Utilizing autoencoders to improve transfer learning when sensor data is sparse

Accurate deep predictive models of wastewater processing plants are important to ensure operational parameters are safe and sustainable. Training such predictive models requires large volumes of data that is hard to find in the wastewater domain. Transfer learning addresses the problem, by training predictive models using data from an adopted domain, and fine-tuning it on the target domain. However, due to the significant distributional shift between the commonly adopted source domains for transfer learning and the target domain of wastewater processes, transfer learning rarely performs at an acceptable level. This paper proposes a method to generate large volumes of training data with a similar distribution to a sample taken from the target domain, to boost transfer learning performance for the task, referred to as AETL. It leverages an autoencoder that systematically augments the target domain data such that synthetic samples it generates closely follow the target domain distribution. The results on the real-world data set establish the efficacy of the proposed framework.