Anomaly detection can prevent damage and accidents

Anomaly detection means distinguishing data that is statistically unusual or different from a large amount of data.

"You want to find cases that stand out among large numbers of 'normal' cases," says Anders Holst, research leader in advanced data analysis at the Centre for Applied AI at RISE.

Statistical machine learning, which is an area of artificial intelligence (AI), is used for anomaly detection. It is a method that has been used for a long time, unlike Deep Learning, which has come on strong in recent years.

"We have worked with deviation detection in close co-operation with different companies since 2001, and it can be applied in a number of different areas," Anders continues.

"One application can be to find faults in equipment at an early stage to avoid more serious faults later on. What you use to find the deviations are various signals from the equipment, such as logged communication messages.

"By analysing the frequency of different log messages for trains or trucks, for example, you can identify unusual anomalies that require action. This could be anything from detecting broken smoke alarms to brake systems that need servicing. "It is important to find remarkable deviations that could otherwise cause major damage or accidents later," says Anders.

Accidents that can affect the environment are an example that society really wants to avoid. This may involve preventing ships from running aground and causing oil spills, in which case detection of anomalous ship movements is used. If an anomalous movement pattern is detected, such as a ship no longer following the fairway but heading towards shallow water, an alarm can be sent to help avoid grounding. Such a system has been developed by RISE in collaboration with the Coast Guard and has been in operation since 2016. The methods are further developed in the EU project AI-ARC.

"The biggest challenge in anomaly detection is choosing which parameters to investigate. Once you define that, the actual AI modelling is easier, and you can find what you are looking for," Anders explains.

Parameters can be all kinds of data - such as height, weight, speed, movement, etc. The basic thing is always to set a normal value, and what range counts as normal. If a value suddenly goes out of range, it may indicate an error. In more advanced anomaly detection, with several interacting parameters, one looks in a similar way at normal and anomalous combinations of parameters.

"We at RISE are method experts, but we always need to collaborate with domain experts who understand what indications may be present there, i.e. which parameters to consider. Together we can find interesting deviations to investigate," says Anders.

Want to know more about anomaly detection and AI?