A whisper riding on a shout
Picture exactly where guide 2 left you. A Wheatstone bridge is excited by, say, 10 V, and a load cell strains its arms ever so slightly. At full load the bridge hands you a difference of only about 20 mV between its two output corners — a genuine whisper. The trouble is what that whisper is sitting on. Each of those two corners hovers near 5 V, half the excitation, because each is the midpoint of a divider across 10 V. So the wires carry roughly 5 V of voltage they share, with your precious 20 mV of difference perched on top. Worse, both wires run side by side through a noisy world and pick up the same 50/60 Hz mains hum, again shared equally by both. This is the heart of signal conditioning: the information lives entirely in the difference between the two wires, and everything they hold in common is to be thrown away.
This is why you cannot just grab one corner of the bridge and feed it to a non-inverting amplifier referred to ground. Such a single-ended amplifier would faithfully magnify the whole 5 V resting voltage and the full hum, drowning the 20 mV you actually want under a hundred times more rubbish. What you need is an amplifier with two inputs that computes the difference between them and amplifies only that — a difference amplifier. Subtract first, amplify second: the 5 V common rest voltage cancels because it is identical on both inputs, the shared hum cancels for the same reason, and the 20 mV difference survives to be made large.
Common-mode rejection: the one number that matters
No real amplifier ignores the common-mode part completely. It has a large differential gain Ad (the gain we want) and a tiny, unwanted common-mode gain Acm (the leak we don't). The figure of merit that compares them is the common-mode rejection ratio (CMRR): the ratio Ad / Acm, almost always quoted in decibels as CMRR(dB) = 20 times log10(Ad / Acm). A bigger number means the amplifier is better at telling your signal apart from the shared muck. A jellybean op-amp manages perhaps 80 dB; a good instrumentation amplifier reaches 100 to 120 dB, meaning it amplifies the difference a hundred thousand to a million times more eagerly than the common-mode.
Make it concrete on our load-cell bridge, amplified by gain Ad = 100. The 5 V common-mode mid-rail leaks to the output as Acm times 5 V. With CMRR = 100 dB the ratio is 10^5, so Acm = Ad / 10^5 = 100 / 100000 = 0.001, and the leak at the output is 0.001 times 5 V = 5 mV. Referred back to the input that is only 5 mV / 100 = 50 uV of error against a 20 mV signal — under a third of a percent, livable. Now imagine an amplifier with only 60 dB of CMRR (ratio 1000): Acm jumps to 100 / 1000 = 0.1, and the same 5 V common-mode dumps 0.1 times 5 V = 0.5 V straight onto the output, swamping the 2 V of real signal with a 25 percent error. The same circuit, the same bridge — only the CMRR changed, and it decided whether the measurement was usable.
Why one op-amp isn't enough
Back in the op-amp rung you built a difference amplifier from one op-amp and four resistors: the signal arrives on both inputs, two resistors set the gain G = R2/R1, and if the four resistors are perfectly matched the common-mode cancels and Vout = G times (V+ minus V-). It is elegant and cheap, and for a stiff, low-impedance source it is exactly right. Hung straight onto a sensor bridge, though, it has two flaws serious enough to disqualify it.
The first flaw is loading. In that one-op-amp circuit the bridge does not look into the near-infinite gate of an op-amp; it looks into the resistor network, so each input presents only a modest, finite input resistance — and, galling for a difference amp, the two inputs do not even present the same resistance. Connecting it to the bridge therefore draws current from the bridge corners, and because the two corners are loaded unequally, it actually unbalances the very bridge you are trying to read. That is the loading effect from the early rungs, biting at the worst possible place.
The second flaw is that its CMRR rests entirely on how well those four resistors match. A handy estimate is CMRR is about (1 plus G) / (4 times t), where t is the fractional resistor tolerance. Build it at unity gain from ordinary 0.1 percent resistors and you get only about 2 / 0.004 = 500, a mere 54 dB — the 25-percent-error disaster from the last section. Pushing past 100 dB would demand resistors matched to a few parts per million, and even then changing the gain means re-matching all four. The cure is to stop relying on hand-matched resistors and let a chip do it — which is exactly the instrumentation amplifier.
The instrumentation amplifier: three op-amps, done right
The instrumentation amplifier (in-amp) fixes both flaws by putting a buffer in front of the difference amp on each input. Each sensor wire goes straight into the high-impedance non-inverting input of its own op-amp, so the in-amp draws essentially no current from the bridge — the loading problem simply vanishes, and because the two input stages are identical they load the bridge equally. Those two input buffers feed a one-op-amp difference amp at the back that does the actual subtraction, with the CMRR-critical resistors trimmed by laser on the silicon, far beyond anything you could match by hand.
THREE-OP-AMP INSTRUMENTATION AMPLIFIER
IN+ o-->[+ \ R R
| A1 >--+-------[R]--+----WWW--+--WWW--+
+[- / | | | |
| | (Va) | \ |
[Rg] <-- ONE resistor | A3 >-+--o OUT
| sets the gain +----| / |
+[- \ | | | + |
| A2 >--+-------[R]--+----WWW | REF
IN- o-->[+ / R +--WWW--+
R (tied to REF)
Stage 1 (A1, A2): matched buffers, NO bridge loading
differential gain = 1 + 2R/Rg
Stage 2 (A3): unity difference amp, kills the common-mode
Overall: Vout - VREF = (1 + 2R/Rg) x (V+ - V-)Here is the prettiest part. The whole gain is set by one external resistor, Rg, wired between the two input buffers, following G = 1 plus 2R/Rg where R is the pair of equal internal feedback resistors. Crucially, Rg sits in the differential path only — it does nothing to the common-mode, which both buffers pass at unity to be cancelled later. So you dial in any gain with a single resistor without touching the matched network, and the CMRR stays intact. One cheap resistor controls the gain; the expensive, carefully-trimmed part stays untouched. That clean split is the whole reason the in-amp exists.
- Find the full-scale signal from guide 2. Our load cell is rated 2 mV/V and runs on 10 V excitation, so full load gives 2 mV/V times 10 V = 20 mV differential between the corners.
- Decide the output swing the next stage wants. To feed a converter whose input range is roughly 0 to 2.5 V, aim for about 2 V full scale to leave headroom. The gain you need is G = 2 V / 20 mV = 100.
- Solve for Rg. With a common internal network of 2R = 49.4 k, set G = 1 + 2R/Rg = 100, so Rg = 49.4 k / (100 - 1) = 49.4 k / 99 = 499 ohm — a standard 1 percent value, no matching required.
- Check the bandwidth you have left. Gain and bandwidth trade off through the gain-bandwidth product, so at G = 100 the usable bandwidth is roughly the part's gain-bandwidth divided by 100 — plenty for a slow load cell, but a warning if you ever need both high gain and high speed.
Dropping it into a real front end
Notice the REF pin in the schematic: the in-amp's output is measured relative to REF, not to ground. That is a quiet gift. Running on a single supply, your amplified signal would otherwise want to swing below 0 V for negative readings and clip. Tie REF to a clean mid-supply reference instead and the whole output floats up around that point, so it can swing both ways inside the rails. That is level-shifting done for free by the part you already have — and pre-positioning the signal for the converter, together with the anti-alias filter that must follow before any sampling, is exactly the work of guide 4.
Respect the in-amp's real limits, all of which live in the datasheet. The common-mode input range must actually contain your Vcm — our 5 V mid-rail has to sit comfortably between the supply rails, or the input buffers saturate and the rejection collapses. The gain error and gain drift (the internal R and your Rg are never perfect and wander with temperature) mean the 100 you dialled is really 100 give-or-take a fraction of a percent. And as the last section warned, CMRR sags with frequency. None of these is a defect to be shocked by; they are the numbers you budget for, the same disciplined way the op-amp rung budgeted offset and bias.
End on the honest framing this whole rung keeps returning to. A trick worth knowing is ratiometric measurement: power the bridge and the converter's reference from the same supply, and any drift in the excitation cancels out in the ratio, so a 5 percent supply wobble does not become a 5 percent error. But the deepest truth is this — the accuracy of the final number is set by the sensor and this conditioning chain, not by the ADC that digitises it. A 24-bit converter cannot recover a signal already buried under poor CMRR, an unbalanced load, or gain drift. Spend your care here, at the front end; that, plus the calibration of guide 5, is what actually decides whether you can trust the reading.