基于元学习的自动化异常检测

Automated Anomaly Detection Based on Meta-Learning

摘要: 在实际的工业生产环境中，常常需要监控机器相关指标的运行状况，对于一个多变量的无监督异常检测任务，由于缺乏带标签的数据，并且同一个检测算法在不同数据集上的性能表现不同，模型设计依赖于人工调整，所以如何高效选择一个异常检测模型并完成超参数调整成为了一个亟待解决的问题。在这篇文章中，我们建立了一个异常检测模型自动选择机制，称为AutoAD (Auto Anomaly Detector)。AutoAD利用了历史数据上异常检测模型的表现和数据集本身的特征，基于元学习的想法，通过深度神经网络自动选择一个有效的异常检测模型并调优，用于新的数据集的异常检测。实验结果表明了在开源数据集上AutoAD在异常检测模型自动选择方面具有有效性。

Abstract: In real industrial production environments, it is often necessary to monitor the operating condi-tions of machine-related indicators. For a multivariate unsupervised anomaly detection task, due to the lack of labeled data and the different performance of the same detection algorithm on different datasets, the design of the model relies on manual tuning, so how to efficiently select an anomaly detection model and complete hyper-parameter tuning has become a pressing problem. In this pa-per, we develop an automatic anomaly detection model selection mechanism called AutoAD (Auto Anomaly Detector). AutoAD utilizes the performance of anomaly detection models on historical data and the characteristics of the dataset itself to automatically select an effective anomaly detection model based on the idea of meta-learning through deep neural networks, and tunes it for anomaly detection on new datasets. The experimental results show that on the open source dataset AutoAD is effective in automatic anomaly detection model selection.