Open set anomaly classification

Abstract

Anomaly classification can provide a discrimination of the type of anomaly at hand, that can help the user to narrow down the search space during troubleshooting. A limitation of most machine learning (ML) classification models is being under the closed set assumption and thereby failing to recognize unknown data. Literature refers to this issue as the Opens Set Recognition (OSR) problem. In this paper, the given use case is the detection of technical faults in HVAC systems in commercial buildings. Here, the OSR problem is of high relevance, because it is very unrealistic that all possible faults are already known beforehand. This paper proposes an ML pipeline for open set anomaly classification. To address the OSR problem, the pipeline makes use of three common neural network architecture models that successively build on each other, namely two autoencoders and one ANN classifier. The advantage of using common well-known models is the ease of implementation for solution developers. The validation is made against an HVAC dataset published by LBNL that contains 7 fault cases, where we assume 3 of them as unknown. Our validation results show that the proposed pipeline can significantly reduce the risk of false class predictions in the presence of unknown data.