Back in the day (1991) we had this coffee pot with a low-res monochrome camera connected to
a frame-grabber and a server in the Computer Laboratory, then in the Austin Building in central
Cambridge. We could type the command xcoffee on our workstations and decide whether to make
the arduous trip around to the coffee machine or lurk in our offices until some worthy colleague had
made a fresh pot of coffee. When the world-wide-web first became a thing we connected the frame grabber
to a web server and were surprised how many people around the world wanted to see a 'webcam' (the word wasn't
yet invented) and the rest is history (see Trojan Room coffee pot).
Well, here we are ~30 years later in a somewhat shinier modern Computer Laboratory out at West Cambridge, but the endemic coffee dependency remains a central theme, with our kitchen looking like this:
So the question is, should we have a new coffee sensor and if so what would it look like in 2019 vs 1991?
Build a sensor set that measures and transmits in real-time the coffee-making and consuming events happening on the Computer Lab Cambridge Coffee Pot.
This is actually monitoring the direct descendant of the original Cambridge Trojan Room coffee pot from 1991 to 2001.
The sensor will take weight measurements as in the sample data illustrated below which was collected using the first prototype.
The sensor will recognize events such as a fresh pot of coffee being placed, a cup being taken, and the pot becoming empty. The objective is that these events are communicated to the server with the minimum latency while the periodic reporting of the weight (e.g. once a minute or so) is primarily a 'watchdog' message confirming the coffee pot is working.
The pot weighs ~0.5 Kg, and when full holds 2Kg of coffee. Each cup taken is ~0.25Kg.
It was unexpected that the press on the plunger to pour a cup of coffee could result in combined downward weight >10Kg.
2019-10-09
Note in this first prototype I'm using a single hx711 to A/D convert both load cells, as the hx711 has two channels (A & B) which are not equally accurate but both good enough for our application. But when reading both A/D values on the Pi from channel A followed by channel B a ~half second latency was required between the two readings as the hx711 is clearly multiplexing some of its circuitry between the two channels.
2019-10-04
In this image you can see I moved the two load cells onto two separate hx711 A/D converters and used the A channel of each. This reduced the single read time of both sensors to a few milliseconds. The A/D sample rate defined in the hx711's is defaulted to 10Hz (the alternative is 80Hz) so there is no point in repeating readings within 100ms.
Here we hit an engineering problem. The weighing platform was reasonable stable 'fore and aft' i.e. along the axis the pot is placed on the sensor (as was the intention and the mountings for the two load cells were aligned this way). However, when the plunger is pressed the tilting forces from side to side were larger than expected and the weighing platform felt too fragile on that axis to survive long in production use. The considered solutions were some sliding pillar arrangement to keep the weighing platform from rocking, or the use of four load cells. We went with the latter.
To keep the height of the housing as small as practicable the electronics are placed in a sealed separate area adjacent to the load cells (as opposed to beneath them). This gives the overall housing the approximate plan dimensions of an A4 sheet of paper. The 'landscape' arrangement of this sensor made poor use of the space available in the kitchen so in the next version a 'portrait' arrangement was used.
2019-10-06
In addition to the four load cells, this prototype uses 4 x hx711 A/D converters to reduce latency in reading as with two.
2019-10-06
The 'portrait' arrangement conflicts with the simple solution of putting the LCD display to the left or right of the load area, as was planned with the 'landscape' design.
The solution was to use transparent acrylic for the weighing platform and mount the LCD under the front left corner of the weighing platform.
Embedded in the coffee pot sensor are two load cells that measure the weight of the pot.
In the data-sensing business, everybody, and I mean everybody, assumes the sensor design is finished after the first few lines of code are written that actually manages to read the parameter being measured and send the results. This is understandable as often they will have spent weeks or months just trying to get the measurement sensor to actually work reliably.
The effect of this is you end up with a 'weight sensor' that either has a built in period for repeatedly sending its reading, or it can be 'polled' by a server somewhere that periodically asks for its reading. Either way you end up with a sensor that doesn't really care about the thing it's measuring, and neither does the server, so long as the data flow adheres to the once-a-minute (or whatever) regular schedule.
Maybe someone comes along and asks for the data with less system-related latency. The sensor / system developer will always, and I mean always, respond with one of two answers: (1) you don't need the data more quickly, or (2) shall I send the data twice as 'fast' (i.e. every 30 seconds).
The truth is the required 'timeliness' (or acceptable latency) of the reading depends considerably on the state of the 'thing' being measured. E.g. A traffic speed sensor on a highway that sends the prevailing traffic speed once a minute is better than no sensor, but there is no good reason it would send the readings "72,75,71,69,0" at regular intervals. Hopefully that last reading of ZERO could be sent within a millisecond of being measured, rather the possibly waiting 59 seconds to send it as a regular update. And a design that simply sends the measured speeds every millisecond is pretty dumb.
This issue regarding the timeliness of sensor data is pervasive, particularly in urban and in-building sensor systems.
Our coffee pot will connect to our existing real-time Intelligent City Platform (which itself can process incoming events without introducing system-related polling latencies) and will
- send the weight periodically, say once per five minutes - this is best thought of as a 'watchdog' which happens to carry a useful payload.
- recognize the following events and transmit them to the platform as soon as they are recognized:
- POT_REMOVED - the pot appears not to be present
- POT_NEW - freshly made coffee seems to have appeared
- POT_POURED - a cup was poured
- POT_EMPTY - pot appears to contain no coffee
- COFFEE_GROUND - by also monitoring coffee grinding machine (with a microphone), it seems coffee has been ground.
This is an open question at the moment. The original coffee pot presented the data as a 128x128px monochrome
image and it was left to the viewer how to interpret it, combined with an implicit assumption the user would
issue the xcoffee command at the time they were interested in the coffee.
The new design can assume the permanent connection of web-based displays (in addition to the 'user request' model) and we will consider (i.e. measure) the latency in the system reporting a fresh pot of coffee.
It has been suggested that any self-respecting coffee machine in 2019 would have a Twitter account. [ref]
git clone https://github.com/ijl20/cambridge_coffee_pot
This repo includes working python libraries for:
- the hx711 D/A chip commonly used to connect load cells (code/hx711_ijl20/hx711.py)
- the st7735 LCD drive chip commonly used with inexpensive small LCD displays (code/st7735_ijl20/st7735.py)
The code directory contains a bunch of other libraries for the hx711 and the st7735 which were a reasonable source of clues but needed improvement.
Python support for http POST of sensor data:
pip install requests
For this one-off experimental sensor we used a Raspberry Pi 3B+, using the GPIO pins to connect the LCD display (via SPI) and the two load cell A/D converters (each needing +Vcc, GND and two data pins)
We interact with this via the st7735_ijl20 library code/st7735_ijl20/st7735.py
E.g. available Amazon UK
Two 5Kg load cells connected via two HX711 A/D converters.
We connect with the A/D converters via the code/hx711_ijl20/hx711.py library
E.g. available Amazon UK
[1] Heidi Howard, in conversation around the Coffee Pot, Cambridge Computer Lab, 2019-10-30