I attached x.  You will find only one channel in x. The complete x contains hundred of channels.  With that, the x(:,17) in my earlier comment should change from x(:, 17) to x.

This x should be the same as the 'ydata' of children 2 of the gcf in any of the ofig. 

figure; plot([1:length(x)]*dt, x);
[tp, ind]=dat2tp(x); hold on;
plot(ind*dt, tp, 'ro');
title('WAFO dat2tp');
hgsave(gcf, 'wafo.ofig');
figure; plot([1:length(x)]*dt, x);
[PKS, LOC, EXTRA] = findpeaks(x,"DoubleSided"); hold on; plot(LOC*dt,PKS, 'ro');title('Octave findpeaks');
hgsave(gcf, 'octave.ofig');

(file #56695)

I was reading bug 66241 and decided to run a data on the findpeaks. 

The specific channel of data used is physically taken on a component.  It is chosen arbitrarily, not specific to trick findpeaks.  

Turning points are widely used in the industry, see nCode or LMS.  In matlab/octave domain, Wafo is a very well established code for fatigue estimation.  https://github.com/wafo-project/wafo

My understanding is that Octave peaks is the upper half side of the turning points. 

From the ofig, you can see that findpeaks misses peaks. 

To generate the ofig, (attached)

dt=1/fs=1.9531e-03.  x(:, 17) is in the ofig
 
figure; plot([1:length(x)]*dt, x(:,17));
[tp, ind]=dat2tp(x(:, 17)); hold on;
plot(ind*dt, tp, 'ro');
title('WAFO dat2tp');
hgsave(gcf, 'wafo.ofig');
figure; plot([1:length(x)]*dt, x(:,17));
[PKS, LOC, EXTRA] = findpeaks(x(:,17),"DoubleSided"); hold on; plot(LOC*dt,PKS, 'ro');title('Octave findpeaks');
hgsave(gcf, 'octave.ofig');