Abstract
Anomaly detection is a critical task in many domains, including the Internet of Things (IoT), where large volumes of sensor data are generated from various devices. Traditional methods based on domain knowledge often fail to capture the complex correlations among the variables and time series. Therefore, there is a need for robust and efficient anomaly detection methods that can learn from data. Recently, deep learning models have shown promising results in detecting anomalies in multivariate time-series data. In particular, transformer-based models and convolutional networks have gained popularity due to their ability to capture long-term dependencies and extract meaningful features from high-dimensional data. In this study, we propose a transformer-based framework for anomaly detection in IoT systems. Our model employs dynamic graph attention to capture the complex correlations among variables and time series, and a transformer architecture with parallel processing ability to extract multidimensional features for downstream prediction tasks. We evaluate our model on several publicly available datasets and demonstrate its superior performance compared to state-of-the-art methods. In summary, our proposed framework is a promising approach for anomaly detection in sensor data for industry predictive maintenance, which can potentially reduce downtime, increase efficiency, and save costs.