Find a portion of a signal using dtw ?

19 次查看（过去 30 天）

David Santos 2024-7-19

0
链接

此问题的直接链接

https://ww2.mathworks.cn/matlabcentral/answers/2138826-find-a-portion-of-a-signal-using-dtw

评论： Umar 2024-7-23

testDTW.mat

I have the two attached signals and I want to know if "ref" is included somehow (time warping could happen) inside "det" signal.

Figure 1 show ref and det signals:

As you can see ref "can fit" inside det (ref is similar to det(1:100) aprox ) so I try to adjust them normalizing signals and using dtw:

refn=ref/max(ref);
detn=det/max(det);
[~, ix ,iy]= dtw(detn,refn);
figure;plot(refn(iy));hold on;plot(detn(ix));
max(normxcorr2(refn(iy),detn(ix)))

Obtaining:

with a normalized correlation of 0.86

I would love to be able to dtw adjust ref signal over det with the nedd of both signal be the same size so the adjust could be better. For example in this case I could have the signals between x=57:204 and get a better correlation (0.99) of both signals:

x=detn(ix);
y=refn(iy);
max(normxcorr2(x(57:204),y(57:204)))

But I'm missing the first part of ref signal (which could fit quita well) and also some samples at the end because dtw always thinks that both signals have to be the same size but I'm looking also for the possibility of being just a fragment of the other.

Any clue on how to get this adjustment of signals when one can be a portion of the other?

All the best

7 个评论
显示 5更早的评论隐藏 5更早的评论

Umar 2024-7-22

编辑：Walter Roberson 2024-7-22

在 MATLAB Online 中打开

Hi David,

Here's a simplified example to get you started. This custom algorithm should allow for the adjustment of warping path constraints to facilitate partial matches between signals, offering a more personalized solution to the specific problem being addressed.

function [distance, warpPath] = customDTW(signal1, signal2, constraint)
% Initialize variables
n = length(signal1);
m = length(signal2);
warpPath = zeros(n+m, 2);
% Fill the warpPath matrix with constraints
for i = 1:n+m
    warpPath(i, 1) = max(1, i-constraint);
    warpPath(i, 2) = min(n, i+constraint);
end
% Calculate the DTW distance
D = inf(n, m);
D(1, 1) = abs(signal1(1) - signal2(1));
for i = 1:n
    for j = max(1, i-constraint):min(m, i+constraint)
        cost = abs(signal1(i) - signal2(j));
        D(i, j) = cost + min([D(i-1, j), D(i, j-1), D(i-1, j-1)]);
    end
end
distance = D(n, m);
end

So, I define a function customDTW that takes two input signals signal1 and signal2, along with a constraint value. The function calculates the DTW distance between the two signals while considering the constraints for the warping path, it also initializes variables and creates a matrix warpPath to store the warping path constraints based on the input constraint value.

It then computes the DTW distance by iteratively calculating the cost of aligning each point in signal1 with each point in signal2 within the specified constraints.The final DTW distance is stored in the variable distance, which is returned along with the warpPath matrix. Please let me know if you have any further questions.

David Santos 2024-7-23

在 MATLAB Online 中打开

Thanks a ot for your answer Umar!

So if I want that signal1 stay unmodified (no warping at all) should I put that constraint to 0?

% Fill the warpPath matrix with constraints
for i = 1:n+m
    warpPath(i, 1) = max(1, i-0);
    warpPath(i, 2) = min(n, i+constraint);
end

And for the indexes (ix, iy in the matlab dtw function) i Guess It must be obtained from the D matrix but do you know how?

All the best

Umar 2024-7-23

Hi David,

To keep signal1 unmodified (no warping), you should set the constraint to 0 in the warpPath matrix. By setting warpPath(i, 1) = max(1, i-0) and warpPath(i, 2) = min(n, i+constraint), it will make sure no warping for signal1.

Now, to obtain indexes (ix, iy) from the D matrix in DTW function, you can find the optimal warping path by backtracking through the accumulated cost matrix D. The indexes can be extracted by following the path with the lowest accumulated cost. It sounds like you are making progress.

请先登录，再进行评论。

请先登录，再回答此问题。