Learning virtual sensors for estimating the scheduling signal of parameter-varying systems

We propose a novel data-driven virtual sensor architecture to reconstruct an unmeasurable scheduling signal of a parameter-varying system from input/output measurements. The key idea is to train and feed an Artificial Neural Network (ANN) with input/output measurements and with data generated by processing such measurements through a bank of linear observers. Special attention is paid to the design of both the ANN and the feature extraction mechanism to keep the architecture as lightweight as possible, so that the resulting virtual sensor can be easily implemented in embedded hardware platforms. As a special case, the proposed virtual sensor can be used for hidden mode reconstruction of switched linear systems. Applications of the proposed approach are geared towards fault detection and isolation, predictive maintenance, and gain-scheduling control.