Anomaly detection is critical in the Internet of Things (IoT) systems due to its wide applications for building smart cities, such as quality control in manufacturing, intrusion detection in system security, fault detection in system monitoring. Many existing schemes are problem specific and supervised approaches, which require domain knowledge and tremendous data labeling efforts. In this paper, we investigate unsupervised anomaly detection on multidimensional time series data in IoT systems, and develops a GRU-based Gaussian Mixture VAE scheme, called GGM-VAE. In particular, we employ Gated Recurrent Unit (GRU) cells to discover the correlations among time series data, and use Gaussian Mixture priors in the latent space to characterize the multimodal data. Several previous works assume simple distributions for Gaussian Mixture priors, resulting in insufficient ability to fully capture the data patterns. To overcome this issue, we design a model selection mechanism during the training process under the guidance of Bayesian Inference Criterion (BIC) to find the model which can well estimate the distribution in the Gaussian Mixture latent space. We conduct extensive simulations on four datasets and observe that our proposed scheme outperforms the state-of-the-art anomaly detection schemes and achieves up to 47.88% improvement in F1 scores on average.

中文翻译：

---

智慧城市物联网系统中的无监督异常检测

异常检测在物联网（IoT）系统中至关重要，这是因为其在构建智能城市中的广泛应用，例如制造中的质量控制，系统安全中的入侵检测，系统监视中的故障检测。许多现有的方案都是针对特定问题且受监督的方法，需要领域知识和巨大的数据标记工作。在本文中，我们研究了物联网系统中多维时间序列数据的无监督异常检测，并开发了一种基于GRU的高斯混合VAE方案，称为GGM-VAE。特别是，我们使用门控循环单元（GRU）单元来发现时间序列数据之间的相关性，并在潜在空间中使用高斯混合先验来表征多峰数据。先前的一些工作假设高斯混合先验的简单分布，导致无法完全捕获数据模式。为了克服这个问题，我们在贝叶斯推理准则（BIC）的指导下设计了训练过程中的模型选择机制，以找到可以很好地估计高斯混合潜空间分布的模型。我们在四个数据集上进行了广泛的仿真，发现我们提出的方案优于最新的异常检测方案，平均F1得分提高了47.88％。