Monitoring the health of tribocontacts requires the study of friction, tribofilm integrity, and wear transitions. This project will miniaturise existing sensing technology, with embedded electronics to overcome signal to noise issues, and use arrayed sensors for augmented sensing, and machine learning.
Condition monitoring and prognosis of machine health is attracting significant research interest, but it is becoming apparent that in order to get the best performance from such systems, the data on which such decisions is made must be of high quality. In this proposal we aim to answer the question “What kind of data is needed to give maximum impact to prognostic systems, and what requirements would the resulting processing system have?”
Fig.1 example bearing damages (a) the outer race (b) rolling elements and (c) inner race.
We believe that it is necessary to investigate measuring more fundamental wear processes in-situ, and therefore propose a vision that the surface of a mechanical component can be continuously monitored using a sensor array, similar in concept to the “fingerprint sensors” used in laptops and mobile phones. The moving surface of a bearing, for example, can be visualised as having a wear “fingerprint” represented by surface charge, repeating with each revolution, but evolving in time, which can be monitored with a linear array of charge sensors. This would be coupled to a learning system to track the evolution of specific features related to wear. This combination of a linear array of sensors feeding machine learning algorithms in order to monitor operational machines represents a new approach to condition monitoring.
The vision, therefore, is one of integrated arrays within tribological components, such as bearings, and of intelligent continuous direct measurement of surfaces which can inform a control system or operator to intervene to protect and extend the productivity and lifetime of machinery, as well as help identify the optimum efficiency operation zone. This project will progress beyond the traditional macro level physical sensors (vibration, speed, temperature and pressure) and will focus on developing capability which provides direct, in-situ measurement of component condition, rather than inferring from gross physical changes in the overall system, to inform the operator why, where and how to mitigate the defect once detected or how long operations can continue. This requires improved temporal and spatial resolution of tribological surfaces to detect evolution of unhealthy surface condition. It will also extend the understanding of the early degradation of tribological contacts from first principles (e.g. mechano-chemical-physics).
Using the friction and electrostatic measurements, an outlier score can be calculated using each stroke as a multivariable input to an outlier detection algorithm, five are used in the video above, feature regression being a simple linear method, the classic one class SVM as well as the more up to date LSTM and Autoencoder.
All give similar outlier score patterns.
