The emission of radio waves from Extensive Air Showers (EAS), initiated by ultrahigh-energy cosmic rays, has been attributed to geomagnetic emission and charge excess processes. At frequencies from 10 to 100 MHz this process leads to coherent radiation. Nowadays, the radio detection technique is used in many experiments consisting in studying EAS. One of them is the Auger Engineering Radio Array (AERA), located at the Pierre Auger Observatory. The frequency band observed by the AERA radio stations is 30-80 MHz. This investigated frequency range is often highly contaminated by human-made and narrow-band radio frequency interferences (RFI). The suppression of this contamination is crucial to lower the rate of spurious triggers.

The paper presents the filter currently in use and proposed:

A. filters currently in use

1) an adaptive filter based on the FFT and median filter in the frequency domain. The filter converts ADC data from the time domain to the frequency domain using the Fast Fourier Transform (FFT) routine provided by Altera. These peaks are removed in the frequency domain by a median filter. With the second, inverse FFT (iFFT), signals are converted back to the time domain. This filter was used in the first radio stations in Cyclone III EP3C80F780C6. It required an implementation of FFT and iFFT power-consuming Mega-Functions and tricks related to aliasing removal. The sampling frequency in ADCs was limited to 180 MHz, which corresponded to only 90 MHz spectrum cut-off according to the Nyquist theorem

The adaptive FFT-Median-iFFT filter was replaced by the non-adaptive IIR notch filter in the next generation of electronics. Adaptive filters allowing an RFI suppression independently of the frequencies of the contamination, are more flexible than non-adaptive filters with fixed parameters usually tuned for the local transmitter frequencies. Such a solution is efficient: however, it cannot react to any additional RFI sources.

2a) IIR non-adaptive filter with fixed coefficients, which suppress four narrow bands. IIR filters are much simpler and power efficient than FIR filters. However, they are generally potentially unstable due to feedbacks. Fixed frequencies, in the AERA IIR filter currently in use, have been selected by investigations of the EM spectrum on the Argentinean pampas in the AERA research region. Coefficients calculated in an external C program have been implemented into the FPGA code as constants. If the frequencies of the contaminations are stable (e.g., correspond to short-wave transmitter carriers) a non-adaptive filter is extremely efficient. However, if the frequencies of contamination do not match the IIR coefficients, the IIR filter is practically useless: the output signals are approximately the same as the input ones.

B. filters already tested in Argentinean pampas

3a) with FIR coefficients calculated in an external ARM processor. This cheap solution uses the existing micro-controller used for the radio station management. However, data exchange between the FPGA and the micro-controller significantly slow downs the operational speed and increases the refresh time of the filter coefficients. The refresh time (~2 s for this filter) is a crucial parameter for quickly fluctuating environmental radio contaminations. The filter coefficients should be updated as fast as possible to follow the contamination changes. During calculations of a new set of coefficients, the filter is “blind” to any occurring changes. This may affect the signal that is the object of detection.

3b) FIR filter supported by the NIOS processor reduces the refresh time to ~600 ms for 32 FIR stages. The NIOS processor needs to solve 32 linear equations in the FPGA. The representation of floating point variables as double provides the same precision as Microsoft Visual C++. This filter was tested in real Argentinean pampas conditions and proved its efficiency and stability.

C. proposed filters

2b) A non-adaptive IIR notch filter has been additionally equipped with the embedded NIOS processor inside the FPGA, calculating the filter coefficients based on the current environmental conditions, providing an adaptive version of the filter. The IIR-notch filter supported by the NIOS internal processor allows refreshing the IIR coefficients every few seconds. As the NIOS processor is relatively power efficient, this filter could adapt itself to changing environmental parameters.

The currently used IIR-filters are a cascade of four sub-filters each tuned to a fixed frequency. If the new fifth ”carrier” appears, only four sub-filters would suppress the RFI (assumed the strongest). Simulations show that a more efficient approach is the suppression of all potential frequency peaks, even if the suppression factor is less than for a fixed IIR-notch solution.

This type of suppression was processed in the FIR filters 3a) and 3b) already tested in the field.

3c) In order to reduce the refresh time the Hardcore Processing System (HPS) inside the FPGA was tested on the development kit with Cyclone V SX SoC5CSXFC6D6F31C6N. A hardware implementation of the micro-controller allows a reduction of the refresh time to only ~20 ms.

3d) A linear prediction method based on the autocovariance matrix calculated from ADC samples. The matrix is of Toeplitz type. Linear equations with a Toeplitz matrix can be solved by much the faster Levinson procedure instead of relatively slow Gaussian elimination. The algorithm of solving 32 linear equations based on the Levinson procedure has been implemented directly in the FPGA fabric. 32 linear equations are solved using 64-bit floating-point Altera FPGA Mega-functions. The refresh time has been reduced to only ~600 microseconds. Splitting the processes for parallel processing improves the calculation efficiency and reduces the refresh time. However, it increases the power consumption, a crucial factor for applications supplied from solar panels. This variant will be tested in the field, especially in seriously contaminated regions. Calculating a set of new coefficients always needs a relatively long interval. A much smarter approach is to correct the existing coefficients in each clock cycle by a procedure of the sigma-delta type. One of the filters has been already tested.

4) An adaptive filter based on the normalized LMS (NLMS) algorithm with the canonical FIR section