Imagine being an operator flooded with excessive and irrelevant SCADA alarms, struggling to make-real time decisions. This is the reality for many, but it doesn't have to be. Not when they have the right tool by their side.
SCADA (Supervisory Control and Data Acquisition) systems play a crucial role in monitoring and controlling any kind of complex equipment found, for instance, in power plants and industrial factories.
To serve this purpose, any SCADA system comprises two crucial components: data and events. The data component enables access to sensor data and set-points, while the event component provides logging information related to equipment activities. Event data are usually defined as:
Thanks to these functionalities, one could assume an operator of a complex piece of equipment (say a wind turbine) is equipped to make well-informed real-time decisions on its operations, right? Unfortunately, that’s not the case.
As a matter of fact, the current state of SCADA alarms is overwhelming operators with excessive and often irrelevant notifications showing the limitations of traditional rule-based SCADA alarms.
In this blog post, we explore and propose a more effective and context-aware approach using machine learning.
Traditional SCADA alarms are based on single sensor threshold-based rules as the ones listed here below:
Although narrowly useful, these alarms are very limited as this approach fails to consider crucial contextual factors such as environmental conditions, operational conditions, and asset-specific characteristics. We can see in Figure 1 that threshold-based rules can very easily lead to either false positive or negative alarms. As a result, operators are exposed to numerous false alarms and an overwhelming amount of noise, leading to alarm fatigue.
Alarm fatigue desensitises operators, causing them to ignore potentially critical alarms. This effect is amplified in sensors with higher variance or fast dynamics. We illustrate this issue through Figure 2, where the excessive number of alarms generated creates more noise than useful information.
To address alarm fatigue and improve the effectiveness of SCADA alarms, we propose a shift towards more complete and context-aware rules. This involves considering seasonality, control options, environmental conditions, and unique characteristics of each asset and sensor which, even if from the same make and model, have slightly different and unique behaviour as seen in Figure 3.
Obviously, configuring and maintaining a large number of complex rules manually is unsustainable and costly so, instead of manually configuring rules, we use machine learning. As a matter of fact, by leveraging historical data and developing digital twins for individual assets, we can create models that understand each asset's behaviour within its unique operational and environmental context.
Based on various SCADA sensors like active power and wind speed, we can model the normal behaviour of various components in wind turbines. These are called “normality models” and provide dynamic and context-aware thresholds, enabling us to create meaningful alarms whenever we detect any sensor’s reading deviating from its normal behaviour.
These alarms, together with other crucial metrics that help prioritise issues (for example by sorting on asset performance), are then made easily accessible to our customers via Canopy, our powerful web application.
Canopy’s context-aware alarms (aka Detections) ensure operators receive fewer but more relevant notifications, reducing alarm fatigue and enabling operators to make well-informed real-time decisions. But, it does not stop there. In Canopy, our users also have multiple functionalities available to further investigate each alarm, down to the individual sensor level!
Figure 5 demonstrates how a Canopy user can easily inspect what was at the origin of a “generator overheating” alarm by using the “Time-series” plotting module. In this case, it shows the behaviour of two generator-bearing temperature sensors together with the normal operating range that is predicted by our models. By simply comparing the two, it is now easy to detect the periods when the sensors deviate from their normal range and therefore, cause an alarm.
The current state of SCADA alarms inundates operators with excessive and often irrelevant notifications, leading to alarm fatigue and an inability to make informed decisions. By shifting towards context-aware alarms using machine learning-based models, we can reduce alarm fatigue and provide operators with meaningful notifications that take into account the unique operational and environmental context of each asset. This approach, which can now be easily adopted by anyone thanks to Canopy, not only improves operator efficiency but also ensures the overall effectiveness of SCADA systems in monitoring and controlling complex equipment like wind turbines, solar panels or even entire industrial processes.
Keep in mind that we can show you around! If you want to learn more about how we're implementing alarm best practices in Canopy, get in touch with our team: sales@jungle.ai
