Multi-scale wavelet Kolmogorov–Arnold networks with adaptive feature refinement for cross-speed spur gearbox fault diagnosis

Accurate spur gearbox fault diagnosis across different rotational speeds remains challenging because gear-mesh and sideband components are shifted and the time–frequency distribution of fault-related vibration signatures is modified by rotational-speed variations, producing distribution shifts between training and deployment conditions. A Multi-Scale Wavelet Kolmogorov–Arnold Network (MWKAN) is proposed for direct cross-speed classification of controlled gear-health states. Complementary discrete and continuous wavelet representations are integrated and processed through multi-kernel convolution, progressive hierarchical feature fusion, KAN-based channel refinement, and a B-spline KAN classifier. Physics-guided auxiliary supervision based on vibration-energy and impulsiveness statistics is further incorporated during training so that signal-relevant information is retained within the shared representation. The framework is evaluated on the XJTU spur gearbox dataset under within-speed and bidirectional cross-speed transfer conditions between 15 and 20 Hz, without target-speed data being used for model development, and further validated on an independent RRL gear-crack dataset. Mean within-speed accuracies of 99.56 ± 0.21% and 99.77 ± 0.18% are obtained at 15 and 20 Hz, respectively, while direct cross-speed accuracies of 96.16 ± 0.48% and 98.46 ± 0.31% are achieved for transfer from 15 to 20 Hz and from 20 to 15 Hz, respectively. On the independent RRL dataset, cross-speed accuracies of 92.3 ± 1.28% and 94.2 ± 1.01% were obtained for the corresponding transfer directions. Complementary analysis of the learned representations and model behavior is provided through stage-wise feature separation analysis, gear-mesh-frequency-domain analysis, Grad-CAM localization, component-wise ablation, and controlled-noise evaluation. The independent RRL evaluation further supports methodological consistency across the investigated datasets, while the observed directional transfer asymmetry is associated with operating-speed-dependent signal characteristics.

Authors

Institutions

Publication Details

Journal
Mechanical Systems and Signal Processing
Published
2026-11-18
DOI
https://doi.org/10.1016/j.ymssp.2026.115187
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

Multi-scale wavelet Kolmogorov–Arnold networks with adaptive feature refinement for cross-speed spur gearbox fault diagnosis

Rajeev Kumar, Ming Jian Zuo, Zhigang Tian
4 links
Mechanical Systems and Signal Processing
Machine Fault Diagnosis Techniques
article

Multi-scale wavelet Kolmogorov–Arnold networks with adaptive feature refinement for cross-speed spur gearbox fault diagnosis

Rajeev Kumar, Ming Jian Zuo, Zhigang Tian
article en

Abstract

Accurate spur gearbox fault diagnosis across different rotational speeds remains challenging because gear-mesh and sideband components are shifted and the time–frequency distribution of fault-related vibration signatures is modified by rotational-speed variations, producing distribution shifts between training and deployment conditions. A Multi-Scale Wavelet Kolmogorov–Arnold Network (MWKAN) is proposed for direct cross-speed classification of controlled gear-health states. Complementary discrete and continuous wavelet representations are integrated and processed through multi-kernel convolution, progressive hierarchical feature fusion, KAN-based channel refinement, and a B-spline KAN classifier. Physics-guided auxiliary supervision based on vibration-energy and impulsiveness statistics is further incorporated during training so that signal-relevant information is retained within the shared representation. The framework is evaluated on the XJTU spur gearbox dataset under within-speed and bidirectional cross-speed transfer conditions between 15 and 20 Hz, without target-speed data being used for model development, and further validated on an independent RRL gear-crack dataset. Mean within-speed accuracies of 99.56 ± 0.21% and 99.77 ± 0.18% are obtained at 15 and 20 Hz, respectively, while direct cross-speed accuracies of 96.16 ± 0.48% and 98.46 ± 0.31% are achieved for transfer from 15 to 20 Hz and from 20 to 15 Hz, respectively. On the independent RRL dataset, cross-speed accuracies of 92.3 ± 1.28% and 94.2 ± 1.01% were obtained for the corresponding transfer directions. Complementary analysis of the learned representations and model behavior is provided through stage-wise feature separation analysis, gear-mesh-frequency-domain analysis, Grad-CAM localization, component-wise ablation, and controlled-noise evaluation. The independent RRL evaluation further supports methodological consistency across the investigated datasets, while the observed directional transfer asymmetry is associated with operating-speed-dependent signal characteristics.

Mechanical Systems and Signal ProcessingVol. 263
University of Alberta (CA), Fuyao University of Science and Technology (CN)
Natural Sciences and Engineering Research Council of Canada
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