Implementation of multi-task learning neural network architectures for robust industrial optical sensing

The simultaneous determination of multiple physical or chemical parameters can be very advantageous in many sensor applications. In some cases, it is unavoidable because the parameters of interest display cross sensitivities or depend on multiple quantities varying simultaneously. One notable example is the determination of oxygen partial pressure via luminescence quenching. The measuring principle is based on the measurement of the luminescence of a specific molecule, whose intensity and decay time are reduced due to collisions with oxygen molecules. Since both the luminescence and the quenching phenomena are strongly temperature-dependent, this type of sensor needs continuous monitoring of the temperature. This is typically achieved by adding temperature sensors and employing a multi-parametric model (Stern{Volmer equation), whose parameters are all temperature-dependent. As a result, the incorrect measurement of the temperature of the indicator is a major source of error. In this work a new approach based on multi-task learning (MTL) artificial neural networks (ANN) was successfully implemented to achieve robust sensing for industrial applications. These were integrated in a sensor that not only does not need the separate detection of temperature but even exploits the intrinsic cross-interferences of the sensing principle to predict simultaneously oxygen partial pressure and temperature. A detailed analysis of the robustness of the method was performed to demonstrate its potential for industrial applications. This type of sensor could in the future signi ficantly simplify the design of the sensor and at the same time increase its performance