Addressing data scarcity in industrial reliability assessment with Physically Informed Echo State Networks

This paper introduces a method for augmenting sensor data using Physically Informed Echo State Networks (ESNs), which facilitates system identification in scenarios with limited sensor data. The approach integrates domain-specific physical knowledge into the learning process of ESNs to generate surrogate time-amplitude signals from the Power Spectral Density (PSD) of the data and a predefined list of system excitation frequencies. This integration proves particularly beneficial during the initial design phases of condition monitoring systems, where empirical data is often sparse. We demonstrate the effectiveness of this method through experiments on a 30 kW jet fan in a road tunnel ventilation system. Results indicate significant improvements in the operational capabilities of condition monitoring systems for newly developed equipment. This method is versatile and applicable across various industrial contexts with insufficient historical operational data.