Gearbox fault and noise detection through audio analysis and deep learning

The achievement of high quality (Better-Best quality) and continuous research in minimizing waste (no Muda) are two of the fundamental points of the ideology of Lean production. ROJ.srl, careful to these considerations, has invested in this thesis for the research and design of an automated system able to test the quality of components purchased from external companies. In particular, the thesis focused on the defects and noise of the gearboxes purchased, aiming to have a system for the detection of gearboxes non-compliant, avoiding the use of these units in the assembly process and thus, improving the optimization of quality and waste of components at the same time. The concept behind this project is to develop a physical system able to catalogue, through audio analysis, the gearboxes tested. For this purpose, the first chapters of the paper report a brief introduction to the reducers, the defects in consideration and hints of sound and signal theory. These topics are useful for the understanding and realization of the project. Subsequently, the design of the test bench was developed, with the plan, to install it during the acceptance phase. For the software implementation, two different approaches were considered: The first one, based on classical programming, has allowed cataloguing, with very high precision, the reducers noisy and defective results using innovative methods for the detection of periodic defects. While, in the second approach, a convolution neural network (CNN) has been trained, with various methods and strategies, for detecting defective reducers. Also in this case, the obtained results were satisfactory, achieving a considerable reduction of the test time. Nevertheless, this solution led to a small loss in reliability, partially obviated, in a second moment, thanks to the implementation of a control method based on multiple frames. This last discussion has been very interesting in future development, as an onboard system could be implemented in such a way as to control, through predictive analysis, the occurrence of possible failures or defects.