The earliest possible detection of pitting damage in gearboxes is a central objective of vibration-based condition monitoring. Machine learning enables the automated analysis of vibration signals, but reliable detection of very early pitting damage requires a detailed understanding of which frequency ranges and frequency resolutions contain damage-relevant information. This work applies machine learning as a data-driven analysis tool to systematically quantify the relevance of frequency-based vibration features for pitting damage detection and pitting size classification. The investigations are based on experiments with three identical single-stage planetary gearboxes and four defined pitting sizes ranging from 0.5 % to 4 %. The measured time signals are transformed into the frequency domain using the Fast Fourier Transform, and the amplitudes of individual frequency bins are used as features. Since the bin width depends on the selected segment length, the influence of frequency resolution on the identification of damage-relevant features is also analyzed. A tree-based gradient boosting algorithm is used for classification, and the importance of individual frequency features is quantified by permutation analysis. The evaluation follows a two-stage approach. First, healthy and damaged states are compared to identify generally relevant frequency features. Second, the healthy state is contrasted separately with each pitting size to determine when specific features become relevant and how their importance changes with increasing damage size. In addition, feature consistency across operating conditions, sensor positions, and the three identical gearboxes is examined. The results support the targeted selection of frequency-based features for subsequent machine-learning-based damage detection and damage size classification and provide guidance for sensor placement, frequency resolution, and measurement system design in application-oriented condition monitoring.
%0 Conference Paper
%1 binanzer2026frequencydomain
%A Binanzer, Lisa
%A Dazer, Martin
%B PHM Society European Conference
%D 2026
%I PHM Society
%K
%N 1
%P 1-12
%R https://doi.org/10.36001/phme.2026.v9i1.5021
%T Frequency-Domain Feature Analysis for Early Gear Damage Detec-tion in Planetary Gearboxes
%V 9
%X The earliest possible detection of pitting damage in gearboxes is a central objective of vibration-based condition monitoring. Machine learning enables the automated analysis of vibration signals, but reliable detection of very early pitting damage requires a detailed understanding of which frequency ranges and frequency resolutions contain damage-relevant information. This work applies machine learning as a data-driven analysis tool to systematically quantify the relevance of frequency-based vibration features for pitting damage detection and pitting size classification. The investigations are based on experiments with three identical single-stage planetary gearboxes and four defined pitting sizes ranging from 0.5 % to 4 %. The measured time signals are transformed into the frequency domain using the Fast Fourier Transform, and the amplitudes of individual frequency bins are used as features. Since the bin width depends on the selected segment length, the influence of frequency resolution on the identification of damage-relevant features is also analyzed. A tree-based gradient boosting algorithm is used for classification, and the importance of individual frequency features is quantified by permutation analysis. The evaluation follows a two-stage approach. First, healthy and damaged states are compared to identify generally relevant frequency features. Second, the healthy state is contrasted separately with each pitting size to determine when specific features become relevant and how their importance changes with increasing damage size. In addition, feature consistency across operating conditions, sensor positions, and the three identical gearboxes is examined. The results support the targeted selection of frequency-based features for subsequent machine-learning-based damage detection and damage size classification and provide guidance for sensor placement, frequency resolution, and measurement system design in application-oriented condition monitoring.
%@ 978-1-936263-44-8
@inproceedings{binanzer2026frequencydomain,
abstract = {The earliest possible detection of pitting damage in gearboxes is a central objective of vibration-based condition monitoring. Machine learning enables the automated analysis of vibration signals, but reliable detection of very early pitting damage requires a detailed understanding of which frequency ranges and frequency resolutions contain damage-relevant information. This work applies machine learning as a data-driven analysis tool to systematically quantify the relevance of frequency-based vibration features for pitting damage detection and pitting size classification. The investigations are based on experiments with three identical single-stage planetary gearboxes and four defined pitting sizes ranging from 0.5 % to 4 %. The measured time signals are transformed into the frequency domain using the Fast Fourier Transform, and the amplitudes of individual frequency bins are used as features. Since the bin width depends on the selected segment length, the influence of frequency resolution on the identification of damage-relevant features is also analyzed. A tree-based gradient boosting algorithm is used for classification, and the importance of individual frequency features is quantified by permutation analysis. The evaluation follows a two-stage approach. First, healthy and damaged states are compared to identify generally relevant frequency features. Second, the healthy state is contrasted separately with each pitting size to determine when specific features become relevant and how their importance changes with increasing damage size. In addition, feature consistency across operating conditions, sensor positions, and the three identical gearboxes is examined. The results support the targeted selection of frequency-based features for subsequent machine-learning-based damage detection and damage size classification and provide guidance for sensor placement, frequency resolution, and measurement system design in application-oriented condition monitoring.},
added-at = {2026-07-14T18:05:14.000+0200},
author = {Binanzer, Lisa and Dazer, Martin},
biburl = {https://puma.ub.uni-stuttgart.de/bibtex/2cbf77a961f6c9061a6ae64f39f3bb4b5/lisabinanzer},
booktitle = {PHM Society European Conference},
doi = {https://doi.org/10.36001/phme.2026.v9i1.5021},
eventdate = {08.-10.07.2026},
eventtitle = {9th European Conference of the PHM Society 2026},
interhash = {f011d71b8d59fb5e047a279ddd1a503a},
intrahash = {cbf77a961f6c9061a6ae64f39f3bb4b5},
isbn = {978-1-936263-44-8},
keywords = {},
month = {7},
note = {(peer-review)},
number = 1,
pages = {1-12},
publisher = {PHM Society},
timestamp = {2026-07-14T18:05:14.000+0200},
title = {Frequency-Domain Feature Analysis for Early Gear Damage Detec-tion in Planetary Gearboxes},
venue = {Oslo, Norway},
volume = 9,
year = 2026
}
