Here are key points in a few words:
UPDATE 1: 20ms is not enough
Tests were performed with some number of flights. This fix don't breaks anything. But it doesn't work with 20ms discharge time. Going to increase it up to 1s, but I have not too much hope it will do. It seems pure software solution will not fix it w/o hardware fixes.
UPDATE 2: It depends on temperature
I have two items of HKPilot32. I flashed item #2 with original ArduCopter-3.2.1 Item #1 already flashed with 1s power reset fix. I put item #1 in refrigerator and started with item #2.
Powering attempts numbered as they were performed in common sequence. Powering cycles were performed by battery disconnection and connection back. To reproduce issue a disconnection time should be less than 1 second. My results are below:
Item #2, original 3.2.1, room temp about 22C
bad: 2 4 5 6 7 8 9 10 11 12 14 15 16 17 18 19
good: 1 3 13 20
Item #2, 1s reset fix for 3.2.1, room temp about 22C
good: all from 21 to 40
WOW! Doesn't it work?! Let's check it with cooled item #1.
I took item #1 out of the refrigerator and put there an item #2.
Item #1, 1s reset fix for 3.2.1, from refrigerator
bad: 43 44 45 46 47 48 49 50 51 52
good: 41 42 53 54 55 56 57 58 59 60
No luck. What about to warm it a bit?
Item #1, 1s reset fix for 3.2.1, after 30 minutes in room temperature
good: all from 61 to 80
It looks like it works for room temperature! But what about cooled item #2?
Item #2, 1s reset fix for 3.2.1, from refrigerator
bad: 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100
Being cooled it doesn't work again.
All logs are here. Next step I going to play with 5sec power reset on cooled FMUs..
UPDATE 3: It mostly works for 5 sec power reset time
Item #1, 5s reset fix for 3.2.1, from refrigerator
good: 101 102 103 104 105 106 108 109 110 111 112 113 114 115 116 117 118 119 120
With 5s reset time it looks like it mostly works for the same item and same temperature.
Also it seems for the attempt #107 I've had really small (about 0.2s) disconnection time. It is an unlikely scenario of actual copter use.
Link to test image is updated, now it have 5sec reset time fix. It would be nice to have test results from people suffering from this problem.
UPDATE 4: It seems temperature dependency is not about capacitors
I measured discharge process on C506 of the same FMU item for two cases: a) FMU have room temperature and b) FMU has been cooled 30 minutes in refrigerator. It appears that measured values difference is really subtle. It is about 20mV difference for discharge level on 5 sec period. So it seems the temperature-capacity dependence is not a factor for this issue.
UPDATE 5: MOSFETs test results
Performed tests with SPICE models and in hardware (a number of MOSFETs were used) demonstrated that issue cannot be resolved reliable w/o hardware fix due to the fact Rds goes high when Vds and Vgs going down. VDD_3V3_SENSORS rail needs to be connected to GND via 220 Ohm resistor and it still need a software fix. As result the EN1 input of mic5332 doing its job perfectly w/o any MOSFETs. But it will not work reliable for the case with resistor connected between any other power line and GND, this way the issue still possible to happen for cases of short power disconnection.
UPDATE 6: hardware fix is enough for most cases
I must admit that it seems working ok even without a software fix. Just performed a number of tests with original ArduCopter-3.2.1 firmware on cooled (all night in refridgerator) board with 220 Ohm resistor which is installed in parallel to C506. As result: by disconnecting and connecting battery back with my hands I was unable do it in enough short time to reproduce issue. All logs from attempt 121 through attempt 140 are showing IMU2 works just fine. Software fix still needed for non-fixed boards only. In case you are able to (de)solder SMD parts the better place for resistor is instead of Q601, you need to desolder Q601 and solder resistor on emitter and collector pads.
Owners of non-modified boards are have to wait for a software fix. It have a chance to be included in ArduCopter-3.3.3.
Here is a long story:
Some time ago I decided to switch from APM to PixHawk, so I bought two items of HKPilot32 FMU from HobbyKing. So far I can see they are clones of PX4 FMU v2.4.5. Boards are labeled as "px4fmu-2.4".
Those days I had not too much experience with PX4 platform, so all intermittent issues and misbehaving were considered as a lack of my skills. When I first stuck with BAD ACCEL HEALTH issue I googled for the issue and found some mentions of power filtering and so on. So I decided to put some ferrite rings on power lines to FMU and FPV, to decouple possible noise and spikes on power lines. It seemed as it helped me for some time and I've seen no more BAD ACCEL HEALTH for a long time.
But later it happened again. So I decided to investigate it in a more details and googled this thread.
While checking my DataFlash logs I realised it wasn't MPU6K issue. It was definitely LSM303D issue. I read through all in the mentioned thread and I collected all important points together in a single document. Also I collected there some important points drom other forums like HobbyKing and Plololu. (All 'hello' and 'thanks' were skipped.)
I discovered that possible key to the problem is in the message by Artem on February 19, 2015 at 11:48am: "first hit is pololu community forum thread. Apparently lsm303d is very sensitive to how it needs to be powered down". So I googled once more and found original message on Pololu forum: "When we were testing the accelerometer on the LSM303, we noticed we could get a similar behavior where the accelerometer constantly reported a single value for all axes. It seems that if you interrupt power to the accelerometer in a certain way (like disconnecting or turning off/on power) so that the voltage falls below a certain amount but not all the way to 0, then it can brown out and get stuck in a bad state".
So I disassembled my HKPilot32 and downloaded all pix4fmu-2.4.5 schemas and pcb layouts. As you can see there are capacitors (C506, C507) on VDD_3V3_SENSORS net close to LSM303D, also there are some capacitors on the same net close to LDO and some capacitors should be close to other sensors:
Having a hope that I would be able to solder SMD resistor on top of C506 I found exact location of C506 and started to measure LSM303D power voltage on C506 pins:
Here is how LSM303D power drops down once battery get disconnected. There are no any other modules connected to FMU or PM (Power Module). As you can see it takes about 50ms to drop on level 0.8V and it takes about 0.1 sec to drop on level 0.6V. And it doesn't fall below 0.344V even after 1 minute since battery get disconnected from PM. It's definitely not good. And it can be even worse in case other modules are connected to FMU or PM.
I was ready to solder SMD 1k resistor on top of C506 when I suddenly discovered Q601. Then I traced VDD_3V3_SENSORS_EN net and..
WOW! It is controlled by PE3 pin of STM32F4 and it is there definitely to discharge sensors power line and to shutdown LDO output. By design it looks like it should do sensors power line shutdown in a proper way. It was designed to do exactly that we need to avoid BAD ACCEL HEALTH.
So why it doesn't work, what I missed? Broken Q601 transistor? PCB issue? Anything else? Well, let's measure it on R620.
To shutdown sensors power line the base of Q601 should go down with logical zero on PE3. And I believe it should happen on FMU initialisation. Ok, FMU is powered, pushing reset button..
During FMU initialisation it goes up for some (really long 5.4sec) time, until PE3 will be configured as an output. Ok, no problem. Then I setup trigger for my oscilloscope on 3V level to catch voltage drop and no luck. It means Q601 doesn't receive logical zero from PE3 on FMU initialisation. At all.
NOTE: Hereafter I will refer to source code of version 3.2.1.
Ok, it's to time to dig the source code. What about to grep PX4Firmware sources for "SENSORS_EN" word? Here it is. It is used in "sensor_reset" method to do exactly what we needed. Let's see where "sensor_reset" methods is used. Nowhere else, for the moment of v3.2.1 released it is used just in the same file "PX4Firmware/src/drivers/px4fmu/fmu.cpp". What about v3.3.1? Checked, the same.
After browsing fmu.cpp file I can state definitely: for versions 3.2.1 and 3.3.1 it is used just to provide ability to run nsh command "fmu sensor_reset" with an optional parameter. Nothing more. And it definitely doesn't used by FMU initialisation process. Ok, let's try to execute "fmu sensor_reset 20" to shutdown power line for 20ms.
It works! As you can see it falls down to 0.8V in 1.8ms and to 0.6V in 10ms.
The bad side is that it doesn't falls down to zero due to Vce(sat) parameter value for Q601. Q601 is BJT transistor of type MMBT3906 and it have Vce(sat) from 0.25V to 0.4V. So VDD_3V3_SENSORS net cannot be discharged below Vce(sat). In case a pure software solution with PE3 and Q601 will not do for BAD ACCELL HEALTH, then design for Q601 should be changed to use some MOSFET instead. It permits to discharge sensors power line much closer to zero level. Also there are low Vce(sat) BJT switching transistors on the market.
The good side is that level of power line discharge isn't an only important factor of issue. The rate of discharge may also affect LSM303D behaviour and we able to fix it right now. Anyway, here is chance it will do so let's try it.
I implemented it this way and performed FMU reset by pressing reset button. Next moment my board is turn on multicolour LED as red and it means a error while initialisation. I connected to PX4 terminal and found following startup error messages:
Starting APM sensors [MS5611_SPI] on SPI bus 1 at 3ms5611: interface init failed
Error in startup
Then I turned on debug messages and realised that SPI is affected:
Starting APM sensors [MS5611_SPI] on SPI bus 1 at 3ms5611: prom all zero
ms5611: prom readout failed
ms5611: interface init failed
Error in startup
After some reading of "sensor_reset" source code I discovered the cause of SPI issue. It doesn't enable back SCK/MOSI/MISO pins after reset. So I fixed it and then FMU loads ok.
As a bottom line:
Comments and corrections are appreciated!
Regards, Dmitry Prokhorov
You shouldn't have done that fix, Dmitry found that it only partially fixes the problems as you noticed yourself. To fix all the gyro and accel health you just had to do the fix i did above, remove Q601, replace the R621 resistor with a 220ohm and make the little bridge.
After doing the fix i can fly in -2 no problems (we haven't had colder weather so far to test).
Dmitry, could you update the main post with the latest working fix, so that people don't try the older fixes you tried that didn't fully fix the issues?
You can use my images to show where to do the fix for the adventurous users.
Both fixes, yours and performed by Nickolay are the same by meaning. Both of them do connection of VDD_3V3_SENSORS rail to GND via resistor. The only difference between your fix and Nickolays fix is a resistor placement on a board.
Also both fixes are hardware-only and should be considered as kind of partial fix due to imperfect reliability, say it able to fix the issue not for all cases, but for MOST OF them. There is a minor chance to still have the issue in some quite-impossible cases. And we still need sowtware part of fix to avoid it completely FOR ALL cases. Adoption of software part of fix is still in the progress. So hardware solution will do for now for most cases.
p.s. Resistor value that was proposed by h/w developers is 220 Ohm.
Thank you for proposal, Johnex!
There are no "older" and "newer" fixes, they are equal. See my response above for details.
I going to update main post as soon as will have news on progress with software part of fix.
Ok Dmitry )
It may work for majority of cases, but hardware part of fix is appreciated.
See my UPDATE 3 to the main post. It is for 5sec software reset w/o hardware fix applied. There are 19 attempts are good and 1 is bad. Success ratio may get worse in more cold temperature conditions.
I did this mod too, but I have not yet powered the board. I have a noob question, are those two big capacitors(?) on top of the solder bridge supposed to be in contact with the bridged pad? (because mine are :) )
Just to be sure do I have a short or is all ok to go..
Tried to get scematics but I don't have software to open those files and nor am I skilled enough to read them anyway..
And thanks Dmitry for great job finding the solution!
The pad doesnt seem to be touching, you can use rubbing alcohol to clean the flux core off the board to see clearly.
Anyway i checked the schematic and those 2 caps are connected to the 3.3 sense line, the one we added the 220ohm resistor + bridge to, so there shouldnt be any issue.
Clean up the board with some alcohol and give it a test :)
Also, never use the continuity tester of your multimeter on a board like the pixhawk, remember the components are connected and you don't know what voltage is given out of your probes, if its 9V it could damage the 3.3V chips. Also it doesn't help much if components are linked together in a complex matter, it can enter one chip and out another in the wrong way or in the reverse through diodes showing a short where there really is none when the thing is powered on.
When i did the mod to my board i just made sure everything else was untouched surrounding the work area.