Biswajit,

I'd be inclined to think about boosting the signal-to-noise ratio by hardware means - what is the sensor and what pressure differences are you trying to sense?

At the outset, I would like to thank you for your attention, James. Boosting the SNR through hardware is not an option as that would require me to do some hardware alterations and I am constrained in that regards for the time being. The sensor is a capacitance type differential pressure transmitter and I have realized (from readings at different points) that the noise gets added at the DAQ ( the DAQ is a sloppy one) and that made me wonder if it would be wise to use a median filter at the output of DAQ to smoothen the output a bit. That, in effect, might reduce some of the chattering in the controller effort. 

Biswajit,

Well, choosing a low noise ADC, shielding the heck out of it, and having a well-understood ground/shield pattern sounds to be the best bet.

If the pressure variation is cyclical, one can do frequency-domain filtering, similarly if the noise has well-understood spectral properties, then an FFT and some filtering may help.

But I'd be very tempted to add an OP05 instrumentation grade op-amp right at the pins of the sensor, drive that with a rock-solid DC line, and enjoy a bigger signal.

Why isn't hardware modification an option?

Try Wavelet Transform, best for noise removal in acquired signals!

James,

The hardware that I am currently working upon in my institute laboratory is warranted against manufacturing defects for a certain period and until then I would not be permitted to alter it. That is why I am seeking software options for the time being.

The noise appears to be in the form of white noise and I would like to know more about the frequency-domain filtering that you have mentioned. How would that fare when compared to smoothing techniques such as moving average or median filtering?

Biswajit

Amit.

Thanks for the idea. I will have a look at it. 

Biwajit,

:(

Sorry to hear about the warranty. Well, there may be some steps to take to cut down EMI: shielding principally, and then ensuring that the cable run to the ADC card uses a twisted pair configuration, with single-point earthing to prevent ground loops.

As for filtering, it may look white, but I bet that there's a spectral shape to it. Noise comes from various sources, and there aren't many low-freq. sources below 50Hz/60Hz (mains EMI) usually.

If the pressure disturbances that you're looking are lower frequency, then a low pass filter would be a starting point (which a moving average approximates).

I would use a Fourier transform of the data and start with a simple low-pass edge filter.

It's more complex to implement than a boxcar, but modern data environments mean that you probably won't have to write a scrap of code.

James,

I have checked the shielding and the twisted pair configuration and they seem to be OK. The vendor has used USB ports to connect the DAQ to the PC and I am not quite sure about the fidelity of this part and or about any possible means of shielding it.

However, I have tried implementing a customized low-pass (5th order) filter provided in the PID palette of LabVIEW and that seems to have reduced the noise only slightly. I am yet to try an averaging or median filter on it but that would happen only after I have exhausted all my options with a lowpass filter. So far, my selection of cut-off for the low pass filter were heuristic but I will now carry out a fourier analysis to exactly determine the cut off. Please advise , James.

Biwajit,

The USB is a digital communication and you should not take care of it's fidelity. 

In the your measurements results you have an aggregation of the process (tank) noise and the measurement noise. First of all you should determine the measurement noise level. In order to do that you have to put inlet pressure ports at the same level and in the same environment. After (a lot) of measurements you get mean and standard deviation of your measurements. The mean value of the measurement helps you to adjust the pressure sensor offset. The standard deviation give the confidence level of your measurements. If you have a large deviation (large noise)  it seems that your measurement system is not good, so you have to verify the cabling track. If you have a small deviation you measurement system is reliable and you have the noise from the process.

If you know the process well enough you can try Kalman filter.

Biswajit,

You've done pretty much what I would have done - as Ioan said, ignore the USB socket - we're concerned about the path from sensor to ADC terminal - and *that* sounds to be okay.

Out of interest, what sort of pressure differentials are you hoping to measure? Pa? tenths of Pa?

I'd look at the frequency spectra of 'useful' data (there'll be a maximum slew rate that you want to sense? Perhaps 10Hz or less?) And then I'd heavily filter everything higher than that.

Is the output of the sensor (electrically spoken) differential or single ended? In case of differential output you must use a differential transmission line and just before the DAQ card you have to convert it to single ended type. Or if the DAQ card has the ability to process differential signals you have to connect the transmission line directly to the DAQ. The input type of your DAQ can be selected in the configuration routine. The most NI DAQ cards can work with differential signals. Pay attention on the shielding, grounding and the ground loops. If the output of the sensor is single ended and perhaps you have a few meters cable between the sensor and DAQ it is preferred to use an voltage to current convertor (voltage to 4-20 mA) , in order to transmit the signal in form of current rather than voltage. This configuration allows higher immunity to the noise. You can use XTR117 from Texas Instruments or similar for example. At the end of the transmission cable you have to read the voltage on a series connected resistor (usually a few tens or hundreds of ohms) and to adjust the offset in your DAQ application.

 Here is the data sheet for XTR117 (it is a few dollars) http://www.ti.com/lit/ds/symlink/xtr117.pdf

See the attached paper for the (IIR) filter coefficients your require (causal smoother). 

https://www.researchgate.net/publication/281103616_Improved_IIR_Low-Pass_Smoothers_and_Differentiators_with_Tunable_Delay

It is my "go to" filter for smoothing noisy sensor data. It works well when the digitizer is fast and the process is slow, so that the data are well approximated by a quadratic curve.

Conference Paper Improved IIR Low-Pass Smoothers and Differentiators with Tunable Delay

Hello, is the output signal of your sensor digital or analog? since you said that it looks like white noise, i suppose that it is analog. In that case i recomend you to put an analog filter between your sensor and the adc. I suppose that your problem is with high frequencies, so i recomend to use a low pass filter (only RC). Try it, it is easy to implement. 

James (and Alejandro), the DPT is of Rosemount Make and has an input range of (0-1000) mmH2O. It gives an analog output in the range of (0 - 5) V and this is converted to digital (0 - 5) V by the DAQ (this is where the trouble starts).

Not much noise can be seen in the raw data coming from the sensor when it is measured with a multimeter. The noise gets added only during the analog to digital conversion process and can perhaps be attributed to quantization noise. So I am left with two options at this moment, using wavelet de-noising or a filtering arrangement although I am yet to figure out the advantages offered by wavelet de-noising technique over traditional low pass or median filters.

Biswajit,

You have a low level pressure sensor. By the multimeter measurement it is possible to have inaccurate measurement because the multimeter has the proper low pass filter. On this low pressure level you can have process noise (eg vibration) or measurement noise. In order to verify your electrical accuracy (noise) or DAQ troubles, replaces the DPT sensor with an resistor. The resistor should have the DPT output impedance value if you know it ,otherwise a simple short.

Biswajit,

> The noise gets added only during the analog to digital conversion process and can >perhaps be attributed to quantization noise.

Can you rule this out? Will the manufacturer tell you the bit-width of the ADC converter?

If your ADC has too few bits, no amount of filtering will help. If the ADC is inherently noisy, there's not really much that can be done.

:(

Do you see any substantial noise on the sensor's output lines?

(a decent storage scope will tell you)

If so, one might try cleaning up the power line to the sensor... and adding a tiny decoupling capacitor.

James, 

>>Will the manufacturer tell you the bit-width of the ADC converter?

I will revert to you shortly on this matter.

>>Do you see any substantial noise on the sensor's output lines?

There is a display connected in the output line of the sensor which gives a fairly stable (read noiseless) analog voltage readout.

>> try adding a tiny decoupling capacitor.

That was the first step I had taken. I already have a 0.1 mF ceramic capacitor installed at the sensor output.

Hi Biswajit,

I agree with Ioan about the DMM having the proper LPF.

Would it be convenient to connect a Digital Storage Oscilloscope at the sensor output and take a look at the Frequency Domain plot and compare the sensor output with and without the connection to the ADC input?

Sources of noise can be very challenging to pin down, and can come from unexpected external sources. For example, when I studied at the University of Technology, Sydney and I built an Op Amp circuit at home and tested it, I got clean waveforms using my old Kikusui 150MHz Bandwidth analog oscilloscope. When I took the circuit to the UTS lab, I got noise that totally compromised the proper operation of my circuit. My fellow students experienced the same situation with their circuits. The lecturer and lab tutor then told us that the Sydney Airport's radar was the major source of the noise; the UTS tower building where the electrical and electronics labs were located, would get regularly swept by the lower lobes of the radar.

I once worked for a now defunct company, where I used a BERT (Bit Error Rate Tester) on digital Telecommunications boards used in the old fashioned ISDN Muxes, and when somebody walked past my lab bench and their cell 'phone would ring, or happen to send the regular alert signal to the nearest cell 'phone tower, the BERT readout would go bananas, and when it first happened I thought the board under test was faulty. After consulting with my fellow engineers, I discovered it was the signal noise from the cell 'phones when their transmitters were activated. I was using the BERT cables that came with the Hewlett Packard equipment I was using.

Later I worked as a Lab Technician in a company's electrical testing laboratory and I had to make some sensitive measurements with a Digital Storage Oscilloscope and the noise from the automated data loggers, swamped the waveform readings I was trying to acquire. I asked for help from one of the Senior Engineers, and he solved it by twisting the DSO coaxial cables together; something I was myself reluctant to do because of my previous instrumentation Industry training, in which I was taught by a cable engineer that twisting coaxial cables was a definite no-no because it can change the cables' characteristic impedance, and may have a detrimental effect on its specified maximum operating frequency. I guess sometimes to tackle noise problems, we have to bend some of the procedural paradigms under which we work, to obtain reliable and accurate data.

If you can measure the noise frequency parameters, these results may assist to identify the source of the noise (there is a small probability it may not be quantization noise), and a more effective remedy may be implemented. Of course the source of the noise may be quantization noise, but from my experiences I discovered its best to consider all possibilities and make careful measurements, and rule out the ones that are not supported by the data.

One area to consider, may be that as Lucian has asked, is the sensor output single ended or differential? Also is the ADC input differential or single ended? If the sensor output's differential and the ADC input is single ended, then the system's CMRR (Common Mode Rejection Ratio) may be compromised.

Here is a link for a pdf explaining the different configurations for ADC interfacing, single ended, pseudo differential and differential:

http://exa.unne.edu.ar/ingenieria/sysistemas/public_html/Archi_pdf/Notas-Aplicacion/understanding%20ADC%20input.pdf

i found another pdf about how to manage noise in a measurement system, that may provide further insight into tracking down the signal noise challenge:

http://www.mccdaq.com/PDFs/specs/Noise-Reduction.pdf

Another potential source of signal noise may be ground loops, which are treated in detail here:

http://www.ti.com/lit/an/slla268/slla268.pdf

The following link deals with a variety of potential noise sources with analog signal transmission, such as with transducers:

http://www.omega.com/literature/transactions/volume2/analogsignal2.html

I have been trying to find a web site, written by a PhD level consulting Engineer who is also in charge of a consulting company that specializes in identifying and resolving issues of electrical and electronics noise. I haven't had much success in locating his web site. I do recall him mentioning that precautions should be taken in mixed signal systems, where analog ground and digital ground should ideally be isolated.

Also all grounds in a circuit or a system, should be joined to the same common earthing/grounding point to have the best chance of reducing ground loops. The PhD consulting Engineer also mentions of cases where other Engineers have wasted hours, days or in some cases weeks, and lots of money trying various solutions such as grounding all shielded cables on both ends, installing ferrite loops around cables, building Faraday cages, and putting everything inside all metal enclosures with ventilation holes that end up allowing noise into the system. And the solution ends up being discovered only when the Engineers decided to start taking careful measurements to identify the source, or they give up after much hair tearing in frustration and call his company for assistance.

Resolving noise issues is both an art and a science, to be approached with the same level of detailed investigation as designing an electronics system or designing the thermal dissipation systems required to keep the electronics cool.

Neil,

The DAQ, as of now, cannot be dismantled by me because it is under warranty. However, I have tried using a median filter in the labVIEW interface and that helped quite a lot.

James,

I have contacted the vendor and they say that its an 8 bit ADC. 

Biswajit