Unsupervised LSTM-Autoencoder Approach for Early Detection of Sensor Failures in High-Dimensional Monitoring Environments

The contemporary industrial systems are growing to be dependent on the high-dimensional sensor networks, thus exposing themselves to unknown faults that threaten the reliability and safety of operations. Conventional threshold based and supervised learning models fail in these settings particularly when labeled fault information is scarce or non-existent. As a remedy to these shortcomings, this paper introduces an unsupervised LSTM-Autoencoder model that can be used to detect abnormalities and learn to identify sensor failures at an early stage without using annotated data. The model is fit on regular-operation sequences only sampled off of the Kaggle Sensor Fault Detection Dataset, where reconstruction error is used as the anomaly score. The results of the experiment prove the constant decrease of reconstruction loss, high detection rate of 95.3%, and significant reduction of false-alarm rates, as well as good early fault detection. The model is useful in identifying slight pattern deviations on a temporal pattern way before major sensor failures take place. Altogether, the suggested LSTM-Autoencoder is a scalable, robust, and label-free tool to facilitate sensor health in a complex and high-dimensional industrial setting.