Raspberry Pi is a very affordable little device, which you can use for various projects. It has outstanding Linux support, and it’s official Linux distribution, Raspbian, runs on all supported Raspberry Pis.
If you develop a project on Raspberry Pi 3 B+ and then you insert your microSD card into Raspberry Pi Zero, your code may still just work. One of the great benefits is a 40-pin GPIO header.
Along with libraries like gpiozero, even beginners can use GPIO without hassle. But Raspberry Pi is not a microcontroller, and Raspbian is not a real-time OS. If you need to perform time-critical tasks, you may get into trouble as we did.
In this blog post, we’ll dive into the development of a stopwatch for firefighter competitions and discuss some of the issues we were facing. You can find the full source code here.
We called our project as “Firefighter Stopwatch.” If you have no clue what’s going on in a firefighter competition, take a look at this video.
There are four checkpoints. When firefighters reach checkpoint 4, the stopwatch will start measuring time. After they reach checkpoint 3, the application will measure split time. Points 2 and 1 work the same, except the checkpoint which the firefighters reach last will stop the clock.
We don’t only measure time. But each time a checkpoint is reached, we measure water flow and pressure inside a hose as well as the engine’s RPM. We also added support for manual data measurement triggered by a dedicated button, and we log the results into a CSV file.
We present all the results during the competition on an external monitor. For this project, we decided to go with Raspberry Pi. It’s affordable; it has a small form factor and has enough performance to handle measuring and displaying data at the same time.
We decided to implement this project in Python 3. This language is widely used and is excellent for fast prototyping. As a GUI library, we choose Tkinter, built on top of a lightweight Tk library. Final UI looks like this:
If you’re not familiar with this library, you create the whole UI via code. This library provides various widgets, like Frame, Label, or Button. You can change the appearance of the widgets with custom styles and position them inside a window using the geometry manager. The most commonly used manager is called a grid. You need to place each widget into a table with columns and rows. The largest widget placed inside the column or row determines the dimensions of the column or row.
And there’s our first problem. If you look at the UI, you can see that when you run our app, there is no data. Widgets take less space, and the window is smaller. As soon as you start measuring values, the window size will change rapidly. Take a look at a screenshot below.
You can see the window collapsing on the first run. And once you start measuring data, it is resized. Here it’s not that much annoying, but here we modeled an unrealistic scenario where all measured values are the same. As a result, the geometry manager resizes our window only once.
But let’s imagine a real-world scenario where captured data is different and in different ranges. When you stop the engine, you’ll measure zero RPM, but this value can climb up to 100,000 RPM very quickly. And then go below 10,000 RPM. In such a case, the geometry manager will resize the window several times. It’s not very pleasant, and we had to deal with it.
Our first approach was to reserve some extra space by putting empty whitespace characters as a placeholder text inside the widget. It didn’t solve our problem, but it was a move in the right direction. Ultimately, we decided that we need to specify the allowed range of values for each measured component. For instance, we’ll enable RPM ranging from 0 to 99,999. In a real-world scenario, RPM won’t reach the maximum value, but now we can ask widget to reserve enough space to fit all input data under all circumstances.
1label = ttk.Label(content_frame, style='Customized.Main.TLabel',2 padding=(30, 10), width=10, anchor='center')3label.grid(column=1, row=9)4label.grid_remove()
You can specify widget size using
width. Please refer to the documentation for more information. You can center the text inside the widget with the
anchor attribute. Assuming the input is smaller than the maximum value. Now the UI is more pleasant.
Another thing we needed to deal with was changing the widget’s visibility. During manual measurement, the UI shows the time of capturing the data and an
M symbol next to it. Then after 2 seconds, this information is hidden. The geometry manager comes to the rescue!
1label = ttk.Label(content_frame, style='Customized.Main.TLabel',2 padding=(30, 10), width=10, anchor='center')34## Position and display widget on the screen5label.grid(column=1, row=9)67# Make widget invisible8label.grid_remove()910# And make it visible again11label.grid()
Note that when we want to make the widget visible again, we don’t need to give a grid position anymore. The grid manager remembers the last place of the widget and will show it up instantly.
Our application consists of several classes, as you can see in a diagram below.
Stopwatch serves to control the clock on user input and to trigger the measurement. The
FlowMeter continuously measures water flow.
PressureTransducer measures pressure from two connected sensors and the
RpmMeter measures the engine’s RPM. All these classes post messages into the MainApp component, which handles an event loop and does all the processing on its UI thread. If you’d try to update the UI widget from a background thread, your application would most likely crash. Android works very similarly.
Handling GPIO inputs with gpiozero
Gpiozero is a small library, which abstracts operations over GPIO. People built it originally on top of RPi.GPIO, but over time they added support for other libraries. It allows, among other things, to use GPIO remotely from your PC as if it had its GPIO on board.
To process GPIO inputs, we used the Button component. This component creates its background thread just to process GPIO input. It is continuously polling for input changes, and it will call your function when such change occurs. We use the Button component not just for handling button presses, but for any signal input. Remember, the button is just an abstraction. And the library configures the input pin as needed.
The following code shows how to attach an input pin to the button and register your function as a callback:
1_STOPWATCH_TRIGGER_PIN = 2623start_button = Button(self._STOPWATCH_TRIGGER_PiN, pull_up=True, bounce_time=0.1)4start_button.when_pressed = lambda: self._start_watch()56def _start_watch(self):7 # Code to run8 pass
You need to connect your callback into
when_pressed property. The button component expects a function with a specific signature. Ideally, you should reference a function without input parameter, or you can pass a function, which assumes one input parameter. When called, it will have a proper reference to the Button object.
We use buttons inside classes. Their functions have
self as their input parameter. To make it work, we use an anonymous lambda function, which calls the code we need.
With this approach, we could build our components from the ground up. The stopwatch component directly handles input signals for controlling the clock; RpmMeter gets triggered each time it receives a pulse from the engine sensor, etc.
Gpiozero supports remote GPIO. We won’t go into details. Let us assure you that it is very convenient, and working with your Python code on your PC instead of your Raspberry Pi is a feature you don’t want to lose once you try it. In some sense, it resembles working with Arduino boards on a PC. Albeit Raspberry Pi is powerful enough for many tasks, developing your code on a much powerful PC saves time.
But when you try to run scripts written for Raspberry Pi on a PC, some libraries don’t expect such an environment and will raise an exception. Therefore, we handled problematic code with a “try…except” block and turned such features off.
One example is I2C communication using CircuitPython’s Busio library. Because this library won’t run on a PC, we won’t get data from the pressure sensor over I2C, but everything else works just fine.
This approach allowed us to debug and resolve most of the issues on a PC and spend the only necessary time developing on a Raspberry Pi directly.
CircuitPython is Adafruit’s flavor of MicroPython. It’s a version of Python for microcontrollers. Previously when you bought a sensor from Adafruit, they provided libraries and guide for the full-featured Python version as well as for MicroPython. This approach is deprecated as it’s worse for maintaining the code. What Adafruit recommends instead is to use its compatibility layer, called Blinka. Through this library, you can use their code written for CircuitPython, and this is an approach we chose as well.
When an event occurs, i.e., clock trigger is pressed, a component, in this case, StopWatch, processes this event and notifies the MainApp by calling
MainApp#post_on_ui_thread(event). MainApp runs an event loop, which processes all events and refreshes the UI when needed. Just a reminder, UI update needs to be handled solely by UI thread. Internally we use synchronized Queue class.
1# Queue for UI thread to update components2self._thread_queue = queue.Queue()
You can put the data into the queue like this:
1def post_on_ui_thread(self, value):2 self._thread_queue.put(value)
And you can read the data during UI refresh like this:
1def _update_ui(self):2 try:3 event = self._thread_queue.get(False)45 # Events without data6 if type(event) == str:7 if event == StopWatch.STOPWATCH_RESET:8 # Reset clock9 pass10 if event == StopWatch.MANUAL_MEASURE_ENDED:11 # Do some work12 pass1314 # Events with data as dicts (key = value)15 elif type(event) == dict:16 checkpoint = None1718 for eventKey, eventValue in event.items():19 if eventKey == StopWatch.SPLIT_TIME_MEASURED:20 # Display split time on a screen21 pass2223 except queue.Empty:24 pass2526 # This line will make sure that update function gets called27 # periodically to pick up events from queue. SCREEN_REFRESH_MS allows28 # us to precisely control, how often this function gets called.29 self._parent.after(self._SCREEN_REFRESH_MS, self._update_ui)
Processing data from sensors
Sensors attached to Raspberry Pi don’t provide values that could be used directly as-is. Instead, we need to compute what we need. As samples arrive, we store them inside of a circular buffer. For that purpose, we use the deque class from the
For instance, if you want to store the last ten samples as they arrive, initialize deque as follows:
1self._samples = deque(maxlen=10)
When a new sample comes, it is appended to the end of the buffer, effectively dropping the oldest sample.
You work with this buffer as with a standard list, which has ten elements. Neat!
Computing RPM from pulses
Our sensor is located right inside the engine. It generates a pulse each time an induction coil throws a spark. On the GPIO pin are received rapid pulses. The faster those pulses are, the higher the RPM is. We capture time (in milliseconds) when each sample arrived. From there, we can compute time delta between samples and subsequently calculate frequency. And finally, we can derive RPM from the rate.
1freq = 1 / ((self._samples[-1] - self._samples) / self._MAX_QUEUE_LENGTH) / self._k_multiplier2 rpm = int(freq * 60)
At the moment, we compute the time delta between 10 samples. Perfect sampling precision on Raspberry Pi is not guaranteed, and this approach will help us mitigate our problem a little bit.
Computing water flow from pulses
It is somewhat similar to the previous case in that we also get pulses on input. To compute flow in liters per second, we just need a different equation.
1freq = 1 / ((self._samples[-1] - self._samples) / self._MAX_QUEUE_LENGTH)2 lpm = int(self._k * (freq + self._q))
Getting pressure data is trickier. Pressure can rise and decline very rapidly. Therefore, we need to compute its moving average.
Implementation for this module differs. The sensor we use communicates through the I2C bus. We don’t use the Button component with its thread. Instead, we need to create our background thread, which will be polling the sensor for new samples periodically. This thread runs the following runnable code.
1def _update_sliding_avg_pressure_thread(self):2 if self._i2c_initialized:3 self._is_measuring = True4 self._voltage_1_samples.append(self._adc_channels.voltage)5 self._voltage_2_samples.append(self._adc_channels.voltage)6 self._is_measuring = False
When we later query for pressure, we compute its value as follows:
1def get_sliding_avg_pressure(self):2 # Sliding average is computed from _MAX_QUEUE_LENGTH samples3 avg_p1 = sum(self._voltage_1_samples) / self._avg_samples_no4 avg_p2 = sum(self._voltage_2_samples) / self._avg_samples_no5 return tuple(map(self._calculate_pressure_from_input_value, [avg_p1, avg_p2]))
When you look closely, you’ll see that in our formulas, we use various user-defined constants. These are defined in a configuration file loaded during application initialization. This configuration file gives us the flexibility we need. We’ve defined fallback default values as constants inside our code.
And now the fun part! Some of the less significant problems we’ve already mentioned. Those were already solved. Now we’ll briefly take a look into issues, which we mitigated only partially.
RPM fluctuates on a steady pulse stream
When we generate pulses with a specific frequency, computed RPM should be constant. In reality, it’s not the case. During testing, our RPM value fluctuated within +-100 range, meaning that if the engine runs at 10,000 RPM, you’ll see values ranging from 9,950 to 10,050 RPM. We ultimately concluded that for this type of project, such precision is acceptable. It doesn’t mean we’re happy with it.
Water flow fluctuates as well
An issue very similar to the one above. Only that water flow fluctuated in a smaller range.
Pressure change isn’t reflected instantly
The change of voltage didn’t reflect into pressure change immediately. We could probably fine-tune pressure computation more. Changing the size of a circular buffer could be effective mitigation. For our case, however, it’s acceptable.
Logging slows down data capture significantly
In our code, we use the logging module extensively. We’ve set its verbosity initially to show all messages with DEBUG priority and above. It proved to be a problem when logging occurs in threads performing data sampling. Ultimately forcing our logger to print only warning and error messages did speed things up.
When we look at the issues, we can see a pattern. Data sampling accuracy is affected because we don’t use hardware designed to work in real-time. Also, Raspbian doesn’t have a Linux kernel fine-tuned for real-time operation. All threads are governed by Linux kernel scheduler. You don’t have a guarantee that a background thread will run on a dedicated CPU core all the time, uninterrupted. It is the reason why we don’t receive the data samples captured in precise moments.
Sometimes the Raspberry doesn’t capture the samples or records their time of arrival with minor skew. When you work with relatively higher frequencies, it does matter. Such small differences cause RPM fluctuation, which is not negligible.
If we started working on a similar project like this one in the future, we would probably consider using a dedicated microcontroller just for getting all the samples at the exact time and computing the target values at its idle time.
Arduino microcontrollers seem like the right choice. They support interrupts; therefore, we can hook up a routine to process samples with the highest priority. And when no sample is processed, this microcontroller still has plenty of CPU time to do all those calculations. It doesn’t run any operating system, after all.
When the Arduino module collects the data, it could periodically send cumulated messages to Raspberry Pi, which will display content on a screen. Even after small hiccups, we still consider Raspberry Pi a perfect tool for handling such simple UIs as our own.
Having GPIO on a Raspberry Pi is a great thing. You can work on many projects without hitting issues as we did. The moral of the story is that if you need maximum precision, deal with it on hardware designed for such a purpose.
Have you ever worked on a project like this? Did you encounter any of the mentioned issues? And how did you deal with them? Please tell us in the comments below.