Tag Archives: education

So You Want to Use PWM, Eh?

PWM Waveform Captured on an OscilloscopePulse-width modulation. It probably sounds a little confusing if you’re new to electronics. Kindof a word mashup, really. What do pulses, width, and modulation have to do with each other anyway? I remember first learning about PWM during my freshman year of college at RPI. I was in a pilot course called “Foundations of Engineering” under the excellent instruction of Professor Kevin Craig (whom I later worked for). I remember thinking later, “Hey, this PWM stuff is pretty clever!” So let’s take a look at PWM and see what we can learn. (If you’re already familiar with the basics of PWM, skip down a few paragraphs for more advanced topics and experiments!)

Say you’ve got a light-emitting diode (LED) and a battery. If you connect the two directly, the LED should produce a lot of light (assuming the voltage of the battery isn’t too high for the LED). But what if you wanted to reduce the amount of light that LED produces? Well, you could add a resistor in series with the LED to reduce the amount of current supplied by the battery. However, this won’t allow for easily adjustable brightness and may waste a bit of energy. That loss may not matter for a single LED, but what if you’re driving several high-power LEDs or light bulbs? This is where pulse-width modulation comes into play.

PWM Graph - 30% Duty CycleImagine you could connect and disconnect the LED and battery multiple times per second, causing the LED to flash or pulse (see graph above). If this ON-OFF cycle is fast enough, you won’t even notice the blinking. In fact, the LED will appear to be continuously lit, but reduced in brightness. In addition, its brightness will be proportional to the ratio of the on and off times. In other words, if the LED is connected for 30% of a pulse cycle, it will appear to be producing about 30% of its full brightness continuously, even though it’s actually turning completely on and off. So to adjust the brightness of the LED, all we need to do is adjust, or modulate, that ON-OFF ratio, also known as the pulse width – hence the name! The ratio between the on and off time is also commonly called the duty cycle.

Now in case you’re imagining yourself frantically flipping switches on and off, or tapping wires against battery terminals, you can stop. Just put a transistor in series with your LED! It can act as a switch which can be controlled by a microcontroller or some type of oscillator circuit (see links below).

Hobby Servo (Commanded via PWM)So what’s PWM good for, anyways? Well, dimming LEDs and other lights is just one of a number of applications (example). You’ll also find PWM used in motor controllers. You can make a very simple DC speed control using a PWM generator and a single transistor (examples – notice the extra diodes in use here to prevent damaging inductive spikes). In addition, PWM is very important for some types of power supplies; specifically the aptly-named “switched-mode” PSUs. This technique can also be used to create a digital to analog converter (DAC) by low-pass filtering the square wave. Finally, pulse-width modulation is sometimes used as a means of digital communication. For example, to command the position of a hobby servo.

Now you may be wondering why I’m writing about PWM all of a sudden. Well, there’s actually a point to all of this background information. By now, you’ve probably seen a car or two with these new-fangled LED tail lights. They’re pretty easy to spot since you can typically make out the individual LEDs within the whole tail light assembly:

Ford LED Tail Light Upgrade - Ain't that a Fancy Photo?
But have you ever noticed that on some cars (e.g. Cadillacs), these lights tend to flicker? You may not see it if you’re looking straight ahead, but if you quickly move your eyes from left to right, you may catch a glimpse of the flicker created by a low-frequency PWM controller. Now, call me strange, but I find this really annoying and distracting. Maybe I just have fast eyes or something, but I hate flicker. Back in the days of CRT monitors I could usually tell the difference between 60Hz and 70Hz refresh rates. But in the case of these tail lights, it sounds like there’s danger for people with photosensitive epilepsy. According to the Epilepsy Foundation, flashing lights in the 5 to 30Hz range can trigger seizures. Obviously, having a seizure while driving would not be a good thing for anyone.

By the way, if you’re ever trying to determine the frequency of a blinking light, just snap a couple pictures while moving your camera (or the light). The one catch is that you need to be able to specify a known shutter speed. Then you just have to count the blinks and divide by the shutter speed (in seconds) to find frequency. Here’s an example:

LED PWM Frequency Comparison

This method can also give you a pretty good indication of duty cycle – in this case it looks to be about 60%. Here’s a second shot I took while on the road one night. You can tell the streetlights are running on 60Hz AC (although they’re not LEDs so they never go completely dark during a cycle), while the green stoplight is likely getting DC:

Pulsing Streetlights

I’m thinking this long-exposure shot might also pass as modern art in some circles.

The Advanced Stuff

So what’s the deal with these awful low-frequency PWM tail lights? Well, one reason you might choose a lower frequency is to save on energy lost during switching. Both LEDs and the transistors used to drive them have parasitic capacitance. In other words, they store a very very small amount of energy (think nanojoules) each time you turn them on. This energy is consumed in addition to the steady-state power drawn by the LED to provide illumination. Furthermore, this stored energy is rapidly dissipated (and thus not recovered) each time the device turns off. Now if you’re turning an LED on and off fifty times per second, it’s probably no big deal. But what if you wanted to eliminate any possibility of flicker by driving the frequency up into the kilohertz range? Would this introduce substantial power loss? I was curious, so setup a simple experiment to find out.

Test Setup
The heart of this test circuit is fairly simple – two bright red LEDs (Model OVLBR4C7) along with 92Ω current-limiting resistors controlled by a BS170 MOSFET. To measure the power consumed by this circuit, I’ve taken a non-traditional approach. Because I was worried that the cheap ammeters I have available would be thrown off by varying PWM frequencies, I decided to measure power consumption based on the discharge time of a supercapacitor. And who doesn’t love supercaps, anyways?

The theory is pretty simple. The energy stored in a capacitor is equal to ½*C*V² (Joules). So all I had to do was charge up the cap, measure its voltage, let the circuit discharge it over a fixed period of time, then measure the final cap voltage. For my 2.5F capacitor (from NessCap), I chose ~60 seconds as my discharge period. Here’s a screenshot of the voltage logging application I used to collect my test data:

IOBoard Test Program
The white line in the graph above plots the capacitor voltage during discharge. The red line indicates the voltage measured across a phototransistor (L14C1). This was used to quantify the amount of light produced by the LEDs at each test point. To get a better measurement I covered the LEDs and phototransistor with an opaque plastic cup, then covered the whole setup with a shoebox and turned off the lights. I was trying to see if, for some reason, the intensity of the LEDs was non-linear with respect to duty cycle or was affected by PWM frequency. Unfortunately this data turned out to be rather boring, but I’ve still included it in my summary spreadsheet which you can download below.

Now before I go on, you’re probably wondering what sort of data acquisition hardware I’m using. Well I doubt you’ve heard of it as it hasn’t yet been commercially released. Right now it’s being called the RPI IOboard. It’s a pretty impressive piece of hardware with dual 12-bit, 1.5MSPS ADCs, dual 14-bit, 1.4MSPS DACs, and a host of digital I/O all powered by a 400Mhz Blackfin processor. For the past few years it’s been developed at RPI and tested at a number of schools across the country. However since the project’s lead professor, Don Millard, left RPI last year, I’m not exactly sure what will become of the board. The screenshot you see above is actually one of several executable VIs I developed as examples for use with the board. Further information on the hardware can be found here.

Test Setup Closeup
So back to the experiment at hand. For my first round of testing, I utilized the IOBoard to generate varying PWM signals for the MOSFET. Thus, the current required to drive the BS170 was not included in my first measurements. I varied both frequency and duty cycle for three pairs of LEDs: white (C513A-WSN), red (OVLBG4C7), and green (OVLBR4C7).

TABLE 1: Data for power consumption tests without gate-drive losses:

Frequency/Duty Cycle (WHITE LED) 30% 60% 90%
50 Hz
36.15 mW 62.08 mW 84.89 mW
300 Hz
36.26 mW 63.50 mW 85.12 mW
10 kHz
38.75 mW 64.25 mW 86.14 mW
100 kHz
38.52 mW 62.80 mW 86.59 mW
Frequency/Duty Cycle (RED LED) 30% 60% 90%
50 Hz
54.70 mW 93.82 mW 123.75 mW
300 Hz
57.76 mW 93.81 mW 125.35 mW
10 kHz
56.99 mW 94.00 mW 126.08 mW
100 kHz
56.61 mW 95.11 mW 125.47 mW
Frequency/Duty Cycle (GREEN LED) 30% 60% 90%
50 Hz
41.49 mW 71.29 mW 91.65 mW
300 Hz
41.93 mW 70.29 mW 91.69 mW
10 kHz
41.90 mW 69.96 mW 93.36 mW
100 kHz
42.57 mW 69.71 mW 93.58 mW

So if you look through the data above, you’ll notice that there is, on average, a slight positive correlation between power consumption and frequency. In other words, the higher the switching frequency, the greater the power consumption. This is just what we would expect. Again, this data does not include losses due to transistor gate capacitance, only losses due to the LEDs’ capacitance and the MOSFET’s output capacitance.

For my next test, I wanted to see what losses might be incurred in driving the MOSFET’s gate. Thus, I called on my trusted 8-bit AVR microcontroller (ATMega644P). I wrote a very simple program (which may be downloaded below) to produce a varying PWM output from one of the MCU’s timer/counter outputs. I then measured the power consumption of the entire circuit, AVR included. For this test I only used a 60% duty cycle:

TABLE 2: Data for the ATMega644 driving a BS170 and two green LEDs:

Test Frequency Total Average Power (mW) Calculated Switching
Losses (mW)
50 Hz
91.741 0.000
300 Hz
92.708 0.000
10 kHz
92.622 0.016
100 kHz
92.978 0.157
1 Mhz 95.789 1.568

TABLE 3: Data for the ATMega644 driving a FDP8860 and two green LEDs:

Test Frequency Total Average Power (mW) Calculated Switching
Losses (mW)
50 Hz
93.475 0.004
300 Hz
95.809 0.021
10 kHz
98.238 0.710
100 kHz
114.526 6.848
1 Mhz 161.657 60.914

TABLE 4: Data for the ATMega644 directly driving two green LEDs:

Test Frequency Total Average Power (mW) Calculated Switching
Losses (mW)
50 Hz
69.278 0.000
300 Hz
67.926 0.000
10 kHz
68.778 0.015
100 kHz
68.534 0.147
1 Mhz 70.708 1.467

Discussion of Results

In Tables 2-4, we’re starting to see a much clearer positive correlation between frequency and power consumption. For these tests I also added a fifth data point not gathered with the IOBoard: a frequency of 1Mhz. This should in theory increase our maximum losses by 10x. The results seem to support with this prediction.

The tables above also include a rudimentary calculation for switching losses based on capacitances. I measured the capacitance of my green LEDs to be about 120pF (this value was not mentioned in the datasheet). The gate capacitance of the BS170 is given in its datasheet as 24pF. Finally, the input capacitance of the FDP8860 (a much beefier power MOSFET) is typically listed as 9200pF. To determine switching losses I again applied the formula for a capacitor’s stored energy (½*C*V²). At each switching interval, the parasitic capacitances in the circuit store and then dissipate this much energy. So to determine how much power is lost, we simply multiply this lost energy by the switching frequency (since 1 watt = 1 joule/sec). It appears that these calculated figures match the measurements fairly well. Isn’t it nice when math agrees with reality? Gives me a fuzzy feeling, that.

Now we can essentially think of the 50Hz test point as a baseline with zero switching loss. For the data in Table 4, the 50Hz power consumption is about 69.3mW. The calculation predicts that at 1Mhz, we’ll lose 1.5mW to parasitic capacitance for a total consumption of 69.3 + 1.5 = 70.8mW. This isn’t that far from our measured 70.7mW.

It’s also interesting to note the substantially higher losses incurred when using the FDP8860. This is largely due to its (relatively) enormous input capacitance of 9200pF. This is nearly 400x the capacitance of the tiny BS170. That’s the price you pay for the ability to sustain larger currents without overheating. For more information on power MOSFETs have a look at this IRF document called “Power MOSFET Basics.”

Summary

Well after all that, I’m going to say that whoever manufactures these tail lights can’t really use efficiency as an excuse for choosing a low switching frequency. Unless they need huge FETs to drive huge currents, switching losses really aren’t so much of an issue. I’m guessing that somehow it was just cheaper to go with a low frequency. I’m pretty sure the components themselves aren’t any cheaper, but perhaps the assembly was less expensive. It may be that some automakers already had a low-frequency module in place to drive old incandescent bulbs and then when LEDs came along they just kept using that same module. Anybody out there care to comment on this?

So my advice to those making LED dimmers: pick a frequency of about 300-500Hz to eliminate flicker while keeping switching loss low. Then find yourself a sufficiently large transistor with low capacitance and low on-resistance. And if you’re working on motor controls or power supplies, things get a lot more interesting, but as a start, try a frequency in the 20+ kHz range to avoid audible whine. Good luck!

  • For further reading on LED losses, try this NI article: Light Emitting Diodes.
  • For more accurate MOSFET swithing loss formulae, try this MAXIM article.
  • Test code for the ATMega644P is available here.
  • A complete spreadsheet containing all data can be downloaded here.

Update (9/22/2010): In the comments below, Jas Strong pointed out that in my switching loss calculations, I’d also neglected the power lost in the MOSFET during turn-on. Jas is absolutely correct about that; I should have mentioned this previously. Essentially, while the gate capacitance of the MOSFET is charging, the resistance between drain and source will pass from very high to very low resistance as the conduction channel is formed. This time period, although short, includes a region of, shall we say, “moderate” resistance which briefly dissipates additional power.

Now, in the case of my two-LED test setup, I neglected the effects of resistive switching loss because they’re quite small. Let’s take a quick look at the numbers. First, we need to know how long it takes Vgs to reach the threshold voltage. For simplicity, I’m going to assume that my AVR drives the gate with a constant current of 40mA (the maximum an AVR will provide per I/O pin). Our worst-case turn-on time will occur with the FDP8860, which has a gate capacitance of 9200pF and a typical threshold voltage of 1.6V. Using the formula ic = C*(dv/dt), I find dv/dt = 4,347,826 which means we reach Vth in 1.6/4,347,826 = 368ns. At a switching frequency of 1Mhz, this represents about 37% of a switching cycle. However, we need to double this since we lose power durning turn-on and turn-off. Thus, we’re losing energy in the MOSFET’s resistance over 74% of a single cycle at 1Mhz. That sounds like a lot, but just how much energy is actually lost?

To determine this loss, I’m going to make a big assumption and say that the MOSFET ramps linearly from 20kΩ down to 0Ω during turn-on. I’m also going to assume the voltage of the diode is constant at 3V and the power supply is constant at 4.2V. Remembering that I have 92Ω resistors in series with the LEDs, the instantaneous power dissipation in the FET becomes 2*Rmos*[(4.2-3)/(92+Rmos)]^2 (based on the fact that I have two LEDs and using the formula P = RI^2 and ohms law, I = V/R). Now I need to integrate to determine an average power dissipation over this interval. If my math is correct (feel free to check me), I get a loss of 0.632mW. Since this occurs during 74% of a cycle, the total loss at 1Mhz will be about 0.468mW. Not too serious in my opinion.

Now of course, the power required by my two-LED setup is piddly in comparison with that drawn by a couple brake lights. Once you start sinking more current into your LEDs, this resistive switching loss, as well as the on-resistance of your MOSFET, is going to start to make a bigger difference. So thanks very much Jas for pointing this out!

Frequency Duty Cycle Start Cap Voltage Start Phototransistor Voltage
50 0.3 4.248407 1.464428967
300 0.3 4.246836767 1.4911225
10000 0.3 4.2389857 1.4911225
100000 0.3 4.243696367 1.538228733

Talking in Toronto: IEEE CASE 2010

Last weekend (August 21-24) the sixth annual IEEE Conference on Automation Science and Engineering (CASE) was held in Toronto, Ontario. So I, being fortunate enough to have the General Chair as my thesis advisor, was invited to attend. Actually, I did more than just attend, I presented a portion of my MS thesis to a group of about 30 people. I have to say, it was pretty exciting. Now I think everything went well, but it’s hard to be objective about your own presentation. Plus, when you’re up there talking (and when you have no means of judging time because your cell phone died), time seems to run faster. Toronto was an interesting place – it reminded me a lot of New York City, actually. This was my view from the hotel room Sunday morning:

Toronto, Ontario

If you’re interested, my paper’s abstract can be found here: CASE 2010 Paper Abstract. It was given the very exciting title, “Application of 6-DOF Sensing for Robotic Disturbance Compensation.” As a quick summary, the paper begins with a discussion of my work developing a 6-DOF (degree-of-freedom) laser-based sensor. This custom sensor measures not only the position of an object in 3D space, but also its orientation – pitch, roll, and yaw. Its positional accuracy is as good as 1mm over a 15m x 10m x 1m area. The end application for this sensor involves measuring the base position of a robot and using that data to stabilize the robot’s end-effector at a desired location. Here’s one example we whipped up to show the robot’s motion with and without compensation:

Disturbance Compensation ExampleThe robot in use here is the Stäubli TX90. It’s a six-axis industrial robot capable of moving a 25kg payload within its 1m reach. In this example it’s mounted to a freely oscillating platform (think diving board) along with our sensor package. The sensors measure the position of the base, then relay that information to the control system. The control system makes the appropriate adjustments and calculations, then sends corrections to the robot.

In the picture above, the robot was first programmed to draw a house shape using a red dry-erase marker without disturbance compensation. That is, its base was oscillating, but the end-effector was not compensating. Then, the compensation system was enabled and a black marker was used to draw the same shape. As you can see from the above image, after an initial “turn-on” squiggle, the black line shows very little error compared with the red line. This is the ultimate goal of the system – to maintain a desired position despite potentially large motions in the base of the robot. In this case, the base was moving by up to 50mm while the end-effector was stabilized to within 2mm.

I’m looking forward to writing more about the inner workings of the sensor system at some point in the future (particularly after our patent clears the USPTO). So, until next time!

P.S. Driving across the US-Canada border is no picnic. That is, unless you brought food and drink, in which case just roll a blanket over the hood of your car and settle down for a bite. Actually, I only had to wait about 30 minutes to get into Canada and 40 minutes to return to the US. My friends on the train, however, were stopped for a full three hours on the return trip while everyone was processed. Anyway, advice for anyone stuck in this situation: don’t change lanes. It never helps, plus it frustrates the agents and pretty much every driver in your vicinity. Yay for courtesy! 🙂

A Shiny Box of Fuel Cells

This week I completed the last of what RPI requires for a Master of Science degree in Electrical Engineering. Since I’d finished my thesis in the spring, I only needed three more credits to graduate. So I opted for a summer independent study. And what did I independently study? Fuel cells! (From an educator’s perspective.) It just so happened that my supervisor had previously purchased a commercial 300W PEM fuel cell stack. He wanted to use it in some sort of educational demonstration, but wasn’t sure exactly how to make that happen. So my project for the summer was to design and build just such a system. Our goals were to make the system:

  • Instructional for students of all ages
  • Portable and self-contained (no need to ever plug into AC power)
  • Visible (both the components and their connections)
  • Interactive (something you’d like to play with)

With this in mind I put together what I felt was a solid design, then went to work constructing it. With the help of some of my lab mates, we built the following:

Fuel Cell Demo System FrontWhat you see here is a large 80/20 box measuring about 30″ x 20″ x 11″ and covered with acrylic panels. As you probably guessed, the large red tank is the hydrogen supply, pressurized to 2000psi when full. That pressure is regulated down to about 6psi for the fuel cell stack. The gas passes through a ball flow meter and solenoid valve before reaching the fuel cell stack – located just to the left of and behind the red LED panel voltmeter.

Our fuel cell stack is produced by Horizon Fuel Cell Technologies, model H-300 (more details here). It’s an air-breathing PEM-type stack, composed of 72 individual fuel cells strung together in series. The voltage produced by the stack varies from 40-60VDC depending on the amount of current drawn. Now this variation is unsuitable for most electronics, so it’s first passed through a DC-DC converter (the largest black box just to the left of center). The converter takes the varying input voltage and steps it down to about 13VDC for use in the rest of the system.

You may also notice a 12V lead-acid battery strapped into the middle of the demo box. This serves two purposes. First, and most importantly, it provides power to the stack’s control module during startup. This is necessary to open the solenoid valve and engage the three fans mounted to the side of the stack. Second, it provides a bit of buffering during transients (e.g. when all the light bulbs are flipped on). One problem with fuel cells is that they cannot respond quickly to changing loads. Batteries, however, can rapidly supply more or less current without significant changes in voltage (large capacitors also have the same effect).

The rest of the system consists of an array of ten 12V, 13W light bulbs, a 120VAC inverter, and equipment to monitor voltages and current at several points within the system. This equipment is mounted to the rear of the system, shown here:

Fuel Cell Demo System Rear DAQ HardwareThe data acquisition (DAQ) module shown above is produced by National Instruments, model USB-6009, and is capable of monitoring eight analog inputs at 14-bit resolution. These analog inputs are fed from a custom PCB I designed, mounted directly below the DAQ module. This PCB is responsible for measuring currents using ACS712 hall-effect sensor ICs. It also performs voltage division so that the system’s voltages are within the measurement range of the DAQ. Last but not least, the PCB allows for computer control of the ten light bulbs using MOSFETs controlled by the DAQ’s digital outputs.

From the start, I knew I wanted to use LabVIEW to monitor and control the system – it’s built for data acquisition and handles simple controls quite well. The only question was, what sort of hardware should I run it with? Since I didn’t need much horesepower and in fact was looking to minimize electrical power consumption, I went with the Asus Eee PC, model 1001PX:

Asus Eee PC 1001PXWith its dual-core Atom processor, the 1001PX actually performs quite well running Windows XP. Its 20-30W power consumption (when charging) is equally impressive. Running LabVIEW 2009 presented no performance problems whatsoever. My only qualm is the lack of screen resolution – 1024 x 600 is just a bit tight most of the time. However, space was no issue since all of my LabVIEW VIs were compiled into executables without scrollbars, menubars, etc. Here’s how the main panel turned out:

LabVIEW Front Panel

From this panel the user can monitor voltage, current, and power at different points throughout the system. The light bulbs can be turned on and off with a single mouse click. I’ve also created VIs for taking polarization curves (voltage vs. current density) and for monitoring the stack’s voltage at high speed (48kHz) during transients. To top it all off, the Eee is loaded with a sample presentation containing the principles of operation for fuel cells as well as diagrams for the demo system itself.

The system has yet to be tested in a real classroom environment. Sadly, I may not be around to see that happen. But I’m pretty confident that it’ll be put to good use. The grand total for all parts in the system? About $4000. Thanks for reading!