Condition Monitoring of a Gear Box by Acoustic Camera and Machine Learning Techniques

In this paper the potentiality of an acoustic camera coupled with a machine learning algorithm to detect possible anomalies of an operating gear box is investigated. First, an experimental campaign was performed for different operating conditions (velocity, amplitude, frequency). During these phases the sound images were collected with the acoustic camera. This is followed by the pre-processing phase in which the acoustic images are prepared to train the network. The next step concerns the creation of a Convolution Neural Network (CNN) suitable for the classification of sound images. The last one involves training and testing of the network created. The analysis of the training plot and the confusion matrix show promising results. Most of the analyzed images are classified correctly with an overall accuracy of the model of 95%, despite the simplicity of the network created. Observing the excellent obtained results, this technique promises to be suitable for non-intrusive monitoring, allowing companies to reduce maintenance costs. The strength of this procedure is that, although the measurements are made in a noisy environment and not in an anechoic chamber, the Convolutional Neural Network is able to classify the images very well.

Automated summary

This paper investigates the potential of an acoustic camera coupled with a machine learning algorithm to detect anomalies in an operating gear box. Experimental data was collected under various operating conditions, and a Convolution Neural Network (CNN) was trained to classify the sound images. The results show promising accuracy of 95%, indicating the suitability of this technique for non-intrusive monitoring.