Explanatory LSTM-AE-based Anomaly Detection for Time Series Data in Marine Transportation

Abstract

Ensuring the normal operation of mechanical equipment is crucial in marine transportation, as data anomalies in these systems can lead to serious safety incidents, environmental pollution, and economic losses. To improve the accuracy and efficiency of anomaly detection in ship equipment data, a long-short-term memory auto-encoder model (LSTM-AE) has been developed, tailored for time-series anomaly detection. The normal behavior patterns of key metrics, such as temperature, pressure, and rotational speed, are learned and captured by this model, enabling abnormal states in these metrics to be accurately identified. The approach is based on the encoder in the long-short-term memory (LSTM) network, where input time-series data is converted into a lower-dimensional, implicit representation, and an attempt is made to reconstruct the original input data via a decoder. Trained exclusively on anomaly-free data, the model ensures a low reconstruction error on normal data. However, when input data that significantly deviates from the training set is encountered, a high reconstruction error is produced, thereby allowing potential anomalies to be flagged. To enhance the interpretability of the results, explainable artificial intelligence (XAI) techniques are incorporated, specifically shapley additive explanations (SHAP) and local interpretable model-agnostic explanations (LIME), to identify which features have the most impact on detected anomalies. The LSTM-AE model shows superior performance compared to other data generation models such as GAN and diffusion models, which have accuracy issues and require high computational cost. In addition, the integration of XAI methods has advantages in the interpretation of the results, solving the problem that these existing methods often lack transparency.