CNN-LSTM transfer learning for sim-to-real broken rotor bar fault diagnosis in induction motors

Broken rotor bar (BRB) faults in three-phase induction motors remain challenging to diagnose reliably, particularly when models trained solely on simulated data are applied to real-world measurements. Conventional methods, such as motor current signature analysis (MCSA) and classical machine learning (ML), are sensitive to load variations, rely on manual feature extraction, and exhibit poor performance under simulation-to-reality domain shift. This paper proposes a convolutional neural network-long short-term memory (CNN-LSTM) transfer learning framework for BRB fault diagnosis. A dq0-based mathematical model is used to generate a comprehensive synthetic dataset covering healthy and faulty conditions across multiple fault severities and load levels. The simulated currents are subjected to empirical noise augmentation to better approximate experimental measurements. Classical ML classi- fiers trained on time-domain features derived from the Park vector modulus (PVM) achieve approximately 99% accuracy on synthetic data but deteriorate to nearly 55% when tested on experimental data, revealing a significant domain gap. The proposed CNN-LSTM model is pre-trained on the synthetic data to learn the spatial and temporal fault signatures and subsequently fine-tuned using limited experimental data. This two-stage approach achieves accuracy and an F1-score exceeding 99% when tested on the experimental dataset from IEEE DataPort. The findings demonstrate that synthetic data pretraining combined with transfer learning can substantially narrow the simulation-to-reality gap under the tested conditions, while reducing the reliance on large volumes of labelled experimental data.

Authors

Institutions

Publication Details

Journal
Proceedings of the Estonian Academy of Sciences
Published
2026-11-16
DOI
https://doi.org/10.3176/proc.2026.4.10
Primary Topic
Machine Fault Diagnosis Techniques
Type
article
Field-Weighted Citation Impact
0.00

Funders

Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
article

CNN-LSTM transfer learning for sim-to-real broken rotor bar fault diagnosis in induction motors

Arun Gnawali, Bishal Silwal, Mehroz Fatima, Aashirbad Bhandari et al.
3 links
Proceedings of the Estonian Academy of Sciences
Machine Fault Diagnosis Techniques
article

CNN-LSTM transfer learning for sim-to-real broken rotor bar fault diagnosis in induction motors

Arun Gnawali, Bishal Silwal, Mehroz Fatima, Aashirbad Bhandari, Ankeet Gadtaula, Aalok Aryal
article en

Abstract

Broken rotor bar (BRB) faults in three-phase induction motors remain challenging to diagnose reliably, particularly when models trained solely on simulated data are applied to real-world measurements. Conventional methods, such as motor current signature analysis (MCSA) and classical machine learning (ML), are sensitive to load variations, rely on manual feature extraction, and exhibit poor performance under simulation-to-reality domain shift. This paper proposes a convolutional neural network-long short-term memory (CNN-LSTM) transfer learning framework for BRB fault diagnosis. A dq0-based mathematical model is used to generate a comprehensive synthetic dataset covering healthy and faulty conditions across multiple fault severities and load levels. The simulated currents are subjected to empirical noise augmentation to better approximate experimental measurements. Classical ML classi- fiers trained on time-domain features derived from the Park vector modulus (PVM) achieve approximately 99% accuracy on synthetic data but deteriorate to nearly 55% when tested on experimental data, revealing a significant domain gap. The proposed CNN-LSTM model is pre-trained on the synthetic data to learn the spatial and temporal fault signatures and subsequently fine-tuned using limited experimental data. This two-stage approach achieves accuracy and an F1-score exceeding 99% when tested on the experimental dataset from IEEE DataPort. The findings demonstrate that synthetic data pretraining combined with transfer learning can substantially narrow the simulation-to-reality gap under the tested conditions, while reducing the reliance on large volumes of labelled experimental data.

Proceedings of the Estonian Academy of SciencesVol. 75(4)
Tallinn University of Technology (EE), Tribhuvan University (NP)
Eesti Teaduste Akadeemia
Industry, innovation and infrastructure
Machine Fault Diagnosis Techniques
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon