X-TSAD: An Explainable Framework for Robust Multivariate Time-Series Anomaly Detection using SHAP and Attention Mechanisms

Abstract

Anomaly detection plays a vital role in industrial systems to improve the quality of product, reducing engine downtime and maintain efficiency. As industries produce huge amount of timeseries data through various sensors. Traditional deep learning and machine learning models can accurately predict the anomalies, however they lack of interpretability, transparency and failed to show the importance of hyperparameters. This makes the domain experts not to trust the traditional models to adopt for industrial systems. This project combines deep learning LSTM model and attention mechanism to detect the anomalies and SHAP mechanism to know about the feature level contribution and critical timesteps that influence anomalies. additionally, the model evaluates the importance of hyperparameter tuning that improves the stability, generalizability and reproducibility. thus the proposed X-TSAD an explainable framework for multivariate time series data using attention and SHAP mechanism not only detect the anomalies, it explains about why the specific anomaly was detected as an anomaly and also evaluate the performance of model using standard metrics. Experimental result on real-world industrial dataset demonstrates about the detection of anomaly pattern by learning from temporal patterns and providing the important timesteps using attention mechanism, quantify the importance of hyperparameter tuning by evaluate the model using metrics such as precision, recall, F1 score and ROC-AUC. Also explain about the features, sensors that cause anomalies which helps to analyses the root cause and improve decision making in real time applications.