Publication

Abstract : The detection of faults in industrial equipment through sound signals is considered to be a difficult task. This is the reason, why it triggers researchers and engineers to dig down into the problem and thus make use of different solutions. This research is aimed at identifying the breakdowns of the equipment in the early stages of major defects based on acoustic emission monitoring. This shall be done by an early detection and correction of the existing defects and by doing that the machines and other mechanical structures will work properly and longer. The system is equipped with a microphone using which sound from the machine is captured. In this research, we propose a system that is based on deep learning technology for the detection of faults in industrial engines where the extracted sounds are used as the source. In the data collection stage, we captured different machine sounds with and without anomalies, e.g. bearings, valves and sliders. These sounds almost constitute 10 classes which include the normal and fault/damaged sounds. Moreover, the dataset is entirely augmented to increase the amount of data in each class and provide variance to the data as there was only a limited number of files in the original dataset. Subsequently, these audio files in WAV format are processed at the feature extraction step using the Mel-Frequency Cepstral Coefficients (MFCC). To detect and classify the fault, RNN and CNN approaches were used, and the results were compared. The CNN approach came out with an exceptional accuracy of 96% while the RNN achieved an 85 % accuracy.