Extract, Compress and Encode: LitNet an Efficient Autoencoder for Noisy Time-Series Data

The Industrial Internet of Things has emerged to enhance massive sensory data collection. Characterized by their large volume and hard interpretability, these data require highly skilled human expertise to compress and extract meaningful information from the raw data using computationally intensive signal processing methods. Deep learning, and autoencoders, in particular, have shown promising results in computer vision and natural language processing. In this paper, different state-of-the-art deep learning blocks are investigated and a lightweight Inception Network, called LitNet, is presented. The experiments are performed with real-world vibration data collected from different CNC machines during production. The models are evaluated based on their compression and feature extraction capabilities. The results show that LitNet outperforms all models and provides a good trade-off between compression and feature extraction. The latent vector analysis shows promising results and that the LitNet encoder can be used as a pre-trained feature extractor for vibration data.