< Terug naar vorige pagina

Publicatie

Unsupervised Machine Learning Methods to Estimate a Health Indicator for Condition Monitoring using Acoustic and Vibration Signals: A Comparison Based On a Toy Data Set From a Coffee Vending Machine

Boekbijdrage - Boekhoofdstuk Conferentiebijdrage

Automating the task of assessing an asset’s status based on sensor data would not only relieve trained engineers from this time intensive task, it would also allow a continuous follow-up of assets, potentially resulting in a fine-grained view on the asset’s status. In this work three unsupervised machine learning approaches that define a Health Indicator (HI) based on acoustic and vibration signals were empirically assessed. Such a HI indicates the similarity of the current measured state to the baseline/normal operational state. The lower the HI score the worse the asset’s condition. In this way the condition of an asset can be automatically monitored. Gaussian mixture models, Variational Autoencoders (VAE) and One Class Support Vector Machine (OC-SVM) were considered for this task. To enable the empirical assessment, a toy data set was created in which vibration and acoustic data was recorded simultaneously from a coffee vending machine with rotating elements in the bean grinder and water pump with relatively fast changing levels in the water and bean containers, and several stages in the coffee making cycle. Experiments were performed to analyse whether subtle changes in the sensor data due to changing container levels could be automatically detected and discriminated. Moreover, it was studied if a change could be rooted back to a cause (being a low level in the water or bean container). A set of temporal and spectral domain features were extracted and considered, while experiments were also performed by fusing the acoustic and vibration signals. The applied models achieved a comparable performance in terms of detecting low and empty container levels, with VAE using convolutional layers and OC-SVM achieving a further better discrimination of the different container levels when using the fused signals. It was also determined that the root cause of a level change can be determined by looking at the HI in the various stages.
Boek: Communications in Computer and Information Science
Pagina's: 145 - 159
Aantal pagina's: 15
ISBN:978-3-030-66770-2
Toegankelijkheid:Open