AutoSS: A Deep Learning-Based Soft Sensor for Handling Time-Series Input Data

Nicolò Bargellesi, Alessandro Beghi, Mirco Rampazzo, Gian Antonio Susto

Abstract: Soft Sensors are data-driven technologies that allow to have estimations of quantities that are impossible or costly to be measured. Unfortunately, the design of effective soft sensors is heavily impacted by time-consuming feature engineering steps that may lead to sub-optimal information, especially when dealing with time-series input data. While domain knowledge may come into help when handling feature extraction in soft sensing applications, the feature extraction typically limits the adoption of such technologies: in this work, we propose AutoSS, a Deep-Learning based approach that allows to overcome such issue. By exploiting autoencoders, dilated convolutions and an ad-hoc defined architecture, AutoSS allows to develop effective soft sensing modules even with time-series input data. The effectiveness of AutoSS is demonstrated on a real-world case study related to Internet of Things equipment.