The Spec Sheet That Lies About Accuracy
The Spec Sheet That Lies About Accuracy: Sensor Selection for People Who Read Datasheets
A sensor can be precise without being accurate, a tight cluster of readings, far from the true value. Accuracy is closeness to the true value; precision is how tightly repeated readings group; resolution is the smallest detectable change. The Levelogger 6 M100's resolution is 0.6 mm on a 100 m full scale, 83 times finer than its ±5 cm accuracy. Averaging fixes random noise; it cannot fix systematic error. That needs calibration against a known reference.
Picture the sensing lead whose product's measurements cannot be trusted. The reading is stable, the same number, every time, hour after hour, but it is wrong, and the customer proved it with a reference instrument. The datasheet number does not survive the field: the accuracy was measured in a laboratory at 25 °C, and the product lives in a box that swings from −10 °C to 60 °C.
Or the calibration budget was never in the plan at all, the sensor was bought on its resolution, the field calibration was assumed, and nobody priced the reference, the lab, or the annual recalibration. The product's measurements cannot be trusted, and the reason is not that the sensor is bad. It is that the datasheet was read as a promise, and the datasheet was not a promise, it was a selection of favourable numbers, with the limitations left out.
This article is how to read a sensor datasheet the way it was written, what accuracy, precision, and resolution actually mean, why a sensor can be precise without being accurate, how the four kinds of specmanship work, why averaging fixes noise and cannot fix systematic error, and how to budget for drift before the product ships.
The words, in plain English
Accuracy. How closely the measured value comes to the true value, how good the data is when compared with a recognised standard. A temperature sensor accurate to 0.001 °C is expected to agree within 0.001 °C with a temperature standard such as a triple-point-of-water cell, or with a platinum resistance thermometer standardised in recognised calibration standards.
Precision (repeatability). How tightly repeated readings group together, the ability of a sensor to repeat a measurement when put back in the same environment. A more precise sensor has a narrow distribution of readings at a given pressure, and produces consistently repeatable measurements. Precision is often directly related to accuracy, but a sensor can be inaccurate yet repeatable.
Resolution. The smallest detectable change, the ability of a sensor to see small differences in readings. Resolution is often controlled by the quantisation in digitising the signal, one bit equal to 0.0005 °C, for example, which is not a function of the sensor, but of the sampling process.
Offset error. The reading the sensor gives when the true value is zero, the sensor's zero point is shifted, so every reading is shifted by the same amount.
Linearity error. The sensor's response is not a straight line, the error changes as the reading moves across the range, so the sensor is accurate at one point and wrong at another.
Hysteresis. A linear up and down input to a sensor results in an output that lags the input, you get one curve on increasing pressure and another on decreasing. Many pressure sensors have this problem; for better ones it can be ignored. It is often seen when the pressure reading on deck after recovery is different from the reading before deployment.
Temperature drift. The sensor's reading changes as the temperature changes, the offset, the sensitivity, or both move with temperature, so the accuracy measured at 25 °C is not the accuracy at 60 °C.
Noise. The random jitter in the readings, the part that changes from one reading to the next even when nothing else changes. Averaging fixes noise, because the noise averages toward zero.
Systematic error. The part of the error that does not change from one reading to the next, the offset, the linearity error, the hysteresis, the temperature drift. Averaging cannot fix systematic error, because the systematic error is the same in every reading, and the average of a thousand wrong readings is a wrong reading.
Random error. The part of the error that changes from one reading to the next, the noise. Averaging fixes random error, because the random errors cancel.
Drift. A low-frequency change in a sensor with time, often associated with electronic ageing of components or reference standards in the sensor. Drift generally decreases with the age of a sensor as the component parts mature. A smoothly drifting sensor can be corrected for drift; drift caused by biofouling cannot be properly corrected for, but is often tried.
Calibration. Comparing the sensor against a known reference and correcting the difference, the only thing that fixes a systematic error. A calibration is tied to the exact sensor, mount, firmware, and operating envelope it was done in.
Full scale (FS). The maximum value of the sensor's range, and the number accuracy is usually stated as a percentage of. An accuracy of ±0.05% FS on a 100 m full scale is ±5 cm; the same percentage on a 10 m full scale is ±0.5 cm.
Specmanship. The practice of presenting product specifications in a way that emphasises favourable values while omitting limitations.
The target and the arrows: four combinations
Accuracy and precision are two different things, and a sensor can have either, both, or neither. The image is a target and four sets of arrows.
A sensor that is accurate and precise puts a tight cluster of arrows in the bullseye. The readings group tightly, and the group is on the true value. This is the sensor the datasheet promises, and the sensor the product needs.
A sensor that is precise but not accurate puts a tight cluster of arrows far from the bullseye. The readings group tightly, the sensor is repeatable, stable, and confident, and every one of them is wrong by the same amount. This is the sensor that fools people, because the reading is stable and the stability looks like trustworthiness.
The error is systematic: an offset, a linearity error, a temperature drift, a calibration that was done on a different unit. Averaging the readings gives a tighter cluster in the same wrong place.
A sensor that is accurate but not precise puts a loose scatter of arrows centred on the bullseye. The average of the readings is right, but any single reading is not. This is the sensor averaging can fix: the noise averages toward zero, and the average converges on the true value.
A sensor that is neither accurate nor precise puts a loose scatter of arrows far from the bullseye. The readings are wrong and they wander. This is the sensor that needs replacing, not calibrating.
The point of the image is the second case, because it is the one that ships. A precise but inaccurate sensor passes every test that measures stability, the readings are the same, hour after hour, and the product looks finished, and it fails the first test that measures truth, which is the customer with a reference instrument.
What accuracy depends on, and what resolution depends on
Accuracy depends on six things: offset error, linearity error, temperature drift, hysteresis, noise, and environmental stress factors. Resolution depends on three: sensor noise characteristics, internal filtering and bandwidth, and ADC bit depth. The two lists overlap in one place, noise, and that overlap is where the confusion lives.
A sensor with low noise has fine resolution, because the noise is what sets the smallest detectable change, and a sensor with low noise can still have a large offset, a large linearity error, and a large temperature drift, because none of those are noise. A sensor may have high resolution but poor accuracy, or vice versa. Resolution describes what a sensor can detect, not how correct the measurement is.
The temperature channel that displays 25.01 °C is not necessarily more accurate than one displaying 25.0 °C. The extra digit may describe formatting rather than useful measurement information, and a stable reading can be consistently offset from a suitable reference. The CO2 sensor that lists 1 ppm resolution on both its versions does not have a ±1 ppm measurement tolerance: the 0–2000 ppm version is specified at ±(40 ppm + 2% of full scale), which is ±80 ppm across the range, and the 0–5000 ppm version is ±(50 ppm + 3% of reading), ±80 ppm at 1000 ppm, ±170 ppm at 4000 ppm.
The resolution is the same in both versions. The accuracy is not, and the accuracy is the number that decides whether the measurement can be trusted.
The worked example: the Levelogger 6 M100
The Solinst Levelogger 6 is a water level datalogger, a Hastelloy piezoresistive pressure sensor, a platinum RTD temperature detector, and a datalogger sealed inside a small waterproof housing, and its datasheet is unusually honest, which makes it the right example. The M100 has a full scale of 100 m. Its accuracy is ±0.05% FS. Its resolution is 0.0006% FS.
The accuracy, worked out: if the Levelogger 6 M100 with a full scale of 100 m reports a pressure of 88.0500 m, and the actual pressure is 88.0000 m, then the difference is 0.0500 m (5 cm); divide 0.0500 m by 100 m and express as a percentage, accuracy (or error) is 0.05%. The Solinst Levelogger 6 and LevelVent 6 have an accuracy of ±0.05% FS. On the M100, that is ±5 cm.
The resolution, worked out: the Levelogger 6 M100 has a resolution of 0.0006% FS. This means with a full scale of 100 m, the sensor will be able to detect a 0.0006 m (0.6 mm) change in pressure. The resolution is 83 times finer than the accuracy, 5 cm divided by 0.6 mm is 83.3. A reading that changes by 1 mm is real (the sensor can see it), but the reading itself may be 5 cm from the true value.
The same shape, in temperature: a temperature sensor may have a resolution of 0.00001 °C but only be accurate to 0.001 °C (MIT OCW), the resolution is 100 times finer than the accuracy, and you can believe the size of relative small changes in temperature, which are smaller than the accuracy of the sensor. Resolution is often controlled by the quantisation in digitising the signal, one bit equal to 0.0005 °C, which is not a function of the sensor, but of the sampling process.
The Levelogger's own temperature sensor is the same shape: accuracy ±0.05 °C, resolution 0.003 °C, the resolution is 16.7 times finer than the accuracy.
And the range matters as much as the percentage, because accuracy is stated as a percentage of full scale. The M5 (5 m full scale) is accurate to ±0.3 cm; the M10 (10 m) to ±0.5 cm; the M20 to ±1 cm; the M30 to ±1.5 cm; the M100 to ±5 cm; the M200 to ±10 cm. A shorter range gives a tighter absolute accuracy, select the shallowest range that covers your expected water column.
The same sensor family, ten times better, by choosing the shallowest range that covers the column.
Specmanship: the four ways a datasheet lies without lying
The engineering term specmanship refers to the practice of presenting product specifications in a way that emphasises favourable values while omitting limitations. In sensor datasheets, specmanship may involve four practices, and the Wikiversity guide on accuracy vs resolution names all four, because they are the four that mislead buyers into overestimating sensor performance.
Presenting resolution as if it were accuracy. The datasheet's headline number is the resolution, 0.6 mm, 0.00001 °C, 1 ppm, and the accuracy is a smaller number, further down the page, in a different unit. The buyer reads the headline and believes the measurement is accurate to 0.6 mm. The measurement is accurate to 5 cm. Finer steps does not equal accuracy.
Reporting accuracy only under nominal laboratory conditions. Many datasheets highlight best-case results in stable laboratory environments, omitting the effects of temperature swings, vibration, and power supply variation. The accuracy was measured at 25 °C, on a bench, with a clean power supply, and the product lives in a box that swings from −10 °C to 60 °C, on a battery that sags under load, next to a radio that transmits. The accuracy in the field is the accuracy at the field temperature, with the field vibration and the field power supply, and the datasheet does not state it.
Basing claims on prototypes instead of production units. The prototype was hand-assembled, hand-calibrated, and measured once. The production units are assembled on a line, calibrated by a procedure, and measured by a sample. The prototype's accuracy is not the production unit's accuracy, and the datasheet does not say which one it measured.
Assuming the customer will perform field calibration. Some datasheets assume that the customer will perform calibration in the field by taking readings in the field (super calibration). While this may improve results for a specific installation, it shifts risk, cost, and accountability to the end user. The field calibration cannot replicate factory-level thermal cycling and precision references, and the customer who was not told to budget for it has not budgeted for the reference, the lab, or the annual recalibration. The calibration budget was never in the plan, because the datasheet assumed it and did not say so.
Averaging fixes noise. It cannot fix systematic error.
Averaging is the most common fix for a noisy sensor, and it works, on the noise. The noise falls as one over the square root of the number of readings: average 100 readings and the noise falls ten times; average 1,000 and it falls 31.6 times. The random errors cancel, because they are random.
Averaging cannot fix systematic error, because the systematic error is the same in every reading. If the hub's pressure reading is offset by +4 cm, a systematic error inside the ±5 cm accuracy band, averaging 1,000 readings gives a reading that is still +4 cm off, just more precisely wrong. The average converges on the offset, not on the true value.
The only thing that fixes a systematic error is calibration against a known reference: compare the sensor against the reference, measure the difference, and correct it, and the calibration is tied to the exact sensor, mount, firmware, and operating envelope it was done in.
The Kalman filter cannot fix it either, from #4: a Kalman filter is a running best guess that weighs each reading by how much it trusts it, and a systematically wrong sensor with a tight noise figure looks trustworthy. The filter trusts the wrong number more, not less, because the tight noise is the evidence the filter uses to decide how much to trust the reading. One sensor is never enough, GPS is honest but slow, an IMU is fast but drifts 15 metres in 30 seconds, and the filter's job is to combine a sensor that is honest with a sensor that is fast, not to make a wrong sensor right.
The gyro demo from #17 is the same shape: a bias of 0.1°/s adds 6° of error every minute, 60° in 10 minutes, 360° in an hour, and no amount of averaging fixes a bias, because the bias is the mean of the readings. A single zero subtraction cannot correct bias, noise, scale factor, and misalignment, because they are four different signatures with four different tests and four different fixes, and bias stability, the 1σ spread read as the minimum of the Allan deviation curve, is a physical chip characteristic that cannot be calibrated away.
Drift: the error that grows with time
Drift is a low-frequency change in a sensor with time, often associated with electronic ageing of components or reference standards in the sensor, and it generally decreases with the age of a sensor as the component parts mature. A smoothly drifting sensor can be corrected for drift. Sea-Bird's temperature sensors are the example: they drift about 1 m°C per year (MIT OCW), and Sea-Bird says temperature generally drifts slowly, at a steady rate and predictably as a simple offset at the rate of about 1–2 millidegrees per year, and the 1995 SBE 3 calibration shows a clearly discernible +1.2 mK per year drift rate that could not be judged in conventional baths typically exhibiting variance at 1 mK+ levels.
Because the drift is smooth and predictable, it can be corrected for, and the correction is a simple offset, applied between calibrations. Sea-Bird's fix is an annual calibration schedule: a profiling CTD is recalibrated once a year, and an annual calibration schedule will usually maintain the CTD accuracy to within 0.01 psu in salinity.
Drift caused by biofouling cannot be properly corrected for, but is often tried, and the Levelogger's Biofoul Screen exists for the same reason: when a Levelogger is deployed for an extended period, there is the risk of biofouling on the pressure sensor and the conductivity cell, which can compromise their readings. The screen reduces the unwanted buildup of microorganisms, plants, algae, and organisms such as barnacles and mussels on the sensors.
Solinst's warning is the general one: all pressure transducers, no matter what they are made of, how expensive they are, or how accurate, are susceptible to sensor drift over time, and pressure sensor drift is a gradual degradation of the sensor and other components that can make readings offset from the original calibrated state. And the drift figures the manufacturers do provide are still based on use in very stable environments, making them somewhat irrelevant, because the drift varies with actual usage and the conditions the sensor is exposed to: pressure change frequency and amplitude, temperature extremes, material responses, and environmental changes.
The hub's calibration budget follows from this: the accuracy at year five is the factory accuracy plus five years of drift, not the factory accuracy. The drift must be budgeted, and the recalibration must be scheduled, before the product ships, because the customer who discovers the drift in year three has discovered that the calibration budget was never in the plan.
The event camera
The event camera, from #2, is the sensor whose spec is a different kind of lie, not because the numbers are wrong, but because the numbers that matter are not the ones on the datasheet. An event camera, also called a dynamic vision sensor (DVS) or an event-based vision sensor, does not take pictures. Each pixel works alone, watches the brightness at its own spot, and fires an event only when the brightness changes enough.
The output is not a video. It is a stream of tiny messages, each one saying where, when, and whether the light went up or down.
The datasheet for an event camera lists a resolution in pixels, because every camera datasheet lists a resolution in pixels, and the resolution in pixels is the least important number on it. What matters is the event rate (how many events per second the sensor can report before it saturates), the latency (how fast a change becomes an event), and the dynamic range (how bright and how dark the scene can be at the same time, without the pixels going blind).
A sensor with a high pixel count and a low event rate will drop events in a fast scene, and the dropped events are the measurements the product needed. The spec that lies is not the resolution. It is the assumption that the resolution is the spec that matters.
Example Labs is a made-up company we use for worked examples.
DEMO
The demo firm, as always: the Example Labs field sensor hub, 100 × 60 × 40 mm, 4-layer board, BLE at 2.4 GHz, STM32H7 with 512 kB SRAM / 2 MB flash, firmware image 384 KiB, CR2032 230 mAh = 2,484 J, deployed in a 10 m water column, with a Levelogger 6 pressure sensor. The arithmetic was run in Python before writing.
Step 1, the resolution trap. The M100's accuracy is ±0.05% FS on a 100 m full scale, ±5 cm, and its resolution is 0.0006% FS, 0.6 mm. The resolution is 83.3 times finer than the accuracy. A reading that changes by 1 mm is real (the sensor can see it), but the reading itself may be 5 cm from the true value. The MIT OCW temperature sensor is the same shape: resolution 0.00001 °C, accuracy 0.001 °C, the resolution is 100 times finer than the accuracy. The Levelogger's own temperature sensor: accuracy ±0.05 °C, resolution 0.003 °C, the resolution is 16.7 times finer than the accuracy.
Step 2, the range choice. Solinst's rule: select the shallowest range that covers your expected water column. The hub is deployed in a 10 m column. The M100 (100 m full scale) is accurate to ±5 cm, 0.50% of the 10 m reading. The M10 (10 m full scale) is accurate to ±0.5 cm, 0.05% of the reading. The M5 (5 m full scale) would be accurate to ±0.3 cm, 0.03%, but it cannot cover a 10 m column. The same sensor family, ten times better, by choosing the shallowest range that covers the column, and the M100 was the wrong choice, because the full scale was chosen for the maximum depth the product might ever see, not the depth it actually measures.
Step 3, averaging fixes noise, not systematic error. The noise falls as one over the square root of the number of readings: average 100 readings and the noise falls 10 times; average 1,000 and it falls 31.6 times. But if the hub's pressure reading is offset by +4 cm, a systematic error inside the ±5 cm accuracy band, averaging 1,000 readings gives a reading that is still +4 cm off, just more precisely wrong. The Kalman filter cannot fix it either: the filter weighs each reading by how much it trusts it, and a systematically wrong sensor with a tight noise figure looks trustworthy. The gyro demo from #17 is the same shape: a bias of 0.1°/s adds 6°/min, 360° in 1 h, and no amount of averaging fixes a bias, because the bias is the mean of the readings.
Step 4, drift, budgeted. The temperature sensor drifts at about 1–2 m°C per year (Sea-Bird), and the drift is a simple offset, predictably. After one year, 1–2 m°C of drift; after three years, 3–6 m°C; after five years, 5–10 m°C; after ten years, 10–20 m°C, against the ±0.05 °C (50 m°C) accuracy band. The drift is inside the accuracy band for temperature, but it is a systematic error, so it adds to the offset, and it does not average away. Sea-Bird's fix is an annual calibration schedule, and an annual calibration schedule will usually maintain the CTD accuracy to within 0.01 psu in salinity. The hub's calibration budget: the accuracy at year five is the factory accuracy plus five years of drift, not the factory accuracy, and the recalibration must be scheduled before the product ships.
Step 5, the barometric correction. The Levelogger measures absolute pressure, so water level data must be barometrically corrected, one Barologger 6 compensates every Levelogger within roughly a 20-mile radius or 1,000 ft of elevation change. The hub's pressure reading is absolute pressure, and without the barometric correction, a weather front moving through changes the reading by the same amount as a change in water level. The reading is stable, the reading is wrong, and the reason is not the sensor, it is the correction that was never applied.
Step 6, the selection. The hub's sensor, selected properly: the M10, not the M100 (the shallowest range that covers the 10 m column, ±0.5 cm instead of ±5 cm); the barometric correction, applied from a Barologger within 20 miles; the calibration against a known reference, done at the factory and scheduled annually; and the accuracy verified in the field conditions, the temperature, the vibration, the power supply, not the laboratory conditions. The sensor is the same family. The selection is the difference between a measurement that can be trusted and a measurement that is stable but wrong.
The decision test: five questions
A sensing lead whose product's measurements cannot be trusted can answer five questions in an afternoon. If the answer to three of them is no, the team has bought a resolution, not an accuracy, and the fix is the calibration budget, not the sensor.
- Is the accuracy stated as a percentage of full scale, and is the full scale the shallowest range that covers what you actually measure? If the accuracy is ±0.05% FS on a 100 m full scale and you measure a 10 m column, the accuracy is ±5 cm, 0.50% of your reading, and the M10 gives you ±0.5 cm for the same sensor family. Select the shallowest range that covers your expected water column.
- Is the headline number the resolution or the accuracy? If the datasheet's headline number is the resolution, 0.6 mm, 0.00001 °C, 1 ppm, the headline is the smallest detectable change, not the closeness to the true value. The resolution is 83 times finer than the accuracy on the M100, and 100 times finer on the MIT OCW temperature sensor. Finer steps does not equal accuracy.
- Was the accuracy measured under your conditions, or under nominal laboratory conditions? If the accuracy was measured at 25 °C, on a bench, with a clean power supply, and your product lives in a box that swings from −10 °C to 60 °C, the accuracy in the field is the accuracy at the field temperature, with the field vibration and the field power supply, and the datasheet does not state it. Ask for the accuracy at your temperature, your vibration, and your power supply.
- Does the datasheet assume you will perform field calibration, and did you budget for it? If the datasheet assumes the customer will perform calibration in the field, it shifts risk, cost, and accountability to the end user, and the field calibration cannot replicate factory-level thermal cycling and precision references. The calibration budget is the reference, the lab, and the annual recalibration, and it was never in the plan, because the datasheet assumed it and did not say so.
- Is the drift budgeted, and is the recalibration scheduled? If the accuracy at year five is the factory accuracy plus five years of drift, the drift must be budgeted before the product ships. A smoothly drifting sensor can be corrected for drift, Sea-Bird's temperature sensors drift 1–2 m°C per year, and the correction is a simple offset, but the correction needs a calibration, and the calibration needs a schedule. An annual calibration schedule will usually maintain the CTD accuracy to within 0.01 psu in salinity.
Objections
"The sensor is very precise, the readings are the same every time." Precision is how tightly repeated readings group, and a sensor can be precise without being accurate, a tight cluster, far from the bullseye. The reading is stable, and the stability looks like trustworthiness, but the error is systematic: an offset, a linearity error, a temperature drift, a calibration done on a different unit. Averaging the readings gives a tighter cluster in the same wrong place. The test that matters is not the stability. It is the comparison against a known reference.
"We average 1,000 readings, the noise is gone." The noise is gone, and the systematic error is still there. Averaging fixes random noise, because the random errors cancel, and it cannot fix systematic error, because the systematic error is the same in every reading. If the reading is offset by +4 cm, averaging 1,000 readings gives a reading that is still +4 cm off, just more precisely wrong. The only thing that fixes a systematic error is calibration against a known reference.
"The Kalman filter will handle it." The Kalman filter weighs each reading by how much it trusts it, and a systematically wrong sensor with a tight noise figure looks trustworthy. The filter trusts the wrong number more, not less, because the tight noise is the evidence the filter uses to decide how much to trust the reading. The filter's job is to combine a sensor that is honest with a sensor that is fast, not to make a wrong sensor right. The gyro demo from #17 is the same shape: a bias of 0.1°/s adds 360° in an hour, and no amount of filtering fixes a bias, because the bias is the mean of the readings.
"The datasheet says ±0.05%, that's good enough." ±0.05% of what? On a 100 m full scale, ±0.05% is ±5 cm, and on a 10 m column, that is 0.50% of the reading. The same percentage on a 10 m full scale is ±0.5 cm. The percentage is not the accuracy. The absolute error, at the reading you actually measure, is the accuracy, and the range choice is the difference between ±5 cm and ±0.5 cm for the same sensor family.
"We'll calibrate in the field." Field calibration can improve results for a specific installation, but it shifts risk, cost, and accountability to the end user, and it cannot replicate factory-level thermal cycling and precision references. The calibration budget is the reference, the lab, and the annual recalibration, and the customer who was not told to budget for it has not budgeted for it. The calibration budget was never in the plan, because the datasheet assumed it and did not say so.
Sources
- Solinst, "Understanding Pressure Sensor Accuracy, Precision, Resolution & Drift"
What we used
Accuracy is how closely the measured value reported by the pressure sensor comes to the actual pressure, stated as the difference between the two values, expressed as a percent of the sensor's full scale (FS); the worked example (if the Levelogger 6 M100 with a full scale of 100 m reports a pressure of 88.0500 m, and the actual pressure is 88.0000 m, then the difference is 0.0500 m (5 cm); divide 0.0500 m by 100 m and express as a percentage, accuracy (or error) is 0.05%; the Solinst Levelogger 6 and LevelVent 6 have an accuracy of ±0.05% FS); precision is how close all the sensor readings are grouped to each other, a more precise sensor has a narrow distribution of readings at a given pressure, and a precise pressure transducer produces consistently repeatable measurements; resolution is the smallest degree of pressure change that can be detected and recorded by the pressure sensor (the Levelogger 6 M100 has a resolution of 0.0006% FS, with a full scale of 100 m, the sensor will be able to detect a 0.0006 m (0.6 mm) change in pressure; the Levelogger 6 M5 has a resolution of 0.001% FS, or 0.05 mm); all pressure transducers, no matter what they are made of, how expensive they are, or how accurate, are susceptible to sensor drift over time, and pressure sensor drift is a gradual degradation of the sensor and other components that can make readings offset from the original calibrated state; some manufacturers provide a value for the expected drift or long term stability, but these numbers are still based on use in very stable environments, making them somewhat irrelevant; the Solinst datalogger calibration procedure ensures linearity and hysteresis specifications are met
- Solinst, Levelogger 6 LTC datasheet
What we used
The accuracy table (M5 ±0.3 cm, resolution 0.001% FS; M10 ±0.5 cm, 0.0006% FS; M20 ±1 cm, 0.0006% FS; M30 ±1.5 cm, 0.0006% FS; M100 ±5 cm, 0.0006% FS; M200 ±10 cm, 0.0006% FS); level sensor accuracy ±0.05% FS, resolution 0.001% FS to 0.0006% FS; temperature sensor accuracy ±0.05 °C, resolution 0.003 °C; clock accuracy ±1 minute/year (−20 °C to 80 °C); the Levelogger measures absolute pressure, so water level data must be barometrically corrected, one Barologger 6 compensates every Levelogger within roughly a 20-mile radius or 1,000 ft of elevation change; a shorter range gives a tighter absolute accuracy, select the shallowest range that covers your expected water column
- Wikiversity, "Accuracy vs resolution" (inclinometer specifications)
What we used
Inclinometer specifications are often reported in ways that can be misleading, some quote only best-case laboratory results, others substitute resolution for true accuracy, and still others assume that end-users will perform field calibration to achieve stated performance; specmanship is the practice of presenting product specifications in a way that emphasizes favorable values while omitting limitations, and in sensor datasheets it may involve presenting resolution as if it were accuracy, reporting accuracy only under nominal laboratory conditions, and basing claims on prototypes instead of production units; many datasheets highlight best-case results in stable laboratory environments, omitting the effects of temperature swings, vibration, and power supply variation; accuracy is closeness of the measurement to the true angle, factoring in offset, linearity, thermal drift, hysteresis, and noise, and is reported only under laboratory conditions; resolution is the smallest detectable change in angle, marketed as accuracy, and finer steps does not equal accuracy; repeatability is the ability to return the same reading under identical conditions, presented as accuracy, and ignores actual error; accuracy depends on offset error, linearity error, temperature drift, hysteresis, noise, and environmental stress factors; resolution depends on sensor noise characteristics, internal filtering and bandwidth, and ADC bit depth; a sensor may have high resolution but poor accuracy, or vice versa, and resolution describes what a sensor can detect, not how correct the measurement is; some datasheets assume that the customer will perform calibration in the field by taking readings in the field (super calibration), and while this may improve results for a specific installation, it shifts risk, cost, and accountability to the end user
- MIT OCW, "Sensor Characteristics"
What we used
(Ocean Instrumentation, Course 13.998, Jim Irish, WHOI, 10 Feb 2005), accuracy is how well the sensor measures the environment in an absolute sense, how good the data is when compared with a recognized standard (a temperature sensor accurate to 0.001 °C is expected to agree within 0.001 °C with a temperature standard such as a triple-point-of-water cell); resolution is the ability of a sensor to see small differences in readings (a temperature sensor may have a resolution of 0.00001 °C, but only be accurate to 0.001 °C, you can believe the size of relative small changes in temperature, which are smaller than the accuracy of the sensor; resolution is often controlled by the quantization in digitizing the signal, one bit is equal to 0.0005 °C, which is not a function of the sensor, but the sampling process); repeatability is the ability of a sensor to repeat a measurement when put back in the same environment, and it is often directly related to accuracy, but a sensor can be inaccurate, yet be repeatable in making observations; drift is the low frequency change in a sensor with time, often associated with electronic aging of components or reference standards in the sensor, and drift generally decreases with the age of a sensor as the component parts mature, a smoothly drifting sensor can be corrected for drift (Sea-Bird temperature sensors that are drifting about 1 m°C/yr, and have been smoothly changing for several years, allow one to correct for the drift and get more accurate readings), and drift is also caused by biofouling that can't be properly corrected for, but is often tried; hysteresis is a linear up and down input to a sensor resulting in an output that lags the input (one curve on increasing pressure and another on decreasing), and many pressure sensors have this problem
- Sea-Bird, "How often do I need to have my instrument and/or auxiliary sensors recalibrated?"
What we used
The general character of sensor drift in Sea-Bird conductivity, temperature, and pressure measurements is well known and predictable, but it is very difficult to know precisely how far a CTD calibration has drifted over time unless you have access to a very sophisticated calibration lab; temperature generally drifts slowly, at a steady rate and predictably as a simple offset at the rate of about 1–2 millidegrees per year (approximately equal to 1–2 parts per million in Salinity error); a profiling CTD is recalibrated once a year, and an annual calibration schedule will usually maintain the CTD accuracy to within 0.01 psu in Salinity; pressure sensor drift is also an offset, and annual comparisons to an accurate barometer to determine offset will generally keep the sensor within specification for several years, particularly as the sensors age over time
- Sea-Bird, Newsletter 5 (August 1995)
What we used
The new calibration system produces temperature calibrations with resolution and accuracy not previously available; the SBE 3 sensor's plots show a drift of less than 0.0004 °C, with a clearly discernible +1.2 mK per year drift rate that could not be judged in conventional baths typically exhibiting variance at 1 mK+ levels; the metrology laboratory achieves temperature precision of 50 µK and accuracy of 0.0005 °C
- NiuBoL, "Sensor Precision vs. Resolution"
What we used
A temperature channel displaying 25.01 °C is not necessarily more accurate than one displaying 25.0 °C, and the extra digit may describe formatting rather than useful measurement information; a stable reading can be consistently offset from a suitable reference; the CO2 example (both versions have 1 ppm resolution, but the 0–2000 ppm version is ±(40 ppm + 2% F.S.) = ±80 ppm, and the 0–5000 ppm version is ±(50 ppm + 3% of reading), ±80 ppm at 1000 ppm, ±170 ppm at 4000 ppm; a 1 ppm output step does not imply a ±1 ppm measurement tolerance)
- Example Labs figures from the cluster: the hub (100 × 60 × 40 mm, 4-layer board, BLE at 2.4 GHz, STM32H7 with 512 kB SRAM / 2 MB flash, firmware image 384 KiB, CR2032 230 mAh = 2,484 J) (#6/#11/#28/#37); the gyro bias of 0.1°/s adds 6°/min, 60° in 10 min, 360° in 1 h, and bias stability is a 1σ spread read as the minimum of the Allan deviation curve that cannot be calibrated away (#17); the Kalman filter weighs each reading by how much it trusts it, and one sensor is never enough
What we used
GPS is honest but slow, an IMU is fast but drifts 15 metres in 30 seconds (#4); the event camera does not take pictures, each pixel fires an event only when the brightness changes enough, and the output is a stream of tiny messages (#2)