My Blog, with Stuff on.
Im in the middle of writing the Optical Disdrometer data sorting code, and I have managed to get the first image out of it. Excel looks the best, but damn, why didnt Microsoft add an “export current chart to image” function. I had to find some VBA and put it in a macro.
The code needs to go through another rewrite before its nice and tidy. Also need to put some failsafe’s in to make sure that the data doesnt trip over itself (E.g. try and overwrite data).
Suprise result here. Drum roll please….
It turns out that the FMinSearch algorithm was the quickest by about a factor of 2! Finally got the lsqnonlin function working by messing with the options. The profiles of the results are shown below:
FMinSearch Profile (2.189s)
FMinUnc Profile (4.355s)
LSqNonLin Profile (3.577s)
mmFSolve Profile (Custom user’s file) (5.814s)
Speeded up the code slightly (0.2s to be exact) by changing the peak detection routine. This new one has come from a PDF entitled “Writing Fast Matlab Code” by Pascal Getreuer in 2006. Below are a local copy of the PDF and the profiles of the slower and faster code.
Writing Fast Matlab Code PDF
Slower Peak Detection Profile
Faster Peak Detection Profile
Once I had got the code working and solving correctly, it was time to move to a circular tank. This would help to mitigate many reflections, since there is hardly any sides for p-waves to directly bounce off into a hydrophone. This should have been quite an easy thing to do in the code, but it turned out it wasnt.
Firstly, after debugging all the silly errors mainly relating to the Order checking routines and variables, the fminsearch to a stupid amount of time to complete. Where it once took 5 seconds, it now to 50. I couldn’t figure it out, and I thought that it was an error in the code. But I came to realise that it was because the initial estimate has just become a lot worse. In the square version, with hydrophones in the corners, there was a quadrant for each hydrophone, and the estimate was placed in the middle of this quadrant. In the new layout, there are now cake sliced quarters of a circle, with the hydrophone at the edge. Using the estimate as the position of the hydrophone (like before) means that half of the data will be more than double the distance away from the initial estimate. This is like placing the estimate for the square tank on one of the walls, rather than in the middle. Therefore, for now, I will place the estimate somewhere in the middle of the centre and the outside hydrophone.
Attached are PDF’s of the Profiler from Matlab.
Profile for the circular tank simulation
Profile for the square tank simulation
Once I’ve finished with the current robot, I would love to get into some sort of UAV design, something like the 4 rotor Helicopter designs ive seen on the internet (see bottom for references).
I’d also like to continue building the EFPD. Possibly place it on a robot that would run away when someoe got close. Make a super-small version on a proper pcb. Another idea is to attatch it to a ring of lights. When someone walks past then they would light up in sequence. It would look so cool up the stairs.
Could also integrate it into the landing lights, or maybe the main downstairs lights so that they turn on when someone comes past. Cool.
Well after a long period where I had no access to my electronic equipment, the time is nearly upon me to get it all back out (just in time for the christmas holidays!). Heres what I need to do:
Over the past couple of days I have been battling with solving the multilateration equations. Today I got them working in the guise of a numerical solution using matlabs fminsolve function. It took a while to figure out how it worked, but when it did, it worked straight away. The next thing was to throw some real data at it, and this is what has taken much of the day up. There must be still something wrong in the main routine, because I am sure that the numerical solutions are ok.
Either way, it takes for ever. The idea was to take an abitrary set of 3 pulses to test the fourth and iterate over all combinations. Then we keep the iteration that provides the smallest (acceptable) error. The problem with this is that when the rain rate gets up, then it is inevitable that 10+ pulses appear in the same time window. Now, for 10, there are somewhere around 200 combinations of 1-10 in sets of 4. The fminsearch takes somewhere around 10ms to compute, so 200*0.01 = 2s. This is only for one set of 4, and thats even if it works out in the end. So when we have maybe 100 pulses, simulation times get loooong.
There are two ways around this, firstly just dont use the estimation. Secondly find another way. What we could also do with the final device is when the rain rate does get too high, then just raise the threshold and ignore all the smaller drops.
I’ll email kevin tommorow and see what he thinks.
Got back from the Paris IPC9 confrence now and overall it was a pretty good experience. I felt really nervous though when I was actually giving the presentation. I thought it went quite badly, but others said it ok; probably being paranoid.
When I got back ive continued with the code, mainly trying to get the multilateration sorted. Its turned out to be a right pain. I cant seem to solve the equations. I tried to do it analytically, but I couldnt find a way. Kevin said to try it numerically which I will try shortly.
Also, finally got a laptop.
Had the TAR meeting today, someone called Michiel Postema, was chairing the meeting with Paulson. Exccentric is not the word. They said the report was good. Background was a bit diluted. Paulson loved the presentation layout. Practised the presentation again; took about 20 mins but I spent too long describing bits.
Rewrote the generation code, was very messy and I think there were a few problems. Also rewrote the check arrival order function. It works ok now, but it doesnt actually delete anything. Just checks that there is 4. We are still getting a pretty high error rate, so I think it would be beneficial to go through it one more time to get those errors down. Once that is done, then it is about time to start the estimation routines we were talking about the other day.
Finished the presentation, do need to practise it a bit more. Got paulson to look at it. Altered some of the content, but likes the look.