Anomaly Detection in Sabisu

Anomaly detection of time series data is a complex problem, especially when the data is subject to measurement error and noise. First you must decide what constitutes an “anomaly”, usually through the construction of a statistical representation of past behaviour. New data is then compared to this statistical model to identify significant changes in behaviour. The anomaly definition is important, as if it is wrong important failures will be missed, or normal behaviour will be flagged as anomalous. Take the signal in Fig. 1, it might be obvious that point A should be classed as an anomaly, but what about point B?

Example time series with anomalies

Fig 1. Example time series with a clear anomaly (A), and a potential anomaly (B).

Sabisu’s anomaly detection system can be used to quickly and reliably identify points of change in time series data. This means that it can be applied to most process data collected on a plant. However, when combined with metasensors®, you have a system that can intelligently determine when something is going wrong anywhere on the site.

Sabisu uses a combination of in-house and cutting edge algorithms to build a statistical picture of your data in real time. It is possible to use these algorithms individually, or combine several of them through a quorum voting system to account for signal noise and variance.

The algorithms can be finely tuned to your process data, or more commonly the metasensors® encoding this data. The application of anomaly detection to metasensors® provides a convenient way to monitor assets, detecting faults at an early stage. The anomaly detection system will then automatically notify the relevant Sabisu users of the potential problem. For continuous monitoring the system can also send regular digest emails summarising the number of anomalies detected across the plant assets.

Contact us

We’re always interested in hearing from you with any comments or suggestions, feel free to get in touch.

Start typing and press Enter to search