High-frequency ground wave radar radio frequency interference slow time domain suppression method (2)

        To use the method in this article to suppress RFI, first divide every 100 frequency sweep cycles into one segment. Since the simulated RFI only exists in the 301 to 400 frequency sweep cycles, only this segment of data needs to be processed to verify the effectiveness of the algorithm. This data is the data of all distance elements and receiving channels in the 301st to 400th sweep cycle. The long-distance metadata is selected to construct the training tensor. Here 497 corresponds to 0 distance elements. The distance element is inverse. Select 457 to 466. Total For the data of 10 distance elements, the three expansion mode matrices obtained from the training tensor are subjected to SVD in sequence. The normalized singular value distribution is shown in Figure 6. The detection threshold is set to 0.1. If the normalized singular value is greater than 0.1, then Large singular values. From Figure 6, we can see that the three expansion pattern matrices have 2, 4, and 3 large singular values ​​respectively. The first expansion pattern matrix corresponds to the first two rows of the right singular matrix, and the second expansion pattern matrix corresponds to the left singular value. Perform FFT on the first four columns of the singular matrix and the first three rows of the third expansion pattern matrix, as shown in Figure 7. It can be seen that the first expansion pattern matrix corresponds to the first row of the right singular matrix, and the second expansion pattern matrix corresponds to the left The first column of the singular matrix and the third expanded mode matrix correspond to the first row of the right singular matrix. There are two obvious spectral peaks in the FFT spectrum. The spectral peak frequencies are -0.48 Hz and 0.35 Hz respectively, which are the same as the simulated RFI Doppler frequency. .

        Table 2 lists the average power and peak power of the FFT spectrum in each row and column. From the table, it can be read that the average power in the first row or column of the singular matrix corresponding to each mode matrix is ​​-14.794 6, -12.629 9 and - 16.002 4 dB, significantly lower than 0 dB. Through verification and analysis of a large amount of data, the following conclusion is reached: If the row vector or column vector corresponds to the noise subspace, after doing FFT, the average spectrum power fluctuates around 0 dB, and the floating range is general. No more than 5 dB. Therefore, the two indicators of spectrum peak power and power mean can be used to determine whether the row vector or column vector corresponds to the noise subspace or the interference subspace, and the Doppler frequency position of RFI on the RD spectrum can be determined by the frequency corresponding to the spectrum peak. To sum up, the first expansion mode matrix corresponds to the first row of the right singular matrix, the second expansion mode matrix corresponds to the first column of the left singular matrix, and the third expansion mode matrix corresponds to the first row of the right singular matrix, which corresponds to the two RFIs of the simulation. The component subspace, and the remaining row and column vectors correspond to the noise subspace, thus determining the parameters of the RFI subspace projection matrix estimated by the first, second, and third expansion mode matrices as d1=1, d2=1, and d3=1 respectively. .

        After estimating the RFI subspace, project it to each processing tensor and eliminate RFI, and draw the RP spectrum and RD spectrum as shown in Figure 8 (a) and (b) respectively. It can be seen from the RP spectrum that between 301 and 400 The RFI added during the frequency sweep cycle is completely eliminated, and the RFI is not visible at all on the RD spectrum, and the positive first-order Bragg peak is highlighted.

3.2 Actual measurement experiment

        The practicality of the RFI suppression algorithm in this article was verified using the measured data of dual-station dual-frequency network detection in Dongshan and Longhai in August 2021. The radar system parameters are shown in Table 3.

        The data of Dongshan high frequency at 07:40 on August 11 was selected. The data in this time period was contaminated by RFI. The spectrum monitoring data for this period were spliced ​​and divided into 12 segments, and FFT was performed on each segment of data to obtain environmental noise spectrum diagrams of different segments. According to Table 3, it can be seen that the radar operating frequency is 12.500 MHz, the sweep period is 30 kHz, and the sweep mode is upsweep, so the detection window range is set to 12.500 MHz to 12.530 MHz, and the noise floor window range is 20~24 MHz. Find the difference between the peak power of the detection window and the average power of the noise floor window, and set the detection threshold to 20 dB. 12 segments of RFI flag bits are obtained, among which the data flag bits of segments 1 to 3 and 8 to 12 are 1, that is, the data in the frequency sweep periods from 1 to 300 and 701 to 1 200 are all contaminated by RFI. Draw the environmental noise spectrum diagrams of the second and fourth sections as shown in Figure 9. The red dotted box indicates the range of the detection window. Comparing the two, you can see that the noise spectrum diagram of the second section has obvious peaks in the detection window, and The noise spectrum diagram of the fourth section has no obvious peaks in the detection window.

         The measured data RP spectrum and RD spectrum are shown in Figure 10 (a) and (b) respectively. It can be seen that the frequency sweep period segments contaminated by RFI in the RP spectrum are consistent with the detection results obtained from the spectrum monitoring data. In the RD spectrum On the top, the RFI appears as a strip of vertical stripes with a Doppler frequency of approximately 0.52 Hz, which is broadened to a certain extent and covers the negative first-order peak area of ​​the ocean echo spectrum received from Longhai Fadongshan. The measured radar data - 80 to 0 distance element is the migration of 0 to 80 distance element, that is, the negative distance element no longer only includes RFI, and because it is a dual-station network, one data includes Dongshan and Longhai dual stations. The echoes in Figure 10(b) are the Bragg peaks at 0 to 15 distance units, which are spontaneously received echoes from Dongshan Station, and the 40 to 55 distance units are ocean echoes received by Dongshan Station from Longhai, and due to the high and low frequency emissions Each has two channels, so Doppler offset is added to distinguish the echoes of the two signals, resulting in a symmetrical distribution of the ocean echo generated by the two signals in the RD spectrum. Ocean echo occupies the majority of distance elements. In order to ensure the RFI suppression effect, distance elements that do not contain ocean echo need to be adaptively selected for training. Set the training range to 10 distance elements. First, read the distance offset of the dual-station dual-frequency through the parameter configuration file. Based on the distance offset, determine the approximate range of the distance elements occupied by the ocean echo of the dual-station. Outside this range, take 10 distance elements for training.

        Taking the second piece of data, that is, the data of 101 to 200 frequency sweep periods, as an example for analysis, the training distance element is selected from 61 to 70, and the normalized singular value distribution of each expansion mode matrix of the training tensor is obtained, as shown in Figure 11, it can be obtained The numbers of large singular values ​​corresponding to the three expansion mode matrices are 3, 3, and 2 respectively. The FFT spectra of each row and column are shown in Figure 12. The average power and peak power are shown in Table 4. The analysis shows that the first expansion The mode matrix corresponds to the first row of the right singular matrix, the second expanded mode matrix corresponds to the first column of the left singular matrix, and the third expanded mode matrix corresponds to the first row of the right singular matrix corresponding to the RFI subspace. From the first two rows of the third expanded mode matrix, FFT It can be seen in the spectrum that there are two obvious peaks, one of which has a peak frequency around 0.52 Hz, corresponding to the RFI component in the RD spectrum, and the other peak frequency is around 0 Hz, which corresponds to the zero-frequency interference in the RD spectrum ( Zero Frequency Interference (ZFI) component can eliminate zero-frequency interference altogether. Determine the parameters of the RFI subspace projection matrix estimated by the first, second, and third expansion mode matrices as d1=1, d2=1, and d3=2 respectively. Project the interference subspace to the tensor to be processed to complete this segment of data. RFI suppression.

        Each segment of data is processed in sequence according to the RFI flag bit, and the RP spectrum and RD spectrum after interference suppression are obtained, as shown in Figure 13 (a) and (b) respectively. It can be seen from the figure that the RFI of each sweep period in the RP spectrum has been suppressed, leaving only a small amount of residual amounts with a low signal-to-noise ratio. It can be seen from the RD spectrum that after the second FFT, these The residual amount is basically reduced to the noise floor level, RFI is completely eliminated, and the masked first-order Bragg peak is highlighted. Figure 14(a) shows the range spectrum comparison chart of the 800th sweep cycle before and after RFI suppression. It can be seen that RFI is suppressed, the ocean echo is not affected, and the signal-to-noise ratio increases by about 10 dB. Figure 14(b) is a comparison chart of the 42nd range element Doppler spectrum before and after RFI suppression. It can be seen that the radio frequency interference with the Doppler frequency around 0.52 Hz and the zero-frequency interference at 0 Hz are completely eliminated, and the ocean echo Completely retained, the signal-to-noise ratio of the Bragg peak is improved.

3.3 Statistical analysis

        The RFI flag bits obtained from the spectrum monitoring data are used to draw the time distribution of the dual-frequency echo data of Dongshan and Longhai dual stations polluted by RFI on August 11, 2021, as shown in Figure 15, among which the proportion of time when Dongshan high-frequency RFI exists 12.93%, Dongshan low-frequency RFI has existed for 22.94% of the time, Longhai high-frequency RFI has existed for 53.68% of the time, and Longhai low-frequency RFI has existed for 0.07% of the time. Overall, Longhai low-frequency data quality is better, and other data are greatly affected by RFI. Using the algorithm in this article to perform indiscriminate suppression of dual-station and dual-frequency data for the whole day on August 11, 2021 took 3 494.531 373 s. After adding the RFI detection algorithm, it took 1 670.540 791 s. Computer GPU model used for testing: Intel(R) Core(TM) i5⁃9400 CPU @ 2.90 GHz. In data batch processing, adding RFI detection algorithm can effectively reduce interference elimination time and improve data batch processing efficiency.

        By viewing the RP spectrum after interference suppression, the remaining RFI time distribution after interference suppression is obtained, as shown in Figure 16. The proportion of RFI existence time is 0.21%, 2.66%, 0.21%, and 0% in order. It can be seen that the interference has not been completely eliminated, and there is also a small amount of RFI in the time period when the original spectrum monitoring data did not detect RFI. Dongshan low-frequency data is still affected by RFI for a long period of time. The analysis reasons are mainly as follows: First, the setting of the RFI detection signal-to-noise ratio threshold is slightly large, resulting in missed alarms, so that RFI is not effectively detected in some time periods; second, RFI is not always composed of a series of single-frequency signals, and they may undergo various modulations, resulting in The distance correlation is weakened and the suppression method of subspace projection fails to work. However, in general, most RFIs are detected and suppressed, proving that the algorithm in this paper is effective.

4 Conclusion

        This paper proposes a slow time domain segmentation detection and suppression method for high-frequency ground wave radar RFI suppression. This method solves the problems of traditional subspace methods without RFI detection, strict judgment criteria for interference-free subspaces, and unfocused right singular matrices. Issues with row vector information. In the preprocessing stage, spectrum monitoring data is used to perform RFI detection on slow time domain segmented data. In the post-processing stage, the HOSVD method is used to process the channel-distance-slow-time dimension data obtained after the first FFT of the radar received data. First, the data to be processed is segmented in the slow-time dimension. For data segments with RFI , adaptively select the appropriate distance metadata to construct the training tensor, and analyze the three expansion mode matrix forms of the training tensor, combined with the left singular matrix column vector or right singular matrix row vector obtained after singular value decomposition of each expansion mode matrix. The frequency information contained provides a parameter determination method for interference subspace estimation, accurately estimates the interference subspace, and then obtains the signal subspace, completing RFI suppression in the slow time domain. Experimental results show that this method can effectively detect and suppress RFI and improve the efficiency of data batch processing.