Condition monitoring and fault diagnosis of gears with vibration signal and artificial neural networks for wear and broken tooth
Loading...
Date
item.page.authors
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
Fault diagnosis plays an important role in condition monitoring to
newlineenhance the machine time. In view of this, the present investigation
newlinefocused on the development of Fault diagnosis system of gearboxes
newlinebased on the vibration signatures and Artificial Neural Networks. In
newlinethe present investigation to generate the vibration signatures an
newlineexperimental set-up has been fabricated with sensing and measuring
newlineequipment. The two prominent faults, wear and broken tooth, of the
newlinegear were practically induced in the present investigation. Vibration
newlinesignatures of the gearbox were collected by transmitting the motion at
newlineconstant speed with gears having no fault, without applying any load.
newlineBy inducing one fault at a time, vibration signatures were collected
newlinewith different degrees of wear on a gear tooth and a gear with a broken
newlinetooth. As the vibration data of maximum amplitudes was found to be
newlineinseparable, fault diagnosis based on this data was not possible. Five
newlineprominent statistical features were extracted based on data pertaining
newlineto maximum amplitudes of vibration and used fault diagnosis. Due
newlineoverlapping of this data, it was decided to use ANN based fault
newlinediagnosis system for the present investigation. The set of statistical
newlinefeatures were extracted based on data pertaining to maximum
newlineamplitudes of vibration and used them as input parameters to the
newlineANN based fault diagnosis system designed and developed in
newlineMATLAB.
newlineThe ANN model was developed in two parts. The first part
newlinenamed as (5I-2O) fault classification system works with five input
newlineii
newlinefeatures and gives one of the two decision outputs as wear or broken
newlinetooth. It is a three layered feed forward network using back
newlinepropagation algorithm. The training parameters of the neural network,
newline(i) learning rate, (ii) number of hidden layer neurons and (iii) number
newlineof hidden layers, are carefully selected as they play a vital role in the
newlineconstruction of the model. The learning rate, number of hidden layer
newlineneurons were varied individually and optimum values were found
newlinebased on the number