RFI MITIGATION AND BURST DETECTION WITH A RECONFIGURABLE DIGITAL RECEIVER

Transcription

1 RFI MITIGATION AND BURST DETECTION WITH A RECONFIGURABLE DIGITAL RECEIVER Cedric Dumez-Viou1,4, Andrée Coffre4, Pierre Colom2, Laurent Denis4, Alain Lecacheux2, Jean-Michel Martin3, Philippe Ravier1, Rodolphe Weber1, Philippe Zarka2 Laboratory of Electronic, Signals, Images, Polytech'Orleans - University of Orleans 12 Rue de Blois BP 6744 Orleans Cedex 2 F France 2 LESIA, Observatoire de Paris/CNRS, 5 Place Jules Janssen, F Meudon, France 3 GEPI, Observatoire de Paris/CNRS, 5 Place Jules Janssen, F Meudon, France 4 Station de Radioastronomie, F Nançay, France 1 Phone: +33 (0) ; cedric.dumez-viou@obs-nancay.fr web: ABSTRACT In radio astronomy, more and more observations are polluted by man-made radio frequency interferences (RFI). The impact of these RFI on spectral measurement ranges from total saturation to tiny distortions of the data. To some extent, the final spectral estimation can be preserved by blanking infected channels in real time. With this aim in view, a complete real time processing line has been implemented on a set of digital signal processing reconfigurable components. The current functionalities of the system are, multi acquisition capabilities, high dynamic range (at least 70 db), band selection facilities (from 875 khz to 14 MHz), high spectral resolution through polyphase filter bank (up to 8192 channels with coefficients) and real time time-frequency blanking with robust threshold detectors. Recent results of real time RFI detection algorithms implemented in the receiver and applied on actual observations are shown. 1 INTRODUCTION Radio astronomy, in common with many others user of the radio spectrum, has the advantage of a few protected frequency bands. However, most scientific questions find their answer in unprotected bands where radio astronomy is not a primary user. Moreover, even in the protected bands, out-of-band emission regulations are not always sufficient to prevent the

2 pollution of astronomical primary bands. As a result, an increasing number of observations become unusable. Indeed, classical receivers were not designed to operate with in such hostile conditions. First, their poor dynamic range induces non-linearity, which spreads the RFI over the whole spectrum. Secondly, the analogue filters used in such systems do not provide enough frequency rejection. Thirdly, their spectral resolution and channel rejection are often too limited to extract the free channels from the corrupted ones. Finally, their hardware architecture is too specific to allow additional functions, such as RFI detection, to be implemented. With this aim in view, new generations of digital receivers have been recently designed [Baan, 2002; Kramer, 2001; Rosolen, 1999]. They differ from each other by their specifications (input bandwidth, number of bits, number of channels ). In this paper, the design of a robust radio astronomy receiver (R3) is presented. It has been specifically designed for the single dish telescopes of the Nançay observatory (France). First, the overall architecture is given. Then, the real time spectral analysis scheme is described. Afterwards, real-time hardware implementations of RFI mitigation algorithms are presented as well as simulations of future hardware implementations. On line RFI mitigation on real data shows that the Signal-Of-Interest (SOI) may be advantageously retrieved. More advanced methods aimed at the detection of brief emissions are briefly discussed at the end. 2 SYSTEM ARCHITECTURE Figure 1 describes the global architecture of the R3. It can be seen as eight parallel receiver lines. An analogue section shifts the signals from the antennas input frequencies to an intermediate frequency (IF) of 70 MHz, providing a final useful bandwidth of 14 MHz. A switch matrix configures the simultaneous connection of three Nançay instruments (Nançay Radio Telescope (NRT), Nançay Decameter Array (NDA) and Nançay Surveillance Antenna (NSA)) to any of the eight receiver line. This functionality is one of the architecture

3 key points because it boots the observational capabilities of the receiver. In particular, 4 bands can be merged so that the bandwidth and the spectral resolution are improved by a factor of 4. Another interesting possibility is to send the same signal on several lines, each running a different process. Thus, the same signal can be analyzed from different points of view. Moreover, interactions between the processes can be implemented. The other key point is the digital section. It corresponds to a succession of digital modules plugged on PCI boards (HEPC9 and HERON modules from Hunt Engineering). Each of the 8 digital processing banks includes a 14 bit ADC, and four reconfigurable digital components for real time signal processing.. A powerful industrial PC is used to drive 2 digital processing banks. The four necessary PCs are connected to a central computer using a private Gigabyte Ethernet that provides configuration information and data output. The central computer embeds a database that schedules the configuration of the system, observations and exportation of data towards public network. The R3 basic configuration includes high resolution spectral analysis and post detection RFI mitigation techniques (see next sections). Besides, all digital processing lines can be reconfigured and merged together to perform any other digital processing, such as specific radio astronomical observations or more complex RFI mitigation techniques. 3 TIME FREQUENCY ANALYSIS CAPABILITIES From the14 MHz bandwidth of the IF, a frequency band (between 14 MHz and 875 khz) is digitally down converted to base band for real time power spectral density estimation (PSD). This spectral analysis has two functions. The first one is to provide spectral information on the SOI to radio astronomers. The second one is to make an ad hoc segmentation of the time frequency plane with a view to performing the best RFI blanking. In practice, given the large flow of data to be processed, classical radio telescope receivers

4 use coarsely quantized correlators to perform this spectral analysis. In our RFI context, this method is not well suited and frequency domain approach has been preferred. Depending on the RFI properties, the time-frequency resolution must be reconfigured (see Figure 2). Thus, two methods have been designed. For a better time resolution, weighted FFT with 50% overlap can be downloaded in the receiver. The number of FFT bins ranges from 256 to When no accumulation is performed, the best time resolution for a 2048 bin FFT (respectively, 256 bin FFT) is 73 µs (respectively, 9 µs). Exact spectral resolution and side-lobe rejection depend on the window applied on the data. For better spectral resolution, an 8192 bin polyphase filter bank [Vaidyanathan, 1993] can be used. The low-pass filter prototype needs coefficients. In Figure 3, the performances in terms of channel rejection and spectral resolution are shown. With the polyphase filter bank, the maximum frequency resolution is 107 Hz for an 875 khz bandwidth. The rejection is better than 70 db. The resulting spectra can be stored on a hard disk or used for further real time embedded processing such as RFI detection (see next section). In the case of disk storage, the dataflow must be reduced due to some output board link limitations. This can be achieved by a preaccumulation of at least 8 instantaneous spectra. 4 REAL TIME RFI DETECTION ALGORITHMS Various methods have been experimented to eliminate those RFI depending on the type of interferences and the type of instruments [Fridman et al., 2001; Bretteil and Weber, 2004]. The present study focuses on time-frequency blanking on data coming from a single dish. From the power time-frequency (T-F) plane generated in the previous processing step, we want to separate all the T-F points corrupted by a RFI (case named H1 hypothesis) from those which are not (case named H0 hypothesis).

5 The simple idea, which has been implemented, is to use a power criterion to perform this discrimination. However, the difficulty is to estimate the ideal power threshold from the power T-F plane, the objectives being the lower false alarm probability and the best detection probability. Under H0 hypothesis, since the SOI is assumed to be a stationary Gaussian noise, each T-F point shows a chi-square distribution with 2 x i degrees of freedom (i being the number of accumulation of a pair of squared Gaussian distributions, i.e. real and imaginary parts of an instantaneous spectrum). Given a probability of false alarm, the calculation of the corresponding threshold, S, is straightforward: S = m + C.s (1) where C is a constant depending on the required discrimination rate, m is the mean of the distribution under H0 hypothesis and s is the absolute distance of the distribution under H0 hypothesis (absolute distance has been preferred to standard deviation because its implementation gives better speed performances). m= s = 1 N N åx (f) t (2) f =1 1.3 N å X ( f ) - m (3) N f =1 t where X t ( f ) is the power density at frequency f and time t on the T-F plane and N is the number of frequency bins used to estimate the statistics. Unfortunately, the T-F points under H0 hypothesis are not known since this is the aim of the algorithm to detect them. To overcome this vicious circle, robust estimation methods of m

6 and s must be implemented (i.e. robust estimators must give the same value for a corrupted signal or a clean one). Besides, those methods have to accommodate the real time constraint of the system. In the next subsections, two examples are given. They are based on some RFI a priori information. The first one makes use of the narrow band properties of some RFI and the second one exploits their impulse like properties. 4.1 NARROW BAND CASE Principle If the RFI is narrow band, only a few frequency bins of the T-F plane are polluted at a given time. Besides, among these bins, only those with high power values may alter the estimation of m and s under H0 hypothesis. Thus, by discarding these extreme values, a robust estimation of m and s can be computed. Two methods have implemented and tested. The first one is based on median filtering and the second on iterative estimation. Median filtering sorts the set of data and returns the value of the sample located halfway in the ranking. Our implementation is based on the Quicksort algorithm [Numerical Recipes] that optimizes sorting for large datasets. Iterative estimation is based on an iterative use of Equ. 2 and Equ. 3. Progressively, extreme points of the original dataset are removed. The algorithm stops when the estimated values are stabilized. It appears that iterative estimation is less accurate than the median filtering but it runs faster, in particular with large dataset. Tests have shown a radio of 4 in favor of iterative estimation while processing 2048 bins spectra. Due to our real time constraint, iterative estimation has been finally implemented for the tests on the OH megamaser IIIZw Application The hydroxyl (OH) ground-state transitions have rest frequencies of 1612, 1665, 1667 and 1720 MHz. Strong extragalactic OH maser emissions have been detected in the central

7 regions of many active and/or starburst galaxies, which are called OH megamaser galaxies. The 1667 MHz main line emission is almost always the strongest line observed in these extragalactic sources. Protected bands have been allocated mainly for the observation of Galactic objects. Since most of the extragalactic emissions are redshifted, their observed spectral lines and continuum emissions may fall into telecommunication-allocated bands. This is the case, for example, of the OH megamaser III Zw 35 whose 1667 MHz line is corrupted by the Iridium satellite phone system. This cosmic source corresponds to a flux of Jansky and it is unobservable with traditional receivers (see Fig. 4.a). Iridium constellation is composed of 66 satellites orbiting on 6 orbital planes at a period of 100 minutes. It communicates with mobile phones using T-FDMA modulation (TimeFrequency Division Multiple Access) in the 1616 MHz MHz band. The digital signal is modulated with QPSK (Quadrature Phase Shift Keying). Figure 5 shows the dynamic spectrum of Iridium RFIs. The Iridium slots are very located in frequency and in time. In our experiment, the mean and the absolute distance are extracted in real time as described previously. The detection threshold level (Equ.1) is computed with C=8. The number of bins per spectrum is A block (Dt, Df ) of X t ( f ) is blanked as soon as one of the corresponding X t ( f ) exceeds the threshold. Two kinds of block pattern have been applied: Full spectrum blanking (Dt = 1, Df = 2048) : the complete spectrum (i.e. all X t ( f ) for given t) is blanked as soon as one or more X t ( f ) exceed the threshold. The loss in data is about 20%. This kind of blanking can be efficient to guarantee very clean observations by removing every possible ripples present in a spectrum. Fine block blanking (Dt = 1, Df = 3) : This method is more time consuming, but the blanking is more accurate. In our example, the loss in data was only 2.5%. Fine block

8 blanking is also well suited for continuous RFIs. In both cases, a few minutes of averaging is sufficient to see 1667 MHz OH main line (see Fig. 4.b). In Figure 5, the data bins which were detected as corrupted are shown in shaded tones. All RFI emissions have been well detected. 4.2 IMPULSE-LIKE CASE Principle Radars RFIs are characterized by short pulses (1µs) that repeat every millisecond (exact values are radar system dependent). Short pulses in time domain results in wide band spectral lobes (see Figure 6) that theoretically follow a square sine cardinal function shape. Neither iterative estimation nor median filtering performs well with radar RFIs. The number of corrupted channels per spectrum is too high to keep those estimators robust. However, when the time resolution of the T-F plane is fine enough (for example with a 256 bin FFT), radar pulses cannot corrupt more than 2 successive spectra. Thus, the idea is to compare the means for 3 or more spectra and to choose the one with the smallest value for the blanking threshold computation. Due to the wide band shape of the radar spectrum, full spectrum blanking (in the example (Dt = 1, Df = 256) ) is the most appropriate Application The NRT is a powerful utility in particular for large surveys of hundreds or thousands sources. In particular, in has been shown during the last decade that the NRT is well adapted to the observation of the neutral hydrogen (HI) component in galaxies, producing the largest set of measurements after the Arecibo telescope. The MHz HI line is also observed redshifted, and more often now than in the past, in the bands below1370 MHz which is the upper limit of the Airport Surveillance Radar band. Efficient RFI mitigation systems are now

9 needed in order to perform redshifted galaxy observations below this limit. Blanking simulations based on real signals acquired for the NRT have shown an effective rejection of radar pulses (see Fig. 7). We are presently testing this algorithm on radio sources located in radar frequency bands in order to precisely evaluate the level of rejection. 5 AUTOMATIC EVENT RECORDER A specific setup of R3 is currently being developed to detect and record any brief event occurring in the analyzed frequency band at a very high time resolution. Obviously, the detection is designed to be robust against RFIs. The objective is to limit disk storage which is a tremendous issue for wide band acquisition system (the current R3 output data rate is 14 Mo/s, including a pre-accumulation by a factor 8). Two R3 boards are used: one board for high resolution recording that will be triggered by a second one dedicated to pattern matching. A FIFO will allow the system to record a few seconds of data before the event occurs. The first intended use of this system is to record radio signatures of Jupiter magnetospheric radio bursts. Some of those radio emissions are predictable in time with a good probability but a good many are not. By this automatic recorder, both high and low probability events can be recorded without any disk storage and time analysis issue (see Fig. 8). Pattern matching for this application is based on the detection of frequency drifts that is one of the characteristic of Jovian emissions. Miniature images consisting of 256 time lines spectra of 256 bins are fed into a 2D-FFT. Oblique structures of frequency drifts create an oblique ellipse in the center of the resulting images (see Fig. 9). The relative contrast of the ellipse over the background is a satisfactory estimator of the presence of Jovian emissions. 6 CONCLUSION In this paper, the digital implementation of a new generation of radio astronomical receivers has been presented. Our system is robust towards RFI by providing improved linearity,

10 higher frequency rejection and better spectral resolution compared to current receiver designs. Thus, the signal integrity can be preserved and real time RFI mitigation techniques can be envisaged. Our system is fully reconfigurable and can be adapted to any RFI context. Besides, real-time implementations of RFI mitigation algorithms have been presented. It seems that some bands can be observed again. Nevertheless, only intermittent parasites have been discussed and much more work has to be done with broad-band and continuous time RFIs. ACKNOWLEDGMENTS The authors gratefully acknowledge people who have made this work possible: L. Amiaud, D.Aubry, L. Bacquart, E. Gérard, E.Thetas.

11 REFERENCE Baan, W.A, P.A. Fridman and R.P. Millenaar (2002), RFI mitigation at WSRT: algorithms, test observations, system implementation, JFC session, paper presented at URSI XXVII General Assembly, Maastricht, The Netherlands, August. Bretteil, S and R. Weber (2004), Comparison between two cyclostationary detectors for RFI mitigation in Radio Astronomy Workshop RFI 2004, Penticton, BC, Canada, July. Fridman, P. A. and W. A. Baan, (2001), RFI Mitigation Methods in Radio Astronomy, Astronomy & Astrophysics, vol. 378, pp Kramer, M., A.G. Lyne, B.C. Joshi and al. (2001), COBRA, a digital receiver at Jodrell Bank, paper presented at RFI mitigation workshop, Bonn, Germany, March. Numerical Recipes, Chap 8, p 342. Rosolen, C., V.Clerc and A. Lecacheux (1999), High dynamic range, interference tolerant, digital receivers for Radio Astronomy, The Radio Science Bulletin, N 291, pp Vaidyanathan, P.P.(1993), Multirate systems and filter banks, Signal Processing Series, Prentice Hall.

12 Figure 1: Basic receiver structure. From the 14 MHz bandwidth of the IF, a frequency band (between 14 MHz and 875 khz) is digitally down converted to base band. The band selection is digitally achieved in two steps. First, an undersampling is applied with a 56 MHz sampling frequency. Then, a digital mixer followed by successive decimation filters selects the band of interest. This process fits into a Xilinx FPGA, VIRTEX II The spectral analysis is performed by another FPGA. Finally, RFI mitigation algorithms are implemented in the two Texas Instrument DSP.

13 Figure 2: Power spectral density measured with the NDA. (up) 12 khz spectral resolution: spectra are unusable (down) 760 Hz spectral resolution: RFI channels are well separated. Thus, RFI mitigation techniques may be applied.

14 Figure 3: Comparison of the spectral resolution and channel rejection between a BlackmanHarris windowing and a polyphase filter bank with coefficients.

15 Figure 4: Spectra of III Zw 35 after a 14 minutes time integration, using (ON-OFF)/OFF display. (a) without blanking. The Y-axis is scaled so that the SOI expected profile (in continuous dash line) can be seen. Some RFI bursts are 26 db stronger than the SOI level. (b) with real time blanking. The III Zw 35 source is clearly visible. The last detection of III Zw 35 was performed in real-time January 8th 2004

16 Figure 5: Dynamic spectrum of Iridium RFIs

17 Figure 6: Dynamic spectrum of radar RFIs

18 Figure 7: (ON-OFF)/OFF display of 1 minute integrations with radar RFIs (a) with and (b) without full spectrum blanking.

19 Figure 8: Values of the Jovian event detection criterion during a long time observation (210 mn). Predictable events are located between 1h30 and 2h10. Some low probability events have been also detected after 3h00 and before 1h30.

20 Figure 9: Dynamic spectrum (a) with and (b) without Jovian emission and the corresponding 2D-FFT transform (c) and (d).

21 Figure 1: Basic receiver structure. From the 14 MHz bandwidth of the IF, a frequency band (between 14 MHz and 875 khz) is digitally down converted to base band. The band selection is digitally achieved in two steps. First, an undersampling is applied with a 56 MHz sampling frequency. Then, a digital mixer followed by successive decimation filters selects the band of interest. This process fits into a Xilinx FPGA, VIRTEX II The spectral analysis is performed by another FPGA. Finally, RFI mitigation algorithms are implemented in the two Texas Instrument DSP. Figure 2: Power spectral density measured with the NDA. (up) 12 khz spectral resolution: spectra are unusable (down) 760 Hz spectral resolution: RFI channels are well separated. Thus, RFI mitigation techniques may be applied. Figure 3: Comparison of the spectral resolution and channel rejection between a BlackmanHarris windowing and a polyphase filter bank with coefficients. Figure 4: Spectra of III Zw 35 after a 14 minutes time integration, using (ON-OFF)/OFF display. (a) without blanking. The Y-axis is scaled so that the SOI expected profile (in continuous dash line) can be seen. Some RFI bursts are 26 db stronger than the SOI level. (b) with real time blanking. The III Zw 35 source is clearly visible. The last detection of III Zw 35 was performed in real-time January 8th 2004 Figure 5: Dynamic spectrum of Iridium RFIs Figure 6: Dynamic spectrum of radar RFIs Figure 7: (ON-OFF)/OFF display of 1 minute integrations with radar RFIs (a) with and (b) without full spectrum blanking.

22 Figure 8: Values of the Jovian event detection criterion during a long time observation (210 mn). Predictable events are located between 1h30 and 2h10. Some low probability events have been also detected after 3h00 and before 1h30. Figure 9: Dynamic spectrum (a) with and (b) without Jovian emission and the corresponding 2D-FFT transform (c) and (d).