ZYNQ-based radar interference signal identification method
By building and deploying a radar jamming signal identification method based on a BP neural network on the ZYNQ platform, the problems of identification accuracy and resource consumption are solved, achieving efficient radar jamming signal identification. This method is applicable to ZYNQ's ARM and FPGA platforms.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-12-31
- Publication Date
- 2026-04-14
AI Technical Summary
Existing technologies suffer from low accuracy and high hardware resource consumption when identifying radar jamming signals, making them difficult to deploy effectively on the resource-constrained ZYNQ ARM and FPGA platforms.
A radar jamming signal identification method based on ZYNQ is adopted. By constructing training and test sets in MATLAB, six feature parameters with the greatest differences are extracted. The model is trained and deployed using a BP neural network and combined with the Levenberg-Marquardt optimization algorithm to achieve deployment and verification on ARM and FPGA platforms.
At a noise-to-interference ratio of 0dB or higher, it achieves an average correct recognition rate of 99%, with low hardware resource consumption and fast recognition response speed. It takes 2245.327us to recognize a pulse signal on the ARM platform and 27.635us on the FPGA platform.
Smart Images
Figure CN121856901A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of radar jamming signal identification technology, and in particular, it is a radar jamming signal identification method based on ZYNQ. Background Technology
[0002] In informationized warfare scenarios, radar plays an irreplaceable and crucial role. However, various new radar jamming technologies are constantly emerging, significantly weakening the normal operational effectiveness of radar. Accurately identifying radar jamming signals can provide reliable decision-making basis for friendly electronic defense deployment and enemy electronic strike operations. The development level of this identification technology has become one of the core indicators for evaluating the advancement of electronic countermeasures equipment.
[0003] The most widely used radar jamming signal identification scheme currently employs a feature extraction-centric technical approach. This technology starts by understanding the generation principles of different types of jamming signals. It first converts the time-domain jamming signals to other signal domains, then uses either manual feature extraction or automatic feature extraction via network models. By comparing the feature differences between different jamming signals, it ultimately achieves accurate identification and classification of various jamming signals. Methods using manually extracted features for identification and classification typically employ machine learning techniques such as decision trees, support vector machines, and backpropagation (BP) neural networks. Methods using automatic feature extraction for identification and classification typically employ deep learning techniques such as convolutional neural networks and LSTM. Manual feature extraction for identification and classification has lower complexity but also relatively lower accuracy; automatic feature extraction via network models has higher accuracy, but the model parameters are often large, making it unsuitable for deployment on resource-constrained hardware platforms.
[0004] Therefore, developing a radar jamming signal identification method that balances recognition accuracy with hardware resource consumption, and enabling its deployment on ZYNQ's ARM and FPGA platforms, is of significant practical importance. Summary of the Invention
[0005] The purpose of this invention is to provide a radar jamming signal identification method based on ZYNQ, which can balance identification accuracy and resource consumption, and can be practically deployed on ZYNQ's ARM platform and FPGA platform for application.
[0006] The technical solution for achieving the objective of this invention is: a method for identifying radar jamming signals based on ZYNQ, the method comprising:
[0007] Step 1: Construct training and test sets for Gaussian white noise and seven types of radar jamming signals under different interference-to-noise ratios in MATLAB. The seven types of radar jamming signals include single-tone jamming, multi-tone jamming, noise amplitude modulation jamming, noise frequency modulation jamming, intermittent sampling direct forwarding jamming, intermittent sampling repeated forwarding jamming, and spectral dispersion jamming.
[0008] Step 2: The signals of the training set and test set are preprocessed by power normalization and Hamming windowing, different feature parameters of the signals are calculated, and the six feature parameters with the greatest differences among them are extracted.
[0009] Step 3: Add the real labels to the six feature parameters of the training set under different interference-to-noise ratios to construct and train the BP neural network model;
[0010] Step 4: Use the test set data to verify the model's recognition accuracy. If the average recognition rate does not exceed the set threshold when the drying ratio is 0dB, repeat steps 1 to 3; otherwise, proceed to step 5.
[0011] Step 5: Deploy the trained BP neural network model to the ARM side of ZYNQ, verify it using test set data, statistically analyze the recognition of various interference signals under different interference-to-noise ratios and record the time taken to recognize a pulse signal, and plot the correct recognition rate curve and confusion matrix.
[0012] Step 6: Deploy the trained BP neural network model to the ZYNQ FPGA, verify it using test set data, statistically analyze the recognition of various interference signals under different interference-to-noise ratios and record the time taken to recognize a pulse signal, and plot the correct recognition rate curve and confusion matrix.
[0013] Furthermore, training and test sets of interference signal data samples were generated. In MATLAB, eight types of signals were modeled, including Gaussian white noise, single-tone interference, multi-tone interference, noise amplitude modulation interference, noise frequency modulation interference, intermittent sampling direct forwarding interference, intermittent sampling repeated forwarding interference, and spectral dispersion interference. Seven of the interference signals were generated under interference-to-noise ratio (IRR) conditions ranging from -5dB to 20dB with a step size of 1dB. Training set data of 500 sample pulses were generated under each IRR condition. Test set data of 100 sample pulses were generated under each IRR condition ranging from -10dB to 20dB with a step size of 2dB.
[0014] Furthermore, we extract multiple feature parameters that differ most significantly from each other; first, we perform power normalization and Hamming windowing on the training and test sets; then, by comparing the parameter differences of different signals at various interference-to-noise ratios, we select the six features that differ most significantly from each other: parameter temporal envelope fluctuation. Time-domain moment kurtosis Frequency domain envelope fluctuation Frequency domain peak factor Frequency domain moment skewness and frequency domain moment kurtosis .
[0015] Furthermore, a backpropagation (BP) neural network is constructed and trained. The BP neural network includes an input layer, a hidden layer, and an output layer; the input layer has 6 neurons, corresponding to six feature parameters; the hidden layer has 13 neurons, used to realize the nonlinear mapping of the input data; the output layer has 8 neurons, corresponding to eight signal labels; the BP neural network is trained using the Levenberg-Marquardt optimization algorithm, which enables the network to converge faster and reduces the possibility of getting trapped in local optima to a certain extent.
[0016] Furthermore, the accuracy of the trained model is verified. Validation is performed using pre-generated test set data. The test set data is input into the BP neural network model, and the correct recognition rate curve and confusion matrix are plotted. A threshold of 98% is set; if the average recognition rate does not exceed 98% when the JNR is 0dB, steps 1 to 3 are repeated; if it does, step 5 is executed.
[0017] Furthermore, the trained BP neural network model was deployed on the ZYNQ ARM platform and statistically validated. The network model parameters were exported and loaded onto the ARM platform. The temporal envelope variability, temporal moment kurtosis, frequency envelope variability, frequency peak factor, frequency moment skewness, and frequency moment kurtosis of the test set data were calculated sequentially. The forward propagation output of the network was calculated using the six calculated feature parameters and the network model parameters. The identification of various interference signals under different interference-to-noise ratios was statistically analyzed, and the time taken to identify a single pulse signal was recorded. Correct identification rate curves and confusion matrices were plotted.
[0018] Furthermore, the trained BP neural network model was deployed on the ZYNQ FPGA platform and statistically validated. The network model parameters were exported and loaded into the FPGA platform ROM core. The temporal envelope ripple, temporal moment kurtosis, frequency envelope ripple, frequency peak factor, frequency moment skewness, and frequency moment kurtosis of the test set data were calculated in parallel. The forward propagation output of the network was calculated using the six calculated feature parameters and the network model parameters. The identification of various interference signals under different interference-to-noise ratios was statistically analyzed, and the time taken to identify a pulse signal was recorded. Correct identification rate curves and confusion matrices were plotted.
[0019] A computer device includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps of the method described above.
[0020] A computer-readable storage medium having a computer program stored thereon that, when executed by a processor, implements the above-described method.
[0021] A computer program product includes a computer program that, when executed by a processor, implements the above-described method.
[0022] Compared with the prior art, the significant advantages of the present invention are:
[0023] (1) The present invention uses only 6 manually extracted feature parameters, but when the noise-to-interference ratio reaches 0dB or above, the average correct recognition rate of the eight signals can reach 99%;
[0024] (2) The identification method designed in this invention can be deployed on ZYNQ's ARM platform and FPGA platform respectively, and the hardware resources used are relatively small;
[0025] (3) The recognition method designed in this invention ensures that the correct recognition rate is not significantly different from the simulation results on ARM and FPGA hardware platforms, while the recognition response rate is fast and can quickly give the recognition result in a short time. The recognition time for a pulse signal on the ARM platform is 2245.327us, and the recognition time for a pulse signal on the FPGA platform is 27.635us. Attached Figure Description
[0026] Figure 1 This is a flowchart of the radar jamming signal identification method based on ZYNQ according to the present invention.
[0027] Figure 2 Flowchart for ARM-based program design.
[0028] Figure 3 Structure diagram for FPGA-side program design and implementation
[0029] Figure 4 The structure diagram for FPGA-side power normalization implementation.
[0030] Figure 5 The structure diagram for implementing the time-domain envelope ripple at the FPGA side is shown.
[0031] Figure 6 This is a structural diagram of the time-domain moment kurtosis implementation on the FPGA side.
[0032] Figure 7 The structure diagram for implementing frequency domain envelope ripple at the FPGA end.
[0033] Figure 8 The structure diagram for implementing the frequency domain peak factor on the FPGA side.
[0034] Figure 9 The structure diagram for the frequency domain moment skewness implementation of the FPGA is shown.
[0035] Figure 10 This is a structural diagram of the frequency domain moment kurtosis implementation on the FPGA side.
[0036] Figure 11 This is a comparison chart of the 15-segment method and the actual value of the sigmoid function.
[0037] Figure 12 This is a graph showing the correct recognition rate of the MATLAB simulation test set.
[0038] Figure 13 This is the confusion matrix for interference signal identification in MATLAB under interference-to-noise ratio conditions of -4 to 20 dB.
[0039] Figure 14 The results are measured at the ARM terminal under a noise-to-interference ratio of 20dB.
[0040] Figure 15 This is a graph showing the correct recognition rate of the ARM-based test set.
[0041] Figure 16 This is a confusion matrix for identifying interference signals on the ARM side under interference-to-noise ratio conditions of -4 to 20dB.
[0042] Figure 17 The image shows the simulation results on the FPGA side.
[0043] Figure 18 This is a graph showing the correct recognition rate of the FPGA simulation test set.
[0044] Figure 19 The interference signal identification confusion matrix is the result of FPGA simulation under interference-to-noise ratio conditions of -4 to 20dB.
[0045] Figure 20 This is a diagram showing the resource consumption on the FPGA side.
[0046] Figure 21 This is a timing report diagram for the FPGA side.
[0047] Figure 22 This is a diagram showing the verification results on the FPGA side of the board. Detailed Implementation
[0048] To further illustrate the technical means and effects adopted by the present invention to achieve the intended purpose, the following describes in detail a radar jamming signal identification method based on ZYNQ proposed by the present invention, in conjunction with the accompanying drawings and specific embodiments.
[0049] Please see Figure 1 The present invention provides a method for identifying radar interference signals based on ZYNQ, the specific steps of which are as follows:
[0050] Step 1: Construct training and test sets of noise and interference signals. In MATLAB, model eight types of signals: Gaussian white noise (no interference), single-tone interference, multi-tone interference, noise amplitude modulation interference, noise frequency modulation interference, intermittent sampling direct forwarding interference, intermittent sampling repeated forwarding interference, and spectral dispersion interference. Add noise to the interference signals to generate a training set of 500 sample pulses for each interference ratio (IRR) condition (-5dB to 20dB, with 1dB steps), and a test set of 100 sample pulses for each IIRR condition (-10dB to 20dB, with 2dB steps).
[0051] Furthermore, the power of the single-tone interference signal is concentrated on a single frequency, which is a single-frequency continuous sine wave, and its complex signal expression is:
[0052]
[0053] in For amplitude, For the center frequency, The initial phase is subject to the following: The uniform distribution.
[0054] Furthermore, the multi-tone interference is composed of the superposition of multiple random single-tone interferences at their center frequencies, and its complex signal expression is:
[0055]
[0056] in , , These represent the amplitude, center frequency, and initial phase of different single-tone interferences. obey They are evenly distributed, uncorrelated, and independent of each other. It represents the number of single-tone interferences.
[0057] Furthermore, the noise amplitude modulation interference modulates the amplitude of the carrier signal using a noise signal without changing the carrier frequency; its complex signal expression is as follows:
[0058]
[0059] in The DC component of the carrier signal. It is Gaussian white noise with a mean of 0 and a variance of . , For carrier frequency, The initial phase is subject to the following: The uniform distribution, and They are unrelated and independent of each other.
[0060] Furthermore, the noise frequency modulation interference modulates the signal frequency using the noise signal without changing the amplitude; its complex signal expression is as follows:
[0061]
[0062] in The DC component of the carrier signal. For carrier frequency, Frequency modulation index, It is Gaussian white noise with a mean of 0 and a variance of . , The initial phase is subject to the following: The uniform distribution, and They are unrelated and independent of each other. It is a Wiener process, subject to Gaussian distribution, frequency modulation index and variance Together, they determine the effective bandwidth of noise frequency modulation. The correspondence is as follows:
[0063]
[0064] Furthermore, the intermittent sampling and direct forwarding slices the intercepted radar transmitted signal with a fixed time width, then amplifies it before direct forwarding, and then intercepts and forwards it again until the signal pulse ends. Assuming the transmitted signal is a linear frequency modulated signal with a bandwidth of [missing information], [missing information]. Pulse width is The starting frequency is The expression is:
[0065]
[0066] The expression for intermittent sampling direct forwarding interference is:
[0067]
[0068] in To capture the total number of slices within a radar pulse, The slice width is determined according to the generation principle. , With radar signal pulse width The relationship is as follows:
[0069]
[0070] Furthermore, the intermittent sampling direct forwarding involves repeatedly forwarding slices of the intercepted radar transmitted signal over a fixed time width until the pulse ends. When the number of repeated forwardings is 1, it is considered direct forwarding. The expression is:
[0071]
[0072] in It represents the number of times the slice is repeatedly forwarded. It is the total number of slices. It is the slice width. , , With pulse width The following relationship must be satisfied:
[0073]
[0074] Furthermore, the spectral dispersion interference consists of multiple identical sub-pulses. These sub-pulses are obtained by decimating the radar transmitted signal at a frequency that is n times the radar transmitted signal sampling rate. Since the sub-pulse decimation frequency is n times the transmitted signal sampling rate, the frequency modulation slope is also n times the transmitted signal sampling rate. The sub-pulse expression is:
[0075]
[0076] The extracted sub-pulses are copied n times to form a spectral dispersion interference, expressed as:
[0077]
[0078] Step 2: The signals in the training and test sets are preprocessed using power normalization and Hamming windowing. Different feature parameters are calculated, and the six most significantly different feature parameters are extracted. To avoid large fluctuations in feature parameter values due to varying interference signal power, which could lead to poor recognition results, power normalization is performed first. The expression is as follows:
[0079]
[0080] in The original signal, The signal is after power normalization. This refers to the signal length. Before extracting frequency domain feature parameters from a time-domain signal, an FFT process is required. Adding a Hamming window can mitigate the impact of spectral leakage caused by FFT processing on finite-length signals, thus improving the recognition effect. The six feature parameters are the time-domain envelope ripple. Time-domain moment kurtosis Frequency domain envelope fluctuation Frequency domain peak factor Frequency domain moment skewness and frequency domain moment kurtosis The feature parameters are determined by comparing the parameter differences of different signals under various interference-to-noise ratios, and are used to improve the discriminative power of signal recognition. The expressions and physical meanings of the six feature parameters are as follows:
[0081]
[0082] in It is the standard deviation of the signal amplitude. It is the mean of the signal amplitude, and the time-domain envelope undulation characterizes the degree of undulation of the signal's time-domain waveform.
[0083]
[0084] These are discrete values of the time-domain signal. It is the standard deviation of the time-domain signal. It is the mean of the time-domain signal. The mean value is taken, and the time-domain moment kurtosis characterizes the steepness of the signal's time-domain waveform.
[0085]
[0086] in It is the standard deviation of the signal spectrum. It is the mean of the signal spectrum, similar to the time-domain envelope variability, while the frequency-domain envelope variability characterizes the degree of fluctuation in the signal spectrum.
[0087]
[0088] It is the amplitude of the signal spectrum. It is the root mean square of the signal spectrum amplitude, and the peak factor characterizes the sharpness of the signal distribution.
[0089]
[0090] It is the amplitude of the signal spectrum. It is the root mean square of the signal's spectral amplitude. It is the average amplitude of the signal spectrum. The mean value is taken, and the frequency domain moment skewness characterizes the degree of asymmetry of the signal spectrum relative to its mean value.
[0091]
[0092] It is the amplitude of the signal spectrum. It is the root mean square of the signal's spectral amplitude. It is the average amplitude of the signal spectrum. The mean value is taken, and the time-domain moment kurtosis characterizes the steepness of the signal spectrum.
[0093] Step 3: Construct and train the BP neural network. The BP neural network consists of an input layer, hidden layers, and an output layer. The input layer has 6 neurons corresponding to six feature parameters, the output layer has 8 neurons corresponding to eight signal labels, and the number of neurons in the hidden layer is determined by... It was determined to be 13. The input layer contains neurons. The BP neural network is trained using the Levenberg-Marquardt optimization algorithm. It exhibits quadratic convergence speed similar to Newton's method when approaching the optimal solution, and performs as stably as gradient descent when far from the optimal solution. A damping coefficient is adjusted to balance these two aspects, enabling the network to converge faster and reducing the likelihood of getting trapped in local optima to some extent. The weight update formula for the LM algorithm is:
[0094]
[0095] in and The values before and after the weight and bias update. It is the Jacobian matrix, i.e., the partial derivative of the error with respect to the weights or biases. It is the introduced damping coefficient. It is the identity matrix. This is the error vector. The optimization objective of a BP neural network is to minimize the sum of squared errors. If the error is after the update This indicates the current situation. Too large, needs to be reduced To make the algorithm closer to Newton's method and accelerate convergence, if the error is updated... This indicates the current situation. Too small, needs to be enlarged This makes the algorithm closer to gradient descent and maintains stability.
[0096] Step 4: Use test set data to test the model's recognition accuracy. Feed the pre-generated test set data into the trained BP neural network, and use the output results to plot the correct recognition rate curve and confusion matrix. If the result differs significantly from the recognition result on the training set, return to Step 1, regenerate the data, extract feature parameters, and train the network. If the difference is not significant, proceed to the next step.
[0097] Step 5: Deploy the trained BP neural network model to the ZYNQ ARM platform and perform statistical validation. The ARM platform programming flow is as follows: Figure 2 As shown, firstly, the model parameters are exported to the ARM, and the pre-generated test set data is written to the ARM's DDR using the command `dow -data C: / Users / Levnon / Desktop / test_data / type4.bin 0x1FFF0000`. Then, the data of one pulse point at a time is read from the DDR for preprocessing and serial calculation of feature parameters. Next, the calculated time-domain envelope fluctuation is... Time-domain moment kurtosis Frequency domain envelope fluctuation Frequency domain peak factor Frequency domain moment skewness and frequency domain moment kurtosis The six feature parameters are normalized using the mapminmax method, and the formula is as follows:
[0098]
[0099] The forward propagation process is calculated serially using the six normalized feature parameters and network model parameters. The recognition results are output and statistically analyzed. Based on the statistical results, the correct recognition rate curves and confusion matrices of various signals under different interference-to-noise ratios are plotted.
[0100] Step 6: Deploy the trained BP neural network model to the ZYNQ FPGA and perform statistical validation. The FPGA program design and implementation structure is as follows: Figure 3 As shown. First, the model parameters and pre-generated test set data are converted to Q10.9 format to reduce the computational complexity of subsequent FPGA calculations. The Q10.9 format includes 1 sign bit, 10 integer bits, and 9 decimal bits. Then, the fixed-point format model parameters and test set data are stored in the ROM core of the FPGA segment. Next, the data of one pulse point is read sequentially from the ROM core for preprocessing and parallel calculation of characteristic parameters. The power normalization implementation structure is as follows. Figure 4 As shown, an FFT is performed before calculating the frequency domain features. The FFT is implemented using the Fast Fourier Transform ip. The calculation structures for the six feature parameters are as follows: Figures 5-10As shown. The calculated time-domain envelope fluctuations... Time-domain moment kurtosis Frequency domain envelope fluctuation Frequency domain peak factor Frequency domain moment skewness and frequency domain moment kurtosis The six feature parameters are normalized using MapMinMax. The forward propagation process is then calculated sequentially using the normalized six feature parameters and the network model parameters. The sigmoid activation function is approximated using the 15-segment method, which saves resources. Approximate values are shown in Table 1. Simulation results obtained using this approximation of the sigmoid function are compared with the actual values. Figure 11 As shown. Output and statistically analyze the recognition results, and plot the correct recognition rate curves and confusion matrices for various signals under different interference-to-noise ratios based on the statistical results.
[0101] Table 1
[0102]
[0103] To verify and visually demonstrate the correctness of the trained network model, it was validated using test set data in MATLAB. Based on the recognition results obtained from the MATLAB simulation, the correct recognition rate curves for various signals under different interference-to-noise ratios were plotted as follows: Figure 12 As shown in the figure. When the interference-to-noise ratio (INR) reaches -2dB, the correct recognition rate of all signals exceeds 92%; when the INR reaches 0dB or above, the correct recognition rate of all signals exceeds 97%, and the average correct recognition rate of the eight signals exceeds 99%. Since the training samples do not include sample data in the range of -10dB to -6dB, the confusion matrix in the range of -4dB to 20dB is plotted as follows. Figure 13 As shown, within the range of -4dB to 20dB, the overall correct recognition rates for noise FM interference and spectral dispersion interference, which had the worst recognition accuracy, were 97.2% and 97.8%, respectively, while the overall correct recognition rates for other signals all exceeded 99%.
[0104] The trained network model was deployed to the ARM platform for verification, with the external crystal oscillator providing a 33.33MHz reference clock for the ARM side. Taking a 20dB interference noise ratio as an example, the measured results are as follows: Figure 14 As shown, the time required to identify a single pulse signal is 1796261.67 ÷ 800 = 2245.327 μs. Based on the statistical results, the correct recognition rate curves for various signals under different interference-to-noise ratios are plotted as follows: Figure 15As shown, when the interference ratio (IRR) reaches -2dB, the correct recognition rate of all signals exceeds 92%. When the IIRR reaches 0dB or higher, the correct recognition rate of all signals exceeds 96%, and the average recognition rate of the eight signals exceeds 99%. There is a slight regression compared to the MATLAB simulation results, which is due to errors caused by the different computational precision of different platforms and is within an acceptable range. The confusion matrix in the range of -4dB to 20dB is shown below. Figure 16 As shown, the overall accuracy rates for noise FM interference and spectral dispersion interference, which had the worst recognition rates, were 97.2% and 97.9%, respectively, while the overall correct recognition rates for other signals all exceeded 99%.
[0105] The trained network model was deployed to an FPGA platform for verification, with the FPGA clock speed at 200MHz. Taking a 20dB interference-to-noise ratio (INR) as an example, the FPGA simulation verification results are as follows: Figure 17 As shown, the time required to identify a pulse signal is 4421600 × 5 ÷ 1000 ÷ 800 = 27.635 μs. Based on the simulation results, the correct recognition rate curves for various signals under different interference-to-noise ratios are plotted as follows: Figure 18 As shown, when the interference ratio (IRR) reaches -2dB, the correct recognition rate of all signals exceeds 84%. When the IRR reaches 0dB or higher, the correct recognition rate of all signals exceeds 95%, and the average correct recognition rate of the eight signals exceeds 99%. There is a slight regression compared to the MATLAB simulation results. This is due to errors caused by the fixed-point quantization using the Q10.9 format during data storage and the truncation of the calculation results, which are unavoidable and within an acceptable range. The confusion matrix in the range of -4dB to 20dB is shown below. Figure 19 As shown, the overall accuracy rates for noise FM interference and spectral dispersion interference, which had the worst recognition rates, were 96.8% and 96.7%, respectively. The overall correct recognition rates for other signals all exceeded 99%. The resource usage and timing violations after FPGA compilation are as follows... Figure 20 and Figure 21 As shown, there were no timing violations. Due to the large amount of signal data, a total of 900 BRAMs were used to store the raw I and Q channels. Apart from this, the usage of other resources did not exceed 20%. Taking a 20dB interference-to-noise ratio as an example, the FPGA board verification results are as follows... Figure 22 As shown, the results are consistent with the FPGA simulation results.
[0106] The above description, in conjunction with specific preferred embodiments, provides a further detailed explanation of the present invention. It should not be construed that the specific implementation of the present invention is limited to these descriptions. For those skilled in the art, various simple deductions or substitutions can be made without departing from the concept of the present invention, and all such modifications and substitutions should be considered within the scope of protection of the present invention.
Claims
1. A method for identifying radar jamming signals based on ZYNQ, characterized in that, It includes the following steps: Step 1: Construct training and test sets for Gaussian white noise and seven types of radar jamming signals under different interference-to-noise ratios in MATLAB. The seven types of radar jamming signals include single-tone jamming, multi-tone jamming, noise amplitude modulation jamming, noise frequency modulation jamming, intermittent sampling direct forwarding jamming, intermittent sampling repeated forwarding jamming, and spectral dispersion jamming. Step 2: The signals of the training set and test set are preprocessed by power normalization and Hamming windowing, different feature parameters of the signals are calculated, and the six feature parameters with the greatest differences among them are extracted. Step 3: Add the real labels to the six feature parameters of the training set under different interference-to-noise ratios to construct and train the BP neural network model; Step 4: Use the test set data to verify the model's recognition accuracy. If the average recognition rate does not exceed the set threshold when the drying ratio is 0dB, repeat steps 1 to 3; otherwise, proceed to step 5. Step 5: Deploy the trained BP neural network model to the ARM side of ZYNQ, verify it using test set data, statistically analyze the recognition of various interference signals under different interference-to-noise ratios and record the time taken to recognize a pulse signal, and plot the correct recognition rate curve and confusion matrix. Step 6: Deploy the trained BP neural network model to the ZYNQ FPGA, verify it using test set data, statistically analyze the recognition of various interference signals under different interference-to-noise ratios and record the time taken to recognize a pulse signal, and plot the correct recognition rate curve and confusion matrix.
2. The radar jamming signal identification method based on ZYNQ according to claim 1, characterized in that, The training and test sets of interference signal data samples were generated. Specifically, eight types of signals were modeled in MATLAB: Gaussian white noise, single-tone interference, multi-tone interference, noise amplitude modulation interference, noise frequency modulation interference, intermittent sampling direct forwarding interference, intermittent sampling repeated forwarding interference, and spectral dispersion interference. Seven of the interference signals were generated under interference-to-noise ratio (IRR) conditions ranging from -5dB to 20dB with a step size of 1dB. Training set data of 500 sample pulses were generated under each IRR condition. Test set data of 100 sample pulses were generated under each IRR condition ranging from -10dB to 20dB with a step size of 2dB.
3. The radar jamming signal identification method based on ZYNQ according to claim 1, characterized in that, Extract multiple feature parameters that differ most from each other; first, perform power normalization and Hamming windowing on the training and test data; By comparing the parameter differences of different signals under various interference-to-noise ratios, six features with the greatest differences among them were selected: time-domain envelope undulation, time-domain moment kurtosis, frequency-domain envelope undulation, frequency-domain peak factor, frequency-domain moment skewness, and frequency-domain moment kurtosis.
4. The radar jamming signal identification method based on ZYNQ according to claim 1, characterized in that, A backpropagation (BP) neural network is constructed and trained. The BP neural network includes an input layer, a hidden layer, and an output layer. The input layer has 6 neurons, corresponding to six feature parameters. The hidden layer has 13 neurons, used to realize the nonlinear mapping of input data. The output layer has 8 neurons, corresponding to eight signal labels. The BP neural network is trained using the Levenberg-Marquardt optimization algorithm.
5. The radar jamming signal identification method based on ZYNQ according to claim 4, characterized in that, Verify the accuracy of the trained model; use pre-generated test set data for verification, input the test set data into the BP neural network model, and plot the correct recognition rate curve and confusion matrix.
6. The radar jamming signal identification method based on ZYNQ according to claim 1, characterized in that, The trained BP neural network model was deployed on the ZYNQ ARM platform and statistically validated. The network model parameters were exported and loaded onto the ARM platform. The temporal envelope undulation, temporal moment kurtosis, frequency envelope undulation, frequency peak factor, frequency moment skewness, and frequency moment kurtosis of the test set data were calculated sequentially. The forward propagation output of the network was calculated using the six calculated feature parameters and the network model parameters. The identification of various interference signals under different interference-to-noise ratios was statistically analyzed, and the time taken to identify a pulse signal was recorded. The correct identification rate curve and confusion matrix were plotted.
7. The radar jamming signal identification method based on ZYNQ according to claim 1, characterized in that, The trained BP neural network model was deployed on the ZYNQ FPGA platform and statistically validated. The network model parameters were exported and loaded into the FPGA platform ROM core. The temporal envelope ripple, temporal moment kurtosis, frequency envelope ripple, frequency peak factor, frequency moment skewness, and frequency moment kurtosis of the test set data were calculated in parallel. The forward propagation output of the network was calculated using the six calculated feature parameters and the network model parameters. The identification of various interference signals under different interference-to-noise ratios was statistically analyzed, and the time taken to identify a pulse signal was recorded. The correct identification rate curve and confusion matrix were plotted.
8. A computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the steps of any of the methods described in claims 1-7.
9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the method described in any one of claims 1-7.
10. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the method described in any one of claims 1-7.