Updating some, but not all, old technology

Lightning strikes. It struck near me, and fried a component of my old X10 home-control system.

X10 is a power-line-based communications network. Data is sent (slowly) over the house power lines by injecting a 120 kHz signal into the 60 Hz AC signal just after the zero-crossing of that signal. One bit per cycle is sent by either adding or withholding a short pulse of the 120 kHz.

System Description

X10 devices are either plugged or wired into the house power lines. Each device has an address, not necessarily unique, which consists of a house code A-P and a unit code 1-16. The original design intent was to have each house in a neighborhood choose a unique house code, to prevent signals from the neighbors from controlling one’s own devices.

The data sent over the power lines corresponds to commands like “A03 AON”, which says to turn on any device with house code A and unit code 3.

    I had been using a TW523 interface which provides:

  • a logic-level zero-crossing synchronization signal
  • a logic-level output which indicates when a 120 kHz carrier is present on the power line
  • a logic-level input which causes the generation of the 120 kHz carrier.

It was controlled by an old Arduino which passed commands between the TW523 and a UDP socket using an Ethernet shield.

A Perl script on another computer provided the intelligence – receiving messages from the X10 system and an auxiliary RF control box, from my MQTT message broker, and from a Perl scheduler.

The Event

A recent thunderstorm provided my next-door neighbor some excitement when a bolt hit his chimney. It provide me with something to do, too…

It fried my TW523.

I subsequently found that the devices aren’t manufactured anymore. Fortunately, Jeff Volp produces a plug-compatible replacement, the XTB-523. I’m not yet using its extended capabilities, but am enjoying the higher reliability produced by its higher-power signals.

The Project

However, being a geek, I had to make other improvements to my system.

I didn’t like the power consumption of the old Arduino and its Ethernet shield – the Arduino’s voltage regulator was always quite hot, and I was afraid it would cook itself. I also didn’t like the solderless breadboard I had cobbled together 20 years ago as an electronics newbie – it was a mess.

So I grabbed a spare Adafruit Huzzah ESP8266 breakout board and went to work on the software. [ I will upload that to my Github account once I figure out why it’s telling me hundreds of my files have changed and I should do a pull but it won’t let me anyways and … wah, wah, wah ]

I also needed to create a better hardware setup – I wanted to plug the Huzzah into a board with its own voltage regulation, level shifting, and an RJ11 jack for the phone wire to connect it to the XTB-523. I entered the schematic into Eagle and laid out the components as if I were going to produce a PCB. I do this even when I’m going to hand-wire a circuit board, just to the point of laying out the components but not laying out the traces. I generally also create and record a color code for the wires, which will help me in future debugging.

This time around I also remembered to add some test points onto the board. I have a 4-pin header for the power (Vcc, 5V, 3.3V, and GND) and one for the 4 connections to the XTB-523.

After some major troubles with the software, I got it all to work. The main software trouble was a bug in the X10 library. I was getting an ESP8266 fatal exception when I enabled receiving. As I tried to create bread crumbs ( Serial.print statements ) and follow them back, I found that the problem arose in an object method that was called from an interrupt service routine (ISR). Now, it’s a really bad idea to do anything in an ISR that takes any significant time at all, because other interrupts want to be serviced, too, and they can’t while one ISR is in control. Print statements take a really long time ( from the perspective of a computer ), so trying to put them in confounded my search for enlightenment. Eventually I got frustrated enough to put some print statements outside the ISR and learned that the object’s .cpp file had to define an object pointer ( so the ISR could call an object method ), but that it wasn’t initialized to point to anything. When you try to dereference a null pointer, bad things happen. Once I found that, I was past the hump.

I almost said, “Once I found that, I was home free.” But I wasn’t. I still had (gasp!) bugs in my own code. They were, however, readily found and fixed.

Now everything works. <contented smile>


Experiment with CircuitPython, Jupyter, and WordPress

I’m trying to embed a Jupyter notebook page as a blog post. I saved it as html, and copied an pasted that html below. I don’t think that’s the best way to do it.

We are exploring the Adafruit Metro M4 Express

Poking around, wanting to see if we can access the hardware timers, we set ourselves a problem:
We want to see how long it takes to update the NeoPixel’s color. Let’s start by exploring the time library:

In [1]:
import time
print ( dir ( time ) )
['__name__', 'monotonic', 'sleep', 'struct_time', 'localtime', 'mktime', 'time']
In [2]:
print ( time.time() )

So it looks like time.time() returns seconds since some fixed epoch – probably midnight, January 1, 1970, which epoch is a standard amongst Unix installations — but I haven’t checked this…

Next we’ll set up the NeoPixel

In [3]:
import board
import neopixel

pixel = neopixel.NeoPixel ( board.NEOPIXEL, 1 )

And now we want to see how many updates we get in one second. We don’t expect to get a very consistent number, because we’re not being really accurate with our timing.

In [4]:
t = time.time()
n = 0
while time.time() == t:
  # do nothing until we get a clock change
t = time.time()
while time.time() == t:
  n += 1
  pixel[0] = [ 2, 2, 2 ]
print ( n )

A better way to time would be to extend the time interval:

In [5]:
durationOfTest_seconds = 5
secondsUsed = 0
numberOfIterationsCounted = 0

t_seconds = time.time()
while time.time() == t_seconds:
  # wait, doing nothing, until we get a clock change

t_seconds = time.time()
while secondsUsed < durationOfTest_seconds:
  pixel[0] = [ 2, 2, 2 ]
  numberOfIterationsCounted += 1
  if time.time() != t_seconds:
    secondsUsed += 1
    t_seconds += 1
print ( numberOfIterationsCounted / durationOfTest_seconds )

Better yet would be to keep running the test until the standard deviation dropped to less than 1. We’ll calculate the variance as v(x)=E(x2)E(x)2




or, because this is sample variance,



In [ ]:
import math

stdevUpperLimit = 20
stdev = stdevUpperLimit + 1

# statistical moments of inertia
n = 0    # number of events summarized
x = 0    # sum of values
x2 = 0   # sum of squares of values

while stdev > stdevUpperLimit:  # keep testing until criterion matched
  t_seconds = time.time()
  while time.time() == t_seconds:
    # wait, doing nothing, until we get a clock change
    pixel[0] = [ 0, 0, 0 ]
  numberOfIterationsCounted = 0
  t_seconds = time.time()
  while time.time() == t_seconds:
    pixel[0] = [ [ 5, 0, 0 ], [ 0, 5, 0 ], [ 0, 0, 5 ] ] [ n % 3 ]
    numberOfIterationsCounted += 1
  # completed 1 test event; summarize statistics
  n += 1
  x += numberOfIterationsCounted
  x2 += numberOfIterationsCounted * numberOfIterationsCounted
  # calculate the mean and standard deviation of all the tests so far
  if n > 0:
    mean = x / n
    if n > 1:
      # variance = e(x2) - e(x)**2
      variance = ( ( x2 * n ) - ( x * x ) ) / ( n * ( n - 1 ) )
      stdev = math.sqrt ( variance )
      print ( "Mean: ", mean, "; Variance: ", variance, "; St Dev: ", stdev )
      # print ( "Insufficient tests for variance to be calculated" )
    print ( "No tests completed!" )

That one stinks — it runs forever if the stdevUpperLimit is set too low. So the next iteration needs to reverse everything – run n iterations of updating the NeoPixel and seeing how long that takes. We can measure elapsed time by using time.monotonic, which will give us a baseless value in ticks. Here we go:

In [6]:
numberOfIterationsCounted = 0
t_ticks = time.monotonic()
while numberOfIterationsCounted < 10000:
  pixel[0] = [ [ 5, 0, 0 ], [ 0, 5, 0 ], [ 0, 0, 5 ] ] [ numberOfIterationsCounted % 3 ]
  numberOfIterationsCounted += 1
t_final_ticks = time.monotonic()
print ( t_final_ticks - t_ticks )

It looks like time.monotonic() is a count in seconds, but with high precision. The number we get, 8.8 or so, from the above routine is consistent with about 1ms updates which were the rough guesses of the previous program. But based on the variation across runs in the result – from 7.8 to 8.9 or so – coupled with the visual impression of there being unexpected pauses in the flashing of the NeoPixel, indicates to me that there’s some overhead in CircuitPython that makes it unwise to count on millisecond-scale timing accuracy.

Thoughts on serial communications

Maybe these issues have been solved in the past. I don’t think there is necessarily a very good solution, though.

The problem is that serial communications is imperfect. Characters get messed up in transit between two devices. When they do, that in turn messes up carefully thought out protocols, like “Capture characters in a buffer until you see the linend; then process the buffer as a single message.”

It’s a small problem if a message character gets corrupted. The message probably won’t make sense; that can be handled with an error message and some recovery process.

But what if the linend itself gets corrupted? Then your program won’t see it as the conclusion of a message, so it will continue capturing characters into the buffer. Then what? Two messages are now corrupted. Still no big deal. But then you have an increased risk of buffer overflow. ( You are checking to avoid buffer overflow, right? ) How to handle that? Flush the buffer and hope the next message is better? But the next message will most likely be broken, ’cause you probably flushed the buffer halfway through it, discarding the first part.

Finally, what do you do if the source goes away? There might be a partial message in the buffer. Should you have a timeout, after which you conclude that the source has crashed, and you need to flush the buffer so the next message that arrives, when the source has recovered, will not also be clobbered?

Think of these possibilities. Writing embedded software is not necessarily as simple as we’d like. But if we consider all the possible ways something can go wrong, and deal with them carefully, we can make software work well and robustly.

Teensy, touch sensing, and ESP8266 – via MQTT and Raspberry Pi


I have a demo/development setup on a small breadboard which powers an ESP-01 ( the small, cheap, 8-pin breakout board for an ESP8266 chip ). In addition to being a programming board, it has 3v3 voltage regulation, a pot, a temperature sensor, a simple LED, and a WS2812 3-color LED. I have the installed ESP8266 running a program to subscribe to an MQTT broker and light the LEDs according to publications on a certain topic.

I’ve developed another small program which listens to a 115200 baud serial connection and accepts JSON messages which instruct it to connect to a WiFi, to publish values to an MQTT topic, or to subscribe to an MQTT topic. It is intended as a general-purpose glue service for an otherwise dumb microcontroller to communicate with remote telemetry.

I have also a Teensy 3.1, which supports capacitive touch sensing, interprets the values as control signals, and via the text-to-MQTT glue service publishes them, whence they will control the aformentioned LEDs. The program that does this is also on Github.

It all works, sort of. I’m considering it to be at the late-alpha stage of development.

Instrumenting and Analyzing My Solar Tracker

I’m an instrumentation and data collection geek. I love the challenge of developing ways to monitor the systems of my house and making sure they’re working properly. I’ve had a couple bad experiences where systems failed and not knowing it soon enough caused problems.

I have a solar tracker system, one which follows the sun when it’s up. It failed once with a hydraulic leak, and it was a couple days before I noticed it. Never again!

I’ve built and installed a monitoring system including its own solar panel, battery, 9 degree of freedom MEMS device normally used for inertial navigation of unmanned aerial vehicles, a microcontroller, and a LoRa radio with enough range to cover the distance to my house.

I’ll go into the details in another post. What I want to put into this post is the 3D plot of the first telemetry data.

The solar tracker is on a pillar out in my yard; the whole rotates around the vertical axis of the pillar (azimuth), and the panel tilts up and down (the motion is called nutation, the result is called elevation). The inertial navigation system (INS) device measures linear acceleration, angular acceleration, and magnetic field, all in 3D. My first task is to develop a coordinate system with one axis parallel to the axis of nutation of the solar tracker.

I have collected thousands of data points, taken at 1 minute intervals, and including the 3 linear accelerations. These mostly reflect the gravity vector. I’ve now plotted these in 3D; they should all lie in a plane, perpendicular to the axis of nutation. I can see this in the plot. (Look at the picture with 3D glasses, red filter for left eye, cyan filter for right eye.)

My next step is to calculate cross products for many pairs of vectors. The cross products will be perpendicular to each of the vectors, and so all the cross products should be collinear with each other, and also with the axis of nutation.

With 40,000 data points, there would be n(n-1) pairs, and hence 1.6 billion cross products. That’s a lot of computation. But I reckon that crossing vectors that are close together will yield poor results, and account for a lot of noise. So I’ll cross only those vectors that are farther apart, by some arbitrary distance.


Whew. I just read LadyAda’s post about using Doxygen to document a project, and setting GitHub up to not only host the files, but also to serve the documentation as html. I managed to do most of it!

I created a separate GitHub project according to her instructions, restructured my code as a library with an example (main) program, and pushed the code up to GitHub.

I read (some of) the Doxygen manual, and added a bunch of structured comments to my library code. I installed and ran Doxygen, fixed the documentation, many times, and got it all working.

I set up a Travis account and followed the instructions, but couldn’t get everything to work. It’s supposed to set up a virtual machine, configure it, compile your code, do some additional diagnostics, and then automatically generate the documentation using its own copy of Doxygen. Unfortunately, this particular code is written for the Teensy platform and won’t run on anything else. The Teensy setup for the Arduino IDE involves first installing the Arduino IDE, then running a Teensy installer on top of it, and I couldn’t figure out how to do all that with command-line scripts. (Probably I could, but didn’t want to take the time to try. Lazy!)

But I could simply upload the documents directly into the gh-docs branch of my project and ask GitHub to serve them. So I did.

They can be found here.

I like this. I think this will be how I document future open-source projects.

Doxygen seems pretty cool

I’ve just (belatedly) added Doxygen commentary to the firmware I’ve written for the OpenEffects Project hardware. It was half a day’s work, but it was reasonably painless to install Doxygen and Graphviz dot, even starting from scratch as I was.

I had blanched at the idea of manually diagramming the class hierarchy and arranging the comments so they made sense. Doxygen handles it well enough that I didn’t have to struggle (much).