Thursday, December 21, 2017

December Flight Test

Paul flew again last Saturday, December 16th from the FAR launch site. If you recall this summer I was fiddling with the Raspberry Pi and Intel Compute Sticks to find the best realtime platform for flight. I had troubles with the Pi, but the first-gen compute stick was able to (barely!) run six trackers. Paul wound up buying a more recent compute stick. This one sports a more recent quad core processor and enough on-board storage to record several test flights (in case of scrub etc.). It comes with Windows 10, and spending a half hour playing with it was actually quite snappy, but ultimately reformatted it and installed Ubuntu Server. The Cherry Trail architecture is fully supported which means both wifi and ethernet work out of the box, even on the console with server.

I meant to have a post on the receiver hardware earlier, but time flies. Briefly, the stick is set up to run gnss-sdr on boot, if a RTLSDR device is plugged into the USB port. gnss-sdr will then attempt real-time tracking while simultaneously dumping the RF signal complex samples to file for post-flight reconstruction. One downside to this is that because of the combination of receivers being used the file dump type is gr_complex which is 32 bits/sample instead of the native 8 bits/sample. The tooling lives here along with the receivers for both realtime and post-flight execution but I need to write up some documentation to make it clear what is what. I tried to use rtl_mus (RTL multi-use server). This would allow one service to stream samples to both gnss_sdr and to file (at native 8 bits/sample, giving 4x the sample storage) but alas I could not get this working reliably and so far as I can tell there is no way to guarantee I and Q pairs as opposed to missing the first I and then getting Q-I pairs.

I sent the stick to Paul. Bench testing weeded out a bad amplifier, once that was replaced we were able to get a nav solution. Paul packaged it all up in the nosecone flight configuration and ... nothing! Likely interference from the electronics, I got the "probably not going to fly tomorrow" email. But then the next day I got the "aluminum tape shielding worked, so I flew it" email!

Unfortunately the on-board realtime tracking got a good fix on the pad, but nothing in flight. Disappointing, but having the RF to post-process is valuable. I tried re-running the code through gnss-sdr several times with several tweaks to the receiver to no avail - spotty tracking on the pad and nothing in flight. So I did what I've done in the past and falled back on SoftGNSS, an offline Matlab-based GPS receiver (link is to my github version of it; updated to run in modern Matlab environment). I had to add the capability to read in complex floating point numbers (prior to this we always logged 8 bit integer samples) but it more or less worked out of the box. As mentioned in previous posts, SoftGNSS is somewhat naive but that naivete means it works where gnss-sdr has to date had trouble, that is with noisy signals. gnss-sdr will drop a satellite receiver if it loses track, whereas SoftGNSS will continue to process that satellite blissfully unawares, which comes in useful when the signal is noisy because yes, you get a few bad samples but then you are back in business.

Seven satellites were tracked through the entire flight. Blue satellites had a low S/N (or were occluded). PRN 5, 21 and 32 were excluded as they had a high S/N but invalid doppler ranges. It is unclear if the code is screwing up the doppler calculation or if the S/N calculation is in error. My gut says the latter because gnss-sdr did pick up PRN 5 and PRN 32 at some point during the flight.

They are plotted below by PRN, the line traces the view of the satellite and the circle is where the satellite started. Axes are azimuth, elevation.

Here are line traces of the range up, east and north, in meters. There is a jump in north position prior to apogee and then in all 3 directions around 100 seconds. Culprit is unclear. Drogue deployment should have been shortly after apogee. It is basically using barometric pressure to find apogee by looking for an inflection in the derivative of pressure. It is possible, but unlikely it fired early. It is more likely that the something went awry in the tracking causing a jump in one of the pseudoranges (I need to dig in). Main deployment is at 1000ft which is where I cut off the plot, because at that point the antenna is occluded (dangling up side down near ground level) and the tracking is all over the place.

And here is a view of the launch site - trace is pretty close to the pad.

And here is the trajectory. It is noisier than I would like but after apogee it is dangling under drogue. I still don't have a good explanation for the kink in the trajectory prior to apogee.

If you want to play with the data:

Tuesday, August 22, 2017

Realtime Tracking on a RPi3 - Update

If you read my last post, my GPS project for this summer was to get gnss-sdr working in realtime on a Raspberry Pi 3 using gnss-sdr. Unfortunately to get right to the punchline, I failed to do so. Using pybombs I built GNURadio and gnss-sdr from scratch along with dependancies that weren't in the apt archive. I tuned the Volk profiles to use the NEON extensions, but my flowgraph suffered from overflows. I tried eliminating the prefiltering and reducing the doppler shift range in the acquisition block and was able to nearly get rid of overflows but by then the doppler range was so small as to not acquire satellites!

I still think I may be doing something wrong. Carles, Javier and others from CTTC in Spain wrote a paper where they document the performance of several platforms including the Raspberry Pi 3. They were able to correlate 6-7 satellites. I was unable to replicate, but there were several confounders. Namely, they used a USRP over network connection whereas I was using a RTLSDR over USB. They are using different source blocks (UHD vs. osmocom) different physical hardware/drivers (Ethernet vs. USB) and it is difficult to tease out the actual culprit.

So, I tried using the first generation Intel Compute Stick we use for the data logging on flight. Like the Raspberry Pi, I had to build GNURadio and gnss-sdr from scratch (it comes loaded with Ubuntu 14.04 and the apt packages don't exist until 16.04) but it only took about 8 hours to complete with two cores. I did disable unit tests when compiling gnss-sdr as gcc would get hung up during unit test compilation (ran out of physical memory and swap, 1GB each). When I first ran it overflowed but just barely. Removing the prefilter and it worked great with four satellites. And five. And six. Seven caused overflows and it couldn't hold a track, so six satellites it is.

The plan is to track a GPS fix in flight sometime in the near future. This platform is acceptable, but barely.  Four satellites are the bare minimum to get a fix, but more is better as it improves the quality of the fix (the root-sum-square [RSS] of the satellites minimizes the influence of a 'bad' invidiual satellite fix). Also in flight during long durations/distances the view in the sky will change and you will necessarily lose satellites and gain new satellies. The process of losing multiple satellites (>2) means you completely lose your fix. Losing 1-2 would merely lose precision. Ideally you can track 8-11 satellites with a good view of the sky.

So, do we stick with the Intel Compute stick or try something a little more computationally advanced? Stay tuned...

Saturday, July 29, 2017

Realtime GPS tracking with a RPi - Summer Project

Long time no post!

Summer gets busy. Work travel. Kids' sports. Tons of excuses but I am back and have a project in mind: realtime GPS tracking with the Raspberry Pi 3.

It has been done before - Carles and Javier from gnss-sdr coauthored a paper which did some performance testing with the Raspberry Pi 3 (along with other platforms). In this case they only looked at correlating the signal not getting a final lat/lon/alt fix but in a personal communcation Carles verified they were able to track 6-7 satellites in real time.

So I set about to replicate this work. I installed Raspbian and attempted to take the "cheater's way" by installing gnuradio using APT then manually compiling gnss-sdr. It worked, but didn't - lots of overruns. Then with a fresh SD card I tried using Pybombs to compile gnuradio and dependancies for gnss-sdr. This broke partway through compile because gcc in the Raspbian repos is compile --with-arch=armv6 which is apropriate for the RPi1/2 but in order to get the full set of NEON extensions you need a compiler --with-arch=armv7-a. This problem and the solution is documented here. Basically you switch back to the debian repos and proceed. I ran into one dependancy issue libgnutls. When repos are switched the package dependancies break - issue a "apt-get remove libgmp10" before proceeding to install with Pybombs.

I was then able to compile gnuradio overnight, and gnss-sdr in an hour. I could have cross compiled but sleep time is free time.

Unfortunately, I still had overflows. Oddly, I don't have a full core occupied or all of the RAM. Not quire sure where the choke point is.

Talked to Carles - one clue he provided is they used a USRP, not a rtlsdr stick, so it's possible there is a driver overhead issue. I do have a LimeSDR and I plan on trying that.

So here's my burndown list

  • Drop sample rate until overflows go away (much less than 2Msps will likely result in failure to track)
  • Try using LimeSDR 
  • Try using an OrangePi - I have a spare from my robotics projects. The board is smaller and half the price, more powerful than a RPi2 but less powerful than a Raspberry Pi, but its unclear at this point the processor is the issue.

Has anyone successfully tracked satellites on a Raspberry Pi using gnss-sdr?