An LSTM-Based Method for Detection and Classification of Sensor Anomalies

Most existing machine learning (ML) based solutions for anomaly detection in sensory data rely on carefully hand-crafted features. This approach has a fundamental limitation since it is often application-specific and requires considerable human effort from domain experts. Deep learning models have been demonstrated to have the ability to abstract relevant high-level features from raw data. Long short-term memory (LSTM) recurrent neural networks have proven effective in complex time-series prediction problems. In this paper, we propose an LSTM-based method for anomaly detection in sensory data. We systematically investigate its effectiveness on raw time-series of real medical sensors measurements and show that it achieves the same level of performance as traditional ML models operating on carefully designed feature vectors. The proposed method achieved micro, macro, and weighted precision, recall, and F1-score of over 0.99.