Noise reduction. Software or hardware?

PCM programmer · Joined: 06 Sep 2003 Posts: 21708

My question are there any other methods of software filtering that you
can recomend, That do not bog the cpu down?
--------------------------------------------------------------

Here are two filter routines. The first one is a median
filter, which is also called a "spike filter". It's useful
because it removes the occasional bad sample ("flyers").
You don't want to include obviously bad data in your average.

The 2nd one is an averaging filter.

With both filters, you just feed it the latest element (ie.,
from an A/D sample, or a CCP sample). It keeps track of the
previous several elements in a static array. So essentially,
these filters can work with streaming data. You give it the
latest reading, and it will return the latest filtered result.

Both filters will automatically adapt for the initial situation,
where they don't have the normal number of elements yet.

I used these in a tachometer project.

Sherpa Doug · Guest

:=
:=Listed below I've got a sub that adds 16 AD values strips of the last bits. It seems to work well enough as a crude averaging method and doesn't require too much cpu. I also have a .1 uf cap as a hardware filter. This is because there is a 3 phase motor controlled by a VFD in the same encloser and I'm trying to reduce the amount of noise. My question are there any other methods of software filtering that you can recomend, That do not bog the cpu down?
:=
:=
:=Update_turret_position()
:={
:=
:=(long)valve_position_accumulator+=Read_ADC; valve_position_accumulator_counter+=1;
:=
:=if(valve_position_accumulator_counter==15)
:= {
:= valve_position.L = (long)valve_position_accumulator>>4;
:= valve_position_accumulator_counter=0;
:= (long)valve_position_accumulator=0;
:= }
:=}

A lot depends on the nature of the noise. If you just get occasional bad readings but most readings are fine then averaging just lets the bad reading spoil the whole lot.

Instead you could try what we call "Olimpic scoring". You take a few readings, as few as 3. You keep track of the maximum, the minimum, and the sum. When you are done subtract the minimum and maximum from the sum and average the rest. If there was one spurious high or low reading it is eliminated and most of the "good" data is averaged. It is a simplified version of what is called a "median filter." If you know all your bad readings will be high (or low) you can simplify further.

Something else to try is to synchronize your A/D readings with the motor power. With a VFD that could be hard though.

___________________________
This message was ported from CCS's old forum
Original Post ID: 13212

Neutone · Joined: 08 Sep 2003 Posts: 839 Location: Houston

:=
:=Listed below I've got a sub that adds 16 AD values strips of the last bits. It seems to work well enough as a crude averaging method and doesn't require too much cpu. I also have a .1 uf cap as a hardware filter. This is because there is a 3 phase motor controlled by a VFD in the same encloser and I'm trying to reduce the amount of noise. My question are there any other methods of software filtering that you can recomend, That do not bog the cpu down?
:=
:=
:=Update_turret_position()
:={
:=
:=(long)valve_position_accumulator+=Read_ADC; valve_position_accumulator_counter+=1;
:=
:=if(valve_position_accumulator_counter==15)
:= {
:= valve_position.L = (long)valve_position_accumulator>>4;
:= valve_position_accumulator_counter=0;
:= (long)valve_position_accumulator=0;
:= }
:=}

This takes advantage of bit shift divide and a simple multiply.

Current_reading = Read_ADC;
Filtered_reading = (Current_reading + (Filtered_reading * 15)) /16;
___________________________
This message was ported from CCS's old forum
Original Post ID: 13214

Eugene Onishi · Guest

:=This takes advantage of bit shift divide and a simple multiply.
:=
:=Current_reading = Read_ADC;
:=Filtered_reading = (Current_reading + (Filtered_reading * 15)) /16;

I didn't know that it would
___________________________
This message was ported from CCS's old forum
Original Post ID: 13215

Hans Wedemeyer · Guest

You did not
I read the replies and no one asked the few important questions.

1) How often do you need to sample the analog data ?

2) Similar to question 1) what is the highest frequency you need to sample or is it a DC level ?

3) You mention a 0.1uF Cap. is it on the input to the ADC, if Yes, is there a resistor in series with the signal before the 0.1uF ? what is the value ?

4) what is the source impedance of the signal ?

Answer these questions and then we can decide what is the best solution.

Some other points to note about assembly.
It is important to make sure the Analog and digital circuits have a common ground at ONE POINT, and not ground loops. In a high noise situation (as you describe) it is possible to inject noise through "poor wiring" and "poor wiring layout".
If it is electromagnetic induced noise, you can do a lot by laying your cables away for other power cables, simple sheilding helps. Using a differential connection to the source of the signal would be best, however that means circuit changes.
___________________________
This message was ported from CCS's old forum
Original Post ID: 13217

georpo · Joined: 18 Nov 2008 Posts: 281 Location: Athens, Greece.

Can you please give an example of how to work with the filter?

What is the filter Filter_Index? Should I keep track of it?

Thanks!!!
_________________
George.

PCM programmer · Joined: 06 Sep 2003 Posts: 21708

You question is about the filter routines posted earlier in this thread.

This is how they are used:

georpo · Joined: 18 Nov 2008 Posts: 281 Location: Athens, Greece.

Thanks a lot!! I will try it at once!
_________________
George.

Ttelmah · Joined: 11 Mar 2010 Posts: 19504

Some more comments:

First it is always better to reduce noise 'earlier'. So hardware filtering
is preferred to software. Ideally probably do both.

As others have said a lot does depend on the nature of the noise. So
for instance, if there is an 'occasional' spike on just one reading, the favoured
filtering method would probably be Olympic averaging.
Basically exactly the same as the original posted filter, but before the
filter load a value called 'max' with 0, and one called 'min' with 65535.
Then instead of adding 16 values, add 18, and at each loop test the new
value against min, and max. If it is above max, load max with the value
if it is below min, load min with the value. Then at the end of the 18
samples, subtract min and max, from the sum, and generate sum/16.
The result is you get an average of sixteen values out of eighteen without
the highest and lowest. This rejects any single value spike much better.

temtronic · Posted: Thu Nov 12, 2020 8:21 am

As Mr. T. says hardware FIRST, then software.

What's unknown, at least to me, is Which PIC and what ADC selection ? There's a HUGE difference in capturing 10, 12 or 16 bit values compared to 8 bit data.
Anything over 8 bits needs real good, no, GREAT! pcb layout, selection of components, grounding, shielding, etc. I designed 16 BIT ADC systems to work INSIDE optical emissions spectrometers (think of a 20KV spark, several 100s of ma), so I have an idea about 'noise'. I got solld 14-15 bits on all 32 channels of data back in late 70s.

Is this just one ADC channel ? Are the others terminated and disabled? A simple R-C front end filter on the channel ? Might need a 3-4 section filter depending on the 'noise'.

If you get the hardware right (within 2 bits, all the time), then software can smooth out the numbers.

I use the 'Olympic average' method.
Take 10 readings ( 12 bit in my case),toss out the max and the min readings which leaves 8 'good' readings. A simple divide by 8 gives the result.

Jay