• Data Logging to Calibrate HMC5883L

    Josh Cole04/25/2021 at 18:43 0 comments

    I have entered this project into the data logging contest because of the work I did to analyze and calibrate the HCM5883L magnetometer. It wasn't the whole point of my project, but I think it's a fun look at how visualizations can help understand what's happening.

    The ultimate goal of my project is to create a two-wheeled robot which is capable of defeating gravity in a real-time battle against the forces of nature. One of the first major decisions to make: which accelerometer do I use? Having a few HMC5883L sensors lying around, I decided to create a small circuit to collect telemetry and see whether it would be a good candidate for my robot.

    RUBBISH

    As a sanity check, I wrote a small program to collect the raw sensor readings and then looked at the numbers as I rotated the device about all of its axis. Much to my surprise, the values were all over the place. Rotating along two axis (what I imagined were X and Y) caused all three numbers to change. That's not a good sign, but looking at raw numbers can be a hard thing for humans to make sense of. It's time to open a Jupyter Notebook and plot the data. 

    IDENTIFYING WHICH AXIS REPRESENT DIRECTIONAL HEADING

    The first actual experiment began by mounting my sensor to a large base and then spinning that base like a turn-table. The idea is that if I can isolate the X/Y coordinates and produce a heading, the remaining plane will indicate rotation. With this information, I can apply virtual offsets to the readings to "normalize" the plane.

    Collecting this data, I was able to plot it in matplotlib. I used a scatterplot to map two axis at a time. Presumably if I find the dimensions which represent "directional heading" that should appear as a perfect circle. 

    And in fact, I was able to find just that. Here are the visualizations:

    X and Z axis are represented by this lovely oval. This would indicate that X and Z are representative of magnetic heading.

    Y and Z are represented by these weird lines. I'm going to assume Y is rotation, and Z is one component of heading. So that means there's variance in the rotation based on which direction the sensor is facing. NOT GOOD! 

    And lastly, when plotting X and Y we can clearly see there is no correlation whatsoever.

    So we now know that X and Z represent directional heading. As we can see by the second graph, the rotation seems to change based on which direction the sensor is facing. An easy win now is to calibrate the rotational axis so it remains flat no matter what the heading is.


    CALIBRATING THE SENSOR

    Here's some python I came up with to demonstrate how to normalize the rotation value. In pseudo-code, this function loops over each point in the dataset and computes the heading. It then stores a dictionary of values where the key is the heading and the value is the inverse of the actual sensor reading.

    If you multiply that value, you will effectively zero out the reading.

    def generate_calibration(cal_list_x, cal_list_y, cal_list_z):
        y_factor = {}
        y_factor_count = {}
        for idx in range(0, len(cal_list_y)):
            px = cal_list_x[idx]
            py = cal_list_y[idx]
            pz = cal_list_z[idx]
            h = round(heading(px,py,pz))
    
            if round(py) == 0:
                factor = 1.0
            else:
                factor = (1 / py)
    
            if h in y_factor:
                y_factor[h] += factor
                y_factor_count[h] += 1.0
            else:
                y_factor[h] = factor
                y_factor_count[h] = 1.0
    
        for k in y_factor.keys():
            y_factor[k] /= y_factor_count[k]
        
        return y_factor
    

    To use the output of this function, you simply ask the robot "what direction am I facing" and then you will get a multiplier that you can apply to the current rotation reading. Any variance detected thereafter should be legitimate.


    TESTING OUT THE CALIBRATION

    To test out the calibration, it is imperative to isolate the variables as much as possible. I 3D printed a jig which has a 10-degree backwards tilt and then mounted my...

    Read more »

  • Thoughts on Calibrating the Compass

    Josh Cole04/10/2021 at 16:49 0 comments

    I've been giving a lot of thought about how to calibrate the magnetic compass (HMC5883L). There are a few interesting qualities that I've discovered. First - the raw readings, even after self-test, are nonsensical. Technically I'm able to get linearly-scaled values. But they are in unexpected ranges. More importantly however, is that the motors themselves cause significant interference with the operational quality of the HMC5883L readings.

    All that to say, I'm thinking of ways I can calibrate this thing. My current plan is very naiive, but I'm going to do some tests and see if it's feasible. At its most basic level, I'd like to discover an offset - and maybe a scale factor - in order to constrain the readings between 0 and 1. For this project, I really only care about the Z axis, as that will represent tilt. Although it would be nice to calibrate all axis so I have more opportunity for advanced functionality in the future.

    Calibrating the baseline

    So to calibrate, I'm thinking I'll mount the sensor roughly near its final resting place (so it can experience interference from the motors) and then I simply need to make sure bbot stays perfectly level, while I rotate the whole machine 360 degrees collecting data from the sensors all the while. Taking the maximum z value as an offset seems like a reasonable way to collect the baseline. Since it'll have been perfectly level, I know that any reading I get will just be potential noise.

    Is that going to be enough? I'm not actually certain. My suspicion is that I'll find it works for some directions, and fails for others. If the values are non-linear, then my approach will need to become infinitely more interesting. I have thoughts on this, but I'll wait until I can prove whether or not they're relevent.

    Calibrating the range

    Once the baseline is collected, that means the sensor should read some static value when it's perfectly level. Next, I need to make sure that tilting the machine produces similarly consistent values based on the degree of rotation. Ultimately I intend to calculate how many steps per degree are required to actuate the motor in order to return to a level state. In this way I won't be over-correcting.

    To calibrate the rotational range, I'm thinking I'll take bbot and rotate forward 90 degrees, then backwards 90 degrees. Repeat this a few times and collecting data all the while. Given these readings I can take the minimum and maximum and derive some kind of range which can then be used to constrain the readings between.

    Tilt = ((reading - min) / min)

    Or something like that. I haven't had enough coffee yet for math.

    Final thoughts

    So that's the plan. Calibrating the baseline, and then calibrating the tilt. It'll require a lot of tests to prove this will work, but I'm hopeful that I can find some way to calibrate the readings accordingly. I'll write more updates as I work through the process.

  • Circuitry

    Josh Cole04/08/2021 at 03:32 0 comments

    Circuit board design is complete!

    The first version of the circuit board is complete!!!! This is my first time using KiCAD and I suspect there will be a lot of random issues, especially with things like hole size. So I'm going to get this fabricated and then we'll see what happens.

    Manually routing wires is really hard!! But at the same time, it's a little zen. Easy to get lost in the routine.

    The good news however is that I can use the general dimensions to finish my 3d printing design. Knowing there will be a 3M mounting hole, and the approximate dimensions, helps a lot.

    In the meantime, the wheels arrive tomorrow! It'll be fun to come up with a strategy to mount them to my stepper motors.

  • CAD work

    Josh Cole04/06/2021 at 03:20 0 comments

    Day 2

    Now that I've got a good handle on the electronics, I switched to SolidWorks for a bit of CAD prototyping. Bbot's chassis is nearly complete. The current design features a NEMA-17 mount, top deck, battery mount, and holes for a wire harness. The only thing I've got to figure out now is where to put the electronics. And for that matter, what form factor the electronics will be in.

    So next I'm going to work on planning out the circuitry which should help inform the last component of the structure.

  • Component check!

    Josh Cole04/05/2021 at 01:22 0 comments

    Day 1

    It's been about a day since I started working on this project, and I've made what I consider tremendous progress. I've written more than half of the code and wired up most of the components - this includes shift-register manipulation, a clean stepper motor interface, and some well documented helper methods. I still need to get the accelerometer hooked up and coded, but that should be relatively easy considering it's i2c. Then it'll be all about the algorithm. I'm happy to have the boilerplate out of the way.

    Next up I'm going to spend some cycles working on the chassis. All code has been checked in to my project github repo, and the CAD files will similarly end up there. 

    Looking forward to having something pretty to show before too long!