Comparison Of Optimization Techniques Of Neural Networks Training For Faults Diagnostic Of Rotating Machinery.

Abstract

Recently sophisticated vibration monitoring techniques have been available to be
used in the monitoring and diagnostics of complexes rotating machinery. Among them, can
relate the artificial intelligence techniques as neural networks, fuzzy logic, expert systems and
so on. The neural networks are tools that have woken up a lot of interest on researchers in the
recent years. They let the monitoring on-line of predictive maintenance aiming the
minimization of the time between the receiving of the information and the diagnosis of the
problem. This paper shows the ability and feasibility of the application of different
optimization techniques of neural networks training in the diagnostic of faults inserted in the
rotating machinery. In the experimental setup are inserted the following faults: defect
electric, mechanical looseness, unbalance + mechanical looseness and unbalance. Several
architectures of neural networks implemented with the Matlab software were trained with
different optimization techniques to provide the best architecture to diagnostic of four faults
inserted in the experimental setup. Results show that the neural networks can be effectively
used in the diagnostic of faults inserted in the experimental set-up with a high performance
and that the Levenberg-Marquardt optimization technique is faster than gradient descent and
gradient descent with momentum for practical problems