Load Cell Amplifier Design: Sigma-Delta ADC vs Instrumentation Amplifier vs Bridge IC

The platform scale in the broiler shed reads 1,842 grams, then 1,846, then 1,839, with the same bird standing still on the pan. By mid-afternoon, when the shed has warmed by ten degrees, the empty pan reads 31 grams. The farm manager does not care about amplifiers. He wants to know whether his birds gained weight this week, and the display cannot tell him.
Nothing is broken. The load cell is doing its job. A cell rated at 2 mV/V and excited at 5 V hands the electronics ten millivolts at full load, whether it is rated for 2 kg or 50 kg. Everything a weighing product does afterwards, from a kitchen scale to a hopper to a poultry platform, is a decision about what to do with those ten millivolts and the noise riding on them.
Most teams pick the amplifier last, after the enclosure and the display, and pick it by price. Then the display flickers, the reading creeps when the shed warms, and someone suggests a bigger filter. The better order is the reverse: decide how many scale divisions you must resolve, convert that to microvolts per division, and choose the cheapest front end whose noise at your data rate is at least four times smaller than one division. This post works through that choice for the three common front ends: a sigma-delta converter with a built-in amplifier, a separate instrumentation amplifier, and a low-cost bridge IC.
What a Load Cell Hands You: Ten Millivolts and a Few Opinions
A strain-gauge load cell is a Wheatstone bridge. Four resistive gauges, bonded to a metal spring element, change resistance by a fraction of a percent when the element flexes. The bridge converts that into a differential voltage proportional to the excitation you apply. The cell datasheet states the result as a sensitivity in millivolts per volt of excitation at rated capacity. Common values are 1 mV/V, 2 mV/V and 3 mV/V; the Analog Devices application notes for their weigh-scale converters use the 2 kg, 2 mV/V, 5 V case that gives 10 mV of full-scale signal.
Three things about that signal shape every later decision.
First, it is small and differential. The output rides on a common-mode level of half the excitation voltage, so the amplifier must reject 2.5 V of common-mode while resolving microvolts of difference. That is why a single-ended ADC input cannot read a load cell directly.
Second, it has an offset. A load cell has a zero-load output, usually called the tare offset or zero balance, that the datasheet allows to be a few percent of full scale. On a 10 mV cell that can be several hundred microvolts with no weight on the pan. Your front end needs headroom for it, or a way to trim it out, or the first gain stage saturates before the scale is even loaded.
Third, it is ratiometric. The output is proportional to the excitation voltage, so if the excitation drifts by 0.1 percent, the output drifts by 0.1 percent. The trick that makes good weighing electronics possible is to make the ADC reference follow the excitation, so the drift cancels. We will come back to that, because it matters more than the gain setting.
Here is the arithmetic that sets the problem. The table gives the signal per scale division for the same 10 mV cell at three division counts.
Scale divisions | Microvolts per division | What it means in practice |
3,000 | 3.33 µV | Kitchen and counter scales, modest front end is enough |
10,000 | 1.00 µV | Upper end of OIML class III trade scales |
30,000 | 0.33 µV | Laboratory and high-resolution industrial work |
A division is the smallest step your display shows. If your noise is 1 µV peak to peak and your division is 1 µV, the last digit flickers constantly. If your noise is a quarter of a division, it does not. That four-to-one ratio is a good working rule, and the rest of this post is how to get there.
Three Ways to Amplify It, Side by Side
There are three architectures in common use, plus a fourth that is really a variant of the third. Each is correct for some product and wrong for another.
Aspect | Discrete in-amp plus ADC | Sigma-delta ADC with PGA | Low-cost integrated bridge IC | Microcontroller with built-in PGA |
Typical gain stage | Instrumentation amplifier, gain set by one resistor | Programmable gain up to 128 inside the converter | Fixed or two-step gain inside the chip | Programmable gain, usually 16 to 64 |
Resolution you can trust | 16 to 20 bits after careful design | 20 to 22 noise-free bits at low data rates | 16 to 19 noise-free bits at its slowest rate | 12 to 16 bits |
Ratiometric reference | You must build it | Reference inputs are differential, so 6-wire sensing is direct | Reference tied to the supply, so no sense lines | Depends on the part, often supply only |
Mains rejection | You build the filter | Digital filter with notch options | Fixed rate, limited rejection | Usually none, filter in software |
Board area and parts | Largest | Small | Smallest | Smallest, but the MCU is on the board anyway |
Cost tier | Highest | Mid | Lowest | Lowest if the MCU is already chosen |
Read the table the way a buyer should: the left two columns buy you precision, the right two buy you simplicity. The question is how much precision the product needs.
A discrete instrumentation amplifier followed by a separate ADC is the older design. It gives you the most control, and it was the only choice before weigh-scale converters existed. Its weak points are offset drift in the amplifier, the resistor matching that sets common-mode rejection, and the need to build your own ratiometric reference. It still wins when you need a very wide supply range, an unusual cell, or an ADC you already have for other channels.
The sigma-delta converter with an on-chip programmable-gain amplifier is the standard for serious weighing. Analog Devices' application notes AN-2505 and AN-2506 describe weigh-scale reference designs built around parts of this kind, with a PGA gain of 128, differential reference inputs, selectable data rates and a digital filter. The gain, the reference and the filter all live in one package, tuned together.
The low-cost integrated bridge IC, sold on small breakout boards by the thousand, trades the flexibility away. You get a fixed data rate, a reference tied to the supply, and no sense lines. For a kitchen scale or a counting scale that is often fine. For anything that must hold calibration through a temperature swing, it is where the problems start, and we will say exactly where.
The fourth column deserves a sentence of its own. Some microcontrollers now carry a PGA and a 16-bit sigma-delta ADC. They are tempting because they remove a chip. They are adequate for 2,000 to 3,000 divisions, and they are noisy neighbours: the digital activity on the same die couples into the analog input. Test before you commit.
A breakout board that works on a bench in January can fail in a shed in May. Judge a front end by its behaviour across temperature and cable length, not by its first reading.
Counting Divisions: How Much Noise Your Scale Can Afford
Noise is the number that decides everything, and it is quoted in confusing ways. Datasheets give RMS noise in nanovolts for each combination of gain and data rate. A display, though, flickers according to peak-to-peak noise, and for Gaussian noise the peak-to-peak value over a few seconds is about 6.6 times the RMS value.
The number of divisions the front end can support is then the full-scale signal divided by that peak-to-peak noise. Call it noise-free divisions. A 10 mV signal with 40 nV RMS noise has 264 nV peak to peak, which supports 10 mV divided by 0.264 µV, or about 37,900 divisions. To use those divisions on a display you want a margin, which is where the four-to-one rule comes from.
The next table applies the formula to an illustrative converter. The noise figures are placeholders of the right order of magnitude for a 24-bit sigma-delta part at PGA gain 128 with a 5 V reference; read your own part's datasheet table, because each converter and filter setting differs.
Output data rate | RMS noise (illustrative) | Noise-free divisions at 10 mV | Divisions at a four-to-one margin |
10 SPS | 40 nV | 37,879 | 9,470 |
50 SPS | 90 nV | 16,835 | 4,209 |
100 SPS | 130 nV | 11,655 | 2,914 |
500 SPS | 300 nV | 5,051 | 1,263 |
1,000 SPS | 450 nV | 3,367 | 842 |
This table explains most of the arguments about weighing electronics. A product that needs 10,000 divisions at a four-to-one margin needs noise-free resolution of 40,000 divisions, and in this illustration only the slowest rate comes close. That is the true cost of speed: every tenfold increase in data rate costs you roughly a factor of three to four in noise-free resolution, because noise bandwidth rises with rate.
Two practical consequences follow.
If the scale only needs to settle in a second or two, run the converter slowly and average. A 10 SPS rate with a short moving average gives you a stable display without any clever tricks. Teams that run at 1,000 SPS and then filter in software are paying for noise they then spend CPU cycles removing, and the software filter rarely matches the converter's own digital filter.
If the scale must be fast, such as a checkweigher with a belt moving at one metre per second and a fraction of a second to measure, the noise table tells you plainly how many divisions you can have. Often the honest answer is fewer than the customer asked for, and it is better to tell them on day one.
The Python below is the function we use for this calculation. It was run with the values in the table and the asserts passed.
import math
def fs_signal_mv(sens_mv_per_v, excitation_v):
return sens_mv_per_v * excitation_v
def counts_per_division(fs_mv, divisions):
"""microvolts of bridge signal per scale division"""
return fs_mv * 1000.0 / divisions
def noise_free_divisions(fs_mv, noise_nv_rms, pp_factor=6.6):
"""full scale over peak-to-peak noise"""
pp_uv = noise_nv_rms * pp_factor / 1000.0
return fs_mv * 1000.0 / pp_uv
def lead_resistance_error(bridge_ohm, cable_len_m, area_mm2, rho=1.72e-8):
"""4-wire cell, reference tied at the cell end: excitation drop in the cable"""
r_per_conductor = rho * cable_len_m / (area_mm2 * 1e-6)
r_total = 2 * r_per_conductor
return r_per_conductor, r_total, r_total / (bridge_ohm + r_total)
def sinc_attenuation_db(f, odr, order=3):
"""ideal sinc^N filter, first null at the output data rate"""
x = math.pi * f / odr
if abs(x) < 1e-12:
return 0.0
h = abs(math.sin(x) / x) ** order
return 20 * math.log10(h) if h > 0 else -300.0
# 2 kg cell, 2 mV/V, 5 V -> 10.0 mV; 10,000 divisions -> 1.0 uV each
# noise_free_divisions(10, 40) -> 37879 ; noise_free_divisions(10, 450) -> 3367
# lead_resistance_error(350, 10, 0.2) -> (0.86 ohm, 1.72 ohm, 0.489 %)
Four Wires, Six Wires, and What a Ten-Metre Cable Does to Your Gain
A four-wire load cell has two excitation wires and two signal wires. A six-wire cell adds two sense wires that connect to the high and low sides of the bridge itself. They carry almost no current, so they do not drop voltage, and they let the ADC measure the voltage that the bridge actually sees.
With a four-wire cell, the ADC reference pins connect to the excitation supply and ground, typically at the board. The Analog Devices notes point out the consequence: the system is no longer completely ratiometric, because there is a voltage drop between the excitation supply and the bridge due to wiring resistance, and another on the ground side.
How large? Take a 350 ohm bridge, 10 m of cable with 0.2 mm² conductors, and copper at 1.72e-8 ohm metres. Each conductor is 0.86 ohm. The excitation current flows out one and back the other, so the loop is 1.72 ohm in series with the 350 ohm bridge. The bridge sees 0.49 percent less excitation than the reference does, and the reading is low by that amount.
That is a gain error, not a noise error, and it is fixed at the moment of calibration. If the calibration weight is applied with the same cable attached, the 0.49 percent disappears into the span adjustment. The trouble is temperature. Copper resistance rises about 0.39 percent per degree Celsius. A cable that swings by 40 °C changes its loop resistance by about 16 percent, which moves the gain error from 0.489 to 0.565 percent. That is a change of 0.076 percent of reading, or 7.6 divisions on a 10,000-division scale. A six-wire connection removes it.
Cable length (0.2 mm², 350 ohm cell) | Loop resistance | Gain error with four wires |
3 m | 0.52 ohm | 0.15 percent |
10 m | 1.72 ohm | 0.49 percent |
30 m | 5.16 ohm | 1.45 percent |
100 m | 17.2 ohm | 4.68 percent |
So the wiring decision is simple. For a cable under about three metres inside one enclosure, four wires with a good calibration is acceptable. For a platform scale in a shed with a 10 to 30 m run back to the indicator, use six wires. For 100 m, either use six wires and a front end that accepts them, or put the amplifier at the cell and send digital data back along the cable.
If you inherit a four-wire cell and a long cable, you can fake sensing by connecting the two sense inputs of your board at the cell end of the cable. That needs two extra conductors and gives you the same benefit. Many existing installations use a cable with spare cores for exactly this reason.
Excitation: The Load Cell Is the Biggest Power Consumer on the Board
A 350 ohm bridge at 5 V draws 14.3 mA and dissipates 71 mW. A 1,000 ohm bridge at 5 V draws 5 mA and dissipates 25 mW. On a mains-powered indicator that is nothing. On a battery-powered scale it can be the largest single load, bigger than the display and the radio.
Four choices reduce it.
Use a higher-resistance cell. A 1,000 ohm bridge draws 35 percent of the current of a 350 ohm bridge at the same excitation for the same millivolts per volt output. Many industrial cells come in 350 ohm; many small cells come in 1,000 ohm.
Lower the excitation. At 2.5 V the 350 ohm bridge draws 7.1 mA and dissipates 17.9 mW, a quarter of the power. The cost is that the signal halves, to 5 mV, and the noise of the amplifier does not halve with it. Your noise-free divisions fall by a factor of two. For a scale that needs 3,000 divisions this is often fine.
Pulse the excitation. Switch the bridge on a few milliseconds before the conversion and off after. The bridge takes microseconds to settle, but the sigma-delta filter needs its own settling time, so the saving depends on the data rate you chose. At one reading per second with a 20 ms on-time you cut the average bridge power by a factor of 50.
Use AC excitation if the converter supports it. Analog Devices' CN0601 reference design uses a converter whose drivers reverse the excitation polarity, which cancels offsets and thermoelectric voltages. The cost is a more complex firmware routine and a longer settling time.
Pulsing the excitation without waiting for the cell and cable to settle gives readings that look stable but sit lower than the true load by a few tenths of a percent. Measure the settling time at your cable length before choosing the delay.
The Sinc Filter and the 50 Hz Hum
The digital filter in a sigma-delta converter is a sinc filter, and it has a property that matters for India and anywhere with 50 Hz mains: it has nulls at multiples of the output data rate. If you set the output rate to 50 SPS, the filter places a null at exactly 50 Hz. At 10 SPS it also places nulls at 50 Hz, because 50 is a multiple of 10.
The arithmetic is simple. For an ideal third-order sinc filter, the attenuation at a frequency f for an output data rate R is the third power of the function sin(pi f / R) divided by (pi f / R). At f equal to R the numerator is zero and the attenuation is infinite in theory. Real converters reach 80 to 100 dB because of other noise sources, but the point is clear.
The mains frequency drifts by perhaps 0.5 Hz on a weak grid, so what matters is the attenuation a little off the null. Running the function above:
Output data rate | Attenuation at 50 Hz | At 49.5 Hz | At 50.5 Hz |
10 SPS | ideal null | 120 dB | 120 dB |
50 SPS | ideal null | 120 dB | 120 dB |
16.7 SPS | 162 dB | 115 dB | 126 dB |
60 SPS | 43 dB | 42 dB | 45 dB |
The last row is the trap. A reference design tuned for 60 Hz mains at 60 SPS gives only 43 dB of rejection at 50 Hz. A product built in India and copied from a US reference code will pick up the mains hum from the lighting, the motors and the cable, and the display will wander at a few counts. Moving to a data rate of 50 SPS or 10 SPS is a one-line configuration change.
When the scale must reject both 50 Hz and 60 Hz, as a product sold worldwide must, there are converter modes that place notches at both. Check the datasheet, because the notch modes cost data rate.
Mains pickup also enters through the cable, which acts as an antenna. The cable shield should be connected at one end only, usually at the indicator, and the input filter should be a differential RC with common-mode capacitors each side. The Analog Devices notes show filters on the analog inputs and capacitors on the reference pins for the same EMC reason. It costs a handful of small parts.
Where the Low-Cost Bridge IC Fits, and Where It Runs Out
For a kitchen scale, a hobby build or a one-off fixture, the low-cost bridge IC is hard to beat: one part, a well-known reference circuit, almost nothing on the bill of materials, and enough resolution for a display that changes in whole grams. The trouble starts when the same part is asked to do a product's job in a shed or a factory.
A bridge IC whose reference is tied to the supply measures your battery voltage as faithfully as it measures your load.
The same chip causes most of the field failures we are asked to diagnose.
Its reference is tied to the supply. There is no ratiometric measurement unless the supply that excites the bridge is also the supply that biases the converter, and then the cable drop and any regulator noise couple straight into the reading. The zero and span drift with the battery voltage unless you compensate in firmware.
In many parts of this class the data rate comes from an internal oscillator, which varies with temperature and supply. The mains notch you thought you had at 10 SPS can slide to 9 SPS in a hot shed and let 50 Hz through.
Its noise at the high-rate setting is far above its noise at the low-rate setting, and many designs run at the high rate because the datasheet's headline resolution is quoted at the low one. Check which rate the reference code uses.
Its gain options are coarse, so a 1 mV/V cell loses half the resolution that a 2 mV/V cell gives.
Its absolute accuracy is not specified in a way a metrology lab accepts. You can still use it in a trade scale if you measure and prove the performance, but you carry the burden.
None of these is a reason to avoid it. They are reasons to know which product you are building. If the scale is battery-powered, outdoors, wired with a long cable or intended for approval, the low-cost chip needs a lot of care and often loses to a proper sigma-delta part by the time the extra firmware is counted.
The 50 kg Platform That Flickered
The following is a representative composite, genericised, and the numbers are illustrative.
A team built a 50 kg platform scale for a livestock application. The cell was 2 mV/V, 350 ohm, excited at 5 V. The target was 10,000 divisions of 5 g. A low-cost bridge IC read the cell through a 12 m cable.
On the bench, the scale was beautiful. In the first shed it flickered by plus or minus two divisions, drifted by five divisions as the day warmed, and read about 0.4 percent low when the cable was extended from 12 m to 20 m.
The calculation was done after the fact. The signal was 10 mV, so one division of 5 g was 1 µV. The chip's peak-to-peak noise at its fast rate was about 1.5 µV, which supports 10 mV divided by 1.5 µV, about 6,700 noise-free divisions, short of the 10,000 target and nowhere near the four-to-one margin. That was the flicker. The cable loop was about 2 ohm at 12 m, giving a gain error near 0.59 percent, and about 0.97 percent at 20 m, a shift of 0.4 percent. That was the low reading. The drift came from a linear regulator near the connector, producing a thermocouple gradient across the input pads of a few tenths of a degree per minute, and from a reference that tracked the supply, not the bridge.
The fix was a front end change, not a filter change: a sigma-delta converter at 10 SPS, a six-wire connection, a differential RC on the inputs, a data rate placing a null at 50 Hz, and the regulator moved 30 mm away from the pads. Peak-to-peak noise fell to about 0.26 µV, which supports about 38,000 noise-free divisions and a margin of about 3.8 to 1 over the target. The flicker stopped, the drift shrank to under one division over a 30 °C warm-up, and the cable length no longer mattered.
It cost a few hundred rupees of components per unit more. The returns stopped.
From Counts to Grams: Calibration, Zero Tracking and Stability
A converter gives you counts. The product needs grams. The calibration is a two-point line through no load and a known weight, and it is worth getting right in firmware, because the display depends on it.
The example below is a small C routine compiled and run. It uses 64-bit arithmetic so that a 24-bit count times a gram span never overflows, rounds to the nearest gram, handles negative readings, tracks zero slowly only when the reading is stable and near zero, and decides stability from the spread of recent readings.
#include <stdint.h>
#include <stdlib.h>
typedef struct {
int32_t zero_counts; /* ADC counts at no load */
int32_t span_counts; /* counts at the calibration weight minus zero */
int32_t span_grams; /* calibration weight in grams */
} cal_t;
static int32_t counts_to_grams(const cal_t *c, int32_t counts)
{
int64_t num = (int64_t)(counts - c->zero_counts) * c->span_grams;
int64_t den = c->span_counts;
int64_t q = (num >= 0) ? (num + den / 2) / den : (num - den / 2) / den;
return (int32_t)q;
}
/* creep the zero toward the reading only while stable and near zero */
static void track_zero(cal_t *c, int32_t counts, int32_t grams,
int32_t window_g, int stable)
{
if (stable && abs(grams) <= window_g)
c->zero_counts += (counts - c->zero_counts) / 16;
}
static int is_stable(const int32_t *h, int n, int32_t limit)
{
int32_t lo = h[0], hi = h[0];
for (int i = 1; i < n; i++) {
if (h[i] < lo) lo = h[i];
if (h[i] > hi) hi = h[i];
}
return (hi - lo) <= limit;
}
/* 2 kg cell, 2 mV/V, 5 V, gain 128, 5 V reference: 2,147,483 counts for 10 mV,
so 1,073.7 counts per gram. Tested: 1000 g -> 1000, 500 g -> 500,
0 g -> 0, -3 g -> -3; a 400-count offset reads 0 g and moves zero by 25 counts. */
The numbers in the comment came from running it. With a 5 V reference and a gain of 128, the converter's range is about plus or minus 39 mV, so the 10 mV signal uses about 2.1 million counts. That is 1,073 counts per gram on a 2 kg cell, so a one-gram division is far above the noise floor of a good front end.
Zero tracking is the feature that makes a scale feel finished. A small amount of material left on the pan, or thermal drift, will shift zero by a fraction of a division over minutes. Tracking the zero slowly inside a window of a few divisions keeps the display at 0.0 without hiding a real load. Make the window small and the tracking slow, and disable it when the reading is not stable. A regulated trade scale has rules about this, and you must follow them.
Stability detection matters for the same reason. A product that commits a weight only when the last eight readings agree within a limit avoids recording the weight of a bird that is still moving. In the example the same eight readings were stable under a limit of 100 counts and unstable under 20, which is the sort of tuning you do against real noise data.
For span linearity, a two-point calibration is enough when the cell is good. If your cell has a nonlinearity near the top of its range, add a third point at half capacity and interpolate piecewise. For temperature, characterise zero and span at three temperatures and store the coefficients.
Tests That Catch a Bad Front End Before the Customer Does
Put the front end through these before the enclosure is designed.
Test | How to run it | What a pass looks like |
Noise histogram | Fixed weight, 10,000 readings at your data rate, plot the histogram | Peak-to-peak under one quarter of a division |
Zero drift | Power on, no load, log for two hours, then heat the board 20 °C with a hot-air gun at 30 cm | Drift under half a division |
Cable length | Calibrate with 1 m, then swap in the longest cable you will ship | Span change under half a division, or six-wire proves immune |
Mains pickup | Run next to a fluorescent fixture and a running motor | No change in noise histogram |
Supply sweep | Vary supply voltage across its allowed range | Reading change under one division |
Creep | Load 80 percent of capacity for 30 minutes | Creep within the cell's datasheet figure |
Run the first two on day one of any design. If they fail, no firmware will rescue the product, and you will have learned it for the price of an afternoon.
The hot-air test is the one teams skip, and it is the one that exposes thermocouple and reference-wiring errors. A scale that is stable at 25 °C and drifts at 45 °C will fail in a shed.
Legal Metrology in India: What Changes If the Scale Is Used for Trade
If the scale will be used to sell goods by weight, the front end is no longer only an engineering choice. The Legal Metrology Act, 2009 and the Legal Metrology (Approval of Models) Rules, 2011 require manufacturers and importers of prescribed weighing and measuring equipment to obtain model approval, and the Legal Metrology (General) Rules, 2011 set specifications for about forty kinds of instrument, including electronic weighing machines. The Department of Consumer Affairs pages describe the procedure.
The practical points an electronics team should know:
The international framework is OIML R 76 for the instrument and OIML R 60 for the load cell. A load cell of accuracy class C is the usual match for a class III instrument, which covers roughly 500 to 10,000 divisions. A class D cell matches class IV at up to about 1,000 divisions. Class III is the common choice for retail and commercial scales.
Part of the instrument's error allowance belongs to the cell, and the electronics must stay within what is left. That is why a front end with a margin of four to one matters, and why a design that merely meets its division count on a good day tends to fail testing.
Sealing and software integrity matter. The approval certificate includes a sealing provision, and the revised procedure asks for a schematic of it, so that vital components and calibration parameters cannot be changed to produce fraudulent weighment. Your firmware needs a protected calibration store and an audit counter.
This is general information, not legal advice. Confirm the current rules, fees and test laboratory with the Legal Metrology authority before you commit a design to trade use.
When None of These Three Is the Right Answer
There are cases where the architecture in the table is the wrong question.
If the weighing is dynamic, such as a belt scale or a checkweigher, the limit is not noise but settling time and vibration. Choose the converter for its filter latency, not its resolution, and invest in mechanical isolation.
If the cell is far away, put the converter at the cell and send digits down the cable on RS-485 or a similar bus. A digital load cell removes the entire cable section of this post, and industrial digital cells exist for that reason.
If the weight changes over weeks, such as a silo, the problem is drift and creep, not resolution. A stable cell and a calibration schedule beat a better converter.
If the product is a one-off prototype, use whatever evaluation board gives you data fastest, and write the noise number down so that you know what you are leaving on the table.
Where to Start on Monday Morning
Write the number of scale divisions the product must resolve, and the data rate it needs.
Convert to microvolts per division for your cell: sensitivity in mV/V times excitation voltage, divided by divisions.
Open the datasheet noise table for the converter, find RMS noise at your gain and rate, multiply by 6.6, and check that it is below a quarter of a division.
Decide on four or six wires from the cable length. Use six wires over about three metres.
Wire the reference to the bridge, not the regulator.
Pick a data rate that puts a notch at 50 Hz, and check the 49.5 Hz and 50.5 Hz attenuation.
Lay out the input pads symmetrically and keep heat sources away.
Run the noise histogram and the hot-air drift test before any other work.
Write the calibration, zero tracking and stability code against measured noise, not assumed noise.
If the scale is for trade, contact the legal metrology authority before the design freezes.
Ten Millivolts, Revisited
Return to the ten millivolts. Nothing about it is exotic. It becomes a good scale or a flickering one by a short list of decisions that can all be made on paper: the divisions, the noise at your data rate, the reference wiring, the cable, the mains notch and the layout.
The scale that flickered in the shed did not need a better filter. It needed the sum done before the board was drawn.
Related service: PCB, Electronics and RF Design.
Related reading: Why Your Load Cell Reading Drifts, Automatic Bird Weighing in Broiler Sheds.
EurthTech is an engineering-first product design and embedded systems company with a development center in Hyderabad. To discuss a project, write to connect@eurthtech.com or see https://www.eurthtech.com/case-studies.




Comments