Within the context of Industry 4.0 the term “anomaly” occurs very often and refers to many different incidents, depending on the branch and the problem.
It might refer to sensor measurements, hacker attacks, credit card fraud or defect and abraded machines. The general definition says that anomalies are rare and distinctively different occurrences.
However, due to differentiated problems the requirements to anomaly detection methods grow. A precise understanding of the process and the data helps divide anomalies into the categories selective,context basedand collective anomalies. Selective anomalies refer to data points which are distinctively different from normal observations. Most often, such data points are seen as outliers which can be identified with statistical tests, threshold values and visualization methods. Context based anomalies show data points which—taken in isolation—occur in a valid range, but in an unusual combination, for example, high energy consumption in spite of a down time. The third type, collective anomalies, occurs within sequential data sets. Here it is also the case that the single data points cannot be identified as anomalies, but the accumulation of specific successive measurements is, considering the whole data set, abnormal. The methods for the identification of the three types are manifold: they range from visualization, clustering and residual approaches to complex method combinations and should/have to be adjusted individually.