Wien Bridge Oscillator. August 2013.

This page describes a switched frequency Wien Bridge oscillator with
j-fet Vo stabilization, buffered Rout = 55r, peak Vo max = +/- 5.1V for
both sine waves and square waves, switched attenuator for 4 levels at
10dB steps below 0dB, and with logarithmic variable Vo. There are 6
ranges between 1Hz and 1MHz, with an F decade each.

There are 4 images,
Fig 1. Sheet 1, basic block diagram schematic of WB oscillator with
details of switches.
Fig 2. Sheet 2, full schematic details of discrete bjt amp for oscillator,
and attenuators, and buffer output stage.
Fig 3. Explanatory diagram of basic Wien Bridge network with two equal
value R and two equal value C.
Fig 4. Schematic of Schmitt Trigger sine to square wave converter with
2 discrete bjts.

I was cleaning up my workshop when I discovered I had been given an old
BWD Electronics 141A audio oscillator which had many problems.
I delayed my life for 2 weeks while I rebuilt the darn thing.

It was made in Australia in the 1960s or 70s and designed to be used with
batteries if need be.
Therefore it contained the bare minimum of current needing devices and
used the bare minimum of other parts. Current and parts cost money, and
bean-counters and CEOs hate any expenses, and BWD had no ambitions
to seriously compete with other expensive brands like HP at the time.

The 141A unit was "budget test gear" and since it offered nothing special,
it is not disrespectful to dismiss ideas used by the original designers and to
consider the unit to be completely expendable in pursuit of better operation,
while keeping the box, and the few parts inside worth actually re-cycling.
I have never ever needed battery operated test gear, except my DMM.

The only way to improve the performance was to revise the circuit inside the
small box completely.
After measuring and making an intelligent appraisal of the existing usability,
and having to give the unit a 1 out of 10 score, I removed all discrete bjt + R + C
and tracks from the circuit board, except 6 pairs of plastic and styrene C and
tracks around the interactive press button F range switches.

The 15k wire wound reverse log dual gang Compton pot for varying was
intermittent and thus faulty and junked. It is impossible to dismantle it and clean
it, or inspect its internals.

Switched resistors were used instead of a pot for F control.

I removed the entire PSU which was not original, but a replacement installed by
a company 'Bodgy & Dumb P/L'.
The new PT and R+C parts of PSU are in a box screwed to the steel plate
rear panel.
I put in a new amp on an additional board for the Wien bridge oscillator.
Another board has the j-fet as variable NFB resistance needed to control
output level.
I built a new Schmitt Trigger square wave circuit where the old one was.
Another small board with a complementary pair of darlington connected bjts
connected as a PP emitter follower buffer preserves the HF content of all
output signals and reduces Rout to 55 ohms for all levels of operation.
I re-designed the output attenuator switch to give 0dB, -10dB, -20dB -30dB
levels which is more usable to me than the original. I replaced the variable
output level pot.

My new circuits contain more parts than the overly simplistic previous original
had, but it work better, and there is now no DC offset at the output, and peak
voltage for sine waves and square waves is 5.1V, and the rise time for square
waves excellent, 104V/uS. There still a few trim-pots to adjust, and 3 trim C but
that is inevitable with much analog test gear where best performance depends
on very small changes to R and C values.

There are 3 schematics :-
Fig 1 shows the switched R+C for C positive FB network and the simplified amp
schematic with output attenuator and the negative FB network with j-fet to vary
the amount of NFB.

Fig 2 shows the full details of the amp using discrete j-fets and bjts which
gave wider BW than most common op-amps.

Fig 3 Gives basic properties of the Wien Bridge network.

Fig 4 shows the Schmitt Trigger square wave schematic.

Fig 1.

The Fig 1 schematic took me many hours to design and construct and test and
adjust and revise before I got flawless operation over the six F ranges covering
1Hz to 1MHz . The original BWD 141A Comptom pot was replaced with a 1950
dual wafer 12 position break-before-make ( BBM ) switch taken from a defunct tube
tester I was given which was quite beyond repair.
I found the amplitude bounced badly at first with this switch so I converted it to
make-before-break ( MBB ). It required careful metal work with 2mm x 3mm brass
plates and a steady hand with soldering iron and Dremel to clean up where needed,
with two pairs of reading glasses.
This reduced the trace bounce on the oscilloscope.

If you build this, RS components sell suitable 40mm dia wafer type switches with
MBB with 12 positions and 1 pole per wafer. I arranged the resistance combinations
"R comb" hanging off the large lugs of the old rugged switch.

Although I show rotary switches for capacitors the original press button spring loaded
interactive switches have been retained and the action is so fast that little trace bounce
occurs from high transient amp voltages at amp Vo.

The much more detailed schematic of the oscillator amp, buffered output etc and PSU is here :-
Fig 2.

The original BWD 141A used an ITT P54 thermistor which is a low wattage type
of 50k cold R and minimum of maybe100r. It was used where the Vo was 2.5Vrms,
and thus 1.66Vrms was across the thermistor and this slightly heated it to reduce its
resistance to approximately 2,000r. It worked well except that the Vo bounced crazily.
It was used where the R12 2k7 above is used between oscillator amp Vo and NFB
input port.
The principle with P54 is that If Vo rises, thermistor gets hotter, reduces its R, and
increases NFB thus reducing Vo until an equilibrium is achieved after the yo-yo of
Vo settles down. THD was at least less than 0.1%.

I didn't want to rely on getting a replacement P54 in future. No modern thermistors
seem designed to work to give a sufficient temperature rise to cause enough useful
resistance change. All mainly seem designed to change their resistance usefully as
a result of an external change in temperature. So modern thermistors are useless as
substitutes for P54.

In the case of the Wien bridge oscillator, we want the thermistor ( or lamp globe )
to operate at a near constant T well above room T and to vary its R due to minute
changes to power and heat generated by Vac or Vdc across the "element" or tiny
glass coated thermistor or tungsten lamp filament within a glass tube.
But lamps are very prone to Vo bounce when switching anything, because like
thermistors, some transient high current causes a delayed rise in or fall or resistance
because temp change is not instant. Therefore, with temperature dependent R
there is an unwanted time constant that can cause LF amplitude modulation of
the Vo signal.

So, I changed the NFB circuit to use a j-fet and rectifier circuit to produce a
negative gate bias of between about -0.8V and -1.1V.
The negative bias voltage for j-fet gate begins to be generated when Vo = exceeds
3.2V peak. The 8 x 1N914 do not begin conducting until their forward voltage = 0.4V
each, or 3.2V across all. The turn on voltage of a silicon diode is not linear, and the
forward voltage across an Si diode does not reach a max of about 0.7V until the max
allowable rated I flows.

The series diode action was found to be a smoother than trying to use a low voltage
zener and series 1 x 1N914.
When gate bias = 0Vdc, the j-fet has minimum Rd = 180r approx.
This then makes ß a lot less than 0.333 so that much less FB is present and
oscillations start and increase Vo very quickly. But as Vo increase, gate bias becomes
more negative and j-fet Rd increases which increases ß so the NFB increases thus
preventing an increase in Vo when the NFB applied to Q2 gate is very nearly the
same amplitude to PFB signal at Q1 gate.
Equilibrium is reached and oscillations and "magically", you get the same Vo amplitude
for all F between 1Hz and 1MHz. The j-fet characteristic curves do not describe their
properties fully but for each value of negative gate bias the "diode" resistance line with
high line slope near vertical id axis varies with gate bias voltage. What isn't shown is
that the Rd operates even if the applied voltage at drain goes negative below the
source voltage.

The amount of current change in Rd across the Rd "diode resistance" has to be kept
small because the diode R line on curves is not straight, and rolls over to a virtually
flat high value R line because above about 2Vds, the Rd maybe thousands of ohms.
If we have Vo at 3.66Vac, then there is about 1.22Vac at both NFB and PFB ports of
the amp.
In fact, the amp has open loop gain at 1kHz = 40,000 approx, and PFB voltage is only
0.09mV above the NFB voltage. The gate to gate voltage cannot easily be measured.

The ideal region of operation for j-fet is where Rd = 300r approx. If I have R12 = 2,700r,
and ß = 0.33, then the whole lower R of the FB network must be 1/2 x 2,700r = 1,350r.
The total R of both R = 1.35k + 2k7 = 4.05k, and if Vo = 3.66V, then Iac = 0.903mA.
Therefore if Rd = 300r, then Vd = 0.271Vac, not much you may say, but little things
have great and fundamental importance.
Therefore the VR1 1k pot plus R6 470r must equal 1,350r - 300r = 880r so pot is set
for 880r - 470r = 410r.
Because the R+C values used in the Wien network cannot be accurate enough to
demand ß always exactly equal 0.33, and because of other slight variables in the
circuit, using a pot for NFB adjustment is imperative.

If Vo goes higher than 3.66Vac, ß would rise above 0.333, and oscillations would stop.
To gain equilibrium and thus Vo there must the pot VR1 1k0 to give
some fine adjustment of NFB so that oscillations do not dither and bounce, or have
too much THD. The VR1 pot controls Vo across a small Vo range. If pot resistance
is set high, oscillations to stop and start and "dither about", but THD is low. If pot R
is too low then you get very stable Vo but high THD, easily exceeding 3% mainly 2H
and seen in the CRO as you twiddle the trim pot with a screw driver.
Between the two extremes there is a sweet spot where action is good, and
THD < 0.5%, and the production level of gate bias with
diodes must be just right
to get the sweet spot.

To further reduce bouncing of the Vo level when switches are used, the Vo has a
voltage clamp using 5.1 zener diodes and 1N4007 seen in Fig 2 below.
During switching, the oscillations try to change momentarily to higher voltage
transient bursts of other F and as square waves.
The zener+diode clamp stops Vo exceeding about 5.7V which is enough to prevent
most Vo bounce. The amp I have uses +/-20V rails and transient Vo peak V can be
up to +/- 12V.
The circuit must not generate excessive gate bias quickly which then subsides
much too slowly to allow oscillations to continue. All operations MUST be free of
delays due to troubles with signal bounce and intermittent oscillations, or delays
between stopped and re-starting.

The major troubles you never see mentioned or quantified in 99% of online
information about Wien Bridge oscillators are the issues of stability when changing F,
and the distortion generated by the darn little j-fet.
Now if you look closely to my above Fig 2 you will see I have a local shunt NFB
network used between drain and gate using R8, C8, R9.
At above 1Hz, the ß = 0.5, and so 1/2 of whatever Vd exists is applied to the gate,
so any THD generated by the j-fet is half applied to the gate, thus it tends to be
amplified to reduce itself. Now the circuit works without R8, C8, R9, and typical
THD will be 2%, mainly 2H, because the naked Rd without NFB is not a linear
resistance but one which generates up to 10% 2H right at the drain.
Part of this feeds into the NFB port and makes the maybe 2% of 2H at Vo.

Now you may think the use of R8, C8, R9 changes the ß but they make negligible
difference to the range of values wanted for Rd because that range is between
about 200r and 1k0, and the loading effect is negligible. The R8 and R9 are 10k
each, and at very low F where C11, C12 are high Z, the value of R17 needs to be
about 10k to prevent excessive FB at LF which makes LF stability very bad because
you have a high open loop gain amp also involved and, where possible, shit happens
that you cannot forecast. I had R12 = 100k at first and LF stability was terrible, but all
calmed down when I made R12 = 10k. But then you need to have sufficiently long time
constants for oscillation down to 1Hz so C11 had to be 220uF and C8, also needed to
be 220uF.
The other needed thing about the NFB loop is that it controls the amp Vdc output offset.
Ideally, Vdc at Vo should = 0.0Vdc, and pigs will fly before you get that. BUT, I have used
fets for the amp inputs which have high Z in and thus I get away with using the
C4 2u2 + R7 1M0 network to allow Vo DC to reach Q2g, without being mauled by the
rest of the NFB network. The loading effect of 1M in parallel with R12 2k7 is negligible.

With R8, C8, R9, the THD drops to less than 0.03%. But other factors affect THD.
The limiting voltage clamp diodes begin to conduct on both + and - wave crests at
Vo when Vo reaches +/- 4.9Vpk.
Such wave crest suppression causes some slight 3H++ to be generated. But in my unit
I ended up getting THD = 0.09% at 1kHz, mainly 3H. This is much better than having
2% 2H, and is much better than many other signal generators and function generators
using chips to generate a fabricated sine wave.

I also tried 6 x series 1N914 in 2 directions ( 12 diodes total ) to act as a voltage clamp
plus VR pot across R12 2k7 but the sharp threshold action of diodes causes high 3H
at maybe 5%, so that idea is hopeless.

I tried using a few 12V x 50mA lamps in series and a variable DC current bias with pnp
bjt from +20Vdc. This made what I thought could be a good variable R for NFB control.
Two more bjts and a rectifier were needed and it looked good in theory but amplitude
stability was appallingly bad, so I quit that idea. THD was about the same as the j-fet
with local FB though.

Now do you see how much everybody else online in the world does not say about their
published electronic junk?
There is always more than ONE simple thing to consider about any single R, C,
or any point of connection anywhere around any analog circuit.
But at least we can get an idea, something you cannot do with a multi-layered PCB
with zillions of multi-pin ICs, all operating digitally.

The Fig 2 shows the whole voltage amp and it looks like a good quality solid state
audio power amp. Indeed it could be a power amp but the load to be driven is a
much higher ohm value so small signal bjts can be used throughout and with
no heat sinks.

There are 4 loads on the amp, in parallel.
1. There is the NFB R network of R10, VR1, R6, and Rd of j-fet.
2. There is the output attenuator, minimum 772 ohms.
3. There is the PFB network of switched R+C.
4. There is the series resistance R1 5k6 Fig 4 from amp to Schmitt trigger square
wave circuit.

1. The NFB resistance network is always 1.5 x R12 = 1.5 x 2k7 = 4,050r.
Why? Because when Vo is stable and 3.66Vac, Q2 gate has 1/3 of Vo present,
or 1.22Vrms. So there is 2.44V across 2k7, and I = 0.904mA so load = Vo / I = 3.66 / 0.904
= 4.05k.
The PFB signal at Q1 gate may be 1.221Vrms and is only just above Q2 gate.
The gate to gate voltage is extremely low where open loop amp gain may be say
40,000 which would make Vg-g = 0.09mV, and quite difficult to measure.

2. The output attenuator load is VR2, 772r and it came from my junk box, and
seemed to have no serious faults so common in old pots caused by dust, dirt,
pollution, corrosion, loose terminals, intermittent tracks.
I always clean out the pot with white spirits and plenty of turning actions then blow
out with compressed air and then compress rivets holding lugs to tracks using long
nose pliers in a vice. I seal up the hole near the lugs with cardboard and tape and
THEN such pots give many more years of trouble free operation.
Most cheap log pots you buy won't be exactly 1k0, and somewhere between 700r
and 1k2, and mine is 772r.
If I ever change to another "1k0" replacement LOG pot, a parallel trimming R will
be needed to ensure the total nominal R value = 772r. If the R is not 772r, the switch
S2 attenuator with its set resistances cannot give correct -10dB reductions of output
signal amplitude because the attenuator R values are calculated for only 772r.
-10dB is a nicely convenient reduction of output level. It means Vo is reduced by
factor 1 / square root 10, or x 0.3162. This is near 1/3. Two such reductions of x 0.3162
give a reduction of x 0.1, and you will eventually find this engineers' way of reducing
outputs and inputs is a pleasure to use.

3. The positive FB circuit is the Wien bridge RC network. For where the WB
network has equal two value C and R, the input load impedance becomes
2.121 x resistance value, and because R is varied between R for highest F
and 10xR for lowest F, the load we need to consider is the lowest because
it needs the highest current and in this case R = 1,590r, and load = 2.121 x 1,590
= becomes 3,372r. So where did the constant 2.121 come from?

The factor 2.121 is the factor calculated from 1.414 x 1.5. Huh? Why?
In all Wien R+C bridges with two equal R and two equal C, the output signal
response is that of a low Q tuned circuit and Vo can only ever rise to 0.333 x Vin
which occurs at only one frequency, that of "resonance" Fo, where phase shift
between Vin and Vo = zero degrees. See Fig 3 below.

Fig 3.

From the above basic Wien bridge when operating at Fo we see Vin = 10Vrms and
V0 = 3.3V. So that 6.67V exists across R1 in series with C1. And we need to know that
phase shift between Vin and Vo = 0 degrees, lest all our statements about measured
voltages are wrong.

At 1kHz Fo, reactance XC = R ohms, and impedance, Z for R and C in series
= 1.414 x R = 2,248 ohms. Therefore current flow = 6.67V / 2,484r = 2.966mA.
Therefore impedance looking into the input = 10V / 2.966mA = 3,370 ohms.
The result validity depends on the the 0 degrees of phase shift between Vi and Vo.
The same 2.966mA flows through the Z of R2 and C2 in parallel, and may be calculated
as 0.707 x R = 1,124r, or we may calculate from observations, Z = 3.3V / 2.966mA = 1,123r.
So total Z between Input and 0V = 2,248r + 1,123r = 3,371r.
The simplest way to calculate Z in for Wien bridge network at Fo is Z in = 2.121 x R.
For decreasing frequencies below Fo, the Z in rises towards infinity at 0.0Hz
because the C become open circuits.
For F increasing above Fo, Z in reduces to a minimum of  R, because the C become short
circuits at HF.

4. There is a feed resistance from oscillator amp and square wave circuit, Fig 4, R1, 5k6.
The voltage at the Q9 base end of  R1 is less than 1/5 of Oscillator output so in fact
effective loading is about 6k2 from this resistance of 5k6.
( See how there is always more than ONE thing to consider about any single part in
an electronic circuit! ).

So, we have 4 loads in parallel = 4,050r // 772r // 3,371r // 6,200r = 501r.

The amp must be able to power the load with little THD caused by loading.
I like my signal generators to make +/-5V peak at least, and in this case 3.66Vrms.
So load current at 3.66V = 3.66V / 501r = 7.3mArms. It could easily be possible to
have idle DC current in Q6 and Q7  to give pure class A operation if Ic = 11mAdc.
But this is not necessary, and the Q6+Q7 may operate in class AB at low enough
THD because the NFB acts to reduce distortion in the amp which has very wide
open loop bandwidth at 3.66Vac thus allowing 1MHz without any attenuation.
But anyone is free to have Idc = say 20mAdc, and perhaps give better HF stability.
I found that capacitive loads on the amp may cause oscillations cease but very little
open loop HF gain reduction was needed with Zobel R+C between
Q4 base and Q4 collector, see R11and VC1, Fig2.

The Vo from the sine wave oscillator in Fig 2 is prone to poor HF performance
especially if there is even a small amount of C, say 100pF between the output side of
R16, 33r, and 0V. This is enough to cause phase shift and HF oscillations above 300kHz.
But the shown switched attenuator S2 and pot VR2 cause negligible C shunting
themselves. To protect the oscillator amp from effects of C shunting of signal due to
cables or C in of gear and from low Rin of gear connected, I decided a buffer output
stage was needed.

For a rugged buffer with wide bandwidth,
I used npn PN100+BD139 and pnp PN200+BD140 connected as a complementary
emitter follower pair which gives gain close to 1.0. Its Rout is set by 22r emitter resistors
and the 47r output resistance. A zener diode + diode 10V clamp prevents high V from
something external to the unit providing a high voltage signal back into the unit output
which would instantly fuse the bjts. The 47r is 1/8 W rated, and fuses open if high voltage
enters the unit. Diode clamps are also placed from 22r junction to rails for further
limiting of stray unwanted back flows of current, like when you accidentally touch an
RCA lead on an anode at +400Vdc.

The buffer works in class A with all loads with Z above about 400r. Now that allows 400pF
as a load at 1MHz.
Class AB action occurs with lower loads. A shorted output with maximum Vo of 3.6Vac
being generated causes a peak current of 74mA, and heats the 47r, but won't blow it, and
causes about 0.5W of heat in each of BD139, BD140 which won't kill them.  Notice the R36,
R37, both 680r emitter resistors for drivers Q9 and Q10. These R make the idle current in
Q9 and Q10 larger than if it was only the base current of Q11 and Q12, and give a path for
discharging turn off current from bjts at HF, something bjts don't do so well.

With the attenuator S2 and VR2 placed before the buffer, there is no attenuation of HF
content from the sine wave oscillator and the HF content of the square wave appears well
preserved. I found the square wave at 1MHz to look well compared to many other units I
have used, and its highest F appears to be 10mhz, and the rise time was calculated to be
104V/uS.

However, most good audio amps are not designed to operate above 65kHz, -3dB pole,
with pure R load for the amp, and the response may be measured with test sine waves of
1Hz to 200kHz. However, to examine reactive behaviours and transformer resonances
it is often useful to have sine wave signals up to 1MHz, or beyond, especially where mosfets
are used in an SS amp which may oscillate at 3MHz+, intentionally, or not intended, or
when provoked into oscillations with a fast enough square wave.

----------------------------------------------------------------------------------------------------------------------------
The Schmitt Trigger square wave circuit in the original BWD 141 had 2 x BC557 pnp
transistors with feeble Idc flow. The good "squareness" of square waves disappeared
above 100kHz, and at LF.
The original 141 had a very low square wave output level and squareness above 100kHz
was poor, ie, much HF content removed by the attenuator network and following cable
capacitance.

I searched the Internet, and found many examples for Schmitt Trigger sine to square
wave converters. None of the pages had any mention about the troubles ppl face with
distortions and odd behaviours before they get a good square wave. I tried a few online
circuits with npn bjts, and then with j-fets, and mosfets, but the npn bjts were just fine using
BC546 or PN100 or 2N2222. During my trials of the devices I played around with variations
on R+C values until I eventually found further experimentation pointless, unless I was to
try something one never sees, such as a pair of Schmitt Trigger circuits with two pairs of npn bjts
and two pairs of pnp bjts wired in parallel with cap coupling. But I did get good enough
performance with the circuit below......

Fig 4.
Schmitt trigger.

The input to the Q13 base network via the above R1 5k6 is permanently
connected to the sine wave amp output in Fig 2, ie, output side of R17, 33r.
The square wave converter works continuously while ever the whole unit is
turned on, so that when switching from sine wave to square wave there is
minimum delay while Vdc rails and working voltages establish themselves
because long time constants are involved.

When switching S1 Fig 2 to sine wave, there is no sign of any leakage of
square wave gate crashing the sine wave party.

As sine wave F rises above 500kHz, there are some difficulty getting enough
energy to trigger the base input at Q13 and Q14 to maintain square wave
symmetry.
To ensure enough triggering voltage at HF appears at Q9 base to give symmetry,
R1a 150r and C1a 100pF have been added as a Zobel network across R1 5k6.
This was found to have a negligible extra loading on the sine wave amp at HF.

The Q15 and Q16 complementary pair emitter follower buffer overcomes the
problem of attenuation of HF on positive going V rise at R9, 1k1.
The 1k1 is not a low enough resistance to overcome any stray C in following
attenuator network, and the attenuator network has low minimum R in of 770r,
so the buffer IS NEEDED.

The Fig 4 schematic has 470uF bypassing R6 15k. In most other schematics
online, R6 is not bypassed, or bypassed with a small value of C, perhaps 100pF.
But after trialling several values, I concluded 470uF was fine.
With no bypass C or low values of C, the delivery of positive FB from Q9 collector
to Q10 base is rather too slow. And I found the square wave's lower horizontal
had a large kink, so the wave looked wrong, and unacceptable.
The RC coupling as I have it works best.

VR3 and R7 must be adjusted carefully to get the wanted level of +/- 5.17Vpk
AND good wave symmetry where both top and bottom horizontals of the square
wave have equal length, and time between verticals is equal. The range of VR3
pot movement needed to get square waves to occur at all is rather small.
The R7 value was adjusted with a temporary pot to get the wanted peak Vo while
VR3 was adjusted for best symmetry as R7 is adjusted.

Nobody else tells you that you MUST adjust these two resistors together
because they are somewhat interactive, and beginners will be frustrated when
they build something they saw online and it fails to work and they stare at it for
10 years without knowing what questions to ask or what to try to get it working.
When using 10kHz sine wave input you should get a splendid looking square
wave and you should find all other F give the same peak Vo and good symmetry.

I have no provision for applying DC offsets at the output, so that in effect a
positive or negative bias voltage cannot be included in the total Vo signal.
Nor can you vary the asymmetry for a weird looking square waves.
I have never needed Vdc offsets, and have always found that having Vdc
content in signal gene outputs is a royal pain in the arse when testing something
where you WANT ONLY 0Vac at the input to something you are testing, and if
you do want some DC, then make up a suitable C+R coupling circuit and add the
Vdc bias manually. If there is to be any biasing of amps under test, it will be done
within the amp, which should not have any Vdc present at its input, say from a
grid which is drawing some slight grid current at idle, and every time you switch
sources you get a loud click.

I do not propose to add more boards for a "triangle wave" or saw tooth wave, or
for AM or FM modulation. To test audio amps, there is no need for more than what
this unit now does, ie, give sine and square waves up to 5.15Vpk level, 1Hz to
1MHz, and with no Vdc content.

For the extra functions an additional box could be made with a suitable circuit you
will have to research and develop after perhaps beginning with something you'll find
online but which probably will not be fully explained properly, or which just won't
work very well. A good triangle wave with straight lines can indicate very basic
linearity of an amp because you can see when the straight lines get a bend if you
have a dual trace CRO.
F distortion can occur, ie, "integration", ie, further action by C&R or L&R circuits
within the tested amp. For measuring THD, it is better to use a dedicated low THD
sine wave and following notch filter as described elsewhere at this website,
thd-measurement.html

Because of long time constants used in the unit, Vdc settlement after turn on
takes about 20 seconds before 0Vdc appears at the output with negligible Vdc
measured. Some variation of the Vdc offset will occur with use and switching F
ranges or from sine to square.
Vdc will eventually settle. Most good amps have a C&R high pass filter at their
input such as say 0.33uF plus 68k, giving LF pole at 7.1Hz. The time constant
= 0.022 seconds. This will cause serious distortion of LF square waves. But you
will find the distortion where square wave horizontals become highly sloped at
below 30Hz is of little concern. With this oscillator unit you can place a bypass
link across the 0.33 uF and then the time constant is set by C24 200uF and 68k
plus R41 47k at unit Vo thus giving time constant of 5.5 seconds, and LF pole at
input at 0.028Hz. The distortion of LF square waves is then negligible.

At the unit's Vo, the R41 47k, gives a long time constant = 9.4 seconds when
unloaded. So where you have a tube amp with 0.22uF and 470k at input, the
TC = 0.103 seconds, LF pole = 1.53Hz and you will still get square wave distortion
at the input. If the 0.22uF was bypassed with say 22uF NP cap, then you get a T
C = 10.3 seconds and square wave at input will keep its integrity.
However, tube amps have maybe 3 internal C&R couplings and at least one R+L
OPT coupling which all cause trouble with square wave inputs. The NFB used will
correct much of the errors in square waves that are inevitable. But do the so called
errors with square waves at LF matter?
Fortunately, nobody has proved the C+R and R+L coupling causes bad LF sound,
and in fact well made tube amps just cope, and sound is excellent.

What is important is the HF square wave performance of any amp and its stability
with C loading or without any load at all.
The amp under test should be able to give a good looking reproduction of square
waves between 100Hz and 5kHz.
There should be very little ringing or overshoot with no amp load or with a pure R
load of nominal value +/- 100%.
So something designed for 8r0 should work OK with 4r0 or 16r0 even though less
max Po is available.
Some overshoot may be allowed to occur with solely C loads between 0.05uF
and 10uF, and where Vo is kept low to avoid overload.

Many amps fail when tested with square waves and they can oscillate badly at
some RF when no load is connected or when a pure C load is connected.
Most dynamic speakers are inductive at HF and many do not have a Zobel
network at their X-over filter to ensure there is a nominal R load when F rises
above 20kHz, such as say 8r2 + 0.47u.

The beauty of square wave testing is that you can adjust the the C value used
to bypass the FB resistance from amp output to its FB input port, and adjust the
HF gain and phase change of the voltage amp to ensure HF gain at F above
20kHz is reduced, along with phase shift so that FB remains negative, and not
positive, even when C loads are used. A CRO and a good signal gene are
essential tools for getting any amplifier to work well as it can, and give optimal
bandwidth and complete unconditional stability when NFB is used.
In amps without NFB, the signal gene is still essential to make sure sine wave
response is as wide as wanted. Without NFB, most amps have no stability or
oscillation problems, but the CRO will tell you their transient response is much
poorer than their brothers with well designed NFB networks.

Back To Education and DIY

Back to Index