Non-cooperative bistatic radar real-time signal processing method based on heterogeneous platform

Abstract

The invention provides a non-cooperative bistatic radar real-time signal processing method based on a heterogeneous platform. The non-cooperative bistatic radar real-time signal processing method comprises the steps that the heterogeneous platform is built; obtaining CPU fixed memory areas corresponding to the number of GPU blocks, and creating a data interaction thread to realize parallel computing; performing data analysis through frame header identifier retrieval to obtain transmitting signal parameter information, multi-channel echo data and reference channel data, and performing calculation to obtain echo beam data; zero padding is carried out on the echo beam data and the reference channel data, matched filtering is carried out on the echo beam data after zero padding, and a distance one-dimensional image is obtained; performing fast Fourier transform calculation in the slow time dimension to obtain data after coherent accumulation; obtaining a target trace point; and sorting the target plots, grouping the sorted plots, screening plot groups by using an M / N algorithm, fusing each screened plot group into a single plot, and outputting a final target plot array. According to the invention, signals can be processed in real time, and target trace points can be output.

Description

Technical Field

[0001] The present invention relates to the technical field of non-cooperative bistatic radar signal processing, and particularly to a real-time signal processing method for non-cooperative bistatic radar based on a heterogeneous platform. Background Art

[0002] The non-cooperative bistatic radar based on radar signals does not radiate signals outward by itself, but realizes target detection through coherent processing of the direct wave signal of the radar radiation source and the target scattered echo signal in a silent reception mode. It is difficult for the other party to detect it through corresponding electronic reconnaissance equipment, and it is even more impossible to locate and track it, which greatly improves the survivability. At the same time, the non-cooperative bistatic radar also has the advantages of low construction and maintenance costs, flexible station layout, and no frequency occupation problem.

[0003] The receiving system of the non-cooperative bistatic radar includes a reference channel and an echo channel, which are respectively used to receive the direct wave signal from the radar radiation source and the echo signal after target scattering. The reference channel extracts parameters such as the carrier frequency, bandwidth, pulse width, and pulse arrival time of the radiation source radar. The echo channel intercepts the echo signal using the pulse arrival time extracted in the reference channel, and performs digital mixing on the echo signal within the intercepted interval in combination with the extracted carrier frequency parameter, converts it into a baseband signal, and completes a series of subsequent signal processing and target detection operations.

[0004] Different from the monostatic radar with integrated transceiver or the cooperative bistatic radar with an information communication channel, the non-cooperative bistatic radar system cannot obtain the working state of the radar radiation source, especially the emission direction and beam scanning mode. To solve the spatial synchronization problem, the simultaneous multi-beam reception method is usually adopted to expand the reception range as much as possible. In addition, since the system has separate transceiver and no information interaction, more complex algorithms are usually required for time-frequency synchronization and data processing, at the cost of higher algorithm complexity and computing amount. Therefore, the non-cooperative bistatic radar system has high requirements for computing power, and how to achieve real-time signal processing is a major difficulty.

[0005] The real-time signal processing of traditional radar systems is carried out through embedded development using Digital Signal Processing (DSP) chips and Field Programmable Gate Arrays (FPGA), and integrating dedicated signal processing boards. However, due to high coupling and complex FPGA programming, it is difficult and time-consuming to adjust and maintain the real-time signal processing platform with the DSP+FPGA architecture.

[0006] Therefore, there is an urgent need for a technical solution to solve the above problems. Summary of the Invention

[0007] Aiming at the defects existing in the prior art, the present invention provides a real-time signal processing method for non-cooperative bistatic radar based on a heterogeneous platform.

[0008] To achieve the above object, the technical solution adopted by the present invention is as follows: The real-time signal processing method for non-cooperative bistatic radar based on a heterogeneous platform provided by the present invention includes the following steps: S1. Build a heterogeneous platform, which is composed of a multi-core CPU and multiple GPUs; S2. Obtain the CPU-fixed memory areas corresponding to the number of GPU blocks, create data interaction threads, and transfer data from each CPU-fixed memory area to the corresponding GPU through the data interaction threads to achieve parallel computing; S3. Perform data parsing through frame header identifier retrieval to obtain transmitted signal parameter information, multi-channel echo data, and reference channel data, and obtain echo beam data based on the transmitted signal parameter information and multi-channel echo data; S4. Zero-pad the echo beam data and reference channel data to the next-nearest higher power of 2 of the original pulse sampling points, and perform matched filtering on the zero-padded echo beam data to obtain the range one-dimensional image; S5. Perform multi-threaded parallel fast Fourier transform calculation in the slow time dimension of the range one-dimensional image to obtain the coherently integrated data; S6. Based on the coherently integrated data, perform parallel power calculation, threshold calculation, constant false alarm rate detection in the range dimension, constant false alarm rate detection in the Doppler dimension, amplitude comparison side-looking, and target positioning to obtain target traces; S7. Sort the target traces in chronological order, group the sorted traces, and use the M / N algorithm to screen the trace groups. Each screened trace group is fused into a single trace by taking the mean value, and the final target trace array is output.

[0009] Further, in S2, the parallel computing is implemented according to the following steps: S21. Create 1 data interaction thread DT, 4 CPU-fixed memory areas, and corresponding data processing threads on the CPU side. The 4 fixed memory areas are named CM1, CM2, CM3, and CM4 respectively, and the corresponding data processing threads are named T1, T2, T3, and T4; S22. Read the data to be processed, and adopt the ping-pong mechanism to alternately read and write the 4 fixed memory areas in the parallel preprocessing space to achieve zero-waiting data switching; S23. Bind four fixed memory regions and the corresponding data processing threads to the GPU to form the corresponding relationships T1-CM1-D1, T2-CM2-D2, T3-CM3-D3, T4-CM4-D4, where D1, D2, D3, D4 are the names of the GPUs corresponding to the fixed memory regions and the corresponding data processing threads; S24. Create four fixed memory regions of equal size on the GPU side for data processing of the four threads, named DM1, DM2, DM3, and DM4 respectively; S25. Bind the four fixed memory regions on the GPU side to the corresponding relationships in S23 to form the final corresponding relationships of T1-CM1-D1-DM1, T2-CM2-D2-DM2, T3-CM3-D3-DM3, T4-CM4-D4-DM4.

[0010] Further, in S3, the echo beam data is obtained according to the following steps: S31. Retrieve the frame header identifier based on the protocol agreed with the front end to obtain the starting position of the frame header of each pulse, and store the position information; S32. Extract the frame header information and the pulse data unfolded in one-dimensional form based on the protocol agreed with the front end and the stored frame header starting position information. The frame header information includes the transmission signal parameters and the receiver parameters, and extract and reorganize the pulse data according to the front-end storage format to obtain the multi-channel echo data and the reference channel data in complex form; S33. Obtain the total number of pulses currently processed according to the number of frame headers extracted , and reorganize the reference channel data into a two-dimensional array: number of pulse sampling points × total number of pulses ; S34. According to and the number of receiving channels stored in the frame header information , reorganize the multi-channel echo data into a three-dimensional array: ; S35. Based on the number of beams to be processed , define the storage space of the three-dimensional array on the GPU side. The three-dimensional array is: ; S36. Set the number of parallel threads of the kernel function on the GPU side , based on calculate the data volume and perform data chunking according to the data volume, call the steering vector matrix pre-loaded on the GPU side, multiply it with the three-dimensional array of the multi-channel echo data to obtain the echo beam data, and put the echo beam data into the storage space of the three-dimensional array defined in S35.

[0011] Further, in S4, the one-dimensional range image is obtained according to the following steps: S41. Based on the number of pulse samples calculate the second-nearest higher power of 2 , and call the kernel function on the GPU to pad zeros to the number of pulse samples dimension of the reference channel data and the echo beam data, so that the number of pulse samples increases from to ; S42. Pre-define the storage space of the three-dimensional array of the matched filtering result on the GPU. The three-dimensional array is: ; S43. Calculate the data volume and perform data chunking according to the data volume. Call the fast Fourier transform, inverse fast Fourier transform, and matrix conjugate multiplication functions to perform matched filtering calculation on the echo beam data after data chunking, and intercept the first points in the number of pulse samples dimension of the calculation result to obtain the one-dimensional range image, and store it in the storage space of the three-dimensional array defined in S42; Call the fast Fourier transform, inverse fast Fourier transform, and matrix conjugate multiplication functions to calculate according to the following formula: ; where is the matched filtering calculation result; is the echo signal; is the reference signal; is the fast Fourier transform; is the inverse fast Fourier transform; n is the n th pulse sample point.

[0012] Further, in S5, the data after coherent integration is obtained according to the following steps: S51. Pre-define the storage space of the three-dimensional array of the coherent integration result on the GPU. The three-dimensional array is: , where is the number of integration pulses; S52. Call the fast Fourier transform function to process the one-dimensional range image in the slow time dimension to obtain the data after coherent integration, and store it in the storage space of the three-dimensional array defined in S52; Process the one-dimensional range image according to the following formula: ; where is the data after coherent integration; is the rd range cell; is the th Doppler cell; is the cumulative number of pulses; is the imaginary unit; e is the base of the natural logarithm; 。

[0013] Further, in S6, the target trace is obtained according to the following steps: S61. Calculate the data volume and perform data chunking on the coherently integrated data according to the data volume. Create GPU processing threads within each data processing main thread to achieve large-scale data parallel computing; S62. Parallelly calculate the detection thresholds of the range dimension and the Doppler dimension through pre-grouping, and construct a cell-averaged constant false alarm rate (CA-CFAR) detection threshold table; calculate the CA-CFAR detection threshold according to the following formula: ; ; where is the total number of training cells; is the constant false alarm rate; is the CA-CFAR detection threshold coefficient; is the CA-CFAR detection threshold; is the n th training cell's power value; S63. Eliminate the corresponding cells to be detected according to the set blind zone range. Numerically compare each cell to be detected with the corresponding detection threshold based on the CA-CFAR detection threshold table. If the power of the cell to be detected is greater than the detection threshold, then determine that the cell to be detected is a preliminary target trace. Then, numerically compare the preliminary target trace with its adjacent cells to be detected. If the power value of the preliminary target trace is greater than the adjacent cells to be detected, then determine it as the initial target trace, and obtain all the target traces to form an array of target traces in the range dimension; S64. Eliminate the part where the Doppler unit is zero frequency in the array of target traces in the range dimension, obtain the processed array of target traces in the range dimension. Numerically compare each initial target trace in the processed array of target traces in the range dimension with the detection threshold based on the CA-CFAR detection threshold table. If the power of the initial target trace is greater than the detection threshold, then determine that the initial target trace is a pre-target trace. Then, numerically compare the pre-target trace with its adjacent Doppler units. If the power value of the pre-target trace is greater than the adjacent Doppler units, then determine it as the target trace, and obtain all the target traces to form an array of target traces in the Doppler dimension; S65. Confirm the amplitude of the beam where the target trace is located. If the amplitude of the beam where the target trace is located is the maximum within the adjacent beam range, then calculate the difference and ratio between the target trace amplitude and the amplitude of the adjacent larger beam, and then load the angle discrimination coefficient matrix. Obtain the target trace angle according to the following formula: ; Among them, is the pointing direction of the beam where the detected target is located; is the corresponding amplitude value of the pointing direction of the beam where the detected target is located; is the pointing direction of the beam with a larger amplitude among adjacent beams; is the corresponding amplitude value of the pointing direction of the beam with a larger amplitude among adjacent beams; , , are the discriminant angle coefficients corresponding to the target; is the difference-sum ratio; is the angle of the target relative to the receiving array surface; Obtain the angles of all target points in the Doppler-dimensional target point array to obtain the target detection result; If the amplitude of the beam where the target point is located is not the maximum value within the range of adjacent beams, then eliminate this target point, and then calculate all target points according to the above formula to obtain the target detection result; S66. According to the bistatic deployment relationship transmitted by the front end and the target detection result, calculate the receiving distance and north deviation angle of all target points according to the following formula to form a target track: ; ; ; ; ; Among them, is the angle between the target and the baseline; is the bistatic baseline distance that has been resolved and received from the front end during the system's target positioning and the north deviation angle of the baseline with the receiving array surface as the reference; is the north deviation angle of the normal line of the receiving array surface; is the detected bistatic distance difference of the target; is the bistatic distance represented by each distance unit; is the receiving end sampling rate; is the speed of light; is the detected target point distance unit; is the distance relative to the receiving array surface; is the north deviation angle of the target.

[0014] Furthermore, in S7, the final target track array is output according to the following steps: S71. Create a track fusion thread PS1 on the CPU side; S72. Input the traces P1, P2, P3, and P4 output by the four main threads T1-CM1-D1-DM1, T2-CM2-D2-DM2, T3-CM3-D3-DM3, and T4-CM4-D4-DM4 into PS1; S73. Extract the time information of all the traces in PS1 on the CPU side, and then call the sorting function to sort the traces in chronological order; S74. Set the grouping criteria based on the distance and angle measurement accuracies of the system. Traverse the sorted traces, and group the traces whose distance differences and angle differences both meet the grouping criteria into the same group to obtain trace groups; S75. Screen the trace groups based on the M / N algorithm, count the number of targets in each group. If the number of targets is greater than the detection times threshold M, then fuse all the traces in the trace group into one trace by taking the average of the distance and angle, otherwise consider the targets in the trace group as false alarms and eliminate them; S76. Obtain all the fused traces, construct the final target trace array and output it.

[0015] Further, in S74, the grouping criteria are as follows: taking the first trace of the sorted traces as the reference trace, comparing each subsequent trace with this trace in terms of distance and angle. If both differences are less than the distance detection accuracy and the angle detection accuracy, it is considered that this trace belongs to the same trace group as the reference trace, otherwise this trace is used as the reference trace for a new group.

[0016] Further, the data volume is calculated according to the following formula: .

[0017] Further, the data block needs to satisfy , ; where is the pulse width of the signal emitted by the external radiation source detected currently as and the pulse repetition interval is the farthest detection distance required by the front end; is the speed of light; is the sampling rate of the receiving end.

[0018] Compared with the prior art, the beneficial technical effects of the present invention are as follows: The real-time signal processing method for non-cooperative bistatic radar based on a heterogeneous platform provided by the present invention makes full use of the heterogeneous server architecture of multi-core CPU + multiple GPUs in parallel. When creating core multi-threads on the CPU side, the ping-pong mechanism is combined to maximize resource utilization, thereby meeting the requirements of low latency and high throughput. By using multiple GPUs and creating the corresponding number of main processing threads on the CPU side for parallel data processing, the processing rate is increased several times. By reorganizing multi-channel echo data into a three-dimensional array, subsequent parallel vectorized operations are facilitated to achieve accelerated computing. By implementing multi-threaded parallel high-speed computing of fast Fourier transform and inverse fast Fourier transform on the GPU side to adapt to large data volume scenarios and achieve real-time signal processing. Then, through coherent integration, a coherent signal vector is constructed in the slow-time dimension and coherently superimposed, thereby achieving an increase in the target signal-to-noise ratio. The present invention calculates the cell-averaging constant false alarm rate detection threshold, constructs a cell-averaging constant false alarm rate detection threshold table to store the detection thresholds in the range dimension and Doppler dimension, and then parallelly implements threshold-crossing detection by looking up the table, thereby optimizing the processing efficiency. By rearranging the traces in chronological order and screening the traces using trace grouping and M / N algorithms after the arrangement, trace fusion is performed based on range detection accuracy and angle detection accuracy, thereby outputting the final target trace array. Therefore, the present invention simplifies the algorithm complexity and improves the operation efficiency through parallel processing, reduces the computing power requirements of the non-cooperative bistatic radar system, thereby achieving real-time signal processing and outputting target traces. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for use in the description of the embodiments or the prior art. Obviously, the following drawings are only some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can also be obtained based on the structures shown in these drawings.

[0020] Figure 1 Flowchart of the real-time signal processing method for non-cooperative bistatic radar based on a heterogeneous platform provided in an embodiment; Figure 2 Schematic diagram of multi-threading with multi-core CPU and multiple GPUs in parallel provided in an embodiment; Figure 3 Comparison diagram of serial computing and parallel computing provided in an embodiment, where Figure 3 (a) Schematic diagram of the operation process of serial GPU, Figure 3 (b) Schematic diagram of the operation of parallel GPU; Figure 4 Comparison diagram of the matched filtering algorithm provided in an embodiment, where Figure 4 (a) Schematic diagram of the existing matched filtering algorithm,Figure 4 (b) Schematic diagram of the matched filtering algorithm with multi-core CPUs and multiple GPUs in parallel; Figure 5 The following figure is a comparison diagram of the constant false alarm rate detection algorithms provided in one embodiment, where Figure 5 (a) Schematic diagram of the existing constant false alarm rate detection algorithm, Figure 5 (b) Schematic diagram of the constant false alarm rate detection algorithm with multiple GPUs in parallel; Figure 6 The following figure is a schematic diagram of bistatic target positioning provided in one embodiment. Detailed implementation manners

[0021] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Apparently, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0022] Refer to Figure 1 , one embodiment provides a non-cooperative bistatic radar real-time signal processing method based on a heterogeneous platform, including the following steps: S1. Build a heterogeneous platform, which is composed of multi-core CPUs and multiple GPUs; S2. Obtain the CPU fixed memory areas corresponding to the number of GPU blocks, create data interaction threads, and transfer data from each CPU fixed memory area to the corresponding GPU through the data interaction threads to achieve parallel computing; S3. Perform data parsing through frame header identifier retrieval to obtain transmitted signal parameter information, multi-channel echo data, and reference channel data, and obtain echo beam data based on the transmitted signal parameter information and multi-channel echo data; S4. Zero-pad the echo beam data and reference channel data to the next adjacent higher power of 2 times the original pulse sampling points, and perform matched filtering on the zero-padded echo beam data to obtain the range one-dimensional image; S5. Perform multi-threaded parallel fast Fourier transform calculation in the slow time dimension of the range one-dimensional image to obtain the coherently integrated data; S6. Based on the coherently integrated data, perform parallel power calculation, threshold calculation, range dimension constant false alarm rate detection, Doppler dimension constant false alarm rate detection, amplitude comparison side-looking, and target positioning to obtain target traces; S7. Sort the target traces in chronological order, group the sorted traces, and use the M / N algorithm to screen the trace groups. Each screened trace group is fused into a single trace by taking the average value, and the final target trace array is output.

[0023] In one embodiment, the heterogeneous platform consists of a multi-core CPU and four GPUs; referring to Figure 2 , in S2, the parallel computing is implemented according to the following steps: S21. Create one data interaction thread DT, four CPU fixed memory areas and corresponding data processing threads at the CPU side. The four fixed memory areas are named CM1, CM2, CM3, and CM4 respectively, and the corresponding data processing threads are named T1, T2, T3, and T4; S22. Read the data to be processed, and adopt the ping-pong mechanism to alternately read and write the four fixed memory areas in the parallel preprocessing space to achieve zero-waiting data switching; S23. Bind the four fixed memory areas and the corresponding data processing threads to the GPU to form the corresponding relationships T1-CM1-D1, T2-CM2-D2, T3-CM3-D3, T4-CM4-D4, where D1, D2, D3, and D4 are the names of the GPUs corresponding to the fixed memory areas and the corresponding data processing threads; S24. Create four fixed memory areas of equal size at the GPU side for data processing of the four threads, and name them DM1, DM2, DM3, and DM4 respectively; S25. Bind the four fixed memory areas at the GPU side to the corresponding relationships in S23 to form the final corresponding relationships of T1-CM1-D1-DM1, T2-CM2-D2-DM2, T3-CM3-D3-DM3, T4-CM4-D4-DM4.

[0024] By partitioning the memory at the CPU side, four fixed memory areas named CM1, CM2, CM3, and CM4 are obtained. After receiving the data collected by the front end, the data is divided into four data packets of equal size and stored in the pre-partitioned CPU fixed memory areas, and the data reading rate is optimized through the ping-pong mechanism. Among them, the front-end collection rate is 1 GB / s, and the data volume of each fixed memory area at the CPU side is 256 MB; referring to Figure 3 , through the data interaction between the CPU and the GPU, the data is transmitted to the corresponding GPU, and four main data processing threads are created at the CPU side for parallel data processing, thereby increasing the processing rate by four times; in the serial working mode, the cumulative calculation time of each part increases; obviously, when processing data of the same size, the present invention can increase the processing rate by several times the number of GPU blocks.

[0025] In one embodiment, in S3, the echo beam data is obtained according to the following steps: S31. Retrieve the frame header identifier based on the protocol agreed with the front end to obtain the starting position of the frame header of each pulse, and store the position information; S32. Extract the frame header information and the pulse data expanded in one-dimensional form based on the protocol agreed with the front end and the stored frame header start position information. The frame header information includes the transmit signal parameters (transmit signal pulse width, pulse repetition period, and carrier frequency) and the receiver parameters (receiver sampling rate, number of echo channels), and extract and reorganize the pulse data according to the front-end storage format to obtain the multi-channel echo data and the reference channel data in complex form; S33. Obtain the total number of pulses currently processed according to the number of extracted frame headers , reorganize the reference channel data into a two-dimensional array: number of pulse samples × total number of pulses ; S34. According to and the number of receiving channels stored in the frame header information , reorganize the multi-channel echo data into a three-dimensional array: ; S35. Based on the number of beams to be processed , define the storage space of the three-dimensional array on the GPU side. The three-dimensional array is: ; S36. Set the number of parallel threads of the kernel function on the GPU side , based on calculate the data volume and perform data chunking according to the data volume, call the steering vector matrix pre-loaded on the GPU side, and perform matrix multiplication with the three-dimensional array of multi-channel echo data to obtain the echo beam data, and put the echo beam data into the storage space of the three-dimensional array defined in S35.

[0026] Since the data collected by the front end writes the radar radiation source parameter information measured based on the direct wave channel into the data frame header of each echo pulse, the present invention performs data parsing through frame header identifier retrieval to obtain the transmit signal parameter information, multi-channel echo data, and reference channel data; reorganize the multi-channel echo data into a three-dimensional array, thereby facilitating subsequent parallel vectorization operations and realizing accelerated calculation. By additionally creating parallel processing threads inside the kernel function on the GPU side of each data processing thread to improve the calculation efficiency, where needs to be an integer multiple of 1024.

[0027] Among them, the data volume is calculated according to the following formula: .

[0028] As can be seen from the above formula, whether the processed echo data is a two-dimensional array or a three-dimensional array, the total number of data samples processed within 1 s needs to be 's integer multiple.

[0029] Moreover, in order not to affect the radar detection performance, the data block needs to satisfy , ; wherein is the pulse width of the external radiation source emission signal currently detected as and the pulse repetition interval is , the farthest detection distance required by the front end; is the speed of light; is the sampling rate of the receiving end.

[0030] In one embodiment, in S4, the one-dimensional range image is obtained according to the following steps: S41. Based on the number of pulse sampling points calculate the second-nearest higher power of 2 , call the kernel function on the GPU side to pad zeros to the number of pulse sampling points dimension of the reference channel data and the echo beam data, so that the number of pulse sampling points increases from to ; S42. Pre-define the storage space of the three-dimensional array of the matched filtering result on the GPU side. The three-dimensional array is: ; S43. Calculate the data volume and perform data block according to the data volume. Call the fast Fourier transform, inverse fast Fourier transform and matrix conjugate multiplication functions to perform matched filtering calculation on the echo beam data after data block, intercept the first points of the number of pulse sampling points dimension in the calculation result to obtain the one-dimensional range image, and store it in the storage space of the three-dimensional array defined in S42; Call the fast Fourier transform, inverse fast Fourier transform and matrix conjugate multiplication functions to calculate according to the following formula: ; wherein is the matched filtering calculation result; is the echo signal; is the reference signal; is the fast Fourier transform; is the inverse fast Fourier transform; n is the n th pulse sampling point.

[0031] Considering the characteristics of the fast Fourier transform algorithm itself, the sidelobes of the pulse compression result of the direct wave pulse in the echo channel after matched filtering will appear at the tail of the signal. In the present invention, zero padding is performed on the data, and then the matched filtering is performed on the zero-padded echo beam data. At the same time, the data is transformed into the frequency domain for calculation, thereby reducing the computational complexity. By implementing multi-threaded parallel high-speed calculation of the fast Fourier transform and the inverse fast Fourier transform on the GPU side, it is adapted to the large data volume scenario and real-time signal processing is realized. Refer to Figure 4 Compared with the existing matched filtering algorithm, the present invention realizes the whole process of matched filtering on the GPU side, thereby avoiding the repeated data copying between the CPU and the GPU and greatly improving the operation rate.

[0032] In one embodiment, in S5, the data after coherent integration is obtained according to the following steps: S51. Define the storage space of the three-dimensional array of the coherent integration result in advance on the GPU side. The three-dimensional array is: , where is the number of accumulated pulses; S52. Call the fast Fourier transform function to process the range one-dimensional image in the slow time dimension, obtain the data after coherent integration, and store it in the storage space of the three-dimensional array defined in S52; The range one-dimensional image is processed according to the following formula: ; where is the data after coherent integration; is the th range cell; is the th Doppler cell; is the number of accumulated pulses; is the imaginary unit; e is the base of the natural logarithm; .

[0033] Through coherent integration, a coherent signal vector is constructed in the slow time dimension and coherently superimposed, thereby realizing the improvement of the target signal-to-noise ratio.

[0034] In one embodiment, in S6, the target trace is obtained according to the following steps: S61. Calculate the data volume and perform data block division on the data after coherent integration according to the data volume. Create GPU processing threads in each data processing main thread to realize large-scale data parallel calculation; S62. Parallelly calculate the detection thresholds in the range dimension and the Doppler dimension by means of pre-grouping, and construct a cell-averaging constant false alarm rate detection threshold table; calculate the cell-averaging constant false alarm rate detection threshold according to the following formula: ; ; wherein, is the total number of training units; is the constant false alarm rate; is the detection threshold coefficient of the average cell constant false alarm rate; is the detection threshold of the average cell constant false alarm rate; is the n th power value of the training unit; In the average cell constant false alarm rate detection, it is necessary to set the protection cell parameters to exclude the main lobe broadening area of the target from the noise estimation area. For a chirp signal with a bandwidth of B , its main lobe width is 1 / B , and the corresponding number of range cells is . The number of protection cells should be at least greater than the number of range cells.

[0035] S63. Exclude the corresponding cells to be detected according to the set blind area range. Based on the average cell constant false alarm rate detection threshold table, numerically compare each cell to be detected with the corresponding detection threshold. If the power of the cell to be detected is greater than the detection threshold, then determine that the cell to be detected is a preliminary target point. Then, numerically compare the preliminary target point with its adjacent cells to be detected. If the power value of the preliminary target point is greater than that of the adjacent cells to be detected, then determine it as an initial target point. Obtain all the target points to form a range dimension target point array; S64. Exclude the part of the range dimension target point array where the Doppler cell is at zero frequency to obtain a processed range dimension target point array. Based on the average cell constant false alarm rate detection threshold table, numerically compare each initial target point in the processed range dimension target point array. If the power of the initial target point is greater than the detection threshold, then determine that the initial target point is a pre-target point. Then, numerically compare the pre-target point with its adjacent Doppler cells. If the power value of the pre-target point is greater than that of the adjacent Doppler cells, then determine it as a target point. Obtain all the target points to form a Doppler dimension target point array; S65. Confirm the amplitude of the beam where the target point is located. If the amplitude of the beam where the target point is located is the maximum value within the adjacent beam range, then calculate the difference and ratio between the target point amplitude and the amplitude of the adjacent larger beam. Then, load the angle discrimination coefficient matrix and obtain the target point angle according to the following formula: ; wherein, is the pointing of the beam where the detected target is located; is the corresponding amplitude value of the pointing of the beam where the detected target is located; is the pointing of the beam with a larger amplitude among the adjacent beams; is the corresponding amplitude value of the pointing of the beam with a larger amplitude among the adjacent beams; , , is the discrimination angle coefficient corresponding to the target; is the difference-sum ratio; is the angle of the target relative to the receiving array surface; Obtain the angles of all target points in the Doppler-dimensional target point array to obtain the target detection result; If the amplitude of the beam where the target point is located is not the maximum within the range of adjacent beams, then eliminate this target point, and then calculate all target points according to the above formula to obtain the target detection result; S66. According to the bistatic deployment relationship and the target detection result transmitted by the front end, calculate the receiving distance and north declination of all target points according to the following formula to form a target track: ; ; ; ; ; where, is the included angle of the target relative to the baseline; is the bistatic baseline distance that has been received and solved from the front end when the system performs target positioning and the north declination of the baseline with the receiving array surface as the reference; is the north declination of the normal line of the receiving array surface; is the detected bistatic range difference of the target; is the bistatic distance represented by each range cell; is the sampling rate of the receiving end; is the speed of light; is the detected target point range cell; is the distance relative to the receiving array surface; is the north declination of the target.

[0036] Referring to Figure 6 , for a non-cooperative bistatic radar, it is necessary to substitute the bistatic deployment relationship to solve the actual position relationship between the target and the receiving array surface, so as to achieve target positioning, where A is the target; R t is the distance between the target and the radiation source; T R is the receiving array. The bistatic deployment relationship is solved based on the GPS positions of the radiation source radar and the receiving array surface.

[0037] By calculating the cell-averaged constant false alarm rate (CFAR) detection threshold, constructing a table of cell-averaged CFAR detection thresholds to store the detection thresholds in the range and Doppler dimensions, and then implementing the cross-threshold detection in parallel by looking up the table, the processing efficiency is optimized. To avoid the same adjacent range cells being detected, it is necessary to compare the power values within a certain range of range cells after threshold discrimination and only retain the range cell with the maximum power value. Moreover, the Doppler-dimension CFAR detection is performed after the range-dimension CFAR detection, reducing the computational load, and additional zero-frequency rejection is also carried out therein. Refer to Figure 5 , compared with the existing CFAR detection algorithms, the present invention realizes parallel computing through a preprocessing method of sliding window grouping in advance, sets the number of training cells and protection cells, simultaneously extracts the range and Doppler training window regions of each cell to be detected, calculates the cell-averaged CFAR detection threshold, and constructs a table of cell-averaged CFAR detection thresholds, and implements the cross-threshold detection in parallel by looking up the table, thereby optimizing the processing efficiency.

[0038] In one embodiment, in S7, the final target track array is output according to the following steps: S71. Create a track fusion thread PS1 on the CPU side; S72. Input the tracks P1, P2, P3, P4 output by the four main threads T1-CM1-D1-DM1, T2-CM2-D2-DM2, T3-CM3-D3-DM3, and T4-CM4-D4-DM4 into PS1; S73. Extract the time information of all tracks in PS1 on the CPU side, and then call the sorting function to sort the tracks according to the time sequence; S74. Based on the distance and angle measurement accuracies of the system, set the grouping criteria, traverse the sorted tracks, and divide the tracks with both distance difference and angle difference meeting the grouping criteria into the same group to obtain track groups; S75. Screen the track groups based on the M / N algorithm, count the number of targets in each group. If the number of targets is greater than the detection times threshold M, fuse all the tracks in the track group into one track by taking the average of the distance and angle, otherwise consider the targets in the track group as false alarms and eliminate them; S76. Obtain all the fused tracks, construct the final target track array and output it.

[0039] In a preferred embodiment, in S74, the grouping criteria are as follows: taking the first track among the sorted tracks as the reference track, comparing each subsequent track with this track in terms of distance and angle. If both differences are less than the distance detection accuracy and the angle detection accuracy, it is considered that this track and the reference track belong to the same track group, otherwise this track is used as the reference track for a new group.

[0040] By rearranging the points in chronological order, and then using point grouping and M / N algorithm to filter the points after arrangement, the points are fused based on the distance detection accuracy and angle detection accuracy to output the final target point array. By doing this, all points can be processed and saved in the normal order.

[0041] The present invention develops algorithms on a heterogeneous platform consisting of a multi-core CPU and multiple GPUs to build a real-time signal processing platform with high flexibility and operability. According to the front-end data acquisition rate, the data processing rate can reach at least 1GB / s, and can adapt to the real-time processing scenario of multi-channel data of non-cooperative dual-base radar based on pulse phased array radar radiation source.

[0042] Matters not covered by the present invention are known technologies.

[0043] The technical features of the above embodiments may be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0044] The above-described embodiments only express several implementation methods of the present application, and the descriptions thereof are relatively specific and detailed, but they cannot be understood as limiting the scope of the invention. It should be pointed out that, for a person of ordinary skill in the art, several modifications and improvements can be made without departing from the concept of the present application, and these all belong to the protection scope of the present application. Therefore, the protection scope of the present application shall be subject to the attached claims.

[0045] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A real-time signal processing method for non-cooperative bistatic radar based on heterogeneous platforms, characterized in that: The following steps are involved: S1. Build a heterogeneous platform, where the heterogeneous platform consists of a multi-core CPU and multiple GPUs; S2, obtain the CPU fixed memory area corresponding to the number of GPU blocks, create a data interaction thread, and transfer data from each CPU fixed memory area to the corresponding GPU through the data interaction thread to achieve parallel computing; S3, performing data parsing by frame header identifier retrieval to obtain transmission signal parameter information, multi-channel echo data and reference channel data, and obtaining echo beam data based on the transmission signal parameter information and multi-channel echo data; S4, pad the echo beam data and the reference channel data with zeros to the second highest power of the original pulse sampling points, and perform matched filtering on the zero-padded echo beam data to obtain a one-dimensional range image; S5, performing multi-threaded parallel fast Fourier transform calculation in the slow time dimension of the distance one-dimensional image to obtain coherently accumulated data; S6. Based on the data after coherent accumulation, parallel power calculation, threshold calculation, distance dimension constant false alarm rate detection, Doppler dimension constant false alarm rate detection, amplitude comparison lateral direction and target positioning are performed to obtain the target point trace; S7. Sort the target traces in chronological order, group the sorted traces, and use the M / N algorithm to filter the trace groups. Each filtered trace group is merged into a single trace by taking the average, and the final target trace array is output.

2. The non-cooperative bistatic radar real-time signal processing method based on heterogeneous platforms according to claim 1, characterized in that: In S2, the parallel computing is implemented according to the following steps: S21. Create a data interaction thread DT, four CPU fixed memory areas and corresponding data processing threads on the CPU side. The four fixed memory areas are named CM1, CM2, CM3 and CM4 respectively, and the corresponding data processing threads are named T1, T2, T3 and T4. S22, read the data to be processed, adopt the ping-pong mechanism, and parallelize the preprocessing space to alternately read and write the four fixed memory areas to achieve zero-wait data switching; S23, binding the four fixed memory areas and the corresponding data processing threads to the GPU to form corresponding relationships T1-CM1-D1, T2-CM2-D2, T3-CM3-D3, and T4-CM4-D4, where D1, D2, D3, and D4 are names of the GPUs that form corresponding relationships with the fixed memory areas and the corresponding data processing threads; S24. Create four fixed memory areas of equal size on the GPU side for data processing of four threads, named DM1, DM2, DM3, and DM4 respectively; S25. Bind the four fixed memory areas on the GPU side with the corresponding relationships in S23 to form the final corresponding relationships of T1-CM1-D1-DM1, T2-CM2-D2-DM2, T3-CM3-D3-DM3, and T4-CM4-D4-DM4.

3. The non-cooperative bistatic radar real-time signal processing method based on heterogeneous platforms according to claim 1, characterized in that: In S3, the echo beam data is obtained according to the following steps: S31, performing frame header identifier retrieval based on the protocol agreed with the front end, obtaining the frame header starting position of each pulse, and storing the position information; S32, extracting frame header information and pulse data expanded in one-dimensional form based on the protocol agreed with the front end and the stored frame header starting position information, wherein the frame header information includes transmission signal parameters and receiver parameters, and extracting and reorganizing the pulse data according to the front end storage format to obtain multi-channel echo data and reference channel data in complex form; S33, obtaining the total number of pulses currently processed according to the number of frame headers extracted , reorganize the reference channel data into a two-dimensional array: the number of pulse sampling points ×Total number of pulses ; S34, according to The number of receiving channels stored in the frame header information , reorganize the multi-channel echo data into a three-dimensional array: ; S35, based on the number of beams to be processed , define the storage space of the three-dimensional array on the GPU side. The three-dimensional array is: ; S36. Set the number of parallel threads for the GPU kernel function ,based on Calculate the data volume and divide the data into blocks according to the data volume, call the steering vector matrix pre-loaded on the GPU side, perform matrix multiplication with the three-dimensional array of multi-channel echo data, obtain the echo beam data, and put the echo beam data into the storage space of the three-dimensional array defined by S35.

4. The non-cooperative bistatic radar real-time signal processing method based on heterogeneous platforms according to claim 1, characterized in that: In S4, the one-dimensional distance image is obtained according to the following steps: S41, based on pulse sampling points Calculate the next nearest higher power of 2 , call the GPU kernel function to fill the pulse sampling point number dimension of the reference channel data and echo beam data with zeros, so that the pulse sampling point number is Increase to ; S42, pre-define the storage space of the three-dimensional array of matched filtering results on the GPU side, the three-dimensional array is: ; S43, calculate the data volume and divide the data into blocks according to the data volume, call the fast Fourier transform, inverse fast Fourier transform and matrix conjugate multiplication function to perform matched filtering calculation on the echo beam data after the data is divided, and intercept the front of the pulse sampling point dimension in the calculation result Points, obtain the one-dimensional image of the distance, and store it in the storage space of the three-dimensional array defined in S42; The Fast Fourier Transform, Inverse Fast Fourier Transform and Matrix Conjugate Multiplication functions are called to perform calculations according to the following formulas: in, It is the result of matched filtering calculation; is the echo signal; is the reference signal; is the fast Fourier transform; is the inverse fast Fourier transform; n For the n Pulse sampling points.

5. The non-cooperative bistatic radar real-time signal processing method based on heterogeneous platform according to claim 1, characterized in that: In S5, the coherently accumulated data is obtained according to the following steps: S51, pre-define the storage space of the three-dimensional array of coherent accumulation results on the GPU side, the three-dimensional array is: ,in, To accumulate the number of pulses; S52, calling the fast Fourier transform function to process the one-dimensional image of the distance in the slow time dimension, obtaining the data after coherent accumulation, and storing it in the storage space of the three-dimensional array defined in S52; The distance one-dimensional image is processed according to the following formula: in, The data after coherent accumulation; For the distance units; For the Doppler units; To accumulate the number of pulses; is an imaginary unit; e is the base of natural logarithms; .

6. The non-cooperative bistatic radar real-time signal processing method based on heterogeneous platforms according to claim 1, characterized in that: In S6, the target point trace is obtained according to the following steps: S61, calculate the data volume and divide the data after coherent accumulation into blocks according to the data volume, and create a GPU processing threads to achieve large-scale data parallel computing; S62, calculating the detection thresholds of the distance dimension and the Doppler dimension in parallel by pre-grouping, and constructing a unit average constant false alarm rate detection threshold table; The unit average constant false alarm rate detection threshold is calculated according to the following formula: in, is the total number of training units; is a constant false alarm rate; is the unit average constant false alarm rate detection threshold coefficient; is the unit average constant false alarm rate detection threshold; For the n The power value of each training unit; S63, according to the set blind area range, the corresponding units to be detected are eliminated, and the values ​​of each unit to be detected and the corresponding detection threshold are compared based on the unit average constant false alarm rate detection threshold table. If the power of the unit to be detected is greater than the detection threshold, the unit to be detected is judged to be a preliminary target point, and then the preliminary target point is numerically compared with its adjacent units to be detected. If the power value of the preliminary target point is greater than that of the adjacent units to be detected, it is determined to be the initial target point, and all the target points are obtained to form a distance dimension target point array; S64, eliminating the part of the distance dimension target point array that is zero frequency in the Doppler unit, obtaining the processed distance dimension target point array, performing numerical comparison on each initial target point in the processed distance dimension target point array based on the unit average constant false alarm rate detection threshold table, if the power of the initial target point is greater than the detection threshold, then judging that the initial target point is a pre-target point, then performing numerical comparison on the pre-target point and its adjacent Doppler unit, if the power value of the pre-target point is greater than the adjacent Doppler unit, then judging that it is a target point, obtaining all the target points to form the Doppler dimension target point array; S65, confirm the amplitude of the beam where the target point is located. If the amplitude of the beam where the target point is located is the maximum value within the adjacent beam range, calculate the difference and ratio between the amplitude of the target point and the amplitude of the adjacent larger beam, and then load it into the angle detection coefficient matrix to obtain the angle of the target point according to the following formula: in, is the direction of the beam where the detected target is located; is the corresponding amplitude value of the direction of the beam where the detected target is located; It is the beam pointing with larger amplitude among adjacent beams; is the corresponding amplitude value of the beam pointing to the adjacent beam with the larger amplitude; , , is the angle detection coefficient corresponding to the target; for differences and ratios; is the angle of the target relative to the receiving array; Obtain the angles of all target points in the Doppler target point array to obtain the target detection result; If the amplitude of the beam where the target point is located is not the maximum value in the adjacent beam range, the target point is eliminated, and then all target points are calculated according to the above formula to obtain the target detection result; S66. Based on the dual-base deployment relationship and target detection results transmitted by the front end, the receiving distances and north deflection angles of all target points are calculated according to the following formula to form target point traces: in, is the angle between the target and the baseline; Receive the resolved dual-base baseline distance from the front end when the system is performing target positioning and the baseline north deflection angle based on the receiving array; is the north deflection angle of the normal of the receiving array; is the detected target bistatic range difference; is the bistatic distance represented by each range cell; is the sampling rate of the receiving end; is the speed of light; is the distance unit of the detected target point; is the distance relative to the receiving array; is the north declination angle of the target.

7. The non-cooperative bistatic radar real-time signal processing method based on heterogeneous platforms according to claim 1, characterized in that: In S7, the final target point trace array is output according to the following steps: S71. Create a point-trace fusion thread PS1 on the CPU side; S72, input the traces P1, P2, P3, and P4 output by the four main threads T1-CM1-D1-DM1, T2-CM2-D2-DM2, T3-CM3-D3-DM3, and T4-CM4-D4-DM4 into PS1; S73, extracting the time information of all traces in PS1 on the CPU side, and then calling the sorting function to sort the traces in chronological order; S74, setting a grouping standard based on the distance and angle measurement accuracy of the system, traversing the sorted traces, and grouping the traces whose distance difference and angle difference both meet the grouping standard into the same group to obtain a trace group; S75, screening the point trace group based on the M / N algorithm, counting the number of targets in each group, if the number of targets is greater than the detection threshold M, then merging all the point traces in the point trace group into one point trace by taking the average of the distance and angle, otherwise the target of the point trace group is considered to be a false alarm and is removed; S76. Obtain all fused traces, construct the final target trace array and output it.

8. The non-cooperative bistatic radar real-time signal processing method based on heterogeneous platforms according to claim 7, characterized in that: In S74, the grouping standard is: take the first point trace after sorting as the reference point trace, compare the distance and angle of each subsequent point trace with the point trace, if the difference between the two is less than the distance detection accuracy and the angle detection accuracy, it is considered that the point trace and the reference point trace belong to the same point trace group, otherwise the point trace is used as the reference point trace of the new group.

9. The method for real-time signal processing of a non-cooperative bistatic radar based on a heterogeneous platform according to any one of claims 3, 4 or 6, characterized in that: The amount of data Calculated according to the following formula: 。 10. The method for real-time signal processing of a non-cooperative bistatic radar based on a heterogeneous platform according to any one of claims 3, 4 or 6, characterized in that: The data blocks must meet , , in, The pulse width of the signal emitted by the currently detected external radiation source is and the pulse repetition interval is When the front end needs to detect the farthest distance; is the speed of light; is the sampling rate at the receiving end.

Images