Gribbestad Magnus, Hassan Muhammad Umair, A Hameed Ibrahim, Sundli Kelvin
Department of ICT and Natural Sciences, Norwegian University of Science and Technology (NTNU), Larsgårdsvegen 2, 6009 Ålesund, Norway.
Cognite AS, Oksenøyveien 10, 1366 Lysaker, Norway.
Entropy (Basel). 2021 Jan 8;23(1):83. doi: 10.3390/e23010083.
Anomaly detection refers to detecting data points, events, or behaviour that do not comply with expected or normal behaviour. For example, a typical problem related to anomaly detection on an industrial level is having little labelled data and a few run-to-failure examples, making it challenging to develop reliable and accurate prognostics and health management systems for fault detection and identification. Certain machine learning approaches for anomaly detection require normal data to train, which reduces the need for historical data with fault labels, where the main task is to differentiate between normal and anomalous behaviour. Several reconstruction-based deep learning approaches are explored in this work and compared towards detecting anomalies in air compressors. Anomalies in such systems are not point-anomalies, but instead, an increasing deviation from the normal condition as the system components start to degrade. In this paper, a descriptive range of the deviation based on the reconstruction-based techniques is proposed. Most anomaly detection approaches are considered black box models, predicting whether an event should be considered an anomaly or not. This paper proposes a method for increasing the transparency and explainability of reconstruction-based anomaly detection to indicate which parts of a system contribute to the deviation from expected behaviour. The results show that the proposed methods detect abnormal behaviour in air compressors accurately and reliably and indicate why it deviates. The proposed approach is capable of detecting faults without the need for historical examples of similar faults. The proposed method for explainable anomaly detection is crucial to any prognostics and health management (PHM) system due to its purpose of detecting deviations and identifying causes.
异常检测是指检测不符合预期或正常行为的数据点、事件或行为。例如，在工业层面上，与异常检测相关的一个典型问题是标记数据很少，且只有少数几个直至故障发生的示例，这使得开发用于故障检测和识别的可靠且准确的预测与健康管理系统具有挑战性。某些用于异常检测的机器学习方法需要正常数据来进行训练，这减少了对带有故障标签的历史数据的需求，其主要任务是区分正常行为和异常行为。在这项工作中，研究了几种基于重构的深度学习方法，并将它们用于检测空气压缩机中的异常情况进行比较。此类系统中的异常并非点异常，而是随着系统组件开始退化，与正常状态的偏差不断增大。本文基于基于重构的技术提出了一种描述偏差范围的方法。大多数异常检测方法被视为黑箱模型，用于预测一个事件是否应被视为异常。本文提出了一种提高基于重构的异常检测的透明度和可解释性的方法，以指明系统的哪些部分导致了与预期行为的偏差。结果表明，所提出的方法能够准确可靠地检测空气压缩机中的异常行为，并指明其偏差原因。所提出的方法能够在无需类似故障历史示例的情况下检测故障。所提出的可解释异常检测方法对于任何预测与健康管理（PHM）系统都至关重要，因为其目的是检测偏差并识别原因。
