Analysis of Persistent RFI Signals Captured Using the CISR Coherent Sampling Mode

Transcription

1 Analysis of Persistent RFI Signals Captured Using the CISR Coherent Sampling Mode S.W. Ellingson and K.H. Lee February 13, 26 Contents 1 Introduction 2 2 Methodology Hardware Configuration and Data Capture Data Analysis Results File 29, Block File 29, Block File 27, Block File 218, Block File 218, Block Conclusions 16 Bradley Dept. of Electrical & Computer Engineering, 34 Whittemore Hall, Virginia Polytechnic Institute & State University, Blacksburg VA 2461 USA. ellingson@vt.edu 1

2 1 Introduction This report summarizes a study of radio frequency interference (RFI) in the CISR capture data sets from airborne experiments previously reported in [1]. In these experiments, the CISR digital backend was integrated with NOAA s PSR/C system in a co-observing configuration [2]. PSR/C itself is strictly a total power (incoherent) sampling radiometer, with time-frequency resolution limited to milliseconds and 1 s of MHz, respectively. The severe RFI environment in C-band significantly degrades the quality of radiometry obtained by PSR/C and other instruments. At the same time, the limited time-frequency resolution of these instruments provides only limited information as to the nature of the RFI. In these experiments, the CISR backend is used in parallel with the PSR/C backend to provide greatly enhanced time-frequency resolution, offering improved ability to identify and characterize RFI. The results may be used to devise RFI mitigation algorithms which could be incorporated into CISR and future C-band radiometers. Our previous report [1] was a preliminary survey of the data to characterize the types and frequency of occurrence of RFI. The signals identified in that report were classified as being either pulsed or persistent. The pulsed signals are relatively easy to recognize, and were judged to be emissions from radars. The persistent signals are more difficult to recognize, and are the focus of this report. Here, we focus on a few examples of persistent signals from the data set. Being persistent, it is likely that at least some of these signals are communications signals. The purpose of this report is to characterize these signals and attempt to identify them. 2 Methodology 2.1 Hardware Configuration and Data Capture The PSR/C front end uses a conical scanning antenna which completes an azimuthal cycle every 3 s. Simultaneously, the front end sweeps from. GHz to 7.7 GHz ( Cband, for the purposes of this discussion) in 22 steps, completing 3.6 sweeps per azimuthal cycle. This signal is downconverted to a 12 MHz intermediate frequency (IF) using a single-sideband ( image rejection ) mixer, which also determines the current center frequency. One limitation of this architecture is that the mixer is known to have image rejection of only about 18 db. Thus, signals observed by CISR at any given frequency may actually be at a frequency 2 MHz away and 18 db greater. The observed density of RFI in the data sets appears to be sufficiently low that this does not spoil the usefulness of the data; however, since some of the data (including RFI) exhibits dynamic range in excess of 2 db, the potential for cross-talk from the image band should be kept in mind. The 12 MHz IF is split into two MHz segments: 7 12 MHz and MHz. The latter is directly digitized by one of CISR s 1-bit 2 MSPS digitizers. The former is frequency-shifted to MHz and digitized in the same manner by CISR s other digitizer. The two half-bands are downconverted, filtered, resampled, 2

3 and combined into a single 1 MSPS complex-valued signal, encoding 8 MHz contiguous bandwidth. This output is captured in blocks of 128K ( ) contiguous samples, yielding a data record 1.3 ms in length. The blocks however are not contiguous; the minimum time between blocks is at least 38 ms, although this separation is often much greater as CISR periodically cycles through other modes of operation, during which time coherent data is not captured. The CISR capture data consists of two distinct data sets. The first is that from the directory set of to 14., identified in [1] as file set 182 ; and the second is that from the directory set of on, identified in [1] as file set 391. All signals considered in this report are from file set Data Analysis Data analysis was conducted in two phases. In the first phase, the data block containing the signal of interest is examined using a procedure from [1] which produces a spectrogram as well as integrated spectrum and a time domain plot. The steps for creating these is as follows: First, the 128K samples in the block are partitioned into 1K sets of 128 samples each. Each of these sets is FFTed (without windowing) to obtain a power spectrum. The frequency response of the instrument is then removed from these spectra by baseline correction. Baseline correction involves normalizing the mean power spectral density (PSD) to unity and then dividing bin-by-bin by a precomputed baseline spectrum which is normalized to unit-mean power spectral density. The baseline spectrum used here is shown in Figure 1, and was computed using a combination of mean normalization and integration over several blocks which appeared to be RFI-free. The 1K baseline-corrected spectra are then assembled into a pixel spectrogram consisting of 128K pixels having time-frequency dimensions of khz and 1.28 µs, respectively. An example of the result in this phase of the analysis is shown in Figure 2, which shows a block containing no obvious RFI. In this report, two auxiliary plots are provided with each spectrogram: One showing PSD averaged over time (i.e., a vertical averaging of the spectrogram) and another showing total power averaged over frequency (i.e., a horizontal averaging of the spectrogram). The auxiliary plots sometimes reveal RFI not apparent in the spectrogram, and vice versa. Note that since the received noise is nominally Gaussian and the mean PSD is now unity, the units of PSD in the spectrogram are standard deviations (σ) with respect to the noise distribution. The second phase of the analysis consists of an attempt to identify the modulation by an examination of the signal s spectral and temporal characteristics, including a harmonic analysis which attempts to generate spectral lines associated with residual carrier, symbol transitions, and other cyclic features of the signal. The method is best illustrated by example. Consider a BPSK signal using mock (random) data at a symbol rate F sy = MHz, centered at +2 MHz relative to zero Hz in the complex baseband representation used here. The signal is corrupted with additive white Gaussian noise resulting in a mean signal-to-noise ratio of 3 db in the analysis bandwidth (1 MHz), or Γ = +7 db in the null-to-null bandwidth, 1 MHz. 128K 3

4 baseline PSD [db] 1 baseband freq [MHz] Figure 1: Baseline spectrum used for baseline correction. The null in the center is due to the stitching together of two separate subbands in the CISR receiver to form a single signal. The reason for the reduced gain and uneven slope of the upper subband is not known, but probably arises in the splitting and recentering of the 12 MHz analog IF. 4

5 2 1 1 baseband freq [MHz] time [µs] Figure 2: A typical spectrogram and auxiliary plots for a block with no obvious RFI.

6 samples are generated in complex baseband format at F S = 1 MHz, yielding data identical in form to that of the CISR datasets. Figure 3 shows the result after the first phase of analysis described above. In the second phase of analysis, the result from Figure 3 is used as follows: First, the occupied bandwidth B s is judged to be 1 MHz, and from this Γ is determined to be +7 db in the occupied bandwidth. The estimated center frequency is estimated using the average of the upper and lower limits of the occupied bandwidth, seen in this case to be equal to the true value of 2 MHz. The second phase of analysis continues as an attempt to determine modulation parameters. First, the estimated center frequency of the signal is spectrally shifted to zero by mixing with complex-valued local oscillator tuned to the estimated center frequency. Next, a lowpass filter is applied in an attempt to improve to the total signal to noise ratio; i.e., to make the signal-to-noise ratio in the analysis bandwidth closer to Γ. Here, we used simple custom filters obtained from the inverse Fourier transform of a weighted version of the desired brick-wall frequency response. The filters are specified in terms of number of taps M, equal to either 128 (for larger bandwidths) or 124 (for smaller bandwidths), and f lp, the low-pass cutoff frequency used in the design process. A triangular window is used to weight the frequency response prior to application of the inverse Fourier transform to obtain the filter coefficients. For this example, we use a filter with M = 128 and f lp = MHz. For digital signals, it is best to select f lp to be as close as possible to the actual symbol rate, as we have chosen here. This removes all sidelobes, leaving only the main lobe of the original spectrum. The reason for using such a severe filtering process is that this dramatically improves the performance of carrier frequency and symbol rate estimation, as discussed below. For this example, the resulting filter is shown in Figure 4. After filtering, we attempt to detect a residual carrier, a symbol rate, or other cyclic features of the signal. This done by calculating the following functions: G (ω) = F {x f (t)x f (t)} 2 ; (1) G 1 (ω) = F { x f (t)x f(t) } 2 ; (2) where x f (t) is the filtered data from the previous step, and F is the Fourier Transform, implemented using the Fast Fourier Transform (FFT). G 1 (t) is simply the power spectral density of the signal times it s conjugate. This has the effect of shifting the carrier frequency to zero and stripping away any part of the modulation which is encoded in phase. What is left is the total power in the specifically in the carrier, represented as the DC component of G 1 (t). However, if x f (t) is a single-carrier digital signal filtered null-to-null as described above, then G 1 (t) also contains spectral components at the symbol rate. This is demonstrated for our simulation example in Figure. G (t) is simply the power spectral density of the signal times itself; i.e., without conjugation. This has the effect of producing a harmonic associated the carrier frequency f c located at 2f c. However, if x(t) is a digital signal filtered null-to-null as described above, then G (t), like G 1 (t), contains harmonics of symbol rate at the 6

7 2 1 1 baseband freq [MHz] time [µs] Figure 3: Simulated BPSK signal used in the data analysis example (see text). 7

8 Tap Value Time Index 1 2 Response [db] freq [MHz] Figure 4: Low-pass filter with M = 128 and f lp = MHz selected for the BPSK simulation example. Top: Filter coefficients, Bottom: Frequency response. 8

9 G 1 (ω) [db] Freq [MHz] Figure : G 1 (ω) for BPSK simulation example. Note harmonic content at multiples of the symbol rate. 9

10 G (ω) [db] Freq [MHz] Figure 6: G (ω) for BPSK simulation example. In this case, a residual carrier frequency of 1 MHz was instroducted into x f (t) for demonstration purposes. Note harmonic content at 2 MHz, associated with the residual carrier, and at 2 MHz ± MHz, associated with the symbol rate. shifted frequencies 2f c ± F Sy. This is demonstrated for our simulation example in Figure 6, in which we have left a residual carrier frequency of 1 MHz after the coarse tuning step above in order to demonstrate the result. If we were confident that all signals in the data were single-carrier linear modulations, then G (ω) would be a sufficient detector as it reveals both the frequency of the residual carrier as well as an indication of the symbol rate. However, we have found that G 1 (ω) sometimes reveals significant harmonic content, possibly associated with symbol rates or other cyclic features of the modulation, when G (ω) does not. We have also identified cases in this report in which G (ω) does not indicate any (single) carrier. Therefore, G (ω) is used as the primary method of characterization and G 1 (ω) can be used to either confirm the presence of spectral content in G (ω) or to identify additional spectral content. Thus, we choose here simply to show both results. Digital modulations are frequently pulse-shaped to improve spectral efficiency. Pulse shaping is essentially a filtering operation which redistributes power from the sidelobes to the main lobe of the spectrum. Because our method for detection of residual carrier and symbol rate depends the sidelobe structure (specifically, what is left after the sidelobes are filtered away), it is prudent to examine the performance 1

11 1 G 1 (ω) [db] G (ω) [db] Freq [MHz] Figure 7: G (ω) and G 1 (ω) for the pulse-shaped version of the BPSK simulation example. of the G (ω) and G 1 (ω) detectors when the signal modulation incorporates pulse shaping. Here, we simply applied the popular raised-cosine pulse shaping filter to the BPSK signal considered above. The filter uses the common roll-off parameter α =.22. G (ω) and G 1 (ω) for the pulse-shaped BPSK signal are shown in Figure 7; note that no significant change in performance is noted. Thus, we are confident that this method works approximately as well for pulse-shaped single-carrier digital modulations as it does for the non-pulse-shaped versions. In contrast, the G (ω) and G 1 (ω) metrics may fail if the signal is composed of multiple carriers; for example, a group of transponder channels which are closely-spaced in frequency, or signals using orthogonal frequency division multiplexing (OFDM). Since the number of carriers present in these cases is large, the G (ω) and G 1 (ω) detectors produce not simply the carrier harmonics but also all the associated intermodulation products, yielding a result which may be indistinguishable from noise. It is believed that multicarrier modulations such as OFDM can be effectively detected using more sophisticated methods including cyclostationarity and higher-order statistics, however in practice these more sophisticated methods experience great difficulty unless Γ is relatively large; i.e., greater than 1 db. 11

12 3 Results 3.1 File 29, Block 8 Figure 8 shows what is perhaps the strongest signal found in the dataset. We determined B s = 2 MHz, Γ = 12 db, and selected a filter with f lp = 8 MHz and M = 128. The results of the harmonic analysis are shown in Figure 9. There is no evidence of a symbol rate or other cyclic feature, and G (ω ) indicates that no residual carrier is present. We also attempted to detect harmonic content using filters with f lp = 6 MHz and 1 MHz, with the same results. Thus, this is probably not a single-carrier digital modulation. We considered whether this signal could possibly be a chirped signal, perhaps a harmonic or intermodulation product associated with a radar, for example. We are certain it is not chirped FM, because there is considerable variation in magnitude of the time-domain signal and no apparent structure in the phase of the time-domain signal. The latter also seems to rule out an amplitude modulated chirp. This signal may possibly be a multicarrier modulation, such as an OFDM signal. We invested this possibility using a cyclostationarity-based analysis, which provided only ambiguous results; perhaps due to insufficient signal-to-noise ratio or perhaps due to some defect in the signal such as a large time-varying frequency component or distortion due to the frequency response associated with the band edge. Given all the available data, our best guess is that this is a multicarrier modulation, possibly OFDM. 3.2 File 29, Block 11 Figure 1 shows a relatively weak persistent signal of moderate bandwidth. We determined B s = MHz, Γ = db, and selected a filter with f lp = 3 MHz and M = 128. The results of the harmonic analysis are shown in Figure 11. There is no evidence of a residual carrier. However, G 1 (ω ) reveals harmonic content having a frequency of 4.4 MHz, with no comparable feature present in G 1 (ω ). This result is repeatable using filters with f lp up to 6 MHz, with the harmonics becoming weaker with increasing bandwidth. No information can be gained by using filters narrower than f lp = 3 MHz, since the spectrum around the harmonic is suppressed in this case. We have no interpretation for the harmonic content, and can conclude only that this is not a single-carrier digital modulation. As in the previous case, we also observed no evidence that this is FM- or AM-modulated chirp. Thus, it seems most likely that this, too, is some form of multicarrier modulation, possibly OFDM. 3.3 File 27, Block 7 Figure 12 shows a strong narrowband signal. We determined B s < 2 khz (using an FFT of the entire 128K-sample dataset). The overall Γ > 1 db, however the signal fades monotonically over the observation time. Neither the spectral analysis or the G (ω) - G 1 (ω ) analysis indicate that the signal is a modulated in any way. Thus, 12

13 2 1 1 baseband freq [MHz] time [µs] Figure 8: File 29, block 8. 13

14 4 3 G 1 (ω) [db] G (ω) [db] Freq [MHz] Figure 9: G (ω) and G 1 (ω) for the strong wideband signal in Block 8 of File 29. it seems most likely that this signal is an unmodulated carrier or simply a spurious product from some other emitter. 3.4 File 218, Block 1 Figure 13 shows a block containing two signals. First, we consider the relatively weak, wideband signal in the upper sideband. We determined B s = 4 MHz, Γ = 2 db, and selected a filter with f lp = 2 MHz and M = 128. The results of the harmonic analysis are shown in Figure 14. There is no evidence of a residual carrier. However, G 1 (ω ) reveals harmonic content at many frequencies, with the strongest being at.28 MHz,.46 MHz,.2 MHz, 1.27 MHz, and 1.61 MHz. This corresponds to periods of 3.6 µs, 2.2 µs, 1.9 µs, 787 ns, and 621 ns, respectively. This result is repeatable using filters having a variety of bandwidths, and so is probably bona fide harmonic content in the signal. As in previous cases, we see no compelling evidence for a residual carrier, and thus speculate that this too is some form of multicarrier modulation, possibly either OFDM or simply multiple independent carriers. Next, we consider the relatively strong, narrowband in the lower sideband. Figure 1 shows a full-resolution view of the spectrum of this signal, which is now revealed to be a carrier with sidetones at a spacing of.6 MHz. We determine B s = 1.2 MHz and Γ = +3 db, and selected a filter with f lp = 1 MHz and M = 124. The results of the harmonic analysis are shown in Figure 16. G 1 (ω) reveals harmonic content at 14

15 2 1 1 baseband freq [MHz] time [µs] Figure 1: File 29, block 11. 1

16 4 3 G 1 (ω) [db] G (ω) [db] Freq [MHz] Figure 11: G (ω) and G 1 (ω) for the signal in Block 11 of File 29. a spacing of 1.2 MHz, which is twice the distance from the carrier to the sidetones. G (ω) reveals the carrier (albeit weakly) and harmonic content at a spacing equal to the distance from the carrier to the sidetones. Comparison of G 1 (ω) and G (ω) suggests a phase relationship between the sidetones. Figure 17 shows the instantaneous phase of this signal (after filtering), revealing that frequency is slowly increasing, perhaps linearly. This suggests a chirp signal; perhaps linear FM (LFM). 3. File 218, Block 12 Figure 18 shows a block containing a very wide but very weak signal. We determined B s = 11 MHz, Γ = 3 db, and selected a filter with f lp = 6 MHz and M = 128. The results of the harmonic analysis are shown in Figure 19. There is no evidence of a residual carrier. However, G 1 (ω ) reveals weak harmonic content at a frequencies of 1.1 MHz and 8.1 MHz. As in previous cases, we are left to speculate that this too is some form of multicarrier modulation, possibly either OFDM or simply multiple independent carriers. 4 Conclusions Six persistent signals from the CISR capture dataset were analyzed in an attempt to identify the modulation. The results are summarized in Figure 2. Overall, the 16

17 2 1 1 baseband freq [MHz] time [µs] Figure 12: File 27, block 7: A strong narrowband signal. 17

18 2 1 1 baseband freq [MHz] time [µs] Figure 13: Two narrowband signals. File 218, block 1. 18

19 4 3 G 1 (ω) [db] G (ω) [db] Freq [MHz] Figure 14: G (ω) and G 1 (ω) for weak wideband signal in upper sideband of file 218, block 1. 19

20 3 2 2 PSD [db/(763 Hz)] f [MHz] Figure 1: High-resolution spectrum of narrowband signal in lower sideband; file 218, block Hz resolution. 2

21 4 3 G 1 (ω) [db] G (ω) [db] Freq [MHz] Figure 16: G (ω) and G 1 (ω) for strong narrowband signal in lower sideband of file 218, block 1. 21

22 1 1 phase [deg] t [ms] Figure 17: Instantaneous phase of narrowband signal in lower sideband; file 218, block 1. 22

23 2 1 1 baseband freq [MHz] time [µs] Figure 18: A weak wideband signal. File 218, block

24 4 3 G 1 (ω) [db] G (ω) [db] Freq [MHz] Figure 19: G (ω) and G 1 (ω) for file 218, block 12. results were disappointing in that in no case were we able to unambiguously identify the modulation and associate it with a signal type known to be in use. However, we are quite confident that none of the signals analyzed were single-carrier linear digital modulations, as we have demonstrated that the techniques used here should be able to readily detect these. As mentioned earlier, more sophisticated techniques such as those based on cyclostationarity or higher-order moments may be effective in revealing relevant characteristics of the signal. We have attempted a cyclostationarity-based analysis of the strong wideband signal in block 8 of file 29, but with only ambiguous results. We suspect that any more advanced technique will be limited in the same way by File Block B s Γ Carrier? Harmonics Best Guess MHz +12 db No det. none Multicarrier? (OFDM?) MHz + db No det. 4.4 MHz Multicarrier? (OFDM?) khz +1 db Unmodulated MHz 2 db No det. multiple Multicarrier? (OFDM?) MHz +3 db Yes Linear FM MHz 3 db No det. multiple Multicarrier? (OFDM?) Figure 2: Summary of findings. 24

25 low signal-to-noise ratio and also by the relatively short duration ( 1.3 ms) of the observation. Better results may be possible with longer observation windows, since the signal-to-noise ratio limitation can be offset by integration time to some extent. Finally, we note the possibility that any of these signals could originate internally; e.g., from some system on the same plane from which the observations were made. 2

26 References [1] S.W. Ellingson, Preliminary Analysis of RFI Signals Captured using the CISR Coherent Sampling Mode, informal project report, March 4, 2. [2] J. T. Johnson et al., Airborne radio frequency interference studies at C-band using a digital receiver, Proc. IEEE Geo. and Remote Sensing Sym., Anchorage, AK, Sep