In technical writings, we often presume readers equipped with basic knowledge about the matter being dealt. Readers of peer-reviewed journals almost agrees on terms and concepts to start with. Veteran readers might agree (and disagree) on selected methods. Researchers in that topics often disagree on interpretation or assumption the authors projected for a final concluding thought. That show a simple pyramid of agreement with a piece of well-written, source-checked and month-to-year of writing and updating.
This post will start with some terms with the device or the information, including:
[Merriam-Webster] /ˈa-kyə-rə-sē/: conformity to truth or to a standard or model.
By far, accurary is one word you should find in any datasheet of a sensor, device or an instrument. For example:
- DS18B20, a low-cost digital temperature sensor stating ± 0.5°C accuracy from -10°C to +85°C
- Sensirion SHT35 indicating typical accuracy of ±1.5 %RH and ±0.1 °C
- Nova Fitness SDS011, a PM2.5 sensor included this line Accurate and Reliable: laser detection, stable, good consistency but no numerical stated.
- Calibration accuracy: a deviation (and correctiveness) of an individual sensor in equilibrium state against a high precision reference
- Hysteresis: the difference of measured value of the same sensor in different environments. This difference is due to a carry-over from the previous condition. For example, when temperature increased, the sensor reading was lower than the true value because it requires sometimes for the whole system is in equilibrium with the new condition. Inside a dust sensor, an old volume of air is not completed change out and the reading concentration reflected almost new sample of air but not entirely.
- Long-term drift: mainly due to aging of parts. Dust sensor, for example, is made of three main parts: a laser source and a photo-detector, and a fan for flow control. The drift can be random or one direction. Laser aged weaker as more in operational so does the sensitivity of the photo-detector. A fan is possibly draw less volume but the bearing is also less predictable making the volume draw less concise. The total variation of different parts of aging is unpredictable.
The accuracy of a device should indicate a range deviation that the reading should fall within with a certain confidence level. Assume that a large number of sensors were tested and the deviation is by the class of the device itself, we could use Gaussian model to assert with 95% confidence level with a range of ± 2σ from the true value. The manufacture of the device should indicate the confidence level associated with an error.
[Merrian-Webster] \pri-ˈsi-zhən\: the degree of refinement with which an operation is performed or a measurement stated.
With ISO 5725, the distinction between accuracy and precision made clear.
Accuracy is about the closeness to the true value and the precision is how each reading or measurement is close to each other. A good device should have high accuracy and precision. That means the bell curve will be tall and thin.
A device can be highly accuracy but not precise (C), precise but not accurate (B), neither (A) or both (D).
A class of device said to precision if the reading are very close to each other. A high precision equipment, such as scales, produce a reading with more decimal point of each reading which means they can measure a very light weight in µg. Precision reading can be confused with repeatability, in which, in the same environmental conditions, the results are identical.
Understanding possible confusion and distinction is not always easy or convenience, but a clear explanation should help us to align technical terms to a general discussion.
[Merriam-Webster] \rō-ˈbəst\: capable of performing without failure under a wide range of conditions.
A robust device or a method is expected be in operation with minimal setup and/or wide-range of condition. For example, light-scattering dust sensor can be considered as robust device in compared to Beta Attenuation Monitor (BAM) in the context that such sensor can be plugged in power source and in operational with a minimal installation. Those sensors can produce readings continually without much adjustment and service during week-to-month operation.
The robustness often is being trade-off with accuracy because the device is designed to work in various conditions, more simple, less if none redundancy and target for wide-range of deloyment rather a reference equipment.
[Merrian-Webster] \ ri-ˈlī-ə-bəl\: giving the same result on successive trials.
giving the same result sounds similar to both accuracy and precision. Reliable is often used to describe a device in operation. A device said to be reliable when in operation, the result is consistent in a wide-range environmental condition. A reliable data source implies the content to be taken at the face value with its history to be trusted. Reliable or reliability is an important character for a system, either a device or source of data, so that other users can used them without a technical expertise.
Reliability is a desired character of any system which giving users the reading or data as accurate, trusted as it is capable of. To make a reliable system, it often requires a complex engineering approaches such as built-in redundancy, high-quality parts and regular preventive maintenance.
In 2020, a whole range of PM2.5 air quality sensors are available to ordinary users such as Nova Fitness SDS011, Plantower PMS7003/5003, Winsen ZH03B, Sensirion SPS30. They are low-cost sensors ranging from $10-30, and could be lower if buying in bulk. These sensors always have the word accurate in the datasheet with measurement range, deviation, technical lifetime. Although, they are all designed for PM2.5, readings between each sensors in the similar environmental condition is not the same. They can can be said to be precision, in the context of reproductivity, or with a new batch of sensors, the reading tends to be close to each other. The accuracy is difficult to determine because the reference equipment is very expensive, both in capital and in operation. The local variation and interval reading add another layers to compare a low-cost sensors to some publicly available PM2.5 stations.
The low-cost sensor is considered to be robust since the build is simple and minimal and can work in different condition with almost "plug-and-play" setup. The robustness is a plus but the down side is that we have to reserve for the accuracy of the reading itself. The reliability of low-cost sensor is one of the biggest disadvantages. The design of the sensor is minimal. When sensor is operation, the regular check and maintenance is at whim of the operators.