Lab Pet Peeves

My Lab Pet Peeves

With Labs 1 and 2 having just been graded, I thought I'd present a summary of common errors in lab practice that drive me crazy. Here they are:

Precision of Measurement

The resolution of a measurement instrument is usually a fixed fraction of its maximum signal range. For example, the scopes we use in lab use an 8-bit analog-to-digital converter to measure the input voltage. So the vertical resolution is one part in 2⁸ = 1/256. For the accuracy in measuring a signal, this is the fractional error if your signal fills the entire range of the instrument. If your signal only fills part of the instrument's range, the resolution error is a larger fraction of your signal.

For the rise time measurement shown in the figure below: note that the signal being measured has a vertical amplitude of only 2-1/2 divisions. Since the total vertical range of 8 divisions represents 2⁸ = 256 quantization levels, there are only 80 quantization levels used by the signal shown. Thus the measurement accuracy on any vertical measurement can be no better than 1.25%. By not using more of the instrument's measurement range, the accuracy of the resulting measurement is degraded.

Additionally, note that the rise time part of the exponential waveform (from the 10% to 90% points) only uses about 2 horizontal divisions. Again there will be reduced accuracy in measuring the times on the horizontal axis since most of the horizontal divisions are not used to display parts of the waveform we care about.

Precision of measurement

To avoid this problem:

THINK about what you're measuring, and what parts of the waveform have the information you need. Adjust your scale factors (vertical V/div, horizontal sec/div) so that the part of the waveform you're interested in uses as much of your instrument's measurement range as possible.

Poor Choice of Data Points for Model Parameter Fitting

The behavior of the MOSFET is so complicated that it is impossible to capture its operation in any kind of simple mathematical representaion. However, as designers we need some kind of tractable mathematical model to work with MOSFETs. ALL of the models we use in ECE4902 are approximations to device behavior. These approximations apply only over a limited range of voltages and/or currents.

The seductive power of numerical software packages like MATLAB is that they will fit a precise line to any set of data points we provide. However, that doesn't necessary mean that the resulting line makes any sense. This is especially true when fitting a model to a set of points where the model doesn't necessarily apply to all the points. An example is shown in the figure below, where the extreme point at V_GS=+5V skews the fit line to give errors at data points near the threshold voltage. Since the point of this exercise is to extract the threshold voltage, it makes sense to throw out a data point far from the region where we are trying to apply the model.

Note that we don't omit the data point from the plot - that would be dishonest. But it is legitimate to omit the point from the fitting process, and then describe in your discussion what you did and why.

Poor choice of points for fitting

If you don't choose the points properly, the model parameters will give large errors. The figure below shows the plot of measured data vs. predicted by the model. Note that the threshold voltage is wrong, and there are large errors in the r_DS(on) values for V_GS near the threshold voltage.

Model prediction errors

To avoid this problem:

LOOK at your data. THINK about what is important in the behavior you are trying to model. Know that the model is an approximation that won't necessarily hold for all points in your data set. Be honest about presenting your data and the limitations of approximate models.

Precision of Fit Parameters

Microsoft Excel will fit a line to a set of data points just fine, using the same least-squares technique that MATLAB does. But if you don't format the display of the fit parameters to use scientific notation, you will have poor resolution on the actual number coming out of the fit procedure.

In the figure below, the slope shown of .0009 has only one significant figure! The actual number could be anywhere between 0.000850 and 0.000949. That's an error of +/- 6%. It's even worse for the y-intercept, which is shown as 0.001: that could be anywhere between 0.00050 and 0.00149. That's an error of +/- 50%. Since you find the x-intercept (the threshold voltage V_tn) as the y-intercept divided by the slope, the error on the final result is +/- 56% !

Poor resolution on fit parameters

To avoid this problem:

Be sure you format the display of fit parameters to have enough significant figures so your accuracy is limited by the precision of your measurement, not the resolution of the number displayed.

Other Bad Excel Behavior

Microsoft Excel is oriented toward management applications - it takes some effort to overcome its business-orietned bias to do good engineering and scientific work. One problem I often see comes from moving too quickly when choosing the format for plotting data. When the "Chart Wizard" gives you a choice of plot formats:

If you choose "Line", you get a plot like this:

Bad Excel Behavior

What is wrong with this picture? Look carefully at the values for the data points on the x-axis, and then the spacing on the x-axis. The student correctly took more data points at finer resolution between 1.7V and 1.9V, to accurately determione the threshold voltage V_TH. BUT in "Line" plot mode, Excel plots the points at uniform spacing regardless of the actual x values! So what should look like a straight line according to the mathematical model doesn't look straight!
To avoid this, choose "XY Scatter" rather than "Line" - or use MATLAB which is better for real engineering.