Impulsive_RFI_Excision:_CASPER_Library_Block - david-macmahon/wiki_convert_test GitHub Wiki

Development Team

This project has been designed as a part of the GMRT upgrade activities by Kaushal D. Buch of the Digital Backend group, Giant Metrewave Radio Telescope, NCRA-TIFR, India.

Testing Support: Shruti Bhatporia, Short Term Engineering Intern, Digital Backend Group, GMRT.

For further details or clarifications regarding this project email on : [email protected], [email protected]

Project Introduction

This RFI mitigation block uses the excision technique for removal of impulsive RFI. Primarily, the block has been designed to mitigate impulsive time domain (broadband) RFI. This type of RFI is pre-dominantly observed due to high voltage power-line and automobile sparking, corona discharges and switching of inductive load. The block operates on Nyquist sampled digital time-series and can be integrated with the CASPER tool-flow along with the other signal processing blocks on the FPGA. The design contains two distinct blocks, one for RFI detection and the other for RFI filtering.

Computing Robust Standard Deviation

RFI manifest as outliers in the statistical distribution of astronomical signal. A conventional estimator of statistical dispersion (standard deviation) gets biased in the presence of outliers and provides an inaccurate estimate in the presence of RFI. Here, Median Absolute Deviation (MAD) has been used for computing standard deviation of the distribution. This estimator can handle up to 50% outliers (theoretically) in a given data set.

MAD = median(|x - median(x)|)

For normal distribution, MAD can be scaled to robust standard deviation using σ = 1.4826 * MAD

RFI detection and filtering threshold can be calculated from robust standard deviation : Threshold = median ± n*σ

Here ‘n’ is ‘aggressiveness’ parameter which is user programmable.

Any value occurring outside this threshold would be considered as RFI.

RFI Detection and Filtering using MAD based Threshold

Once the RFI is detected, the sample can be filtered by replacing it with either the median or a constant value or a sample drawn from a Gaussian distribution (noise). Alternatively the samples can simply be clipped at their detection threshold. The option can be selected by the user.

In case of replacement with noise, a sample with standard deviation similar to that of the distribution has to be generated. This is done with the help of digital noise source developed at the GMRT. More details of the digital noise source are available on: https://casper.berkeley.edu/wiki/Variable_Correlation_Digital_Noise_Source_for_FPGA

FPGA Implementation

Since MAD requires recursive median computation, it is difficult to implement the approach in real-time. It was found that the histogram based approach for median calculation is best suitable for real-time FPGA implementation and is used in this block.

Currently, the design has fixed parameters and can work with 8-bit signed data and a window size of 1024 samples up to a maximum clock of 250 MHz. Sliding window MAD is computed for contiguous 1024 sample windows. The resource utilization for design in its current form is ~20 % slices and 3% Block RAM.

A separate block computes the threshold and filters the RFI based on the option selected by the user.

Project Files

Design file: Version 1: June 2014:

Note: This version is undergoing rigorous testing and improvements/changes would be reflected in the subsequent versions.

Test Results

The design was integrated with the GMRT wideband backend (GWB) and was tested with antenna signals. The results presented here provide a glimpse of the broadband RFI removal capability of the design.

The tests described here were carried out with the following setup:

RF: 600 MHz RF BW: 100 MHz Processing BW: 200 MHz Antenna: GMRT Central Square Antenna - C04 No. of Spectral Channels: 2048 GWB mode: Incoherent Array Integration time: 1.3ms The plots shown below are Total Intensity over time for one spectral channel corresponding to RF of ~651 MHz. The first subplot shows the filtered time series at 3 sigma threshold (filtered values replaced by zero), the second subplot is the original time series (without filtering) and the third subplot shows overlap plot of the filtered and unfiltered time series.

Queries, Suggestions and Feedback

Kaushal Buch, Digital Backend Group, GMRT, NCRA-TIFR, India ([email protected], [email protected])