Machine diagnostics by neural networks (DTN.3584)

 The project aims at the development of a general, flexible machine monitoring and diagnosis system that can be used for predictive maintenance of mechanical equipment. The system will be based on the characteristic properties of the fingerprints in mainly the mechanical vibration patterns caused by the rotating and moving parts. A neural network classifier will be developed that indicates deviations from the normal state of operation and classifies some characteristic types of failure. The neural network will be fed with the spectra retrieved from a varietv of sensors. Vibration, position, temperature, sound and pressure are some of the accessible parameters that can be measured at varying intervals, normalized and compared to limits to reach an accurate assessment of the concept of 'machine health'. A first, PC-based prototype will be built that is interfaced to a set of sensors that yields a real time monitoring of a trained machine fingerprint.

During the development of the system a laboratory machine test bench will be used in which faults and artifacts can be generated artificially. The following aspects can be studied: loose foundation, unbalance, parallel and angular misalignment, incorrect gear meshing, bearing wear and effects of variable load. The key part of the project will be the development of a trainable neural network system integrated with the sensory devices covering issues such as architecture, input definition (signal preprocessing), trainability, output interpretation. Special hardware for interfacing, preprocessing and achieving the real time computational speed may be needed. A final evaluation on a real application is planned.

Veel meer informatie kan worden gevonden op de Website van de vakgroep

Er zijn 8 bedrijven bij dit project betrokken.

Dr.ir. R.P.W. Duin
Technische Universiteit Delft
Faculteit Technische Natuurwetenschappen,
subfaculteit Technische natuurkunde