Diagnósticos de Fallos en Cajas de Engranajes Basado en una Red Neuronal Artificial Perceptrón Multicapa

This paper presents an algorithm for automated fault diagnostics in spur gearboxes, based in a multilayer perceptron (MLP) Artificial Neural Network (ANN). Four different fault classes were tested with different load and speed conditions: gear tooth breakage in three percentages, gear misalignment, pinion with face wear in one tooth, and pinion pitting. Vibration signals were acquired using an accelerometer and a data acquisition board. These were preprocessed using statistical measures from the time domain and frequency domain signal, also the signal spectrum was divided in energy bands so as to maintain its shape at the dominant peaks. Linear discriminant analysis (LDA) was used with the objective of reducing the dimensionality of input data to the ANN. Results obtained showed that a computationally efficient algorithm can be obtained due to the MLP ANN architecture, that makes diagnoses of proposed faults with a high relative margin of correct identification, which was evaluated by using the confusion matrix and ROC curves.