How might we identify drifts or inaccuracies in sensor readings?

Accurate and well-calibrated sensors are critical for the operation of the water treatment plants run by PUB, especially in the area of water quality control. Over time, sensors measuring crucial parameters, such as pH and total residual chlorine, could drift in their readings. It is difficult for operators to detect these changes in a timely manner, especially if the drift happens very slowly. Current methods of detection include doing a comparison test with the spare sensor(s), verifying the reading through lab tests, and manually analysing the sensor data. Sensor drifts affect the operators' decision making and even the process control systems.

To mitigate the problem of sensor drift, calibrations are done at a predetermined interval. However, as sensors age or if the water they are analysing has variations in water quality, the sensor drift could occur before they are due for calibration. In addition, an improper calibration that is not noticed by staff may also lead to premature inaccuracies in the sensor.

We are looking for solutions that can help to identify inaccurate sensor readings in near real-time and verify if calibration has been done properly. We are open to various solutions, including those which use computational or statistical correlation-based methods (i.e. soft sensors) and pattern-based matching methods to model sensor readings by processing the combined historical data of multiple adjacent and associated sensors in that part of the water treatment process. The soft sensors would assist to validate the readings produced by the physical sensors and pinpoint those that require calibration or further checks.

The proposals also need to consider secured methods of extracting the data from the existing supervisory control and data acquisition (SCADA) system, if this is required.

• There is a large variety of sensors in a typical water treatment plant. They measure a range of physical and chemical parameters (e.g. pH, turbidity, conductivity, chlorine concentration, etc.) at various points in the overall water treatment process. The solution should be able to model the behaviours of many different types of online/real-time sensors.
• The inter-relationships between different physical or chemical parameters are not always known. The solution must be able to learn and develop its own mathematical or network models and identify the relationships or patterns of data with little or no prior knowledge.
• Where present, the solution could make use of the data from a backup or parallel sensor to validate whether a particular sensor is drifting, but it must also be able to operate in the absence of a backup or parallel sensor, in the event that no such sensor is available.
• The solution should be intuitive, requiring little operator intervention.

After Selection

• PUB water treatment plants as test sites
• Sensor datasets from the chosen test sites

Water Supply (Plants) Department

Total project period - less than 12 months
Data analysis and solution development - 4 months
Testing and review of the solution - 6 months

A solution that is verified and validated at a test site within PUB, with a clear plan to scale to other PUB sites.

Adasa, Spain

Mott MacDonald, United Kingdom