An experimental PWM circuit with 74HCU04

[mawyatt]

I started developing a MASH-based PWM voltage calibrator a couple of years back, and got fine results. This technique ensures high resolution and very good linearity at the same time, using just one PWM string. One has to make sure to use a proper lowpass filter to filter out not just the PWM but also the rumble that occurs from the MASH-process. I made a DC-accurate lowpass filter and added linearization of the switching stage, and tested the calibrator's linearity against a HP 3458A. The result from 16 voltage sweeps from 1 V to 9 V (201 steps of 40 mV) is shown the the enclosed graph. The graph shows the deviation from a linear fit. The GUI and the MASH driver for the switches was originally made in LabVIEW, for reasons of flexilibity, but I'm moving the design to a controller in order to have a stand-alone solution. Presently, the voltage setting has a resolution of 32 bits, and it's being MASHed into 16 bits of PWM. There are still improvements to be done, but the results so far are promising. MASH with PWM is definitely a thing if you are looking at resolution and linearity, and multi-string PWMs will probably never be able to reach the same levels of linearity without very elaborate trimming.

That's impressive

Best

[arvee]

Here are some notes about the PWM software.

The PWM is full range.
The output voltage is between 0 volt and the 7 volt reference voltage.

#1
The unit for the width is one clock cycle (1/16MHz).
If the count is zero, all pulses have width 0.
If the count is increased then pulses with width 1 appear.
This continues until there are 300 pulses with width 1.
If the count is increased beyond 300 then pulses with width 2 appear.
This continues until all 300 pulses have width 2.

There are 3 operating regions:
1. In the first region pulses with width 1 appear.
    The width of these pulses is NOT exactly 1 clock cycle.
    The level shifter between 5 volt and the reference voltage is not ideal.
    The rise and fall time can differ.
    What matters is the actual width.
    The region is very small: 12V / 53333 = 225 uV.
2. In the second region the width of pulses is increased by 1.
    The increment of the width is exactly 1 clock cycle.
    The pulses are longer, with a constant width error.
3. In the third region, at the top of the range, pulses disappear.
    It is the inverse of the first region.
   
Each of the 3 regions has a different slope.
The error can be corrected with one parameter.

The processor frequency is 16 MHz, the period is 62.5 ns.
The PWM frequency is 300 Hz, the PWM period is 3.3 ms.
A change in the width by 3.3 ns gives 1e-6 error.

#2
The pulse width can be set from 1 to the period count.
Zero width is not included.
Zero width can be implemented by setting the output to 0.
But then PWM interpolation between 0 and 1 is not possible.
The solution is to invert both the count and the polarity of the output.

#3
The polarity of the output voltage is reversed using two relays.
There are two signed zeros depending on the state of the relays.
The voltage difference between the two zeros is twice the output offset of the buffer.
It is an error of the voltage source.
The correction should be with done by a shift over the x-axis.
It is easier to do on the y-axis without any loss of accuracy.
The error is removed by shifting the positive and negative readings over the y-axis towards each other.

The relays are Coto 3402.
The datasheet has a picture of the offset over time after the relay is energized.
The drift in offset was 10 uV. The measured shift of relays used is less than 1 uV.

#4
With both relays energized the output is shorted.
The buffer's output is also shorted. Therefore the PWM voltage is set to zero volts.
In addition there is a 1 k ohm resistor in series with the buffer output.

#5
The calculation is done with integers.
The numerical accuracy is improved with guard digits.