Published February 27, 2026 | Version v1
Journal Open

PHYSICS-INFORMED DOMAIN ADAPTATION: A DIGITAL TWIN APPROACH TO FAULT DIAGNOSIS IN DATA-SCARCE INDUSTRIAL ENVIRONMENTS.

Authors/Creators

Description

Although the viability of data-driven Predictive Maintenance (PdM) solutions is often thwarted by
the lack of available faulty data, leading to challenged training for deep learning models. This
paper presents a Physics-Informed Domain Adaptation (PIDA) approach, which combines a high
precision Digital Twin model for a direct current (DC) motor and a one-dimensional Convolutional
Neural Network (1D-CNN), for successful Sim-to-Real fault diagnosis in a data-scarce setting.
This Digital Twin representation is established using a system of coupled differential equations,
which model electromechanical phenomena, with simulated system parameters set to R= 2.0 Ω
and B = 0.001 Nm· s/rad for healthy cases. Fault conditions can be simulated by parameter
variations, such as R = 0.5 Ω (short circuit in stator) and B = 0.05 (bearing friction fault). The
concept of Domain Randomization is applied by adding a zero-mean Gaussian noise process 𝑁 (0,
0.1) that allows overcoming differences between simulated and actual signals coming from
physical sensors. Using the aforementioned hybrid approach, which was trained on 100%
simulated data and 50% vibration data from real cases, a perfect accuracy (100%) with precision
= 1.00 and a perfect recall measure (100%) was obtained in classifying the CWRU-bearing test
data set.

Files

PHYSICS-INFORMED DOMAIN ADAPTATION A DIGITAL TWIN APPROACH TO FAULT DIAGNOSIS IN DATA-SCARCE INDUSTRIAL ENVIRONMENTS..pdf

Files (465.9 kB)