Hunt-Pain, George
2022.
Application of machine learning to gearbox condition monitoring with rate of change of torque sensors.
PhD Thesis,
Cardiff University.
Item availability restricted. |
![]() |
PDF (PhD thesis)
- Accepted Post-Print Version
Restricted to Repository staff only until 22 May 2028 due to copyright restrictions. Download (66MB) |
![]() |
PDF (Cardiff University Electronic Publication Form)
- Supplemental Material
Restricted to Repository staff only Download (284kB) |
Abstract
Gearbox condition monitoring is an integral part of high-performance motorsport. In the world of Formula 1, gearboxes are designed to carry enormous loads, be tightly packaged and as lightweight as possible. As such the ability to reliably understand the condition of the gearbox, and conduct preventative maintenance, can mean the difference between winning and losing a championship. Traditional Condition Monitoring techniques in this field use accelerometers to measure the vibration of the system, however they have significant flaws in regards to rotating machinery - being prone to background noise, difficult to package and sensitive to their placement. This thesis investigates an alternative technique which uses a non-contact magneto-elastic sensor to monitor the Rate-of-Change of Torque (ROC), which provides a direct measurement of the driving interactions inside the gearbox and can be placed anywhere inside the torque path. The work in this thesis extends the traditional monitoring techniques to ROC sensors and implements machine learning algorithms to improve detection accuracy and robustness of bearing and gear damage recorded in a complex geared transmission. A new bearing defect dataset is recorded as part of this work and a comparison of vibration and ROC measurements is presented with respect to bearing damage. The techniques developed are then extended to a real-world defect and an existing gear defect dataset. The new dataset recorded seeded bearing defects on the outer raceway that replicated a real-world distributed defect for a range of pit sizes and bearings across the gearbox assembly. ROC and vibration were recorded simultaneously at high frequencies to provide a direct comparison between the two technologies. The seeded bearing defect dataset showed good agreement between ROC and accelerometer measurements, however accelerometers had more low-level noise and content more specific to the shaft on which they were mounted - ROC had a more balanced frequency spectrum, showing its ability to capture information across the entire assembly regardless of its placement. A range of Machine Learning studies were created to understand if models were capable of learning the distinction between damaged and healthy conditions from a simple frequency spectrum, ROC outperforming accelerometers by a maximum of 7% accuracy. A range of techniques were investigated, including automated feature extractions using autoencoders and unsupervised anomaly detection. The analysis of the real-world dataset showed that the frequency spectrum holds sufficient information to reliably detect bearing damage across an entire gearbox. Unsupervised techniques were demonstrated to show great potential in this field, leveraging the vast array of healthy data to learn deviations from normal operation with great sensitivity. The analysis of an existing gear defect dataset extended the work into machine learning and vast success was found in a process called Transfer Learning that used pretrained Convolutional Neural Networks. The process used Convolutional Wavelet I Condition Monitoring with Rate of Change of Torque Sensors Transforms to generate 2D images that would then be used to fine tune state-of-the-art image recognition models to detect the impulsive events related to gear damage in ROC signals. This was again extended to an unsupervised method, where a convolutional variational autoencoder was trained to understand the healthy condition, the reconstruction error resulting from the model was shown to be an excellent measure of damage severity. This thesis concludes that ROC is capable of capturing information across the entire gearbox to produce a reliable condition monitoring algorithm, with a clear benefit of ROC over traditional methods. Machine Learning can reliably identify defects that manual analysis cannot and the techniques developed have been proven on a real-world defect. This has shown great potential and the outstanding results of unsupervised learning shows that an intelligent condition monitoring algorithm can be deployed to the real-world without ever being trained on specific fault conditions. This thesis paves the way for future developments to start implementing a complete and robust online condition monitoring algorithm for the entire gearbox using ROC sensors.
Item Type: | Thesis (PhD) |
---|---|
Date Type: | Completion |
Status: | Unpublished |
Schools: | Engineering |
Uncontrolled Keywords: | . Gearbox condition monitoring 2). Rate-of-Change of Torque (ROC) 3). Unsupervised anomaly detection 4). Transfer Learning for damage detection 5). Autoencoders for feature extraction 6)._ Machine learning algorithms |
Date of First Compliant Deposit: | 22 May 2023 |
Last Modified: | 22 May 2023 14:55 |
URI: | https://orca.cardiff.ac.uk/id/eprint/159788 |
Actions (repository staff only)
![]() |
Edit Item |