• adversarial networks
• business use case
• deep learning
• discriminative training
• LSTM
• regression
• signal virtualization
• unsupervised domain adaptation

Virtualization of a signal is a common task in industrial sectors, where machinery needs many sensors to monitor its behaviour and health. The ability to virtualize a sensor with a machine learning approach would represent a significant breakthrough and would overcome the shortcomings of the limited complexity one can afford to simulate with physics-based models. It is also common in the industrial sector to need unsupervised domain adaptation methods to correct the domain shift that occurs when the virtual sensor is trained on a machine in a prototype state and used to make inferences in a machine in a fleet state. Despite this, there are few references in the literature when the domain shift is involved in a regression task. This thesis intends to fill this gap, proposing an approach inspired by adversarial architectures consisting of recurrent networks and an adversarial domain classifier. Experiments using the proposed approach show how the virtualized sensor improves reliability when it is based on features that combine discriminativeness and domain invariance.