Detecting Anomalous Behavior Towards Predictive Maintenance

Abstract

A key Industry 4.0 element is predictive maintenance, which leverages machine learning, IoT and big data applications to ensure that the required equipment is fully functional at all times. In this work, we present a case study of smart maintenance in a real-world setting. The rationale is to depart from model-based and simple rule-based techniques and adopt an approach, which detects anomalous events in an unsupervised manner. Further, we explore how incorporation of domain knowledge can assist the unsupervised anomaly detection process and we discuss practical issues.

Appendix

Distance-Based Outlier Definition. A data point that has less than k neighbours inside a radius R, is called a distance-based outlier [7]. Figure 3 shows an example of a dataset that has two outliers with k = 4. The points o1 and o2 are outliers since they have 3 and 1 neighbours, respectively inside the R radius. In a data stream, we assume that we keep in a sliding window the most recent points, and the challenge is to continuously report all the outliers among the objects in that window. Apart from the technique in [9], additional streaming solutions are proposed in proposals, such as [3, 5].

Fig. 3.

Outliers example.

Fig. 4.

MCOD - Micro-clusters.

MCOD Algorithm. Finding the neighbours of each alive data object in a streaming scenario is a particularly computation-intensive process. The MCOD algorithm uses micro-clusters, as depicted in Fig. 4 (i.e. MC1. MC2, MC3) of radius $R{/}2$, inside which all the data points are inliers. Hence it alleviates the need of (re-)computing distances between all the data points on every window movement. The rationale is that, normally, most data points fall into one of such clusters and thus need not be further processed. Therefore, only a small portion of all objects needs to be examined. More details about the algorithm’s functionality are provided in [9].