If you were a patient at a hospital, wouldn’t you want the medical staff to respond swiftly to machines bleeping at your bedside? I would. However, as it turns out, care givers are exposed to up to 1,000 alarms per day, with up to 99% of those not requiring clinical intervention. Medical staff shows its experience and expertise not by responding adequately to all these bleep-bleeps and ding-dongs, but by knowing which alarms to follow up immediately, which ones to postpone, and which ones to ignore altogether.
The same is true for online water quality monitoring. Over the years, more and more online monitoring equipment has been installed to keep a watchful eye on our water quality. When certain thresholds are exceeded, an alarm is sounded, either via a central control room or directly on your phone. But like in hospitals, staff for following up on such alarms is limited, and – also like in hospitals – water utility or authority staff find themselves prioritising alarms into categories ranging from ‘follow up urgently’ to ‘ignore’. Several weeks ago, Sensileau co-organised a workshop on the installation and daily operation of online water quality sensors, and several participants raised this issue. It strikes me as very odd.
The purpose of using online monitoring equipment is to collect relevant water quality information, and to receive a signal when a serious water quality issue occurs, so that actions can be taken adequately and in a timely fashion. The whole idea of this setup is to filter out irrelevant information and focus attention on issues of real and immediate concern. So why are we still bombarded with alarms which do not require our (immediate) attention? Just to be clear: these are not false alarms, these are real alarms indicating a (rapid) change in water quality. But because an estimated 75-90% of all alarms signify a water quality event for which no meaningful action can be taken, we are still left with the task of manually selecting the alarms we need to respond to.
Besides being a nuisance, all this unnecessary use of alarm bells adds to a perceived workload, and can be risky. A multitude of alarms lures us into thinking that the system is safely up-and-running, but in fact it creates a false sense of security. We think we will be able to pick out the alarms that really matter, but are we? What is the risk of serious alarms being overlooked simply because there are so many alarms going off? And what if an alarm pulls us away from other important duties, only to find out there was nothing we could have done about it after all? Or even worse: what if we receive so many alarms that we stop responding altogether? In that case, we are diagnosed with a condition called alarm fatigue.
Unfortunately, in many cases, we have ourselves to blame. When installing online monitoring equipment and setting alarm thresholds, we tend to think “better safe than sorry.” As a result, alarm thresholds are set more broadly than necessary resulting in alarms which do not require human intervention. Besides reducing the efficiency of alarm handling and adding to the workload of a limited number of staff, this may actually increase the risk of serious water quality events going unnoticed rather than reducing them, for reasons as described above.
The solution to this sounds easier than it is: determine the right thresholds and set the alarm limits accordingly. However, it often proves difficult to work out what exactly the right alarm thresholds are, and this may be an iterative process requiring a substantial amount of time to get it right. If anything, it requires regular and thorough data evaluation over a prolonged period of time in order to determine when a change in water quality requires human action. Yet, it is still worth investing the time to do this well so that the sheer number of alarms can be reduced, making the relevant ones stand out. A positive side effect of this is that we will start to take the alarms that do occur more seriously.
After the number of alarms has been reduced to the minimum by setting the right alarm thresholds, we can start working on increasing the value and relevance of serious alarms. I will discuss two ways in which this can be done, although there are probably more.
The first option is – again – thorough data evaluation; but this time in comparison to other datasets. By comparing online sensor data with e.g. a database of weather and precipitation data, we can filter out events related to rainfall events affecting surface water quality. A water quality sensor in a river is generally not installed in order to determine when it rains, and when water quality changes occur due to rainfall in the area, we may not necessarily want to know. By combining our sensor data with weather data, we can potentially determine specific patterns related to rainfall events, and label them accordingly. We may also be able to set our alarm thresholds such that alarms labelled ‘rainfall event’ are not forwarded to our phones for follow-up. This further reduces the total number of alarms, and at the same time increases the importance of the remaining alarms.
The second option is to compare sensor data of multiple instruments, be it different parameters measured at a single location, or one parameter measured at various locations within a certain area. If more than one parameter at a certain location shows a (rapid) change or if similar sensors within an area show changes at the same time (or according to water flow), it can be given a higher priority than on an individual basis alone.
Whether you go for the first or the second option (or both), getting some help is always a good idea. The water industry is rapidly catching up on the digital transformation, and machine learning is quickly becoming a very powerful support tool for online water quality monitoring. Machine learning is closely related to computational statistics and prediction modelling, and allows a machine (the computer) to learn from experience while evaluating (large) datasets. In this way, the computer can learn to distinguish specific patterns in water quality data autonomously, such as the abovementioned example of the rainfall events. The more datasets the computer evaluates, the better it becomes at detecting specific patterns and labelling them accordingly. If you are interested in machine learning and how to apply it to water data, you can check e.g. Muharemi et al. (2019) or Di et al. (2019).
If you feel you are receiving too many water quality alarms, or are at risk of developing alarm fatigue, now is the time to do something about it. Take a closer look at the sensor(s) causing too many alarms. Is it the right sensor for this purpose? You can check this using one of our step-by-step guides. And if it is the right type of equipment, are you sure it is in the right location? If all this is the case, take a closer look at the data produced by this sensor to see if and how the number of alarms can be reduced. Feel free to ask us for help if you need it. And although this requires a considerable amount of effort, I am confident that in the end, you will agree that it’s worth it. Let me know how you are getting on.