Acceleration to Displacement - Filtering Issue

[sanchezb]

I am using fast fourier transforms to convert raw acceleration to displacement. I am able to get extremely close to expected results; however, I am having issues at the beginning of the dataset. I believe there may be an issue with the filter I am using.

Here is the code for the butterworth filter I am using

[B,A] = butter(5,0.5/(Fs/2),'high');

I apply it to the raw data and after conversion.

Here are the results, I move the device +/- 10cm, +/- 7.5cm, and +/- 5cm as seen. The green line represents actual movement and the blue line represents the dataset produced through the conversion algorithm.

Any suggestions for better filtering techniques? I can provide the code and some data upon request.

 

[sanchezb]

Here is the code

clear
clc
home

load ACCfilter
load ElapsedTime

%definitions

Fs = 32;
t = ElapsedTime; % sampling range
yno = ACCfilter;

[B,A] = butter(5,0.5/(Fs/2),'high');
y = filter(B,A,yno);

%plot in time domain

subplot(2, 1, 1); %
plot(t, y,'r'); grid on % plot with grid
title('Acceleration vs Time')
xlabel('Time (s)'); % time expressed in seconds
hold on
plot(t,yno,'g')
%FFT section

Y = fft(y); % compute Fourier transform enter
n = length(y); %
Amp = abs(Y )/n; % absolute value and normalize


%if n is even the fft will be symmetric
%the first n/2 + 1 points will be unique and the rest are symmetrically
%redundant. Point 1 is the DC component. Point n/2 +1 is the nyquist
%component

%1 2 3 4       5           6  7  8
%4 3 2 1      0          -1 -2 -3


% if n is odd the nyquist component is not evaluated. The number of unique
% points is (n+1)/2

%1 2 3 4 5  6  7  8  9
%5 4 3 2 1 -1 -2 -3 -4


%create the frequency axis

NumUniquePts=ceil((n+1)/2);

% the positive frequencies   

freq=(0:NumUniquePts-1)*Fs/(n-1);

%mirror the positive frequencies but make them negative then adjust some things
%to look right

freq2= -1*freq(end:-1:1);
if mod (n,2)==0
    freq2(1)=[];
end

freq3=[freq, freq2];
freq3';
size(freq3);

if mod(n,2)==1
freq3(n+1)=[];
end
if mod(n,2)==0
    freq3(n)=[];
end

%store freq3 in a column

freq3c = freq3';

%plot acceleration in frequency domain

subplot(2, 1, 2);
plot(freq, Amp(1 : NumUniquePts),'b.'); grid on % plot amplitude spectrum enter
xlabel('Frequency (Hz)'); % 1 Herz = number of cycles per second enter
ylabel('Acceleration Amplitude'); % amplitude as function of frequency enter


%Omega Arithmetic to convert acceleration to velocity
%uses complete frequency vector from above
V=Y./(2*pi*freq3c*1i);
%result=[freq3',Y',G']   %use this line to look at freq3, Y, and G next to each other 


% Get rid of infinities resulting from divisions by zero to prepare for
% ifft
V(1)=[1];
if mod(n,2)==0
    V(n)=[1];
end



%go back to time domain with the velocity

inversedVno=ifft(V);

[B,A] = butter(5,0.5/(Fs/2),'high');
inversedV = filter(B,A,inversedVno);


figure
plot(t,real(inversedV),'b')
hold on
title('Velocity vs Time'); % amplitude as function of time
xlabel('Time (s)'); % time expressed in seconds
ylabel('FFT Velocity');

% FsdirV=100; %sampling rate
% dtdirV = 1/Fs; %
% etdirV = 3; % end of the interval
% tdirV = 0 : dt : et; % sampling range
% ydirV = -cos(4*pi*t)/(2*pi);% define the signal
% plot(t,ydirV,'g')


%Omega Arithmetic to convert velocity to displacement
D=V./(2*pi*freq3c*1i);

% Get rid of infinities resulting from divisions by zero to prepare for
% ifft
D(1)=[1];
if mod(n,2)==0
    D(n)=[1];
end

%go back to time domain with the displacement

inversedDno=ifft(D);

[B,A] = butter(5,0.5/(Fs/2),'high');
inversedD = filter(B,A,inversedDno);

 

[IRstuff]

PLEASE DO NOT DOUBLE POST

TTFN
faq731-376

 

[FeX32]

Yes never double post!!!!

Whenever you use a high-pass filter there will be some magnitude and phase warp. The warp or "lag" is frequency dependent. Consequently, the estimated displacement is just that, an estimate. You can predict how different the actual from estimated will be by several methods. Firstly, look at a bode plot of the filter input-outputs.

It is evident from your pic that this is what is happening. First easy thing to try is increase the cutt-off frequency of your butterworth filter.


Fe (IronX32)

 

[btrueblood]

Your accel data is not very well windowed, i.e. it starts at a non-zero value and instantaneously reaches a very high level. The FFT as a result is giving you some weird harmonics (aliasing? not the right term, but like that) near the start.

 

[GregLocock]

You could try integrating in the time domain. I'll have a look at your code if you post a zip with ACCfilter and ElapsedTime in it.

Cheers

Greg Locock


New here? Try reading these, they might help FAQ731-376

http://eng-tips.com/market.cfm?

 

[FeX32]

Hmm, or filtering in the frequency domain?


Fe (IronX32)

 

[IRstuff]

I ask again, what did you expect, and why do you think this is wrong?

TTFN
faq731-376

 

[MikeyP]

You could try running the signal through the filter twice - once forward and once backward - to remove filter phase effects. Also you should check that you are doing the transforms correctly: a useful check is to see if you have any imaginary part in your final result (you may see some imag part down at machine precision level - that is OK - but anything significant means you have gone wrong somewhere).

M

--
Dr Michael F Platten

 

[hacksaw]

why do you feel it necessary to use a high pass filter on accelerometer data?

agree with gL, a low pass filter might be more suited to your problem

 

[FeX32]

I didn't even notice it was high-pass! For sure that is one of the problems...


Fe (IronX32)

 

[MikeyP]

You high pass filter when doing this sort of thing to avoid drift problems when you integrate. Just removing the DC isn't usually enough.

Ta

M

--
Dr Michael F Platten

 

[FeX32]

What about a band-pass? Seems appropriate.


Fe (IronX32)

 

[sanchezb]

I'll try a bandpass