Attention based Anomaly Detection using hybrid models in Multivariate Time Series Data

Anomaly detection in time series data is vital for identifying deviations from normal behavior, enabling proactive measures in domains such as finance, healthcare, and IoT. Conventional techniques struggle with handling temporal dependencies and complex, multivariate data, resulting in sub-optimal accuracy in detecting rare or faint anomalies. Introduction of a combination model of deep learning based on CNN, LSTM networks, Temporal Convolutional Networks (TCN), and an attention mechanism aims to mitigate these constraints. The CNN detects local anomalies at certain timestamps. LSTM remembers the long-term temporal patterns, and TCNs further improves the sequence modeling with the help of causal convolutions to detect complex temporal patterns. The attention mechanism further improves the process by taking account of only the most important data points for further improvement in accuracy of anomaly detection. The proposed model is verified on univariate and multivariate time series datasets with significantly less validation loss compared to existing methods. With the integration of multiple models, the given framework offers a robust and scalable method for anomaly detection across multiple real-world scenarios with complex time series data.