A physics-informed autoencoder for system health state assessment based on energy-oriented system performance

Abstract

Health Indicators (HIs) have been widely used for health state assessments. In many applications, HI with physical meaning is a preferred choice to assist system health management due to its inherent nature of objectively and accurately representing the system health state. However, in many cases, the true value of HI with physical meaning is difficult to obtain due to the difficulty in measuring them, which means, the HI is hidden from the user during system operation. It results in difficulty in training HI construction methods. In light of these challenges, we propose a physics-informed autoencoder for HI construction by fusing the physics-based model with deep learning (DL) approaches. In this framework, we redefine the conventional HI construction process with autoencoders into a new paradigm: mapping the sensor readings to a degradation-represented latent space by a DL model and reconstructing the sensor readings by a physics-based model. The latent variable, bridging the connection between the encoder and decoder, works as the HI and is meticulously designed with an energy-oriented perspective, thus ensuring its applicability across various systems. Furthermore, a novel training strategy is proposed for this framework to be well-trained. The superiority and effectiveness of the proposed framework are validated on the CALCE battery dataset and electromechanical actuator simulation data. In the two examples, the SOH of batteries and the energy efficiency of electromechanical actuators can both be estimated using the proposed method.