Hello All, I need to detect the bursts of the wide band noise in the sampled signal. The detector should ignore the narrow band pulses or the stationary signals. This is intended for blanking of the signal processing in the instrument. A solution could be based on the FFT, LPC or some sort of adaptive filter. However I am looking for less demanding method, which can be very approximate. Can you suggest the idea of such method or a book dealing with those methods. Vladimir Vassilevsky DSP and Mixed Signal Design Consultant http://www.abvolt.com
Detect wide band noise
Started by ●May 13, 2009
Reply by ●May 13, 20092009-05-13
On May 13, 10:46�am, Vladimir Vassilevsky <antispam_bo...@hotmail.com> wrote:> Hello All, > > I need to detect the bursts of the wide band noise in the sampled > signal. The detector should ignore the narrow band pulses or the > stationary signals. This is intended for blanking of the signal > processing in the instrument. > � A solution could be based on the FFT, LPC or some sort of adaptive > filter. However I am looking for less demanding method, which can be > very approximate. Can you suggest the idea of such method or a book > dealing with those methods. > > Vladimir Vassilevsky > DSP and Mixed Signal Design Consultanthttp://www.abvolt.comPerhaps a bank of low-order IIR filters followed by envelope detectors would be appropriate. Look at the bank of envelope detectors over time and look for times at which many of the detectors see a jump in their output magnitude. Pulses of energy out of a relatively small number of the filters would be rejected as narrowband. Spacing and bandwidth of the filters in the bank would be determined by the bandwidth and locations of the pulses that you would expect to see and do not want to blank out. Jason
Reply by ●May 13, 20092009-05-13
On May 13, 7:46 am, Vladimir Vassilevsky <antispam_bo...@hotmail.com> wrote:> Hello All, > > I need to detect the bursts of the wide band noise in the sampled > signal. The detector should ignore the narrow band pulses or the > stationary signals. This is intended for blanking of the signal > processing in the instrument. > A solution could be based on the FFT, LPC or some sort of adaptive > filter. However I am looking for less demanding method, which can be > very approximate. Can you suggest the idea of such method or a book > dealing with those methods. > > Vladimir Vassilevsky > DSP and Mixed Signal Design Consultanthttp://www.abvolt.comHere's one way: 1) Block data and FFT 2) Calculate power spectrum 3) Exponentially average each bin 4) Divide current power spectrum by long term average 5) Run a sliding median filter across the frequency bins 6) Sum the median filter output across the desired frequency band 7) Threshold to detect pulses Step 4 removes the stationary components Step 5 removes the narrowband pulses The sliding median filter filter can be performed with a dual heap algorithm to avoid sorting. An even faster approximation is to perform one or two passes of a 3-point sliding median filter where the three points are separated instead of adjacent. Blobs of spectral power narrower than the separation will be removed. The 3-point median only takes a few comparisons per point. Implement it at as low a level as possible. Common alternatives to the median are sliding window and sliding split window, but they are nowhere near as robust. FFT size, averaging period and filter width can be tuned to match your definitions of stationary and narrowband. The threshold is determined for probability of detection vs false alarm rate. Dale B. Dalrymple
Reply by ●May 13, 20092009-05-13
dbd wrote:> On May 13, 7:46 am, Vladimir Vassilevsky <antispam_bo...@hotmail.com> > wrote: > >>Hello All, >> >>I need to detect the bursts of the wide band noise in the sampled >>signal. The detector should ignore the narrow band pulses or the >>stationary signals. This is intended for blanking of the signal >>processing in the instrument. >> A solution could be based on the FFT, LPC or some sort of adaptive >>filter. However I am looking for less demanding method, which can be >>very approximate. Can you suggest the idea of such method or a book >>dealing with those methods. >>> > Here's one way: > > 1) Block data and FFT > 2) Calculate power spectrum > 3) Exponentially average each bin > 4) Divide current power spectrum by long term average > 5) Run a sliding median filter across the frequency bins > 6) Sum the median filter output across the desired frequency band > 7) Threshold to detect pulsesThank you for your reply, Dale. It can certainly be done in that way; however I am looking for a quick and dirty method. Such as compute the total energy and the zero crossing rate variance; or do it all at once using the frequency discriminator. Vladimir Vassilevsky DSP and Mixed Signal Design Consultant http://www.abvolt.com
Reply by ●May 13, 20092009-05-13
cincydsp@gmail.com wrote:> On May 13, 10:46 am, Vladimir Vassilevsky <antispam_bo...@hotmail.com> > wrote: > >>I need to detect the bursts of the wide band noise in the sampled >>signal. The detector should ignore the narrow band pulses or the >>stationary signals. This is intended for blanking of the signal >>processing in the instrument. >> A solution could be based on the FFT, LPC or some sort of adaptive >>filter. However I am looking for less demanding method, which can be >>very approximate. Can you suggest the idea of such method or a book >>dealing with those methods. >> > > Perhaps a bank of low-order IIR filters followed by envelope detectors > would be appropriate.That can be done even more efficient using Goertzel algorithm, however I would like to have something simple and crude, based on the envelope and phase distribution. I can split a signal into the basic high and low bands by a simple first order filter, but I can't afford much processing. Vladimir Vassilevsky DSP and Mixed Signal Design Consultant http://www.abvolt.com
Reply by ●May 13, 20092009-05-13
On May 13, 3:43 pm, Vladimir Vassilevsky <antispam_bo...@hotmail.com> wrote:> dbd wrote: > > On May 13, 7:46 am, Vladimir Vassilevsky <antispam_bo...@hotmail.com> > > wrote: > > >>Hello All, > > >>I need to detect the bursts of the wide band noise in the sampled > >>signal. The detector should ignore the narrow band pulses or the > >>stationary signals. This is intended for blanking of the signal > >>processing in the instrument. > >> A solution could be based on the FFT, LPC or some sort of adaptive > >>filter. However I am looking for less demanding method, which can be > >>very approximate. Can you suggest the idea of such method or a book > >>dealing with those methods. > > > Here's one way: > > > 1) Block data and FFT > > 2) Calculate power spectrum > > 3) Exponentially average each bin > > 4) Divide current power spectrum by long term average > > 5) Run a sliding median filter across the frequency bins > > 6) Sum the median filter output across the desired frequency band > > 7) Threshold to detect pulses > > Thank you for your reply, Dale. It can certainly be done in that way; > however I am looking for a quick and dirty method. Such as compute the > total energy and the zero crossing rate variance; or do it all at once > using the frequency discriminator. > > Vladimir Vassilevsky > DSP and Mixed Signal Design Consultanthttp://www.abvolt.comThe method I outlined is particularly convenient only if the processing to be blanked makes use of the Fourier coefficients so they are available for free. If you do not need any of the information typically generated from the Fourier coefficients, one way to speed up the processing chain I outlined is to subsample the signal bandwidth if the aliasing will lose no useful information. The greatest efficiency that could be achieved in general depends, as we are accustomed to telling beginning questioners, on the details of the signals and the other processing that you haven't given us. Dale B. Dalrymple
Reply by ●May 14, 20092009-05-14
> I am looking for a quick and dirty method. Auto Correlation would concentrate noise in the samples near zero. There are variants to speed it up by eliminating the multiplications( Otnes, Enochson "Digital Time Series Analysis" Chapter 6 ). Assuming the original data is 8 bit they reduce the samples in one or both frames to: 1 bit by 1 bit "binary" ( usually not working ) 1 bit by 8 bit "clipped autocorrelation" 4 bit by 4 bit multiplication by table "Clipped autocorrelation" has been used in the speech recognition phone of Origin Technology Inc. in 1989 ( US Patent 5,007,081 Speech activated Telephone ). Hardware: 6502 controller ( 2 MHz clock, 32kbyte SRAM, 32kbyte EPROM ). 8 bit A/D, samplerate 7,2kHz, frame 144 bytes ( ca. 20msec). MfG JRD
Reply by ●May 14, 20092009-05-14
On May 13, 9:46�am, Vladimir Vassilevsky <antispam_bo...@hotmail.com> wrote:> Hello All, > > I need to detect the bursts of the wide band noise in the sampled > signal. The detector should ignore the narrow band pulses or the > stationary signals. This is intended for blanking of the signal > processing in the instrument. > � A solution could be based on the FFT, LPC or some sort of adaptive > filter. However I am looking for less demanding method, which can be > very approximate. Can you suggest the idea of such method or a book > dealing with those methods. > > Vladimir Vassilevsky > DSP and Mixed Signal Design Consultanthttp://www.abvolt.comJust a thought, Take the absolute value of the sum of the last two samples and the absolute value of the difference of the last two samples and compare this to a threshold. Ie, if ( abs(x[n] + x[n-1]) + abs(x[n]-x[n-1]) ) > THRESHOLD then [trigger] -Martin
