Can a Low-Cost Sensor Catch Adulterated Milk at a Collection Centre? Eight Objections, Answered
3.2 and 8.3. Those are the minimum fat and solids-not-fat percentages that FSSAI's standard sets for cow milk, and a collection centre's analyser reports exactly those two numbers in a few seconds. A milk can pass both after water has been added, if powder or something else lifts the solids back up. A can can also fail both without a drop of adulterant, because the cow is early in lactation or the season has turned. The distance between "the analyser says the milk is low" and "the milk has been adulterated" is where disputes, fines and trust are decided, and it is the distance this post is about.
The claim under test is one that gets repeated in tenders and brochures: a low-cost sensor kit at a collection centre will catch adulterated milk. The honest answer is that sensors screen for some things, miss others, and are easily misread for the rest. This post puts eight objections to the claim, answers each from the standards and the physics, and ends with a screening workflow that a cooperative can defend to a farmer and to a regulator, plus code that shows what a fat and SNF screen can and cannot say.
A scope note before the argument. This is an engineering discussion of screening, not legal advice or a substitute for a laboratory. The standards quoted come from FSSAI's published regulations and its manual of analytical methods, and each cooperative should confirm the current text and its own legal position on sampling and enforcement.
A screening device tells you which milk to look at more closely. It does not prove adulteration, and a decision that costs a farmer his income should never rest on one reading.
What the Standard Says, and Where It Applies
FSSAI's standard for milk sets minimum fat and solids-not-fat (SNF) for each class. The table lists the classes most relevant to collection centres. The standard also notes that the compositional requirements apply at the points of sale, which matters because a collection centre is upstream of that point and uses the numbers as an acceptance and pricing reference.
Class of milk | Minimum milk fat | Minimum SNF |
Buffalo milk | 5.0 % | 9.0 % |
Cow milk | 3.2 % | 8.3 % |
Goat milk | 3.0 % | 8.0 % |
Mixed milk | 4.5 % | 8.5 % |
Standardised milk | 4.5 % | 8.5 % |
Toned milk | 3.0 % | 8.5 % |
Double toned milk | 1.3 % (amended 2023) | 9.0 % |
The same standard limits total urea in milk to 700 ppm and total sodium to 650 mg per 100 g of SNF. FSSAI's analytical manual explains that urea occurs naturally in milk at around 20 to 70 mg per 100 ml, and that more than 70 mg per 100 ml suggests added urea. The two statements match: 700 ppm is 70 mg per 100 ml. The manual also lists chemical tests for a long catalogue of adulterants: cane sugar, starch, cellulose, urea, ammonium compounds, sulphates, glucose, sodium chloride, foreign fat, nitrates, neutralisers, hypochlorites, quaternary ammonium compounds, detergents, skimmed milk powder, gelatine, formalin, hydrogen peroxide and salicylic acid, plus reference methods for fat (Gerber and Rose-Gottlieb) and protein (Kjeldahl).
That catalogue is the first reply to the claim. Each adulterant on the list has its own chemistry, and most of the tests are reagent tests run on a sample. Almost none of them is a physical property that a cheap sensor reads.
What a Cheap Instrument Measures
Collection-centre instruments read physical properties of milk and infer composition from them. Knowing the property is the first step to knowing the blind spots.
Instrument | Property it reads | What it estimates | What shifts the reading that is not adulteration |
Thermometer | Temperature | Sample condition and correction for other readings | Sample age, mixing |
Lactometer (hydrometer) | Density | Corrected lactometer reading (CLR), used with fat to estimate SNF | Temperature, fat content, foam |
Conductivity probe | Electrical conductivity | Ionic content; a general flag for water or salts | Mastitis (raises it), season, feed, temperature |
Ultrasonic analyser | Sound velocity and attenuation | Fat, SNF, density, protein estimates, added water | Species calibration, temperature, fat state, air bubbles |
Optical or infrared analyser | Absorption at selected wavelengths | Fat, protein, lactose, solids | Calibration and sample preparation; usually costs more |
Strip or reagent kit | A chemical reaction | One named adulterant or class | Reagent age, technique, interfering substances |
Conductivity, ultrasonic and density readings are indirect. They respond to changes in the aqueous phase, the fat or the solids, and many different changes can produce the same reading. That is the source of both their usefulness and their traps.
Why a Fixed Rule Misleads: Season, Breed and Herd
A regulatory minimum is a line drawn through a distribution. Milk from healthy animals falls on both sides of it, and the position of an individual farmer's milk relative to the line changes with factors that have nothing to do with adulteration. A screening rule that ignores them will produce false alarms in predictable places and times.
Factor | General effect on composition | What it means for screening |
Species and breed | Buffalo milk is richer in fat than cow milk; breeds differ in fat and solids | Calibrate and judge by species, and by breed mix where it is known |
Stage of lactation | Composition changes across the lactation | A farmer's early-lactation animals may sit lower than his mid-lactation herd |
Season and heat stress | Hot spells are commonly associated with lower solids in milk | Expect more borderline results in summer, and look at the whole herd's trend |
Feed and fodder | Energy and fibre in the ration change fat and solids | A change of fodder can move a farmer's results without any adulteration |
Mastitis | Raises conductivity and sodium and chloride, and lowers solids | A high conductivity reading should prompt a health check as well as a quality check |
Milking time | Morning and evening milk differ in fat | Compare like with like, and sample the same milking each day |
Pooled supply | A can may hold milk from several animals | The result is a blend; one low animal can pull a lot down |
The practical remedy is to measure each farmer against his own history as well as against the standard. A farmer whose milk has averaged 8.6 percent SNF for a year and reads 7.9 today is telling you something even if 7.9 is above a lower local threshold, and a farmer whose milk has always sat at 8.3 and reads 8.3 is not. A rule that flags sudden departures from a farmer's own pattern catches the abrupt change that adulteration causes and tolerates the slow drifts that season and feed cause. Keep the standard as the legal reference, and use the history as the sensitivity.
Judge a delivery against the standard and against the farmer's own record. A sudden step from his normal is worth a second look; a seasonal drift shared across the village is the weather.
Objection 1: "A conductivity probe will catch urea"
It will not. Conductivity measures the movement of ions. Urea is a neutral molecule that does not ionise in water, so adding it to milk changes the conductivity very little. A probe that is excellent at flagging salt, many neutralisers and some watering will sit silent while urea goes in. The same holds for sugar, starch, glucose and most of the organic adulterants on the FSSAI list.
The reason the claim is repeated is that conductivity does move for a handful of adulterants (salt, some neutralisers) and for a natural condition, mastitis, in which sodium and chloride rise and the milk's conductivity climbs. A conductivity alarm is therefore a flag for "something ionic changed", which could be disease, salt or water, and none of those is urea.
Urea needs a chemical test (the manual describes a colour reaction using para-dimethylaminobenzaldehyde) or an enzymatic or spectroscopic method. Because milk naturally contains urea, any test must also compare the result with the natural range, so the decision threshold sits at the standard's 700 ppm, not at zero.
Verdict: conductivity is a useful general-purpose flag, and a poor adulteration detector for non-ionic additives. Do not tell a farmer his milk is free of urea because the conductivity is normal.
Objection 2: "An ultrasonic analyser tells you the composition"
It estimates the composition, using calibration curves. An ultrasonic analyser sends sound through a milk sample and measures how quickly and how strongly it travels. Fat, solids and water affect those quantities, and the instrument converts them into reported fat, SNF, protein and density using equations stored in its memory. The result is only as good as the calibration, and the calibration assumes the milk is a natural milk of the species it was set up for.
Several practical consequences follow. A device calibrated on cow milk will misread buffalo milk, whose fat globules and protein differ, and one calibrated on mixed milk will misread a pure sample. Temperature must be controlled or corrected. Air bubbles and cream that has risen scatter the sound and distort the reading. And the device cannot tell why a number is low: it can only say that the fat or solids reading is low.
Vendors quote errors of a fraction of a percent for fat and SNF under good conditions. Treat that as a claim to test: run 30 samples of your own milk on the analyser and by the Gerber method for fat, and compare. The difference is the instrument's real uncertainty in your conditions, and it is the number the screening rule needs.
Verdict: an ultrasonic analyser is a good screening tool for composition, within its calibrated domain, and a bad witness for a dispute.
Objection 3: "If the SNF is low, the milk has been watered"
Low SNF is consistent with added water, and it is also consistent with several natural causes. SNF varies with breed, stage of lactation, feed, season and heat stress, with disease (mastitis changes the composition), and with the herd's genetics. A cow in the heat of summer on poor fodder can give milk below the standard's minimum with nothing added. The standard's minimums are a regulatory floor, and individual healthy animals can fall below them.
The usual estimate of SNF from a lactometer reading and fat uses a relation of the Richmond type: total solids equal a quarter of the corrected lactometer reading plus 1.2 times the fat percentage plus 0.14, and SNF is total solids minus fat. In code the SNF relation is a single line. The more interesting part is how a screen should behave around the threshold.
# milk_screen.py -- SNF from CLR and fat, with an uncertainty band so borderline results are not accusations
MIN = {"cow": (3.2, 8.3), "buffalo": (5.0, 9.0)} # (fat %, SNF %) FSSAI minimums
def snf_from_clr(clr, fat):
return clr / 4.0 + 0.2 * fat + 0.14 # SNF = TS - fat, with TS = 0.25*CLR + 1.2*fat + 0.14
def screen(species, fat, clr, unc=0.15):
fmin, smin = MIN[species]
snf = snf_from_clr(clr, fat)
flags = []
if fat < fmin - 0.1: flags.append("fat low")
if snf < smin - unc: flags.append("FAIL snf")
elif snf < smin: flags.append("borderline snf (within analyser uncertainty)")
return round(snf, 2), flags or ["pass"]
def added_water_pct(snf_measured, snf_standard):
return max(0.0, (1 - snf_measured / snf_standard) * 100)
Running the screen on a few cases shows the three outcomes the rule should distinguish.
Case | Computed SNF | Screen result |
Cow, fat 3.8 %, CLR 29.6 | 8.30 % | Pass |
Cow, fat 3.8 %, CLR 29.0 | 8.15 % | Borderline, within analyser uncertainty |
Cow, fat 3.8 %, CLR 27.0 | 7.65 % | Fail on SNF |
Buffalo, fat 6.0 %, CLR 31.0 | 9.09 % | Pass |
Buffalo, fat 6.0 %, CLR 29.5 | 8.71 % | Fail on SNF |
A borderline result should trigger a retest and a second method, not a penalty. A fail should trigger a confirmation on a calibrated analyser and, if it is contested, a reference test. The uncertainty band is the difference between a screen and an accusation.
Verdict: low SNF is a reason to look further. It is not proof of watering, and a system that treats it as proof will punish honest farmers in bad weather.
Objection 4: "Water and powder cannot fool the analyser"
They can, and this is the objection that matters most to a cooperative. Added water lowers fat and SNF together. Adding skimmed milk powder or another source of solids raises the SNF again. The standard's own manual notes that skimmed milk powder may not be used to adjust SNF in cow, buffalo or mixed milk, which tells you the practice exists and that a separate test is required to find it. The analyser sees the compensated composition.
A worked example with the same code shows why composition alone is not evidence. Take cow milk at 4.0 percent fat and a CLR of 30. Dilute it by 10 percent with water, which lowers fat to 3.64 percent and the CLR to about 27, and the screen fails the SNF at 7.62 percent. Now add enough powder to bring the CLR back to 30. The screen on fat 3.64 percent and CLR 30 gives an SNF of 8.37 percent and a pass, even though the milk has been diluted and reconstituted.
Milk | Fat | CLR | Computed SNF | Screen result |
Natural cow milk (starting point) | 4.00 % | 30.0 | 8.44 % | Pass |
Diluted by 10 % with water | 3.64 % | 27.0 | 7.62 % | Fail on SNF |
Diluted, then powder added to restore CLR | 3.64 % | 30.0 | 8.37 % | Pass |
Catching the second case needs a different kind of test: a protein or lactose measurement that does not move with added powder in the same proportions, a freezing-point measurement of the water content, or a reagent test for skimmed milk powder as described in the manual. Cheap screens do not do that.
Verdict: a milk that passes fat and SNF may still be adulterated. Treat a pass as "not flagged by this method", and run periodic confirmatory tests that look at different properties.
A passing analyser reading means the milk was not caught by that method. It does not mean the milk is clean.
Objection 5: "A sensor can flag detergent, formalin, peroxide and neutralisers"
Not the cheap ones, and not by physical property. Detergent, formalin, hydrogen peroxide, hypochlorite and neutralisers are chemical adulterants, and FSSAI's manual detects them with specific reagent tests: colour changes, precipitates or titrations. Some of them (neutralisers, salts) will move conductivity or pH, which gives a hint, but the confirmation is a chemical test.
Rapid strip kits exist for several of these, and they are useful as a second tier when a reason to suspect exists. They have limits of their own. Reagents expire, technique matters, other components of the milk can interfere, and a faint colour is open to dispute. A cooperative that uses strips should record the batch, expiry and operator, and should treat a positive as a trigger for a laboratory confirmation when the consequence is serious.
Adulterant | Can a physical-property sensor flag it? | What does |
Added water | Yes, indirectly, through fat, SNF, density, conductivity | Freezing-point measurement for confirmation |
Urea | No | Chemical or enzymatic test against the 700 ppm limit |
Salt, neutralisers | Partly, through conductivity and pH | Specific reagent tests (chloride, neutraliser tests in the manual) |
Detergent | No | Specific reagent test |
Formalin, hydrogen peroxide | No | Specific reagent test |
Starch, sugar, glucose | No | Reagent tests and some spectroscopic methods |
Skimmed milk powder | Hidden by compensated composition | Specific test in the manual; protein and lactose analysis |
Foreign (vegetable) fat | Not by a basic analyser | Butyro-refractometer after fat extraction, fatty acid profile in a laboratory |
Verdict: a collection centre's analyser covers one row of this table well and the rest poorly. The honest design is a tiered one.
Objection 6: "Temperature does not matter much"
It matters a great deal, in two ways. Density and sound velocity both change with temperature, so an analyser or a lactometer reading must be corrected to a standard temperature, and the correction factor depends on the instrument and its calibration; follow the manufacturer's table, not a rule of thumb. The second way is more subtle: a warm sample also ages, which affects acidity and reagent tests, and fat behaves differently above and below the melting range of milk fat, so a sample that has been cooled and warmed unevenly may not be homogeneous.
A screening rule should therefore refuse to read a sample outside a temperature window, and should record the temperature with every result. The rule should also require mixing the sample until homogeneous before analysis, in line with the sample preparation step in FSSAI's manual (warm the sample, mix thoroughly and bring it to room temperature before withdrawing for analysis, and reject a sample that is clotted or curdled).
Verdict: temperature is part of the measurement. A reading without a temperature is a number, not a result.
Objection 7: "One sample from the can speaks for the lot"
Only if the sample is representative, and in milk that takes discipline. Fat rises as cream in a can that stands. A sample taken from the top is rich, one from the bottom is lean, and a sample taken before the can has been stirred will not match the average of the lot. A farmer who brings several cans of different dates in one delivery, or who blends milk from several animals, presents a composite whose fat and solids are not those of any single animal.
The remedy is procedural: stir or plunge the can for a set time, take the sample at a standard point, take it before the milk has been agitated or cooled unevenly, and use the same method for every farmer. For disputes, keep a duplicate sample sealed and labelled, stored cold, so that a laboratory can test the same milk. Where the collection centre pools milk in a chiller, sample the bulk as well as the individual deliveries, because a pooled sample that fails after individual samples passed is itself a signal.
Verdict: the best analyser in the world measures the sample in the cup. Make sure the sample in the cup is the milk in the can.
Objection 8: "Digital records will stop cheating"
Records stop some cheating and move the rest. A device that logs every reading with a timestamp, farmer identifier and operator makes silent edits harder. It does not protect against a sample that was prepared to pass, a calibration that was set to favour a supplier, or an operator who enters the wrong farmer. In a cooperative where operators and farmers know each other, collusion is a more likely failure than a hack.
A defensible design treats the data as evidence that must itself be defended. Use a device that stores readings without allowing deletion, and sync them to a central record. Require that calibration and verification events are logged and reviewed by someone outside the collection centre. Check the analyser against a reference each day with a known sample, and publish the results. Rotate operators, audit random deliveries, and give farmers access to their own readings, because a farmer who can see the history of his milk can spot a bad reading before the cooperative does.
Combine the milk readings with the weight. Many cooperatives pay on fat and SNF as well as quantity, and the weighing instrument is subject to its own legal-metrology requirements, as our post on load-cell drift discusses. A reading and a weight tied to one farmer identifier and one timestamp make a record that is much harder to dispute than either alone.
Verdict: logging helps if the logs are protected, the calibration is audited and the people are rotated. It does not replace honest procedure.
A Three-Tier Workflow a Cooperative Can Defend
The answers to the eight objections point to the same structure: screen widely and cheaply, confirm narrowly and specifically, and reserve the laboratory for decisions that carry consequences.
Tier | What it does | Tools | Outcome |
1. Screening at every delivery | Check that the milk looks like milk of the declared species | Temperature, sensory check, analyser for fat and SNF (with the species calibration), conductivity flag | Pass, borderline (retest) or flag |
2. Targeted checks on flagged or random samples | Test for the adulterants that screening cannot see | Reagent strips or kits for urea, neutralisers, detergent, peroxide, formalin and similar, per the FSSAI methods | Negative, or positive for confirmation |
3. Reference laboratory | Settle disputes and confirm positives | Accredited laboratory using the standard methods (Gerber or Rose-Gottlieb for fat, Kjeldahl for protein, and the specific tests for adulterants) | A result a regulator will accept |
Several rules hold the workflow together. A farmer is never penalised on a tier 1 result alone. Borderline results are retested on a second reading and, if needed, a second instrument. Positives at tier 2 trigger a sealed duplicate sample to the laboratory before any penalty. Random tier 2 checks, a fixed share of deliveries chosen by a rule the operator cannot influence, protect against the compensation tricks that tier 1 cannot see. Everything is logged, and farmers can see their own records and the process for appeal.
A Delivery Record and the Decision Rule
The workflow becomes defensible when every delivery leaves a record that shows what was measured, how, by whom and what was decided. A record small enough to fill in at the bench is better than a perfect one that is skipped.
Field | Why it is there |
Delivery ID, date, time, milking (morning or evening) | Compare like with like, and anchor the record |
Farmer ID and species declared | Select the calibration and the history |
Weight or volume (linked to the weighing record) | One record per delivery |
Sample temperature and mixing confirmed | Objection 6 and 7 |
Instrument ID, calibration ID, date of last reference check | Trace a reading to a device in a known state |
Fat, SNF, density or CLR, conductivity | The screening readings |
Screen result: pass, borderline or fail | The tier 1 outcome |
Retest reading and result, if any | The borderline path |
Targeted test, batch and expiry of the reagent, result | The tier 2 outcome |
Sealed duplicate sample ID, if taken | Dispute route |
Operator ID | Accountability and rotation |
Decision and reason | What was done and why |
The decision rule can then be written down and checked by a script instead of left to memory. The function below encodes the three-tier logic from this post: a pass is accepted, a borderline is retested and not penalised, a fail is confirmed with targeted tests and a sealed sample, and no penalty is applied without laboratory confirmation.
# decide.py -- the tiered rule as code, so every delivery gets the same treatment
def decide(tier1, retest=None, strip=None, lab=None):
if tier1 == "pass":
return "accept; random tier 2 may still apply"
if tier1 == "borderline":
if retest == "pass":
return "accept"
return "retest on a second reading or instrument; no penalty yet"
if tier1 == "fail":
if retest == "pass":
return "accept; log analyser disagreement"
if strip is None:
return "run targeted strips; seal duplicate sample"
if strip == "negative":
return "price on measured composition; no adulteration finding"
if strip == "positive" and lab is None:
return "send sealed sample to laboratory; hold penalty"
if lab == "confirmed":
return "apply policy; give farmer the record and appeal route"
return "no penalty: laboratory did not confirm"
Running the eight representative paths through the function gives the outputs in the table, which also make a handy training sheet for operators.
Tier 1 | Retest | Strip | Laboratory | Decision |
Pass | none | none | none | Accept; random tier 2 may still apply |
Borderline | none | none | none | Retest on a second reading or instrument; no penalty yet |
Borderline | Pass | none | none | Accept |
Fail | Fail | none | none | Run targeted strips; seal duplicate sample |
Fail | Fail | Negative | none | Price on measured composition; no adulteration finding |
Fail | Fail | Positive | none | Send sealed sample to laboratory; hold penalty |
Fail | Fail | Positive | Confirmed | Apply policy; give the farmer the record and the appeal route |
Fail | Fail | Positive | Not confirmed | No penalty: laboratory did not confirm |
Note the second row from the bottom of the logic: a fail that the strips do not support is priced on its measured composition, not punished. That separates the question of what the milk is worth from the question of what the farmer did, and it is the line a cooperative most needs to hold.
What the Hardware Should Do
If a cooperative is buying an analyser or building a station, the hardware requirements follow from the objections.
Requirement | Reason |
Species-specific calibrations (cow, buffalo, mixed) selectable and logged | Calibration is the dominant source of error |
Temperature measurement and correction with every reading, and refusal to read outside a window | Objection 6 |
Uncertainty reporting, not just a number | Borderline decisions need an uncertainty band |
Tamper-evident, append-only logs with farmer ID, operator ID, time and temperature | Objection 8 |
Daily check against a reference sample, with the result stored | Detects drift before it harms a farmer |
Easy cleaning and a rinse routine between samples | Carry-over distorts the next reading |
Integration with the weighing instrument and farmer ID | One record per delivery |
A path to export raw readings for audit | Lets an outside party recheck the data |
A clear statement of what it cannot detect | Prevents overclaiming by operators and sellers |
Questions Farmers Ask, and Honest Answers
A screening system is accepted or rejected in conversation with farmers, and the questions are predictable. Prepared answers, agreed in advance, stop operators improvising.
Question | An honest answer |
Why was my milk flagged when I added nothing? | The screen flagged composition below the standard. Several natural causes lower it, so we retest and check before any decision. |
Can I see my readings? | Yes. Every reading, with the instrument, time and temperature, is in your record. |
Is the machine always right? | No. It has an uncertainty, which we publish, and a borderline result is retested. |
Who checks the machine? | It is checked daily against a reference sample, and the results are reviewed by someone outside the centre. |
What happens if I disagree? | We take a sealed duplicate sample to an independent laboratory, and you will see the result. |
Does one bad reading cost me money? | Not by itself. No penalty is applied on a single screening result. |
Why do you test some milk for chemicals and not all? | The tests are specific and slower. We test a fixed share chosen by rule, plus any flagged sample. |
Will the machine treat buffalo and cow milk differently? | Yes, because they differ. Each farmer's species is set in his record. |
Answers like these do more for compliance than any penalty. They tell the farmer that the system is a measurement with rules, which a farmer can understand and can contest, and not a machine that decides his income.
A farmer who understands why his milk was flagged and how to appeal will cooperate with the rules. One who receives a verdict from a machine will look for the machine's weak point.
A Cooperative's First Six Months
A representative composite, genericised, shows the shape. A village dairy cooperative with several hundred members installed ultrasonic analysers and a conductivity flag at its collection centres, and began paying partly on fat and SNF. In the first month the analyser flagged a high share of morning samples as low in SNF, and members complained. A review found three causes: the analyser's cow calibration had been applied to a buffalo-heavy intake, the sample temperature varied widely between the first and last deliveries of the morning, and during a hot spell genuinely lean milk was common.
The cooperative fixed its procedure before it touched the penalties. It set species calibrations per farmer, recorded temperature and refused out-of-window samples, added the borderline band and retest rule, and began a daily reference-sample check. It started random tier 2 strip tests on a fixed share of deliveries and sent a few flagged samples to a laboratory to learn what the screens were missing. Over the following months the share of flagged samples fell sharply, most of the early flags having been calibration and procedure problems, and a small number of confirmed cases, including a compensated dilution that tier 1 had passed, were found by the random checks. Members who had been angry about automated verdicts accepted the system once they could see their own readings and knew about the appeal route. The figures in this account are illustrative of the pattern. The pattern is that a screening system's first job is to fix its own errors, and that trust comes from procedure.
When Not to Deploy a Screening Kit
When there is no laboratory to confirm disputes. A screen without a reference route turns every borderline into a quarrel.
When the cooperative cannot calibrate by species and check daily. An uncalibrated analyser creates false positives and false negatives together.
When penalties are automatic. Do not link a single reading to a fine.
When operators cannot be rotated or audited. A system that depends on one person's honesty is only as good as that person.
When the intent is to prove adulteration to a regulator. Enforcement needs a method the regulator accepts, which means a laboratory.
Starting on Monday
List the adulterants that matter for your supply chain, and sort them by whether a physical sensor can see them (water) or a chemical test is needed (the rest).
Compare your analyser with the Gerber method on 30 samples of your own milk, by species, and write down the uncertainty.
Set species calibrations and a temperature window, and log both with every reading.
Write the screening rule with a borderline band, a retest step and no automatic penalty.
Add random targeted tests on a fixed share of deliveries, with a documented method from the FSSAI manual.
Agree a reference laboratory and a sealed duplicate-sample routine for any contested result.
Make the logs tamper-evident and rotate operators; let farmers see their own records.
Check the analyser daily against a reference sample and review the results weekly.
What a Sensor Can Promise
A low-cost sensor can promise that it will measure a property of the milk consistently, and a well-run collection centre can promise that it will act on the measurement fairly. Neither can promise that the milk is clean. Build the system for what it can do: a fast, cheap first look that is honest about its uncertainty, backed by specific tests and a laboratory for the cases that matter. The farmer who is trusted with an explanation will accept most rules, and the one who is handed an unexplained verdict will not.
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