Milk Analyser Buying Guide for Collection Centres: Accuracy, Calibration and Cost per Test
Buy for calibration, not for the price on the quotation or the accuracy printed on the brochure.
A milk analyser costs a fraction of a rupee per test to own and run. A small bias in its readings costs lakhs of rupees a year. That inversion is the whole of this guide: the money in this purchase sits in the calibration, the service and the honesty of the numbers, and almost none of it sits in the price of the box.
Three verdicts, before the reasoning. For most village collection centres and bulk milk chillers in India, an ultrasonic analyser is the right instrument, provided it can be calibrated for cow, buffalo and mixed milk and its calibration cannot be changed casually. Where you need to screen for several adulterants in one pass, report protein, or serve as a reference point for many centres, an infrared analyser earns its higher price. And whichever type you buy, insist on a calibration and service arrangement that you have read in writing before you pay.
Our earlier guide compared the methods used to detect adulteration in milk, test by test. This one answers the question that follows it: once you have decided to measure fat and solids-not-fat electronically, which instrument do you buy, and what do you ask before you sign? The arithmetic is shown, the code runs, and where a figure comes from a published datasheet or a public tender, it is marked as such.
What a Milk Analyser Measures, and How
Three technologies cover almost every analyser sold for collection centres, and they measure different things in different ways.
An ultrasonic analyser passes sound through a small sample and measures how fast it travels and how much it is weakened. Sound speed in milk depends on its composition and temperature, and the instrument turns those measurements into fat, solids-not-fat, density and an estimate of added water, using calibration equations stored for each milk type. It is compact, fast, uses no chemicals and has no moving optics to go out of alignment. Its readings are inferences from a physical property, so they are only as good as the equations behind them and the calibration set they were built on.
An infrared analyser measures how the sample absorbs light across a range of wavelengths. Each component of milk, fat, protein, lactose and others, absorbs in characteristic regions, so the instrument can read several components directly and can be extended to screen for substances such as urea, sugars and added starches. It is the method used in dairy laboratories and many processing plants. It is more complex, costs several times as much, and needs more careful maintenance and trained operators.
A near-infrared or optical sensor, often built into a smaller or cheaper instrument, uses a narrower band of wavelengths. It can be effective for fat and some other parameters, and its accuracy depends heavily on calibration against a reference.
Technology | What it measures | What it infers | Typical use |
Ultrasonic | Sound speed and attenuation, temperature | Fat, SNF, density, added water, sometimes protein, lactose and salts, through stored equations | Village collection centres, bulk milk chillers |
Mid-infrared | Absorption across the spectrum | Fat, protein, lactose and solids directly, with adulterant screening on some models | Chilling plants, dairies, reference points |
Near-infrared or optical | Absorption in selected bands | Fat and a few other parameters | Compact instruments, screening |
One distinction matters for every purchase. Some of the numbers on the display are measured and others are derived. Fat and density are usually closest to measured. Solids-not-fat, protein and added water are often computed from them. A derived value carries the errors of every input and of the formula, which is why the next sections spend time on calibration.
Typical Published Figures, and How to Read Them
Datasheets for ultrasonic analysers sold in India show a consistent pattern. The figures below are typical ranges drawn from several manufacturers' published datasheets and one public tender specification, and are intended to help you read a quotation, not to describe any one product.
Parameter | Typical published range | Typical published accuracy |
Fat | About 0.5 to 12 percent, some to 15 | Plus or minus 0.1 percent, with plus or minus 0.2 at the high end |
Solids-not-fat | About 6 to 15 percent | Plus or minus 0.15 percent, some quoting 0.2 |
Added water | 0 to about 60 percent | Plus or minus 3 to 5 percent |
Density | About 1.022 to 1.036 grams per cubic centimetre | Plus or minus 0.0005 |
Sample temperature | About 1 to 40 degrees Celsius | Plus or minus 1 degree |
Measuring time | About 25 to 40 seconds per sample | Not applicable |
Sample volume | About 15 millilitres | Not applicable |
Repeatability | Fat plus or minus 0.03 to 0.07 percent, SNF plus or minus 0.05 to 0.1 | Not applicable |
Power | About 30 watts on average, 90 at peak | Not applicable |
Reading these needs three habits.
First, separate accuracy from repeatability. Repeatability says how closely the instrument agrees with itself on the same sample. Accuracy says how closely it agrees with the truth. An instrument can repeat to 0.03 percent and still read 0.2 percent high on every sample. Repeatability is easy to achieve and easy to advertise, and accuracy is the number that decides money.
Second, ask accurate against what. Accuracy is always relative to a reference method and a set of samples. A datasheet that says plus or minus 0.1 percent for fat, with no statement of the reference method, the number of samples and the milk types, has told you little. Ask for the validation report.
Third, notice what the range excludes. A fat range that stops at 12 percent will not read rich buffalo or cream samples, and a density range that ends at 1.036 will not read heavily concentrated milk. For the milk you actually receive, check that your normal values sit well inside the stated ranges.
Repeatability is easy to advertise. Ask for accuracy against a named reference method, on named samples, in a document you can read before you pay.
The Money Is in the Bias
A small systematic error, called bias, is the most expensive thing an analyser can have, because it repeats on every can for as long as it goes unnoticed. Payment systems that pay on fat and solids-not-fat turn a decimal place of error into rupees at once.
The calculation below prices a litre of milk from its fat and SNF, using illustrative rates of 600 rupees per kilogram of fat and 350 rupees per kilogram of SNF. Your dairy's real rates will differ, and the logic is the same.
def price_per_litre(fat, snf, fat_rate_kg=600.0, snf_rate_kg=350.0, density=1.03):
return density * (fat / 100 * fat_rate_kg + snf / 100 * snf_rate_kg)
base = price_per_litre(4.0, 8.5) # 55.36 rupees per litre
for df, ds in ((0.1, 0.0), (0.0, 0.15), (0.1, 0.15)):
d = price_per_litre(4.0 + df, 8.5 + ds) - base
print(df, ds, round(d, 3), round(d * 1000 * 365 / 1e5, 2))
# 0.1 0.0 0.618 rupees per litre, 2.26 lakh a year at 1,000 litres a day
# 0.0 0.15 0.541 rupees per litre, 1.97 lakh a year
# 0.1 0.15 1.159 rupees per litre, 4.23 lakh a year
def cost_per_test(capex, life_years, tests_per_day, days=330, consumable=0.05,
calib_per_year=5000, watts=30, hours_per_day=8, tariff=9.0):
tests = tests_per_day * days * life_years
energy = watts / 1000 * hours_per_day * days * life_years * tariff
return (capex + consumable * tests + calib_per_year * life_years + energy) / tests
print(round(cost_per_test(80000, 5, 300), 3)) # 0.269 rupees per test at 300 tests a day
print(round(cost_per_test(80000, 5, 60), 3)) # 1.147 rupees per test at 60 tests a day
The result is stark. A bias of 0.1 percent in fat alone is worth about 62 paise per litre, which on a centre collecting 1,000 litres a day is about 2.26 lakh rupees a year. A solids-not-fat bias of 0.15 percent is about 2 lakh. Both together are over 4 lakh rupees a year, flowing from farmers to the dairy or the reverse, depending on the direction of the error.
Set that beside the cost of the test. Assuming an illustrative 80,000 rupees for the analyser, a five-year life and 300 tests a day, each test costs about 27 paise including the instrument, consumables, calibration and electricity. At a quiet centre doing 60 tests a day it is about 1.15 rupees.
The analyser costs paise per test and a bias costs lakhs per year. Spend the effort on calibration and service, and not on bargaining down the price of the box.
Bias also has a fairness side. An instrument that reads high pays farmers more than their milk is worth, and the dairy absorbs the loss. One that reads low shortchanges farmers, who notice sooner than any auditor. Disputes at the weighing counter usually begin with a number that two instruments disagree about, which is why the dairy's laboratory, the centre's analyser and the farmer's own sample must agree within a stated tolerance.
Calibration: The Requirement That Decides Everything
Calibration is the process of teaching the instrument what its readings should mean. For an ultrasonic analyser it means adjusting the stored equations so that the readings on a set of reference samples match their laboratory values, for each milk type. Everything the instrument reports afterwards depends on it.
Four questions about calibration should be answered in the quotation, in writing.
Who calibrates it, and where? Options range from the operator at the centre, to a service engineer who visits, to a central or remote process managed by the manufacturer or the dairy. Each has a different risk. Operator calibration is flexible and open to error and abuse. Engineer visits are slow and costly. Central calibration, done remotely and logged, is the arrangement that best protects against both mistakes and tampering, and some manufacturers now offer it through a mobile application or a cloud link. A public tender specification published by an Indian district administration in 2026 asks for exactly this: secure calibration through a Bluetooth or Wi-Fi connected mobile device.
How is the calibration protected? An analyser that any operator can recalibrate from the keypad invites a result that suits somebody. Look for access control, a log of every calibration with the time and the person, and a way for the dairy to see the log.
What reference samples are used? A calibration is only as good as its samples. They should span the fat and solids range you receive, for each milk type, with laboratory values from an accredited method. Ask how many samples, from where and by what reference method.
How often is it checked? Calibration drifts with temperature, wear, cleaning habits and the age of components. A check against a control sample, a retained sample of known composition, should be a daily routine, and a full recalibration scheduled on a calendar and after any repair.
Calibration arrangement | Strength | Weakness | Suits |
Operator at the centre | Fast, no travel | Open to error and to tampering | Only with a strict logged procedure |
Service engineer visit | Skilled, independent | Slow and costly across many centres | Small networks |
Central or remote with a logged record | Consistent and auditable | Needs connectivity and a trusted provider | Dairies with many centres |
Cow, Buffalo and Mixed Milk: Why One Curve Is Not Enough
Milk from different species differs in fat globule size, protein and mineral composition, and the equations that convert sound speed to composition differ with them. An instrument calibrated on cow milk reads buffalo milk wrongly, and mixed milk, common at village centres, sits between and shifts with the proportions.
Datasheets and tenders ask for mode selection for cow, buffalo and mixed milk, usually with a single curve per type. That is the minimum. Check four things.
Which curves are included, and was each built on samples from your region? Milk composition varies with breed, feed and season, and a curve built elsewhere can carry an offset.
How does the instrument handle mixed milk? A single mixed curve assumes an average proportion, which may not match what arrives. Some centres solve this by routing cow and buffalo milk to separate lanes or times.
Can you add a local curve? Instruments that let the manufacturer or the dairy build an additional curve from your own reference samples can improve accuracy substantially, provided the process is controlled.
How do you check? Run a monthly set of retained samples from each milk type against the dairy laboratory, and record the differences by type. A bias that appears for buffalo milk and not for cow milk is a calibration problem, not a fault.
Speed and Throughput: Do the Queue Arithmetic
Brochures quote a measuring time of 25 to 40 seconds per sample and sometimes a speed of 90 to 120 samples an hour. Those figures describe the instrument's cycle, and the queue at the counter depends on the whole routine: mixing the can, drawing the sample, loading the analyser, waiting for the result, recording it, rinsing and moving to the next farmer.
The calculation below turns that into stations needed. It assumes 40 seconds of measurement and 20 seconds of handling per can, and caps planned utilisation at 80 percent so that a rush does not form a queue.
import math
def stations_needed(cans_per_hour, measure_s=40, handling_s=20, utilisation_cap=0.8):
per_hour = 3600 / (measure_s + handling_s)
return math.ceil(cans_per_hour / (per_hour * utilisation_cap)), round(per_hour)
for c in (40, 80, 120, 200):
print(c, stations_needed(c))
# 40 (1, 60) 80 (2, 60) 120 (3, 60) 200 (5, 60)
Each station handles about 60 cans an hour once handling time is counted, not the 90 to 120 of the brochure. A centre receiving 80 cans an hour in its morning peak needs two stations, not one, if it wants to keep queues short. At 200 cans an hour, as at a large chilling point, it needs five, and at that scale a faster instrument or a more automated sampling arrangement starts to repay its cost.
The peak, not the average, sets the number. Most village centres receive the bulk of their milk in a window of an hour or two, and a single instrument that copes with the day's total may still cause a queue in that window. Measure your peak for a week before you decide.
Ultrasonic or Infrared: A Decision Table
The choice between the two main technologies follows from what you need the instrument to do.
If you need | Prefer | Why |
Fast fat and SNF at a village centre, low running cost | Ultrasonic | Compact, quick, no chemicals, lower price |
Protein reported directly for payment or processing | Infrared | Protein is measured, not inferred |
Screening for several adulterants in one pass | Infrared with an adulterant module | Spectral models can flag urea, sugars and others, subject to validation |
A reference instrument for many centres | Infrared | Stable, traceable, used in laboratories |
The lowest purchase price and simplest maintenance | Ultrasonic | Fewer parts and simpler service |
High throughput at a chilling plant | Infrared or several ultrasonic stations | Depends on queue arithmetic |
Infrared instruments aimed at collection centres are published with speeds of about 45 seconds per sample and up to about 80 samples an hour, and some advertise screening for water, urea, sucrose, maltodextrin and ammonium salts. Treat the screening claims as claims. Ask for the validation data: how many samples, which concentrations, what false-alarm rate and what detection limit. An adulterant screen that cries wolf is switched off by the third week.
A hybrid pattern works well. Ultrasonic analysers at the many village centres, with an infrared instrument at the chilling centre or the dairy laboratory as the reference and for periodic audits of the centres. The reference instrument also serves as the calibration anchor for the others.
What a Public Tender Asks For
A public tender is a useful template, because someone has already thought about what to require. A 2026 specification for milk analysers published by an Indian district administration lists, among other items, these.
A fat range of about 0.5 to 12 percent with accuracy of 0.1, widening to 0.2 above 12 percent. An SNF range of about 6 to 15 percent with accuracy of 0.15. Added water from 0 to 99 percent with an accuracy of 3 percent. Protein, lactose and salts reported. A measurement time of 25 to 40 seconds, a sample volume of about 15 millilitres, a speed of 90 to 120 samples an hour and continuous operation. Repeatability of about 0.07 percent for fat and 0.1 for SNF. Mode selection for cow, buffalo and mixed milk, with a single curve and cleaning, calibration and an error list. A detachable ultrasonic stirrer to remove air from the sample before testing. A performance certificate approved by the national dairy development agency. Secured calibration through a mobile device connected over Bluetooth or Wi-Fi. A three-year warranty. And a stainless steel enclosure of stated grade and thickness.
Use that list as a starting point for your own, and add what public tenders often omit: the right to inspect the calibration log, a stated maximum response time for service, a supply of cleaning solution and spares for the warranty period, a training commitment for operators, and a data-export format you can read without the vendor's software.
Treat a tender specification as a floor. It describes what a district buyer needs. Your centre may need more, such as operation at higher ambient temperatures, tolerance of voltage swings, or integration with your own payment software.
What an Analyser Cannot See
An analyser reports numbers for fat, solids and added water. It does not tell you what the solids are. This is the point that connects this guide to our earlier piece on adulteration tests.
A substance added to restore density, such as urea, salt or sugar, raises the density and the apparent solids, which means an ultrasonic or density-based reading can look normal while the milk is weaker and, in the case of urea, unsafe. The added-water figure is an estimate that depends on the instrument assuming the rest of the milk is normal, and it can be offset by adding something dense after the water.
For that reason, an analyser is a front-line filter and not a certificate. A sound protocol pairs it with supplier history, so that a reading is judged against the same supplier's usual values, with chemical tests for flagged cans and laboratory confirmation for disputes. The reading-plausibility check below is a small example of the kind of cross-check that catches inconsistent numbers before they are paid for.
def plausibility(fat, snf, density, temp_c):
flags = []
if not (0.5 <= fat <= 12):
flags.append("fat outside range")
if not (6 <= snf <= 12.5):
flags.append("SNF outside plausible range")
clr = (density - 1) * 1000
snf_from_clr = clr / 4 + 0.2 * fat + 0.14 # common field formula, constants vary by source
if abs(snf_from_clr - snf) > 0.6:
flags.append("SNF and density disagree by %.1f" % abs(snf_from_clr - snf))
if temp_c > 35 or temp_c < 5:
flags.append("sample temperature outside range")
return flags
print(plausibility(4.0, 8.5, 1.0305, 20)) # []
print(plausibility(4.0, 8.5, 1.0240, 20)) # ['SNF and density disagree by 1.6']
print(plausibility(4.0, 8.5, 1.0305, 38)) # ['sample temperature outside range']
The check cannot prove milk is pure. It can notice a reading that contradicts itself, which usually means a dirty sensor, a bubble, a warm sample or something added, and each of those deserves a second look before a payment is made.
Data, Software and Payment Integration
An analyser at a collection centre is one component in a chain: the weighing scale, the analyser, a printer and a billing system that turns a weight and a composition into a payment slip. How the pieces talk to each other affects accuracy, speed and trust.
Ask how readings reach the billing software. Automatic transfer by cable or Bluetooth avoids transcription errors, which are a surprisingly large source of disputes. Manual entry reintroduces them. Ask whether each record carries a timestamp, the farmer's identifier and the instrument's identifier, and whether the record can be edited after the fact. A tamper-evident log protects everyone.
Ask where the data lives. Cloud connectivity gives the dairy visibility across centres and allows central calibration, and it also creates dependence on a connection and a provider. Check that the instrument works offline and uploads later, and that you can export your data in an open format without paying for it.
Consider privacy and law. Records of what each farmer delivered and was paid are commercial and personal information. Under India's data-protection rules, now being phased in, the dairy has duties as the party that decides how the data is used. A vendor who holds the data should be bound by a clear agreement on access, retention and deletion.
Whoever controls the calibration and the data log controls the trust in the whole system. Decide who that should be before you sign, and write it into the contract.
Power, Environment and Daily Care
Collection centres are not laboratories. Heat, dust, damp and unreliable electricity are the normal environment, and an instrument that works in a showroom may not survive a village shed.
Power is the first issue. Datasheets commonly quote operation at 230 volts with a tolerance of 15 to 20 percent, and rural supply can swing outside that. A voltage stabiliser or a small uninterruptible supply is cheap protection, and a sudden loss of power during a measurement can corrupt a calibration or a stored record.
Temperature is the second. Sample temperature has a stated range, usually 1 to 40 degrees, and fresh milk arriving warm from the animal or after a hot journey can fall outside it or at its edge. The reading is temperature compensated, and the compensation is least reliable at the extremes. Cooling samples to a stable temperature before testing, or at least recording the temperature with each reading, helps.
Cleaning is the third and the one most often neglected. Milk residue in the measuring cell causes drift. A rinse routine after every sample, a deeper clean with the recommended solution at the end of each session and a log of both keep the readings steady. Air bubbles in the sample also distort readings, which is why tenders ask for an ultrasonic stirrer or a gentle mixing routine before testing.
Spares and service matter as much as the instrument. Ask where the nearest service engineer is, what the response time is, which parts wear out and what they cost, and whether a loan unit is available during repair.
Five-Year Cost of Ownership
The purchase price is a minor part of what the instrument costs over its life. A fair comparison puts every item next to every other.
Cost item | What to ask the vendor | A common surprise |
Purchase and installation | A line-by-line price including the stirrer, the printer and software | Accessories listed as optional that you will need |
Calibration and service | Annual charge, visit frequency and remote options | Calibration charged per visit, per curve |
Cleaning solution and consumables | Price per month and the supply commitment | Proprietary consumables with a single supplier |
Warranty and extended cover | What is covered, what is excluded, response time | Sensors and cells excluded from the warranty |
Software and cloud fees | Subscription cost and what happens if you stop paying | Data inaccessible after a lapse |
Power protection | Stabiliser or UPS cost | Not in the quotation at all |
Downtime | Loan units, response time | A week without an instrument in the peak season |
Training | Operator training at installation and later | One session for the first operator only |
Compute cost per test for each quotation, using your real daily volume, as the code above does. Two quotations that differ by 20,000 rupees in price can differ by much more over five years once service and consumables are included, and the instrument that costs more to buy often costs less per test.
A dairy that logs every calibration change can answer a dispute in minutes. A dairy that does not can only apologise and recalibrate.
A Composite: The Centre That Chose the Cheaper Analyser
This is a representative composite, genericised, from several dairy networks. The numbers are illustrative.
A dairy equipped two groups of collection centres with ultrasonic analysers from two suppliers. The first group of units was priced about a quarter lower. They calibrated from the keypad, without a log, and the vendor's service was by visit.
For two months nothing seemed wrong. In the hot weeks, farmers at several centres of the first group began to complain that their fat readings were lower than at neighbouring centres for similar milk. The dairy laboratory tested retained samples and found that the analysers read between 0.1 and 0.2 percent low on fat. Over the season the instruments had drifted, and some centre operators had adjusted them on their own, in different directions, to match what they thought were correct values.
The dairy could not tell who had changed what, because nothing was logged. It recalibrated every instrument against reference samples, which took three weeks of engineer visits, and settled farmer claims for the affected period. The cost of the exercise, in settlements, engineer time and lost goodwill, was several times the saving on the purchase price.
The second group, with logged central calibration, showed a small drift, caught by a daily control sample and corrected remotely in an afternoon. The dairy replaced the first group over the following year and wrote logged, access-controlled calibration into every later tender.
A Buyer's Checklist
Question | Why it matters | What a good answer looks like |
Which milk types are calibrated, and on what samples | One curve does not fit all | Separate curves for cow, buffalo and mixed, built on regional samples |
Who can change the calibration and is every change logged | Prevents drift and tampering | Access-controlled, logged and visible to the dairy |
What is the validation report against a reference method | Turns a claim into evidence | A document with method, sample numbers and results |
What is repeatability versus accuracy | They are different | Both stated, with the reference |
What are the service response time and the nearest engineer | Downtime costs money | A stated time and a local engineer |
How does the data reach billing, and can I export it | Transcription errors and lock-in | Automatic transfer, open export format |
What is the warranty and what is excluded | Sensors often are | Written scope, spares commitment |
What does it need in power and environment | Rural conditions | Stated tolerance, with a stabiliser or UPS option |
Is an adulterant screen validated, with its false-alarm rate | Marketing versus evidence | Data on concentrations and false alarms |
What is the training and the manual language | Operators decide quality | On-site training, manual in the local language |
Acceptance Testing on Day One
Do not accept an instrument on the strength of the brochure. A day of testing before payment protects you.
Collect retained samples that span your range: low and high fat, low and high SNF, for each milk type you receive. Have the dairy laboratory or an accredited laboratory analyse them by the reference method, and keep the results sealed.
Run each sample on the new analyser, ten times, in randomised order, at stable temperature. Record every reading. From them compute the average and the spread. The average against the laboratory value gives the bias. The spread gives the repeatability.
Compare with your acceptance limits, written in advance. A reasonable starting point is that the average of ten readings falls within the instrument's stated accuracy of the laboratory value for nearly every sample, and that repeatability is no worse than stated. Test across a range of temperatures within the stated limits, and test after power-cycling the instrument and after a cleaning cycle to check that readings return to the same values.
Record everything and sign it. The acceptance record becomes the baseline for every later calibration check.
Starting Monday
List the milk types, daily volumes and peak cans per hour at each centre.
Decide which centres need ultrasonic units and where an infrared reference is worth having.
Write the specification, adding calibration control, a service response time and data export to the tender template.
Ask each vendor for a validation report, a calibration procedure and a five-year cost quotation.
Compute cost per test and stations needed for each quotation.
Run the day-one acceptance test on a sample unit before ordering in bulk.
Set up the daily control sample, the monthly laboratory check and the calendar for full recalibration.
The Box and the Number
The analyser on the counter is a box that turns a sample into a number, and the number turns into money. Buying one well is mostly a matter of controlling that number: how it is calibrated, who can change it, how it is checked and how it travels to the payment slip.
The price of the box is the smallest part of the decision. Ask for the calibration, the service and the data in writing, test before you pay, and measure every month.
Related service: Dairy and Milk Quality Monitoring
Related case study: NIR Spectrometer for Food Quality Monitoring
Related reading: How to Detect Milk Adulteration: Methods Compared, From Lactometer to Analyser and Precision Agriculture in India: What to Measure First on a 1.08-Hectare Farm.
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