|dc.description.abstract||Anomaly detection techniques are supposed to identify anomalies from loads of seemingly homogeneous data and being able to do so can lead us to timely, pivotal and actionable decisions, saving us from potential human, financial and informational loss. In anomaly detection, an often encountered situation is the absence of prior knowledge about the nature of anomalies. Such circumstances advocate for ‘unsupervised’ learning-based anomaly detection techniques. Compared to its ‘supervised’ counterpart, which possesses the luxury to utilize a labeled training dataset containing both normal and anomalous samples, unsupervised problems are far more difficult. Moreover, high dimensional streaming data from tons of interconnected sensors present in modern day industries makes the task more challenging. To carry out an investigative effort to address these challenges is the overarching theme of this dissertation.
In this dissertation, the fundamental issue of similarity measure among observations, which is a central piece in any anomaly detection techniques, is reassessed. Manifold hypotheses suggests the possibility of low dimensional manifold structure embedded in high dimensional data. In the presence of such structured space, traditional similarity measures fail to measure the true intrinsic similarity. In light of this revelation, reevaluating the notion of similarity measure seems more pressing rather than providing incremental improvements over any of the existing techniques. A graph theoretic similarity measure is proposed to differentiate and thus identify the anomalies from normal observations. Specifically, the minimum spanning tree (MST), a graph-based approach is proposed to approximate the similarities among data points in the presence of high dimensional structured space. It can track the structure of the embedded manifold better than the existing measures and help to distinguish the anomalies from normal observations. This dissertation investigates further three different aspects of the anomaly detection problem and develops three sets of solution approaches with all of them revolving around the newly proposed MST based similarity measure.
In the first part of the dissertation, a local MST (LoMST) based anomaly detection approach is proposed to detect anomalies using the data in the original space. A two-step procedure is developed to detect both cluster and point anomalies. The next two sets of methods are proposed in the subsequent two parts of the dissertation, for anomaly detection in reduced data space. In the second part of the dissertation, a neighborhood structure assisted version of the nonnegative matrix factorization approach (NS-NMF) is proposed. To detect anomalies, it uses the neighborhood information captured by a sparse MST similarity matrix along with the original attribute information. To meet the industry demands, the online version of both LoMST and NS-NMF is also developed for real-time anomaly detection. In the last part of the dissertation, a graph regularized autoencoder is proposed which uses an MST regularizer in addition to the original loss function and is thus capable of maintaining the local invariance property. All of the approaches proposed in the dissertation are tested on 20 benchmark datasets and one real-life hydropower dataset. When compared with the state of art approaches, all three approaches produce statistically significant better outcomes.
“Industry 4.0” is a reality now and it calls for anomaly detection techniques capable of processing a large amount of high dimensional data generated in real-time. The proposed MST based similarity measure followed by the individual techniques developed in this dissertation are equipped to tackle each of these issues and provide an effective and reliable real-time anomaly identification platform.||