Enhancing Anomaly Detection in Multivariate Time Series Data Using Hybrid Deep Learning Architectures with Attention Mechanisms and Feature Engineering

Abstract

Multivariate time series anomaly detection is a challenging task that is encountered in most industries like industrial processes, computer security, and health care. Traditional approaches cannot capture within-variable rich temporal relationships as well as cross-stream correlations in high-dimensional streams. This paper presented a new hybrid deep learning model for detecting anomalies in multivariate time series using Long Short-Term Memory (LSTM) networks and Convolutional Neural Networks (CNNs) augmented with applications of attention. Within the hybrid model, the LSTM learns specifically temporal dependencies among the multivariate time series, and the CNN learn spatial features embed within the time series. Use of attention mechanism picks out significant features and time steps, which improve improved anomaly detection. Moreover, we proposed a feature engineering technique to extract useful features from raw time series data, which were then inputted into the deep learning model. We empirically demonstrated the proposed method with several benchmark datasets compared to state-of-the-art anomaly detection techniques. The results showed that our proposed model, the integration of attention and feature engineering improved the performance of anomaly detection as a whole (precision, recall, and F1-score).