Anomaly Detection – A precursor to Predictive Maintenance?

Anomaly detection plays a crucial role through identifying deviations in telemetry data that may indicate an issue emerging, but not all anomalies signify failure – some are just noise that’s inherent when dealing with data. The real challenge lies in distinguishing between benign and problematic anomalies to generate the high quality training data required for PdM models.

The Role of Anomaly Detection

Anomaly Detection involves continuously monitoring asset telemetry to identify any irregularities that could signal potential failure. It uses sensors and machine learning models to detect deviations away from expected behaviour. However, anomaly detection on its own does not predict when failures will occur – it simply flags abnormalities that require further investigation.

One of the key issues with traditional anomaly detection is the high rate of false positives. Not all detected anomalies indicate failure or impending failure, and responding to every anomaly can lead to unnecessary actions and increased costs.

Taking for example the graph shown below, which visualises a number of telemetry variables against each other. The red dots representing detected anomalies while the blue dots representing normal operation. There are two distinct groupings of anomalies in general – but without further context, we can’t determine which ones are truly problematic.

Figure 1: Pairs Plot of refrigeration telemetry variables anomaly detection results without severity scoring (Blue = Non Anomalous / Red = Anomalous)

How to truly add value through Anomaly Detection

To make anomaly detection more actionable for us, applying a layer of analytics over the top is critical. An effective approach is a severity scoring system, such as a red, amber and green framework which helps prioritise which anomalies require attention. This reduces the unnecessary response whilst ensuring genuine faults are addressed and helps us build that high quality data set for further use.

With this layered approach, the true problematic anomalies become more distinct and easier to target as seen in the improved and enhanced graph below.

Figure 3: Pairs Plot of refrigeration telemetry variables anomaly detection results with severity scoring (Blue = Non Anomalous / Green = Low Severity / Amber = Medium Severity / Red = High Severity)

The Transition to Predictive Maintenance

While anomaly detection provides immediate operational insights, its true value lies in its ability to serve as a precursor to predictive maintenance. By generating high-quality training data for machine learning models, anomaly detection lays the foundation for predictive algorithms that can estimate failure timelines and recommend optimal intervention points.

Any supervised machine learning approach relies on well-labelled failure data to develop robust predictive models and a layered approach to anomaly detection serves this purpose. Without reliable anomaly detection, predictive maintenance efforts can be hindered unless 3^rd party data exists from CAFM systems for example.

Summary

Anomaly detection is more than just a tool for monitoring, but a critical stepping stone towards predictive maintenance efforts. Through refining anomaly detection with a layered approach, businesses can generate high quality failure data that’s needed to train the predictive models. This self-contained approach eliminates reliance on external failure data sources and enhances the overall efficiency of getting predictive maintenance in place.

As industry continues to integrate artificial intelligence and machine learning insights into maintenance strategies, the shift from anomaly detection to predictive maintenance represents a major leap towards minimising downtime, reducing costs and maximising assets performance.

The key take away here is therefore – before you can actually predict failures, you first need to understand your failures – anomaly detection can help you with this.