Thesis Details

This work presents a computational model for fault identification in a gearbox setup of a commercial wind turbine. The signals used for model development were taken from a test conducted by the technology center CARTIF, to measure vibrations four piezoelectric accelerometers and a data acquisition module from National Instruments were used. Two coupled reducers simulate the gearbox on the test platform, therefore the accelerometers were placed orthogonally in parallel and planetary stage of the reducers.

The different conditions simulated in the test-bed were without fault, unbalance, misalignment and a failure with the maximum values of unbalance and misalignment. To extract features of different mechanical conditions, the computational model has two processing options, the first option is using linear prediction coefficients (LPC) and the second one makes use of the cepstral coefficients.

The analysis by LPC coefficients and the cepstrum have been used for the diagnosis of incipient failures in rotary machines. However, actually there have been no reported research that makes use of them for values of variable speed and load as the employed ones. The main motivation of this work focuses on the use of these methods in order to test their effectiveness in this type of conditions and to reduce the processing time which have other approaches reported in the literature.

Signal processing was performed using MATLAB with different methods and tests of the neural models, for the posterior comparison of the different studied methods both in processing time and the percentage of classification of different neural models, a multilayer perceptron and radial basis artificial network for LPC and cepstrum, respectively.

With the graphical environment software LabVIEW, virtual instruments that allow reading the vibration signals from the used database to diagnose the actual conditions of the analyzed gearbox were generated.

Honorable Mention - Given by the academic comitee evaluating the thesis.