|
ABSTRACT. THIS PAPER DEALS WITH THE DEVELOPMENT OF A COMPUTATIONAL ALGORITHM FOR FAULT CLASSIFICATION IN GEARBOX, USING ACOUSTICAL NOISE. THIS SYSTEM WAS SET UP TO RECOGNISE THREE DIFFERENT PATTERNS INDEPENDENTLY OF THE GEARBOX SHAFT SPEED. A NEURAL NETWORK SYSTEM IS USED FOR THIS PURPOSE. TWO DIFFERENT TYPES OF CONVERGENCE ALGORITHM FOR THE TRAINING PHASE, NAMELY THE CONJUGATE GRADIENT AND MARQUARDT, ARE COMPARED. CONCERNING THE CLASSIFICATION TASK, TWO DIFFERENT STRATEGIES ARE DESCRIBED AND ADOPTED. AS PRE-PROCESSING METHODS WE USED A COMBINATION OF TWO STATISTICAL PARAMETERS (RMS AND KURTOSIS), AND A SPECTRAL REPRESENTATION. THE RESULTS SHOW THAT IS POSSIBLE TO OBTAIN A VERY HIGH RELIABILITY NETWORK SYSTEM TO CLASSIFICATION PURPOSE.
KEYWORDS: NEURAL NETWORKS, ACOUSTICS, FAULT CLASSIFICATION, GEARBOX
|
