Facoltà: Ingegneria Gestionale - classe LM-31
Collocazione: 20001
Autore: Pierobon Tommaso
Data: 20/12/2021
Titolo: The sound of maintenance
Relatore: Pirovano Giovanni Luca
Autorizzazione per la consultazione: SI
Le tesi si possono consultare unicamente in sede
Productivity is a key strategy for manufacturing companies to stay competitive in a continuously growing global market. Higher availability can lead to higher productivity. To maintain a high level of availability, the maintenance of the production assets is essential. Sensors detecting predominantly vibration, temperature, pressure, and acceleration are commonly used in established predictive maintenance approaches to monitor system’s conditions and spot failure with critical parameters. Most of these solutions require deep integration with the monitored system, which implies high costs and significant installation efforts. Constraints imposed by the law are also crucial: the invasiveness of predictive maintenance approaches frequently violates warranty and leasing conditions of production machines. Thus, with these prerogatives, it is evident how a non-intrusive, lightweight and generic solution approach is best desired. In general, every system generates a unique acoustic footprint. In a production environment different operation stages, such as running, standby, offline, and failure, generate acoustic sound profiles that are typically recognisable by skilled and experienced operators. Starting from this point, the idea of developing a digital system able to recognise the state of a production machine based on the emitted sound was born and will be presented throughout this paper. This can have a lot of different applications: it can be used as a predictive maintenance tool or it can be applied as a performance KPI indicator. Moreover, microphones can record data without requiring a physical connection to the monitored production machine. In this way the approach is non-invasive, cheaper and, preserves the structural integrity of the observed system, addressing therefore the legal difficulties stated above. In order to achieve such a goal, a new dataset of machine’s operating sounds in a real production environment will be created by using an off-the-shelf microphone and a Raspberry Pi. At this point, three different machine learning frameworks based on this dataset will be proposed and a baseline will be defined. In particular, the MFEC was selected as the input of our models and we applied a Convolutional Autoencoder, a Random Forest and an SVM in different ways. Overall this study has shown encouraging results in relations to using sound to determine the conditions of industrial machinery. The proposed methodologies can be applied to different contexts where sound can provide useful information. Alternatively, some inherent characteristics of the machine, including its complexity, seem to have a considerable impact on the results obtained, making our algorithms non-marketable for predictive maintenance purposes but suitable as a performance KPI tool.
