Dr. Haneefa,

If your sources are narrowband and if the noise is white, you may try to use either ESPRIT or MUSIC (both are parametric estimation techniques).

Best regards,

Since you can control the source, you may use a lock-in amplifier to extract small signals buried in a noisy background.

In the time-domain, you may use a boxcar averager & integrator.

Ricardo

Are these just carrier signals? Is there modulation present? Give us some numbers, please. Engineers & scientists need numbers! ;-) What is the minimum SNR you wish to detect? How much certainty do you want in detection? What is your tolerance for false alarms?

If the signals are all tones and there are a small number of frequencies that you are looking at, you may use narrowband filters. The maximum possible SNR gain is limited by the ratio of the bandwidth of one signal to the total bandwidth. Within a given narrowband filter, maximum SNR gain is approximated by the time-bandwidth product of the signal; search on "matched filtering" for this limit.

If the signals are all tones of many or unknown frequencies, you can use a digital filter bank implemented with one or more weighted FFTs. If the signals are tones with varying frequencies but the rates of variation are all the same, you can convert them to tones by using a complementary frequency variation in your demodulation chain. For multiple rates of variation, you will need multiple demodulation chains.

Receiver, processor, and detection theory is a field of endeavor, and can easily be more than a field of endeavor. Surveillance receivers are not the same as communications receivers, for example. More specific information requires more specific requirements, which is to say more information on the signals, the SNRs, what is to be done with the signals (QPSK demodulation, simple present or not present decisions with specified decision error probabilities, etc.).

why not try a maximum entropy burg or maximum likelihood estimator which should give a decent result so long as the sepaparate sources are uncorrelated.ive used these to good effect in the past. Music sounds good to But never used it.

Thank you dear all for your kind replies and support. I will try to see each of the methods you all mentioned. In the mean time, for more clarity I am just posting one source signal data plotted in matlab which are collected at two different sensor positions and its denoised versions using EMD (I don't know whether EMD helps here because the amplitude levels are very less). Just explaining a bit more about the problem I am trying to solve below.

Like this one source, there are two other source signals also present which are collected by the above two sensors. Hence a total of 7 combinations (source 1 present, source2 not, source 3 present etc) is possible in the final received signal, and at any position in a fixed time span.

My ultimate aim is to detect which all sources are present in the received signal, and then to identify the location of the sources in the time domain received signal. The three sources have three different frequency ranges, but two sources frequency ranges slightly overlaps although the dominant frequency components are different. I am planning to use a classifier like svm or so to classify. But how to identify the feature set, and how to proceed with the same is what I am really confused with. This is what I thought of doing initially, although I will change that now based on the inputs you all provided.

1. Denoise using emd.

2. FFT or Power spectrum of the sources (The amplitude levels are coming very less).

3. Use a BPF to separate out the sources in the received signal.

4. Design a matched filter for each of the sources and if required for each of the sensors also.

5. Perform matched filter detection to identify whether a source is present or not.

6. Perform cross correlation between a source signal and the received signal to identify the location.

7. Use a classifier based on the parameters obtained from the above steps to classify the received signal.

Any thoughts, inputs and helps is deeply appreciated. I am trying to do this in matlab.

Thanks

Niyas

I think the first step if you don't know how many sources you have in the mixture, is to estimate the number of impinging sources by using AIC or MDL algorithms. After that, you can use MUSIC or any of the proposed techniques.

Your problem is common to many people, and sonar/radar communities invest a lot to improve the tools! It comes very different:

- if the signal is stationary or a pulse (your graphics suggest a pulse),

- if you can access a copy of the original signal or not,

- if the propagation medium is dispersive or not (meaning it would alter the signal frequency content so to say shape) (if the travelling distance is short, this becomes a non-issue),

- if the medium is reverberant or not (the original signal will create many echoes interfering all together)

If you have access to the original signals, use them to filter the signal of your sensors and separate the various contributions! (methods based on convolution, adaptative filtering etc)

At a glance,

- your SNR is excellent on the 2 sensors, as long as you see the pulses in the time domain on a raw signal I would say the same!

- there is no reverberation

- your EMD is totally irrelevant as it does not extract the pulse presence as sharply as your eyes!

- I understand that your challenge is to separate several pulses with close-by frequencies converging almost simultaneously to the sensor(s). Because they are pulses the frequency content is not a dirac and frequency contents overlap; and you have to quantify their respective contribution - is that right?

The usual method of detecting pulses that are identical except for time of arrival and Doppler is a matched filter bank.

Bernard Garnier - Yes Sir, I have to identify first which all sources are present in the original signal, and then to identify their locations in time..

A typical matched filter bank implementation begins with a defined bandwidth for processing, which is downconverted and sampled to provide a complex data stream. If the bandwidth is too wide for one channel, multiple parallel channels are downconverted and processed in parallel. The complex data stream is processed by FFTs to provide the convolution that implements the matched filter stream. The outputs are threasholded to provide a specified probability of detection versus SNR along with a specified probability of false alarm versus SNR. Time extents processed in the FFTs are overlapped to avoid missing signals that come in at times that are not entirely within a particular FFT dwell time. Parameters such as sample rate, bit length of samples, number of samples per FFT dwell, detection threshold setting, etc. are traded off to provide a system design point that meets all requirements.

Other processing techniques such as QPSK demodulation, FSK signal detection and demodulation, etc. may or may not be processed in this way, depending on the signal type, range of Doppler shift, etc.

More focused techniques such as MUSIC are used for specific signal types such as long duration unmodulated tones at very low SNR and known frequency range but unknown exact frequency, and often have very long processing dwell times.