storage module is connected to the processor system via a DDR controller, serving as the working memory for algorithm execution. It stores raw PDW data, distance matrices, nearest neighbor lists, natural nearest neighbor lists, local density vectors, relative distance vectors, cluster label vectors, and other data. The programmable logic directly accesses the PS-side DDR4 storage module via an AXI high-speed interface. The power management module, clock distribution module, and external communication interface module are responsible for board-level power supply, clock distribution and synchronization, and outputting the final sorting results, respectively.

[0038] II. Specific Implementation of Method Steps The following section provides a more detailed explanation of steps 1) to 11) of the method, taking into account the hardware platform architecture.

[0039] Step 1) The platform receives a batch of N PDW data through the PDW data input interface. Each PDW data includes three feature dimensions: angle of arrival (DOA), carrier frequency (CF), and pulse width (PW). The data is stored in the pre-allocated working memory in the PS-side DDR4 memory.

[0040] The PS reads PDW data from the DDR4 memory, independently traverses the batch of data for each feature dimension, and calculates the maximum value for that dimension. and minimum value And calculate the normalized reciprocal constant based on floating-point operations. ;Will and After fixed-point normalization, the data is written to the configuration register of the PL normalization module via the AXI-Lite interface. The PL normalization module then performs parallel normalization operations on this batch of PDW data along its feature dimensions based on the aforementioned constants. The three dimensions of DOA, CF, and PW are processed in parallel using three independent multiply-accumulate pipelines.

[0041] Step 2) The potential energy field radiation factor of this batch of data is determined at the PS end based on the criterion of minimizing potential energy entropy. .

[0042] The definition of potential energy entropy is: (1) in For the first Each PDW data point in the radiation factor The potential energy value below, This is the normalization factor.

[0043] The PS terminal determines the target using a three-way search method within a preset search range. Optimal radiation factor that achieves minimum value The collaborative timing for each iteration is as follows: the PS end determines the timing based on the two triangulation points of the current iteration. , Calculate separately and The value, after being normalized, is written to the configuration register of the PL potential energy field calculation unit via the AXI-Lite interface; the PL distance matrix calculation unit calculates the squared distance of each pulse pair online based on the normalized coordinates. The Gaussian kernel lookup table (LUT) in the PL potential field calculation unit is based on an index. Approximating the output Gaussian kernel value For each pulse Accumulate all The corresponding Gaussian kernel value yields the potential energy value. In this step, the PL does not write the distance matrix to the DDR4 memory. In each iteration, the PL distance matrix calculation unit recalculates the distance value online and immediately sends it to the downstream calculation unit, avoiding the memory access overhead caused by large-scale off-chip memory read and write.

[0044] After the PL sends the potential energy array back to the PS via the AXI interface, the PS performs precise calculations based on floating-point arithmetic. and The search interval is shrunk into three parts based on the size relationship between the two and enters the next iteration. The iteration terminates when the length of the search interval is less than the preset threshold. After the calculation is completed, the PS end will The value is written to the PL terminal, which then calculates the potential energy value of each pulse based on this constant. The calculation is performed and the result is sent back to the PS end.

[0045] Step 3) The PS terminal marks pulses with potential energy values ​​less than a preset threshold as false alarm pulses and discards them. The set of valid pulses after rejection is denoted as... Its scale is A mapping table is simultaneously established from each pulse index in the valid pulse set to the corresponding pulse index in the original batch. The mapping table is used to backfill the final sorting results in step 11).

[0046] Step 4) PS terminal sets of valid pulses The maximum and minimum value statistics for each feature dimension are re-executed, and the reciprocal normalization constant is recalculated and written into the configuration register of the PL normalization module; the PL side reuses the normalization module from step 1) for... Re-execute the normalization operation.

[0047] Step 5) The PL distance matrix calculation unit receives the renormalized valid pulse set PDW data and stores the normalized coordinates into two types of internal BRAMs: sequential storage BRAM for reading the pulse coordinates corresponding to the current row, and multi-interleaved BRAM for reading the coordinates of multiple comparison pulses in parallel each cycle. The parallel expansion structure of the distance matrix calculation unit is as follows: the three dimensions of DOA, CF, and PW adopt a three-stage pipeline structure of subtraction-square-accumulation. Multiple parallel processing units simultaneously calculate the squared Euclidean distance between the current pulse and multiple comparison pulses in the three-dimensional feature space, and output a transmission beat containing multiple sets of squared distance values ​​each cycle.

[0048] Each completed line (total) The distance matrix is ​​calculated (each distance value). The PL end temporarily stores the result in the internal row buffer BRAM, and then writes it to the pre-allocated contiguous physical address space in the PS end's DDR4 memory via the AXI main interface in burst write mode. The distance matrix is ​​stored contiguously row by row in fixed-point format, and both the PL and PS ends can randomly access the distance matrix through the PS end's DDR4 memory module.

[0049] Step 6) The PL nearest neighbor sorting unit reads the distance matrix row-by-row burst from the PS-side DDR4 storage module via the AXI main interface. Internally, this unit maintains an ordered register array with a depth equal to the preset maximum neighborhood size plus one, sorted in ascending order of distance values. Upon receiving a transmission beat, a multi-channel parallel comparator simultaneously determines the relationship between the distance values ​​of each path and the maximum value at the tail of the ordered array, marking all paths with values ​​smaller than the tail value as awaiting insertion. The distance value and column index of the path to be inserted are then used in a parallel shift insertion network to insert it one by one into the correct position in the ordered array. After processing each row, the first row in the ordered array is... After packaging the nearest neighbor entries (including distance values ​​and pulse indices), they are sent back to the working memory of the PS via DMA through the AXI-Stream interface.

[0050] Based on the aforementioned nearest neighbor information, the PS performs a natural nearest neighbor search to adaptively determine the local density neighborhood size of this batch of data. The iteration starts with a neighborhood size of 1 and increments incrementally. In each round, a K-nearest neighbor list and an inverse K-nearest neighbor list are constructed based on the current neighborhood size, and the intersection of the two is taken to obtain the natural nearest neighbor list. After convergence according to the preset iteration termination criterion, the result is output. And the corresponding natural nearest neighbor table.

[0051] Step 7) The PS terminal calculates the local density value of each pulse based on floating-point operations, forming a density array. This density is determined by the pulse at... The value consists of two parts: the exponential function value of the average distance within the nearest neighbor range and the number of natural nearest neighbors of the pulse. The exponential and floating-point accumulation operations involved in this step are precisely performed by the PS terminal, avoiding the resource overhead and precision degradation problems faced by implementing exponential functions in programmable logic.

[0052] The PS end sorts the obtained density array in descending order to obtain a density descending index table, and sends this index table to the PL relative distance calculation unit through the AXI-Stream interface. The PL relative distance calculation unit internally maintains a high-density pulse bitmap, traversing each pulse in descending density order: when traversing to the current pulse, it first sets the corresponding position of the previous pulse in the bitmap, and then reads the distance matrix row corresponding to the pulse from the PS end's DDR4 via the AXI main interface, using the current pulse index as the base address; for each received transmission pulse, it filters out the effective distance value corresponding to the traversed high-density pulses through the bitmap mask, then calculates the minimum distance within that transmission pulse and merges it with the minimum value already obtained in the current row; after processing the row, it outputs the relative distance value of the pulse and the corresponding nearest high-density neighbor index, forming a relative distance array and a nearest high-density neighbor index array. The final result is sent back to the PS end via DMA through the AXI-Stream interface. For the pulse with the highest density, the PS end supplements the value with the maximum value among all relative distances.

[0053] Step 8) The PS end packages the obtained density array and relative distance array and sends them to the normalization module reused by the PL end for recalculation. and The normalized reciprocal constants of (density and relative distance) are written into the normalization module configuration register; after the PL performs parallel normalization operation, the normalization result is sent back to the PS via DMA as the basis data for the decision graph of subsequent cluster center selection.

[0054] Step 9) Based on the decision graph data, the PS terminal automatically selects the cluster centers for this batch of data through second-order polynomial regression and confidence interval determination. This selection process includes two independent decision paths: the first path uses normalized density as the independent variable and normalized relative distance as the dependent variable to perform second-order polynomial fitting, and for each pulse, it uses Student- The distribution is judged based on the 95% prediction interval. Pulses whose relative distance exceeds the upper limit of the prediction interval and whose density is greater than a preset ratio threshold are added to the initial candidate center set. The second path uses a second-order polynomial fitting on the descending sequence of normalized relative distance to normalized density ratio as the object, adding pulses that deviate from the prediction interval to the candidate exclusion set, and appending the pulse with the lowest density to the end of this set. The final cluster center set is obtained by performing a difference operation on the initial candidate center set and the candidate exclusion set.

[0055] This step involves solving the least squares normal equation, matrix inversion, and Student-... Critical value lookup and confidence interval determination are all performed precisely by the PS terminal based on floating-point operations, avoiding the resource overhead and precision degradation problems faced by programmable logic in implementing complex floating-point matrix operations.

[0056] Step 10) The PS end performs three-step cluster allocation based on the final cluster center set: First, using each cluster center in the final cluster center set as a seed point, a breadth-first search-based core pulse allocation is performed along the natural nearest neighbor list, assigning clusters that have a natural neighbor relationship with the seed point and whose natural nearest neighbors satisfy the following conditions. The first step is to assign the pulses to the corresponding clusters; the second step is to count the pulses that are still unassigned. The number of pulses in each assigned cluster in the nearest neighbor is iteratively allocated according to the principle of the maximum number of votes until no pulses can be allocated. In the third step, the remaining unassigned pulses are traversed in descending order of density, and each unassigned pulse is assigned to the cluster to which its nearest high-density neighbor belongs. After the three-step allocation is completed, the initial cluster label vector of this batch of data is obtained.

[0057] Step 11) For each cluster, the PS terminal calculates the distribution of the number of natural nearest neighbors of all pulses within that cluster and then calculates the mean of this distribution. with standard deviation The number of natural nearest neighbors is less than The pulses marked are the boundary pulses of the cluster, and the rest are marked as the core pulses of the cluster.

[0058] For each pair of clusters on the PS side ,statistics Among the natural nearest neighbors of cluster core pulses, those belonging to The number of pulses in the cluster, as Cluster to The number of bridges between clusters is used as the merging threshold. The smaller of the two clusters is multiplied by a preset ratio and then rounded down. If the number of bidirectional bridges between two clusters exceeds this merging threshold, the cluster pair is marked as a mergeable pair. The PS performs transitive merging on all mergeable pairs, unifying the clusters involved in all mergeable pairs to the same cluster number, thus obtaining the final cluster partitioning result for this batch of data.

[0059] Based on the index mapping table established in step 3), the PS terminal fills the final cluster division result back into the original batch pulse index position: the cluster label corresponding to the position of the false alarm pulse that was removed in the original batch is set to zero, and for each pulse in the valid pulse set, its cluster number in the final cluster division result is written into the original index position pointed to by the index mapping table. Finally, the radiation source cluster label vector of each pulse in the original batch is output.

[0060] III. Examples of Measured Data The method provided in this embodiment is implemented under typical operating conditions, including: a simulated mixed pulse stream containing 15 radar radiation sources, a false alarm pulse ratio of 10%, a missed detection pulse ratio of 10%, and a single batch PDW input scale of approximately 7000. The measured performance data of the method provided in this embodiment are as follows: Figure 3 As shown: the end-to-end processing latency for a single batch is approximately 915 milliseconds; the resource utilization rates of the programmable logic terminal (LUT) are approximately 15.40%, LUTRAM approximately 1.30%, FF approximately 6.82%, BRAM approximately 10.32%, DSP approximately 21.84%, IO approximately 3.29%, and BUFG approximately 0.43%; the operating frequency of the programmable logic terminal is 200MHz; and the sorting accuracy is higher than 95%. Even under extended operating conditions such as increased false alarm and missed detection pulse ratios, larger single-batch PDW input sizes, and continuous batch processing exceeding several batches, the sorting accuracy of the method provided in this embodiment remains at a high level.

[0061] Example 2 This application also provides an adaptive radar signal sorting system based on an RFSoC heterogeneous platform, implemented using the above method. The system includes: The first normalization module is used to calculate the normalization constants of each feature dimension based on the input PDW data, and to perform parallel normalization processing on the PDW data based on the normalization constants. The pulse potential energy value calculation module is used to perform a three-way search of the potential energy field radiation factor and complete the calculation of the potential energy value of each pulse. A pulse index mapping table module is established to remove false alarm pulses based on potential energy, obtain the effective pulse set after removing false alarms, and establish a mapping between each pulse index in the effective pulse set and the corresponding pulse index in the original set. The second normalization module is used to recalculate the normalization constants of each feature dimension of the effective pulse set, and perform normalization processing on the effective pulse set based on the normalization constants; The distance matrix calculation module is used to calculate the Euclidean distance matrix between pulse pairs of the effective pulse set; The local density neighborhood size calculation module is used to perform nearest neighbor sorting on the distance matrix, perform iterative termination determination of natural nearest neighbor search based on the sorting results, and determine the local density neighborhood size of this batch of data. The relative distance calculation module is used to calculate the local density of each pulse based on floating-point operations, obtain a density sorting index table in descending order of density, and calculate the relative distance of each pulse based on the density sorting index table and the distance matrix. The decision graph basic data calculation module is used to complete the normalization processing of local density vectors and relative distance vectors to obtain the decision graph basic data; The cluster center calculation module is used to automatically select the cluster centers of the current batch of data based on the decision graph data through second-order polynomial regression and confidence interval determination. The initial cluster partitioning module is used to sequentially perform core pulse allocation, unallocated pulse voting allocation, and remaining pulse back-off allocation based on the selected cluster centers to obtain the initial cluster partitioning results for this batch of data; The cluster merging module is used to adaptively calculate the merging threshold based on the distribution of the number of natural nearest neighbors within each cluster, perform cluster merging on the initial cluster division results, and output the final sorting results after backfilling the final cluster division results to the original batch pulse index positions based on the index mapping table.

[0062] This application may also provide a computer device, including: at least one processor, memory, at least one network interface, and a user interface. The various components in this device are coupled together via a bus system. It is understood that the bus system is used to implement communication between these components. In addition to a data bus, the bus system also includes a power bus, a control bus, and a status signal bus.

[0063] The user interface can include a display, keyboard, or clicking device. Examples include a mouse, trackball, touchpad, or touchscreen.

[0064] It is understood that the memory in the embodiments disclosed in this application may be volatile memory or non-volatile memory, or may include both volatile and non-volatile memory. The non-volatile memory may be read-only memory (ROM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), or flash memory. The volatile memory may be random access memory (RAM), which is used as an external cache. By way of example, but not limitation, many forms of RAM are available, such as Static Random Access Memory (SRAM), Dynamic Random Access Memory (DRAM), Synchronous DRAM (SDRAM), Double Data Rate SDRAM (DDRSDRAM), Enhanced Synchronous DRAM (ESDRAM), Synchlink DRAM (SLDRAM), and Direct Rambus RAM (DRRAM). The memories described herein are intended to include, but are not limited to, these and any other suitable types of memory.

[0065] In some implementations, the memory stores elements such as executable modules or data structures, or subsets thereof, or extended sets thereof: operating systems and applications.

[0066] The operating system includes various system programs, such as the framework layer, core library layer, and driver layer, used to implement various basic business functions and handle hardware-based tasks. The application programs include various applications, such as media players and browsers, used to implement various application functions. Programs implementing the methods of the embodiments of this disclosure can be included in the application programs.

[0067] In the above embodiments, the processor can also invoke programs or instructions stored in memory, specifically programs or instructions stored in an application program, for the following purposes: Follow the steps described above.

[0068] The above methods can be applied to or implemented by a processor. The processor may be an integrated circuit chip with signal processing capabilities. During implementation, each step of the above methods can be completed by integrated logic circuits in the processor's hardware or by software instructions. The processor can be a general-purpose processor, a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components. It can implement or execute the methods, steps, and logic diagrams disclosed above. The general-purpose processor can be a microprocessor or any conventional processor. The steps of the disclosed methods can be directly implemented by a hardware decoding processor, or by a combination of hardware and software modules in the decoding processor. The software modules can reside in random access memory, flash memory, read-only memory, programmable read-only memory, electrically erasable programmable memory, registers, or other mature storage media in the art. This storage medium is located in memory; the processor reads information from the memory and, in conjunction with its hardware, completes the steps of the above methods.

[0069] It is understood that the embodiments described in this application can be implemented using hardware, software, firmware, middleware, microcode, or a combination thereof. For hardware implementation, the processing unit can be implemented in one or more application-specific integrated circuits (ASICs), digital signal processors (DSPs), digital signal processing devices (DSPDs), programmable logic devices (PLDs), field-programmable gate arrays (FPGAs), general-purpose processors, controllers, microcontrollers, microprocessors, other electronic units for performing the functions described in this application, or combinations thereof.

[0070] For software implementation, the technology of this application can be implemented by executing the functional modules (e.g., procedures, functions, etc.) of this application. The software code can be stored in memory and executed by a processor. The memory can be implemented in the processor or outside the processor.

[0071] This application may also provide a non-volatile storage medium for storing a computer program. When the computer program is executed by a processor, it can implement the steps in the above method embodiments.

[0072] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application and are not intended to limit it. Although this application has been described in detail with reference to the embodiments, those skilled in the art should understand that modifications or equivalent substitutions to the technical solutions of this application do not depart from the spirit and scope of the technical solutions of this application, and should all be covered within the scope of the claims of this application.

Claims

1. An adaptive radar signal sorting method based on an RFSoC heterogeneous platform, comprising: Step 1: The PS side calculates the normalization constants for each feature dimension based on the input PDW data, and the PL side performs parallel normalization processing on the PDW data based on the normalization constants; Step 2: The PS and PL ends work together to perform a three-way search for the potential energy field radiation factor: In each iteration, the PS end sends the current candidate radiation factor to the PL end. The PL end calculates the potential energy value based on the pulse pair distance and Gaussian kernel lookup table calculated online and then sends it back to the PS end. The PS end calculates the potential energy entropy based on floating-point operations and shrinks the search interval accordingly. After the iteration ends, the PL end calculates the potential energy value of each pulse based on the optimal radiation factor. Step 3: The PS terminal removes false alarm pulses based on potential energy to obtain the effective pulse set after removing false alarms, and establishes a mapping between each pulse index in the effective pulse set and the corresponding pulse index in the original set; Step 4: The PS end recalculates the normalization constants of each feature dimension for the effective pulse set, and the PL end performs normalization processing on the effective pulse set based on the normalization constants; Step 5: The PL terminal calculates the pulse pair Euclidean distance matrix of the effective pulse set and writes it into the PS terminal memory; Step 6: The PL end performs nearest neighbor sorting on the distance matrix and sends it back to the PS end. The PS end performs iterative termination determination of natural nearest neighbor search based on the sorting results to determine the size of the local density neighborhood of this batch of data. Step 7: The PS terminal calculates the local density of each pulse based on floating-point operations, obtains the density sorting index table in descending order of density, and sends it to the PL terminal; the PL terminal calculates the relative distance of each pulse based on the density sorting index table and the distance matrix in the PS terminal's memory, and sends it back to the PS terminal. Step 8: The PS and PL ends work together to normalize the local density vector and relative distance vector to obtain the basic data of the decision graph; Step 9: Based on the decision graph data, the PS terminal automatically selects the cluster centers of this batch of data through second-order polynomial regression and confidence interval determination; Step 10: The PS terminal sequentially performs core pulse allocation, unallocated pulse voting allocation, and remaining pulse rollback allocation based on the selected cluster centers to obtain the initial cluster partitioning results for this batch of data; Step 11: The PS end adaptively calculates the merging threshold based on the distribution of natural nearest neighbors within each cluster, performs cluster merging on the initial cluster division results, and outputs the final sorting results after backfilling the final cluster division results to the original batch pulse index position based on the index mapping table.

2. The adaptive radar signal sorting method based on an RFSoC heterogeneous platform according to claim 1, characterized in that, Step 2 includes: The PS terminal determines the potential energy entropy within a preset search interval using a ternary search method. Optimal radiation factor that achieves minimum value The collaborative timing for each iteration is as follows: the PS end determines the timing based on the two triangulation points of the current iteration. , Calculate separately and The value is written into the configuration register of the PL potential energy field calculation unit after being processed by fixed-point conversion; The PL distance matrix calculation unit calculates the squared distance of each pulse pair online based on normalized coordinates. The Gaussian kernel lookup table (LUT) in the PL potential field calculation unit is based on an index. Approximating the output Gaussian kernel value The potential energy value is obtained by summing the Gaussian kernel values ​​corresponding to all pulse pairs for each pulse. This forms a potential energy array that is transmitted back to the PS terminal; among which, Radiation factor; The PS side calculates potential energy entropy based on floating-point operations. and Based on the relationship between the two, the search interval is narrowed to three parts to enter the next iteration. The iteration terminates when the length of the search interval is less than the preset threshold. After the calculation is completed, the PS end will... Write to the PL terminal, and the PL terminal is based on Complete the potential energy values ​​of each pulse The calculation is then sent back to the PS end; Among them, potential energy entropy The expression is: ； in, For the first Each PDW data point in the radiation factor The potential energy value below; Normalization factor; N This represents the number of data entries in the PDW.

3. The adaptive radar signal sorting method based on an RFSoC heterogeneous platform according to claim 1, characterized in that, Step 5 includes: The PL distance matrix calculation unit receives the normalized effective pulse set PDW data and stores the normalized coordinates into two types of internal BRAMs: sequential storage BRAM for reading the pulse coordinates corresponding to the current row, and multi-interleaved BRAM for reading the coordinates of multiple comparison pulses in parallel each cycle; the parallel expansion structure of the distance matrix calculation unit is: the three dimensions of DOA, CF, and PW adopt a three-stage pipeline structure of subtraction-square-accumulation, and multiple parallel processing units simultaneously calculate the squared Euclidean distance between the current pulse and multiple comparison pulses in the three-dimensional feature space, and output a transmission beat containing multiple sets of squared distance values ​​each cycle; After each row of calculation is completed, the PL terminal temporarily stores the result of that row in the internal row buffer BRAM, and then writes it into the pre-allocated contiguous physical address space in the PS terminal memory in a burst write mode; the distance matrix is ​​stored contiguously row by row in a fixed-point format.

4. The adaptive radar signal sorting method based on an RFSoC heterogeneous platform according to claim 1, characterized in that, Step 6 includes: The nearest neighbor sorting unit (PL) reads the distance matrix from the PS-side memory in row-by-row bursts. Internally, the PL maintains an ordered register array with a depth equal to the preset maximum neighborhood size plus one, sorted in ascending order of distance values. Upon receiving a transmission beat, a multi-channel parallel comparator simultaneously determines the relationship between the distance values ​​of each path and the maximum value at the tail of the ordered array, marking all paths with values ​​smaller than the tail value as awaiting insertion. The distance value and column index of the path to be inserted are then used in a parallel shift insertion network to insert it one by one into the correct position in the ordered array. After processing each row, the first row in the ordered array is... The nearest neighbor entries are packaged and sent back to the working memory of the PS; among them... Indicates the preset maximum neighborhood size; The PS (Power Supply) performs a natural nearest neighbor search based on nearest neighbor information to adaptively determine the local density neighborhood size of this batch of data. The iteration starts with a neighborhood size of 1 and increments incrementally. In each round, a K-nearest neighbor list and an inverse K-nearest neighbor list are constructed based on the current neighborhood size, and the intersection of the two is taken to obtain the natural nearest neighbor list. After convergence according to the preset iteration termination criterion, the result is output. And the corresponding natural nearest neighbor table.

5. The adaptive radar signal sorting method based on an RFSoC heterogeneous platform according to claim 1, characterized in that, Step 7 includes: The PS (Power Sequencer) calculates the local density value of each pulse using floating-point arithmetic, forming a density array. This density is determined by the pulse's density in its local density neighborhood. The value consists of two parts: the exponential function value of the average distance within the nearest neighbor range and the number of natural nearest neighbors of the pulse. The PS end sorts the obtained density array in descending order to obtain a density descending index table, and sends the index table to the PL relative distance calculation unit; The PL relative distance calculation unit maintains a high-density pulse bitmap internally, traversing each pulse in descending density order: when traversing to the current pulse, the corresponding position of the previous pulse in the bitmap is first set, and then the distance matrix row corresponding to the pulse is read from the PS-end memory using the current pulse index as the base address; for each received transmission beat, the effective distance value corresponding to the traversed high-density pulse is filtered out through the bitmap mask, and then the minimum distance within the transmission beat is calculated and merged with the minimum value already obtained in the current row; after the row is processed, the relative distance value of the pulse and the corresponding nearest high-density neighbor index are output to form a relative distance array and a nearest high-density neighbor index array; the final result is sent back to the PS end; The PS end provides the maximum value among all relative distances for the pulse complement with the highest density.

6. The adaptive radar signal sorting method based on an RFSoC heterogeneous platform according to claim 1, characterized in that, Step 9 includes: Based on the decision graph data, the PS terminal automatically selects the cluster centers for this batch of data through second-order polynomial regression and confidence interval determination. This selection process includes two independent decision paths: the first path uses normalized density as the independent variable and normalized relative distance as the dependent variable to perform second-order polynomial fitting, and for each pulse, it uses Student-... The distribution is judged based on the 95% prediction interval. Pulses whose relative distance exceeds the upper limit of the prediction interval and whose density is greater than the preset ratio threshold are added to the initial candidate center set. The second path uses the descending sequence of normalized relative distance to normalized density ratio as the object to perform second-order polynomial fitting. Pulses that deviate from the prediction interval are added to the candidate exclusion set, and the pulse with the lowest density is appended to the end of the set. Finally, the cluster center set is obtained by performing a difference operation on the initial candidate center set and the candidate exclusion set.

7. The adaptive radar signal sorting method based on an RFSoC heterogeneous platform according to claim 1, characterized in that, Step 10 includes: The PS side performs a three-step cluster allocation based on the final set of cluster centers: First, using each cluster center in the final set as a seed point, a breadth-first search is performed along the natural nearest neighbor list to allocate core pulses, identifying clusters that have natural neighbor relationships with the seed point and whose natural nearest neighbors satisfy a certain condition. Local density neighborhood size The first step is to assign the pulses to the corresponding clusters; the second step is to count the pulses that are still unassigned. The number of pulses in each assigned cluster in the nearest neighbor is iteratively allocated according to the principle of the maximum number of votes until there are no more pulses to be allocated; in the third step, the remaining unassigned pulses are traversed in descending order of density, and each unassigned pulse is assigned to the cluster to which its nearest high-density neighbor belongs; after the three-step allocation is completed, the initial cluster label vector of this batch of data is obtained.

8. The adaptive radar signal sorting method based on an RFSoC heterogeneous platform according to claim 1, characterized in that, Step 11 includes: For each cluster, the PS terminal calculates the distribution of the number of natural nearest neighbors of all pulses within that cluster and then calculates the mean of this distribution. with standard deviation The number of natural nearest neighbors is less than The pulses marked are the boundary pulses of the cluster, and the rest are marked as the core pulses of the cluster; For each pair of clusters on the PS side ,statistics Among the natural nearest neighbors of cluster core pulses, those belonging to The number of pulses in the cluster, as Cluster to The number of bridges between clusters; the smaller of the two clusters is multiplied by a preset ratio and then rounded down to the nearest integer as the merging threshold. If the number of bidirectional bridges between the two clusters exceeds the merging threshold, the cluster pair is marked as a mergeable pair; the PS performs transitive merging on all mergeable pairs, unifying the clusters involved in all mergeable pairs to the same cluster number, and obtaining the final cluster partitioning result of this batch of data. The PS terminal uses the index mapping table to backfill the final cluster division result to the original batch pulse index position: the cluster label corresponding to the position of the false alarm pulse that was removed in the original batch is set to zero, and the cluster number of each pulse in the effective pulse set is written into the original index position pointed to by the index mapping table; finally, the radiation source cluster label vector of each pulse in the original batch is output.

9. An adaptive radar signal sorting system based on an RFSoC heterogeneous platform, implemented according to the method of any one of claims 1-8, characterized in that, The system includes: The first normalization module is used to calculate the normalization constants of each feature dimension based on the input PDW data, and to perform parallel normalization processing on the PDW data based on the normalization constants. The pulse potential energy value calculation module is used to perform a three-way search of the potential energy field radiation factor and complete the calculation of the potential energy value of each pulse. A pulse index mapping table module is established to remove false alarm pulses based on potential energy, obtain the effective pulse set after removing false alarms, and establish a mapping between each pulse index in the effective pulse set and the corresponding pulse index in the original set. The second normalization module is used to recalculate the normalization constants of each feature dimension of the effective pulse set, and perform normalization processing on the effective pulse set based on the normalization constants; The distance matrix calculation module is used to calculate the Euclidean distance matrix between pulse pairs of the effective pulse set; The local density neighborhood size calculation module is used to perform nearest neighbor sorting on the distance matrix, perform iterative termination determination of natural nearest neighbor search based on the sorting results, and determine the local density neighborhood size of this batch of data. The relative distance calculation module is used to calculate the local density of each pulse based on floating-point operations, obtain a density sorting index table in descending order of density, and calculate the relative distance of each pulse based on the density sorting index table and the distance matrix. The decision graph basic data calculation module is used to complete the normalization processing of local density vectors and relative distance vectors to obtain the decision graph basic data; The cluster center calculation module is used to automatically select the cluster centers of the current batch of data based on the decision graph data through second-order polynomial regression and confidence interval determination. The initial cluster partitioning module is used to sequentially perform core pulse allocation, unallocated pulse voting allocation, and remaining pulse backoff allocation based on the selected cluster centers, thereby obtaining the initial cluster partitioning results for this batch of data; and The cluster merging module is used to adaptively calculate the merging threshold based on the distribution of the number of natural nearest neighbors within each cluster, perform cluster merging on the initial cluster division results, and output the final sorting results after backfilling the final cluster division results to the original batch pulse index positions based on the index mapping table.

Images