Gas station without pumps

2014 May 30

Class-D amplifier lab done, EKG block diagrams begun

Filed under: Circuits course — gasstationwithoutpumps @ 21:08
Tags: , , , , ,

Yesterday’s lab ran long (as expected), because students had not gotten enough done in Tuesday’s lab.  But everyone in the lab did get a working class-D power amplifier.  Several also managed to measure the turn on and turn off times for the comparators driving the FETs, though that required some hands-on guidance in using the digital scopes (setting the trigger level to the FET threshold voltage, then looking to see how long the rise or fall was before reaching the trigger level.  As expected, everyone had chosen values that made the pFETs turn on and off quickly, but it was difficult to get the nFETs to turn off quickly.  I don’t know whether anyone managed to equal turn on and turn off times for the nFET (they turned on fairly fast), but several groups managed to keep their FETs cool.  Even those with warm FETs did not dissipate so much in them that they got dangerously hot.

I’ll be reading the design reports over the weekend, and I’ll see whether the students really understood PWM or not.  I suspect that about half the groups understood what they were doing well enough, and the other half got part of the ideas.  There should be time on Monday to review the idea of PWM and to explain again why it is a good choice for efficient power delivery, particularly for inductive loads.

Today, I returned the quiz 2s redone as homework.  Students did fairly well on them as homework (range 18.5 to 31.5 out of 36, up from 7 to 17 on the timed quiz).  The biggest difficulty was with the last problem, which asked them to design a simple amplifier, giving both a block diagram and a schematic.  A lot of students did not understand the question as I phrased it, perhaps because I had not been clear enough earlier in the quarter about what a block diagram means and how to use it.

Students have not yet internalized the idea of something having inputs and outputs, and a block diagram being a refinement of an I/O spec into I/O specs for subunits.  I may need to use that language more explicitly earlier next year. I’m thinking also that I need to add more text to the lab handouts next year and refer to them as a draft textbook rather than as lab handouts.  How many pages do I have so far?

handout pages
01-thermistor  11
02-microphone  9
03-hysteresis  11
04-electrodes  7
05a-loudspeaker  8
03b-sampling  7
06-audio-amp  6
07-pulse-monitor  11
08-pressure-sensor  8
09-power-amp  13
09-power-amp-addendum  6
10-EKG  5
total  100

One hundred pages is a bit short for a textbook, but there is a lot of explanatory material still missing (most of which I provided in class or in lab). If I worked on it diligently over the summer, I could probably create a book with most of what the students need that would be around 200 pages. Do I have the energy to turn this into a textbook? Is it worth the effort?

After going over the block diagram of the quiz problem, I helped the students develop an EKG block diagram.They did get to the realization that the unknown but potentially large DC offset from the EKG electrode half cells limits the gain that they can ask from the first stage of the EKG, and that they’ll have to high-pass filter and add more gain.  The design is similar to their pressure sensor instrumentation amp, but the gain needs to be higher (1000 to 1500, rather than 100 to 250), and the pressure sensor amplifier had to go down to DC, so did not include a high-pass filter.

I was a little worried that I may have suggested too high a lower end for the passband (0.1Hz to 40Hz).  They’ll get less baseline drift with a 0.5Hz cutoff instead of 0.1Hz.  My EKG designs have used 0.05Hz—88.4Hz and 1.0Hz–7.2Hz for the blinky EKG.  Both worked ok, but I now think that the 7.2Hz cutoff is too low (it was adequate for blinking an LED, but not for recording the waveform).  Since I did not have much problem with a 0.05Hz corner frequency, I think they’ll be ok with a 0.1Hz one.  The blinky EKG circuit has an adjustable gain (needed to make the R spike large enough to light the LED), but it is probably better to have a fixed gain.

It would be really nice if they could finish the EKG on Tuesday, since the annual undergraduate poster symposium is scheduled for the same time as the Thursday lab, and I always like to spend an hour or so looking at the posters.

2014 May 28

Cardiac action potential

Filed under: Circuits course — gasstationwithoutpumps @ 22:27
Tags: , ,

At the beginning of today’s class I took some questions about the quiz that the students had just redone as homework, and some about the class-D power amplifier lab that they had worked on yesterday and will work on more tomorrow.  I don’t think that most of the groups got far enough in the lab yesterday—they barely finished the pre-amp and had not yet wired up the comparators—I’m a little worried about how long the lab might take tomorrow. Other than taking orders for T-shirts, I spent most of the rest of the time on cardiac action potentials.

This year I decided to try to give a lecture on action potentials myself, rather than having a guest lecturer, as I did last year.  In part I wanted to try to strengthen my own knowledge (which comes mainly from Wikipedia’s cardiac action potential article), and in part I wanted to cover muscle action potentials, rather than neural ones, so I needed to bring calcium ions into the picture.

Fortunately for me, one of the students in the class had had a neuroscience course (all the students have had more biology than me), so could remember words I had forgotten (like sarcoplasmic reticulum for the organelle that stores Ca++ ions in cardiac cells).

I was able to give rough electronic analogs for some of the components (like a capacitor for the membrane and an FET for the voltage-gated ion channels), but I did not include some of the things that the bio professor had included last year (like references to the Nernst equation).

I tried to explain how the EKG signal we record corresponds to the dipole that results from the movement of the waveform across the heart, but I think I did a poor job of it. I got confused about whether the Sinoatrial node was in the upper right or upper left atrium (it is the upper right), and did not remember the polarity of Lead I (which is left arm – right arm). The P and R depolarizations are upright (positive voltage) in Leads I and II, because the depolarizations travel from upper right to lower left of the heart.

I think I’m still a little confused why the dipole points the way it does.  As best I can figure out, the leading edge of the depolarization wave lowers the voltage outside the cell quickly, giving a + potential in front of the wave and a – potential during it.  The trailing edge of the depolarization wave is much slower and so does not produce a strong dipole?   I’m not quite sure how this dipole interacts with the volume conduction of the body to produce the waveforms we observe on the EKG electrodes, though.

I did get out a sketch of the PQRST waveform and a brief interpretation of it.

There are some cool animations at that I did not get a chance to use. Perhaps I can make use of them on Friday. I should also point students to the filter information in the Wikipedia ECG page, because they will need to set up their 1-lead EKGs in “monitor mode”.

2014 May 24

Class-D amplifier lecture 2

Filed under: Circuits course — gasstationwithoutpumps @ 17:41
Tags: , , , , ,

Yesterday, in the last lecture before the long weekend (hence before they start wiring their Class-D power amplifiers), I covered three topics:

  • open-collector outputs.  I had hoped to avoid that this year by switching to the TLC3702 comparators, but I couldn’t get the TLC3702 comparators to work with the FETs, so I went back to using the LM2903 comparators that the students got in their parts kits.  Because I had already cut the open-collector information from the lab handout, I had to cover it in lecture, and I’ll have to write an addendum to the handout today.
  • LC filters for the loudspeaker.  This was a pretty rushed job.  I didn’t even have time to get them to derive the LC filter behavior with no load (the standard voltage divider formula \frac{Z_{down}}{Z_{up}+Z_{down}} becomes \frac{1/(j \omega C)}{j\omega L + 1/(j \omega C)}=\frac{1}{1-\omega^2 LC}, which goes to infinity at \omega=1/\sqrt{LC} and 0 at DC and infinite frequency.  Instead I pointed them to a gnuplot script that they needtomodify to see how much power the LC filter delivers to the loudspeaker at different frequencies, with different choices of capacitor (no choices for them on the inductor—it is a 220µH 0.252Ω AIUR-06-221 inductor).
    LC filter and loudspeaker model used for sizing the capacitor in the gnuplot script.  The students need to replace the loudspeaker model with the one that they created in an earlier lab.  I think that next year I may use a much simpler loudspeaker model in the script, so that I can tell whether the students have replaced the model or not.

    LC filter and loudspeaker model used for sizing the capacitor in the gnuplot script. The students need to replace the loudspeaker model with the one that they created in an earlier lab. I think that next year I may use a much simpler loudspeaker model in the script, so that I can tell whether the students have replaced the model or not.

    The output of the gnuplot script looks something like this:

    Output of the gnuplot script, note that too small a capacitor results in a spike near 61kHz, and nowhere near enough suppression of their 50kHz–100kHz PWM frequency. Too large a capacitor results in a big boost in gain around 3.8kHz, which would a different sort of problem.

    Output of the gnuplot script. Note that too small a capacitor results in a spike near 61kHz and nowhere near enough suppression of the 50kHz–100kHz PWM frequency. Too large a capacitor results in a big boost in gain around 3.8kHz, which would a different sort of problem. Click
    here for a PDF version of this plot.

  • Block diagram for the class-D amplifier.  I had originally planned to spend almost the entire lecture having the students develop the block diagram, but the addition of almost 45 minutes on open-collector outputs and review of the FET driver stage left me with little time for the block diagram. I did get some participation from the class in developing the diagram, but almost entirely from one student.  I felt bad about presenting, rather than getting them to participate in creating the block diagram, but they had to have a block diagram to do their more detailed design over the weekend.One important point I that I emphasized, based on common problems on the second quiz, is that a block diagram is not a simple pipeline, but can have merging and splitting.  The class-D amplifier has merging of voltage generators for DC bias with the signal path and merging of the triangle wave generator with the signal path.
    One good thing came out of the block diagram discussion: in putting together the design they had the output of the preamp centered at 3v but the input of the comparators centered a 0v.  I could point out that putting range information on the signal lines allowed them to catch this error early (before even doing the schematic).  Most of the class was able to come up with the standard solution for changing the C bias: adding a high-pass RC filter.  I don’t know whether they can choose a corner frequency appropriately, but we’ll see that on Tuesday.

I didn’t get a chance to teach them about real power, either, which the LC script computes.  I’ll have to go over that next Wednesday, between the two halves of the lab.  There is a writeup of the concept in the lab handout, but my experience has been that students in this class don’t learn from written material.

I’ve also told the students that they need to get all their required “REDO” assignments turned in by Wednesday.  It seems that this year’s class does not have the time-management skills to handle open deadlines—they keep putting off redoing the assignments.  Given that they sometimes don’t fix the assignments sufficiently when they turn them in again, leaving the assignments to the last week is really dangerous.  Next year I’ll have to make one-week deadlines for redoing the assignments, though I can be generous about extending deadlines on request.

Overall, the lecture was way too rushed, because of the extra coverage needed for open collectors. Next year I’ll have to make sure to allow 3 lectures for the class-D power amp, which means not having it on the week with Memorial Day.  I’ll probably want to move Quiz 2 a little earlier also, so that it isn’t the week before the class-D power amp.

Segmenting filtered signals

Last August, I posted on Segmenting noisy signals from nanopores, and in January I updated that with improved parameterization in More on segmenting noisy signals.  Last week, Jacob Schreiber (the programmer for the nanopore group) found some problems with the method. He had re-implemented the method in Cython (rather than Python, as I had prototyped it), and got the same results as me when applied to the same examples I had used. But when he applied it to some new data, he got serious oversegmentation, with hundred of segments chosen where only one should have been.

I tracked down the problem to a difference in how we were applying the method.  I was applying it to data that had been low-pass filtered at 5kHz, recorded at a sampling frequency of 100kHz, then downsampled to 10kHz.  He took the same data without downsampling, then further filtered it with a low-pass filter of 1kHz or 20kHz.  I confirmed that the same data would be correctly segmented with my downsampling and severely oversampled with his extra filtering.  He conjectured that the problem was with the sampling frequency and that the problem could be fixed by just changing the desired gain in log Prob based on sampling frequency.

His argument didn’t seem quite right, as the underlying math was not dependent on sampling frequency.  What it did assume was that the samples within a segment were independent Gaussian-distributed values (more correctly, that the difference between the samples and the model being fitted was Gaussian).  Putting in a low-pass filter removes the independence assumption.  Indeed, with a very low cutoff frequency you expect to see the values changing only slowly, so the segmenter would see it as highly desirable to chop the signal up into short segments, each of which has low variance.

I confirmed that I could get serious oversegmentation even with very low sampling frequency, if the filter cutoff was set much lower.  I got reasonable segmentation results with only minor variations across a wide range of sampling frequencies if the filter cutoff frequency was close to the Nyquist frequency (half the sampling frequency), but oversegmentation if I filtered to 0.1 Nyquist frequency.

There are two ways to address the problem:

  • When filtering, always resample at twice the cutoff frequency.
  • Adjust the thresholds used according to the cutoff frequency and sampling frequency.

Adjusting the thresholds is a cleaner approach, if it can be made to work.  Today I I tried filtering Gaussian noise with a 5th-order Bessel filter (implemented by scipy.signal.lfilter) and looking at the distribution of the log-odds that we threshold:

L_{i}=\log \frac{P_{\theta_{1}}(s:i) P_{\theta_{2}}(i:t)}{P_{\theta_{0}}(s:e)} = \log P_{\theta_{1}}(s:i) + \log P_{\theta_{2}}(i:t) - \log P_{\theta_{0}}(s:t),

where i is the breakpoint, θ1 and θ2 are the parameters for the two segments, and θ0 is the parameter setting for the whole segment.  For stepwise Gaussian models this simplifies to L_{i} = (s-i)\ln \sigma_{1} + (i-t)\ln \sigma_{2} - (s-t)\ln \sigma_{0}.   (To avoid the extra computation of taking square roots, the segmenter actually computes twice this, using variances instead of standard deviations.)

I found that the distributions were still very well fit by exponential distributions, but that the scaling factor changed with the cutoff frequency.   Furthermore, the log-odds seemed to scale linearly with the ratio of Nyquist frequency to cutoff frequency.  Reversing that scaling makes the curves superimpose nicely:

    If I scale the log-odds by filter cutoff frequency as a fraction of the Nyquist frequency, the distributions seem to be almost identical. Unfiltered Gaussians have slightly lower log-odds values. Filtering, then downsampling by averaging blocks of 10 samples has slightly higher log-odds, probably because the averaging does a little more low-pass filtering.

If I scale the log-odds by filter cutoff frequency as a fraction of the Nyquist frequency, the distributions seem to be almost identical. Unfiltered Gaussians have slightly lower log-odds values. Filtering, then downsampling by averaging blocks of 10 samples has slightly higher log-odds, probably because the averaging does a little more low-pass filtering.

Each curve of the plot was computed from only 1000 sequences of 10000 samples, so the results are vary a lot in the tail, due to not having enough data to get the tail exactly.

Because the segmentation method is not very sensitive to the threshold set, simply scaling thresholds by Nyquist frequency/cutoff frequency should work well.   The default parameters had been set for the downsampled 5kHz, so to match the old results in signals  that aren’t downsampled, the thresholds should probably be scaled by about 1.1*Nyquist frequency/cutoff frequency.  (The analog filter cutoff frequency is probably in the metatdata in the data files, but the abf format is messy enough that I’ve not tried to extract it.  The lower cutoff frequency of the analog filter or a subsequent digital filter should be used.)


2014 May 22

Class-D lab revision didn’t work

In Long weekend, I discussed what I was planning to do about anticipated problems with the class-D amplifier lab, specifically

  • Replace the AOI518 nFET with one that has a lower input capacitance, such as the PSMN022-30PL,127.  The gate charge at VGS of 4.5v is 4.4nC, about half that of the AOI518.
  • Replace the open-collector comparator with one that has push-pull output, like the TLC3072, which can provide ±20mA current (more than the LM2903, even before we allow for the current through the pullup resistor).

I did a neat version of the schematics last night using the TLC3072 comparators and the AOI518 nFET. This year I remembered to include an adjustable gain stage in the preamp, so that I could more easily control the volume. Today in the lab, while the students were soldering up their instrumentation amps for the pressure sensor, I wired up the class-D amplifier, one stage at a time, confirming that each stage worked using the oscilloscope before moving on to the next. The build took me longer than I had expected—almost 2 hours.

Everything worked fine until I connected the drains of the two FETs together.  Initially it worked ok, but after about 20 seconds the shoot-through current increased, causing the current limitation of the bench power supply to kick in.  Then the voltage on the lower power rail moved up  close to ground, and the input voltage on the comparator was swinging below the negative rail.  I think that this damaged a couple of my TLC3072 chips—I’ve marked them, and I’ll have to test them before using them.

Replacing the AOI518 transistor with the smaller  PSMN022-30PL,127 did not help.

I finally borrowed an LM2093P chip from one of the students (I’d left mine at home, by mistake) and tried replacing the TLC3072 chip with the LM2093P. They have the same pinout, but the LM2093P is an open-collector output, so I had to add pullup resistors.  I guessed a couple of values, based on vague recollections of last year’s design, and the amplifier worked.

Initially I could only run the amplifier up to ±7v on the power supply, without the FETs getting too warm—there was still too much shoot-through current during transitions.  I switched to a lower resistance for the pullup on the pFET gate, to make the voltage swing less and the turn-off faster.  At that point the amplifier worked quite happily with a ±8v swing without the transistors getting warm.The circuit worked with either of the nFET transistors, so I’ll just have the students stick with the AOI518 in their parts kit.

I couldn’t crank up the volume on the speaker, though, because I got feedback squeal whenever the gain got too high.  Perhaps I should make a long speaker extension cable for students to do testing next week.  I seem to be out of speaker wire, though.

The class-D amplifier design will be a tough one for the students, and I’ll need to do a supplemental handout on open-collector outputs (I’d cut that material from the handout when I thought that we would be using the TLC3072 comparators).

Last week I thought that the students could start on the class-D amplifier in lab today, having finished the soldering for the pressure-sensor amp on Tuesday, but it took almost the whole lab time today for students to finish the soldering, even though everyone had working breadboards on Tuesday.  The lab ran over by almost 2 hours for one group, taking a total of 8 hours instead of 6 (last year the same lab took only 4 hours for the slowest group, probably because last year’s class came to lab more prepared).

The layout took longer than students expected, as did the soldering.  Everyone did (eventually) get working soldered instrumentation amps, though for a couple of groups I had to point out that their wiring did not match their schematic (they had called me over to help debugging).

In one case they had connected a resistor to the wrong point in their circuit.  I found the bug by tracing where their virtual ground was connected, and asking them to identify each component. Even after I showed them both resistors connected to their virtual ground, where only one was supposed to be, it took them a long time to realize what the discrepancy was. They had wired exactly what they had laid out, but the bug was in their layout, and they had not done a thorough enough job of checking against their schematic.

Another group had a working circuit but with too little gain. After checking a few of the DC voltage levels with them, I compared each of their resistors to the schematic.  In one place they had wired in a 1kΩ resistor where the schematic called for a 10kΩ resistor.  They unsoldered the incorrect resistor and soldered in the resistor from the design, which salvaged the circuit.

I also returned Wednesday’s quiz in lab today—pretty much like last year, the scores were much better on the second quiz than the first one, though still only half what I think the students should be able to do at this point of the quarter.  I’m once again assigning the students to redo the quiz as homework.  I need to decide soon whether to give them another quiz during the final exam time.

Next Page »

%d bloggers like this: