HUMS2023 Data Challenge Result Submission
Team Name: NRC-AERO-SMPL-01
Team Members: Emma Seabrook, Catherine Cheung
Institutions: National Research Council Canada
Publishable: yes
1. Summary of Findings
Data from DST Group’s Planet Gear Fault Propagation test [1] was analysed using anomaly
detection and regression machine learning (ML) algorithms. Given data from four vibration channels
positioned at four different locations on the gearbox [1], both the crack initiation time as well as the
propagation trend for the crack on the planet gear rim were determined. In order to increase the
dimensionality of the dataset and provide the machine learning model with more information, gear
condition indicators (CI) based on vibration characteristics in the frequency domain were calculated [2],
detailed in Section 6.
To detect the point of crack initiation, an ensemble of anomaly detection algorithms were used to
find outliers within the data. The algorithms selected files that correspond to anomalies in the gear
condition indicators. Since the data point for file 150 aligns with a peak in the condition indicators ‘FM0’
and ‘FM4’, we believe that the crack initiated at this point, summarized in Table 1. The trend for crack
propagation was generated using a semi-supervised regression ML model. The model was able to produce
a regression from a basic training trend and showed that there is a clear trend in crack growth from file
170 which accelerates around file 327. This is consistent with trends in the ‘Crest Factor’, ‘FM0’, and ‘FM4’
condition indicators.
Table 1: Summary of Analysis Results
Consistent detection on at least one signal channel; i.e. the fault
indicators remain consistently above the threshold.
Confirmed detection on at least two signal channels; i.e. the fault
indicators remain consistently above the threshold.
Clear multi-channel indication of the characteristic fault features;
i.e. faulty planet gear meshing with both the ring and sun gears.
Confirmed trend of fault progression; i.e. a consistent increasing
trend started from which file number/name.
Confirmed trend of accelerated fault progression; i.e. a consistent
exponential increasing trend started from which file number/name
