What is Anomaly Detection?
Anomaly detection is a method used to identify irregular or unusual patterns in a complex environment. It is recommended for systems and environments that measure multiple, interdependent variables, and where it makes a difference to catch non-conformances before they result in system failure, damaged goods and parts, or the restart of expensive procedures.
How does this differ from threshold monitoring?
Threshold monitoring looks for anomalous values that are above or below user-defined levels. This can be extended to more complex logic, but is generally for making sure the data values perform within their expected envelopes. This is good for catching single point failures.
In contrast, anomaly detection takes in multiple variables from across a range of sources, and trains machine learning algorithms to identify regular patterns within the datasets based on a statistical understanding of their performance. It identifies issues that human observation and threshold monitoring will normally miss or identify too late, due to the complexity of the signal interactions. It is good for catching system-wide issues and identifying local failures before they cause damage.
What are some examples of anomaly detection?
One example is within a mechanical system containing rotating parts. It may contain several vibration sensors, temperature sensors, and rotational speed sensors, among others. If a gearbox were to fail, threshold alerts would pick up the issue upon failure, or just before failure. However, with anomaly detection, no thresholds would be exceeded at first, but the combination of vibration, temperature, output and time of day may be considered unusual by the detector, triggering an inspection and potentially preventing catastrophic failure. This may even occur earlier than an equivalent detection picked up by a dedicated human observer.
Another example is in an urban farm environment, where it is important to carefully maintain the environmental variables in the system (e.g. temperature, humidity, CO2). Imagine a door is opened on a hot, humid day, and the control systems in place cannot cope with the change in conditions, sending temperature, humidity and CO2 towards the edges of their thresholds. Although no threshold is breached, this is no longer an ideal scenario because:
1. The conditions are no longer optimal for the plants, and
2. The control system is struggling and may need repairing or modifying.
However, a good anomaly detection unit would catch very early on that all variables have begun deviating at the same time, and will send a warning despite no threshold being reached. This would prevent suboptimal plant growth, provide evidence of a faulty or struggling control system, and prevent expensive loss of product.
Should I implement anomaly detection in my system?
- Have a high number of sensors
- Have a system where the signals come from similar regions and environments or show signs of interdependency
- Can collect enough data over time to train the detector
then your system would probably benefit from anomaly detection implementation.
At Datch, we recommend that you define a metric, such as downtime, failure rate, or yield. When you deploy your anomaly detection tool, you can then observe the change in your metric through its use, and find out what difference it has made.
If you would like to talk to us about implementing anomaly detection into your system, please contact us at firstname.lastname@example.org.