Research and Development of Vibration Diagnostics Methods for Electric Motors Using Machine Learning

Electric motors are the principal power units of industry, and their failures cause cascading stoppages of process equipment. The vibration signal carries the most complete information about the condition of rotating assemblies, yet machine-learning diagnostics is constrained by the shortage of labelled fault data. This work proposes and verifies a methodology for generating a physically grounded synthetic dataset of vibration signals: analytical models of imbalance, misalignment, rolling-bearing defects (BPFO, BPFI, BSF, FTF) and mechanical looseness produced 8000 one-second fragments of vibration acceleration covering eight classes of technical state, with variation of the rotational frequency (±5%), signal-to-noise ratio (5–30 dB) and defect amplitude. The synthetic signals were verified against reference Case Western Reserve University (CWRU) records: the outer- and inner-ring defect frequencies agree within 0.22%, and the envelope spectra are class-specific for these defects. Six classifiers were compared in three stages: Random Forest, XGBoost, a 1D-CNN (WDCNN), two 2D-CNNs on STFT and CWT representations, and a PatchTST transformer. Under the full set of variations, models reached a Macro F1 of 99.6–99.8%. On the real CWRU records, separate models were trained on the dataset based on real data from other operating loads. On the unseen loads, the Random Forest reached a Macro F1 of 95.5% and the 1D-CNN 93.2%, against 21–42% for the same models trained on synthetic data alone, and with one real fragment per class, the difference between the two training datasets was already 17–32 p.p. High accuracy on synthetic data therefore did not carry over to zero-shot diagnosis on the real bench.

Authors

Institutions

Publication Details

Journal
Automation
Published
2026-11-16
DOI
https://doi.org/10.3390/automation7060190
Primary Topic
Machine Fault Diagnosis Techniques
Type
article
Field-Weighted Citation Impact
0.00
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
article

Research and Development of Vibration Diagnostics Methods for Electric Motors Using Machine Learning

Yuriy K. Kozhubaev, Dmitry Kazanin, Georgy Molodtsov, Roman Dorovskikh et al.
2 links
Automation
Machine Fault Diagnosis Techniques
article

Research and Development of Vibration Diagnostics Methods for Electric Motors Using Machine Learning

Yuriy K. Kozhubaev, Dmitry Kazanin, Georgy Molodtsov, Roman Dorovskikh, Pavel Dolgiy
article en

Abstract

Electric motors are the principal power units of industry, and their failures cause cascading stoppages of process equipment. The vibration signal carries the most complete information about the condition of rotating assemblies, yet machine-learning diagnostics is constrained by the shortage of labelled fault data. This work proposes and verifies a methodology for generating a physically grounded synthetic dataset of vibration signals: analytical models of imbalance, misalignment, rolling-bearing defects (BPFO, BPFI, BSF, FTF) and mechanical looseness produced 8000 one-second fragments of vibration acceleration covering eight classes of technical state, with variation of the rotational frequency (±5%), signal-to-noise ratio (5–30 dB) and defect amplitude. The synthetic signals were verified against reference Case Western Reserve University (CWRU) records: the outer- and inner-ring defect frequencies agree within 0.22%, and the envelope spectra are class-specific for these defects. Six classifiers were compared in three stages: Random Forest, XGBoost, a 1D-CNN (WDCNN), two 2D-CNNs on STFT and CWT representations, and a PatchTST transformer. Under the full set of variations, models reached a Macro F1 of 99.6–99.8%. On the real CWRU records, separate models were trained on the dataset based on real data from other operating loads. On the unseen loads, the Random Forest reached a Macro F1 of 95.5% and the 1D-CNN 93.2%, against 21–42% for the same models trained on synthetic data alone, and with one real fragment per class, the difference between the two training datasets was already 17–32 p.p. High accuracy on synthetic data therefore did not carry over to zero-shot diagnosis on the real bench.

AutomationVol. 7(6)
Peter the Great St. Petersburg Polytechnic University (RU), Saint Petersburg Mining University (RU)
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