Category: Motor Controllers

Some more random boards: The sm00thsensor Project

chorl3 Comments

In keeping with my promise of not building anything else huge and parking-requiring this semester, I’ve been investigating in more detail some other potentially useful electronics projects which have been the products of discussions with peers as well as my uncalibrated imagination. This is one idea that has been in on-and-off development for a while.

(Also, buy my stuff.)

One thing that we like to do around this neck o’ the silly vehicle woods is to append Hall Effect rotor sensors onto motors which don’t have them. The majority of industrial brushless motors have 3 Hall sensors embedded in the windings in order to pick up the position of the rotor magnets – this is so that the controller can switch the windings in the correct order, even at extremely low speeds. The sensors essentially become a 6-state absolute encoder. The positioning afforded by this is extremely coarse (for instance, the example 3-slot, 2 magnet pole motor in the link will only have 60 degree positioning clicks), so for high-precision applications like robot arms and CNC servos, this is usually backed up (or totally replaced) by a sinusoidal resolver or much finer optical absolute encoder.

Plain old sensors are well-suited for vehicle traction purposes, though, where the motion is generally in 1 direction and whose timescale of change is much greater than 1 motor electrical cycle – the motor usually has to apply a few seconds of effort to get the vehicle up to speed. Pretty much all e-bike hub motors combine Hall sensors and a very high tooth/magnet poles count to get super-smooth motion at low rotational speeds.

sensors on r/c motors

For indirect drives that use stock or modified R/C style outrunners, adding sensors needs some creativity. There’s usually two approaches: putting them inside the motor and mounting them externally.

I generally favor the inside approach, in part because I usually construct hub motors (with few exceptions, you can’t sense externally with a hub motor). However, fixing the sensors inside also has disadvantages.

  • You have to know how the windings are disposed. The sensors need to be in key locations between phases (putting them accidentally in between 2 teeth of the same phase means they are always mistimed by 30 electrical degrees) and this is only possible to find out if 1. you wound the motor, or 2. you have a sensitive enough magnetometer to check each phase to see which direction the teeth’s fields are pointing.So, it’s riskier than externally mounting them because you have to crack open the motor again in order to fix it if you got it wrong – plus, the motor will totally run like this if the right combination of phase and sensor wires are achieved, but it will sound terrible and draw much more current to produce the same torque, because some of the current is wasted on firing the windings at the wrong time.
  • There is no ability to fine tune the sensor location to account for timing errors caused by sensor hysteresis.

The sensor boards that I’ve been using and which are occasionally seen on this site are custom made with angular slots to allow easier “combo finding”, since I did not wind those outrunners and therefore their internal layout is a mystery. They were developed for “2.00EV“, which ran entirely off the principle of using R/C outrunners, and I judged sensored control to be the more reliable method.

The second issue is one which is concerning for vehicle traction that needs to be bidirectional (i.e. have a reverse). I’m not sure if it’s just not been noticed because most R/C motor modified vehicles don’t ever go backwards (e.g. bicycles) or what, but on vehicles like Landbearshark and tinykart the issue of hysteresis-induced timing error can be quite severe.

The majority of hall sensors employed on motors are “bipolar latches”, which means they have 2 states (outputing high/low, coresponding to ~5v and ~0v for the usual 5v system), and they flip between those states whenever the magnetic flux density sensed goes above or below internal thresholds:


from the Allegro A1250LUA datasheet

Above is the state diagram for the A1250, a common hall sensor. For this sensor, Bop is approximately 5 gauss – when the field strength exceeds this amount, the output will pop low. It will not return high again until the field reverses directions and exceeds -5 gauss. These are “typical” specs – the maximum band is +/- 25 gauss, which leads to a hysteresis band of 50 gauss.

With the sensors being on the outside of the motor and trying to sense the magnets through a steel can, every little bit of field counts. I’m lucky that R/C motors are made so cheaply that the rotor is not thick enough to fully contain the magnetic flux. The field from the motors looks vaguely like this:

…by the way, here’s an actual magnetic field simulation made using the FEMM software. You don’t see as much leakage in that one because the rotor is thick enough that pretty much the entirety of the field is contained. (I’ve also wondered how much more torque per amp you can get just by wrapping another layer of steel around the typical R/C motor, but that is an aside.)

The message that my sketch is meaning to convey is that at each point where you can measure the flux density (“strength of the field”), there is a tangential component and a radial component. The only thing that the Hall sensors detect is this radial component. This is where the hysteresis of the sensor can introduce timing errors.

Say that the radial field at the location of the arrow head is -25 Gauss, and is where the hall sensor switches states. For the sensor to switchback to its original state, it has to travel all the way across to where the radial field is +25G. In the above drawing, the mechanical distance (in degrees) between these 2 points is about 5-6 degrees. Not a big deal, right? But because most R/C motors are 14-pole “LRK” motors, the equivalent electrical degrees of that error is 5 * 7 (pole pairs) = 35 electrical degrees.

What that means is if I moved the sensors to trigger exactly on-time for one direction of rotation, it will be 35 degrees too late for the other direction. It’s pretty much ‘magnetic backlash’ to use a mechanical analogy. The ideal hysteresis-free sensor would switch exactly as it crosses the midplane between the two magnets with no delay. This can actually be a pretty big deal. A few degrees of timing error might just result in a lowered top speed and more current draw, but 35+ degrees might make your motor just straight up not reverse or have issues starting. 25 gauss is also really goodfor a Hall sensor latch – the ATS177 type I typically use actually have triggering thresholds of +/- 70G with some specifications having +/- 100G.

Shane did a very thorough investigation using a strobe tachometer of this phenomenon when it is applied to externally-sensed r/c outrunners in traction applications if you want to see how bad it can get on video.

how fix

One idea I thought of which would fix this would be to use hysteresis-less sensors.

Yeah.

Smart, huh? It’s what I went to MIT for, duh.

Well, the real explanation is slightly more complicated. Instead of using digital latches with their own hysteresis bands built-in, I would instead use analog (linear) hall sensors which output a varying voltage, typically 1 to 2 millivolts per gauss centered around half of the logic supply voltage (VCC/2).

However, just using them by themselves is not necessarily enough, because the logic that is reading the sensors usually has digital hysteresis by itself. For instance, the ATMega328 microcontroller’s input thresholds are 0.3*Vcc for low-going transitions and 0.6*Vcc for high-going. Not only is this not symmetric around 0.5*Vcc, but the band of 0.3 Vccs is 1.5 entire volts for a 5v system. If my linear hall sensors are 1mv/gauss, then that’s a hysteresis band of 1500 gauss. That ruins the point. I’m fairly certain that is actually way worse than using digital sensors. The asymmetry would mean that the Halls are not “on” or “off” equally, so the motor will prefer several states over others.

The solution I settled upon was to pass the linear hall sensor outputs through a fast, low-hysteresis comparator which checks the input voltage against a Vcc/2 resistor divider made of 0.1% ultra-tight tolerance resistors.

The comparator I chose for this task was the MAX9034 4-channel comparator, which was selected mostly by stochastic Digikey-ing (I have no particular attachment to this part number), and the hall sensors are Honeywell SS49E linear types. I had to go with a TSSOP chip, which will be bitchy to solder but was the only size that could really fit on the board.

The idea is that the comparator’s internal hysteresis is the only discontinuous element in the system, and that is almost vanishingly small. For the 9034, it is 4 millivolts, which translates to 4 gauss or less depending on the actual sensitivity of the Halls. The output of the comparator will be a fast-edged square wave feeding into the logic and skipping over its hysteresis band.

Here’s a picture of the board. I’m starting with 63mm motors since tinykart is an excellent plant for validation of the idea, and it uses the 63mm class outrunners formerly retailed by Hobbyking (now retailed by Leader Hobby).

It uses the same adjustable slots as the current generation of sensor boards so it should be a drop-in replacement in most cases.

I sent the board to MyroPCB about 2 weeks ago and had them panelize it:

And last week, they arrived!

Sweet, that means I can get to testing them, right?

I wish. Alot of this design and thinking was performed several months ago, and I ordered all the parts then, but not the boards. Problem is, I’ve since lost the Hall sensors! That was like 50 bucks in sensors…

I had to reorder them, but they did’t get here in time for the weekend. Ideally I’ll have the sensors in-hand tomrrow and can get to populating these boards. Then, testing to see if I am bullshitting myself.

other sensor boards

Besides these super special ones, I’ve also updated the designs of the 2.00EV boards. Not only am I going to make another run of these boards for next semester’s 2.00gokart where everybody will need them, but I fully intend on selling these on this site too. These are just plain digital latch boards, suitable for one directional movement. I’m also working on some mounting rings that fit the 59-80mm common outrunners for the complete value-added package.

f

 

The Triumphant Return of the Ragebridge: RageBridgeTwo

chorl

First, in accordance with the Charles Z. Guan Strategic Stuff Reduction Act of 2012, I’ve opened up a listing page of things I’m clearing out. Let me know if you want a thing. As I continue mining through my multiple mid-dens, more things may be added.

Now, onto the real content.

With the version 1 of the Ragebridge controllers having proven themselves at Robot Battles 2012 (after being the stuff of nightmares prior to that), and now that I understand what I’ve been doing wrong the whole time, it’s time for me to make an update which addresses the little shortcomings and fine details that were lacking in version 1. Recall the major hardware issues:

  1. Gate driver bypass caps were placed incorrectly, thus causing the gate drive power traces to have higher impedance to pulsed currents, with the resultant voltage spikes some times causing instability.
  2. Gate drive current return path was long and loopy: through the main power ground, and back in through the narrow logic regulator trace.
  3. Non-rotationally symmetric current sensor placement meant there was a major current bottleneck for one side
  4. Long power traces in general made for more bottlenecks
  5. The logic regulator inductor was (and apparently has been in many of my past designs) too small which caused the regulator to become unstable.

Luckily they were all solvable with component-level hacks (no more wire cutting and trace jumping!) so I can easily roll them up into an upgrade. Next, there were a few little gripes I wanted to resolve which didn’t impact the function of the board so much as was just annoying.

  1. None of the 3-pin digital input headers had the correct orientation for standard servo cables. I had intended to make 2 of them “potentiometer” style inputs, reading [5v] [Signal] [Ground] from one side to the other, and 2 of them R/C style inputs reading [Signal][5V][Ground], but messed up (and didn’t check), making the latter two [Signal][Ground][5V] instead. Basically this meant I had to hard-solder pigtails to the board so as to not plug in a normal servo cable and explode everything.
  2. The board is 2.2″ wide, which precluded it from fitting vertically (space-saving) in a 2″ frame, fairly standard for the 12 and 30lb classes. Only Clocker could fit it vertically in a custom little “rack”.
  3. It still uses that ATMEGA328 breakout board called the Arduino Nano. Not only is that kind of unnecesary, but the thru-hole package of the Nano actually makes routing pretty difficult since I have to fit signal traces between the thru-holes.
  4. I could only read 60A peak with the “single bypassed” ACS714 current sensors. With a little help from a fan (and maybe 4oz copper) I don’t doubt this board can flow more than that.

With these issues in mind, the goals for RB2 were clear:

  • Squish the design down to 2″ wide maximum
  • Rebuild the schematic from scratch to eliminate possible cross-generation schematic copying problems, which is what led to my incorrectly placed buscaps and ignorance of the logic regulator problem for so long.
  • Widen all the power traces and make sure all the gate drive traces have their own low-impedance path back to the driver chip and the power supply
  • Make the current sensors “double bypassable” to read peak currents up to 90A and make the layout as much of a 180 degree rotation as possible to preserve the bottleneck-free layout.
  • Board-mount the ATmega328! It’s still going to be Arduino, but I’ll forego the extra $12-20 breakout board.

It’s also important to note that RB has been printed on 1oz copper using the cheap prototyping service from MyroPCB. Even moving to 2oz copper would help alot with the current bottlenecks. Ideally I’d try to get it fabbed from 4oz copper.

Here’s the first shot of the layout progress. The master layout remains essentially the same. The board is now 4.5″ long and 2″ wide:

Beginning the layout is usually the hardest for me because that’s when the problem is totally unconstrained. I usually jiggle the major power components around in a rough grid snap (to ensure reproduceability) first. After that, I usually work from the “outside in” – gate drive components get arranged next, then the logic fills in the loose volume. I’ve gotten all the way to the “so, where to put the logic components?” stage here, so they remain in a formless blob in the center. It does show that I have space to play with however.

For the ATMEGA chip, I knocked a footprint and schematic symbol from the Sparkfun Electronics library. The one difference is that I added to the ATMEGA’s pin names in the schematic their equivalent analog and digital pins when the chip is brainwashed into Arduino mode. This just makes mapping from old to new schematic easier. I’ll update my EAGLE Library under the Useful Stuff / references page with this device.

There’s one more little detail about this board, but I’ll keep going for now. Let’s see if you can figure it out.

I’ve now added some of the major power planes and a logic ground plane. One of the things I’ve never done in the past is a logic ground plane, for some reason. It makes trying to tie all the ground nets together so easy! And a wide plane is lower impedance than narrow 10 mil wide traces hopping through 9 vias. I just had to be careful with the bottom-layer traces not cutting the plane in half or totally isolating a ground pad, which would defeat the purpose.

The big SMT 6 pin header has disappeared – it just took up too much floorspace.

A few hours later, the first round of iteration is complete. Really I should send the board out to manufacture right now…

…but I decided to wait a while this time. I didn’t do anything with the board for about 2 days, upon which I returned and started optimizing. The differences are very minor, mostly in component fine placement. There’s now plenty of room to attach standoffs to the board at its mounting holes. Furthermore, the mounting holes have been made into a slightly smaller square that is the same size as a 52mm DC fan’s mounting pattern. This means I should be able to get some active…. microcontroller? cooling going on with this board.

Even though I say microcontroller (the ATMega would be receiving the best airflow in the house), almost any airflow at all would make the FETs’ thermal resistance to ambient air lower and hence increase the current rating of the board. Maybe there’s enough space to fiddle with putting a temp sensor on it…

Alright, one more layer of optimization. I’ve organized the top row of parts, which are 15v logic supply regulator related, into a better arrangement. Previously, they were extremely close together (soldering would have been difficult and hell if an IR lamp can reflow the pads in the dark alleyways of my L/C skyscrapers). I’ve space them apart alot more, but that involved moving the regulator chip itself down. This in turn meant I needed to shift some passives around. Whatever, the component density has only increased as a result.

(Now, none of this would be a problem if I would only use 0603 or 0402 SMT parts, but I’m too heavily invested in 1206 and fuck people with good eyesight.)

That is pretty much the final version of the board as it stands. With each of these revisions, I was sending it off to Advanced Circuit’s FreeDFM service in order to double check that it can actually be printed. I also ran a full Design Rule Check every time I changed a bunch of things between iterations. This was mostly me clicking “Approve” to all my vias-in-pads and insufficient solder mask clearances, but it did catch a few too-close-spacings.

Speaking of vias-in-pads, apparently they’re generally discouraged in automated PCB assembly, so in the interest of possibly getting this board professionally stuffed, I went through all the difficult tunneling in the gate drive and tried to un-pad my vias, at least connecting them to the pad with a short neck trace if possible.

This is a picture showing the VSS (everything-power-return) highlighted. Each gate drive now has its own private access to VSS at the location of the FET. This makes sure that 8 switching current pulses aren’t trying to octo-penetrate the thin logic regulator supply trace at the right in order to get back to the regulator. While it “worked” in RB version 1, it’s definitely not optimal. This general idea of not having signals mix with power is called Kelvin connecting, after the dude who first used it to sense very small voltages.

Highlighted now is the 15 volt distribution bus, which I’ve made strictly tree shaped this time. I try to adhere to the rule that any time power is being dealt with, closed loops are a bad idea.

The 5 volt bus follows a similar tree structure.

The large ground plane, which is the logic return path, is actually not a closed loop either. On the lower left corner, it was purposefully broken by judicious gate drive trace routing. It’s topologically a big mirrored C shape with the ends of the C right at the regulator output.

I’m pulling out all the stops I know to make sure that this board works on the first try. Ideally I’d get this sent to manufacturing through Myro tomorrow and have it ship right before Maker Faire 2012 in New York. This will probably be my on and off diddling project for most of the semester as a result. If it works out, I am now much less opposed to letting a few out “in the wild” for testing and subsequently investigating how much it would cost for Myro to just assemble all the parts, minus cables and through-hole pin headers and capacitors, for me.

To make it easier for them, I’ve placedallthe parts on the top side. That was the surprise – go back and look at the first few pictures now. Notice how all the parts are red (top layer)?

Now, all top side parts doesn’t instantly mean manufacturable. Once I build a few I might find where problem areas are. Say, anyone know if you can tell pick-and-place machines to put on the short and flat parts first?