In the last years, data driven approaches have been frequently used to analyse real time data coming from the sensors installed on Oil & Gas plants to detect anomalous behaviours or diagnose potential faults of equipment. However, due to the large amount of installed sensors, it is often required to reduce the number of variables before building machine learning models. With the increased popularity of neural networks, traditional methods used for the reduction of the dimensionality of variables, such as Principal Component Analysis, have been replaced by Neural Autoencoders (NA), which provide the capability to better capture the nonlinear relationships among the variables, and therefore allow to use fewer variables for the analysis of the behaviour of the equipment.
In this paper, a Quantum Computing (QC) approach is described to implement a NA. As case study, this paper focuses on data coming from a separator within the Stabilization Unit of an Oil & Gas plant. We describe the research challenges that have been overcome to perform a QC modelling of the NA. In particular, input data have to be encoded in the phases of the qubit states representing the computational basis. Then, up to 4 classical input data are squeezed into the quantum state of a single qubit through an "encoder" quantum circuit. Compressed information can then be retrieved by suitably measuring this latter qubit. The paper shows that the reversed algorithm allows to reconstruct the original state with large fidelity, comparing well with the performance of a traditional NA run on a classical computer. The quantum autoencoder has been tested through classical simulations of quantum computing hardware, as well as run on an actual quantum computer made available through the IBM Quantum Experience. Due to the noisy character of state-of-art quantum processors, the statistical distribution of real data could be reconstructed only upon a certain degree of approximation. This first demonstration sets the basis for the concrete development of quantum autoencoding strategies and their practical use on real data sets for actual industrial applications.
